Update

app.py

@@ -1,505 +1,508 @@
-import gradio as gr
-import numpy as np
-import os
-import torch
-import random
-
-from accelerate import infer_auto_device_map, load_checkpoint_and_dispatch, init_empty_weights
-from PIL import Image
-
-from data.data_utils import add_special_tokens, pil_img2rgb
-from data.transforms import ImageTransform
-from inferencer import InterleaveInferencer
-from modeling.autoencoder import load_ae
-from modeling.bagel.qwen2_navit import NaiveCache
-from modeling.bagel import (
-    BagelConfig, Bagel, Qwen2Config, Qwen2ForCausalLM,
-    SiglipVisionConfig, SiglipVisionModel
-)
-from modeling.qwen2 import Qwen2Tokenizer
-
-
-# Model Initialization
-model_path = "/path/to/BAGEL-7B-MoT/weights" #Download from https://huggingface.co/ByteDance-Seed/BAGEL-7B-MoT
-
-llm_config = Qwen2Config.from_json_file(os.path.join(model_path, "llm_config.json"))
-llm_config.qk_norm = True
-llm_config.tie_word_embeddings = False
-llm_config.layer_module = "Qwen2MoTDecoderLayer"
-
-vit_config = SiglipVisionConfig.from_json_file(os.path.join(model_path, "vit_config.json"))
-vit_config.rope = False
-vit_config.num_hidden_layers -= 1
-
-vae_model, vae_config = load_ae(local_path=os.path.join(model_path, "ae.safetensors"))
-
-config = BagelConfig(
-    visual_gen=True,
-    visual_und=True,
-    llm_config=llm_config, 
-    vit_config=vit_config,
-    vae_config=vae_config,
-    vit_max_num_patch_per_side=70,
-    connector_act='gelu_pytorch_tanh',
-    latent_patch_size=2,
-    max_latent_size=64,
-)
-
-with init_empty_weights():
-    language_model = Qwen2ForCausalLM(llm_config)
-    vit_model      = SiglipVisionModel(vit_config)
-    model          = Bagel(language_model, vit_model, config)
-    model.vit_model.vision_model.embeddings.convert_conv2d_to_linear(vit_config, meta=True)
-
-tokenizer = Qwen2Tokenizer.from_pretrained(model_path)
-tokenizer, new_token_ids, _ = add_special_tokens(tokenizer)
-
-vae_transform = ImageTransform(1024, 512, 16)
-vit_transform = ImageTransform(980, 224, 14)
-
-# Model Loading and Multi GPU Infernece Preparing
-device_map = infer_auto_device_map(
-    model,
-    max_memory={i: "80GiB" for i in range(torch.cuda.device_count())},
-    no_split_module_classes=["Bagel", "Qwen2MoTDecoderLayer"],
-)
-
-same_device_modules = [
-    'language_model.model.embed_tokens',
-    'time_embedder',
-    'latent_pos_embed',
-    'vae2llm',
-    'llm2vae',
-    'connector',
-    'vit_pos_embed'
-]
-
-if torch.cuda.device_count() == 1:
-    first_device = device_map.get(same_device_modules[0], "cuda:0")
-    for k in same_device_modules:
-        if k in device_map:
-            device_map[k] = first_device
-        else:
-            device_map[k] = "cuda:0"
-else:
-    first_device = device_map.get(same_device_modules[0])
-    for k in same_device_modules:
-        if k in device_map:
-            device_map[k] = first_device
-            
-model = load_checkpoint_and_dispatch(
-    model,
-    checkpoint=os.path.join(model_path, "ema.safetensors"),
-    device_map=device_map,
-    offload_buffers=True,
-    dtype=torch.bfloat16,
-    force_hooks=True,
-).eval()
-
-
-# Inferencer Preparing 
-inferencer = InterleaveInferencer(
-    model=model,
-    vae_model=vae_model,
-    tokenizer=tokenizer,
-    vae_transform=vae_transform,
-    vit_transform=vit_transform,
-    new_token_ids=new_token_ids,
-)
-
-def set_seed(seed):
-    """Set random seeds for reproducibility"""
-    if seed > 0:
-        random.seed(seed)
-        np.random.seed(seed)
-        torch.manual_seed(seed)
-        if torch.cuda.is_available():
-            torch.cuda.manual_seed(seed)
-            torch.cuda.manual_seed_all(seed)
-        torch.backends.cudnn.deterministic = True
-        torch.backends.cudnn.benchmark = False
-    return seed
-
-# Text to Image function with thinking option and hyperparameters
-def text_to_image(prompt, show_thinking=False, cfg_text_scale=4.0, cfg_interval=0.4, 
-                 timestep_shift=3.0, num_timesteps=50, 
-                 cfg_renorm_min=1.0, cfg_renorm_type="global", 
-                 max_think_token_n=1024, do_sample=False, text_temperature=0.3,
-                 seed=0, image_ratio="1:1"):
-    # Set seed for reproducibility
-    set_seed(seed)
-
-    if image_ratio == "1:1":
-        image_shapes = (1024, 1024)
-    elif image_ratio == "4:3":
-        image_shapes = (768, 1024)
-    elif image_ratio == "3:4":
-        image_shapes = (1024, 768) 
-    elif image_ratio == "16:9":
-        image_shapes = (576, 1024)
-    elif image_ratio == "9:16":
-        image_shapes = (1024, 576) 
-    
-    # Set hyperparameters
-    inference_hyper = dict(
-        max_think_token_n=max_think_token_n if show_thinking else 1024,
-        do_sample=do_sample if show_thinking else False,
-        text_temperature=text_temperature if show_thinking else 0.3,
-        cfg_text_scale=cfg_text_scale,
-        cfg_interval=[cfg_interval, 1.0],  # End fixed at 1.0
-        timestep_shift=timestep_shift,
-        num_timesteps=num_timesteps,
-        cfg_renorm_min=cfg_renorm_min,
-        cfg_renorm_type=cfg_renorm_type,
-        image_shapes=image_shapes,
-    )
-    
-    # Call inferencer with or without think parameter based on user choice
-    result = inferencer(text=prompt, think=show_thinking, **inference_hyper)
-    return result["image"], result.get("text", None)
-
-
-# Image Understanding function with thinking option and hyperparameters
-def image_understanding(image: Image.Image, prompt: str, show_thinking=False, 
-                        do_sample=False, text_temperature=0.3, max_new_tokens=512):
-    if image is None:
-        return "Please upload an image."
-
-    if isinstance(image, np.ndarray):
-        image = Image.fromarray(image)
-
-    image = pil_img2rgb(image)
-    
-    # Set hyperparameters
-    inference_hyper = dict(
-        do_sample=do_sample,
-        text_temperature=text_temperature,
-        max_think_token_n=max_new_tokens, # Set max_length
-    )
-    
-    # Use show_thinking parameter to control thinking process
-    result = inferencer(image=image, text=prompt, think=show_thinking, 
-                        understanding_output=True, **inference_hyper)
-    return result["text"]
-
-
-# Image Editing function with thinking option and hyperparameters
-def edit_image(image: Image.Image, prompt: str, show_thinking=False, cfg_text_scale=4.0, 
-              cfg_img_scale=2.0, cfg_interval=0.0, 
-              timestep_shift=3.0, num_timesteps=50, cfg_renorm_min=1.0, 
-              cfg_renorm_type="text_channel", max_think_token_n=1024, 
-              do_sample=False, text_temperature=0.3, seed=0):
-    # Set seed for reproducibility
-    set_seed(seed)
-    
-    if image is None:
-        return "Please upload an image.", ""
-
-    if isinstance(image, np.ndarray):
-        image = Image.fromarray(image)
-
-    image = pil_img2rgb(image)
-    
-    # Set hyperparameters
-    inference_hyper = dict(
-        max_think_token_n=max_think_token_n if show_thinking else 1024,
-        do_sample=do_sample if show_thinking else False,
-        text_temperature=text_temperature if show_thinking else 0.3,
-        cfg_text_scale=cfg_text_scale,
-        cfg_img_scale=cfg_img_scale,
-        cfg_interval=[cfg_interval, 1.0],  # End fixed at 1.0
-        timestep_shift=timestep_shift,
-        num_timesteps=num_timesteps,
-        cfg_renorm_min=cfg_renorm_min,
-        cfg_renorm_type=cfg_renorm_type,
-    )
-    
-    # Include thinking parameter based on user choice
-    result = inferencer(image=image, text=prompt, think=show_thinking, **inference_hyper)
-    return result["image"], result.get("text", "")
-
-
-# Helper function to load example images
-def load_example_image(image_path):
-    try:
-        return Image.open(image_path)
-    except Exception as e:
-        print(f"Error loading example image: {e}")
-        return None
-
-
-# Gradio UI 
-with gr.Blocks() as demo:
-    gr.Markdown("""
-<div>
-  <img src="https://lf3-static.bytednsdoc.com/obj/eden-cn/nuhojubrps/banner.png" alt="BAGEL" width="380"/>
-</div>
-""")
-
-    with gr.Tab("📝 Text to Image"):
-        txt_input = gr.Textbox(
-            label="Prompt", 
-            value="A female cosplayer portraying an ethereal fairy or elf, wearing a flowing dress made of delicate fabrics in soft, mystical colors like emerald green and silver. She has pointed ears, a gentle, enchanting expression, and her outfit is adorned with sparkling jewels and intricate patterns. The background is a magical forest with glowing plants, mystical creatures, and a serene atmosphere."
-        )
-        
-        with gr.Row():
-            show_thinking = gr.Checkbox(label="Thinking", value=False)
-        
-        # Add hyperparameter controls in an accordion
-        with gr.Accordion("Inference Hyperparameters", open=False):
-            # 参数一排两个布局
-            with gr.Group():
-                with gr.Row():
-                    seed = gr.Slider(minimum=0, maximum=1000000, value=0, step=1, 
-                                   label="Seed", info="0 for random seed, positive for reproducible results")
-                    image_ratio = gr.Dropdown(choices=["1:1", "4:3", "3:4", "16:9", "9:16"], 
-                                                value="1:1", label="Image Ratio", 
-                                                info="The longer size is fixed to 1024")
-                    
-                with gr.Row():
-                    cfg_text_scale = gr.Slider(minimum=1.0, maximum=8.0, value=4.0, step=0.1, interactive=True,
-                                             label="CFG Text Scale", info="Controls how strongly the model follows the text prompt (4.0-8.0)")
-                    cfg_interval = gr.Slider(minimum=0.0, maximum=1.0, value=0.4, step=0.1, 
-                                           label="CFG Interval", info="Start of CFG application interval (end is fixed at 1.0)")
-                
-                with gr.Row():
-                    cfg_renorm_type = gr.Dropdown(choices=["global", "local", "text_channel"], 
-                                                value="global", label="CFG Renorm Type", 
-                                                info="If the genrated image is blurry, use 'global'")
-                    cfg_renorm_min = gr.Slider(minimum=0.0, maximum=1.0, value=0.0, step=0.1, interactive=True,
-                                             label="CFG Renorm Min", info="1.0 disables CFG-Renorm")
-                
-                with gr.Row():
-                    num_timesteps = gr.Slider(minimum=10, maximum=100, value=50, step=5, interactive=True,
-                                            label="Timesteps", info="Total denoising steps")
-                    timestep_shift = gr.Slider(minimum=1.0, maximum=5.0, value=3.0, step=0.5, interactive=True,
-                                             label="Timestep Shift", info="Higher values for layout, lower for details")
-                
-                # Thinking parameters in a single row
-                thinking_params = gr.Group(visible=False)
-                with thinking_params:
-                    with gr.Row():
-                        do_sample = gr.Checkbox(label="Sampling", value=False, info="Enable sampling for text generation")
-                        max_think_token_n = gr.Slider(minimum=64, maximum=4006, value=1024, step=64, interactive=True,
-                                                    label="Max Think Tokens", info="Maximum number of tokens for thinking")
-                        text_temperature = gr.Slider(minimum=0.1, maximum=1.0, value=0.3, step=0.1, interactive=True,
-                                                  label="Temperature", info="Controls randomness in text generation")
-        
-        thinking_output = gr.Textbox(label="Thinking Process", visible=False)
-        img_output = gr.Image(label="Generated Image")
-        gen_btn = gr.Button("Generate")
-        
-        # Dynamically show/hide thinking process box and parameters
-        def update_thinking_visibility(show):
-            return gr.update(visible=show), gr.update(visible=show)
-        
-        show_thinking.change(
-            fn=update_thinking_visibility,
-            inputs=[show_thinking],
-            outputs=[thinking_output, thinking_params]
-        )
-        
-        # Process function based on thinking option and hyperparameters
-        def process_text_to_image(prompt, show_thinking, cfg_text_scale, 
-                                 cfg_interval, timestep_shift, 
-                                 num_timesteps, cfg_renorm_min, cfg_renorm_type, 
-                                 max_think_token_n, do_sample, text_temperature, seed, image_ratio):
-            image, thinking = text_to_image(
-                prompt, show_thinking, cfg_text_scale, cfg_interval,
-                timestep_shift, num_timesteps, 
-                cfg_renorm_min, cfg_renorm_type,
-                max_think_token_n, do_sample, text_temperature, seed, image_ratio
-            )
-            return image, thinking if thinking else ""
-        
-        gen_btn.click(
-            fn=process_text_to_image,
-            inputs=[
-                txt_input, show_thinking, cfg_text_scale, 
-                cfg_interval, timestep_shift, 
-                num_timesteps, cfg_renorm_min, cfg_renorm_type,
-                max_think_token_n, do_sample, text_temperature, seed, image_ratio
-            ],
-            outputs=[img_output, thinking_output]
-        )
-
-    with gr.Tab("🖌️ Image Edit"):
-        with gr.Row():
-            with gr.Column(scale=1):
-                edit_image_input = gr.Image(label="Input Image", value=load_example_image('test_images/women.jpg'))
-                edit_prompt = gr.Textbox(
-                    label="Prompt",
-                    value="She boards a modern subway, quietly reading a folded newspaper, wearing the same clothes."
-                )
-            
-            with gr.Column(scale=1):
-                edit_image_output = gr.Image(label="Result")
-                edit_thinking_output = gr.Textbox(label="Thinking Process", visible=False)
-        
-        with gr.Row():
-            edit_show_thinking = gr.Checkbox(label="Thinking", value=False)
-        
-        # Add hyperparameter controls in an accordion
-        with gr.Accordion("Inference Hyperparameters", open=False):
-            with gr.Group():
-                with gr.Row():
-                    edit_seed = gr.Slider(minimum=0, maximum=1000000, value=0, step=1, interactive=True,
-                                        label="Seed", info="0 for random seed, positive for reproducible results")
-                    edit_cfg_text_scale = gr.Slider(minimum=1.0, maximum=8.0, value=4.0, step=0.1, interactive=True,
-                                                  label="CFG Text Scale", info="Controls how strongly the model follows the text prompt")
-                
-                with gr.Row():
-                    edit_cfg_img_scale = gr.Slider(minimum=1.0, maximum=4.0, value=2.0, step=0.1, interactive=True,
-                                                 label="CFG Image Scale", info="Controls how much the model preserves input image details")
-                    edit_cfg_interval = gr.Slider(minimum=0.0, maximum=1.0, value=0.0, step=0.1, interactive=True,
-                                                label="CFG Interval", info="Start of CFG application interval (end is fixed at 1.0)")
-                    
-                with gr.Row():
-                    edit_cfg_renorm_type = gr.Dropdown(choices=["global", "local", "text_channel"], 
-                                                     value="text_channel", label="CFG Renorm Type", 
-                                                     info="If the genrated image is blurry, use 'global")
-                    edit_cfg_renorm_min = gr.Slider(minimum=0.0, maximum=1.0, value=0.0, step=0.1, interactive=True,
-                                                  label="CFG Renorm Min", info="1.0 disables CFG-Renorm")
-                
-                with gr.Row():
-                    edit_num_timesteps = gr.Slider(minimum=10, maximum=100, value=50, step=5, interactive=True,
-                                                 label="Timesteps", info="Total denoising steps")
-                    edit_timestep_shift = gr.Slider(minimum=1.0, maximum=10.0, value=3.0, step=0.5, interactive=True,
-                                                  label="Timestep Shift", info="Higher values for layout, lower for details")
-                
-                
-                # Thinking parameters in a single row
-                edit_thinking_params = gr.Group(visible=False)
-                with edit_thinking_params:
-                    with gr.Row():
-                        edit_do_sample = gr.Checkbox(label="Sampling", value=False, info="Enable sampling for text generation")
-                        edit_max_think_token_n = gr.Slider(minimum=64, maximum=4006, value=1024, step=64, interactive=True,
-                                                         label="Max Think Tokens", info="Maximum number of tokens for thinking")
-                        edit_text_temperature = gr.Slider(minimum=0.1, maximum=1.0, value=0.3, step=0.1, interactive=True,
-                                                        label="Temperature", info="Controls randomness in text generation")
-        
-        edit_btn = gr.Button("Submit")
-        
-        # Dynamically show/hide thinking process box for editing
-        def update_edit_thinking_visibility(show):
-            return gr.update(visible=show), gr.update(visible=show)
-        
-        edit_show_thinking.change(
-            fn=update_edit_thinking_visibility,
-            inputs=[edit_show_thinking],
-            outputs=[edit_thinking_output, edit_thinking_params]
-        )
-        
-        # Process editing with thinking option and hyperparameters
-        def process_edit_image(image, prompt, show_thinking, cfg_text_scale, 
-                              cfg_img_scale, cfg_interval, 
-                              timestep_shift, num_timesteps, cfg_renorm_min, 
-                              cfg_renorm_type, max_think_token_n, do_sample, 
-                              text_temperature, seed):
-            edited_image, thinking = edit_image(
-                image, prompt, show_thinking, cfg_text_scale, cfg_img_scale, 
-                cfg_interval, timestep_shift, 
-                num_timesteps, cfg_renorm_min, cfg_renorm_type,
-                max_think_token_n, do_sample, text_temperature, seed
-            )
-            
-            return edited_image, thinking if thinking else ""
-        
-        edit_btn.click(
-            fn=process_edit_image,
-            inputs=[
-                edit_image_input, edit_prompt, edit_show_thinking, 
-                edit_cfg_text_scale, edit_cfg_img_scale, edit_cfg_interval,
-                edit_timestep_shift, edit_num_timesteps, 
-                edit_cfg_renorm_min, edit_cfg_renorm_type,
-                edit_max_think_token_n, edit_do_sample, edit_text_temperature, edit_seed
-            ],
-            outputs=[edit_image_output, edit_thinking_output]
-        )
-
-    with gr.Tab("🖼️ Image Understanding"):
-        with gr.Row():
-            with gr.Column(scale=1):
-                img_input = gr.Image(label="Input Image", value=load_example_image('test_images/meme.jpg'))
-                understand_prompt = gr.Textbox(
-                    label="Prompt", 
-                    value="Can someone explain what's funny about this meme??"
-                )
-            
-            with gr.Column(scale=1):
-                txt_output = gr.Textbox(label="Result", lines=20)
-        
-        with gr.Row():
-            understand_show_thinking = gr.Checkbox(label="Thinking", value=False)
-        
-        # Add hyperparameter controls in an accordion
-        with gr.Accordion("Inference Hyperparameters", open=False):
-            with gr.Row():
-                understand_do_sample = gr.Checkbox(label="Sampling", value=False, info="Enable sampling for text generation")
-                understand_text_temperature = gr.Slider(minimum=0.0, maximum=1.0, value=0.3, step=0.05, interactive=True,
-                                                     label="Temperature", info="Controls randomness in text generation (0=deterministic, 1=creative)")
-                understand_max_new_tokens = gr.Slider(minimum=64, maximum=4096, value=512, step=64, interactive=True,
-                                                   label="Max New Tokens", info="Maximum length of generated text, including potential thinking")
-        
-        img_understand_btn = gr.Button("Submit")
-        
-        # Process understanding with thinking option and hyperparameters
-        def process_understanding(image, prompt, show_thinking, do_sample, 
-                                 text_temperature, max_new_tokens):
-            result = image_understanding(
-                image, prompt, show_thinking, do_sample, 
-                text_temperature, max_new_tokens
-            )
-            return result
-        
-        img_understand_btn.click(
-            fn=process_understanding,
-            inputs=[
-                img_input, understand_prompt, understand_show_thinking,
-                understand_do_sample, understand_text_temperature, understand_max_new_tokens
-            ],
-            outputs=txt_output
-        )
-
-    gr.Markdown("""
-<div style="display: flex; justify-content: flex-start; flex-wrap: wrap; gap: 10px;">
-  <a href="https://bagel-ai.org/">
-    <img
-      src="https://img.shields.io/badge/BAGEL-Website-0A66C2?logo=safari&logoColor=white"
-      alt="BAGEL Website"
-    />
-  </a>
-  <a href="https://arxiv.org/abs/2505.14683">
-    <img
-      src="https://img.shields.io/badge/BAGEL-Paper-red?logo=arxiv&logoColor=red"
-      alt="BAGEL Paper on arXiv"
-    />
-  </a>
-  <a href="https://huggingface.co/ByteDance-Seed/BAGEL-7B-MoT">
-    <img 
-        src="https://img.shields.io/badge/BAGEL-Hugging%20Face-orange?logo=huggingface&logoColor=yellow" 
-        alt="BAGEL on Hugging Face"
-    />
-  </a>
-  <a href="https://demo.bagel-ai.org/">
-    <img
-      src="https://img.shields.io/badge/BAGEL-Demo-blue?logo=googleplay&logoColor=blue"
-      alt="BAGEL Demo"
-    />
-  </a>
-  <a href="https://discord.gg/Z836xxzy">
-    <img
-      src="https://img.shields.io/badge/BAGEL-Discord-5865F2?logo=discord&logoColor=purple"
-      alt="BAGEL Discord"
-    />
-  </a>
-  <a href="mailto:[email protected]">
-    <img
-      src="https://img.shields.io/badge/BAGEL-Email-D14836?logo=gmail&logoColor=red"
-      alt="BAGEL Email"
-    />
-  </a>
-</div>
-""")
-
-demo.launch(share=True)
+import spaces
+import gradio as gr
+import numpy as np
+import os
+import torch
+import random
+import subprocess
+subprocess.run(
+    "pip install flash-attn --no-build-isolation",
+    env={"FLASH_ATTENTION_SKIP_CUDA_BUILD": "TRUE"},
+    shell=True,
+)
+
+from accelerate import infer_auto_device_map, load_checkpoint_and_dispatch, init_empty_weights
+from PIL import Image
+
+from data.data_utils import add_special_tokens, pil_img2rgb
+from data.transforms import ImageTransform
+from inferencer import InterleaveInferencer
+from modeling.autoencoder import load_ae
+from modeling.bagel.qwen2_navit import NaiveCache
+from modeling.bagel import (
+    BagelConfig, Bagel, Qwen2Config, Qwen2ForCausalLM,
+    SiglipVisionConfig, SiglipVisionModel
+)
+from modeling.qwen2 import Qwen2Tokenizer
+
+from huggingface_hub import snapshot_download
+
+save_dir = "./model"
+repo_id = "ByteDance-Seed/BAGEL-7B-MoT"
+cache_dir = save_dir + "/cache"
+
+snapshot_download(cache_dir=cache_dir,
+  local_dir=save_dir,
+  repo_id=repo_id,
+  local_dir_use_symlinks=False,
+  resume_download=True,
+  allow_patterns=["*.json", "*.safetensors", "*.bin", "*.py", "*.md", "*.txt"],
+)
+
+# Model Initialization
+model_path = "./model" #Download from https://huggingface.co/ByteDance-Seed/BAGEL-7B-MoT
+
+llm_config = Qwen2Config.from_json_file(os.path.join(model_path, "llm_config.json"))
+llm_config.qk_norm = True
+llm_config.tie_word_embeddings = False
+llm_config.layer_module = "Qwen2MoTDecoderLayer"
+
+vit_config = SiglipVisionConfig.from_json_file(os.path.join(model_path, "vit_config.json"))
+vit_config.rope = False
+vit_config.num_hidden_layers -= 1
+
+vae_model, vae_config = load_ae(local_path=os.path.join(model_path, "ae.safetensors"))
+
+config = BagelConfig(
+    visual_gen=True,
+    visual_und=True,
+    llm_config=llm_config, 
+    vit_config=vit_config,
+    vae_config=vae_config,
+    vit_max_num_patch_per_side=70,
+    connector_act='gelu_pytorch_tanh',
+    latent_patch_size=2,
+    max_latent_size=64,
+)
+
+with init_empty_weights():
+    language_model = Qwen2ForCausalLM(llm_config)
+    vit_model      = SiglipVisionModel(vit_config)
+    model          = Bagel(language_model, vit_model, config)
+    model.vit_model.vision_model.embeddings.convert_conv2d_to_linear(vit_config, meta=True)
+
+tokenizer = Qwen2Tokenizer.from_pretrained(model_path)
+tokenizer, new_token_ids, _ = add_special_tokens(tokenizer)
+
+vae_transform = ImageTransform(1024, 512, 16)
+vit_transform = ImageTransform(980, 224, 14)
+
+# Model Loading and Multi GPU Infernece Preparing
+device_map = infer_auto_device_map(
+    model,
+    max_memory={i: "80GiB" for i in range(torch.cuda.device_count())},
+    no_split_module_classes=["Bagel", "Qwen2MoTDecoderLayer"],
+)
+
+same_device_modules = [
+    'language_model.model.embed_tokens',
+    'time_embedder',
+    'latent_pos_embed',
+    'vae2llm',
+    'llm2vae',
+    'connector',
+    'vit_pos_embed'
+]
+
+if torch.cuda.device_count() == 1:
+    first_device = device_map.get(same_device_modules[0], "cuda:0")
+    for k in same_device_modules:
+        if k in device_map:
+            device_map[k] = first_device
+        else:
+            device_map[k] = "cuda:0"
+else:
+    first_device = device_map.get(same_device_modules[0])
+    for k in same_device_modules:
+        if k in device_map:
+            device_map[k] = first_device
+            
+model = load_checkpoint_and_dispatch(
+    model,
+    checkpoint=os.path.join(model_path, "ema.safetensors"),
+    device_map=device_map,
+    offload_buffers=True,
+    dtype=torch.bfloat16,
+    force_hooks=True,
+).eval()
+
+
+# Inferencer Preparing 
+inferencer = InterleaveInferencer(
+    model=model,
+    vae_model=vae_model,
+    tokenizer=tokenizer,
+    vae_transform=vae_transform,
+    vit_transform=vit_transform,
+    new_token_ids=new_token_ids,
+)
+
+def set_seed(seed):
+    """Set random seeds for reproducibility"""
+    if seed > 0:
+        random.seed(seed)
+        np.random.seed(seed)
+        torch.manual_seed(seed)
+        if torch.cuda.is_available():
+            torch.cuda.manual_seed(seed)
+            torch.cuda.manual_seed_all(seed)
+        torch.backends.cudnn.deterministic = True
+        torch.backends.cudnn.benchmark = False
+    return seed
+
+# Text to Image function with thinking option and hyperparameters
+@spaces.GPU(duration=90)
+def text_to_image(prompt, show_thinking=False, cfg_text_scale=4.0, cfg_interval=0.4, 
+                 timestep_shift=3.0, num_timesteps=50, 
+                 cfg_renorm_min=1.0, cfg_renorm_type="global", 
+                 max_think_token_n=1024, do_sample=False, text_temperature=0.3,
+                 seed=0, image_ratio="1:1"):
+    # Set seed for reproducibility
+    set_seed(seed)
+
+    if image_ratio == "1:1":
+        image_shapes = (1024, 1024)
+    elif image_ratio == "4:3":
+        image_shapes = (768, 1024)
+    elif image_ratio == "3:4":
+        image_shapes = (1024, 768) 
+    elif image_ratio == "16:9":
+        image_shapes = (576, 1024)
+    elif image_ratio == "9:16":
+        image_shapes = (1024, 576) 
+    
+    # Set hyperparameters
+    inference_hyper = dict(
+        max_think_token_n=max_think_token_n if show_thinking else 1024,
+        do_sample=do_sample if show_thinking else False,
+        text_temperature=text_temperature if show_thinking else 0.3,
+        cfg_text_scale=cfg_text_scale,
+        cfg_interval=[cfg_interval, 1.0],  # End fixed at 1.0
+        timestep_shift=timestep_shift,
+        num_timesteps=num_timesteps,
+        cfg_renorm_min=cfg_renorm_min,
+        cfg_renorm_type=cfg_renorm_type,
+        image_shapes=image_shapes,
+    )
+
+    result = {}
+    
+    # Call inferencer with or without think parameter based on user choice
+    for i in inferencer(text=prompt, think=show_thinking, **inference_hyper):
+        if type(i) == str:
+            result["text"] += i
+        elif type(i) == Image.Image:
+            result["image"] = i
+
+        yield result["image"], result.get("text", None)
+
+
+# Image Understanding function with thinking option and hyperparameters
+@spaces.GPU(duration=90)
+def image_understanding(image: Image.Image, prompt: str, show_thinking=False, 
+                        do_sample=False, text_temperature=0.3, max_new_tokens=512):
+    if image is None:
+        return "Please upload an image."
+
+    if isinstance(image, np.ndarray):
+        image = Image.fromarray(image)
+
+    image = pil_img2rgb(image)
+    
+    # Set hyperparameters
+    inference_hyper = dict(
+        do_sample=do_sample,
+        text_temperature=text_temperature,
+        max_think_token_n=max_new_tokens, # Set max_length
+    )
+    
+    result = {}
+    # Use show_thinking parameter to control thinking process
+    for i in inferencer(image=image, text=prompt, think=show_thinking, 
+                        understanding_output=True, **inference_hyper):
+        if type(i) == str:
+            result["text"] += i
+        elif type(i) == Image.Image:
+            result["image"] = i
+        yield result["text"]
+
+
+# Image Editing function with thinking option and hyperparameters
+@spaces.GPU(duration=90)
+def edit_image(image: Image.Image, prompt: str, show_thinking=False, cfg_text_scale=4.0, 
+              cfg_img_scale=2.0, cfg_interval=0.0, 
+              timestep_shift=3.0, num_timesteps=50, cfg_renorm_min=1.0, 
+              cfg_renorm_type="text_channel", max_think_token_n=1024, 
+              do_sample=False, text_temperature=0.3, seed=0):
+    # Set seed for reproducibility
+    set_seed(seed)
+    
+    if image is None:
+        return "Please upload an image.", ""
+
+    if isinstance(image, np.ndarray):
+        image = Image.fromarray(image)
+
+    image = pil_img2rgb(image)
+    
+    # Set hyperparameters
+    inference_hyper = dict(
+        max_think_token_n=max_think_token_n if show_thinking else 1024,
+        do_sample=do_sample if show_thinking else False,
+        text_temperature=text_temperature if show_thinking else 0.3,
+        cfg_text_scale=cfg_text_scale,
+        cfg_img_scale=cfg_img_scale,
+        cfg_interval=[cfg_interval, 1.0],  # End fixed at 1.0
+        timestep_shift=timestep_shift,
+        num_timesteps=num_timesteps,
+        cfg_renorm_min=cfg_renorm_min,
+        cfg_renorm_type=cfg_renorm_type,
+    )
+    
+    # Include thinking parameter based on user choice
+    result = {}
+    for i in inferencer(image=image, text=prompt, think=show_thinking, **inference_hyper):
+        if type(i) == str:
+            result["text"] += i
+        elif type(i) == Image.Image:
+            result["image"] = i
+
+        yield result["image"], result.get("text", "")
+
+# Helper function to load example images
+def load_example_image(image_path):
+    try:
+        return Image.open(image_path)
+    except Exception as e:
+        print(f"Error loading example image: {e}")
+        return None
+
+
+# Gradio UI 
+with gr.Blocks() as demo:
+    gr.Markdown("""
+<div>
+  <img src="https://lf3-static.bytednsdoc.com/obj/eden-cn/nuhojubrps/banner.png" alt="BAGEL" width="380"/>
+</div>
+""")
+
+    with gr.Tab("📝 Text to Image"):
+        txt_input = gr.Textbox(
+            label="Prompt", 
+            value="A female cosplayer portraying an ethereal fairy or elf, wearing a flowing dress made of delicate fabrics in soft, mystical colors like emerald green and silver. She has pointed ears, a gentle, enchanting expression, and her outfit is adorned with sparkling jewels and intricate patterns. The background is a magical forest with glowing plants, mystical creatures, and a serene atmosphere."
+        )
+        
+        with gr.Row():
+            show_thinking = gr.Checkbox(label="Thinking", value=False)
+        
+        # Add hyperparameter controls in an accordion
+        with gr.Accordion("Inference Hyperparameters", open=False):
+            # 参数一排两个布局
+            with gr.Group():
+                with gr.Row():
+                    seed = gr.Slider(minimum=0, maximum=1000000, value=0, step=1, 
+                                   label="Seed", info="0 for random seed, positive for reproducible results")
+                    image_ratio = gr.Dropdown(choices=["1:1", "4:3", "3:4", "16:9", "9:16"], 
+                                                value="1:1", label="Image Ratio", 
+                                                info="The longer size is fixed to 1024")
+                    
+                with gr.Row():
+                    cfg_text_scale = gr.Slider(minimum=1.0, maximum=8.0, value=4.0, step=0.1, interactive=True,
+                                             label="CFG Text Scale", info="Controls how strongly the model follows the text prompt (4.0-8.0)")
+                    cfg_interval = gr.Slider(minimum=0.0, maximum=1.0, value=0.4, step=0.1, 
+                                           label="CFG Interval", info="Start of CFG application interval (end is fixed at 1.0)")
+                
+                with gr.Row():
+                    cfg_renorm_type = gr.Dropdown(choices=["global", "local", "text_channel"], 
+                                                value="global", label="CFG Renorm Type", 
+                                                info="If the genrated image is blurry, use 'global'")
+                    cfg_renorm_min = gr.Slider(minimum=0.0, maximum=1.0, value=0.0, step=0.1, interactive=True,
+                                             label="CFG Renorm Min", info="1.0 disables CFG-Renorm")
+                
+                with gr.Row():
+                    num_timesteps = gr.Slider(minimum=10, maximum=100, value=50, step=5, interactive=True,
+                                            label="Timesteps", info="Total denoising steps")
+                    timestep_shift = gr.Slider(minimum=1.0, maximum=5.0, value=3.0, step=0.5, interactive=True,
+                                             label="Timestep Shift", info="Higher values for layout, lower for details")
+                
+                # Thinking parameters in a single row
+                thinking_params = gr.Group(visible=False)
+                with thinking_params:
+                    with gr.Row():
+                        do_sample = gr.Checkbox(label="Sampling", value=False, info="Enable sampling for text generation")
+                        max_think_token_n = gr.Slider(minimum=64, maximum=4006, value=1024, step=64, interactive=True,
+                                                    label="Max Think Tokens", info="Maximum number of tokens for thinking")
+                        text_temperature = gr.Slider(minimum=0.1, maximum=1.0, value=0.3, step=0.1, interactive=True,
+                                                  label="Temperature", info="Controls randomness in text generation")
+        
+        thinking_output = gr.Textbox(label="Thinking Process", visible=False)
+        img_output = gr.Image(label="Generated Image")
+        gen_btn = gr.Button("Generate")
+        
+        # Dynamically show/hide thinking process box and parameters
+        def update_thinking_visibility(show):
+            return gr.update(visible=show), gr.update(visible=show)
+        
+        show_thinking.change(
+            fn=update_thinking_visibility,
+            inputs=[show_thinking],
+            outputs=[thinking_output, thinking_params]
+        )
+        
+        gen_btn.click(
+            fn=text_to_image,
+            inputs=[
+                txt_input, show_thinking, cfg_text_scale, 
+                cfg_interval, timestep_shift, 
+                num_timesteps, cfg_renorm_min, cfg_renorm_type,
+                max_think_token_n, do_sample, text_temperature, seed, image_ratio
+            ],
+            outputs=[img_output, thinking_output]
+        )
+
+    with gr.Tab("🖌️ Image Edit"):
+        with gr.Row():
+            with gr.Column(scale=1):
+                edit_image_input = gr.Image(label="Input Image", value=load_example_image('test_images/women.jpg'))
+                edit_prompt = gr.Textbox(
+                    label="Prompt",
+                    value="She boards a modern subway, quietly reading a folded newspaper, wearing the same clothes."
+                )
+            
+            with gr.Column(scale=1):
+                edit_image_output = gr.Image(label="Result")
+                edit_thinking_output = gr.Textbox(label="Thinking Process", visible=False)
+        
+        with gr.Row():
+            edit_show_thinking = gr.Checkbox(label="Thinking", value=False)
+        
+        # Add hyperparameter controls in an accordion
+        with gr.Accordion("Inference Hyperparameters", open=False):
+            with gr.Group():
+                with gr.Row():
+                    edit_seed = gr.Slider(minimum=0, maximum=1000000, value=0, step=1, interactive=True,
+                                        label="Seed", info="0 for random seed, positive for reproducible results")
+                    edit_cfg_text_scale = gr.Slider(minimum=1.0, maximum=8.0, value=4.0, step=0.1, interactive=True,
+                                                  label="CFG Text Scale", info="Controls how strongly the model follows the text prompt")
+                
+                with gr.Row():
+                    edit_cfg_img_scale = gr.Slider(minimum=1.0, maximum=4.0, value=2.0, step=0.1, interactive=True,
+                                                 label="CFG Image Scale", info="Controls how much the model preserves input image details")
+                    edit_cfg_interval = gr.Slider(minimum=0.0, maximum=1.0, value=0.0, step=0.1, interactive=True,
+                                                label="CFG Interval", info="Start of CFG application interval (end is fixed at 1.0)")
+                    
+                with gr.Row():
+                    edit_cfg_renorm_type = gr.Dropdown(choices=["global", "local", "text_channel"], 
+                                                     value="text_channel", label="CFG Renorm Type", 
+                                                     info="If the genrated image is blurry, use 'global")
+                    edit_cfg_renorm_min = gr.Slider(minimum=0.0, maximum=1.0, value=0.0, step=0.1, interactive=True,
+                                                  label="CFG Renorm Min", info="1.0 disables CFG-Renorm")
+                
+                with gr.Row():
+                    edit_num_timesteps = gr.Slider(minimum=10, maximum=100, value=50, step=5, interactive=True,
+                                                 label="Timesteps", info="Total denoising steps")
+                    edit_timestep_shift = gr.Slider(minimum=1.0, maximum=10.0, value=3.0, step=0.5, interactive=True,
+                                                  label="Timestep Shift", info="Higher values for layout, lower for details")
+                
+                
+                # Thinking parameters in a single row
+                edit_thinking_params = gr.Group(visible=False)
+                with edit_thinking_params:
+                    with gr.Row():
+                        edit_do_sample = gr.Checkbox(label="Sampling", value=False, info="Enable sampling for text generation")
+                        edit_max_think_token_n = gr.Slider(minimum=64, maximum=4006, value=1024, step=64, interactive=True,
+                                                         label="Max Think Tokens", info="Maximum number of tokens for thinking")
+                        edit_text_temperature = gr.Slider(minimum=0.1, maximum=1.0, value=0.3, step=0.1, interactive=True,
+                                                        label="Temperature", info="Controls randomness in text generation")
+        
+        edit_btn = gr.Button("Submit")
+        
+        # Dynamically show/hide thinking process box for editing
+        def update_edit_thinking_visibility(show):
+            return gr.update(visible=show), gr.update(visible=show)
+        
+        edit_show_thinking.change(
+            fn=update_edit_thinking_visibility,
+            inputs=[edit_show_thinking],
+            outputs=[edit_thinking_output, edit_thinking_params]
+        )
+        
+        edit_btn.click(
+            fn=edit_image,
+            inputs=[
+                edit_image_input, edit_prompt, edit_show_thinking, 
+                edit_cfg_text_scale, edit_cfg_img_scale, edit_cfg_interval,
+                edit_timestep_shift, edit_num_timesteps, 
+                edit_cfg_renorm_min, edit_cfg_renorm_type,
+                edit_max_think_token_n, edit_do_sample, edit_text_temperature, edit_seed
+            ],
+            outputs=[edit_image_output, edit_thinking_output]
+        )
+
+    with gr.Tab("🖼️ Image Understanding"):
+        with gr.Row():
+            with gr.Column(scale=1):
+                img_input = gr.Image(label="Input Image", value=load_example_image('test_images/meme.jpg'))
+                understand_prompt = gr.Textbox(
+                    label="Prompt", 
+                    value="Can someone explain what's funny about this meme??"
+                )
+            
+            with gr.Column(scale=1):
+                txt_output = gr.Textbox(label="Result", lines=20)
+        
+        with gr.Row():
+            understand_show_thinking = gr.Checkbox(label="Thinking", value=False)
+        
+        # Add hyperparameter controls in an accordion
+        with gr.Accordion("Inference Hyperparameters", open=False):
+            with gr.Row():
+                understand_do_sample = gr.Checkbox(label="Sampling", value=False, info="Enable sampling for text generation")
+                understand_text_temperature = gr.Slider(minimum=0.0, maximum=1.0, value=0.3, step=0.05, interactive=True,
+                                                     label="Temperature", info="Controls randomness in text generation (0=deterministic, 1=creative)")
+                understand_max_new_tokens = gr.Slider(minimum=64, maximum=4096, value=512, step=64, interactive=True,
+                                                   label="Max New Tokens", info="Maximum length of generated text, including potential thinking")
+        
+        img_understand_btn = gr.Button("Submit")
+        
+        img_understand_btn.click(
+            fn=image_understanding,
+            inputs=[
+                img_input, understand_prompt, understand_show_thinking,
+                understand_do_sample, understand_text_temperature, understand_max_new_tokens
+            ],
+            outputs=txt_output
+        )
+
+    gr.Markdown("""
+<div style="display: flex; justify-content: flex-start; flex-wrap: wrap; gap: 10px;">
+  <a href="https://bagel-ai.org/">
+    <img
+      src="https://img.shields.io/badge/BAGEL-Website-0A66C2?logo=safari&logoColor=white"
+      alt="BAGEL Website"
+    />
+  </a>
+  <a href="https://arxiv.org/abs/2505.14683">
+    <img
+      src="https://img.shields.io/badge/BAGEL-Paper-red?logo=arxiv&logoColor=red"
+      alt="BAGEL Paper on arXiv"
+    />
+  </a>
+  <a href="https://huggingface.co/ByteDance-Seed/BAGEL-7B-MoT">
+    <img 
+        src="https://img.shields.io/badge/BAGEL-Hugging%20Face-orange?logo=huggingface&logoColor=yellow" 
+        alt="BAGEL on Hugging Face"
+    />
+  </a>
+  <a href="https://demo.bagel-ai.org/">
+    <img
+      src="https://img.shields.io/badge/BAGEL-Demo-blue?logo=googleplay&logoColor=blue"
+      alt="BAGEL Demo"
+    />
+  </a>
+  <a href="https://discord.gg/Z836xxzy">
+    <img
+      src="https://img.shields.io/badge/BAGEL-Discord-5865F2?logo=discord&logoColor=purple"
+      alt="BAGEL Discord"
+    />
+  </a>
+  <a href="mailto:[email protected]">
+    <img
+      src="https://img.shields.io/badge/BAGEL-Email-D14836?logo=gmail&logoColor=red"
+      alt="BAGEL Email"
+    />
+  </a>
+</div>
+""")
+
+demo.launch()

modeling/__init__.py

@@ -1,4 +1,4 @@
-# Copyright 2025 Bytedance Ltd. and/or its affiliates.
-# SPDX-License-Identifier: Apache-2.0
-
+# Copyright 2025 Bytedance Ltd. and/or its affiliates.
+# SPDX-License-Identifier: Apache-2.0
+
 from . import bagel, qwen2, siglip, autoencoder

modeling/autoencoder.py

@@ -1,361 +1,361 @@
-# Copyright (c) 2024 Black Forest Labs.
-# Copyright (c) 2025 Bytedance Ltd. and/or its affiliates.
-# SPDX-License-Identifier: Apache-2.0
-#
-# This file has been modified by ByteDance Ltd. and/or its affiliates. on 2025-05-20.
-#
-# Original file was released under Apache-2.0, with the full license text
-# available at https://github.com/black-forest-labs/flux/blob/main/LICENSE.
-#
-# This modified file is released under the same license.
-
-from dataclasses import dataclass
-
-import torch
-from einops import rearrange
-from torch import Tensor, nn
-from huggingface_hub import hf_hub_download
-from safetensors.torch import load_file as load_sft
-
-
-@dataclass
-class AutoEncoderParams:
-    resolution: int
-    in_channels: int
-    downsample: int
-    ch: int
-    out_ch: int
-    ch_mult: list[int]
-    num_res_blocks: int
-    z_channels: int
-    scale_factor: float
-    shift_factor: float
-
-
-def swish(x: Tensor) -> Tensor:
-    return x * torch.sigmoid(x)
-
-
-class AttnBlock(nn.Module):
-    def __init__(self, in_channels: int):
-        super().__init__()
-        self.in_channels = in_channels
-
-        self.norm = nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True)
-
-        self.q = nn.Conv2d(in_channels, in_channels, kernel_size=1)
-        self.k = nn.Conv2d(in_channels, in_channels, kernel_size=1)
-        self.v = nn.Conv2d(in_channels, in_channels, kernel_size=1)
-        self.proj_out = nn.Conv2d(in_channels, in_channels, kernel_size=1)
-
-    def attention(self, h_: Tensor) -> Tensor:
-        h_ = self.norm(h_)
-        q = self.q(h_)
-        k = self.k(h_)
-        v = self.v(h_)
-
-        b, c, h, w = q.shape
-        q = rearrange(q, "b c h w -> b 1 (h w) c").contiguous()
-        k = rearrange(k, "b c h w -> b 1 (h w) c").contiguous()
-        v = rearrange(v, "b c h w -> b 1 (h w) c").contiguous()
-        h_ = nn.functional.scaled_dot_product_attention(q, k, v)
-
-        return rearrange(h_, "b 1 (h w) c -> b c h w", h=h, w=w, c=c, b=b)
-
-    def forward(self, x: Tensor) -> Tensor:
-        return x + self.proj_out(self.attention(x))
-
-
-class ResnetBlock(nn.Module):
-    def __init__(self, in_channels: int, out_channels: int):
-        super().__init__()
-        self.in_channels = in_channels
-        out_channels = in_channels if out_channels is None else out_channels
-        self.out_channels = out_channels
-
-        self.norm1 = nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True)
-        self.conv1 = nn.Conv2d(in_channels, out_channels, kernel_size=3, stride=1, padding=1)
-        self.norm2 = nn.GroupNorm(num_groups=32, num_channels=out_channels, eps=1e-6, affine=True)
-        self.conv2 = nn.Conv2d(out_channels, out_channels, kernel_size=3, stride=1, padding=1)
-        if self.in_channels != self.out_channels:
-            self.nin_shortcut = nn.Conv2d(in_channels, out_channels, kernel_size=1, stride=1, padding=0)
-
-    def forward(self, x):
-        h = x
-        h = self.norm1(h)
-        h = swish(h)
-        h = self.conv1(h)
-
-        h = self.norm2(h)
-        h = swish(h)
-        h = self.conv2(h)
-
-        if self.in_channels != self.out_channels:
-            x = self.nin_shortcut(x)
-
-        return x + h
-
-
-class Downsample(nn.Module):
-    def __init__(self, in_channels: int):
-        super().__init__()
-        # no asymmetric padding in torch conv, must do it ourselves
-        self.conv = nn.Conv2d(in_channels, in_channels, kernel_size=3, stride=2, padding=0)
-
-    def forward(self, x: Tensor):
-        pad = (0, 1, 0, 1)
-        x = nn.functional.pad(x, pad, mode="constant", value=0)
-        x = self.conv(x)
-        return x
-
-
-class Upsample(nn.Module):
-    def __init__(self, in_channels: int):
-        super().__init__()
-        self.conv = nn.Conv2d(in_channels, in_channels, kernel_size=3, stride=1, padding=1)
-
-    def forward(self, x: Tensor):
-        x = nn.functional.interpolate(x, scale_factor=2.0, mode="nearest")
-        x = self.conv(x)
-        return x
-
-
-class Encoder(nn.Module):
-    def __init__(
-        self,
-        resolution: int,
-        in_channels: int,
-        ch: int,
-        ch_mult: list[int],
-        num_res_blocks: int,
-        z_channels: int,
-    ):
-        super().__init__()
-        self.ch = ch
-        self.num_resolutions = len(ch_mult)
-        self.num_res_blocks = num_res_blocks
-        self.resolution = resolution
-        self.in_channels = in_channels
-        # downsampling
-        self.conv_in = nn.Conv2d(in_channels, self.ch, kernel_size=3, stride=1, padding=1)
-
-        curr_res = resolution
-        in_ch_mult = (1,) + tuple(ch_mult)
-        self.in_ch_mult = in_ch_mult
-        self.down = nn.ModuleList()
-        block_in = self.ch
-        for i_level in range(self.num_resolutions):
-            block = nn.ModuleList()
-            attn = nn.ModuleList()
-            block_in = ch * in_ch_mult[i_level]
-            block_out = ch * ch_mult[i_level]
-            for _ in range(self.num_res_blocks):
-                block.append(ResnetBlock(in_channels=block_in, out_channels=block_out))
-                block_in = block_out
-            down = nn.Module()
-            down.block = block
-            down.attn = attn
-            if i_level != self.num_resolutions - 1:
-                down.downsample = Downsample(block_in)
-                curr_res = curr_res // 2
-            self.down.append(down)
-
-        # middle
-        self.mid = nn.Module()
-        self.mid.block_1 = ResnetBlock(in_channels=block_in, out_channels=block_in)
-        self.mid.attn_1 = AttnBlock(block_in)
-        self.mid.block_2 = ResnetBlock(in_channels=block_in, out_channels=block_in)
-
-        # end
-        self.norm_out = nn.GroupNorm(num_groups=32, num_channels=block_in, eps=1e-6, affine=True)
-        self.conv_out = nn.Conv2d(block_in, 2 * z_channels, kernel_size=3, stride=1, padding=1)
-
-    def forward(self, x: Tensor) -> Tensor:
-        # downsampling
-        hs = [self.conv_in(x)]
-        for i_level in range(self.num_resolutions):
-            for i_block in range(self.num_res_blocks):
-                h = self.down[i_level].block[i_block](hs[-1])
-                if len(self.down[i_level].attn) > 0:
-                    h = self.down[i_level].attn[i_block](h)
-                hs.append(h)
-            if i_level != self.num_resolutions - 1:
-                hs.append(self.down[i_level].downsample(hs[-1]))
-
-        # middle
-        h = hs[-1]
-        h = self.mid.block_1(h)
-        h = self.mid.attn_1(h)
-        h = self.mid.block_2(h)
-        # end
-        h = self.norm_out(h)
-        h = swish(h)
-        h = self.conv_out(h)
-        return h
-
-
-class Decoder(nn.Module):
-    def __init__(
-        self,
-        ch: int,
-        out_ch: int,
-        ch_mult: list[int],
-        num_res_blocks: int,
-        in_channels: int,
-        resolution: int,
-        z_channels: int,
-    ):
-        super().__init__()
-        self.ch = ch
-        self.num_resolutions = len(ch_mult)
-        self.num_res_blocks = num_res_blocks
-        self.resolution = resolution
-        self.in_channels = in_channels
-        self.ffactor = 2 ** (self.num_resolutions - 1)
-
-        # compute in_ch_mult, block_in and curr_res at lowest res
-        block_in = ch * ch_mult[self.num_resolutions - 1]
-        curr_res = resolution // 2 ** (self.num_resolutions - 1)
-        self.z_shape = (1, z_channels, curr_res, curr_res)
-
-        # z to block_in
-        self.conv_in = nn.Conv2d(z_channels, block_in, kernel_size=3, stride=1, padding=1)
-
-        # middle
-        self.mid = nn.Module()
-        self.mid.block_1 = ResnetBlock(in_channels=block_in, out_channels=block_in)
-        self.mid.attn_1 = AttnBlock(block_in)
-        self.mid.block_2 = ResnetBlock(in_channels=block_in, out_channels=block_in)
-
-        # upsampling
-        self.up = nn.ModuleList()
-        for i_level in reversed(range(self.num_resolutions)):
-            block = nn.ModuleList()
-            attn = nn.ModuleList()
-            block_out = ch * ch_mult[i_level]
-            for _ in range(self.num_res_blocks + 1):
-                block.append(ResnetBlock(in_channels=block_in, out_channels=block_out))
-                block_in = block_out
-            up = nn.Module()
-            up.block = block
-            up.attn = attn
-            if i_level != 0:
-                up.upsample = Upsample(block_in)
-                curr_res = curr_res * 2
-            self.up.insert(0, up)  # prepend to get consistent order
-
-        # end
-        self.norm_out = nn.GroupNorm(num_groups=32, num_channels=block_in, eps=1e-6, affine=True)
-        self.conv_out = nn.Conv2d(block_in, out_ch, kernel_size=3, stride=1, padding=1)
-
-    def forward(self, z: Tensor) -> Tensor:
-        # z to block_in
-        h = self.conv_in(z)
-
-        # middle
-        h = self.mid.block_1(h)
-        h = self.mid.attn_1(h)
-        h = self.mid.block_2(h)
-
-        # upsampling
-        for i_level in reversed(range(self.num_resolutions)):
-            for i_block in range(self.num_res_blocks + 1):
-                h = self.up[i_level].block[i_block](h)
-                if len(self.up[i_level].attn) > 0:
-                    h = self.up[i_level].attn[i_block](h)
-            if i_level != 0:
-                h = self.up[i_level].upsample(h)
-
-        # end
-        h = self.norm_out(h)
-        h = swish(h)
-        h = self.conv_out(h)
-        return h
-
-
-class DiagonalGaussian(nn.Module):
-    def __init__(self, sample: bool = True, chunk_dim: int = 1):
-        super().__init__()
-        self.sample = sample
-        self.chunk_dim = chunk_dim
-
-    def forward(self, z: Tensor) -> Tensor:
-        mean, logvar = torch.chunk(z, 2, dim=self.chunk_dim)
-        if self.sample:
-            std = torch.exp(0.5 * logvar)
-            return mean + std * torch.randn_like(mean)
-        else:
-            return mean
-
-
-class AutoEncoder(nn.Module):
-    def __init__(self, params: AutoEncoderParams):
-        super().__init__()
-        self.encoder = Encoder(
-            resolution=params.resolution,
-            in_channels=params.in_channels,
-            ch=params.ch,
-            ch_mult=params.ch_mult,
-            num_res_blocks=params.num_res_blocks,
-            z_channels=params.z_channels,
-        )
-        self.decoder = Decoder(
-            resolution=params.resolution,
-            in_channels=params.in_channels,
-            ch=params.ch,
-            out_ch=params.out_ch,
-            ch_mult=params.ch_mult,
-            num_res_blocks=params.num_res_blocks,
-            z_channels=params.z_channels,
-        )
-        self.reg = DiagonalGaussian()
-
-        self.scale_factor = params.scale_factor
-        self.shift_factor = params.shift_factor
-
-    def encode(self, x: Tensor) -> Tensor:
-        z = self.reg(self.encoder(x))
-        z = self.scale_factor * (z - self.shift_factor)
-        return z
-
-    def decode(self, z: Tensor) -> Tensor:
-        z = z / self.scale_factor + self.shift_factor
-        return self.decoder(z)
-
-    def forward(self, x: Tensor) -> Tensor:
-        return self.decode(self.encode(x))
-
-
-def print_load_warning(missing: list[str], unexpected: list[str]) -> None:
-    if len(missing) > 0 and len(unexpected) > 0:
-        print(f"Got {len(missing)} missing keys:\n\t" + "\n\t".join(missing))
-        print("\n" + "-" * 79 + "\n")
-        print(f"Got {len(unexpected)} unexpected keys:\n\t" + "\n\t".join(unexpected))
-    elif len(missing) > 0:
-        print(f"Got {len(missing)} missing keys:\n\t" + "\n\t".join(missing))
-    elif len(unexpected) > 0:
-        print(f"Got {len(unexpected)} unexpected keys:\n\t" + "\n\t".join(unexpected))
-
-
-def load_ae(local_path: str) -> AutoEncoder:
-    ae_params = AutoEncoderParams(
-            resolution=256,
-            in_channels=3,
-            downsample=8,
-            ch=128,
-            out_ch=3,
-            ch_mult=[1, 2, 4, 4],
-            num_res_blocks=2,
-            z_channels=16,
-            scale_factor=0.3611,
-            shift_factor=0.1159,
-    )
-
-    # Loading the autoencoder
-    ae = AutoEncoder(ae_params)
-
-    if local_path is not None:
-        sd = load_sft(local_path)
-        missing, unexpected = ae.load_state_dict(sd, strict=False, assign=True)
-        print_load_warning(missing, unexpected)
-    return ae, ae_params
+# Copyright (c) 2024 Black Forest Labs.
+# Copyright (c) 2025 Bytedance Ltd. and/or its affiliates.
+# SPDX-License-Identifier: Apache-2.0
+#
+# This file has been modified by ByteDance Ltd. and/or its affiliates. on 2025-05-20.
+#
+# Original file was released under Apache-2.0, with the full license text
+# available at https://github.com/black-forest-labs/flux/blob/main/LICENSE.
+#
+# This modified file is released under the same license.
+
+from dataclasses import dataclass
+
+import torch
+from einops import rearrange
+from torch import Tensor, nn
+from huggingface_hub import hf_hub_download
+from safetensors.torch import load_file as load_sft
+
+
+@dataclass
+class AutoEncoderParams:
+    resolution: int
+    in_channels: int
+    downsample: int
+    ch: int
+    out_ch: int
+    ch_mult: list[int]
+    num_res_blocks: int
+    z_channels: int
+    scale_factor: float
+    shift_factor: float
+
+
+def swish(x: Tensor) -> Tensor:
+    return x * torch.sigmoid(x)
+
+
+class AttnBlock(nn.Module):
+    def __init__(self, in_channels: int):
+        super().__init__()
+        self.in_channels = in_channels
+
+        self.norm = nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True)
+
+        self.q = nn.Conv2d(in_channels, in_channels, kernel_size=1)
+        self.k = nn.Conv2d(in_channels, in_channels, kernel_size=1)
+        self.v = nn.Conv2d(in_channels, in_channels, kernel_size=1)
+        self.proj_out = nn.Conv2d(in_channels, in_channels, kernel_size=1)
+
+    def attention(self, h_: Tensor) -> Tensor:
+        h_ = self.norm(h_)
+        q = self.q(h_)
+        k = self.k(h_)
+        v = self.v(h_)
+
+        b, c, h, w = q.shape
+        q = rearrange(q, "b c h w -> b 1 (h w) c").contiguous()
+        k = rearrange(k, "b c h w -> b 1 (h w) c").contiguous()
+        v = rearrange(v, "b c h w -> b 1 (h w) c").contiguous()
+        h_ = nn.functional.scaled_dot_product_attention(q, k, v)
+
+        return rearrange(h_, "b 1 (h w) c -> b c h w", h=h, w=w, c=c, b=b)
+
+    def forward(self, x: Tensor) -> Tensor:
+        return x + self.proj_out(self.attention(x))
+
+
+class ResnetBlock(nn.Module):
+    def __init__(self, in_channels: int, out_channels: int):
+        super().__init__()
+        self.in_channels = in_channels
+        out_channels = in_channels if out_channels is None else out_channels
+        self.out_channels = out_channels
+
+        self.norm1 = nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True)
+        self.conv1 = nn.Conv2d(in_channels, out_channels, kernel_size=3, stride=1, padding=1)
+        self.norm2 = nn.GroupNorm(num_groups=32, num_channels=out_channels, eps=1e-6, affine=True)
+        self.conv2 = nn.Conv2d(out_channels, out_channels, kernel_size=3, stride=1, padding=1)
+        if self.in_channels != self.out_channels:
+            self.nin_shortcut = nn.Conv2d(in_channels, out_channels, kernel_size=1, stride=1, padding=0)
+
+    def forward(self, x):
+        h = x
+        h = self.norm1(h)
+        h = swish(h)
+        h = self.conv1(h)
+
+        h = self.norm2(h)
+        h = swish(h)
+        h = self.conv2(h)
+
+        if self.in_channels != self.out_channels:
+            x = self.nin_shortcut(x)
+
+        return x + h
+
+
+class Downsample(nn.Module):
+    def __init__(self, in_channels: int):
+        super().__init__()
+        # no asymmetric padding in torch conv, must do it ourselves
+        self.conv = nn.Conv2d(in_channels, in_channels, kernel_size=3, stride=2, padding=0)
+
+    def forward(self, x: Tensor):
+        pad = (0, 1, 0, 1)
+        x = nn.functional.pad(x, pad, mode="constant", value=0)
+        x = self.conv(x)
+        return x
+
+
+class Upsample(nn.Module):
+    def __init__(self, in_channels: int):
+        super().__init__()
+        self.conv = nn.Conv2d(in_channels, in_channels, kernel_size=3, stride=1, padding=1)
+
+    def forward(self, x: Tensor):
+        x = nn.functional.interpolate(x, scale_factor=2.0, mode="nearest")
+        x = self.conv(x)
+        return x
+
+
+class Encoder(nn.Module):
+    def __init__(
+        self,
+        resolution: int,
+        in_channels: int,
+        ch: int,
+        ch_mult: list[int],
+        num_res_blocks: int,
+        z_channels: int,
+    ):
+        super().__init__()
+        self.ch = ch
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.resolution = resolution
+        self.in_channels = in_channels
+        # downsampling
+        self.conv_in = nn.Conv2d(in_channels, self.ch, kernel_size=3, stride=1, padding=1)
+
+        curr_res = resolution
+        in_ch_mult = (1,) + tuple(ch_mult)
+        self.in_ch_mult = in_ch_mult
+        self.down = nn.ModuleList()
+        block_in = self.ch
+        for i_level in range(self.num_resolutions):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_in = ch * in_ch_mult[i_level]
+            block_out = ch * ch_mult[i_level]
+            for _ in range(self.num_res_blocks):
+                block.append(ResnetBlock(in_channels=block_in, out_channels=block_out))
+                block_in = block_out
+            down = nn.Module()
+            down.block = block
+            down.attn = attn
+            if i_level != self.num_resolutions - 1:
+                down.downsample = Downsample(block_in)
+                curr_res = curr_res // 2
+            self.down.append(down)
+
+        # middle
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock(in_channels=block_in, out_channels=block_in)
+        self.mid.attn_1 = AttnBlock(block_in)
+        self.mid.block_2 = ResnetBlock(in_channels=block_in, out_channels=block_in)
+
+        # end
+        self.norm_out = nn.GroupNorm(num_groups=32, num_channels=block_in, eps=1e-6, affine=True)
+        self.conv_out = nn.Conv2d(block_in, 2 * z_channels, kernel_size=3, stride=1, padding=1)
+
+    def forward(self, x: Tensor) -> Tensor:
+        # downsampling
+        hs = [self.conv_in(x)]
+        for i_level in range(self.num_resolutions):
+            for i_block in range(self.num_res_blocks):
+                h = self.down[i_level].block[i_block](hs[-1])
+                if len(self.down[i_level].attn) > 0:
+                    h = self.down[i_level].attn[i_block](h)
+                hs.append(h)
+            if i_level != self.num_resolutions - 1:
+                hs.append(self.down[i_level].downsample(hs[-1]))
+
+        # middle
+        h = hs[-1]
+        h = self.mid.block_1(h)
+        h = self.mid.attn_1(h)
+        h = self.mid.block_2(h)
+        # end
+        h = self.norm_out(h)
+        h = swish(h)
+        h = self.conv_out(h)
+        return h
+
+
+class Decoder(nn.Module):
+    def __init__(
+        self,
+        ch: int,
+        out_ch: int,
+        ch_mult: list[int],
+        num_res_blocks: int,
+        in_channels: int,
+        resolution: int,
+        z_channels: int,
+    ):
+        super().__init__()
+        self.ch = ch
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.resolution = resolution
+        self.in_channels = in_channels
+        self.ffactor = 2 ** (self.num_resolutions - 1)
+
+        # compute in_ch_mult, block_in and curr_res at lowest res
+        block_in = ch * ch_mult[self.num_resolutions - 1]
+        curr_res = resolution // 2 ** (self.num_resolutions - 1)
+        self.z_shape = (1, z_channels, curr_res, curr_res)
+
+        # z to block_in
+        self.conv_in = nn.Conv2d(z_channels, block_in, kernel_size=3, stride=1, padding=1)
+
+        # middle
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock(in_channels=block_in, out_channels=block_in)
+        self.mid.attn_1 = AttnBlock(block_in)
+        self.mid.block_2 = ResnetBlock(in_channels=block_in, out_channels=block_in)
+
+        # upsampling
+        self.up = nn.ModuleList()
+        for i_level in reversed(range(self.num_resolutions)):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_out = ch * ch_mult[i_level]
+            for _ in range(self.num_res_blocks + 1):
+                block.append(ResnetBlock(in_channels=block_in, out_channels=block_out))
+                block_in = block_out
+            up = nn.Module()
+            up.block = block
+            up.attn = attn
+            if i_level != 0:
+                up.upsample = Upsample(block_in)
+                curr_res = curr_res * 2
+            self.up.insert(0, up)  # prepend to get consistent order
+
+        # end
+        self.norm_out = nn.GroupNorm(num_groups=32, num_channels=block_in, eps=1e-6, affine=True)
+        self.conv_out = nn.Conv2d(block_in, out_ch, kernel_size=3, stride=1, padding=1)
+
+    def forward(self, z: Tensor) -> Tensor:
+        # z to block_in
+        h = self.conv_in(z)
+
+        # middle
+        h = self.mid.block_1(h)
+        h = self.mid.attn_1(h)
+        h = self.mid.block_2(h)
+
+        # upsampling
+        for i_level in reversed(range(self.num_resolutions)):
+            for i_block in range(self.num_res_blocks + 1):
+                h = self.up[i_level].block[i_block](h)
+                if len(self.up[i_level].attn) > 0:
+                    h = self.up[i_level].attn[i_block](h)
+            if i_level != 0:
+                h = self.up[i_level].upsample(h)
+
+        # end
+        h = self.norm_out(h)
+        h = swish(h)
+        h = self.conv_out(h)
+        return h
+
+
+class DiagonalGaussian(nn.Module):
+    def __init__(self, sample: bool = True, chunk_dim: int = 1):
+        super().__init__()
+        self.sample = sample
+        self.chunk_dim = chunk_dim
+
+    def forward(self, z: Tensor) -> Tensor:
+        mean, logvar = torch.chunk(z, 2, dim=self.chunk_dim)
+        if self.sample:
+            std = torch.exp(0.5 * logvar)
+            return mean + std * torch.randn_like(mean)
+        else:
+            return mean
+
+
+class AutoEncoder(nn.Module):
+    def __init__(self, params: AutoEncoderParams):
+        super().__init__()
+        self.encoder = Encoder(
+            resolution=params.resolution,
+            in_channels=params.in_channels,
+            ch=params.ch,
+            ch_mult=params.ch_mult,
+            num_res_blocks=params.num_res_blocks,
+            z_channels=params.z_channels,
+        )
+        self.decoder = Decoder(
+            resolution=params.resolution,
+            in_channels=params.in_channels,
+            ch=params.ch,
+            out_ch=params.out_ch,
+            ch_mult=params.ch_mult,
+            num_res_blocks=params.num_res_blocks,
+            z_channels=params.z_channels,
+        )
+        self.reg = DiagonalGaussian()
+
+        self.scale_factor = params.scale_factor
+        self.shift_factor = params.shift_factor
+
+    def encode(self, x: Tensor) -> Tensor:
+        z = self.reg(self.encoder(x))
+        z = self.scale_factor * (z - self.shift_factor)
+        return z
+
+    def decode(self, z: Tensor) -> Tensor:
+        z = z / self.scale_factor + self.shift_factor
+        return self.decoder(z)
+
+    def forward(self, x: Tensor) -> Tensor:
+        return self.decode(self.encode(x))
+
+
+def print_load_warning(missing: list[str], unexpected: list[str]) -> None:
+    if len(missing) > 0 and len(unexpected) > 0:
+        print(f"Got {len(missing)} missing keys:\n\t" + "\n\t".join(missing))
+        print("\n" + "-" * 79 + "\n")
+        print(f"Got {len(unexpected)} unexpected keys:\n\t" + "\n\t".join(unexpected))
+    elif len(missing) > 0:
+        print(f"Got {len(missing)} missing keys:\n\t" + "\n\t".join(missing))
+    elif len(unexpected) > 0:
+        print(f"Got {len(unexpected)} unexpected keys:\n\t" + "\n\t".join(unexpected))
+
+
+def load_ae(local_path: str) -> AutoEncoder:
+    ae_params = AutoEncoderParams(
+            resolution=256,
+            in_channels=3,
+            downsample=8,
+            ch=128,
+            out_ch=3,
+            ch_mult=[1, 2, 4, 4],
+            num_res_blocks=2,
+            z_channels=16,
+            scale_factor=0.3611,
+            shift_factor=0.1159,
+    )
+
+    # Loading the autoencoder
+    ae = AutoEncoder(ae_params)
+
+    if local_path is not None:
+        sd = load_sft(local_path)
+        missing, unexpected = ae.load_state_dict(sd, strict=False, assign=True)
+        print_load_warning(missing, unexpected)
+    return ae, ae_params

modeling/bagel/__init__.py

@@ -1,18 +1,18 @@
-# Copyright 2025 Bytedance Ltd. and/or its affiliates.
-# SPDX-License-Identifier: Apache-2.0
-
-
-from .bagel import BagelConfig, Bagel
-from .qwen2_navit import Qwen2Config, Qwen2Model, Qwen2ForCausalLM
-from .siglip_navit import SiglipVisionConfig, SiglipVisionModel
-
-
-__all__ = [
-    'BagelConfig',
-    'Bagel',
-    'Qwen2Config',
-    'Qwen2Model', 
-    'Qwen2ForCausalLM',
-    'SiglipVisionConfig',
-    'SiglipVisionModel',
-]
+# Copyright 2025 Bytedance Ltd. and/or its affiliates.
+# SPDX-License-Identifier: Apache-2.0
+
+
+from .bagel import BagelConfig, Bagel
+from .qwen2_navit import Qwen2Config, Qwen2Model, Qwen2ForCausalLM
+from .siglip_navit import SiglipVisionConfig, SiglipVisionModel
+
+
+__all__ = [
+    'BagelConfig',
+    'Bagel',
+    'Qwen2Config',
+    'Qwen2Model', 
+    'Qwen2ForCausalLM',
+    'SiglipVisionConfig',
+    'SiglipVisionModel',
+]

modeling/bagel/bagel.py

@@ -1,1026 +1,1026 @@
-# Copyright 2025 Bytedance Ltd. and/or its affiliates.
-# SPDX-License-Identifier: Apache-2.0
-
-import copy
-from typing import List, Tuple, Optional
-import matplotlib.pyplot as plt
-
-from PIL import Image
-import torch
-import torch.nn.functional as F
-from torch import nn
-from torch.nn.attention.flex_attention import create_block_mask
-from transformers.configuration_utils import PretrainedConfig
-from transformers.modeling_utils import PreTrainedModel
-
-from data.data_utils import (
-    create_sparse_mask, 
-    get_flattened_position_ids_extrapolate, 
-    get_flattened_position_ids_interpolate,
-    patchify, 
-)
-from .qwen2_navit import NaiveCache
-from .modeling_utils import MLPconnector, TimestepEmbedder, PositionEmbedding
-
-
-class BagelConfig(PretrainedConfig):
-    def __init__(
-        self,
-        visual_gen=True,
-        visual_und=True,
-        llm_config=None,
-        vit_config=None,
-        vae_config=None,
-        latent_patch_size=2,
-        max_latent_size=32,
-        vit_max_num_patch_per_side=70,
-        connector_act="gelu_pytorch_tanh",
-        interpolate_pos=False,
-        timestep_shift=1.0,
-        **kwargs
-    ):
-        super().__init__(**kwargs)
-        self.visual_gen = visual_gen
-        self.visual_und = visual_und
-        self.llm_config = llm_config
-        self.vit_config = vit_config
-        self.vae_config = vae_config
-        self.latent_patch_size = latent_patch_size
-        self.max_latent_size = max_latent_size
-        self.vit_max_num_patch_per_side = vit_max_num_patch_per_side
-        self.connector_act = connector_act
-        self.interpolate_pos = interpolate_pos
-        self.timestep_shift = timestep_shift
-
-
-class Bagel(PreTrainedModel):
-    config_class = BagelConfig
-    base_model_prefix = 'bagel'
-
-    def __init__(self, language_model, vit_model, config: BagelConfig):
-        super().__init__(config)    
-        self.language_model = language_model
-        self.hidden_size = config.llm_config.hidden_size
-        self.use_moe = "Mo" in config.llm_config.layer_module
-        self.num_heads = config.llm_config.num_attention_heads
-
-        if config.visual_gen:
-            self.latent_patch_size = config.latent_patch_size
-            self.timestep_shift = config.timestep_shift
-            self.latent_downsample = config.vae_config.downsample * config.latent_patch_size
-            self.max_latent_size = config.max_latent_size
-            self.latent_channel = config.vae_config.z_channels
-            self.patch_latent_dim = self.latent_patch_size ** 2 * self.latent_channel
-            self.time_embedder = TimestepEmbedder(self.hidden_size)
-            self.vae2llm = nn.Linear(self.patch_latent_dim, self.hidden_size)
-            self.llm2vae = nn.Linear(self.hidden_size, self.patch_latent_dim)
-            self.latent_pos_embed = PositionEmbedding(self.max_latent_size, self.hidden_size)
-
-        if config.visual_und:
-            self.vit_model = vit_model
-            self.vit_patch_size = config.vit_config.patch_size
-            self.vit_max_num_patch_per_side = config.vit_max_num_patch_per_side
-            self.vit_hidden_size = config.vit_config.hidden_size
-            self.connector = MLPconnector(self.vit_hidden_size, self.hidden_size, config.connector_act)
-            self.vit_pos_embed = PositionEmbedding(self.vit_max_num_patch_per_side, self.hidden_size)
-
-        if config.interpolate_pos:
-            self.get_flattened_position_ids = get_flattened_position_ids_interpolate
-        else:
-            self.get_flattened_position_ids = get_flattened_position_ids_extrapolate
-
-        self.config = config
-        self._init_weights()
-
-    def _init_weights(self):
-        if self.config.visual_gen:
-            nn.init.constant_(self.llm2vae.weight, 0)
-            nn.init.constant_(self.llm2vae.bias, 0)
-
-    def forward(
-        self,
-        sequence_length: int,
-        packed_text_ids: torch.LongTensor,
-        packed_text_indexes: torch.LongTensor,
-        sample_lens: List[int],
-        packed_position_ids: torch.LongTensor,
-        nested_attention_masks: List[torch.Tensor] = None,
-        split_lens: List[int] = None,
-        attn_modes: List[str] = None,
-        # for visual understanding
-        ce_loss_indexes: Optional[torch.BoolTensor] = None,
-        packed_label_ids: Optional[torch.LongTensor] = None,
-        packed_vit_tokens: Optional[torch.Tensor] = None,
-        packed_vit_token_indexes: Optional[torch.LongTensor] = None,
-        packed_vit_position_ids: Optional[torch.LongTensor] = None,
-        vit_token_seqlens: Optional[torch.IntTensor] = None,
-        # for visual generation
-        padded_latent: Optional[torch.Tensor] = None,
-        patchified_vae_latent_shapes: Optional[List[Tuple[int, int]]] = None,
-        packed_latent_position_ids: Optional[torch.LongTensor] = None,
-        packed_vae_token_indexes: Optional[torch.LongTensor] = None,
-        packed_timesteps: Optional[torch.LongTensor] = None,
-        mse_loss_indexes: Optional[torch.BoolTensor] = None,
-    ) -> torch.Tensor:
-        """
-        Args:
-            sequence_length: length of sequence.
-            packed_text_ids: 1-D int tensor, packed text token ids.
-            packed_text_indexes: 1-D int tensor, packed text token indexes in sequence.
-            sample_lens: A list of N ints, length of each sample in packed_sequence.
-            nested_attention_masks: A list of N 2-D float tensor,  where 0.0 means attention and 
-                -inf means ignore.
-            packed_position_ids: packed 1-D positions, an image has only one global position shared
-                by all latent tokens.
-
-            packed_vit_tokens: packed patchified image tokens for vit model.
-            packed_vit_position_ids: 1-D int tensor, the position of each token for vit model.
-            packed_vit_token_indexes: 1-D int tensor, packed vit token indexes in sequence.
-            vit_token_seqlens: 1-D int tensor, the length of each image tokens for vit model.
-            packed_label_ids: 1-D int tensor, packed label token ids.
-            ce_loss_indexes: 1-D bool tensor, where to compute ce loss.
-
-            padded_latent: padded latent from VAE encoder.
-            patchified_vae_latent_shapes: A list of (h, w) tuples, patchfied latent shapes of each image.
-            packed_latent_position_ids: 1-D int tensor, the position of each token for latent.
-            packed_vae_token_indexes: 1-D int tensor, padded image token indexes in sequence.
-            packed_timesteps: 1-D float tensor, flow timesteps. 0 indicates use clean image.
-            mse_loss_indexes: 1-D bool tensor, where to compute mse loss.
-        """
-        packed_text_embedding = self.language_model.model.embed_tokens(packed_text_ids)
-        packed_sequence = packed_text_embedding.new_zeros(size=(sequence_length, self.hidden_size))
-        packed_sequence[packed_text_indexes] = packed_text_embedding
-
-        if nested_attention_masks is None:
-            sparse_mask = create_sparse_mask(sample_lens, split_lens, attn_modes, packed_text_embedding.device)
-            seqlen = sum(sample_lens)
-            block_mask = create_block_mask(
-                sparse_mask, B=1, H=self.num_heads, Q_LEN=seqlen, KV_LEN=seqlen, 
-                device=packed_text_embedding.device, BLOCK_SIZE=128, _compile=True
-            )
-            attention_mask = block_mask
-        else:
-            attention_mask = nested_attention_masks
-
-        if self.config.visual_und:
-            cu_seqlens = torch.nn.functional.pad(torch.cumsum(vit_token_seqlens, dim=0), (1, 0))
-            cu_seqlens = cu_seqlens.to(torch.int32)
-            max_seqlen = torch.max(vit_token_seqlens).item()
-            packed_vit_token_embed = self.vit_model(
-                packed_pixel_values=packed_vit_tokens, 
-                packed_flattened_position_ids=packed_vit_position_ids,
-                cu_seqlens=cu_seqlens,
-                max_seqlen=max_seqlen,
-            )
-            packed_vit_token_embed = self.connector(packed_vit_token_embed)
-            vit_token_pos_emb = self.vit_pos_embed(packed_vit_position_ids)
-            packed_vit_token_embed = packed_vit_token_embed + vit_token_pos_emb
-            packed_sequence[packed_vit_token_indexes] = packed_vit_token_embed
-
-        if self.config.visual_gen:
-            p = self.latent_patch_size
-            packed_latent = []
-            for latent, (h, w) in zip(padded_latent, patchified_vae_latent_shapes):
-                latent = latent[:, :h * p, :w * p].reshape(self.latent_channel, h, p, w, p)
-                latent = torch.einsum("chpwq->hwpqc", latent).reshape(-1, p * p * self.latent_channel)
-                packed_latent.append(latent)
-            packed_latent_clean = torch.cat(packed_latent, dim=0)
-
-            noise = torch.randn_like(packed_latent_clean)
-            packed_timesteps = torch.sigmoid(packed_timesteps)
-            packed_timesteps = self.timestep_shift * packed_timesteps / (1 + (self.timestep_shift - 1) * packed_timesteps)
-            packed_latent = (1 - packed_timesteps[:, None]) * packed_latent_clean + packed_timesteps[:, None] * noise
-            packed_timestep_embeds = self.time_embedder(packed_timesteps)
-            latent_token_pos_emb = self.latent_pos_embed(packed_latent_position_ids)
-            packed_latent = self.vae2llm(packed_latent) + packed_timestep_embeds + latent_token_pos_emb
-            packed_sequence[packed_vae_token_indexes] = packed_latent
-
-        extra_inputs = {}
-        if self.use_moe:
-            packed_und_token_indexes = packed_text_indexes
-            if packed_vit_token_indexes is not None:
-                packed_und_token_indexes=torch.cat([packed_text_indexes, packed_vit_token_indexes], dim=0)
-            extra_inputs.update(
-                packed_und_token_indexes=packed_und_token_indexes,
-                packed_gen_token_indexes=packed_vae_token_indexes,
-            )
-
-        last_hidden_state = self.language_model(
-            packed_sequence=packed_sequence,
-            sample_lens=sample_lens,
-            attention_mask=attention_mask,
-            packed_position_ids=packed_position_ids,
-            **extra_inputs,
-        )
-
-        mse = None
-        if self.config.visual_gen:
-            packed_mse_preds = self.llm2vae(last_hidden_state[mse_loss_indexes])
-            target = noise - packed_latent_clean # NOTE: v_t=dx_t/dt=x_1-x_0, pointing from data to noise
-            has_mse = packed_timesteps > 0
-            mse = (packed_mse_preds - target[has_mse]) ** 2
-
-        ce = None
-        if ce_loss_indexes is not None:
-            packed_ce_preds = self.language_model.lm_head(last_hidden_state[ce_loss_indexes])
-            ce = F.cross_entropy(packed_ce_preds, packed_label_ids, reduction="none")
-
-        return dict(mse=mse, ce=ce)
-
-
-    def prepare_prompts(self, curr_kvlens, curr_rope, prompts, tokenizer, new_token_ids):
-        packed_text_ids = list()
-        packed_text_position_ids = list()
-        text_token_lens = list()
-        packed_text_indexes = list()
-        packed_key_value_indexes = list()
-
-        curr = 0
-        newlens, new_rope = list(), list()
-        for prompt, curr_kvlen, curr_position_id in zip(prompts, curr_kvlens, curr_rope):
-            packed_key_value_indexes.extend(range(curr, curr + curr_kvlen))
-            curr += curr_kvlen
-
-            text_ids = tokenizer.encode(prompt)
-            text_ids = [new_token_ids['bos_token_id']] + text_ids + [new_token_ids['eos_token_id']]
-            text_token_lens.append(len(text_ids))
-            packed_text_ids.extend(text_ids)
-            packed_text_position_ids.extend(range(curr_position_id, curr_position_id + len(text_ids)))
-            packed_text_indexes.extend(range(curr, curr + len(text_ids)))
-            newlens.append(curr_kvlen + len(text_ids))
-            new_rope.append(curr_position_id + len(text_ids))
-            curr += len(text_ids)
-
-        generation_input = {
-            "text_token_lens": torch.tensor(text_token_lens, dtype=torch.int),
-            "packed_text_ids": torch.tensor(packed_text_ids, dtype=torch.long),
-            "packed_text_position_ids": torch.tensor(packed_text_position_ids, dtype=torch.long),
-            "packed_text_indexes": torch.tensor(packed_text_indexes, dtype=torch.long),
-            "packed_key_value_indexes": torch.tensor(packed_key_value_indexes, dtype=torch.long),
-            "key_values_lens": torch.tensor(curr_kvlens, dtype=torch.int),
-        }
-
-        return generation_input, newlens, new_rope
-
-    @torch.no_grad
-    def forward_cache_update_text(
-        self,
-        past_key_values: NaiveCache,
-        packed_text_ids: torch.IntTensor,
-        packed_text_position_ids: torch.LongTensor,
-        text_token_lens: torch.LongTensor,
-        packed_text_indexes: torch.LongTensor,
-        packed_key_value_indexes: torch.LongTensor,
-        key_values_lens: torch.IntTensor,
-    ):
-        packed_text_embedding = self.language_model.model.embed_tokens(packed_text_ids)
-
-        extra_inputs = {}
-        if self.use_moe:
-            extra_inputs = {"mode": "und"}
-
-        output = self.language_model.forward_inference(
-            packed_query_sequence=packed_text_embedding,
-            query_lens=text_token_lens,
-            packed_query_position_ids=packed_text_position_ids,
-            packed_query_indexes=packed_text_indexes,
-            past_key_values=past_key_values,
-            packed_key_value_indexes=packed_key_value_indexes,
-            key_values_lens=key_values_lens,
-            update_past_key_values=True,
-            is_causal=True,
-            **extra_inputs,
-        )
-        past_key_values = output.past_key_values
-
-        return past_key_values
-
-    def prepare_vit_images(self, curr_kvlens, curr_rope, images, transforms, new_token_ids):
-        packed_vit_token_indexes = list()
-        vit_token_seqlens, packed_vit_tokens, packed_vit_position_ids = list(), list(), list()
-        packed_text_ids, packed_text_indexes = list(), list()
-        packed_seqlens, packed_position_ids, packed_indexes = list(), list(), list()
-        packed_key_value_indexes = list()
-
-        _curr = curr = 0
-        newlens, new_rope = list(), list()
-        for image, curr_kvlen, curr_position_id in zip(images, curr_kvlens, curr_rope):
-            packed_key_value_indexes.extend(range(curr, curr + curr_kvlen))
-            curr += curr_kvlen
-
-            packed_text_ids.append(new_token_ids['start_of_image'])
-            packed_text_indexes.append(_curr)
-            packed_indexes.append(curr)
-            curr += 1
-            _curr += 1
-
-            image_tensor = transforms(image)
-            vit_position_ids = self.get_flattened_position_ids(
-                image_tensor.size(1), image_tensor.size(2), 
-                self.vit_patch_size, 
-                max_num_patches_per_side=self.vit_max_num_patch_per_side
-            )
-            vit_tokens = patchify(image_tensor, self.vit_patch_size)
-            packed_vit_tokens.append(vit_tokens)
-            num_img_tokens = vit_tokens.shape[0]
-            packed_vit_position_ids.append(vit_position_ids)
-            vit_token_seqlens.append(num_img_tokens)
-            packed_vit_token_indexes.extend(range(_curr, _curr + num_img_tokens))
-            packed_indexes.extend(range(curr, curr + num_img_tokens))
-            curr += num_img_tokens
-            _curr += num_img_tokens
-
-            packed_text_ids.append(new_token_ids['end_of_image'])
-            packed_text_indexes.append(_curr)
-            packed_indexes.append(curr)
-            curr += 1
-            _curr += 1
-
-            packed_position_ids.extend([curr_position_id] * (num_img_tokens + 2))
-            packed_seqlens.append(num_img_tokens + 2)
-            newlens.append(curr_kvlen + num_img_tokens + 2)
-            new_rope.append(curr_position_id + 1)
-
-        generation_input = {
-            "packed_text_ids": torch.tensor(packed_text_ids, dtype=torch.long),
-            "packed_text_indexes": torch.tensor(packed_text_indexes, dtype=torch.long),
-            "vit_token_seqlens": torch.tensor(vit_token_seqlens, dtype=torch.int),
-            "packed_vit_tokens": torch.cat(packed_vit_tokens, dim=0),
-            "packed_vit_position_ids": torch.cat(packed_vit_position_ids, dim=0),
-            "packed_vit_token_indexes": torch.tensor(packed_vit_token_indexes, dtype=torch.long),
-            "packed_position_ids": torch.tensor(packed_position_ids, dtype=torch.long),
-            "packed_seqlens": torch.tensor(packed_seqlens, dtype=torch.int),
-            "packed_indexes": torch.tensor(packed_indexes, dtype=torch.long),
-            "packed_key_value_indexes": torch.tensor(packed_key_value_indexes, dtype=torch.long),
-            "key_values_lens": torch.tensor(curr_kvlens, dtype=torch.int),
-        }
-
-        return generation_input, newlens, new_rope
-
-    @torch.no_grad
-    def forward_cache_update_vit(
-        self,
-        past_key_values: NaiveCache,
-        packed_text_ids: torch.LongTensor,
-        packed_text_indexes: torch.LongTensor,
-        packed_vit_tokens: torch.Tensor,
-        packed_vit_token_indexes: torch.LongTensor,
-        packed_vit_position_ids: torch.LongTensor,
-        vit_token_seqlens: torch.IntTensor,
-        packed_position_ids: torch.LongTensor,
-        packed_seqlens: torch.IntTensor,
-        packed_indexes: torch.LongTensor,
-        packed_key_value_indexes: torch.LongTensor,
-        key_values_lens: torch.IntTensor,
-    ):
-        packed_text_embedding = self.language_model.model.embed_tokens(packed_text_ids)
-        packed_sequence = packed_text_embedding.new_zeros((sum(packed_seqlens), self.hidden_size))
-        packed_sequence[packed_text_indexes] = packed_text_embedding
-
-        cu_seqlens = torch.nn.functional.pad(torch.cumsum(vit_token_seqlens, dim=0), (1, 0))
-        cu_seqlens = cu_seqlens.to(torch.int32)
-        max_seqlen = torch.max(vit_token_seqlens).item()
-        packed_vit_token_embed = self.vit_model(
-            packed_pixel_values=packed_vit_tokens, 
-            packed_flattened_position_ids=packed_vit_position_ids,
-            cu_seqlens=cu_seqlens,
-            max_seqlen=max_seqlen,
-        )
-        packed_vit_token_embed = self.connector(packed_vit_token_embed)
-        pos_emb = self.vit_pos_embed(packed_vit_position_ids)
-        packed_vit_token_embed = packed_vit_token_embed + pos_emb
-        packed_sequence[packed_vit_token_indexes] = packed_vit_token_embed
-
-        extra_inputs = {}
-        if self.use_moe:
-            extra_inputs = {"mode": "und"}
-
-        output = self.language_model.forward_inference(
-            packed_query_sequence=packed_sequence,
-            query_lens=packed_seqlens,
-            packed_query_position_ids=packed_position_ids,
-            packed_query_indexes=packed_indexes,
-            past_key_values=past_key_values,
-            packed_key_value_indexes=packed_key_value_indexes,
-            key_values_lens=key_values_lens,
-            update_past_key_values=True,
-            is_causal=False,
-            **extra_inputs,
-        )
-        past_key_values = output.past_key_values
-
-        return past_key_values
-
-    def prepare_vae_images(self, curr_kvlens, curr_rope, images, transforms, new_token_ids, timestep=0):
-        patchified_vae_latent_shapes, packed_vae_position_ids = list(), list()
-        packed_vae_token_indexes = list()
-        packed_text_ids, packed_text_indexes = list(), list()
-        packed_seqlens, packed_position_ids, packed_indexes = list(), list(), list()
-        packed_key_value_indexes = list()
-
-        _curr = curr = 0
-        vae_image_tensors = list()
-        newlens, new_rope = list(), list()
-        for image, curr_kvlen, curr_position_id in zip(images, curr_kvlens, curr_rope):
-            packed_key_value_indexes.extend(range(curr, curr + curr_kvlen))
-            curr += curr_kvlen
-
-            packed_text_ids.append(new_token_ids['start_of_image'])
-            packed_text_indexes.append(_curr)
-            packed_indexes.append(curr)
-            curr += 1
-            _curr += 1
-
-            image_tensor = transforms(image)
-            vae_image_tensors.append(image_tensor)
-            vae_posiiton_ids = self.get_flattened_position_ids(
-                image_tensor.size(1), image_tensor.size(2),
-                self.latent_downsample, 
-                max_num_patches_per_side=self.max_latent_size
-            )
-            packed_vae_position_ids.append(vae_posiiton_ids)
-            H, W = image_tensor.shape[1:]
-            h = H // self.latent_downsample
-            w = W // self.latent_downsample
-            patchified_vae_latent_shapes.append((h, w))
-
-            num_img_tokens = w * h
-            packed_vae_token_indexes.extend(range(_curr, _curr + num_img_tokens))
-            packed_indexes.extend(range(curr, curr + num_img_tokens))
-            curr += num_img_tokens
-            _curr += num_img_tokens
-
-            packed_text_ids.append(new_token_ids['end_of_image'])
-            packed_text_indexes.append(_curr)
-            packed_indexes.append(curr)
-            curr += 1
-            _curr += 1
-
-            packed_position_ids.extend([curr_position_id] * (num_img_tokens + 2))
-            packed_seqlens.append(num_img_tokens + 2)
-            newlens.append(curr_kvlen + num_img_tokens + 2)
-            new_rope.append(curr_position_id + 1)
-
-        image_sizes = [item.shape for item in vae_image_tensors]
-        max_image_size = [max(item) for item in list(zip(*image_sizes))]
-        padded_images = torch.zeros(size=(len(vae_image_tensors), *max_image_size))
-        for i, image_tensor in enumerate(vae_image_tensors):
-            padded_images[i, :, :image_tensor.shape[1], :image_tensor.shape[2]] = image_tensor
-
-        generation_input = {
-            "padded_images": padded_images,
-            "patchified_vae_latent_shapes": patchified_vae_latent_shapes,
-            "packed_vae_position_ids": torch.cat(packed_vae_position_ids, dim=0),
-            "packed_timesteps": torch.tensor([timestep]),
-            "packed_vae_token_indexes": torch.tensor(packed_vae_token_indexes, dtype=torch.long),
-            "packed_text_ids": torch.tensor(packed_text_ids, dtype=torch.long),
-            "packed_text_indexes": torch.tensor(packed_text_indexes, dtype=torch.long),
-            "packed_position_ids": torch.tensor(packed_position_ids, dtype=torch.long),
-            "packed_seqlens": torch.tensor(packed_seqlens, dtype=torch.int),
-            "packed_indexes": torch.tensor(packed_indexes, dtype=torch.long),
-            "packed_key_value_indexes": torch.tensor(packed_key_value_indexes, dtype=torch.long),
-            "key_values_lens": torch.tensor(curr_kvlens, dtype=torch.int),
-        }
-
-        return generation_input, newlens, new_rope
-
-    @torch.no_grad
-    def forward_cache_update_vae(
-        self,
-        vae_model,
-        past_key_values: NaiveCache,
-        padded_images: torch.Tensor,
-        patchified_vae_latent_shapes: List,
-        packed_vae_position_ids: torch.LongTensor,
-        packed_timesteps: torch.Tensor,
-        packed_vae_token_indexes: torch.LongTensor,
-        packed_text_ids: torch.LongTensor,
-        packed_text_indexes: torch.LongTensor,
-        packed_position_ids: torch.LongTensor,
-        packed_seqlens: torch.IntTensor,
-        packed_indexes: torch.LongTensor,
-        key_values_lens: torch.IntTensor,
-        packed_key_value_indexes: torch.Tensor,
-    ):
-        packed_text_embedding = self.language_model.model.embed_tokens(packed_text_ids)
-        packed_sequence = packed_text_embedding.new_zeros((sum(packed_seqlens), self.hidden_size))
-        packed_sequence[packed_text_indexes] = packed_text_embedding
-
-        padded_latent = vae_model.encode(padded_images)
-
-        p = self.latent_patch_size
-        packed_latent = list()
-        for latent, (h, w) in zip(padded_latent, patchified_vae_latent_shapes):
-            latent = latent[:, :h * p, :w * p].reshape(self.latent_channel, h, p, w, p)
-            latent = torch.einsum("chpwq->hwpqc", latent).reshape(-1, p * p * self.latent_channel)
-            packed_latent.append(latent)
-        packed_latent = torch.cat(packed_latent, dim=0)
-        packed_pos_embed = self.latent_pos_embed(packed_vae_position_ids)
-        packed_timestep_embeds = self.time_embedder(packed_timesteps)
-        packed_latent = self.vae2llm(packed_latent) + packed_timestep_embeds + packed_pos_embed
-        packed_sequence[packed_vae_token_indexes] = packed_latent
-
-        extra_inputs = {}
-        if self.use_moe:
-            extra_inputs = {
-                "mode": "gen",
-                "packed_vae_token_indexes": packed_vae_token_indexes,
-                "packed_text_indexes": packed_text_indexes
-            }
-
-        output = self.language_model.forward_inference(
-            packed_query_sequence=packed_sequence,
-            query_lens=packed_seqlens,
-            packed_query_position_ids=packed_position_ids,
-            packed_query_indexes=packed_indexes,
-            past_key_values=past_key_values,
-            key_values_lens=key_values_lens,
-            packed_key_value_indexes=packed_key_value_indexes,
-            update_past_key_values=True,
-            is_causal=False,
-            **extra_inputs,
-        )
-        past_key_values = output.past_key_values
-
-        return past_key_values
-
-    def prepare_vae_latent(self, curr_kvlens, curr_rope, image_sizes, new_token_ids):
-        packed_text_ids, packed_text_indexes = list(), list()
-        packed_vae_position_ids, packed_vae_token_indexes, packed_init_noises = list(), list(), list()
-        packed_position_ids, packed_seqlens, packed_indexes = list(), list(), list()
-        packed_key_value_indexes = list()
-
-        query_curr = curr = 0
-        for (H, W), curr_kvlen, curr_position_id in zip(image_sizes, curr_kvlens, curr_rope):
-            packed_key_value_indexes.extend(range(curr, curr + curr_kvlen))
-            curr += curr_kvlen
-
-            packed_text_ids.append(new_token_ids['start_of_image'])
-            packed_text_indexes.append(query_curr)
-            packed_indexes.append(curr)
-            curr += 1
-            query_curr += 1
-
-            vae_posiiton_ids = self.get_flattened_position_ids(
-                H, W,
-                self.latent_downsample, 
-                max_num_patches_per_side=self.max_latent_size
-            )
-            packed_vae_position_ids.append(vae_posiiton_ids)
-
-            h, w = H // self.latent_downsample, W // self.latent_downsample
-            num_image_tokens = h * w
-            packed_init_noises.append(
-                torch.randn(num_image_tokens, self.latent_channel * self.latent_patch_size ** 2)
-            )
-            packed_vae_token_indexes.extend(range(query_curr, query_curr + num_image_tokens))
-            packed_indexes.extend(range(curr, curr + num_image_tokens))
-            curr += num_image_tokens
-            query_curr += num_image_tokens
-
-            packed_text_ids.append(new_token_ids['end_of_image'])
-            packed_text_indexes.append(query_curr)
-            packed_indexes.append(curr)
-            curr += 1
-            query_curr += 1
-
-            packed_position_ids.extend([curr_position_id] * (num_image_tokens + 2))
-            packed_seqlens.append(num_image_tokens + 2)
-
-        generation_input = {
-            "packed_text_ids": torch.tensor(packed_text_ids, dtype=torch.long),
-            "packed_text_indexes": torch.tensor(packed_text_indexes, dtype=torch.long),
-            "packed_init_noises": torch.cat(packed_init_noises, dim=0),
-            "packed_vae_position_ids": torch.cat(packed_vae_position_ids, dim=0),
-            "packed_vae_token_indexes": torch.tensor(packed_vae_token_indexes, dtype=torch.long),
-            "packed_seqlens": torch.tensor(packed_seqlens, dtype=torch.int),
-            "packed_position_ids": torch.tensor(packed_position_ids, dtype=torch.long),
-            "key_values_lens": torch.tensor(curr_kvlens, dtype=torch.int),
-            "packed_indexes": torch.tensor(packed_indexes, dtype=torch.long),
-            "packed_key_value_indexes": torch.tensor(packed_key_value_indexes, dtype=torch.long),
-        }
-
-        return generation_input
-
-    def prepare_vae_latent_cfg(self, curr_kvlens, curr_rope, image_sizes):
-        packed_position_ids, packed_indexes, packed_key_value_indexes = list(), list(), list()
-
-        query_curr = curr = 0
-        for (H, W), curr_kvlen, curr_position_id in zip(image_sizes, curr_kvlens, curr_rope):
-            packed_key_value_indexes.extend(range(curr, curr + curr_kvlen))
-            curr += curr_kvlen
-
-            packed_indexes.append(curr)
-            curr += 1
-            query_curr += 1
-
-            h, w = H // self.latent_downsample, W // self.latent_downsample
-            num_image_tokens = h * w
-            packed_indexes.extend(range(curr, curr + num_image_tokens))
-            curr += num_image_tokens
-            query_curr += num_image_tokens
-
-            packed_indexes.append(curr)
-            curr += 1
-            query_curr += 1
-
-            packed_position_ids.extend([curr_position_id] * (num_image_tokens + 2))
-
-        generation_input = {
-            "cfg_packed_position_ids": torch.tensor(packed_position_ids, dtype=torch.long),
-            "cfg_key_values_lens": torch.tensor(curr_kvlens, dtype=torch.int),
-            "cfg_packed_query_indexes": torch.tensor(packed_indexes, dtype=torch.long),
-            "cfg_packed_key_value_indexes": torch.tensor(packed_key_value_indexes, dtype=torch.long),
-        }
-
-        return generation_input
-
-    @torch.no_grad
-    def generate_image(
-        self,
-        packed_text_ids: torch.LongTensor,
-        packed_text_indexes: torch.LongTensor,
-        packed_init_noises: torch.Tensor,
-        packed_vae_position_ids: torch.LongTensor,
-        packed_vae_token_indexes: torch.LongTensor,
-        packed_seqlens: torch.IntTensor,
-        packed_position_ids: torch.LongTensor,
-        packed_indexes: torch.LongTensor,
-        past_key_values: NaiveCache,
-        key_values_lens: torch.IntTensor,
-        packed_key_value_indexes: torch.LongTensor,
-        num_timesteps: int = 24,
-        timestep_shift: float = 1.0,
-        cfg_renorm_min: float = 0.0,
-        cfg_renorm_type: str = "global",
-        cfg_interval: Optional[Tuple[float, float]] = [0, 1],
-        # cfg_text
-        cfg_text_scale: float = 1.0,
-        cfg_text_packed_query_indexes: Optional[torch.LongTensor] = None,
-        cfg_text_packed_position_ids: Optional[torch.LongTensor] = None,
-        cfg_text_past_key_values: Optional[NaiveCache] = None,
-        cfg_text_key_values_lens: Optional[torch.IntTensor] = None,
-        cfg_text_packed_key_value_indexes: Optional[torch.LongTensor] = None,
-        # cfg_img
-        cfg_img_scale: float = 1.0,
-        cfg_img_packed_query_indexes: Optional[torch.LongTensor] = None,
-        cfg_img_packed_position_ids: Optional[torch.LongTensor] = None,
-        cfg_img_past_key_values: Optional[NaiveCache] = None,
-        cfg_img_key_values_lens: Optional[torch.IntTensor] = None,
-        cfg_img_packed_key_value_indexes: Optional[torch.LongTensor] = None,
-        cfg_type: str = "parallel",
-    ):
-        x_t = packed_init_noises
-
-        timesteps = torch.linspace(1, 0, num_timesteps, device=x_t.device)
-        timesteps = timestep_shift * timesteps / (1 + (timestep_shift - 1) * timesteps)
-        dts =  timesteps[:-1] - timesteps[1:]
-        timesteps = timesteps[:-1]
-
-        for i, t in enumerate(timesteps):
-
-            timestep = torch.tensor([t] * x_t.shape[0], device=x_t.device)
-            if t > cfg_interval[0] and t <= cfg_interval[1]:
-                cfg_text_scale_ = cfg_text_scale
-                cfg_img_scale_ = cfg_img_scale
-            else:
-                cfg_text_scale_ = 1.0
-                cfg_img_scale_ = 1.0
-            v_t = self._forward_flow(
-                x_t=x_t,
-                timestep=timestep, 
-                packed_vae_token_indexes=packed_vae_token_indexes,
-                packed_vae_position_ids=packed_vae_position_ids,
-                packed_text_ids=packed_text_ids,
-                packed_text_indexes=packed_text_indexes,
-                packed_position_ids=packed_position_ids,
-                packed_indexes=packed_indexes,
-                packed_seqlens=packed_seqlens,
-                key_values_lens=key_values_lens,
-                past_key_values=past_key_values,
-                packed_key_value_indexes=packed_key_value_indexes,
-                cfg_renorm_min=cfg_renorm_min,
-                cfg_renorm_type=cfg_renorm_type,
-                # cfg_text
-                cfg_text_scale=cfg_text_scale_,
-                cfg_text_packed_position_ids=cfg_text_packed_position_ids,
-                cfg_text_packed_query_indexes=cfg_text_packed_query_indexes,
-                cfg_text_key_values_lens=cfg_text_key_values_lens,
-                cfg_text_past_key_values=cfg_text_past_key_values,
-                cfg_text_packed_key_value_indexes=cfg_text_packed_key_value_indexes,
-                # cfg_img
-                cfg_img_scale=cfg_img_scale_,
-                cfg_img_packed_position_ids=cfg_img_packed_position_ids,
-                cfg_img_packed_query_indexes=cfg_img_packed_query_indexes,
-                cfg_img_key_values_lens=cfg_img_key_values_lens,
-                cfg_img_past_key_values=cfg_img_past_key_values,
-                cfg_img_packed_key_value_indexes=cfg_img_packed_key_value_indexes,
-                cfg_type=cfg_type,
-            )
-
-            x_t = x_t - v_t.to(x_t.device) * dts[i] # velocity pointing from data to noise
-
-        unpacked_latent = x_t.split((packed_seqlens - 2).tolist())
-        return unpacked_latent
-
-    @torch.no_grad
-    def _forward_flow(
-        self,
-        x_t: torch.Tensor,
-        timestep: torch.LongTensor,
-        packed_vae_token_indexes: torch.LongTensor,
-        packed_vae_position_ids: torch.LongTensor,
-        packed_text_ids: torch.LongTensor,
-        packed_text_indexes: torch.LongTensor,
-        packed_indexes: torch.LongTensor,
-        packed_position_ids: torch.LongTensor,
-        packed_seqlens: torch.IntTensor,
-        key_values_lens: torch.IntTensor,
-        past_key_values: NaiveCache,
-        packed_key_value_indexes: torch.LongTensor,
-        cfg_renorm_min: float = 0.0,
-        cfg_renorm_type: str = "global",
-        # cfg_text
-        cfg_text_scale: float = 1.0,
-        cfg_text_packed_position_ids: Optional[torch.LongTensor] = None,
-        cfg_text_packed_query_indexes: Optional[torch.LongTensor] = None,
-        cfg_text_key_values_lens: Optional[torch.Tensor] = None,
-        cfg_text_past_key_values: Optional[NaiveCache] = None,
-        cfg_text_packed_key_value_indexes: Optional[torch.LongTensor] = None,
-        # cfg_img
-        cfg_img_scale: float = 1.0,
-        cfg_img_packed_position_ids: Optional[torch.LongTensor] = None,
-        cfg_img_packed_query_indexes: Optional[torch.LongTensor] = None,
-        cfg_img_key_values_lens: Optional[torch.Tensor] = None,
-        cfg_img_past_key_values: Optional[NaiveCache] = None,
-        cfg_img_packed_key_value_indexes: Optional[torch.LongTensor] = None,
-        cfg_type: str = "parallel",
-    ):
-        packed_text_embedding = self.language_model.model.embed_tokens(packed_text_ids)
-        packed_sequence = packed_text_embedding.new_zeros((sum(packed_seqlens), self.hidden_size))
-        packed_sequence[packed_text_indexes] = packed_text_embedding
-
-        assert timestep.unique().shape[0] == 1
-        packed_pos_embed = self.latent_pos_embed(packed_vae_position_ids)
-        packed_timestep_embeds = self.time_embedder(timestep)
-        x_t = self.vae2llm(x_t) + packed_timestep_embeds + packed_pos_embed
-        packed_sequence[packed_vae_token_indexes] = x_t
-
-        extra_inputs = {}
-        if self.use_moe:
-            extra_inputs = {
-                "mode": "gen",
-                "packed_vae_token_indexes": packed_vae_token_indexes,
-                "packed_text_indexes": packed_text_indexes
-            }
-
-        output = self.language_model.forward_inference(
-            packed_query_sequence=packed_sequence,
-            query_lens=packed_seqlens,
-            packed_query_position_ids=packed_position_ids,
-            packed_query_indexes=packed_indexes,
-            past_key_values=past_key_values,
-            key_values_lens=key_values_lens,
-            packed_key_value_indexes=packed_key_value_indexes,
-            update_past_key_values=False,
-            is_causal=False,
-            **extra_inputs,
-        )
-        v_t = self.llm2vae(output.packed_query_sequence)
-        v_t = v_t[packed_vae_token_indexes]
-
-        if cfg_text_scale > 1.0:
-            cfg_text_output = self.language_model.forward_inference(
-                packed_query_sequence=packed_sequence,
-                query_lens=packed_seqlens,
-                packed_query_position_ids=cfg_text_packed_position_ids,
-                packed_query_indexes=cfg_text_packed_query_indexes,
-                past_key_values=cfg_text_past_key_values,
-                key_values_lens=cfg_text_key_values_lens,
-                packed_key_value_indexes=cfg_text_packed_key_value_indexes,
-                update_past_key_values=False,
-                is_causal=False,
-                **extra_inputs,
-            )
-            cfg_text_v_t = self.llm2vae(cfg_text_output.packed_query_sequence)
-            cfg_text_v_t = cfg_text_v_t[packed_vae_token_indexes]
-
-        if cfg_img_scale > 1.0:
-            cfg_img_output = self.language_model.forward_inference(
-                packed_query_sequence=packed_sequence,
-                query_lens=packed_seqlens,
-                packed_query_position_ids=cfg_img_packed_position_ids,
-                packed_query_indexes=cfg_img_packed_query_indexes,
-                past_key_values=cfg_img_past_key_values,
-                key_values_lens=cfg_img_key_values_lens,
-                packed_key_value_indexes=cfg_img_packed_key_value_indexes,
-                update_past_key_values=False,
-                is_causal=False,
-                **extra_inputs,
-            )
-            cfg_img_v_t = self.llm2vae(cfg_img_output.packed_query_sequence)
-            cfg_img_v_t = cfg_img_v_t[packed_vae_token_indexes]
-
-        if cfg_text_scale > 1.0:
-            if cfg_renorm_type == "text_channel":
-                v_t_text_ = cfg_text_v_t + cfg_text_scale * (v_t - cfg_text_v_t)
-                norm_v_t = torch.norm(v_t, dim=-1, keepdim=True)
-                norm_v_t_text_ = torch.norm(v_t_text_, dim=-1, keepdim=True)
-                scale = (norm_v_t / (norm_v_t_text_ + 1e-8)).clamp(min=cfg_renorm_min, max=1.0)
-                v_t_text = v_t_text_ * scale
-                if cfg_img_scale > 1.0:
-                    v_t = cfg_img_v_t + cfg_img_scale * (v_t_text - cfg_img_v_t)
-                else:
-                    v_t = v_t_text
-            else:
-                v_t_text_ = cfg_text_v_t + cfg_text_scale * (v_t - cfg_text_v_t)
-                
-                if cfg_img_scale > 1.0:
-                    v_t_ = cfg_img_v_t + cfg_img_scale * (v_t_text_ - cfg_img_v_t)
-                else:
-                    v_t_ = v_t_text_
-
-                # NOTE norm is computed over all dimensions, thus currently only supports batch_size = 1 with navit
-                if cfg_renorm_type == "global":
-                    norm_v_t = torch.norm(v_t)
-                    norm_v_t_ = torch.norm(v_t_)
-                elif cfg_renorm_type == "channel":
-                    norm_v_t = torch.norm(v_t, dim=-1, keepdim=True)
-                    norm_v_t_ = torch.norm(v_t_, dim=-1, keepdim=True)
-                else:
-                    raise NotImplementedError(f"{cfg_renorm_type} is not suppoprted")
-                scale = (norm_v_t / (norm_v_t_ + 1e-8)).clamp(min=cfg_renorm_min, max=1.0)
-                v_t = v_t_ * scale
-        else:
-            # No CFG
-            pass
-
-        return v_t
-
-    def prepare_start_tokens(self, curr_kvlens, curr_rope, new_token_ids):
-        packed_start_tokens, packed_key_value_indexes = list(), list()
-        packed_query_position_ids = list()
-
-        curr = 0
-        for curr_kvlen, curr_position_id in zip(curr_kvlens, curr_rope):
-            packed_key_value_indexes.extend(range(curr, curr + curr_kvlen))
-            packed_start_tokens.append(new_token_ids['bos_token_id'])
-            packed_query_position_ids.append(curr_position_id)
-            curr += curr_kvlen
-
-        generation_input = {
-            "packed_start_tokens": torch.tensor(packed_start_tokens, dtype=torch.long),
-            "packed_query_position_ids": torch.tensor(packed_query_position_ids, dtype=torch.long),
-            "key_values_lens": torch.tensor(curr_kvlens, dtype=torch.int),
-            "packed_key_value_indexes": torch.tensor(packed_key_value_indexes, dtype=torch.long),
-        }
-
-        return generation_input
-
-    @torch.no_grad
-    def generate_text(
-        self,
-        past_key_values: NaiveCache,
-        packed_key_value_indexes: torch.LongTensor,
-        key_values_lens: torch.IntTensor,
-        packed_start_tokens: torch.LongTensor,
-        packed_query_position_ids: torch.LongTensor,
-        max_length: int,
-        do_sample: bool = False,
-        temperature: float = 1.0,
-        end_token_id: int = None,
-    ):
-        step = 0
-        generated_sequence = []
-        curr_tokens = packed_start_tokens
-        while step < max_length:
-            generated_sequence.append(curr_tokens)
-            packed_text_embedding = self.language_model.model.embed_tokens(curr_tokens)
-            query_lens = torch.ones_like(curr_tokens)
-            packed_query_indexes = torch.cumsum(key_values_lens, dim=0) + torch.arange(
-                0, len(key_values_lens), 
-                device=key_values_lens.device, 
-                dtype=key_values_lens.dtype
-            )
-
-            uppacked = list(packed_key_value_indexes.split(key_values_lens.tolist(), dim=0))
-            for i in range(len(uppacked)):
-                uppacked[i] += i
-            packed_key_value_indexes = torch.cat(uppacked, dim=0)
-
-            extra_inputs = {}
-            if self.use_moe:
-                extra_inputs = {"mode": "und"}
-
-            output = self.language_model.forward_inference(
-                packed_query_sequence=packed_text_embedding,
-                query_lens=query_lens,
-                packed_query_position_ids=packed_query_position_ids,
-                packed_query_indexes=packed_query_indexes,
-                past_key_values=past_key_values,
-                key_values_lens=key_values_lens,
-                packed_key_value_indexes=packed_key_value_indexes,
-                update_past_key_values=True,
-                is_causal=True,
-                **extra_inputs,
-            )
-            past_key_values = output.past_key_values
-            packed_query_sequence = output.packed_query_sequence
-            pred_logits = self.language_model.lm_head(packed_query_sequence)
-
-            if do_sample:
-                probs = nn.functional.softmax(pred_logits / temperature, dim=-1)
-                curr_tokens = torch.multinomial(probs, num_samples=1).squeeze(1)
-            else:
-                curr_tokens = torch.argmax(pred_logits, dim=-1)
-
-            uppacked = list(packed_key_value_indexes.split(key_values_lens.tolist(), dim=0))
-            for i in range(len(uppacked)):
-                uppacked[i] = torch.cat(
-                    [uppacked[i], torch.tensor([uppacked[i][-1] + 1], device=uppacked[i].device)], dim=0
-                )
-            packed_key_value_indexes = torch.cat(uppacked, dim=0)
-            key_values_lens = key_values_lens + 1
-            packed_query_position_ids = packed_query_position_ids + 1
-            step += 1
-
-            if end_token_id is not None and curr_tokens[0] == end_token_id: # only support batch=1
-                break
-
-        output_device = generated_sequence[0].device
-        return torch.stack([i.to(output_device) for i in generated_sequence], dim=0)
-
-    # for evaluation
-    @torch.no_grad()
-    def chat(
-        self,
-        tokenizer,
-        new_token_ids,
-        image_transform,
-        images,
-        prompt,
-        max_length: int,
-        do_sample: bool = False,
-        temperature: float = 1.0,
-    ):
-        device = next(self.parameters()).device
-
-        if isinstance(new_token_ids, dict):
-            for k, v in new_token_ids.items():
-                if torch.is_tensor(v):
-                    new_token_ids[k] = v.to(device)
-        elif torch.is_tensor(new_token_ids):
-            new_token_ids = new_token_ids.to(device)
-
-        # prefill
-        past_key_values = NaiveCache(self.config.llm_config.num_hidden_layers)
-        newlens = [0]
-        new_rope = [0]
-
-        # add images
-        for image in images:
-            generation_input, newlens, new_rope = self.prepare_vit_images(
-                curr_kvlens=newlens,
-                curr_rope=new_rope, 
-                images=[image], 
-                transforms=image_transform,
-                new_token_ids=new_token_ids,
-            )
-            for k, v in generation_input.items():
-                if torch.is_tensor(v):
-                    generation_input[k] = v.to(device)
-            with torch.amp.autocast("cuda", enabled=True, dtype=torch.bfloat16):
-                past_key_values = self.forward_cache_update_vit(past_key_values, **generation_input)
-
-        # add text
-        generation_input, newlens, new_rope = self.prepare_prompts(
-            curr_kvlens=newlens,
-            curr_rope=new_rope, 
-            prompts=[prompt],
-            tokenizer=tokenizer, 
-            new_token_ids=new_token_ids,
-        )
-        for k, v in generation_input.items():
-            if torch.is_tensor(v):
-                generation_input[k] = v.to(device)
-        with torch.amp.autocast("cuda", enabled=True, dtype=torch.bfloat16):
-            past_key_values = self.forward_cache_update_text(past_key_values, **generation_input)
-
-        # decode
-        generation_input = self.prepare_start_tokens(newlens, new_rope, new_token_ids)
-        for k, v in generation_input.items():
-            if torch.is_tensor(v):
-                generation_input[k] = v.to(device)
-        with torch.amp.autocast("cuda", enabled=True, dtype=torch.bfloat16):
-            unpacked_latent = self.generate_text(
-                past_key_values=past_key_values,
-                max_length=max_length,
-                do_sample=do_sample,
-                temperature=temperature,
-                end_token_id=new_token_ids['eos_token_id'],
-                **generation_input,
-            )
-        output = tokenizer.decode(unpacked_latent[:,0])
-        output = output.split('<|im_end|>')[0].split('<|im_start|>')[1]
-
-        return output
+# Copyright 2025 Bytedance Ltd. and/or its affiliates.
+# SPDX-License-Identifier: Apache-2.0
+
+import copy
+from typing import List, Tuple, Optional
+import matplotlib.pyplot as plt
+
+from PIL import Image
+import torch
+import torch.nn.functional as F
+from torch import nn
+from torch.nn.attention.flex_attention import create_block_mask
+from transformers.configuration_utils import PretrainedConfig
+from transformers.modeling_utils import PreTrainedModel
+
+from data.data_utils import (
+    create_sparse_mask, 
+    get_flattened_position_ids_extrapolate, 
+    get_flattened_position_ids_interpolate,
+    patchify, 
+)
+from .qwen2_navit import NaiveCache
+from .modeling_utils import MLPconnector, TimestepEmbedder, PositionEmbedding
+
+
+class BagelConfig(PretrainedConfig):
+    def __init__(
+        self,
+        visual_gen=True,
+        visual_und=True,
+        llm_config=None,
+        vit_config=None,
+        vae_config=None,
+        latent_patch_size=2,
+        max_latent_size=32,
+        vit_max_num_patch_per_side=70,
+        connector_act="gelu_pytorch_tanh",
+        interpolate_pos=False,
+        timestep_shift=1.0,
+        **kwargs
+    ):
+        super().__init__(**kwargs)
+        self.visual_gen = visual_gen
+        self.visual_und = visual_und
+        self.llm_config = llm_config
+        self.vit_config = vit_config
+        self.vae_config = vae_config
+        self.latent_patch_size = latent_patch_size
+        self.max_latent_size = max_latent_size
+        self.vit_max_num_patch_per_side = vit_max_num_patch_per_side
+        self.connector_act = connector_act
+        self.interpolate_pos = interpolate_pos
+        self.timestep_shift = timestep_shift
+
+
+class Bagel(PreTrainedModel):
+    config_class = BagelConfig
+    base_model_prefix = 'bagel'
+
+    def __init__(self, language_model, vit_model, config: BagelConfig):
+        super().__init__(config)    
+        self.language_model = language_model
+        self.hidden_size = config.llm_config.hidden_size
+        self.use_moe = "Mo" in config.llm_config.layer_module
+        self.num_heads = config.llm_config.num_attention_heads
+
+        if config.visual_gen:
+            self.latent_patch_size = config.latent_patch_size
+            self.timestep_shift = config.timestep_shift
+            self.latent_downsample = config.vae_config.downsample * config.latent_patch_size
+            self.max_latent_size = config.max_latent_size
+            self.latent_channel = config.vae_config.z_channels
+            self.patch_latent_dim = self.latent_patch_size ** 2 * self.latent_channel
+            self.time_embedder = TimestepEmbedder(self.hidden_size)
+            self.vae2llm = nn.Linear(self.patch_latent_dim, self.hidden_size)
+            self.llm2vae = nn.Linear(self.hidden_size, self.patch_latent_dim)
+            self.latent_pos_embed = PositionEmbedding(self.max_latent_size, self.hidden_size)
+
+        if config.visual_und:
+            self.vit_model = vit_model
+            self.vit_patch_size = config.vit_config.patch_size
+            self.vit_max_num_patch_per_side = config.vit_max_num_patch_per_side
+            self.vit_hidden_size = config.vit_config.hidden_size
+            self.connector = MLPconnector(self.vit_hidden_size, self.hidden_size, config.connector_act)
+            self.vit_pos_embed = PositionEmbedding(self.vit_max_num_patch_per_side, self.hidden_size)
+
+        if config.interpolate_pos:
+            self.get_flattened_position_ids = get_flattened_position_ids_interpolate
+        else:
+            self.get_flattened_position_ids = get_flattened_position_ids_extrapolate
+
+        self.config = config
+        self._init_weights()
+
+    def _init_weights(self):
+        if self.config.visual_gen:
+            nn.init.constant_(self.llm2vae.weight, 0)
+            nn.init.constant_(self.llm2vae.bias, 0)
+
+    def forward(
+        self,
+        sequence_length: int,
+        packed_text_ids: torch.LongTensor,
+        packed_text_indexes: torch.LongTensor,
+        sample_lens: List[int],
+        packed_position_ids: torch.LongTensor,
+        nested_attention_masks: List[torch.Tensor] = None,
+        split_lens: List[int] = None,
+        attn_modes: List[str] = None,
+        # for visual understanding
+        ce_loss_indexes: Optional[torch.BoolTensor] = None,
+        packed_label_ids: Optional[torch.LongTensor] = None,
+        packed_vit_tokens: Optional[torch.Tensor] = None,
+        packed_vit_token_indexes: Optional[torch.LongTensor] = None,
+        packed_vit_position_ids: Optional[torch.LongTensor] = None,
+        vit_token_seqlens: Optional[torch.IntTensor] = None,
+        # for visual generation
+        padded_latent: Optional[torch.Tensor] = None,
+        patchified_vae_latent_shapes: Optional[List[Tuple[int, int]]] = None,
+        packed_latent_position_ids: Optional[torch.LongTensor] = None,
+        packed_vae_token_indexes: Optional[torch.LongTensor] = None,
+        packed_timesteps: Optional[torch.LongTensor] = None,
+        mse_loss_indexes: Optional[torch.BoolTensor] = None,
+    ) -> torch.Tensor:
+        """
+        Args:
+            sequence_length: length of sequence.
+            packed_text_ids: 1-D int tensor, packed text token ids.
+            packed_text_indexes: 1-D int tensor, packed text token indexes in sequence.
+            sample_lens: A list of N ints, length of each sample in packed_sequence.
+            nested_attention_masks: A list of N 2-D float tensor,  where 0.0 means attention and 
+                -inf means ignore.
+            packed_position_ids: packed 1-D positions, an image has only one global position shared
+                by all latent tokens.
+
+            packed_vit_tokens: packed patchified image tokens for vit model.
+            packed_vit_position_ids: 1-D int tensor, the position of each token for vit model.
+            packed_vit_token_indexes: 1-D int tensor, packed vit token indexes in sequence.
+            vit_token_seqlens: 1-D int tensor, the length of each image tokens for vit model.
+            packed_label_ids: 1-D int tensor, packed label token ids.
+            ce_loss_indexes: 1-D bool tensor, where to compute ce loss.
+
+            padded_latent: padded latent from VAE encoder.
+            patchified_vae_latent_shapes: A list of (h, w) tuples, patchfied latent shapes of each image.
+            packed_latent_position_ids: 1-D int tensor, the position of each token for latent.
+            packed_vae_token_indexes: 1-D int tensor, padded image token indexes in sequence.
+            packed_timesteps: 1-D float tensor, flow timesteps. 0 indicates use clean image.
+            mse_loss_indexes: 1-D bool tensor, where to compute mse loss.
+        """
+        packed_text_embedding = self.language_model.model.embed_tokens(packed_text_ids)
+        packed_sequence = packed_text_embedding.new_zeros(size=(sequence_length, self.hidden_size))
+        packed_sequence[packed_text_indexes] = packed_text_embedding
+
+        if nested_attention_masks is None:
+            sparse_mask = create_sparse_mask(sample_lens, split_lens, attn_modes, packed_text_embedding.device)
+            seqlen = sum(sample_lens)
+            block_mask = create_block_mask(
+                sparse_mask, B=1, H=self.num_heads, Q_LEN=seqlen, KV_LEN=seqlen, 
+                device=packed_text_embedding.device, BLOCK_SIZE=128, _compile=True
+            )
+            attention_mask = block_mask
+        else:
+            attention_mask = nested_attention_masks
+
+        if self.config.visual_und:
+            cu_seqlens = torch.nn.functional.pad(torch.cumsum(vit_token_seqlens, dim=0), (1, 0))
+            cu_seqlens = cu_seqlens.to(torch.int32)
+            max_seqlen = torch.max(vit_token_seqlens).item()
+            packed_vit_token_embed = self.vit_model(
+                packed_pixel_values=packed_vit_tokens, 
+                packed_flattened_position_ids=packed_vit_position_ids,
+                cu_seqlens=cu_seqlens,
+                max_seqlen=max_seqlen,
+            )
+            packed_vit_token_embed = self.connector(packed_vit_token_embed)
+            vit_token_pos_emb = self.vit_pos_embed(packed_vit_position_ids)
+            packed_vit_token_embed = packed_vit_token_embed + vit_token_pos_emb
+            packed_sequence[packed_vit_token_indexes] = packed_vit_token_embed
+
+        if self.config.visual_gen:
+            p = self.latent_patch_size
+            packed_latent = []
+            for latent, (h, w) in zip(padded_latent, patchified_vae_latent_shapes):
+                latent = latent[:, :h * p, :w * p].reshape(self.latent_channel, h, p, w, p)
+                latent = torch.einsum("chpwq->hwpqc", latent).reshape(-1, p * p * self.latent_channel)
+                packed_latent.append(latent)
+            packed_latent_clean = torch.cat(packed_latent, dim=0)
+
+            noise = torch.randn_like(packed_latent_clean)
+            packed_timesteps = torch.sigmoid(packed_timesteps)
+            packed_timesteps = self.timestep_shift * packed_timesteps / (1 + (self.timestep_shift - 1) * packed_timesteps)
+            packed_latent = (1 - packed_timesteps[:, None]) * packed_latent_clean + packed_timesteps[:, None] * noise
+            packed_timestep_embeds = self.time_embedder(packed_timesteps)
+            latent_token_pos_emb = self.latent_pos_embed(packed_latent_position_ids)
+            packed_latent = self.vae2llm(packed_latent) + packed_timestep_embeds + latent_token_pos_emb
+            packed_sequence[packed_vae_token_indexes] = packed_latent
+
+        extra_inputs = {}
+        if self.use_moe:
+            packed_und_token_indexes = packed_text_indexes
+            if packed_vit_token_indexes is not None:
+                packed_und_token_indexes=torch.cat([packed_text_indexes, packed_vit_token_indexes], dim=0)
+            extra_inputs.update(
+                packed_und_token_indexes=packed_und_token_indexes,
+                packed_gen_token_indexes=packed_vae_token_indexes,
+            )
+
+        last_hidden_state = self.language_model(
+            packed_sequence=packed_sequence,
+            sample_lens=sample_lens,
+            attention_mask=attention_mask,
+            packed_position_ids=packed_position_ids,
+            **extra_inputs,
+        )
+
+        mse = None
+        if self.config.visual_gen:
+            packed_mse_preds = self.llm2vae(last_hidden_state[mse_loss_indexes])
+            target = noise - packed_latent_clean # NOTE: v_t=dx_t/dt=x_1-x_0, pointing from data to noise
+            has_mse = packed_timesteps > 0
+            mse = (packed_mse_preds - target[has_mse]) ** 2
+
+        ce = None
+        if ce_loss_indexes is not None:
+            packed_ce_preds = self.language_model.lm_head(last_hidden_state[ce_loss_indexes])
+            ce = F.cross_entropy(packed_ce_preds, packed_label_ids, reduction="none")
+
+        return dict(mse=mse, ce=ce)
+
+
+    def prepare_prompts(self, curr_kvlens, curr_rope, prompts, tokenizer, new_token_ids):
+        packed_text_ids = list()
+        packed_text_position_ids = list()
+        text_token_lens = list()
+        packed_text_indexes = list()
+        packed_key_value_indexes = list()
+
+        curr = 0
+        newlens, new_rope = list(), list()
+        for prompt, curr_kvlen, curr_position_id in zip(prompts, curr_kvlens, curr_rope):
+            packed_key_value_indexes.extend(range(curr, curr + curr_kvlen))
+            curr += curr_kvlen
+
+            text_ids = tokenizer.encode(prompt)
+            text_ids = [new_token_ids['bos_token_id']] + text_ids + [new_token_ids['eos_token_id']]
+            text_token_lens.append(len(text_ids))
+            packed_text_ids.extend(text_ids)
+            packed_text_position_ids.extend(range(curr_position_id, curr_position_id + len(text_ids)))
+            packed_text_indexes.extend(range(curr, curr + len(text_ids)))
+            newlens.append(curr_kvlen + len(text_ids))
+            new_rope.append(curr_position_id + len(text_ids))
+            curr += len(text_ids)
+
+        generation_input = {
+            "text_token_lens": torch.tensor(text_token_lens, dtype=torch.int),
+            "packed_text_ids": torch.tensor(packed_text_ids, dtype=torch.long),
+            "packed_text_position_ids": torch.tensor(packed_text_position_ids, dtype=torch.long),
+            "packed_text_indexes": torch.tensor(packed_text_indexes, dtype=torch.long),
+            "packed_key_value_indexes": torch.tensor(packed_key_value_indexes, dtype=torch.long),
+            "key_values_lens": torch.tensor(curr_kvlens, dtype=torch.int),
+        }
+
+        return generation_input, newlens, new_rope
+
+    @torch.no_grad
+    def forward_cache_update_text(
+        self,
+        past_key_values: NaiveCache,
+        packed_text_ids: torch.IntTensor,
+        packed_text_position_ids: torch.LongTensor,
+        text_token_lens: torch.LongTensor,
+        packed_text_indexes: torch.LongTensor,
+        packed_key_value_indexes: torch.LongTensor,
+        key_values_lens: torch.IntTensor,
+    ):
+        packed_text_embedding = self.language_model.model.embed_tokens(packed_text_ids)
+
+        extra_inputs = {}
+        if self.use_moe:
+            extra_inputs = {"mode": "und"}
+
+        output = self.language_model.forward_inference(
+            packed_query_sequence=packed_text_embedding,
+            query_lens=text_token_lens,
+            packed_query_position_ids=packed_text_position_ids,
+            packed_query_indexes=packed_text_indexes,
+            past_key_values=past_key_values,
+            packed_key_value_indexes=packed_key_value_indexes,
+            key_values_lens=key_values_lens,
+            update_past_key_values=True,
+            is_causal=True,
+            **extra_inputs,
+        )
+        past_key_values = output.past_key_values
+
+        return past_key_values
+
+    def prepare_vit_images(self, curr_kvlens, curr_rope, images, transforms, new_token_ids):
+        packed_vit_token_indexes = list()
+        vit_token_seqlens, packed_vit_tokens, packed_vit_position_ids = list(), list(), list()
+        packed_text_ids, packed_text_indexes = list(), list()
+        packed_seqlens, packed_position_ids, packed_indexes = list(), list(), list()
+        packed_key_value_indexes = list()
+
+        _curr = curr = 0
+        newlens, new_rope = list(), list()
+        for image, curr_kvlen, curr_position_id in zip(images, curr_kvlens, curr_rope):
+            packed_key_value_indexes.extend(range(curr, curr + curr_kvlen))
+            curr += curr_kvlen
+
+            packed_text_ids.append(new_token_ids['start_of_image'])
+            packed_text_indexes.append(_curr)
+            packed_indexes.append(curr)
+            curr += 1
+            _curr += 1
+
+            image_tensor = transforms(image)
+            vit_position_ids = self.get_flattened_position_ids(
+                image_tensor.size(1), image_tensor.size(2), 
+                self.vit_patch_size, 
+                max_num_patches_per_side=self.vit_max_num_patch_per_side
+            )
+            vit_tokens = patchify(image_tensor, self.vit_patch_size)
+            packed_vit_tokens.append(vit_tokens)
+            num_img_tokens = vit_tokens.shape[0]
+            packed_vit_position_ids.append(vit_position_ids)
+            vit_token_seqlens.append(num_img_tokens)
+            packed_vit_token_indexes.extend(range(_curr, _curr + num_img_tokens))
+            packed_indexes.extend(range(curr, curr + num_img_tokens))
+            curr += num_img_tokens
+            _curr += num_img_tokens
+
+            packed_text_ids.append(new_token_ids['end_of_image'])
+            packed_text_indexes.append(_curr)
+            packed_indexes.append(curr)
+            curr += 1
+            _curr += 1
+
+            packed_position_ids.extend([curr_position_id] * (num_img_tokens + 2))
+            packed_seqlens.append(num_img_tokens + 2)
+            newlens.append(curr_kvlen + num_img_tokens + 2)
+            new_rope.append(curr_position_id + 1)
+
+        generation_input = {
+            "packed_text_ids": torch.tensor(packed_text_ids, dtype=torch.long),
+            "packed_text_indexes": torch.tensor(packed_text_indexes, dtype=torch.long),
+            "vit_token_seqlens": torch.tensor(vit_token_seqlens, dtype=torch.int),
+            "packed_vit_tokens": torch.cat(packed_vit_tokens, dim=0),
+            "packed_vit_position_ids": torch.cat(packed_vit_position_ids, dim=0),
+            "packed_vit_token_indexes": torch.tensor(packed_vit_token_indexes, dtype=torch.long),
+            "packed_position_ids": torch.tensor(packed_position_ids, dtype=torch.long),
+            "packed_seqlens": torch.tensor(packed_seqlens, dtype=torch.int),
+            "packed_indexes": torch.tensor(packed_indexes, dtype=torch.long),
+            "packed_key_value_indexes": torch.tensor(packed_key_value_indexes, dtype=torch.long),
+            "key_values_lens": torch.tensor(curr_kvlens, dtype=torch.int),
+        }
+
+        return generation_input, newlens, new_rope
+
+    @torch.no_grad
+    def forward_cache_update_vit(
+        self,
+        past_key_values: NaiveCache,
+        packed_text_ids: torch.LongTensor,
+        packed_text_indexes: torch.LongTensor,
+        packed_vit_tokens: torch.Tensor,
+        packed_vit_token_indexes: torch.LongTensor,
+        packed_vit_position_ids: torch.LongTensor,
+        vit_token_seqlens: torch.IntTensor,
+        packed_position_ids: torch.LongTensor,
+        packed_seqlens: torch.IntTensor,
+        packed_indexes: torch.LongTensor,
+        packed_key_value_indexes: torch.LongTensor,
+        key_values_lens: torch.IntTensor,
+    ):
+        packed_text_embedding = self.language_model.model.embed_tokens(packed_text_ids)
+        packed_sequence = packed_text_embedding.new_zeros((sum(packed_seqlens), self.hidden_size))
+        packed_sequence[packed_text_indexes] = packed_text_embedding
+
+        cu_seqlens = torch.nn.functional.pad(torch.cumsum(vit_token_seqlens, dim=0), (1, 0))
+        cu_seqlens = cu_seqlens.to(torch.int32)
+        max_seqlen = torch.max(vit_token_seqlens).item()
+        packed_vit_token_embed = self.vit_model(
+            packed_pixel_values=packed_vit_tokens, 
+            packed_flattened_position_ids=packed_vit_position_ids,
+            cu_seqlens=cu_seqlens,
+            max_seqlen=max_seqlen,
+        )
+        packed_vit_token_embed = self.connector(packed_vit_token_embed)
+        pos_emb = self.vit_pos_embed(packed_vit_position_ids)
+        packed_vit_token_embed = packed_vit_token_embed + pos_emb
+        packed_sequence[packed_vit_token_indexes] = packed_vit_token_embed
+
+        extra_inputs = {}
+        if self.use_moe:
+            extra_inputs = {"mode": "und"}
+
+        output = self.language_model.forward_inference(
+            packed_query_sequence=packed_sequence,
+            query_lens=packed_seqlens,
+            packed_query_position_ids=packed_position_ids,
+            packed_query_indexes=packed_indexes,
+            past_key_values=past_key_values,
+            packed_key_value_indexes=packed_key_value_indexes,
+            key_values_lens=key_values_lens,
+            update_past_key_values=True,
+            is_causal=False,
+            **extra_inputs,
+        )
+        past_key_values = output.past_key_values
+
+        return past_key_values
+
+    def prepare_vae_images(self, curr_kvlens, curr_rope, images, transforms, new_token_ids, timestep=0):
+        patchified_vae_latent_shapes, packed_vae_position_ids = list(), list()
+        packed_vae_token_indexes = list()
+        packed_text_ids, packed_text_indexes = list(), list()
+        packed_seqlens, packed_position_ids, packed_indexes = list(), list(), list()
+        packed_key_value_indexes = list()
+
+        _curr = curr = 0
+        vae_image_tensors = list()
+        newlens, new_rope = list(), list()
+        for image, curr_kvlen, curr_position_id in zip(images, curr_kvlens, curr_rope):
+            packed_key_value_indexes.extend(range(curr, curr + curr_kvlen))
+            curr += curr_kvlen
+
+            packed_text_ids.append(new_token_ids['start_of_image'])
+            packed_text_indexes.append(_curr)
+            packed_indexes.append(curr)
+            curr += 1
+            _curr += 1
+
+            image_tensor = transforms(image)
+            vae_image_tensors.append(image_tensor)
+            vae_posiiton_ids = self.get_flattened_position_ids(
+                image_tensor.size(1), image_tensor.size(2),
+                self.latent_downsample, 
+                max_num_patches_per_side=self.max_latent_size
+            )
+            packed_vae_position_ids.append(vae_posiiton_ids)
+            H, W = image_tensor.shape[1:]
+            h = H // self.latent_downsample
+            w = W // self.latent_downsample
+            patchified_vae_latent_shapes.append((h, w))
+
+            num_img_tokens = w * h
+            packed_vae_token_indexes.extend(range(_curr, _curr + num_img_tokens))
+            packed_indexes.extend(range(curr, curr + num_img_tokens))
+            curr += num_img_tokens
+            _curr += num_img_tokens
+
+            packed_text_ids.append(new_token_ids['end_of_image'])
+            packed_text_indexes.append(_curr)
+            packed_indexes.append(curr)
+            curr += 1
+            _curr += 1
+
+            packed_position_ids.extend([curr_position_id] * (num_img_tokens + 2))
+            packed_seqlens.append(num_img_tokens + 2)
+            newlens.append(curr_kvlen + num_img_tokens + 2)
+            new_rope.append(curr_position_id + 1)
+
+        image_sizes = [item.shape for item in vae_image_tensors]
+        max_image_size = [max(item) for item in list(zip(*image_sizes))]
+        padded_images = torch.zeros(size=(len(vae_image_tensors), *max_image_size))
+        for i, image_tensor in enumerate(vae_image_tensors):
+            padded_images[i, :, :image_tensor.shape[1], :image_tensor.shape[2]] = image_tensor
+
+        generation_input = {
+            "padded_images": padded_images,
+            "patchified_vae_latent_shapes": patchified_vae_latent_shapes,
+            "packed_vae_position_ids": torch.cat(packed_vae_position_ids, dim=0),
+            "packed_timesteps": torch.tensor([timestep]),
+            "packed_vae_token_indexes": torch.tensor(packed_vae_token_indexes, dtype=torch.long),
+            "packed_text_ids": torch.tensor(packed_text_ids, dtype=torch.long),
+            "packed_text_indexes": torch.tensor(packed_text_indexes, dtype=torch.long),
+            "packed_position_ids": torch.tensor(packed_position_ids, dtype=torch.long),
+            "packed_seqlens": torch.tensor(packed_seqlens, dtype=torch.int),
+            "packed_indexes": torch.tensor(packed_indexes, dtype=torch.long),
+            "packed_key_value_indexes": torch.tensor(packed_key_value_indexes, dtype=torch.long),
+            "key_values_lens": torch.tensor(curr_kvlens, dtype=torch.int),
+        }
+
+        return generation_input, newlens, new_rope
+
+    @torch.no_grad
+    def forward_cache_update_vae(
+        self,
+        vae_model,
+        past_key_values: NaiveCache,
+        padded_images: torch.Tensor,
+        patchified_vae_latent_shapes: List,
+        packed_vae_position_ids: torch.LongTensor,
+        packed_timesteps: torch.Tensor,
+        packed_vae_token_indexes: torch.LongTensor,
+        packed_text_ids: torch.LongTensor,
+        packed_text_indexes: torch.LongTensor,
+        packed_position_ids: torch.LongTensor,
+        packed_seqlens: torch.IntTensor,
+        packed_indexes: torch.LongTensor,
+        key_values_lens: torch.IntTensor,
+        packed_key_value_indexes: torch.Tensor,
+    ):
+        packed_text_embedding = self.language_model.model.embed_tokens(packed_text_ids)
+        packed_sequence = packed_text_embedding.new_zeros((sum(packed_seqlens), self.hidden_size))
+        packed_sequence[packed_text_indexes] = packed_text_embedding
+
+        padded_latent = vae_model.encode(padded_images)
+
+        p = self.latent_patch_size
+        packed_latent = list()
+        for latent, (h, w) in zip(padded_latent, patchified_vae_latent_shapes):
+            latent = latent[:, :h * p, :w * p].reshape(self.latent_channel, h, p, w, p)
+            latent = torch.einsum("chpwq->hwpqc", latent).reshape(-1, p * p * self.latent_channel)
+            packed_latent.append(latent)
+        packed_latent = torch.cat(packed_latent, dim=0)
+        packed_pos_embed = self.latent_pos_embed(packed_vae_position_ids)
+        packed_timestep_embeds = self.time_embedder(packed_timesteps)
+        packed_latent = self.vae2llm(packed_latent) + packed_timestep_embeds + packed_pos_embed
+        packed_sequence[packed_vae_token_indexes] = packed_latent
+
+        extra_inputs = {}
+        if self.use_moe:
+            extra_inputs = {
+                "mode": "gen",
+                "packed_vae_token_indexes": packed_vae_token_indexes,
+                "packed_text_indexes": packed_text_indexes
+            }
+
+        output = self.language_model.forward_inference(
+            packed_query_sequence=packed_sequence,
+            query_lens=packed_seqlens,
+            packed_query_position_ids=packed_position_ids,
+            packed_query_indexes=packed_indexes,
+            past_key_values=past_key_values,
+            key_values_lens=key_values_lens,
+            packed_key_value_indexes=packed_key_value_indexes,
+            update_past_key_values=True,
+            is_causal=False,
+            **extra_inputs,
+        )
+        past_key_values = output.past_key_values
+
+        return past_key_values
+
+    def prepare_vae_latent(self, curr_kvlens, curr_rope, image_sizes, new_token_ids):
+        packed_text_ids, packed_text_indexes = list(), list()
+        packed_vae_position_ids, packed_vae_token_indexes, packed_init_noises = list(), list(), list()
+        packed_position_ids, packed_seqlens, packed_indexes = list(), list(), list()
+        packed_key_value_indexes = list()
+
+        query_curr = curr = 0
+        for (H, W), curr_kvlen, curr_position_id in zip(image_sizes, curr_kvlens, curr_rope):
+            packed_key_value_indexes.extend(range(curr, curr + curr_kvlen))
+            curr += curr_kvlen
+
+            packed_text_ids.append(new_token_ids['start_of_image'])
+            packed_text_indexes.append(query_curr)
+            packed_indexes.append(curr)
+            curr += 1
+            query_curr += 1
+
+            vae_posiiton_ids = self.get_flattened_position_ids(
+                H, W,
+                self.latent_downsample, 
+                max_num_patches_per_side=self.max_latent_size
+            )
+            packed_vae_position_ids.append(vae_posiiton_ids)
+
+            h, w = H // self.latent_downsample, W // self.latent_downsample
+            num_image_tokens = h * w
+            packed_init_noises.append(
+                torch.randn(num_image_tokens, self.latent_channel * self.latent_patch_size ** 2)
+            )
+            packed_vae_token_indexes.extend(range(query_curr, query_curr + num_image_tokens))
+            packed_indexes.extend(range(curr, curr + num_image_tokens))
+            curr += num_image_tokens
+            query_curr += num_image_tokens
+
+            packed_text_ids.append(new_token_ids['end_of_image'])
+            packed_text_indexes.append(query_curr)
+            packed_indexes.append(curr)
+            curr += 1
+            query_curr += 1
+
+            packed_position_ids.extend([curr_position_id] * (num_image_tokens + 2))
+            packed_seqlens.append(num_image_tokens + 2)
+
+        generation_input = {
+            "packed_text_ids": torch.tensor(packed_text_ids, dtype=torch.long),
+            "packed_text_indexes": torch.tensor(packed_text_indexes, dtype=torch.long),
+            "packed_init_noises": torch.cat(packed_init_noises, dim=0),
+            "packed_vae_position_ids": torch.cat(packed_vae_position_ids, dim=0),
+            "packed_vae_token_indexes": torch.tensor(packed_vae_token_indexes, dtype=torch.long),
+            "packed_seqlens": torch.tensor(packed_seqlens, dtype=torch.int),
+            "packed_position_ids": torch.tensor(packed_position_ids, dtype=torch.long),
+            "key_values_lens": torch.tensor(curr_kvlens, dtype=torch.int),
+            "packed_indexes": torch.tensor(packed_indexes, dtype=torch.long),
+            "packed_key_value_indexes": torch.tensor(packed_key_value_indexes, dtype=torch.long),
+        }
+
+        return generation_input
+
+    def prepare_vae_latent_cfg(self, curr_kvlens, curr_rope, image_sizes):
+        packed_position_ids, packed_indexes, packed_key_value_indexes = list(), list(), list()
+
+        query_curr = curr = 0
+        for (H, W), curr_kvlen, curr_position_id in zip(image_sizes, curr_kvlens, curr_rope):
+            packed_key_value_indexes.extend(range(curr, curr + curr_kvlen))
+            curr += curr_kvlen
+
+            packed_indexes.append(curr)
+            curr += 1
+            query_curr += 1
+
+            h, w = H // self.latent_downsample, W // self.latent_downsample
+            num_image_tokens = h * w
+            packed_indexes.extend(range(curr, curr + num_image_tokens))
+            curr += num_image_tokens
+            query_curr += num_image_tokens
+
+            packed_indexes.append(curr)
+            curr += 1
+            query_curr += 1
+
+            packed_position_ids.extend([curr_position_id] * (num_image_tokens + 2))
+
+        generation_input = {
+            "cfg_packed_position_ids": torch.tensor(packed_position_ids, dtype=torch.long),
+            "cfg_key_values_lens": torch.tensor(curr_kvlens, dtype=torch.int),
+            "cfg_packed_query_indexes": torch.tensor(packed_indexes, dtype=torch.long),
+            "cfg_packed_key_value_indexes": torch.tensor(packed_key_value_indexes, dtype=torch.long),
+        }
+
+        return generation_input
+
+    @torch.no_grad
+    def generate_image(
+        self,
+        packed_text_ids: torch.LongTensor,
+        packed_text_indexes: torch.LongTensor,
+        packed_init_noises: torch.Tensor,
+        packed_vae_position_ids: torch.LongTensor,
+        packed_vae_token_indexes: torch.LongTensor,
+        packed_seqlens: torch.IntTensor,
+        packed_position_ids: torch.LongTensor,
+        packed_indexes: torch.LongTensor,
+        past_key_values: NaiveCache,
+        key_values_lens: torch.IntTensor,
+        packed_key_value_indexes: torch.LongTensor,
+        num_timesteps: int = 24,
+        timestep_shift: float = 1.0,
+        cfg_renorm_min: float = 0.0,
+        cfg_renorm_type: str = "global",
+        cfg_interval: Optional[Tuple[float, float]] = [0, 1],
+        # cfg_text
+        cfg_text_scale: float = 1.0,
+        cfg_text_packed_query_indexes: Optional[torch.LongTensor] = None,
+        cfg_text_packed_position_ids: Optional[torch.LongTensor] = None,
+        cfg_text_past_key_values: Optional[NaiveCache] = None,
+        cfg_text_key_values_lens: Optional[torch.IntTensor] = None,
+        cfg_text_packed_key_value_indexes: Optional[torch.LongTensor] = None,
+        # cfg_img
+        cfg_img_scale: float = 1.0,
+        cfg_img_packed_query_indexes: Optional[torch.LongTensor] = None,
+        cfg_img_packed_position_ids: Optional[torch.LongTensor] = None,
+        cfg_img_past_key_values: Optional[NaiveCache] = None,
+        cfg_img_key_values_lens: Optional[torch.IntTensor] = None,
+        cfg_img_packed_key_value_indexes: Optional[torch.LongTensor] = None,
+        cfg_type: str = "parallel",
+    ):
+        x_t = packed_init_noises
+
+        timesteps = torch.linspace(1, 0, num_timesteps, device=x_t.device)
+        timesteps = timestep_shift * timesteps / (1 + (timestep_shift - 1) * timesteps)
+        dts =  timesteps[:-1] - timesteps[1:]
+        timesteps = timesteps[:-1]
+
+        for i, t in enumerate(timesteps):
+
+            timestep = torch.tensor([t] * x_t.shape[0], device=x_t.device)
+            if t > cfg_interval[0] and t <= cfg_interval[1]:
+                cfg_text_scale_ = cfg_text_scale
+                cfg_img_scale_ = cfg_img_scale
+            else:
+                cfg_text_scale_ = 1.0
+                cfg_img_scale_ = 1.0
+            v_t = self._forward_flow(
+                x_t=x_t,
+                timestep=timestep, 
+                packed_vae_token_indexes=packed_vae_token_indexes,
+                packed_vae_position_ids=packed_vae_position_ids,
+                packed_text_ids=packed_text_ids,
+                packed_text_indexes=packed_text_indexes,
+                packed_position_ids=packed_position_ids,
+                packed_indexes=packed_indexes,
+                packed_seqlens=packed_seqlens,
+                key_values_lens=key_values_lens,
+                past_key_values=past_key_values,
+                packed_key_value_indexes=packed_key_value_indexes,
+                cfg_renorm_min=cfg_renorm_min,
+                cfg_renorm_type=cfg_renorm_type,
+                # cfg_text
+                cfg_text_scale=cfg_text_scale_,
+                cfg_text_packed_position_ids=cfg_text_packed_position_ids,
+                cfg_text_packed_query_indexes=cfg_text_packed_query_indexes,
+                cfg_text_key_values_lens=cfg_text_key_values_lens,
+                cfg_text_past_key_values=cfg_text_past_key_values,
+                cfg_text_packed_key_value_indexes=cfg_text_packed_key_value_indexes,
+                # cfg_img
+                cfg_img_scale=cfg_img_scale_,
+                cfg_img_packed_position_ids=cfg_img_packed_position_ids,
+                cfg_img_packed_query_indexes=cfg_img_packed_query_indexes,
+                cfg_img_key_values_lens=cfg_img_key_values_lens,
+                cfg_img_past_key_values=cfg_img_past_key_values,
+                cfg_img_packed_key_value_indexes=cfg_img_packed_key_value_indexes,
+                cfg_type=cfg_type,
+            )
+
+            x_t = x_t - v_t.to(x_t.device) * dts[i] # velocity pointing from data to noise
+
+        unpacked_latent = x_t.split((packed_seqlens - 2).tolist())
+        return unpacked_latent
+
+    @torch.no_grad
+    def _forward_flow(
+        self,
+        x_t: torch.Tensor,
+        timestep: torch.LongTensor,
+        packed_vae_token_indexes: torch.LongTensor,
+        packed_vae_position_ids: torch.LongTensor,
+        packed_text_ids: torch.LongTensor,
+        packed_text_indexes: torch.LongTensor,
+        packed_indexes: torch.LongTensor,
+        packed_position_ids: torch.LongTensor,
+        packed_seqlens: torch.IntTensor,
+        key_values_lens: torch.IntTensor,
+        past_key_values: NaiveCache,
+        packed_key_value_indexes: torch.LongTensor,
+        cfg_renorm_min: float = 0.0,
+        cfg_renorm_type: str = "global",
+        # cfg_text
+        cfg_text_scale: float = 1.0,
+        cfg_text_packed_position_ids: Optional[torch.LongTensor] = None,
+        cfg_text_packed_query_indexes: Optional[torch.LongTensor] = None,
+        cfg_text_key_values_lens: Optional[torch.Tensor] = None,
+        cfg_text_past_key_values: Optional[NaiveCache] = None,
+        cfg_text_packed_key_value_indexes: Optional[torch.LongTensor] = None,
+        # cfg_img
+        cfg_img_scale: float = 1.0,
+        cfg_img_packed_position_ids: Optional[torch.LongTensor] = None,
+        cfg_img_packed_query_indexes: Optional[torch.LongTensor] = None,
+        cfg_img_key_values_lens: Optional[torch.Tensor] = None,
+        cfg_img_past_key_values: Optional[NaiveCache] = None,
+        cfg_img_packed_key_value_indexes: Optional[torch.LongTensor] = None,
+        cfg_type: str = "parallel",
+    ):
+        packed_text_embedding = self.language_model.model.embed_tokens(packed_text_ids)
+        packed_sequence = packed_text_embedding.new_zeros((sum(packed_seqlens), self.hidden_size))
+        packed_sequence[packed_text_indexes] = packed_text_embedding
+
+        assert timestep.unique().shape[0] == 1
+        packed_pos_embed = self.latent_pos_embed(packed_vae_position_ids)
+        packed_timestep_embeds = self.time_embedder(timestep)
+        x_t = self.vae2llm(x_t) + packed_timestep_embeds + packed_pos_embed
+        packed_sequence[packed_vae_token_indexes] = x_t
+
+        extra_inputs = {}
+        if self.use_moe:
+            extra_inputs = {
+                "mode": "gen",
+                "packed_vae_token_indexes": packed_vae_token_indexes,
+                "packed_text_indexes": packed_text_indexes
+            }
+
+        output = self.language_model.forward_inference(
+            packed_query_sequence=packed_sequence,
+            query_lens=packed_seqlens,
+            packed_query_position_ids=packed_position_ids,
+            packed_query_indexes=packed_indexes,
+            past_key_values=past_key_values,
+            key_values_lens=key_values_lens,
+            packed_key_value_indexes=packed_key_value_indexes,
+            update_past_key_values=False,
+            is_causal=False,
+            **extra_inputs,
+        )
+        v_t = self.llm2vae(output.packed_query_sequence)
+        v_t = v_t[packed_vae_token_indexes]
+
+        if cfg_text_scale > 1.0:
+            cfg_text_output = self.language_model.forward_inference(
+                packed_query_sequence=packed_sequence,
+                query_lens=packed_seqlens,
+                packed_query_position_ids=cfg_text_packed_position_ids,
+                packed_query_indexes=cfg_text_packed_query_indexes,
+                past_key_values=cfg_text_past_key_values,
+                key_values_lens=cfg_text_key_values_lens,
+                packed_key_value_indexes=cfg_text_packed_key_value_indexes,
+                update_past_key_values=False,
+                is_causal=False,
+                **extra_inputs,
+            )
+            cfg_text_v_t = self.llm2vae(cfg_text_output.packed_query_sequence)
+            cfg_text_v_t = cfg_text_v_t[packed_vae_token_indexes]
+
+        if cfg_img_scale > 1.0:
+            cfg_img_output = self.language_model.forward_inference(
+                packed_query_sequence=packed_sequence,
+                query_lens=packed_seqlens,
+                packed_query_position_ids=cfg_img_packed_position_ids,
+                packed_query_indexes=cfg_img_packed_query_indexes,
+                past_key_values=cfg_img_past_key_values,
+                key_values_lens=cfg_img_key_values_lens,
+                packed_key_value_indexes=cfg_img_packed_key_value_indexes,
+                update_past_key_values=False,
+                is_causal=False,
+                **extra_inputs,
+            )
+            cfg_img_v_t = self.llm2vae(cfg_img_output.packed_query_sequence)
+            cfg_img_v_t = cfg_img_v_t[packed_vae_token_indexes]
+
+        if cfg_text_scale > 1.0:
+            if cfg_renorm_type == "text_channel":
+                v_t_text_ = cfg_text_v_t + cfg_text_scale * (v_t - cfg_text_v_t)
+                norm_v_t = torch.norm(v_t, dim=-1, keepdim=True)
+                norm_v_t_text_ = torch.norm(v_t_text_, dim=-1, keepdim=True)
+                scale = (norm_v_t / (norm_v_t_text_ + 1e-8)).clamp(min=cfg_renorm_min, max=1.0)
+                v_t_text = v_t_text_ * scale
+                if cfg_img_scale > 1.0:
+                    v_t = cfg_img_v_t + cfg_img_scale * (v_t_text - cfg_img_v_t)
+                else:
+                    v_t = v_t_text
+            else:
+                v_t_text_ = cfg_text_v_t + cfg_text_scale * (v_t - cfg_text_v_t)
+                
+                if cfg_img_scale > 1.0:
+                    v_t_ = cfg_img_v_t + cfg_img_scale * (v_t_text_ - cfg_img_v_t)
+                else:
+                    v_t_ = v_t_text_
+
+                # NOTE norm is computed over all dimensions, thus currently only supports batch_size = 1 with navit
+                if cfg_renorm_type == "global":
+                    norm_v_t = torch.norm(v_t)
+                    norm_v_t_ = torch.norm(v_t_)
+                elif cfg_renorm_type == "channel":
+                    norm_v_t = torch.norm(v_t, dim=-1, keepdim=True)
+                    norm_v_t_ = torch.norm(v_t_, dim=-1, keepdim=True)
+                else:
+                    raise NotImplementedError(f"{cfg_renorm_type} is not suppoprted")
+                scale = (norm_v_t / (norm_v_t_ + 1e-8)).clamp(min=cfg_renorm_min, max=1.0)
+                v_t = v_t_ * scale
+        else:
+            # No CFG
+            pass
+
+        return v_t
+
+    def prepare_start_tokens(self, curr_kvlens, curr_rope, new_token_ids):
+        packed_start_tokens, packed_key_value_indexes = list(), list()
+        packed_query_position_ids = list()
+
+        curr = 0
+        for curr_kvlen, curr_position_id in zip(curr_kvlens, curr_rope):
+            packed_key_value_indexes.extend(range(curr, curr + curr_kvlen))
+            packed_start_tokens.append(new_token_ids['bos_token_id'])
+            packed_query_position_ids.append(curr_position_id)
+            curr += curr_kvlen
+
+        generation_input = {
+            "packed_start_tokens": torch.tensor(packed_start_tokens, dtype=torch.long),
+            "packed_query_position_ids": torch.tensor(packed_query_position_ids, dtype=torch.long),
+            "key_values_lens": torch.tensor(curr_kvlens, dtype=torch.int),
+            "packed_key_value_indexes": torch.tensor(packed_key_value_indexes, dtype=torch.long),
+        }
+
+        return generation_input
+
+    @torch.no_grad
+    def generate_text(
+        self,
+        past_key_values: NaiveCache,
+        packed_key_value_indexes: torch.LongTensor,
+        key_values_lens: torch.IntTensor,
+        packed_start_tokens: torch.LongTensor,
+        packed_query_position_ids: torch.LongTensor,
+        max_length: int,
+        do_sample: bool = False,
+        temperature: float = 1.0,
+        end_token_id: int = None,
+    ):
+        step = 0
+        # generated_sequence = [] # Removed for streaming
+        curr_tokens = packed_start_tokens
+        while step < max_length:
+            # generated_sequence.append(curr_tokens) # Removed for streaming
+            yield curr_tokens # Yield current tokens
+
+            packed_text_embedding = self.language_model.model.embed_tokens(curr_tokens)
+            query_lens = torch.ones_like(curr_tokens)
+            packed_query_indexes = torch.cumsum(key_values_lens, dim=0) + torch.arange(
+                0, len(key_values_lens), 
+                device=key_values_lens.device, 
+                dtype=key_values_lens.dtype
+            )
+
+            uppacked = list(packed_key_value_indexes.split(key_values_lens.tolist(), dim=0))
+            for i in range(len(uppacked)):
+                uppacked[i] += i
+            packed_key_value_indexes = torch.cat(uppacked, dim=0)
+
+            extra_inputs = {}
+            if self.use_moe:
+                extra_inputs = {"mode": "und"}
+
+            output = self.language_model.forward_inference(
+                packed_query_sequence=packed_text_embedding,
+                query_lens=query_lens,
+                packed_query_position_ids=packed_query_position_ids,
+                packed_query_indexes=packed_query_indexes,
+                past_key_values=past_key_values,
+                key_values_lens=key_values_lens,
+                packed_key_value_indexes=packed_key_value_indexes,
+                update_past_key_values=True,
+                is_causal=True,
+                **extra_inputs,
+            )
+            past_key_values = output.past_key_values
+            packed_query_sequence = output.packed_query_sequence
+            pred_logits = self.language_model.lm_head(packed_query_sequence)
+
+            if do_sample:
+                probs = nn.functional.softmax(pred_logits / temperature, dim=-1)
+                curr_tokens = torch.multinomial(probs, num_samples=1).squeeze(1)
+            else:
+                curr_tokens = torch.argmax(pred_logits, dim=-1)
+
+            uppacked = list(packed_key_value_indexes.split(key_values_lens.tolist(), dim=0))
+            for i in range(len(uppacked)):
+                uppacked[i] = torch.cat(
+                    [uppacked[i], torch.tensor([uppacked[i][-1] + 1], device=uppacked[i].device)], dim=0
+                )
+            packed_key_value_indexes = torch.cat(uppacked, dim=0)
+            key_values_lens = key_values_lens + 1
+            packed_query_position_ids = packed_query_position_ids + 1
+            step += 1
+
+            if end_token_id is not None and curr_tokens[0] == end_token_id: # only support batch=1
+                break
+
+        # output_device = generated_sequence[0].device # Removed for streaming
+        # return torch.stack([i.to(output_device) for i in generated_sequence], dim=0) # Removed for streaming
+
+    # for evaluation
+    @torch.no_grad()
+    def chat(
+        self,
+        tokenizer,
+        new_token_ids,
+        image_transform,
+        images,
+        prompt,
+        max_length: int,
+        do_sample: bool = False,
+        temperature: float = 1.0,
+    ):
+        device = next(self.parameters()).device
+
+        if isinstance(new_token_ids, dict):
+            for k, v in new_token_ids.items():
+                if torch.is_tensor(v):
+                    new_token_ids[k] = v.to(device)
+        elif torch.is_tensor(new_token_ids):
+            new_token_ids = new_token_ids.to(device)
+
+        # prefill
+        past_key_values = NaiveCache(self.config.llm_config.num_hidden_layers)
+        newlens = [0]
+        new_rope = [0]
+
+        # add images
+        for image in images:
+            generation_input, newlens, new_rope = self.prepare_vit_images(
+                curr_kvlens=newlens,
+                curr_rope=new_rope, 
+                images=[image], 
+                transforms=image_transform,
+                new_token_ids=new_token_ids,
+            )
+            for k, v in generation_input.items():
+                if torch.is_tensor(v):
+                    generation_input[k] = v.to(device)
+            with torch.amp.autocast("cuda", enabled=True, dtype=torch.bfloat16):
+                past_key_values = self.forward_cache_update_vit(past_key_values, **generation_input)
+
+        # add text
+        generation_input, newlens, new_rope = self.prepare_prompts(
+            curr_kvlens=newlens,
+            curr_rope=new_rope, 
+            prompts=[prompt],
+            tokenizer=tokenizer, 
+            new_token_ids=new_token_ids,
+        )
+        for k, v in generation_input.items():
+            if torch.is_tensor(v):
+                generation_input[k] = v.to(device)
+        with torch.amp.autocast("cuda", enabled=True, dtype=torch.bfloat16):
+            past_key_values = self.forward_cache_update_text(past_key_values, **generation_input)
+
+        # decode
+        generation_input = self.prepare_start_tokens(newlens, new_rope, new_token_ids)
+        for k, v in generation_input.items():
+            if torch.is_tensor(v):
+                generation_input[k] = v.to(device)
+        with torch.amp.autocast("cuda", enabled=True, dtype=torch.bfloat16):
+            for unpacked_latent in self.generate_text(
+                past_key_values=past_key_values,
+                max_length=max_length,
+                do_sample=do_sample,
+                temperature=temperature,
+                end_token_id=new_token_ids['eos_token_id'],
+                **generation_input,
+            ):
+                output = tokenizer.decode(unpacked_latent[:,0])
+                yield output

modeling/bagel/modeling_utils.py

@@ -1,144 +1,144 @@
-# Copyright (c) 2022 Facebook, Inc. and its affiliates.
-# Copyright (c) 2025 Bytedance Ltd. and/or its affiliates.
-# SPDX-License-Identifier: CC BY-NC 4.0
-#
-# This file has been modified by ByteDance Ltd. and/or its affiliates. on 2025-05-20.
-#
-# Original file was released under CC BY-NC 4.0, with the full license text
-# available at https://github.com/facebookresearch/DiT/blob/main/LICENSE.txt.
-#
-# This modified file is released under the same license.
-
-import math
-
-import numpy as np
-import torch
-from torch import nn
-from transformers.activations import ACT2FN
-
-# --------------------------------------------------------
-# 2D sine-cosine position embedding
-# References:
-# DiT: https://github.com/facebookresearch/DiT/blob/main/models.py
-# --------------------------------------------------------
-def get_2d_sincos_pos_embed(embed_dim, grid_size, cls_token=False, extra_tokens=0):
-    grid_h = np.arange(grid_size, dtype=np.float32)
-    grid_w = np.arange(grid_size, dtype=np.float32)
-    grid = np.meshgrid(grid_w, grid_h)  # here w goes first
-    grid = np.stack(grid, axis=0)
-
-    grid = grid.reshape([2, 1, grid_size, grid_size])
-    pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid)
-    if cls_token and extra_tokens > 0:
-        pos_embed = np.concatenate([np.zeros([extra_tokens, embed_dim]), pos_embed], axis=0)
-    return pos_embed
-
-
-def get_2d_sincos_pos_embed_from_grid(embed_dim, grid):
-    assert embed_dim % 2 == 0
-
-    # use half of dimensions to encode grid_h
-    emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0])  # (H*W, D/2)
-    emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1])  # (H*W, D/2)
-
-    emb = np.concatenate([emb_h, emb_w], axis=1) # (H*W, D)
-    return emb
-
-
-def get_1d_sincos_pos_embed_from_grid(embed_dim, pos):
-    """
-    embed_dim: output dimension for each position
-    pos: a list of positions to be encoded: size (M,)
-    out: (M, D)
-    """
-    assert embed_dim % 2 == 0
-    omega = np.arange(embed_dim // 2, dtype=np.float64)
-    omega /= embed_dim / 2.
-    omega = 1. / 10000**omega  # (D/2,)
-
-    pos = pos.reshape(-1)  # (M,)
-    out = np.einsum('m,d->md', pos, omega)  # (M, D/2), outer product
-
-    emb_sin = np.sin(out) # (M, D/2)
-    emb_cos = np.cos(out) # (M, D/2)
-
-    emb = np.concatenate([emb_sin, emb_cos], axis=1)  # (M, D)
-    return emb
-
-
-# --------------------------------------------------------
-# TimestepEmbedder
-# Reference:
-# DiT: https://github.com/facebookresearch/DiT/blob/main/models.py
-# --------------------------------------------------------
-class TimestepEmbedder(nn.Module):
-    """
-    Embeds scalar timesteps into vector representations.
-    """
-    def __init__(self, hidden_size, frequency_embedding_size=256):
-        super().__init__()
-        self.mlp = nn.Sequential(
-            nn.Linear(frequency_embedding_size, hidden_size, bias=True),
-            nn.SiLU(),
-            nn.Linear(hidden_size, hidden_size, bias=True),
-        )
-        self.frequency_embedding_size = frequency_embedding_size
-
-    @staticmethod
-    def timestep_embedding(t, dim, max_period=10000):
-        """
-        Create sinusoidal timestep embeddings.
-        :param t: a 1-D Tensor of N indices, one per batch element.
-                          These may be fractional.
-        :param dim: the dimension of the output.
-        :param max_period: controls the minimum frequency of the embeddings.
-        :return: an (N, D) Tensor of positional embeddings.
-        """
-        half = dim // 2
-        freqs = torch.exp(
-            -math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half
-        ).to(device=t.device)
-        args = t[:, None].float() * freqs[None]
-        embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
-        if dim % 2:
-            embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
-        return embedding
-
-    def forward(self, t):
-        t_freq = self.timestep_embedding(t, self.frequency_embedding_size)
-        t_emb = self.mlp(t_freq)
-        return t_emb
-
-
-class MLPconnector(nn.Module):
-    def __init__(self, in_dim: int, out_dim: int, hidden_act: str):
-        super().__init__()
-        self.activation_fn = ACT2FN[hidden_act]
-        self.fc1 = nn.Linear(in_dim, out_dim)
-        self.fc2 = nn.Linear(out_dim, out_dim)
-
-    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
-        hidden_states = self.fc1(hidden_states)
-        hidden_states = self.activation_fn(hidden_states)
-        hidden_states = self.fc2(hidden_states)
-        return hidden_states
-
-
-class PositionEmbedding(nn.Module):
-    def __init__(self, max_num_patch_per_side, hidden_size):
-        super().__init__()
-        self.max_num_patch_per_side = max_num_patch_per_side
-        self.hidden_size = hidden_size
-        self.pos_embed = nn.Parameter(
-            torch.zeros(max_num_patch_per_side ** 2, hidden_size), 
-            requires_grad=False
-        )
-        self._init_weights()
-
-    def _init_weights(self):
-        # Initialize (and freeze) pos_embed by sin-cos embedding:
-        pos_embed = get_2d_sincos_pos_embed(self.hidden_size, self.max_num_patch_per_side)
-        self.pos_embed.data.copy_(torch.from_numpy(pos_embed).float())
-
-    def forward(self, position_ids):
+# Copyright (c) 2022 Facebook, Inc. and its affiliates.
+# Copyright (c) 2025 Bytedance Ltd. and/or its affiliates.
+# SPDX-License-Identifier: CC BY-NC 4.0
+#
+# This file has been modified by ByteDance Ltd. and/or its affiliates. on 2025-05-20.
+#
+# Original file was released under CC BY-NC 4.0, with the full license text
+# available at https://github.com/facebookresearch/DiT/blob/main/LICENSE.txt.
+#
+# This modified file is released under the same license.
+
+import math
+
+import numpy as np
+import torch
+from torch import nn
+from transformers.activations import ACT2FN
+
+# --------------------------------------------------------
+# 2D sine-cosine position embedding
+# References:
+# DiT: https://github.com/facebookresearch/DiT/blob/main/models.py
+# --------------------------------------------------------
+def get_2d_sincos_pos_embed(embed_dim, grid_size, cls_token=False, extra_tokens=0):
+    grid_h = np.arange(grid_size, dtype=np.float32)
+    grid_w = np.arange(grid_size, dtype=np.float32)
+    grid = np.meshgrid(grid_w, grid_h)  # here w goes first
+    grid = np.stack(grid, axis=0)
+
+    grid = grid.reshape([2, 1, grid_size, grid_size])
+    pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid)
+    if cls_token and extra_tokens > 0:
+        pos_embed = np.concatenate([np.zeros([extra_tokens, embed_dim]), pos_embed], axis=0)
+    return pos_embed
+
+
+def get_2d_sincos_pos_embed_from_grid(embed_dim, grid):
+    assert embed_dim % 2 == 0
+
+    # use half of dimensions to encode grid_h
+    emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0])  # (H*W, D/2)
+    emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1])  # (H*W, D/2)
+
+    emb = np.concatenate([emb_h, emb_w], axis=1) # (H*W, D)
+    return emb
+
+
+def get_1d_sincos_pos_embed_from_grid(embed_dim, pos):
+    """
+    embed_dim: output dimension for each position
+    pos: a list of positions to be encoded: size (M,)
+    out: (M, D)
+    """
+    assert embed_dim % 2 == 0
+    omega = np.arange(embed_dim // 2, dtype=np.float64)
+    omega /= embed_dim / 2.
+    omega = 1. / 10000**omega  # (D/2,)
+
+    pos = pos.reshape(-1)  # (M,)
+    out = np.einsum('m,d->md', pos, omega)  # (M, D/2), outer product
+
+    emb_sin = np.sin(out) # (M, D/2)
+    emb_cos = np.cos(out) # (M, D/2)
+
+    emb = np.concatenate([emb_sin, emb_cos], axis=1)  # (M, D)
+    return emb
+
+
+# --------------------------------------------------------
+# TimestepEmbedder
+# Reference:
+# DiT: https://github.com/facebookresearch/DiT/blob/main/models.py
+# --------------------------------------------------------
+class TimestepEmbedder(nn.Module):
+    """
+    Embeds scalar timesteps into vector representations.
+    """
+    def __init__(self, hidden_size, frequency_embedding_size=256):
+        super().__init__()
+        self.mlp = nn.Sequential(
+            nn.Linear(frequency_embedding_size, hidden_size, bias=True),
+            nn.SiLU(),
+            nn.Linear(hidden_size, hidden_size, bias=True),
+        )
+        self.frequency_embedding_size = frequency_embedding_size
+
+    @staticmethod
+    def timestep_embedding(t, dim, max_period=10000):
+        """
+        Create sinusoidal timestep embeddings.
+        :param t: a 1-D Tensor of N indices, one per batch element.
+                          These may be fractional.
+        :param dim: the dimension of the output.
+        :param max_period: controls the minimum frequency of the embeddings.
+        :return: an (N, D) Tensor of positional embeddings.
+        """
+        half = dim // 2
+        freqs = torch.exp(
+            -math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half
+        ).to(device=t.device)
+        args = t[:, None].float() * freqs[None]
+        embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
+        if dim % 2:
+            embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
+        return embedding
+
+    def forward(self, t):
+        t_freq = self.timestep_embedding(t, self.frequency_embedding_size)
+        t_emb = self.mlp(t_freq)
+        return t_emb
+
+
+class MLPconnector(nn.Module):
+    def __init__(self, in_dim: int, out_dim: int, hidden_act: str):
+        super().__init__()
+        self.activation_fn = ACT2FN[hidden_act]
+        self.fc1 = nn.Linear(in_dim, out_dim)
+        self.fc2 = nn.Linear(out_dim, out_dim)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.fc1(hidden_states)
+        hidden_states = self.activation_fn(hidden_states)
+        hidden_states = self.fc2(hidden_states)
+        return hidden_states
+
+
+class PositionEmbedding(nn.Module):
+    def __init__(self, max_num_patch_per_side, hidden_size):
+        super().__init__()
+        self.max_num_patch_per_side = max_num_patch_per_side
+        self.hidden_size = hidden_size
+        self.pos_embed = nn.Parameter(
+            torch.zeros(max_num_patch_per_side ** 2, hidden_size), 
+            requires_grad=False
+        )
+        self._init_weights()
+
+    def _init_weights(self):
+        # Initialize (and freeze) pos_embed by sin-cos embedding:
+        pos_embed = get_2d_sincos_pos_embed(self.hidden_size, self.max_num_patch_per_side)
+        self.pos_embed.data.copy_(torch.from_numpy(pos_embed).float())
+
+    def forward(self, position_ids):
         return self.pos_embed[position_ids]

modeling/bagel/siglip_navit.py

@@ -1,402 +1,402 @@
-# Copyright (c) 2024 The HuggingFace Inc. team.
-# Copyright (c) 2025 Bytedance Ltd. and/or its affiliates.
-# SPDX-License-Identifier: Apache-2.0
-#
-# This file has been modified by ByteDance Ltd. and/or its affiliates. on 2025-05-20.
-#
-# Original file was released under Apache-2.0, with the full license text
-# available at https://github.com/huggingface/transformers/blob/main/LICENSE.
-#
-# This modified file is released under the same license.
-
-import torch
-from torch import nn
-
-from transformers.activations import ACT2FN
-from modeling.siglip.configuration_siglip import SiglipVisionConfig as _SiglipVisionConfig
-from modeling.siglip.modeling_siglip import SiglipAttention, SiglipPreTrainedModel
-from flash_attn import flash_attn_varlen_func
-
-
-class SiglipVisionConfig(_SiglipVisionConfig):
-    r"""
-    This is the configuration class to store the configuration of a [`SiglipVisionModel`]. It is used to instantiate a
-    Siglip vision encoder according to the specified arguments, defining the model architecture. Instantiating a
-    configuration with the defaults will yield a similar configuration to that of the vision encoder of the Siglip
-    [google/siglip-base-patch16-224](https://huggingface.co/google/siglip-base-patch16-224) architecture.
-
-    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
-    documentation from [`PretrainedConfig`] for more information.
-
-    Args:
-        hidden_size (`int`, *optional*, defaults to 768):
-            Dimensionality of the encoder layers and the pooler layer.
-        intermediate_size (`int`, *optional*, defaults to 3072):
-            Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
-        num_hidden_layers (`int`, *optional*, defaults to 12):
-            Number of hidden layers in the Transformer encoder.
-        num_attention_heads (`int`, *optional*, defaults to 12):
-            Number of attention heads for each attention layer in the Transformer encoder.
-        num_channels (`int`, *optional*, defaults to 3):
-            Number of channels in the input images.
-        image_size (`int`, *optional*, defaults to 224):
-            The size (resolution) of each image.
-        patch_size (`int`, *optional*, defaults to 16):
-            The size (resolution) of each patch.
-        hidden_act (`str` or `function`, *optional*, defaults to `"gelu_pytorch_tanh"`):
-            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
-            `"relu"`, `"selu"` and `"gelu_new"` `"quick_gelu"` are supported.
-        layer_norm_eps (`float`, *optional*, defaults to 1e-06):
-            The epsilon used by the layer normalization layers.
-        attention_dropout (`float`, *optional*, defaults to 0.0):
-            The dropout ratio for the attention probabilities.
-
-    Example:
-
-    ```python
-    >>> from transformers import SiglipVisionConfig, SiglipVisionModel
-
-    >>> # Initializing a SiglipVisionConfig with google/siglip-base-patch16-224 style configuration
-    >>> configuration = SiglipVisionConfig()
-
-    >>> # Initializing a SiglipVisionModel (with random weights) from the google/siglip-base-patch16-224 style configuration
-    >>> model = SiglipVisionModel(configuration)
-
-    >>> # Accessing the model configuration
-    >>> configuration = model.config
-    ```"""
-
-    model_type = "siglip_vision_model"
-
-    def __init__(
-        self,
-        hidden_size=768,
-        intermediate_size=3072,
-        num_hidden_layers=12,
-        num_attention_heads=12,
-        num_channels=3,
-        image_size=224,
-        patch_size=16,
-        hidden_act="gelu_pytorch_tanh",
-        layer_norm_eps=1e-6,
-        attention_dropout=0.0,
-        rope=True,
-        **kwargs,
-    ):
-        super().__init__(
-            hidden_size=hidden_size,
-            intermediate_size=intermediate_size,
-            num_hidden_layers=num_hidden_layers,
-            num_attention_heads=num_attention_heads,
-            num_channels=num_channels,
-            image_size=image_size,
-            patch_size=patch_size,
-            hidden_act=hidden_act,
-            layer_norm_eps=layer_norm_eps,
-            attention_dropout=attention_dropout,
-            **kwargs)
-        
-        self.rope = rope
-
-
-class RotaryEmbedding2D(torch.nn.Module):
-    def __init__(self, dim, max_h, max_w, base=10000):
-        super().__init__()
-        freq = torch.arange(0, dim, 2, dtype=torch.int64).float() / dim
-        inv_freq = 1.0 / (base ** freq)
-
-        grid_h = torch.arange(0, max_h)
-        grid_h = grid_h.to(inv_freq.dtype)
-        grid_h = grid_h[:, None].repeat(1, max_w)
-
-        grid_w = torch.arange(0, max_w)
-        grid_w = grid_w.to(inv_freq.dtype)
-        grid_w = grid_w[None, :].repeat(max_h, 1)
-
-        cos_h, sin_h = self._forward_one_side(grid_h, inv_freq)
-        cos_w, sin_w = self._forward_one_side(grid_w, inv_freq)
-
-        self.register_buffer("cos_h", cos_h)
-        self.register_buffer("sin_h", sin_h)
-        self.register_buffer("cos_w", cos_w)
-        self.register_buffer("sin_w", sin_w)
-
-    def _forward_one_side(self, grid, inv_freq):
-        freqs = grid[..., None] * inv_freq[None, None, :]
-        emb = torch.cat((freqs, freqs), dim=-1).flatten(0, 1)
-        return emb.cos(), emb.sin()
-
-
-def rotate_half(x):
-    x1 = x[..., : x.shape[-1] // 2]
-    x2 = x[..., x.shape[-1] // 2 :]
-    return torch.cat((-x2, x1), dim=-1)
-
-
-def apply_rotary_pos_emb(q, k, cos, sin):
-    # unsqueeze due to the head dimension
-    cos = cos.unsqueeze(1)
-    sin = sin.unsqueeze(1)
-    q_embed = (q * cos) + (rotate_half(q) * sin)
-    k_embed = (k * cos) + (rotate_half(k) * sin)
-    return q_embed, k_embed
-
-
-class SiglipVisionEmbeddings(nn.Module):
-    def __init__(self, config: SiglipVisionConfig):
-        super().__init__()
-        self.config = config
-        self.embed_dim = config.hidden_size
-        self.image_size = config.image_size
-        self.patch_size = config.patch_size
-
-        self.patch_embedding = nn.Conv2d(
-            in_channels=config.num_channels,
-            out_channels=self.embed_dim,
-            kernel_size=self.patch_size,
-            stride=self.patch_size,
-            padding="valid",
-        )
-
-        self.num_patches_per_side = self.image_size // self.patch_size
-        self.num_patches = self.num_patches_per_side**2
-        self.num_positions = self.num_patches
-        if not config.rope:
-            self.position_embedding = nn.Embedding(self.num_positions, self.embed_dim)
-
-    def convert_conv2d_to_linear(self, config, meta=False):
-        if meta:
-            linear_patch_embedding = nn.Linear(
-                config.num_channels * self.patch_size ** 2, self.embed_dim, bias=True, device='meta'
-            )
-        else:
-            linear_patch_embedding = nn.Linear(
-                config.num_channels * self.patch_size ** 2, self.embed_dim, bias=True
-            )
-        W = self.patch_embedding.weight.permute(0, 2, 3, 1).reshape(
-            self.embed_dim, config.num_channels * self.patch_size ** 2
-        )
-        linear_patch_embedding.weight.data = W
-        linear_patch_embedding.bias.data = self.patch_embedding.bias.data
-        del self.patch_embedding
-        self.patch_embedding = linear_patch_embedding
-
-    def forward(
-        self, 
-        packed_pixel_values: torch.FloatTensor, 
-        packed_flattened_position_ids: torch.LongTensor
-    ) -> torch.Tensor:
-
-        patch_embeds = self.patch_embedding(packed_pixel_values)
-        if not self.config.rope:
-            embeddings = patch_embeds + self.position_embedding(packed_flattened_position_ids)
-        else:
-            embeddings = patch_embeds
-        return embeddings
-
-
-class SiglipFlashAttention2(SiglipAttention):
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-
-    def forward(
-        self,
-        hidden_states: torch.Tensor,
-        cu_seqlens: torch.IntTensor,
-        max_seqlen: int,
-        cos_h: torch.Tensor = None,
-        sin_h: torch.Tensor = None,
-        cos_w: torch.Tensor = None,
-        sin_w: torch.Tensor = None,
-        **kwargs,
-    ) -> torch.Tensor:
-
-        total_q_len, _ = hidden_states.size()
-
-        query_states = self.q_proj(hidden_states)
-        key_states = self.k_proj(hidden_states)
-        value_states = self.v_proj(hidden_states)
-
-        query_states = query_states.view(total_q_len, self.num_heads, self.head_dim)
-        key_states = key_states.view(total_q_len, self.num_heads, self.head_dim)
-        value_states = value_states.view(total_q_len, self.num_heads, self.head_dim)
-
-        if self.config.rope:
-            qh, qw = query_states[:, :, :self.head_dim // 2], query_states[:, :, self.head_dim // 2:] 
-            kh, kw = key_states[:, :, :self.head_dim // 2], key_states[:, :, self.head_dim // 2:]
-            qh, kh = apply_rotary_pos_emb(qh, kh, cos_h, sin_h)
-            qw, kw = apply_rotary_pos_emb(qw, kw, cos_w, sin_w)
-            query_states = torch.cat([qh, qw], dim=-1)
-            key_states = torch.cat([kh, kw], dim=-1)
-
-        attn_output = flash_attn_varlen_func(
-            query_states.to(torch.bfloat16),
-            key_states.to(torch.bfloat16),
-            value_states.to(torch.bfloat16),
-            cu_seqlens_q=cu_seqlens,
-            cu_seqlens_k=cu_seqlens,
-            max_seqlen_q=max_seqlen,
-            max_seqlen_k=max_seqlen,
-            causal=False,
-        )
-
-        attn_output = self.out_proj(attn_output.reshape(total_q_len, -1))
-        return attn_output
-
-
-class SiglipMLP(nn.Module):
-    def __init__(self, config):
-        super().__init__()
-        self.config = config
-        self.activation_fn = ACT2FN[config.hidden_act]
-        self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size)
-        self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size)
-
-    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
-        hidden_states = self.fc1(hidden_states)
-        hidden_states = self.activation_fn(hidden_states)
-        hidden_states = self.fc2(hidden_states)
-        return hidden_states
-
-
-class SiglipEncoderLayer(nn.Module):
-    def __init__(self, config: SiglipVisionConfig):
-        super().__init__()
-        self.embed_dim = config.hidden_size
-        self.self_attn = SiglipFlashAttention2(config)
-        self.layer_norm1 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
-        self.mlp = SiglipMLP(config)
-        self.layer_norm2 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
-
-    def forward(
-        self,
-        hidden_states: torch.Tensor,
-        cu_seqlens: torch.IntTensor,
-        max_seqlen: int,
-        cos_h: torch.Tensor = None,
-        sin_h: torch.Tensor = None,
-        cos_w: torch.Tensor = None,
-        sin_w: torch.Tensor = None
-    ) -> torch.Tensor:
-        residual = hidden_states
-
-        hidden_states = self.layer_norm1(hidden_states)
-        hidden_states = self.self_attn(
-            hidden_states=hidden_states,
-            cu_seqlens=cu_seqlens,
-            max_seqlen=max_seqlen,
-            cos_h=cos_h,
-            sin_h=sin_h,
-            cos_w=cos_w,
-            sin_w=sin_w
-        )
-        hidden_states = residual + hidden_states
-
-        residual = hidden_states
-        hidden_states = self.layer_norm2(hidden_states)
-        hidden_states = self.mlp(hidden_states)
-        hidden_states = residual + hidden_states
-
-        return hidden_states
-
-
-class SiglipEncoder(nn.Module):
-    def __init__(self, config: SiglipVisionConfig):
-        super().__init__()
-        self.config = config
-        self.layers = nn.ModuleList(
-            [SiglipEncoderLayer(config) for _ in range(config.num_hidden_layers)]
-        )
-
-    def forward(
-        self,
-        inputs_embeds: torch.Tensor,
-        cu_seqlens: torch.IntTensor,
-        max_seqlen: int,
-        cos_h: torch.Tensor = None,
-        sin_h: torch.Tensor = None,
-        cos_w: torch.Tensor = None,
-        sin_w: torch.Tensor = None,
-    ) -> torch.Tensor:
-
-        hidden_states = inputs_embeds
-        for encoder_layer in self.layers:
-            hidden_states = encoder_layer(hidden_states, cu_seqlens, max_seqlen,
-                                          cos_h=cos_h, sin_h=sin_h, cos_w=cos_w, sin_w=sin_w)
-
-        return hidden_states
-
-
-class SiglipVisionTransformer(nn.Module):
-    def __init__(self, config: SiglipVisionConfig):
-        super().__init__()
-        self.config = config
-        embed_dim = config.hidden_size
-
-        self.embeddings = SiglipVisionEmbeddings(config)
-        if config.rope:
-            max_size = config.image_size // config.patch_size
-            dim_head = config.hidden_size // config.num_attention_heads
-            self.rope = RotaryEmbedding2D(dim_head // 2, max_size, max_size)
-
-        self.encoder = SiglipEncoder(config)
-        self.post_layernorm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)
-
-    def forward(
-        self,
-        packed_pixel_values: torch.Tensor,
-        packed_flattened_position_ids: torch.LongTensor,
-        cu_seqlens: torch.IntTensor,
-        max_seqlen: int,
-    ) -> torch.Tensor:
-        hidden_states = self.embeddings(
-            packed_pixel_values=packed_pixel_values, 
-            packed_flattened_position_ids=packed_flattened_position_ids
-        )
-
-        extra_inputs = {}
-        if self.config.rope:
-            extra_inputs.update(
-                cos_h = self.rope.cos_h[packed_flattened_position_ids],
-                sin_h = self.rope.sin_h[packed_flattened_position_ids],
-                cos_w = self.rope.cos_w[packed_flattened_position_ids],
-                sin_w = self.rope.sin_w[packed_flattened_position_ids]
-            )
-
-        last_hidden_state = self.encoder(
-            inputs_embeds=hidden_states, cu_seqlens=cu_seqlens, max_seqlen=max_seqlen, 
-            **extra_inputs
-        )
-        last_hidden_state = self.post_layernorm(last_hidden_state)
-        return last_hidden_state
-
-
-class SiglipVisionModel(SiglipPreTrainedModel):
-    config_class = SiglipVisionConfig
-    main_input_name = "packed_pixel_values"
-
-    def __init__(self, config: SiglipVisionConfig):
-        super().__init__(config)
-
-        self.vision_model = SiglipVisionTransformer(config)
-
-        # Initialize weights and apply final processing
-        self.post_init()
-
-    def get_input_embeddings(self) -> nn.Module:
-        return self.vision_model.embeddings.patch_embedding
-
-    def forward(
-        self,
-        packed_pixel_values: torch.Tensor,
-        packed_flattened_position_ids: torch.LongTensor,
-        cu_seqlens: torch.IntTensor,
-        max_seqlen: int,
-    ) -> torch.Tensor:
-
-        return self.vision_model(
-            packed_pixel_values=packed_pixel_values,
-            packed_flattened_position_ids=packed_flattened_position_ids,
-            cu_seqlens=cu_seqlens,
-            max_seqlen=max_seqlen,
-        )
+# Copyright (c) 2024 The HuggingFace Inc. team.
+# Copyright (c) 2025 Bytedance Ltd. and/or its affiliates.
+# SPDX-License-Identifier: Apache-2.0
+#
+# This file has been modified by ByteDance Ltd. and/or its affiliates. on 2025-05-20.
+#
+# Original file was released under Apache-2.0, with the full license text
+# available at https://github.com/huggingface/transformers/blob/main/LICENSE.
+#
+# This modified file is released under the same license.
+
+import torch
+from torch import nn
+
+from transformers.activations import ACT2FN
+from modeling.siglip.configuration_siglip import SiglipVisionConfig as _SiglipVisionConfig
+from modeling.siglip.modeling_siglip import SiglipAttention, SiglipPreTrainedModel
+from flash_attn import flash_attn_varlen_func
+
+
+class SiglipVisionConfig(_SiglipVisionConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`SiglipVisionModel`]. It is used to instantiate a
+    Siglip vision encoder according to the specified arguments, defining the model architecture. Instantiating a
+    configuration with the defaults will yield a similar configuration to that of the vision encoder of the Siglip
+    [google/siglip-base-patch16-224](https://huggingface.co/google/siglip-base-patch16-224) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        hidden_size (`int`, *optional*, defaults to 768):
+            Dimensionality of the encoder layers and the pooler layer.
+        intermediate_size (`int`, *optional*, defaults to 3072):
+            Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
+        num_hidden_layers (`int`, *optional*, defaults to 12):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 12):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        num_channels (`int`, *optional*, defaults to 3):
+            Number of channels in the input images.
+        image_size (`int`, *optional*, defaults to 224):
+            The size (resolution) of each image.
+        patch_size (`int`, *optional*, defaults to 16):
+            The size (resolution) of each patch.
+        hidden_act (`str` or `function`, *optional*, defaults to `"gelu_pytorch_tanh"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
+            `"relu"`, `"selu"` and `"gelu_new"` `"quick_gelu"` are supported.
+        layer_norm_eps (`float`, *optional*, defaults to 1e-06):
+            The epsilon used by the layer normalization layers.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+
+    Example:
+
+    ```python
+    >>> from transformers import SiglipVisionConfig, SiglipVisionModel
+
+    >>> # Initializing a SiglipVisionConfig with google/siglip-base-patch16-224 style configuration
+    >>> configuration = SiglipVisionConfig()
+
+    >>> # Initializing a SiglipVisionModel (with random weights) from the google/siglip-base-patch16-224 style configuration
+    >>> model = SiglipVisionModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "siglip_vision_model"
+
+    def __init__(
+        self,
+        hidden_size=768,
+        intermediate_size=3072,
+        num_hidden_layers=12,
+        num_attention_heads=12,
+        num_channels=3,
+        image_size=224,
+        patch_size=16,
+        hidden_act="gelu_pytorch_tanh",
+        layer_norm_eps=1e-6,
+        attention_dropout=0.0,
+        rope=True,
+        **kwargs,
+    ):
+        super().__init__(
+            hidden_size=hidden_size,
+            intermediate_size=intermediate_size,
+            num_hidden_layers=num_hidden_layers,
+            num_attention_heads=num_attention_heads,
+            num_channels=num_channels,
+            image_size=image_size,
+            patch_size=patch_size,
+            hidden_act=hidden_act,
+            layer_norm_eps=layer_norm_eps,
+            attention_dropout=attention_dropout,
+            **kwargs)
+        
+        self.rope = rope
+
+
+class RotaryEmbedding2D(torch.nn.Module):
+    def __init__(self, dim, max_h, max_w, base=10000):
+        super().__init__()
+        freq = torch.arange(0, dim, 2, dtype=torch.int64).float() / dim
+        inv_freq = 1.0 / (base ** freq)
+
+        grid_h = torch.arange(0, max_h)
+        grid_h = grid_h.to(inv_freq.dtype)
+        grid_h = grid_h[:, None].repeat(1, max_w)
+
+        grid_w = torch.arange(0, max_w)
+        grid_w = grid_w.to(inv_freq.dtype)
+        grid_w = grid_w[None, :].repeat(max_h, 1)
+
+        cos_h, sin_h = self._forward_one_side(grid_h, inv_freq)
+        cos_w, sin_w = self._forward_one_side(grid_w, inv_freq)
+
+        self.register_buffer("cos_h", cos_h)
+        self.register_buffer("sin_h", sin_h)
+        self.register_buffer("cos_w", cos_w)
+        self.register_buffer("sin_w", sin_w)
+
+    def _forward_one_side(self, grid, inv_freq):
+        freqs = grid[..., None] * inv_freq[None, None, :]
+        emb = torch.cat((freqs, freqs), dim=-1).flatten(0, 1)
+        return emb.cos(), emb.sin()
+
+
+def rotate_half(x):
+    x1 = x[..., : x.shape[-1] // 2]
+    x2 = x[..., x.shape[-1] // 2 :]
+    return torch.cat((-x2, x1), dim=-1)
+
+
+def apply_rotary_pos_emb(q, k, cos, sin):
+    # unsqueeze due to the head dimension
+    cos = cos.unsqueeze(1)
+    sin = sin.unsqueeze(1)
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return q_embed, k_embed
+
+
+class SiglipVisionEmbeddings(nn.Module):
+    def __init__(self, config: SiglipVisionConfig):
+        super().__init__()
+        self.config = config
+        self.embed_dim = config.hidden_size
+        self.image_size = config.image_size
+        self.patch_size = config.patch_size
+
+        self.patch_embedding = nn.Conv2d(
+            in_channels=config.num_channels,
+            out_channels=self.embed_dim,
+            kernel_size=self.patch_size,
+            stride=self.patch_size,
+            padding="valid",
+        )
+
+        self.num_patches_per_side = self.image_size // self.patch_size
+        self.num_patches = self.num_patches_per_side**2
+        self.num_positions = self.num_patches
+        if not config.rope:
+            self.position_embedding = nn.Embedding(self.num_positions, self.embed_dim)
+
+    def convert_conv2d_to_linear(self, config, meta=False):
+        if meta:
+            linear_patch_embedding = nn.Linear(
+                config.num_channels * self.patch_size ** 2, self.embed_dim, bias=True, device='meta'
+            )
+        else:
+            linear_patch_embedding = nn.Linear(
+                config.num_channels * self.patch_size ** 2, self.embed_dim, bias=True
+            )
+        W = self.patch_embedding.weight.permute(0, 2, 3, 1).reshape(
+            self.embed_dim, config.num_channels * self.patch_size ** 2
+        )
+        linear_patch_embedding.weight.data = W
+        linear_patch_embedding.bias.data = self.patch_embedding.bias.data
+        del self.patch_embedding
+        self.patch_embedding = linear_patch_embedding
+
+    def forward(
+        self, 
+        packed_pixel_values: torch.FloatTensor, 
+        packed_flattened_position_ids: torch.LongTensor
+    ) -> torch.Tensor:
+
+        patch_embeds = self.patch_embedding(packed_pixel_values)
+        if not self.config.rope:
+            embeddings = patch_embeds + self.position_embedding(packed_flattened_position_ids)
+        else:
+            embeddings = patch_embeds
+        return embeddings
+
+
+class SiglipFlashAttention2(SiglipAttention):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        cu_seqlens: torch.IntTensor,
+        max_seqlen: int,
+        cos_h: torch.Tensor = None,
+        sin_h: torch.Tensor = None,
+        cos_w: torch.Tensor = None,
+        sin_w: torch.Tensor = None,
+        **kwargs,
+    ) -> torch.Tensor:
+
+        total_q_len, _ = hidden_states.size()
+
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+        query_states = query_states.view(total_q_len, self.num_heads, self.head_dim)
+        key_states = key_states.view(total_q_len, self.num_heads, self.head_dim)
+        value_states = value_states.view(total_q_len, self.num_heads, self.head_dim)
+
+        if self.config.rope:
+            qh, qw = query_states[:, :, :self.head_dim // 2], query_states[:, :, self.head_dim // 2:] 
+            kh, kw = key_states[:, :, :self.head_dim // 2], key_states[:, :, self.head_dim // 2:]
+            qh, kh = apply_rotary_pos_emb(qh, kh, cos_h, sin_h)
+            qw, kw = apply_rotary_pos_emb(qw, kw, cos_w, sin_w)
+            query_states = torch.cat([qh, qw], dim=-1)
+            key_states = torch.cat([kh, kw], dim=-1)
+
+        attn_output = flash_attn_varlen_func(
+            query_states.to(torch.bfloat16),
+            key_states.to(torch.bfloat16),
+            value_states.to(torch.bfloat16),
+            cu_seqlens_q=cu_seqlens,
+            cu_seqlens_k=cu_seqlens,
+            max_seqlen_q=max_seqlen,
+            max_seqlen_k=max_seqlen,
+            causal=False,
+        )
+
+        attn_output = self.out_proj(attn_output.reshape(total_q_len, -1))
+        return attn_output
+
+
+class SiglipMLP(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.activation_fn = ACT2FN[config.hidden_act]
+        self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size)
+        self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.fc1(hidden_states)
+        hidden_states = self.activation_fn(hidden_states)
+        hidden_states = self.fc2(hidden_states)
+        return hidden_states
+
+
+class SiglipEncoderLayer(nn.Module):
+    def __init__(self, config: SiglipVisionConfig):
+        super().__init__()
+        self.embed_dim = config.hidden_size
+        self.self_attn = SiglipFlashAttention2(config)
+        self.layer_norm1 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
+        self.mlp = SiglipMLP(config)
+        self.layer_norm2 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        cu_seqlens: torch.IntTensor,
+        max_seqlen: int,
+        cos_h: torch.Tensor = None,
+        sin_h: torch.Tensor = None,
+        cos_w: torch.Tensor = None,
+        sin_w: torch.Tensor = None
+    ) -> torch.Tensor:
+        residual = hidden_states
+
+        hidden_states = self.layer_norm1(hidden_states)
+        hidden_states = self.self_attn(
+            hidden_states=hidden_states,
+            cu_seqlens=cu_seqlens,
+            max_seqlen=max_seqlen,
+            cos_h=cos_h,
+            sin_h=sin_h,
+            cos_w=cos_w,
+            sin_w=sin_w
+        )
+        hidden_states = residual + hidden_states
+
+        residual = hidden_states
+        hidden_states = self.layer_norm2(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+
+        return hidden_states
+
+
+class SiglipEncoder(nn.Module):
+    def __init__(self, config: SiglipVisionConfig):
+        super().__init__()
+        self.config = config
+        self.layers = nn.ModuleList(
+            [SiglipEncoderLayer(config) for _ in range(config.num_hidden_layers)]
+        )
+
+    def forward(
+        self,
+        inputs_embeds: torch.Tensor,
+        cu_seqlens: torch.IntTensor,
+        max_seqlen: int,
+        cos_h: torch.Tensor = None,
+        sin_h: torch.Tensor = None,
+        cos_w: torch.Tensor = None,
+        sin_w: torch.Tensor = None,
+    ) -> torch.Tensor:
+
+        hidden_states = inputs_embeds
+        for encoder_layer in self.layers:
+            hidden_states = encoder_layer(hidden_states, cu_seqlens, max_seqlen,
+                                          cos_h=cos_h, sin_h=sin_h, cos_w=cos_w, sin_w=sin_w)
+
+        return hidden_states
+
+
+class SiglipVisionTransformer(nn.Module):
+    def __init__(self, config: SiglipVisionConfig):
+        super().__init__()
+        self.config = config
+        embed_dim = config.hidden_size
+
+        self.embeddings = SiglipVisionEmbeddings(config)
+        if config.rope:
+            max_size = config.image_size // config.patch_size
+            dim_head = config.hidden_size // config.num_attention_heads
+            self.rope = RotaryEmbedding2D(dim_head // 2, max_size, max_size)
+
+        self.encoder = SiglipEncoder(config)
+        self.post_layernorm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)
+
+    def forward(
+        self,
+        packed_pixel_values: torch.Tensor,
+        packed_flattened_position_ids: torch.LongTensor,
+        cu_seqlens: torch.IntTensor,
+        max_seqlen: int,
+    ) -> torch.Tensor:
+        hidden_states = self.embeddings(
+            packed_pixel_values=packed_pixel_values, 
+            packed_flattened_position_ids=packed_flattened_position_ids
+        )
+
+        extra_inputs = {}
+        if self.config.rope:
+            extra_inputs.update(
+                cos_h = self.rope.cos_h[packed_flattened_position_ids],
+                sin_h = self.rope.sin_h[packed_flattened_position_ids],
+                cos_w = self.rope.cos_w[packed_flattened_position_ids],
+                sin_w = self.rope.sin_w[packed_flattened_position_ids]
+            )
+
+        last_hidden_state = self.encoder(
+            inputs_embeds=hidden_states, cu_seqlens=cu_seqlens, max_seqlen=max_seqlen, 
+            **extra_inputs
+        )
+        last_hidden_state = self.post_layernorm(last_hidden_state)
+        return last_hidden_state
+
+
+class SiglipVisionModel(SiglipPreTrainedModel):
+    config_class = SiglipVisionConfig
+    main_input_name = "packed_pixel_values"
+
+    def __init__(self, config: SiglipVisionConfig):
+        super().__init__(config)
+
+        self.vision_model = SiglipVisionTransformer(config)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self) -> nn.Module:
+        return self.vision_model.embeddings.patch_embedding
+
+    def forward(
+        self,
+        packed_pixel_values: torch.Tensor,
+        packed_flattened_position_ids: torch.LongTensor,
+        cu_seqlens: torch.IntTensor,
+        max_seqlen: int,
+    ) -> torch.Tensor:
+
+        return self.vision_model(
+            packed_pixel_values=packed_pixel_values,
+            packed_flattened_position_ids=packed_flattened_position_ids,
+            cu_seqlens=cu_seqlens,
+            max_seqlen=max_seqlen,
+        )

modeling/qwen2/__init__.py

@@ -1,68 +1,68 @@
-# Copyright 2024 The Qwen Team and The HuggingFace Inc. team.
-# SPDX-License-Identifier: Apache-2.0
-
-from typing import TYPE_CHECKING
-
-from transformers.utils import (
-    OptionalDependencyNotAvailable,
-    _LazyModule,
-    is_tokenizers_available,
-    is_torch_available,
-)
-
-
-_import_structure = {
-    "configuration_qwen2": ["Qwen2Config"],
-    "tokenization_qwen2": ["Qwen2Tokenizer"],
-}
-
-try:
-    if not is_tokenizers_available():
-        raise OptionalDependencyNotAvailable()
-except OptionalDependencyNotAvailable:
-    pass
-else:
-    _import_structure["tokenization_qwen2_fast"] = ["Qwen2TokenizerFast"]
-
-try:
-    if not is_torch_available():
-        raise OptionalDependencyNotAvailable()
-except OptionalDependencyNotAvailable:
-    pass
-else:
-    _import_structure["modeling_qwen2"] = [
-        "Qwen2ForCausalLM",
-        "Qwen2Model",
-        "Qwen2PreTrainedModel",
-    ]
-
-
-if TYPE_CHECKING:
-    from .configuration_qwen2 import Qwen2Config
-    from .tokenization_qwen2 import Qwen2Tokenizer
-
-    try:
-        if not is_tokenizers_available():
-            raise OptionalDependencyNotAvailable()
-    except OptionalDependencyNotAvailable:
-        pass
-    else:
-        from .tokenization_qwen2_fast import Qwen2TokenizerFast
-
-    try:
-        if not is_torch_available():
-            raise OptionalDependencyNotAvailable()
-    except OptionalDependencyNotAvailable:
-        pass
-    else:
-        from .modeling_qwen2 import (
-            Qwen2ForCausalLM,
-            Qwen2Model,
-            Qwen2PreTrainedModel,
-        )
-
-
-else:
-    import sys
-
-    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)
+# Copyright 2024 The Qwen Team and The HuggingFace Inc. team.
+# SPDX-License-Identifier: Apache-2.0
+
+from typing import TYPE_CHECKING
+
+from transformers.utils import (
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    is_tokenizers_available,
+    is_torch_available,
+)
+
+
+_import_structure = {
+    "configuration_qwen2": ["Qwen2Config"],
+    "tokenization_qwen2": ["Qwen2Tokenizer"],
+}
+
+try:
+    if not is_tokenizers_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["tokenization_qwen2_fast"] = ["Qwen2TokenizerFast"]
+
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["modeling_qwen2"] = [
+        "Qwen2ForCausalLM",
+        "Qwen2Model",
+        "Qwen2PreTrainedModel",
+    ]
+
+
+if TYPE_CHECKING:
+    from .configuration_qwen2 import Qwen2Config
+    from .tokenization_qwen2 import Qwen2Tokenizer
+
+    try:
+        if not is_tokenizers_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .tokenization_qwen2_fast import Qwen2TokenizerFast
+
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .modeling_qwen2 import (
+            Qwen2ForCausalLM,
+            Qwen2Model,
+            Qwen2PreTrainedModel,
+        )
+
+
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)

modeling/qwen2/configuration_qwen2.py

@@ -1,179 +1,179 @@
-# Copyright 2024 The Qwen Team and The HuggingFace Inc. team.
-# SPDX-License-Identifier: Apache-2.0
-
-"""Qwen2 model configuration"""
-
-from transformers.configuration_utils import PretrainedConfig
-from transformers.modeling_rope_utils import rope_config_validation
-from transformers.utils import logging
-
-
-logger = logging.get_logger(__name__)
-
-
-class Qwen2Config(PretrainedConfig):
-    r"""
-    This is the configuration class to store the configuration of a [`Qwen2Model`]. It is used to instantiate a
-    Qwen2 model according to the specified arguments, defining the model architecture. Instantiating a configuration
-    with the defaults will yield a similar configuration to that of
-    Qwen2-7B-beta [Qwen/Qwen2-7B-beta](https://huggingface.co/Qwen/Qwen2-7B-beta).
-
-    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
-    documentation from [`PretrainedConfig`] for more information.
-
-
-    Args:
-        vocab_size (`int`, *optional*, defaults to 151936):
-            Vocabulary size of the Qwen2 model. Defines the number of different tokens that can be represented by the
-            `inputs_ids` passed when calling [`Qwen2Model`]
-        hidden_size (`int`, *optional*, defaults to 4096):
-            Dimension of the hidden representations.
-        intermediate_size (`int`, *optional*, defaults to 22016):
-            Dimension of the MLP representations.
-        num_hidden_layers (`int`, *optional*, defaults to 32):
-            Number of hidden layers in the Transformer encoder.
-        num_attention_heads (`int`, *optional*, defaults to 32):
-            Number of attention heads for each attention layer in the Transformer encoder.
-        num_key_value_heads (`int`, *optional*, defaults to 32):
-            This is the number of key_value heads that should be used to implement Grouped Query Attention. If
-            `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
-            `num_key_value_heads=1` the model will use Multi Query Attention (MQA) otherwise GQA is used. When
-            converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
-            by meanpooling all the original heads within that group. For more details checkout [this
-            paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to `32`.
-        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
-            The non-linear activation function (function or string) in the decoder.
-        max_position_embeddings (`int`, *optional*, defaults to 32768):
-            The maximum sequence length that this model might ever be used with.
-        initializer_range (`float`, *optional*, defaults to 0.02):
-            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
-        rms_norm_eps (`float`, *optional*, defaults to 1e-06):
-            The epsilon used by the rms normalization layers.
-        use_cache (`bool`, *optional*, defaults to `True`):
-            Whether or not the model should return the last key/values attentions (not used by all models). Only
-            relevant if `config.is_decoder=True`.
-        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
-            Whether the model's input and output word embeddings should be tied.
-        rope_theta (`float`, *optional*, defaults to 10000.0):
-            The base period of the RoPE embeddings.
-        rope_scaling (`Dict`, *optional*):
-            Dictionary containing the scaling configuration for the RoPE embeddings. NOTE: if you apply new rope type
-            and you expect the model to work on longer `max_position_embeddings`, we recommend you to update this value
-            accordingly.
-            Expected contents:
-                `rope_type` (`str`):
-                    The sub-variant of RoPE to use. Can be one of ['default', 'linear', 'dynamic', 'yarn', 'longrope',
-                    'llama3'], with 'default' being the original RoPE implementation.
-                `factor` (`float`, *optional*):
-                    Used with all rope types except 'default'. The scaling factor to apply to the RoPE embeddings. In
-                    most scaling types, a `factor` of x will enable the model to handle sequences of length x *
-                    original maximum pre-trained length.
-                `original_max_position_embeddings` (`int`, *optional*):
-                    Used with 'dynamic', 'longrope' and 'llama3'. The original max position embeddings used during
-                    pretraining.
-                `attention_factor` (`float`, *optional*):
-                    Used with 'yarn' and 'longrope'. The scaling factor to be applied on the attention
-                    computation. If unspecified, it defaults to value recommended by the implementation, using the
-                    `factor` field to infer the suggested value.
-                `beta_fast` (`float`, *optional*):
-                    Only used with 'yarn'. Parameter to set the boundary for extrapolation (only) in the linear
-                    ramp function. If unspecified, it defaults to 32.
-                `beta_slow` (`float`, *optional*):
-                    Only used with 'yarn'. Parameter to set the boundary for interpolation (only) in the linear
-                    ramp function. If unspecified, it defaults to 1.
-                `short_factor` (`List[float]`, *optional*):
-                    Only used with 'longrope'. The scaling factor to be applied to short contexts (<
-                    `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
-                    size divided by the number of attention heads divided by 2
-                `long_factor` (`List[float]`, *optional*):
-                    Only used with 'longrope'. The scaling factor to be applied to long contexts (<
-                    `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
-                    size divided by the number of attention heads divided by 2
-                `low_freq_factor` (`float`, *optional*):
-                    Only used with 'llama3'. Scaling factor applied to low frequency components of the RoPE
-                `high_freq_factor` (`float`, *optional*):
-                    Only used with 'llama3'. Scaling factor applied to high frequency components of the RoPE
-        use_sliding_window (`bool`, *optional*, defaults to `False`):
-            Whether to use sliding window attention.
-        sliding_window (`int`, *optional*, defaults to 4096):
-            Sliding window attention (SWA) window size. If not specified, will default to `4096`.
-        max_window_layers (`int`, *optional*, defaults to 28):
-            The number of layers that use SWA (Sliding Window Attention). The bottom layers use SWA while the top use full attention.
-        attention_dropout (`float`, *optional*, defaults to 0.0):
-            The dropout ratio for the attention probabilities.
-
-    ```python
-    >>> from transformers import Qwen2Model, Qwen2Config
-
-    >>> # Initializing a Qwen2 style configuration
-    >>> configuration = Qwen2Config()
-
-    >>> # Initializing a model from the Qwen2-7B style configuration
-    >>> model = Qwen2Model(configuration)
-
-    >>> # Accessing the model configuration
-    >>> configuration = model.config
-    ```"""
-
-    model_type = "qwen2"
-    keys_to_ignore_at_inference = ["past_key_values"]
-
-    def __init__(
-        self,
-        vocab_size=151936,
-        hidden_size=4096,
-        intermediate_size=22016,
-        num_hidden_layers=32,
-        num_attention_heads=32,
-        num_key_value_heads=32,
-        hidden_act="silu",
-        max_position_embeddings=32768,
-        initializer_range=0.02,
-        rms_norm_eps=1e-6,
-        use_cache=True,
-        tie_word_embeddings=False,
-        rope_theta=10000.0,
-        rope_scaling=None,
-        use_sliding_window=False,
-        sliding_window=4096,
-        max_window_layers=28,
-        attention_dropout=0.0,
-        is_causal=True,
-        _attn_implementation="flash_attention_2",
-        **kwargs,
-    ):
-        self.vocab_size = vocab_size
-        self.max_position_embeddings = max_position_embeddings
-        self.hidden_size = hidden_size
-        self.intermediate_size = intermediate_size
-        self.num_hidden_layers = num_hidden_layers
-        self.num_attention_heads = num_attention_heads
-        self.use_sliding_window = use_sliding_window
-        self.sliding_window = sliding_window if use_sliding_window else None
-        self.max_window_layers = max_window_layers
-
-        # for backward compatibility
-        if num_key_value_heads is None:
-            num_key_value_heads = num_attention_heads
-
-        self.num_key_value_heads = num_key_value_heads
-        self.hidden_act = hidden_act
-        self.initializer_range = initializer_range
-        self.rms_norm_eps = rms_norm_eps
-        self.use_cache = use_cache
-        self.rope_theta = rope_theta
-        self.rope_scaling = rope_scaling
-        self.attention_dropout = attention_dropout
-        self.is_causal = is_causal
-        self._attn_implementation = _attn_implementation
-
-        # Validate the correctness of rotary position embeddings parameters
-        # BC: if there is a 'type' field, move it to 'rope_type'.
-        if self.rope_scaling is not None and "type" in self.rope_scaling:
-            self.rope_scaling["rope_type"] = self.rope_scaling["type"]
-        rope_config_validation(self)
-
-        super().__init__(
-            tie_word_embeddings=tie_word_embeddings,
-            **kwargs,
-        )
+# Copyright 2024 The Qwen Team and The HuggingFace Inc. team.
+# SPDX-License-Identifier: Apache-2.0
+
+"""Qwen2 model configuration"""
+
+from transformers.configuration_utils import PretrainedConfig
+from transformers.modeling_rope_utils import rope_config_validation
+from transformers.utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+
+class Qwen2Config(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`Qwen2Model`]. It is used to instantiate a
+    Qwen2 model according to the specified arguments, defining the model architecture. Instantiating a configuration
+    with the defaults will yield a similar configuration to that of
+    Qwen2-7B-beta [Qwen/Qwen2-7B-beta](https://huggingface.co/Qwen/Qwen2-7B-beta).
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 151936):
+            Vocabulary size of the Qwen2 model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`Qwen2Model`]
+        hidden_size (`int`, *optional*, defaults to 4096):
+            Dimension of the hidden representations.
+        intermediate_size (`int`, *optional*, defaults to 22016):
+            Dimension of the MLP representations.
+        num_hidden_layers (`int`, *optional*, defaults to 32):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 32):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        num_key_value_heads (`int`, *optional*, defaults to 32):
+            This is the number of key_value heads that should be used to implement Grouped Query Attention. If
+            `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
+            `num_key_value_heads=1` the model will use Multi Query Attention (MQA) otherwise GQA is used. When
+            converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
+            by meanpooling all the original heads within that group. For more details checkout [this
+            paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to `32`.
+        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
+            The non-linear activation function (function or string) in the decoder.
+        max_position_embeddings (`int`, *optional*, defaults to 32768):
+            The maximum sequence length that this model might ever be used with.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        rms_norm_eps (`float`, *optional*, defaults to 1e-06):
+            The epsilon used by the rms normalization layers.
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values attentions (not used by all models). Only
+            relevant if `config.is_decoder=True`.
+        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether the model's input and output word embeddings should be tied.
+        rope_theta (`float`, *optional*, defaults to 10000.0):
+            The base period of the RoPE embeddings.
+        rope_scaling (`Dict`, *optional*):
+            Dictionary containing the scaling configuration for the RoPE embeddings. NOTE: if you apply new rope type
+            and you expect the model to work on longer `max_position_embeddings`, we recommend you to update this value
+            accordingly.
+            Expected contents:
+                `rope_type` (`str`):
+                    The sub-variant of RoPE to use. Can be one of ['default', 'linear', 'dynamic', 'yarn', 'longrope',
+                    'llama3'], with 'default' being the original RoPE implementation.
+                `factor` (`float`, *optional*):
+                    Used with all rope types except 'default'. The scaling factor to apply to the RoPE embeddings. In
+                    most scaling types, a `factor` of x will enable the model to handle sequences of length x *
+                    original maximum pre-trained length.
+                `original_max_position_embeddings` (`int`, *optional*):
+                    Used with 'dynamic', 'longrope' and 'llama3'. The original max position embeddings used during
+                    pretraining.
+                `attention_factor` (`float`, *optional*):
+                    Used with 'yarn' and 'longrope'. The scaling factor to be applied on the attention
+                    computation. If unspecified, it defaults to value recommended by the implementation, using the
+                    `factor` field to infer the suggested value.
+                `beta_fast` (`float`, *optional*):
+                    Only used with 'yarn'. Parameter to set the boundary for extrapolation (only) in the linear
+                    ramp function. If unspecified, it defaults to 32.
+                `beta_slow` (`float`, *optional*):
+                    Only used with 'yarn'. Parameter to set the boundary for interpolation (only) in the linear
+                    ramp function. If unspecified, it defaults to 1.
+                `short_factor` (`List[float]`, *optional*):
+                    Only used with 'longrope'. The scaling factor to be applied to short contexts (<
+                    `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
+                    size divided by the number of attention heads divided by 2
+                `long_factor` (`List[float]`, *optional*):
+                    Only used with 'longrope'. The scaling factor to be applied to long contexts (<
+                    `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
+                    size divided by the number of attention heads divided by 2
+                `low_freq_factor` (`float`, *optional*):
+                    Only used with 'llama3'. Scaling factor applied to low frequency components of the RoPE
+                `high_freq_factor` (`float`, *optional*):
+                    Only used with 'llama3'. Scaling factor applied to high frequency components of the RoPE
+        use_sliding_window (`bool`, *optional*, defaults to `False`):
+            Whether to use sliding window attention.
+        sliding_window (`int`, *optional*, defaults to 4096):
+            Sliding window attention (SWA) window size. If not specified, will default to `4096`.
+        max_window_layers (`int`, *optional*, defaults to 28):
+            The number of layers that use SWA (Sliding Window Attention). The bottom layers use SWA while the top use full attention.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+
+    ```python
+    >>> from transformers import Qwen2Model, Qwen2Config
+
+    >>> # Initializing a Qwen2 style configuration
+    >>> configuration = Qwen2Config()
+
+    >>> # Initializing a model from the Qwen2-7B style configuration
+    >>> model = Qwen2Model(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "qwen2"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    def __init__(
+        self,
+        vocab_size=151936,
+        hidden_size=4096,
+        intermediate_size=22016,
+        num_hidden_layers=32,
+        num_attention_heads=32,
+        num_key_value_heads=32,
+        hidden_act="silu",
+        max_position_embeddings=32768,
+        initializer_range=0.02,
+        rms_norm_eps=1e-6,
+        use_cache=True,
+        tie_word_embeddings=False,
+        rope_theta=10000.0,
+        rope_scaling=None,
+        use_sliding_window=False,
+        sliding_window=4096,
+        max_window_layers=28,
+        attention_dropout=0.0,
+        is_causal=True,
+        _attn_implementation="flash_attention_2",
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.use_sliding_window = use_sliding_window
+        self.sliding_window = sliding_window if use_sliding_window else None
+        self.max_window_layers = max_window_layers
+
+        # for backward compatibility
+        if num_key_value_heads is None:
+            num_key_value_heads = num_attention_heads
+
+        self.num_key_value_heads = num_key_value_heads
+        self.hidden_act = hidden_act
+        self.initializer_range = initializer_range
+        self.rms_norm_eps = rms_norm_eps
+        self.use_cache = use_cache
+        self.rope_theta = rope_theta
+        self.rope_scaling = rope_scaling
+        self.attention_dropout = attention_dropout
+        self.is_causal = is_causal
+        self._attn_implementation = _attn_implementation
+
+        # Validate the correctness of rotary position embeddings parameters
+        # BC: if there is a 'type' field, move it to 'rope_type'.
+        if self.rope_scaling is not None and "type" in self.rope_scaling:
+            self.rope_scaling["rope_type"] = self.rope_scaling["type"]
+        rope_config_validation(self)
+
+        super().__init__(
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )

modeling/qwen2/modeling_qwen2.py

@@ -1,929 +1,929 @@
-# Copyright 2024 The Qwen Team and The HuggingFace Inc. team.
-# SPDX-License-Identifier: Apache-2.0
-
-"""PyTorch Qwen2 model."""
-
-import math
-from typing import List, Optional, Tuple, Union
-
-import torch
-import torch.utils.checkpoint
-from torch import nn
-
-from transformers.activations import ACT2FN
-from transformers.cache_utils import Cache, DynamicCache
-from transformers.generation import GenerationMixin
-from transformers.modeling_outputs import (
-    BaseModelOutputWithPast,
-    CausalLMOutputWithPast,
-)
-from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS
-from transformers.modeling_utils import PreTrainedModel
-from transformers.utils import (
-    add_start_docstrings,
-    add_start_docstrings_to_model_forward,
-    is_flash_attn_2_available,
-    is_flash_attn_greater_or_equal_2_10,
-    logging,
-    replace_return_docstrings,
-)
-from .configuration_qwen2 import Qwen2Config
-
-
-if is_flash_attn_2_available():
-    from transformers.modeling_flash_attention_utils import _flash_attention_forward
-
-
-logger = logging.get_logger(__name__)
-
-
-_CHECKPOINT_FOR_DOC = "Qwen/Qwen2-7B"
-_CONFIG_FOR_DOC = "Qwen2Config"
-
-
-# Copied from transformers.models.llama.modeling_llama.LlamaRMSNorm with Llama->Qwen2
-class Qwen2RMSNorm(nn.Module):
-    def __init__(self, hidden_size, eps=1e-6):
-        """
-        Qwen2RMSNorm is equivalent to T5LayerNorm
-        """
-        super().__init__()
-        self.weight = nn.Parameter(torch.ones(hidden_size))
-        self.variance_epsilon = eps
-
-    def forward(self, hidden_states):
-        input_dtype = hidden_states.dtype
-        hidden_states = hidden_states.to(torch.float32)
-        variance = hidden_states.pow(2).mean(-1, keepdim=True)
-        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
-        return self.weight * hidden_states.to(input_dtype)
-
-    def extra_repr(self):
-        return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"
-
-
-# Copied from transformers.models.llama.modeling_llama.LlamaRotaryEmbedding with Llama->Qwen2
-class Qwen2RotaryEmbedding(nn.Module):
-    def __init__(
-        self,
-        dim=None,
-        max_position_embeddings=2048,
-        base=10000,
-        device=None,
-        scaling_factor=1.0,
-        rope_type="default",
-        config: Optional[Qwen2Config] = None,
-    ):
-        super().__init__()
-        # TODO (joao): remove the `if` below, only used for BC
-        self.rope_kwargs = {}
-        if config is None:
-            logger.warning_once(
-                "`Qwen2RotaryEmbedding` can now be fully parameterized by passing the model config through the "
-                "`config` argument. All other arguments will be removed in v4.46"
-            )
-            self.rope_kwargs = {
-                "rope_type": rope_type,
-                "factor": scaling_factor,
-                "dim": dim,
-                "base": base,
-                "max_position_embeddings": max_position_embeddings,
-            }
-            self.rope_type = rope_type
-            self.max_seq_len_cached = max_position_embeddings
-            self.original_max_seq_len = max_position_embeddings
-        else:
-            # BC: "rope_type" was originally "type"
-            if config.rope_scaling is not None:
-                self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))
-            else:
-                self.rope_type = "default"
-            self.max_seq_len_cached = config.max_position_embeddings
-            self.original_max_seq_len = config.max_position_embeddings
-
-        self.config = config
-        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
-
-        inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device, **self.rope_kwargs)
-        self.register_buffer("inv_freq", inv_freq, persistent=False)
-        self.original_inv_freq = self.inv_freq
-
-    def _dynamic_frequency_update(self, position_ids, device):
-        """
-        dynamic RoPE layers should recompute `inv_freq` in the following situations:
-        1 - growing beyond the cached sequence length (allow scaling)
-        2 - the current sequence length is in the original scale (avoid losing precision with small sequences)
-        """
-        seq_len = torch.max(position_ids) + 1
-        if seq_len > self.max_seq_len_cached:  # growth
-            inv_freq, self.attention_scaling = self.rope_init_fn(
-                self.config, device, seq_len=seq_len, **self.rope_kwargs
-            )
-            self.register_buffer("inv_freq", inv_freq, persistent=False)  # TODO joao: may break with compilation
-            self.max_seq_len_cached = seq_len
-
-        if seq_len < self.original_max_seq_len and self.max_seq_len_cached > self.original_max_seq_len:  # reset
-            self.register_buffer("inv_freq", self.original_inv_freq, persistent=False)
-            self.max_seq_len_cached = self.original_max_seq_len
-
-    @torch.no_grad()
-    def forward(self, x, position_ids):
-        if "dynamic" in self.rope_type:
-            self._dynamic_frequency_update(position_ids, device=x.device)
-
-        # Core RoPE block
-        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
-        position_ids_expanded = position_ids[:, None, :].float()
-        # Force float32 (see https://github.com/huggingface/transformers/pull/29285)
-        device_type = x.device.type
-        device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu"
-        with torch.autocast(device_type=device_type, enabled=False):
-            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
-            emb = torch.cat((freqs, freqs), dim=-1)
-            cos = emb.cos()
-            sin = emb.sin()
-
-        # Advanced RoPE types (e.g. yarn) apply a post-processing scaling factor, equivalent to scaling attention
-        cos = cos * self.attention_scaling
-        sin = sin * self.attention_scaling
-
-        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
-
-
-# Copied from transformers.models.llama.modeling_llama.rotate_half
-def rotate_half(x):
-    """Rotates half the hidden dims of the input."""
-    x1 = x[..., : x.shape[-1] // 2]
-    x2 = x[..., x.shape[-1] // 2 :]
-    return torch.cat((-x2, x1), dim=-1)
-
-
-# Copied from transformers.models.llama.modeling_llama.apply_rotary_pos_emb
-def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
-    """Applies Rotary Position Embedding to the query and key tensors.
-
-    Args:
-        q (`torch.Tensor`): The query tensor.
-        k (`torch.Tensor`): The key tensor.
-        cos (`torch.Tensor`): The cosine part of the rotary embedding.
-        sin (`torch.Tensor`): The sine part of the rotary embedding.
-        position_ids (`torch.Tensor`, *optional*):
-            Deprecated and unused.
-        unsqueeze_dim (`int`, *optional*, defaults to 1):
-            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
-            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
-            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
-            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
-            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
-            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
-    Returns:
-        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
-    """
-    cos = cos.unsqueeze(unsqueeze_dim)
-    sin = sin.unsqueeze(unsqueeze_dim)
-    q_embed = (q * cos) + (rotate_half(q) * sin)
-    k_embed = (k * cos) + (rotate_half(k) * sin)
-    return q_embed, k_embed
-
-
-# Copied from transformers.models.mistral.modeling_mistral.MistralMLP with Mistral->Qwen2
-class Qwen2MLP(nn.Module):
-    def __init__(self, config):
-        super().__init__()
-        self.hidden_size = config.hidden_size
-        self.intermediate_size = config.intermediate_size
-        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
-        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
-        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
-        self.act_fn = ACT2FN[config.hidden_act]
-
-    def forward(self, hidden_state):
-        return self.down_proj(self.act_fn(self.gate_proj(hidden_state)) * self.up_proj(hidden_state))
-
-
-# Copied from transformers.models.llama.modeling_llama.repeat_kv
-def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
-    """
-    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
-    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
-    """
-    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
-    if n_rep == 1:
-        return hidden_states
-    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
-    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
-
-
-class Qwen2Attention(nn.Module):
-    """
-    Multi-headed attention from 'Attention Is All You Need' paper. Modified to use sliding window attention: Longformer
-    and "Generating Long Sequences with Sparse Transformers".
-    """
-
-    def __init__(self, config: Qwen2Config, layer_idx: Optional[int] = None):
-        super().__init__()
-        self.config = config
-        self.layer_idx = layer_idx
-        if layer_idx is None:
-            logger.warning_once(
-                f"Instantiating {self.__class__.__name__} without passing `layer_idx` is not recommended and will "
-                "to errors during the forward call, if caching is used. Please make sure to provide a `layer_idx` "
-                "when creating this class."
-            )
-
-        self.hidden_size = config.hidden_size
-        self.num_heads = config.num_attention_heads
-        self.head_dim = self.hidden_size // self.num_heads
-        self.num_key_value_heads = config.num_key_value_heads
-        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
-        self.max_position_embeddings = config.max_position_embeddings
-        self.rope_theta = config.rope_theta
-        self.is_causal = config.is_causal
-        self.attention_dropout = config.attention_dropout
-
-        if (self.head_dim * self.num_heads) != self.hidden_size:
-            raise ValueError(
-                f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
-                f" and `num_heads`: {self.num_heads})."
-            )
-        self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=True)
-        self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=True)
-        self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=True)
-        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
-
-    def forward(
-        self,
-        hidden_states: torch.Tensor,
-        attention_mask: Optional[torch.Tensor] = None,
-        position_ids: Optional[torch.LongTensor] = None,
-        past_key_value: Optional[Cache] = None,
-        output_attentions: bool = False,
-        use_cache: bool = False,
-        cache_position: Optional[torch.LongTensor] = None,
-        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # will become mandatory in v4.46
-    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
-        bsz, q_len, _ = hidden_states.size()
-
-        query_states = self.q_proj(hidden_states)
-        key_states = self.k_proj(hidden_states)
-        value_states = self.v_proj(hidden_states)
-
-        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
-        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
-        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
-
-        if position_embeddings is None:
-            logger.warning_once(
-                "The attention layers in this model are transitioning from computing the RoPE embeddings internally "
-                "through `position_ids` (2D tensor with the indexes of the tokens), to using externally computed "
-                "`position_embeddings` (Tuple of tensors, containing cos and sin). In v4.46 `position_ids` will be "
-                "removed and `position_embeddings` will be mandatory."
-            )
-            cos, sin = self.rotary_emb(value_states, position_ids)
-        else:
-            cos, sin = position_embeddings
-        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
-
-        if past_key_value is not None:
-            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}  # Specific to RoPE models
-            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
-
-        # repeat k/v heads if n_kv_heads < n_heads
-        key_states = repeat_kv(key_states, self.num_key_value_groups)
-        value_states = repeat_kv(value_states, self.num_key_value_groups)
-
-        attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
-        if attention_mask is not None:  # no matter the length, we just slice it
-            causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
-            attn_weights = attn_weights + causal_mask
-
-        # upcast attention to fp32
-        attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
-        attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training)
-        attn_output = torch.matmul(attn_weights, value_states)
-
-        if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
-            raise ValueError(
-                f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
-                f" {attn_output.size()}"
-            )
-
-        attn_output = attn_output.transpose(1, 2).contiguous()
-        attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
-
-        attn_output = self.o_proj(attn_output)
-
-        if not output_attentions:
-            attn_weights = None
-
-        return attn_output, attn_weights, past_key_value
-
-
-class Qwen2FlashAttention2(Qwen2Attention):
-    """
-    Qwen2 flash attention module, following Qwen2 attention module. This module inherits from `Qwen2Attention`
-    as the weights of the module stays untouched. The only required change would be on the forward pass
-    where it needs to correctly call the public API of flash attention and deal with padding tokens
-    in case the input contains any of them. Additionally, for sliding window attention, we apply SWA only to the bottom
-    config.max_window_layers layers.
-    """
-
-    # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2.__init__
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-
-        # TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1.
-        # flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignement, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0.
-        # Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left).
-        self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()
-
-    def forward(
-        self,
-        hidden_states: torch.Tensor,
-        attention_mask: Optional[torch.Tensor] = None,
-        position_ids: Optional[torch.LongTensor] = None,
-        past_key_value: Optional[Cache] = None,
-        output_attentions: bool = False,
-        use_cache: bool = False,
-        cache_position: Optional[torch.LongTensor] = None,
-        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # will become mandatory in v4.46
-    ):
-        bsz, q_len, _ = hidden_states.size()
-
-        query_states = self.q_proj(hidden_states)
-        key_states = self.k_proj(hidden_states)
-        value_states = self.v_proj(hidden_states)
-
-        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
-        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
-        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
-
-        if position_embeddings is None:
-            logger.warning_once(
-                "The attention layers in this model are transitioning from computing the RoPE embeddings internally "
-                "through `position_ids` (2D tensor with the indexes of the tokens), to using externally computed "
-                "`position_embeddings` (Tuple of tensors, containing cos and sin). In v4.46 `position_ids` will be "
-                "removed and `position_embeddings` will be mandatory."
-            )
-            cos, sin = self.rotary_emb(value_states, position_ids)
-        else:
-            cos, sin = position_embeddings
-        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
-
-        if past_key_value is not None:
-            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}  # Specific to RoPE models
-            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
-
-        # repeat k/v heads if n_kv_heads < n_heads
-        key_states = repeat_kv(key_states, self.num_key_value_groups)
-        value_states = repeat_kv(value_states, self.num_key_value_groups)
-        dropout_rate = 0.0 if not self.training else self.attention_dropout
-
-        # In PEFT, usually we cast the layer norms in float32 for training stability reasons
-        # therefore the input hidden states gets silently casted in float32. Hence, we need
-        # cast them back in float16 just to be sure everything works as expected.
-        input_dtype = query_states.dtype
-        if input_dtype == torch.float32:
-            if torch.is_autocast_enabled():
-                target_dtype = torch.get_autocast_gpu_dtype()
-            # Handle the case where the model is quantized
-            elif hasattr(self.config, "_pre_quantization_dtype"):
-                target_dtype = self.config._pre_quantization_dtype
-            else:
-                target_dtype = self.q_proj.weight.dtype
-
-            logger.warning_once(
-                f"The input hidden states seems to be silently casted in float32, this might be related to"
-                f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in"
-                f" {target_dtype}."
-            )
-
-            query_states = query_states.to(target_dtype)
-            key_states = key_states.to(target_dtype)
-            value_states = value_states.to(target_dtype)
-
-        # Reashape to the expected shape for Flash Attention
-        query_states = query_states.transpose(1, 2)
-        key_states = key_states.transpose(1, 2)
-        value_states = value_states.transpose(1, 2)
-
-        if (
-            self.config.use_sliding_window
-            and getattr(self.config, "sliding_window", None) is not None
-            and self.layer_idx >= self.config.max_window_layers
-        ):
-            sliding_window = self.config.sliding_window
-        else:
-            sliding_window = None
-
-        attn_output = _flash_attention_forward(
-            query_states,
-            key_states,
-            value_states,
-            attention_mask,
-            q_len,
-            position_ids=position_ids,
-            dropout=dropout_rate,
-            sliding_window=sliding_window,
-            is_causal=self.is_causal,
-            use_top_left_mask=self._flash_attn_uses_top_left_mask,
-        )
-
-        attn_output = attn_output.reshape(bsz, q_len, self.hidden_size).contiguous()
-        attn_output = self.o_proj(attn_output)
-
-        if not output_attentions:
-            attn_weights = None
-
-        return attn_output, attn_weights, past_key_value
-
-
-QWEN2_ATTENTION_CLASSES = {
-    "eager": Qwen2Attention,
-    "flash_attention_2": Qwen2FlashAttention2,
-}
-
-
-class Qwen2DecoderLayer(nn.Module):
-    def __init__(self, config: Qwen2Config, layer_idx: int):
-        super().__init__()
-        self.hidden_size = config.hidden_size
-
-        if config.sliding_window and config._attn_implementation != "flash_attention_2":
-            logger.warning_once(
-                f"Sliding Window Attention is enabled but not implemented for `{config._attn_implementation}`; "
-                "unexpected results may be encountered."
-            )
-        self.self_attn = QWEN2_ATTENTION_CLASSES[config._attn_implementation](config, layer_idx)
-
-        self.mlp = Qwen2MLP(config)
-        self.input_layernorm = Qwen2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
-        self.post_attention_layernorm = Qwen2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
-
-    def forward(
-        self,
-        hidden_states: torch.Tensor,
-        attention_mask: Optional[torch.Tensor] = None,
-        position_ids: Optional[torch.LongTensor] = None,
-        past_key_value: Optional[Tuple[torch.Tensor]] = None,
-        output_attentions: Optional[bool] = False,
-        use_cache: Optional[bool] = False,
-        cache_position: Optional[torch.LongTensor] = None,
-        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # will become mandatory in v4.46
-        **kwargs,
-    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
-        """
-        Args:
-            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
-            attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
-                `(batch, sequence_length)` where padding elements are indicated by 0.
-            output_attentions (`bool`, *optional*):
-                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
-                returned tensors for more detail.
-            use_cache (`bool`, *optional*):
-                If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
-                (see `past_key_values`).
-            past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
-            cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
-                Indices depicting the position of the input sequence tokens in the sequence.
-            position_embeddings (`Tuple[torch.FloatTensor, torch.FloatTensor]`, *optional*):
-                Tuple containing the cosine and sine positional embeddings of shape `(batch_size, seq_len, head_dim)`,
-                with `head_dim` being the embedding dimension of each attention head.
-            kwargs (`dict`, *optional*):
-                Arbitrary kwargs to be ignored, used for FSDP and other methods that injects code
-                into the model
-        """
-
-        residual = hidden_states
-
-        hidden_states = self.input_layernorm(hidden_states)
-
-        # Self Attention
-        hidden_states, self_attn_weights, present_key_value = self.self_attn(
-            hidden_states=hidden_states,
-            attention_mask=attention_mask,
-            position_ids=position_ids,
-            past_key_value=past_key_value,
-            output_attentions=output_attentions,
-            use_cache=use_cache,
-            cache_position=cache_position,
-            position_embeddings=position_embeddings,
-        )
-        hidden_states = residual + hidden_states
-
-        # Fully Connected
-        residual = hidden_states
-        hidden_states = self.post_attention_layernorm(hidden_states)
-        hidden_states = self.mlp(hidden_states)
-        hidden_states = residual + hidden_states
-
-        outputs = (hidden_states,)
-
-        if output_attentions:
-            outputs += (self_attn_weights,)
-
-        if use_cache:
-            outputs += (present_key_value,)
-
-        return outputs
-
-
-QWEN2_START_DOCSTRING = r"""
-    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
-    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
-    etc.)
-
-    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
-    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
-    and behavior.
-
-    Parameters:
-        config ([`Qwen2Config`]):
-            Model configuration class with all the parameters of the model. Initializing with a config file does not
-            load the weights associated with the model, only the configuration. Check out the
-            [`~PreTrainedModel.from_pretrained`] method to load the model weights.
-"""
-
-
-@add_start_docstrings(
-    "The bare Qwen2 Model outputting raw hidden-states without any specific head on top.",
-    QWEN2_START_DOCSTRING,
-)
-class Qwen2PreTrainedModel(PreTrainedModel):
-    config_class = Qwen2Config
-    base_model_prefix = "model"
-    supports_gradient_checkpointing = True
-    _no_split_modules = ["Qwen2DecoderLayer"]
-    _skip_keys_device_placement = "past_key_values"
-    _supports_flash_attn_2 = True
-    _supports_cache_class = True
-    _supports_quantized_cache = True
-    _supports_static_cache = True
-
-    def _init_weights(self, module):
-        std = self.config.initializer_range
-        if isinstance(module, nn.Linear):
-            module.weight.data.normal_(mean=0.0, std=std)
-            if module.bias is not None:
-                module.bias.data.zero_()
-        elif isinstance(module, nn.Embedding):
-            module.weight.data.normal_(mean=0.0, std=std)
-            if module.padding_idx is not None:
-                module.weight.data[module.padding_idx].zero_()
-
-
-QWEN2_INPUTS_DOCSTRING = r"""
-    Args:
-        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
-            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
-            it.
-
-            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
-            [`PreTrainedTokenizer.__call__`] for details.
-
-            [What are input IDs?](../glossary#input-ids)
-        attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
-            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
-
-            - 1 for tokens that are **not masked**,
-            - 0 for tokens that are **masked**.
-
-            [What are attention masks?](../glossary#attention-mask)
-
-            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
-            [`PreTrainedTokenizer.__call__`] for details.
-
-            If `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see
-            `past_key_values`).
-
-            If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
-            and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
-            information on the default strategy.
-
-            - 1 indicates the head is **not masked**,
-            - 0 indicates the head is **masked**.
-        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
-            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
-            config.n_positions - 1]`.
-
-            [What are position IDs?](../glossary#position-ids)
-        past_key_values (`Cache` or `tuple(tuple(torch.FloatTensor))`, *optional*):
-            Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
-            blocks) that can be used to speed up sequential decoding. This typically consists in the `past_key_values`
-            returned by the model at a previous stage of decoding, when `use_cache=True` or `config.use_cache=True`.
-
-            Two formats are allowed:
-            - a [`~cache_utils.Cache`] instance, see our
-            [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache);
-            - Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of
-            shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`). This is also known as the legacy
-            cache format.
-
-            The model will output the same cache format that is fed as input. If no `past_key_values` are passed, the
-            legacy cache format will be returned.
-
-            If `past_key_values` are used, the user can optionally input only the last `input_ids` (those that don't
-            have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `input_ids`
-            of shape `(batch_size, sequence_length)`.
-        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
-            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
-            is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
-            model's internal embedding lookup matrix.
-        use_cache (`bool`, *optional*):
-            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
-            `past_key_values`).
-        output_attentions (`bool`, *optional*):
-            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
-            tensors for more detail.
-        output_hidden_states (`bool`, *optional*):
-            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
-            more detail.
-        return_dict (`bool`, *optional*):
-            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
-        cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
-            Indices depicting the position of the input sequence tokens in the sequence. Contrarily to `position_ids`,
-            this tensor is not affected by padding. It is used to update the cache in the correct position and to infer
-            the complete sequence length.
-"""
-
-
-@add_start_docstrings(
-    "The bare Qwen2 Model outputting raw hidden-states without any specific head on top.",
-    QWEN2_START_DOCSTRING,
-)
-class Qwen2Model(Qwen2PreTrainedModel):
-    """
-    Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`Qwen2DecoderLayer`]
-
-    Args:
-        config: Qwen2Config
-    """
-
-    def __init__(self, config: Qwen2Config):
-        super().__init__(config)
-        self.padding_idx = config.pad_token_id
-        self.vocab_size = config.vocab_size
-
-        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
-        self.layers = nn.ModuleList(
-            [Qwen2DecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
-        )
-        self._attn_implementation = config._attn_implementation
-        self.norm = Qwen2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
-        self.rotary_emb = Qwen2RotaryEmbedding(config=config)
-
-        self.gradient_checkpointing = False
-        # Initialize weights and apply final processing
-        self.post_init()
-
-    def get_input_embeddings(self):
-        return self.embed_tokens
-
-    def set_input_embeddings(self, value):
-        self.embed_tokens = value
-
-    @add_start_docstrings_to_model_forward(QWEN2_INPUTS_DOCSTRING)
-    def forward(
-        self,
-        input_ids: torch.LongTensor = None,
-        attention_mask: Optional[torch.Tensor] = None,
-        position_ids: Optional[torch.LongTensor] = None,
-        past_key_values: Optional[List[torch.FloatTensor]] = None,
-        inputs_embeds: Optional[torch.FloatTensor] = None,
-        use_cache: Optional[bool] = None,
-        output_attentions: Optional[bool] = None,
-        output_hidden_states: Optional[bool] = None,
-        return_dict: Optional[bool] = None,
-        cache_position: Optional[torch.LongTensor] = None,
-    ) -> Union[Tuple, BaseModelOutputWithPast]:
-        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
-        output_hidden_states = (
-            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
-        )
-        use_cache = use_cache if use_cache is not None else self.config.use_cache
-
-        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
-
-        if (input_ids is None) ^ (inputs_embeds is not None):
-            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
-
-        if self.gradient_checkpointing and self.training:
-            if use_cache:
-                logger.warning_once(
-                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
-                )
-                use_cache = False
-
-        # kept for BC (non `Cache` `past_key_values` inputs)
-        return_legacy_cache = False
-        if use_cache and not isinstance(past_key_values, Cache):
-            return_legacy_cache = True
-            if past_key_values is None:
-                past_key_values = DynamicCache()
-            else:
-                past_key_values = DynamicCache.from_legacy_cache(past_key_values)
-                logger.warning_once(
-                    "We detected that you are passing `past_key_values` as a tuple of tuples. This is deprecated and "
-                    "will be removed in v4.47. Please convert your cache or use an appropriate `Cache` class "
-                    "(https://huggingface.co/docs/transformers/kv_cache#legacy-cache-format)"
-                )
-
-        if inputs_embeds is None:
-            inputs_embeds = self.embed_tokens(input_ids)
-
-        if cache_position is None:
-            past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
-            cache_position = torch.arange(
-                past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
-            )
-        if position_ids is None:
-            position_ids = cache_position.unsqueeze(0)
-
-        if attention_mask is not None and 0.0 in attention_mask:
-            causal_mask = attention_mask
-        else:
-            causal_mask = None
-
-        hidden_states = inputs_embeds
-        # create position embeddings to be shared across the decoder layers
-        position_embeddings = self.rotary_emb(hidden_states, position_ids)
-
-        # decoder layers
-        all_hidden_states = () if output_hidden_states else None
-        all_self_attns = () if output_attentions else None
-        next_decoder_cache = None
-
-        for decoder_layer in self.layers:
-            if output_hidden_states:
-                all_hidden_states += (hidden_states,)
-
-            if self.gradient_checkpointing and self.training:
-                layer_outputs = self._gradient_checkpointing_func(
-                    decoder_layer.__call__,
-                    hidden_states,
-                    causal_mask,
-                    position_ids,
-                    past_key_values,
-                    output_attentions,
-                    use_cache,
-                    cache_position,
-                    position_embeddings,
-                )
-            else:
-                layer_outputs = decoder_layer(
-                    hidden_states,
-                    attention_mask=causal_mask,
-                    position_ids=position_ids,
-                    past_key_value=past_key_values,
-                    output_attentions=output_attentions,
-                    use_cache=use_cache,
-                    cache_position=cache_position,
-                    position_embeddings=position_embeddings,
-                )
-
-            hidden_states = layer_outputs[0]
-
-            if use_cache:
-                next_decoder_cache = layer_outputs[2 if output_attentions else 1]
-
-            if output_attentions:
-                all_self_attns += (layer_outputs[1],)
-
-        hidden_states = self.norm(hidden_states)
-
-        # add hidden states from the last decoder layer
-        if output_hidden_states:
-            all_hidden_states += (hidden_states,)
-
-        next_cache = next_decoder_cache if use_cache else None
-        if return_legacy_cache:
-            next_cache = next_cache.to_legacy_cache()
-
-        if not return_dict:
-            return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
-        return BaseModelOutputWithPast(
-            last_hidden_state=hidden_states,
-            past_key_values=next_cache,
-            hidden_states=all_hidden_states,
-            attentions=all_self_attns,
-        )
-
-
-class Qwen2ForCausalLM(Qwen2PreTrainedModel, GenerationMixin):
-    _tied_weights_keys = ["lm_head.weight"]
-
-    def __init__(self, config):
-        super().__init__(config)
-        self.model = Qwen2Model(config)
-        self.vocab_size = config.vocab_size
-        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
-
-        # Initialize weights and apply final processing
-        self.post_init()
-
-    def get_input_embeddings(self):
-        return self.model.embed_tokens
-
-    def set_input_embeddings(self, value):
-        self.model.embed_tokens = value
-
-    def get_output_embeddings(self):
-        return self.lm_head
-
-    def set_output_embeddings(self, new_embeddings):
-        self.lm_head = new_embeddings
-
-    def set_decoder(self, decoder):
-        self.model = decoder
-
-    def get_decoder(self):
-        return self.model
-
-    @add_start_docstrings_to_model_forward(QWEN2_INPUTS_DOCSTRING)
-    @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
-    def forward(
-        self,
-        input_ids: torch.LongTensor = None,
-        attention_mask: Optional[torch.Tensor] = None,
-        position_ids: Optional[torch.LongTensor] = None,
-        past_key_values: Optional[List[torch.FloatTensor]] = None,
-        inputs_embeds: Optional[torch.FloatTensor] = None,
-        labels: Optional[torch.LongTensor] = None,
-        use_cache: Optional[bool] = None,
-        output_attentions: Optional[bool] = None,
-        output_hidden_states: Optional[bool] = None,
-        return_dict: Optional[bool] = None,
-        cache_position: Optional[torch.LongTensor] = None,
-        num_logits_to_keep: int = 0,
-        **loss_kwargs,
-    ) -> Union[Tuple, CausalLMOutputWithPast]:
-        r"""
-        Args:
-            labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
-                Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
-                config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
-                (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
-
-            num_logits_to_keep (`int`, *optional*):
-                Calculate logits for the last `num_logits_to_keep` tokens. If `0`, calculate logits for all
-                `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that
-                token can save memory, which becomes pretty significant for long sequences or large vocabulary size.
-
-        Returns:
-
-        Example:
-
-        ```python
-        >>> from transformers import AutoTokenizer, Qwen2ForCausalLM
-
-        >>> model = Qwen2ForCausalLM.from_pretrained(PATH_TO_CONVERTED_WEIGHTS)
-        >>> tokenizer = AutoTokenizer.from_pretrained(PATH_TO_CONVERTED_TOKENIZER)
-
-        >>> prompt = "Hey, are you conscious? Can you talk to me?"
-        >>> inputs = tokenizer(prompt, return_tensors="pt")
-
-        >>> # Generate
-        >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
-        >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
-        "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
-        ```"""
-
-        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
-        output_hidden_states = (
-            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
-        )
-        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
-
-        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
-        outputs = self.model(
-            input_ids=input_ids,
-            attention_mask=attention_mask,
-            position_ids=position_ids,
-            past_key_values=past_key_values,
-            inputs_embeds=inputs_embeds,
-            use_cache=use_cache,
-            output_attentions=output_attentions,
-            output_hidden_states=output_hidden_states,
-            return_dict=return_dict,
-            cache_position=cache_position,
-        )
-
-        hidden_states = outputs[0]
-        # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
-        logits = self.lm_head(hidden_states[:, -num_logits_to_keep:, :])
-
-        loss = None
-        if labels is not None:
-            loss = self.loss_function(logits, labels, self.vocab_size, **loss_kwargs)
-
-        if not return_dict:
-            output = (logits,) + outputs[1:]
-            return (loss,) + output if loss is not None else output
-
-        return CausalLMOutputWithPast(
-            loss=loss,
-            logits=logits,
-            past_key_values=outputs.past_key_values,
-            hidden_states=outputs.hidden_states,
-            attentions=outputs.attentions,
-        )
+# Copyright 2024 The Qwen Team and The HuggingFace Inc. team.
+# SPDX-License-Identifier: Apache-2.0
+
+"""PyTorch Qwen2 model."""
+
+import math
+from typing import List, Optional, Tuple, Union
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+
+from transformers.activations import ACT2FN
+from transformers.cache_utils import Cache, DynamicCache
+from transformers.generation import GenerationMixin
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPast,
+    CausalLMOutputWithPast,
+)
+from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS
+from transformers.modeling_utils import PreTrainedModel
+from transformers.utils import (
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    is_flash_attn_2_available,
+    is_flash_attn_greater_or_equal_2_10,
+    logging,
+    replace_return_docstrings,
+)
+from .configuration_qwen2 import Qwen2Config
+
+
+if is_flash_attn_2_available():
+    from transformers.modeling_flash_attention_utils import _flash_attention_forward
+
+
+logger = logging.get_logger(__name__)
+
+
+_CHECKPOINT_FOR_DOC = "Qwen/Qwen2-7B"
+_CONFIG_FOR_DOC = "Qwen2Config"
+
+
+# Copied from transformers.models.llama.modeling_llama.LlamaRMSNorm with Llama->Qwen2
+class Qwen2RMSNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-6):
+        """
+        Qwen2RMSNorm is equivalent to T5LayerNorm
+        """
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+
+    def forward(self, hidden_states):
+        input_dtype = hidden_states.dtype
+        hidden_states = hidden_states.to(torch.float32)
+        variance = hidden_states.pow(2).mean(-1, keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
+        return self.weight * hidden_states.to(input_dtype)
+
+    def extra_repr(self):
+        return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"
+
+
+# Copied from transformers.models.llama.modeling_llama.LlamaRotaryEmbedding with Llama->Qwen2
+class Qwen2RotaryEmbedding(nn.Module):
+    def __init__(
+        self,
+        dim=None,
+        max_position_embeddings=2048,
+        base=10000,
+        device=None,
+        scaling_factor=1.0,
+        rope_type="default",
+        config: Optional[Qwen2Config] = None,
+    ):
+        super().__init__()
+        # TODO (joao): remove the `if` below, only used for BC
+        self.rope_kwargs = {}
+        if config is None:
+            logger.warning_once(
+                "`Qwen2RotaryEmbedding` can now be fully parameterized by passing the model config through the "
+                "`config` argument. All other arguments will be removed in v4.46"
+            )
+            self.rope_kwargs = {
+                "rope_type": rope_type,
+                "factor": scaling_factor,
+                "dim": dim,
+                "base": base,
+                "max_position_embeddings": max_position_embeddings,
+            }
+            self.rope_type = rope_type
+            self.max_seq_len_cached = max_position_embeddings
+            self.original_max_seq_len = max_position_embeddings
+        else:
+            # BC: "rope_type" was originally "type"
+            if config.rope_scaling is not None:
+                self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))
+            else:
+                self.rope_type = "default"
+            self.max_seq_len_cached = config.max_position_embeddings
+            self.original_max_seq_len = config.max_position_embeddings
+
+        self.config = config
+        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+
+        inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device, **self.rope_kwargs)
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self.original_inv_freq = self.inv_freq
+
+    def _dynamic_frequency_update(self, position_ids, device):
+        """
+        dynamic RoPE layers should recompute `inv_freq` in the following situations:
+        1 - growing beyond the cached sequence length (allow scaling)
+        2 - the current sequence length is in the original scale (avoid losing precision with small sequences)
+        """
+        seq_len = torch.max(position_ids) + 1
+        if seq_len > self.max_seq_len_cached:  # growth
+            inv_freq, self.attention_scaling = self.rope_init_fn(
+                self.config, device, seq_len=seq_len, **self.rope_kwargs
+            )
+            self.register_buffer("inv_freq", inv_freq, persistent=False)  # TODO joao: may break with compilation
+            self.max_seq_len_cached = seq_len
+
+        if seq_len < self.original_max_seq_len and self.max_seq_len_cached > self.original_max_seq_len:  # reset
+            self.register_buffer("inv_freq", self.original_inv_freq, persistent=False)
+            self.max_seq_len_cached = self.original_max_seq_len
+
+    @torch.no_grad()
+    def forward(self, x, position_ids):
+        if "dynamic" in self.rope_type:
+            self._dynamic_frequency_update(position_ids, device=x.device)
+
+        # Core RoPE block
+        inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
+        position_ids_expanded = position_ids[:, None, :].float()
+        # Force float32 (see https://github.com/huggingface/transformers/pull/29285)
+        device_type = x.device.type
+        device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu"
+        with torch.autocast(device_type=device_type, enabled=False):
+            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
+            emb = torch.cat((freqs, freqs), dim=-1)
+            cos = emb.cos()
+            sin = emb.sin()
+
+        # Advanced RoPE types (e.g. yarn) apply a post-processing scaling factor, equivalent to scaling attention
+        cos = cos * self.attention_scaling
+        sin = sin * self.attention_scaling
+
+        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
+
+
+# Copied from transformers.models.llama.modeling_llama.rotate_half
+def rotate_half(x):
+    """Rotates half the hidden dims of the input."""
+    x1 = x[..., : x.shape[-1] // 2]
+    x2 = x[..., x.shape[-1] // 2 :]
+    return torch.cat((-x2, x1), dim=-1)
+
+
+# Copied from transformers.models.llama.modeling_llama.apply_rotary_pos_emb
+def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
+    """Applies Rotary Position Embedding to the query and key tensors.
+
+    Args:
+        q (`torch.Tensor`): The query tensor.
+        k (`torch.Tensor`): The key tensor.
+        cos (`torch.Tensor`): The cosine part of the rotary embedding.
+        sin (`torch.Tensor`): The sine part of the rotary embedding.
+        position_ids (`torch.Tensor`, *optional*):
+            Deprecated and unused.
+        unsqueeze_dim (`int`, *optional*, defaults to 1):
+            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
+            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
+            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
+            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
+            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
+            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
+    Returns:
+        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
+    """
+    cos = cos.unsqueeze(unsqueeze_dim)
+    sin = sin.unsqueeze(unsqueeze_dim)
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return q_embed, k_embed
+
+
+# Copied from transformers.models.mistral.modeling_mistral.MistralMLP with Mistral->Qwen2
+class Qwen2MLP(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = config.intermediate_size
+        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
+        self.act_fn = ACT2FN[config.hidden_act]
+
+    def forward(self, hidden_state):
+        return self.down_proj(self.act_fn(self.gate_proj(hidden_state)) * self.up_proj(hidden_state))
+
+
+# Copied from transformers.models.llama.modeling_llama.repeat_kv
+def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
+    """
+    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
+    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
+    """
+    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
+    if n_rep == 1:
+        return hidden_states
+    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
+    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
+
+
+class Qwen2Attention(nn.Module):
+    """
+    Multi-headed attention from 'Attention Is All You Need' paper. Modified to use sliding window attention: Longformer
+    and "Generating Long Sequences with Sparse Transformers".
+    """
+
+    def __init__(self, config: Qwen2Config, layer_idx: Optional[int] = None):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+        if layer_idx is None:
+            logger.warning_once(
+                f"Instantiating {self.__class__.__name__} without passing `layer_idx` is not recommended and will "
+                "to errors during the forward call, if caching is used. Please make sure to provide a `layer_idx` "
+                "when creating this class."
+            )
+
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.hidden_size // self.num_heads
+        self.num_key_value_heads = config.num_key_value_heads
+        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
+        self.max_position_embeddings = config.max_position_embeddings
+        self.rope_theta = config.rope_theta
+        self.is_causal = config.is_causal
+        self.attention_dropout = config.attention_dropout
+
+        if (self.head_dim * self.num_heads) != self.hidden_size:
+            raise ValueError(
+                f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
+                f" and `num_heads`: {self.num_heads})."
+            )
+        self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=True)
+        self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=True)
+        self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=True)
+        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # will become mandatory in v4.46
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        bsz, q_len, _ = hidden_states.size()
+
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+
+        if position_embeddings is None:
+            logger.warning_once(
+                "The attention layers in this model are transitioning from computing the RoPE embeddings internally "
+                "through `position_ids` (2D tensor with the indexes of the tokens), to using externally computed "
+                "`position_embeddings` (Tuple of tensors, containing cos and sin). In v4.46 `position_ids` will be "
+                "removed and `position_embeddings` will be mandatory."
+            )
+            cos, sin = self.rotary_emb(value_states, position_ids)
+        else:
+            cos, sin = position_embeddings
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
+
+        if past_key_value is not None:
+            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}  # Specific to RoPE models
+            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+        # repeat k/v heads if n_kv_heads < n_heads
+        key_states = repeat_kv(key_states, self.num_key_value_groups)
+        value_states = repeat_kv(value_states, self.num_key_value_groups)
+
+        attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
+        if attention_mask is not None:  # no matter the length, we just slice it
+            causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
+            attn_weights = attn_weights + causal_mask
+
+        # upcast attention to fp32
+        attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
+        attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training)
+        attn_output = torch.matmul(attn_weights, value_states)
+
+        if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
+            raise ValueError(
+                f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
+                f" {attn_output.size()}"
+            )
+
+        attn_output = attn_output.transpose(1, 2).contiguous()
+        attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
+
+        attn_output = self.o_proj(attn_output)
+
+        if not output_attentions:
+            attn_weights = None
+
+        return attn_output, attn_weights, past_key_value
+
+
+class Qwen2FlashAttention2(Qwen2Attention):
+    """
+    Qwen2 flash attention module, following Qwen2 attention module. This module inherits from `Qwen2Attention`
+    as the weights of the module stays untouched. The only required change would be on the forward pass
+    where it needs to correctly call the public API of flash attention and deal with padding tokens
+    in case the input contains any of them. Additionally, for sliding window attention, we apply SWA only to the bottom
+    config.max_window_layers layers.
+    """
+
+    # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2.__init__
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        # TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1.
+        # flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignement, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0.
+        # Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left).
+        self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # will become mandatory in v4.46
+    ):
+        bsz, q_len, _ = hidden_states.size()
+
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+
+        if position_embeddings is None:
+            logger.warning_once(
+                "The attention layers in this model are transitioning from computing the RoPE embeddings internally "
+                "through `position_ids` (2D tensor with the indexes of the tokens), to using externally computed "
+                "`position_embeddings` (Tuple of tensors, containing cos and sin). In v4.46 `position_ids` will be "
+                "removed and `position_embeddings` will be mandatory."
+            )
+            cos, sin = self.rotary_emb(value_states, position_ids)
+        else:
+            cos, sin = position_embeddings
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
+
+        if past_key_value is not None:
+            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}  # Specific to RoPE models
+            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+        # repeat k/v heads if n_kv_heads < n_heads
+        key_states = repeat_kv(key_states, self.num_key_value_groups)
+        value_states = repeat_kv(value_states, self.num_key_value_groups)
+        dropout_rate = 0.0 if not self.training else self.attention_dropout
+
+        # In PEFT, usually we cast the layer norms in float32 for training stability reasons
+        # therefore the input hidden states gets silently casted in float32. Hence, we need
+        # cast them back in float16 just to be sure everything works as expected.
+        input_dtype = query_states.dtype
+        if input_dtype == torch.float32:
+            if torch.is_autocast_enabled():
+                target_dtype = torch.get_autocast_gpu_dtype()
+            # Handle the case where the model is quantized
+            elif hasattr(self.config, "_pre_quantization_dtype"):
+                target_dtype = self.config._pre_quantization_dtype
+            else:
+                target_dtype = self.q_proj.weight.dtype
+
+            logger.warning_once(
+                f"The input hidden states seems to be silently casted in float32, this might be related to"
+                f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in"
+                f" {target_dtype}."
+            )
+
+            query_states = query_states.to(target_dtype)
+            key_states = key_states.to(target_dtype)
+            value_states = value_states.to(target_dtype)
+
+        # Reashape to the expected shape for Flash Attention
+        query_states = query_states.transpose(1, 2)
+        key_states = key_states.transpose(1, 2)
+        value_states = value_states.transpose(1, 2)
+
+        if (
+            self.config.use_sliding_window
+            and getattr(self.config, "sliding_window", None) is not None
+            and self.layer_idx >= self.config.max_window_layers
+        ):
+            sliding_window = self.config.sliding_window
+        else:
+            sliding_window = None
+
+        attn_output = _flash_attention_forward(
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            q_len,
+            position_ids=position_ids,
+            dropout=dropout_rate,
+            sliding_window=sliding_window,
+            is_causal=self.is_causal,
+            use_top_left_mask=self._flash_attn_uses_top_left_mask,
+        )
+
+        attn_output = attn_output.reshape(bsz, q_len, self.hidden_size).contiguous()
+        attn_output = self.o_proj(attn_output)
+
+        if not output_attentions:
+            attn_weights = None
+
+        return attn_output, attn_weights, past_key_value
+
+
+QWEN2_ATTENTION_CLASSES = {
+    "eager": Qwen2Attention,
+    "flash_attention_2": Qwen2FlashAttention2,
+}
+
+
+class Qwen2DecoderLayer(nn.Module):
+    def __init__(self, config: Qwen2Config, layer_idx: int):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+
+        if config.sliding_window and config._attn_implementation != "flash_attention_2":
+            logger.warning_once(
+                f"Sliding Window Attention is enabled but not implemented for `{config._attn_implementation}`; "
+                "unexpected results may be encountered."
+            )
+        self.self_attn = QWEN2_ATTENTION_CLASSES[config._attn_implementation](config, layer_idx)
+
+        self.mlp = Qwen2MLP(config)
+        self.input_layernorm = Qwen2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = Qwen2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        output_attentions: Optional[bool] = False,
+        use_cache: Optional[bool] = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # will become mandatory in v4.46
+        **kwargs,
+    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
+        """
+        Args:
+            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
+            attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
+                `(batch, sequence_length)` where padding elements are indicated by 0.
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+            use_cache (`bool`, *optional*):
+                If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
+                (see `past_key_values`).
+            past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
+            cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
+                Indices depicting the position of the input sequence tokens in the sequence.
+            position_embeddings (`Tuple[torch.FloatTensor, torch.FloatTensor]`, *optional*):
+                Tuple containing the cosine and sine positional embeddings of shape `(batch_size, seq_len, head_dim)`,
+                with `head_dim` being the embedding dimension of each attention head.
+            kwargs (`dict`, *optional*):
+                Arbitrary kwargs to be ignored, used for FSDP and other methods that injects code
+                into the model
+        """
+
+        residual = hidden_states
+
+        hidden_states = self.input_layernorm(hidden_states)
+
+        # Self Attention
+        hidden_states, self_attn_weights, present_key_value = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_value=past_key_value,
+            output_attentions=output_attentions,
+            use_cache=use_cache,
+            cache_position=cache_position,
+            position_embeddings=position_embeddings,
+        )
+        hidden_states = residual + hidden_states
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (self_attn_weights,)
+
+        if use_cache:
+            outputs += (present_key_value,)
+
+        return outputs
+
+
+QWEN2_START_DOCSTRING = r"""
+    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
+    etc.)
+
+    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
+    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
+    and behavior.
+
+    Parameters:
+        config ([`Qwen2Config`]):
+            Model configuration class with all the parameters of the model. Initializing with a config file does not
+            load the weights associated with the model, only the configuration. Check out the
+            [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+
+@add_start_docstrings(
+    "The bare Qwen2 Model outputting raw hidden-states without any specific head on top.",
+    QWEN2_START_DOCSTRING,
+)
+class Qwen2PreTrainedModel(PreTrainedModel):
+    config_class = Qwen2Config
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["Qwen2DecoderLayer"]
+    _skip_keys_device_placement = "past_key_values"
+    _supports_flash_attn_2 = True
+    _supports_cache_class = True
+    _supports_quantized_cache = True
+    _supports_static_cache = True
+
+    def _init_weights(self, module):
+        std = self.config.initializer_range
+        if isinstance(module, nn.Linear):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=std)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+
+
+QWEN2_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+            Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
+            it.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            If `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see
+            `past_key_values`).
+
+            If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
+            and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
+            information on the default strategy.
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+        position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.n_positions - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        past_key_values (`Cache` or `tuple(tuple(torch.FloatTensor))`, *optional*):
+            Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
+            blocks) that can be used to speed up sequential decoding. This typically consists in the `past_key_values`
+            returned by the model at a previous stage of decoding, when `use_cache=True` or `config.use_cache=True`.
+
+            Two formats are allowed:
+            - a [`~cache_utils.Cache`] instance, see our
+            [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache);
+            - Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of
+            shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`). This is also known as the legacy
+            cache format.
+
+            The model will output the same cache format that is fed as input. If no `past_key_values` are passed, the
+            legacy cache format will be returned.
+
+            If `past_key_values` are used, the user can optionally input only the last `input_ids` (those that don't
+            have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `input_ids`
+            of shape `(batch_size, sequence_length)`.
+        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
+            is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
+            model's internal embedding lookup matrix.
+        use_cache (`bool`, *optional*):
+            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
+            `past_key_values`).
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+        cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
+            Indices depicting the position of the input sequence tokens in the sequence. Contrarily to `position_ids`,
+            this tensor is not affected by padding. It is used to update the cache in the correct position and to infer
+            the complete sequence length.
+"""
+
+
+@add_start_docstrings(
+    "The bare Qwen2 Model outputting raw hidden-states without any specific head on top.",
+    QWEN2_START_DOCSTRING,
+)
+class Qwen2Model(Qwen2PreTrainedModel):
+    """
+    Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`Qwen2DecoderLayer`]
+
+    Args:
+        config: Qwen2Config
+    """
+
+    def __init__(self, config: Qwen2Config):
+        super().__init__(config)
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
+        self.layers = nn.ModuleList(
+            [Qwen2DecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
+        )
+        self._attn_implementation = config._attn_implementation
+        self.norm = Qwen2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.rotary_emb = Qwen2RotaryEmbedding(config=config)
+
+        self.gradient_checkpointing = False
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.embed_tokens = value
+
+    @add_start_docstrings_to_model_forward(QWEN2_INPUTS_DOCSTRING)
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+    ) -> Union[Tuple, BaseModelOutputWithPast]:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if (input_ids is None) ^ (inputs_embeds is not None):
+            raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
+
+        if self.gradient_checkpointing and self.training:
+            if use_cache:
+                logger.warning_once(
+                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+                )
+                use_cache = False
+
+        # kept for BC (non `Cache` `past_key_values` inputs)
+        return_legacy_cache = False
+        if use_cache and not isinstance(past_key_values, Cache):
+            return_legacy_cache = True
+            if past_key_values is None:
+                past_key_values = DynamicCache()
+            else:
+                past_key_values = DynamicCache.from_legacy_cache(past_key_values)
+                logger.warning_once(
+                    "We detected that you are passing `past_key_values` as a tuple of tuples. This is deprecated and "
+                    "will be removed in v4.47. Please convert your cache or use an appropriate `Cache` class "
+                    "(https://huggingface.co/docs/transformers/kv_cache#legacy-cache-format)"
+                )
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        if cache_position is None:
+            past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
+            cache_position = torch.arange(
+                past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
+            )
+        if position_ids is None:
+            position_ids = cache_position.unsqueeze(0)
+
+        if attention_mask is not None and 0.0 in attention_mask:
+            causal_mask = attention_mask
+        else:
+            causal_mask = None
+
+        hidden_states = inputs_embeds
+        # create position embeddings to be shared across the decoder layers
+        position_embeddings = self.rotary_emb(hidden_states, position_ids)
+
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        next_decoder_cache = None
+
+        for decoder_layer in self.layers:
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            if self.gradient_checkpointing and self.training:
+                layer_outputs = self._gradient_checkpointing_func(
+                    decoder_layer.__call__,
+                    hidden_states,
+                    causal_mask,
+                    position_ids,
+                    past_key_values,
+                    output_attentions,
+                    use_cache,
+                    cache_position,
+                    position_embeddings,
+                )
+            else:
+                layer_outputs = decoder_layer(
+                    hidden_states,
+                    attention_mask=causal_mask,
+                    position_ids=position_ids,
+                    past_key_value=past_key_values,
+                    output_attentions=output_attentions,
+                    use_cache=use_cache,
+                    cache_position=cache_position,
+                    position_embeddings=position_embeddings,
+                )
+
+            hidden_states = layer_outputs[0]
+
+            if use_cache:
+                next_decoder_cache = layer_outputs[2 if output_attentions else 1]
+
+            if output_attentions:
+                all_self_attns += (layer_outputs[1],)
+
+        hidden_states = self.norm(hidden_states)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        next_cache = next_decoder_cache if use_cache else None
+        if return_legacy_cache:
+            next_cache = next_cache.to_legacy_cache()
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
+        return BaseModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=next_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+        )
+
+
+class Qwen2ForCausalLM(Qwen2PreTrainedModel, GenerationMixin):
+    _tied_weights_keys = ["lm_head.weight"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = Qwen2Model(config)
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def set_decoder(self, decoder):
+        self.model = decoder
+
+    def get_decoder(self):
+        return self.model
+
+    @add_start_docstrings_to_model_forward(QWEN2_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        num_logits_to_keep: int = 0,
+        **loss_kwargs,
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+        r"""
+        Args:
+            labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+                config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+                (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+            num_logits_to_keep (`int`, *optional*):
+                Calculate logits for the last `num_logits_to_keep` tokens. If `0`, calculate logits for all
+                `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that
+                token can save memory, which becomes pretty significant for long sequences or large vocabulary size.
+
+        Returns:
+
+        Example:
+
+        ```python
+        >>> from transformers import AutoTokenizer, Qwen2ForCausalLM
+
+        >>> model = Qwen2ForCausalLM.from_pretrained(PATH_TO_CONVERTED_WEIGHTS)
+        >>> tokenizer = AutoTokenizer.from_pretrained(PATH_TO_CONVERTED_TOKENIZER)
+
+        >>> prompt = "Hey, are you conscious? Can you talk to me?"
+        >>> inputs = tokenizer(prompt, return_tensors="pt")
+
+        >>> # Generate
+        >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
+        >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+        "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
+        ```"""
+
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+        outputs = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            cache_position=cache_position,
+        )
+
+        hidden_states = outputs[0]
+        # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
+        logits = self.lm_head(hidden_states[:, -num_logits_to_keep:, :])
+
+        loss = None
+        if labels is not None:
+            loss = self.loss_function(logits, labels, self.vocab_size, **loss_kwargs)
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return (loss,) + output if loss is not None else output
+
+        return CausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )

modeling/qwen2/tokenization_qwen2.py

@@ -1,328 +1,328 @@
-# Copyright 2024 The Qwen Team and The HuggingFace Inc. team.
-# SPDX-License-Identifier: Apache-2.0
-
-"""Tokenization classes for Qwen2."""
-
-import json
-import os
-import unicodedata
-from functools import lru_cache
-from typing import Optional, Tuple
-
-import regex as re
-
-from transformers.tokenization_utils import AddedToken, PreTrainedTokenizer
-from transformers.utils import logging
-
-
-logger = logging.get_logger(__name__)
-
-VOCAB_FILES_NAMES = {
-    "vocab_file": "vocab.json",
-    "merges_file": "merges.txt",
-}
-
-
-MAX_MODEL_INPUT_SIZES = {"qwen/qwen-tokenizer": 32768}
-
-PRETOKENIZE_REGEX = r"""(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\r\n\p{L}\p{N}]?\p{L}+|\p{N}| ?[^\s\p{L}\p{N}]+[\r\n]*|\s*[\r\n]+|\s+(?!\S)|\s+"""
-
-
-@lru_cache()
-# Copied from transformers.models.gpt2.tokenization_gpt2.bytes_to_unicode
-def bytes_to_unicode():
-    """
-    Returns list of utf-8 byte and a mapping to unicode strings. We specifically avoids mapping to whitespace/control
-    characters the bpe code barfs on.
-
-    The reversible bpe codes work on unicode strings. This means you need a large # of unicode characters in your vocab
-    if you want to avoid UNKs. When you're at something like a 10B token dataset you end up needing around 5K for
-    decent coverage. This is a significant percentage of your normal, say, 32K bpe vocab. To avoid that, we want lookup
-    tables between utf-8 bytes and unicode strings.
-    """
-    bs = (
-        list(range(ord("!"), ord("~") + 1)) + list(range(ord("¡"), ord("¬") + 1)) + list(range(ord("®"), ord("ÿ") + 1))
-    )
-    cs = bs[:]
-    n = 0
-    for b in range(2**8):
-        if b not in bs:
-            bs.append(b)
-            cs.append(2**8 + n)
-            n += 1
-    cs = [chr(n) for n in cs]
-    return dict(zip(bs, cs))
-
-
-# Copied from transformers.models.gpt2.tokenization_gpt2.get_pairs
-def get_pairs(word):
-    """
-    Return set of symbol pairs in a word.
-
-    Word is represented as tuple of symbols (symbols being variable-length strings).
-    """
-    pairs = set()
-    prev_char = word[0]
-    for char in word[1:]:
-        pairs.add((prev_char, char))
-        prev_char = char
-    return pairs
-
-
-class Qwen2Tokenizer(PreTrainedTokenizer):
-    """
-    Construct a Qwen2 tokenizer. Based on byte-level Byte-Pair-Encoding.
-
-    Same with GPT2Tokenizer, this tokenizer has been trained to treat spaces like parts of the tokens so a word will
-    be encoded differently whether it is at the beginning of the sentence (without space) or not:
-
-    ```python
-    >>> from transformers import Qwen2Tokenizer
-
-    >>> tokenizer = Qwen2Tokenizer.from_pretrained("Qwen/Qwen-tokenizer")
-    >>> tokenizer("Hello world")["input_ids"]
-    [9707, 1879]
-
-    >>> tokenizer(" Hello world")["input_ids"]
-    [21927, 1879]
-    ```
-    This is expected.
-
-    You should not use GPT2Tokenizer instead, because of the different pretokenization rules.
-
-    This tokenizer inherits from [`PreTrainedTokenizer`] which contains most of the main methods. Users should refer to
-    this superclass for more information regarding those methods.
-
-    Args:
-        vocab_file (`str`):
-            Path to the vocabulary file.
-        merges_file (`str`):
-            Path to the merges file.
-        errors (`str`, *optional*, defaults to `"replace"`):
-            Paradigm to follow when decoding bytes to UTF-8. See
-            [bytes.decode](https://docs.python.org/3/library/stdtypes.html#bytes.decode) for more information.
-        unk_token (`str`, *optional*, defaults to `"<|endoftext|>"`):
-            The unknown token. A token that is not in the vocabulary cannot be converted to an ID and is set to be this
-            token instead.
-        bos_token (`str`, *optional*):
-            The beginning of sequence token. Not applicable for this tokenizer.
-        eos_token (`str`, *optional*, defaults to `"<|endoftext|>"`):
-            The end of sequence token.
-        pad_token (`str`, *optional*, defaults to `"<|endoftext|>"`):
-            The token used for padding, for example when batching sequences of different lengths.
-        clean_up_tokenization_spaces (`bool`, *optional*, defaults to `False`):
-            Whether or not the model should cleanup the spaces that were added when splitting the input text during the
-            tokenization process. Not applicable to this tokenizer, since tokenization does not add spaces.
-        split_special_tokens (`bool`, *optional*, defaults to `False`):
-            Whether or not the special tokens should be split during the tokenization process. The default behavior is
-            to not split special tokens. This means that if `<|endoftext|>` is the `eos_token`, then `tokenizer.tokenize("<|endoftext|>") =
-            ['<|endoftext|>`]. Otherwise, if `split_special_tokens=True`, then `tokenizer.tokenize("<|endoftext|>")` will be give `['<',
-            '|', 'endo', 'ft', 'ext', '|', '>']`. This argument is only supported for `slow` tokenizers for the moment.
-    """
-
-    vocab_files_names = VOCAB_FILES_NAMES
-    model_input_names = ["input_ids", "attention_mask"]
-
-    def __init__(
-        self,
-        vocab_file,
-        merges_file,
-        errors="replace",
-        unk_token="<|endoftext|>",
-        bos_token=None,
-        eos_token="<|endoftext|>",
-        pad_token="<|endoftext|>",
-        clean_up_tokenization_spaces=False,
-        split_special_tokens=False,
-        **kwargs,
-    ):
-        # Qwen vocab does not contain control tokens; added tokens need to be special
-        bos_token = (
-            AddedToken(bos_token, lstrip=False, rstrip=False, special=True, normalized=False)
-            if isinstance(bos_token, str)
-            else bos_token
-        )
-        eos_token = (
-            AddedToken(eos_token, lstrip=False, rstrip=False, special=True, normalized=False)
-            if isinstance(eos_token, str)
-            else eos_token
-        )
-        unk_token = (
-            AddedToken(unk_token, lstrip=False, rstrip=False, special=True, normalized=False)
-            if isinstance(unk_token, str)
-            else unk_token
-        )
-        pad_token = (
-            AddedToken(pad_token, lstrip=False, rstrip=False, special=True, normalized=False)
-            if isinstance(pad_token, str)
-            else pad_token
-        )
-
-        with open(vocab_file, encoding="utf-8") as vocab_handle:
-            self.encoder = json.load(vocab_handle)
-        self.decoder = {v: k for k, v in self.encoder.items()}
-        self.errors = errors  # how to handle errors in decoding
-        self.byte_encoder = bytes_to_unicode()
-        self.byte_decoder = {v: k for k, v in self.byte_encoder.items()}
-        bpe_merges = []
-        with open(merges_file, encoding="utf-8") as merges_handle:
-            for i, line in enumerate(merges_handle):
-                line = line.strip()
-                if (i == 0 and line.startswith("#version:")) or not line:
-                    continue
-                bpe_merges.append(tuple(line.split()))
-        self.bpe_ranks = dict(zip(bpe_merges, range(len(bpe_merges))))
-        # NOTE: the cache can grow without bound and will get really large for long running processes
-        # (esp. for texts of language that do not use space between word, e.g. Chinese); technically
-        # not a memory leak but appears as one.
-        # GPT2Tokenizer has the same problem, so let's be consistent.
-        self.cache = {}
-
-        self.pat = re.compile(PRETOKENIZE_REGEX)
-
-        if kwargs.get("add_prefix_space", False):
-            logger.warning_once(
-                f"{self.__class__.__name} does not support `add_prefix_space`, setting it to True has no effect."
-            )
-
-        super().__init__(
-            errors=errors,
-            bos_token=bos_token,
-            eos_token=eos_token,
-            pad_token=pad_token,
-            unk_token=unk_token,
-            clean_up_tokenization_spaces=clean_up_tokenization_spaces,
-            split_special_tokens=split_special_tokens,
-            **kwargs,
-        )
-
-    @property
-    def vocab_size(self) -> int:
-        return len(self.encoder)
-
-    # Copied from transformers.models.gpt2.tokenization_gpt2.GPT2Tokenizer.get_vocab
-    def get_vocab(self):
-        return dict(self.encoder, **self.added_tokens_encoder)
-
-    # Copied from transformers.models.gpt2.tokenization_gpt2.GPT2Tokenizer.bpe
-    def bpe(self, token):
-        if token in self.cache:
-            return self.cache[token]
-        word = tuple(token)
-        pairs = get_pairs(word)
-
-        if not pairs:
-            return token
-
-        while True:
-            bigram = min(pairs, key=lambda pair: self.bpe_ranks.get(pair, float("inf")))
-            if bigram not in self.bpe_ranks:
-                break
-            first, second = bigram
-            new_word = []
-            i = 0
-            while i < len(word):
-                try:
-                    j = word.index(first, i)
-                except ValueError:
-                    new_word.extend(word[i:])
-                    break
-                else:
-                    new_word.extend(word[i:j])
-                    i = j
-
-                if word[i] == first and i < len(word) - 1 and word[i + 1] == second:
-                    new_word.append(first + second)
-                    i += 2
-                else:
-                    new_word.append(word[i])
-                    i += 1
-            new_word = tuple(new_word)
-            word = new_word
-            if len(word) == 1:
-                break
-            else:
-                pairs = get_pairs(word)
-        word = " ".join(word)
-        self.cache[token] = word
-        return word
-
-    # Copied from transformers.models.gpt2.tokenization_gpt2.GPT2Tokenizer._tokenize
-    def _tokenize(self, text):
-        """Tokenize a string."""
-        bpe_tokens = []
-        for token in re.findall(self.pat, text):
-            token = "".join(
-                self.byte_encoder[b] for b in token.encode("utf-8")
-            )  # Maps all our bytes to unicode strings, avoiding control tokens of the BPE (spaces in our case)
-            bpe_tokens.extend(bpe_token for bpe_token in self.bpe(token).split(" "))
-        return bpe_tokens
-
-    # Copied from transformers.models.gpt2.tokenization_gpt2.GPT2Tokenizer._convert_token_to_id
-    def _convert_token_to_id(self, token):
-        """Converts a token (str) in an id using the vocab."""
-        return self.encoder.get(token, self.encoder.get(self.unk_token))
-
-    # Copied from transformers.models.gpt2.tokenization_gpt2.GPT2Tokenizer._convert_id_to_token
-    def _convert_id_to_token(self, index):
-        """Converts an index (integer) in a token (str) using the vocab."""
-        return self.decoder.get(index)
-
-    # Copied from transformers.models.gpt2.tokenization_gpt2.GPT2Tokenizer.convert_tokens_to_string
-    def convert_tokens_to_string(self, tokens):
-        """Converts a sequence of tokens (string) in a single string."""
-        text = "".join(tokens)
-        text = bytearray([self.byte_decoder[c] for c in text]).decode("utf-8", errors=self.errors)
-        return text
-
-    def decode(
-        self,
-        token_ids,
-        skip_special_tokens: bool = False,
-        clean_up_tokenization_spaces: Optional[bool] = False,
-        spaces_between_special_tokens: bool = False,
-        **kwargs,
-    ) -> str:
-        # `spaces_between_special_tokens` defaults to True for _decode in slow tokenizers
-        # and cannot be configured elsewhere, but it should default to False for Qwen2Tokenizer
-        return super().decode(
-            token_ids,
-            skip_special_tokens=skip_special_tokens,
-            clean_up_tokenization_spaces=clean_up_tokenization_spaces,
-            spaces_between_special_tokens=spaces_between_special_tokens,
-            **kwargs,
-        )
-
-    # Copied from transformers.models.gpt2.tokenization_gpt2.GPT2Tokenizer.save_vocabulary
-    def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> Tuple[str]:
-        if not os.path.isdir(save_directory):
-            logger.error(f"Vocabulary path ({save_directory}) should be a directory")
-            return
-        vocab_file = os.path.join(
-            save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["vocab_file"]
-        )
-        merge_file = os.path.join(
-            save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["merges_file"]
-        )
-
-        with open(vocab_file, "w", encoding="utf-8") as f:
-            f.write(json.dumps(self.encoder, indent=2, sort_keys=True, ensure_ascii=False) + "\n")
-
-        index = 0
-        with open(merge_file, "w", encoding="utf-8") as writer:
-            writer.write("#version: 0.2\n")
-            for bpe_tokens, token_index in sorted(self.bpe_ranks.items(), key=lambda kv: kv[1]):
-                if index != token_index:
-                    logger.warning(
-                        f"Saving vocabulary to {merge_file}: BPE merge indices are not consecutive."
-                        " Please check that the tokenizer is not corrupted!"
-                    )
-                    index = token_index
-                writer.write(" ".join(bpe_tokens) + "\n")
-                index += 1
-
-        return vocab_file, merge_file
-
-    def prepare_for_tokenization(self, text, **kwargs):
-        text = unicodedata.normalize("NFC", text)
-        return (text, kwargs)
+# Copyright 2024 The Qwen Team and The HuggingFace Inc. team.
+# SPDX-License-Identifier: Apache-2.0
+
+"""Tokenization classes for Qwen2."""
+
+import json
+import os
+import unicodedata
+from functools import lru_cache
+from typing import Optional, Tuple
+
+import regex as re
+
+from transformers.tokenization_utils import AddedToken, PreTrainedTokenizer
+from transformers.utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+VOCAB_FILES_NAMES = {
+    "vocab_file": "vocab.json",
+    "merges_file": "merges.txt",
+}
+
+
+MAX_MODEL_INPUT_SIZES = {"qwen/qwen-tokenizer": 32768}
+
+PRETOKENIZE_REGEX = r"""(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\r\n\p{L}\p{N}]?\p{L}+|\p{N}| ?[^\s\p{L}\p{N}]+[\r\n]*|\s*[\r\n]+|\s+(?!\S)|\s+"""
+
+
+@lru_cache()
+# Copied from transformers.models.gpt2.tokenization_gpt2.bytes_to_unicode
+def bytes_to_unicode():
+    """
+    Returns list of utf-8 byte and a mapping to unicode strings. We specifically avoids mapping to whitespace/control
+    characters the bpe code barfs on.
+
+    The reversible bpe codes work on unicode strings. This means you need a large # of unicode characters in your vocab
+    if you want to avoid UNKs. When you're at something like a 10B token dataset you end up needing around 5K for
+    decent coverage. This is a significant percentage of your normal, say, 32K bpe vocab. To avoid that, we want lookup
+    tables between utf-8 bytes and unicode strings.
+    """
+    bs = (
+        list(range(ord("!"), ord("~") + 1)) + list(range(ord("¡"), ord("¬") + 1)) + list(range(ord("®"), ord("ÿ") + 1))
+    )
+    cs = bs[:]
+    n = 0
+    for b in range(2**8):
+        if b not in bs:
+            bs.append(b)
+            cs.append(2**8 + n)
+            n += 1
+    cs = [chr(n) for n in cs]
+    return dict(zip(bs, cs))
+
+
+# Copied from transformers.models.gpt2.tokenization_gpt2.get_pairs
+def get_pairs(word):
+    """
+    Return set of symbol pairs in a word.
+
+    Word is represented as tuple of symbols (symbols being variable-length strings).
+    """
+    pairs = set()
+    prev_char = word[0]
+    for char in word[1:]:
+        pairs.add((prev_char, char))
+        prev_char = char
+    return pairs
+
+
+class Qwen2Tokenizer(PreTrainedTokenizer):
+    """
+    Construct a Qwen2 tokenizer. Based on byte-level Byte-Pair-Encoding.
+
+    Same with GPT2Tokenizer, this tokenizer has been trained to treat spaces like parts of the tokens so a word will
+    be encoded differently whether it is at the beginning of the sentence (without space) or not:
+
+    ```python
+    >>> from transformers import Qwen2Tokenizer
+
+    >>> tokenizer = Qwen2Tokenizer.from_pretrained("Qwen/Qwen-tokenizer")
+    >>> tokenizer("Hello world")["input_ids"]
+    [9707, 1879]
+
+    >>> tokenizer(" Hello world")["input_ids"]
+    [21927, 1879]
+    ```
+    This is expected.
+
+    You should not use GPT2Tokenizer instead, because of the different pretokenization rules.
+
+    This tokenizer inherits from [`PreTrainedTokenizer`] which contains most of the main methods. Users should refer to
+    this superclass for more information regarding those methods.
+
+    Args:
+        vocab_file (`str`):
+            Path to the vocabulary file.
+        merges_file (`str`):
+            Path to the merges file.
+        errors (`str`, *optional*, defaults to `"replace"`):
+            Paradigm to follow when decoding bytes to UTF-8. See
+            [bytes.decode](https://docs.python.org/3/library/stdtypes.html#bytes.decode) for more information.
+        unk_token (`str`, *optional*, defaults to `"<|endoftext|>"`):
+            The unknown token. A token that is not in the vocabulary cannot be converted to an ID and is set to be this
+            token instead.
+        bos_token (`str`, *optional*):
+            The beginning of sequence token. Not applicable for this tokenizer.
+        eos_token (`str`, *optional*, defaults to `"<|endoftext|>"`):
+            The end of sequence token.
+        pad_token (`str`, *optional*, defaults to `"<|endoftext|>"`):
+            The token used for padding, for example when batching sequences of different lengths.
+        clean_up_tokenization_spaces (`bool`, *optional*, defaults to `False`):
+            Whether or not the model should cleanup the spaces that were added when splitting the input text during the
+            tokenization process. Not applicable to this tokenizer, since tokenization does not add spaces.
+        split_special_tokens (`bool`, *optional*, defaults to `False`):
+            Whether or not the special tokens should be split during the tokenization process. The default behavior is
+            to not split special tokens. This means that if `<|endoftext|>` is the `eos_token`, then `tokenizer.tokenize("<|endoftext|>") =
+            ['<|endoftext|>`]. Otherwise, if `split_special_tokens=True`, then `tokenizer.tokenize("<|endoftext|>")` will be give `['<',
+            '|', 'endo', 'ft', 'ext', '|', '>']`. This argument is only supported for `slow` tokenizers for the moment.
+    """
+
+    vocab_files_names = VOCAB_FILES_NAMES
+    model_input_names = ["input_ids", "attention_mask"]
+
+    def __init__(
+        self,
+        vocab_file,
+        merges_file,
+        errors="replace",
+        unk_token="<|endoftext|>",
+        bos_token=None,
+        eos_token="<|endoftext|>",
+        pad_token="<|endoftext|>",
+        clean_up_tokenization_spaces=False,
+        split_special_tokens=False,
+        **kwargs,
+    ):
+        # Qwen vocab does not contain control tokens; added tokens need to be special
+        bos_token = (
+            AddedToken(bos_token, lstrip=False, rstrip=False, special=True, normalized=False)
+            if isinstance(bos_token, str)
+            else bos_token
+        )
+        eos_token = (
+            AddedToken(eos_token, lstrip=False, rstrip=False, special=True, normalized=False)
+            if isinstance(eos_token, str)
+            else eos_token
+        )
+        unk_token = (
+            AddedToken(unk_token, lstrip=False, rstrip=False, special=True, normalized=False)
+            if isinstance(unk_token, str)
+            else unk_token
+        )
+        pad_token = (
+            AddedToken(pad_token, lstrip=False, rstrip=False, special=True, normalized=False)
+            if isinstance(pad_token, str)
+            else pad_token
+        )
+
+        with open(vocab_file, encoding="utf-8") as vocab_handle:
+            self.encoder = json.load(vocab_handle)
+        self.decoder = {v: k for k, v in self.encoder.items()}
+        self.errors = errors  # how to handle errors in decoding
+        self.byte_encoder = bytes_to_unicode()
+        self.byte_decoder = {v: k for k, v in self.byte_encoder.items()}
+        bpe_merges = []
+        with open(merges_file, encoding="utf-8") as merges_handle:
+            for i, line in enumerate(merges_handle):
+                line = line.strip()
+                if (i == 0 and line.startswith("#version:")) or not line:
+                    continue
+                bpe_merges.append(tuple(line.split()))
+        self.bpe_ranks = dict(zip(bpe_merges, range(len(bpe_merges))))
+        # NOTE: the cache can grow without bound and will get really large for long running processes
+        # (esp. for texts of language that do not use space between word, e.g. Chinese); technically
+        # not a memory leak but appears as one.
+        # GPT2Tokenizer has the same problem, so let's be consistent.
+        self.cache = {}
+
+        self.pat = re.compile(PRETOKENIZE_REGEX)
+
+        if kwargs.get("add_prefix_space", False):
+            logger.warning_once(
+                f"{self.__class__.__name} does not support `add_prefix_space`, setting it to True has no effect."
+            )
+
+        super().__init__(
+            errors=errors,
+            bos_token=bos_token,
+            eos_token=eos_token,
+            pad_token=pad_token,
+            unk_token=unk_token,
+            clean_up_tokenization_spaces=clean_up_tokenization_spaces,
+            split_special_tokens=split_special_tokens,
+            **kwargs,
+        )
+
+    @property
+    def vocab_size(self) -> int:
+        return len(self.encoder)
+
+    # Copied from transformers.models.gpt2.tokenization_gpt2.GPT2Tokenizer.get_vocab
+    def get_vocab(self):
+        return dict(self.encoder, **self.added_tokens_encoder)
+
+    # Copied from transformers.models.gpt2.tokenization_gpt2.GPT2Tokenizer.bpe
+    def bpe(self, token):
+        if token in self.cache:
+            return self.cache[token]
+        word = tuple(token)
+        pairs = get_pairs(word)
+
+        if not pairs:
+            return token
+
+        while True:
+            bigram = min(pairs, key=lambda pair: self.bpe_ranks.get(pair, float("inf")))
+            if bigram not in self.bpe_ranks:
+                break
+            first, second = bigram
+            new_word = []
+            i = 0
+            while i < len(word):
+                try:
+                    j = word.index(first, i)
+                except ValueError:
+                    new_word.extend(word[i:])
+                    break
+                else:
+                    new_word.extend(word[i:j])
+                    i = j
+
+                if word[i] == first and i < len(word) - 1 and word[i + 1] == second:
+                    new_word.append(first + second)
+                    i += 2
+                else:
+                    new_word.append(word[i])
+                    i += 1
+            new_word = tuple(new_word)
+            word = new_word
+            if len(word) == 1:
+                break
+            else:
+                pairs = get_pairs(word)
+        word = " ".join(word)
+        self.cache[token] = word
+        return word
+
+    # Copied from transformers.models.gpt2.tokenization_gpt2.GPT2Tokenizer._tokenize
+    def _tokenize(self, text):
+        """Tokenize a string."""
+        bpe_tokens = []
+        for token in re.findall(self.pat, text):
+            token = "".join(
+                self.byte_encoder[b] for b in token.encode("utf-8")
+            )  # Maps all our bytes to unicode strings, avoiding control tokens of the BPE (spaces in our case)
+            bpe_tokens.extend(bpe_token for bpe_token in self.bpe(token).split(" "))
+        return bpe_tokens
+
+    # Copied from transformers.models.gpt2.tokenization_gpt2.GPT2Tokenizer._convert_token_to_id
+    def _convert_token_to_id(self, token):
+        """Converts a token (str) in an id using the vocab."""
+        return self.encoder.get(token, self.encoder.get(self.unk_token))
+
+    # Copied from transformers.models.gpt2.tokenization_gpt2.GPT2Tokenizer._convert_id_to_token
+    def _convert_id_to_token(self, index):
+        """Converts an index (integer) in a token (str) using the vocab."""
+        return self.decoder.get(index)
+
+    # Copied from transformers.models.gpt2.tokenization_gpt2.GPT2Tokenizer.convert_tokens_to_string
+    def convert_tokens_to_string(self, tokens):
+        """Converts a sequence of tokens (string) in a single string."""
+        text = "".join(tokens)
+        text = bytearray([self.byte_decoder[c] for c in text]).decode("utf-8", errors=self.errors)
+        return text
+
+    def decode(
+        self,
+        token_ids,
+        skip_special_tokens: bool = False,
+        clean_up_tokenization_spaces: Optional[bool] = False,
+        spaces_between_special_tokens: bool = False,
+        **kwargs,
+    ) -> str:
+        # `spaces_between_special_tokens` defaults to True for _decode in slow tokenizers
+        # and cannot be configured elsewhere, but it should default to False for Qwen2Tokenizer
+        return super().decode(
+            token_ids,
+            skip_special_tokens=skip_special_tokens,
+            clean_up_tokenization_spaces=clean_up_tokenization_spaces,
+            spaces_between_special_tokens=spaces_between_special_tokens,
+            **kwargs,
+        )
+
+    # Copied from transformers.models.gpt2.tokenization_gpt2.GPT2Tokenizer.save_vocabulary
+    def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> Tuple[str]:
+        if not os.path.isdir(save_directory):
+            logger.error(f"Vocabulary path ({save_directory}) should be a directory")
+            return
+        vocab_file = os.path.join(
+            save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["vocab_file"]
+        )
+        merge_file = os.path.join(
+            save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["merges_file"]
+        )
+
+        with open(vocab_file, "w", encoding="utf-8") as f:
+            f.write(json.dumps(self.encoder, indent=2, sort_keys=True, ensure_ascii=False) + "\n")
+
+        index = 0
+        with open(merge_file, "w", encoding="utf-8") as writer:
+            writer.write("#version: 0.2\n")
+            for bpe_tokens, token_index in sorted(self.bpe_ranks.items(), key=lambda kv: kv[1]):
+                if index != token_index:
+                    logger.warning(
+                        f"Saving vocabulary to {merge_file}: BPE merge indices are not consecutive."
+                        " Please check that the tokenizer is not corrupted!"
+                    )
+                    index = token_index
+                writer.write(" ".join(bpe_tokens) + "\n")
+                index += 1
+
+        return vocab_file, merge_file
+
+    def prepare_for_tokenization(self, text, **kwargs):
+        text = unicodedata.normalize("NFC", text)
+        return (text, kwargs)

modeling/qwen2/tokenization_qwen2_fast.py

@@ -1,123 +1,123 @@
-# Copyright 2024 The Qwen Team and The HuggingFace Inc. team.
-# SPDX-License-Identifier: Apache-2.0
-
-"""Tokenization classes for Qwen2."""
-
-from typing import Optional, Tuple
-
-from transformers.tokenization_utils import AddedToken
-from transformers.tokenization_utils_fast import PreTrainedTokenizerFast
-from transformers.utils import logging
-from .tokenization_qwen2 import Qwen2Tokenizer
-
-
-logger = logging.get_logger(__name__)
-
-VOCAB_FILES_NAMES = {
-    "vocab_file": "vocab.json",
-    "merges_file": "merges.txt",
-    "tokenizer_file": "tokenizer.json",
-}
-
-
-MAX_MODEL_INPUT_SIZES = {"qwen/qwen-tokenizer": 32768}
-
-
-class Qwen2TokenizerFast(PreTrainedTokenizerFast):
-    """
-    Construct a "fast" Qwen2 tokenizer (backed by HuggingFace's *tokenizers* library). Based on byte-level
-    Byte-Pair-Encoding.
-
-    Same with GPT2Tokenizer, this tokenizer has been trained to treat spaces like parts of the tokens so a word will
-    be encoded differently whether it is at the beginning of the sentence (without space) or not:
-
-    ```python
-    >>> from transformers import Qwen2TokenizerFast
-
-    >>> tokenizer = Qwen2TokenizerFast.from_pretrained("Qwen/Qwen-tokenizer")
-    >>> tokenizer("Hello world")["input_ids"]
-    [9707, 1879]
-
-    >>> tokenizer(" Hello world")["input_ids"]
-    [21927, 1879]
-    ```
-    This is expected.
-
-    This tokenizer inherits from [`PreTrainedTokenizerFast`] which contains most of the main methods. Users should
-    refer to this superclass for more information regarding those methods.
-
-    Args:
-        vocab_file (`str`, *optional*):
-            Path to the vocabulary file.
-        merges_file (`str`, *optional*):
-            Path to the merges file.
-        tokenizer_file (`str`, *optional*):
-            Path to [tokenizers](https://github.com/huggingface/tokenizers) file (generally has a .json extension) that
-            contains everything needed to load the tokenizer.
-        unk_token (`str`, *optional*, defaults to `"<|endoftext|>"`):
-            The unknown token. A token that is not in the vocabulary cannot be converted to an ID and is set to be this
-            token instead. Not applicable to this tokenizer.
-        bos_token (`str`, *optional*):
-            The beginning of sequence token. Not applicable for this tokenizer.
-        eos_token (`str`, *optional*, defaults to `"<|endoftext|>"`):
-            The end of sequence token.
-        pad_token (`str`, *optional*, defaults to `"<|endoftext|>"`):
-            The token used for padding, for example when batching sequences of different lengths.
-    """
-
-    vocab_files_names = VOCAB_FILES_NAMES
-    model_input_names = ["input_ids", "attention_mask"]
-    slow_tokenizer_class = Qwen2Tokenizer
-
-    def __init__(
-        self,
-        vocab_file=None,
-        merges_file=None,
-        tokenizer_file=None,
-        unk_token="<|endoftext|>",
-        bos_token=None,
-        eos_token="<|endoftext|>",
-        pad_token="<|endoftext|>",
-        **kwargs,
-    ):
-        # We need to at least pass vocab_file and merges_file to base class
-        # in case a slow tokenizer needs to be initialized; other can be
-        # configured through files.
-        # following GPT2TokenizerFast, also adding unk_token, bos_token, and eos_token
-
-        bos_token = (
-            AddedToken(bos_token, lstrip=False, rstrip=False, special=True, normalized=False)
-            if isinstance(bos_token, str)
-            else bos_token
-        )
-        eos_token = (
-            AddedToken(eos_token, lstrip=False, rstrip=False, special=True, normalized=False)
-            if isinstance(eos_token, str)
-            else eos_token
-        )
-        unk_token = (
-            AddedToken(unk_token, lstrip=False, rstrip=False, special=True, normalized=False)
-            if isinstance(unk_token, str)
-            else unk_token
-        )
-        pad_token = (
-            AddedToken(pad_token, lstrip=False, rstrip=False, special=True, normalized=False)
-            if isinstance(pad_token, str)
-            else pad_token
-        )
-
-        super().__init__(
-            vocab_file=vocab_file,
-            merges_file=merges_file,
-            tokenizer_file=tokenizer_file,
-            unk_token=unk_token,
-            bos_token=bos_token,
-            eos_token=eos_token,
-            pad_token=pad_token,
-            **kwargs,
-        )
-
-    # Copied from transformers.models.gpt2.tokenization_gpt2_fast.GPT2TokenizerFast.save_vocabulary
-    def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> Tuple[str]:
-        files = self._tokenizer.model.save(save_directory, name=filename_prefix)
-        return tuple(files)
+# Copyright 2024 The Qwen Team and The HuggingFace Inc. team.
+# SPDX-License-Identifier: Apache-2.0
+
+"""Tokenization classes for Qwen2."""
+
+from typing import Optional, Tuple
+
+from transformers.tokenization_utils import AddedToken
+from transformers.tokenization_utils_fast import PreTrainedTokenizerFast
+from transformers.utils import logging
+from .tokenization_qwen2 import Qwen2Tokenizer
+
+
+logger = logging.get_logger(__name__)
+
+VOCAB_FILES_NAMES = {
+    "vocab_file": "vocab.json",
+    "merges_file": "merges.txt",
+    "tokenizer_file": "tokenizer.json",
+}
+
+
+MAX_MODEL_INPUT_SIZES = {"qwen/qwen-tokenizer": 32768}
+
+
+class Qwen2TokenizerFast(PreTrainedTokenizerFast):
+    """
+    Construct a "fast" Qwen2 tokenizer (backed by HuggingFace's *tokenizers* library). Based on byte-level
+    Byte-Pair-Encoding.
+
+    Same with GPT2Tokenizer, this tokenizer has been trained to treat spaces like parts of the tokens so a word will
+    be encoded differently whether it is at the beginning of the sentence (without space) or not:
+
+    ```python
+    >>> from transformers import Qwen2TokenizerFast
+
+    >>> tokenizer = Qwen2TokenizerFast.from_pretrained("Qwen/Qwen-tokenizer")
+    >>> tokenizer("Hello world")["input_ids"]
+    [9707, 1879]
+
+    >>> tokenizer(" Hello world")["input_ids"]
+    [21927, 1879]
+    ```
+    This is expected.
+
+    This tokenizer inherits from [`PreTrainedTokenizerFast`] which contains most of the main methods. Users should
+    refer to this superclass for more information regarding those methods.
+
+    Args:
+        vocab_file (`str`, *optional*):
+            Path to the vocabulary file.
+        merges_file (`str`, *optional*):
+            Path to the merges file.
+        tokenizer_file (`str`, *optional*):
+            Path to [tokenizers](https://github.com/huggingface/tokenizers) file (generally has a .json extension) that
+            contains everything needed to load the tokenizer.
+        unk_token (`str`, *optional*, defaults to `"<|endoftext|>"`):
+            The unknown token. A token that is not in the vocabulary cannot be converted to an ID and is set to be this
+            token instead. Not applicable to this tokenizer.
+        bos_token (`str`, *optional*):
+            The beginning of sequence token. Not applicable for this tokenizer.
+        eos_token (`str`, *optional*, defaults to `"<|endoftext|>"`):
+            The end of sequence token.
+        pad_token (`str`, *optional*, defaults to `"<|endoftext|>"`):
+            The token used for padding, for example when batching sequences of different lengths.
+    """
+
+    vocab_files_names = VOCAB_FILES_NAMES
+    model_input_names = ["input_ids", "attention_mask"]
+    slow_tokenizer_class = Qwen2Tokenizer
+
+    def __init__(
+        self,
+        vocab_file=None,
+        merges_file=None,
+        tokenizer_file=None,
+        unk_token="<|endoftext|>",
+        bos_token=None,
+        eos_token="<|endoftext|>",
+        pad_token="<|endoftext|>",
+        **kwargs,
+    ):
+        # We need to at least pass vocab_file and merges_file to base class
+        # in case a slow tokenizer needs to be initialized; other can be
+        # configured through files.
+        # following GPT2TokenizerFast, also adding unk_token, bos_token, and eos_token
+
+        bos_token = (
+            AddedToken(bos_token, lstrip=False, rstrip=False, special=True, normalized=False)
+            if isinstance(bos_token, str)
+            else bos_token
+        )
+        eos_token = (
+            AddedToken(eos_token, lstrip=False, rstrip=False, special=True, normalized=False)
+            if isinstance(eos_token, str)
+            else eos_token
+        )
+        unk_token = (
+            AddedToken(unk_token, lstrip=False, rstrip=False, special=True, normalized=False)
+            if isinstance(unk_token, str)
+            else unk_token
+        )
+        pad_token = (
+            AddedToken(pad_token, lstrip=False, rstrip=False, special=True, normalized=False)
+            if isinstance(pad_token, str)
+            else pad_token
+        )
+
+        super().__init__(
+            vocab_file=vocab_file,
+            merges_file=merges_file,
+            tokenizer_file=tokenizer_file,
+            unk_token=unk_token,
+            bos_token=bos_token,
+            eos_token=eos_token,
+            pad_token=pad_token,
+            **kwargs,
+        )
+
+    # Copied from transformers.models.gpt2.tokenization_gpt2_fast.GPT2TokenizerFast.save_vocabulary
+    def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> Tuple[str]:
+        files = self._tokenizer.model.save(save_directory, name=filename_prefix)
+        return tuple(files)

modeling/siglip/__init__.py

@@ -1,98 +1,98 @@
-# Copyright 2024 The HuggingFace Inc. team.
-# SPDX-License-Identifier: Apache-2.0
-
-from typing import TYPE_CHECKING
-
-from transformers.utils import (
-    OptionalDependencyNotAvailable,
-    _LazyModule,
-    is_sentencepiece_available,
-    is_torch_available,
-    is_vision_available,
-)
-
-
-_import_structure = {
-    "configuration_siglip": [
-        "SiglipConfig",
-        "SiglipTextConfig",
-        "SiglipVisionConfig",
-    ],
-    "processing_siglip": ["SiglipProcessor"],
-}
-
-try:
-    if not is_sentencepiece_available():
-        raise OptionalDependencyNotAvailable()
-except OptionalDependencyNotAvailable:
-    pass
-else:
-    _import_structure["tokenization_siglip"] = ["SiglipTokenizer"]
-
-
-try:
-    if not is_vision_available():
-        raise OptionalDependencyNotAvailable()
-except OptionalDependencyNotAvailable:
-    pass
-else:
-    _import_structure["image_processing_siglip"] = ["SiglipImageProcessor"]
-
-try:
-    if not is_torch_available():
-        raise OptionalDependencyNotAvailable()
-except OptionalDependencyNotAvailable:
-    pass
-else:
-    _import_structure["modeling_siglip"] = [
-        "SiglipModel",
-        "SiglipPreTrainedModel",
-        "SiglipTextModel",
-        "SiglipVisionModel",
-        "SiglipForImageClassification",
-    ]
-
-
-if TYPE_CHECKING:
-    from .configuration_siglip import (
-        SiglipConfig,
-        SiglipTextConfig,
-        SiglipVisionConfig,
-    )
-    from .processing_siglip import SiglipProcessor
-
-    try:
-        if not is_sentencepiece_available():
-            raise OptionalDependencyNotAvailable()
-    except OptionalDependencyNotAvailable:
-        pass
-    else:
-        from .tokenization_siglip import SiglipTokenizer
-
-    try:
-        if not is_vision_available():
-            raise OptionalDependencyNotAvailable()
-    except OptionalDependencyNotAvailable:
-        pass
-    else:
-        from .image_processing_siglip import SiglipImageProcessor
-
-    try:
-        if not is_torch_available():
-            raise OptionalDependencyNotAvailable()
-    except OptionalDependencyNotAvailable:
-        pass
-    else:
-        from .modeling_siglip import (
-            SiglipForImageClassification,
-            SiglipModel,
-            SiglipPreTrainedModel,
-            SiglipTextModel,
-            SiglipVisionModel,
-        )
-
-
-else:
-    import sys
-
-    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)
+# Copyright 2024 The HuggingFace Inc. team.
+# SPDX-License-Identifier: Apache-2.0
+
+from typing import TYPE_CHECKING
+
+from transformers.utils import (
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    is_sentencepiece_available,
+    is_torch_available,
+    is_vision_available,
+)
+
+
+_import_structure = {
+    "configuration_siglip": [
+        "SiglipConfig",
+        "SiglipTextConfig",
+        "SiglipVisionConfig",
+    ],
+    "processing_siglip": ["SiglipProcessor"],
+}
+
+try:
+    if not is_sentencepiece_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["tokenization_siglip"] = ["SiglipTokenizer"]
+
+
+try:
+    if not is_vision_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["image_processing_siglip"] = ["SiglipImageProcessor"]
+
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["modeling_siglip"] = [
+        "SiglipModel",
+        "SiglipPreTrainedModel",
+        "SiglipTextModel",
+        "SiglipVisionModel",
+        "SiglipForImageClassification",
+    ]
+
+
+if TYPE_CHECKING:
+    from .configuration_siglip import (
+        SiglipConfig,
+        SiglipTextConfig,
+        SiglipVisionConfig,
+    )
+    from .processing_siglip import SiglipProcessor
+
+    try:
+        if not is_sentencepiece_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .tokenization_siglip import SiglipTokenizer
+
+    try:
+        if not is_vision_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .image_processing_siglip import SiglipImageProcessor
+
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .modeling_siglip import (
+            SiglipForImageClassification,
+            SiglipModel,
+            SiglipPreTrainedModel,
+            SiglipTextModel,
+            SiglipVisionModel,
+        )
+
+
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)

modeling/siglip/configuration_siglip.py

@@ -1,287 +1,287 @@
-# Copyright 2024 The HuggingFace Inc. team.
-# SPDX-License-Identifier: Apache-2.0
-
-"""Siglip model configuration"""
-
-import os
-from typing import Union
-
-from transformers.configuration_utils import PretrainedConfig
-from transformers.utils import logging
-
-
-logger = logging.get_logger(__name__)
-
-
-class SiglipTextConfig(PretrainedConfig):
-    r"""
-    This is the configuration class to store the configuration of a [`SiglipTextModel`]. It is used to instantiate a
-    Siglip text encoder according to the specified arguments, defining the model architecture. Instantiating a
-    configuration with the defaults will yield a similar configuration to that of the text encoder of the Siglip
-    [google/siglip-base-patch16-224](https://huggingface.co/google/siglip-base-patch16-224) architecture.
-
-    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
-    documentation from [`PretrainedConfig`] for more information.
-
-    Args:
-        vocab_size (`int`, *optional*, defaults to 32000):
-            Vocabulary size of the Siglip text model. Defines the number of different tokens that can be represented by
-            the `inputs_ids` passed when calling [`SiglipModel`].
-        hidden_size (`int`, *optional*, defaults to 768):
-            Dimensionality of the encoder layers and the pooler layer.
-        intermediate_size (`int`, *optional*, defaults to 3072):
-            Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
-        num_hidden_layers (`int`, *optional*, defaults to 12):
-            Number of hidden layers in the Transformer encoder.
-        num_attention_heads (`int`, *optional*, defaults to 12):
-            Number of attention heads for each attention layer in the Transformer encoder.
-        max_position_embeddings (`int`, *optional*, defaults to 64):
-            The maximum sequence length that this model might ever be used with. Typically set this to something large
-            just in case (e.g., 512 or 1024 or 2048).
-        hidden_act (`str` or `function`, *optional*, defaults to `"gelu_pytorch_tanh"`):
-            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
-            `"relu"`, `"selu"` and `"gelu_new"` `"quick_gelu"` are supported.
-        layer_norm_eps (`float`, *optional*, defaults to 1e-06):
-            The epsilon used by the layer normalization layers.
-        attention_dropout (`float`, *optional*, defaults to 0.0):
-            The dropout ratio for the attention probabilities.
-        pad_token_id (`int`, *optional*, defaults to 1):
-            The id of the padding token in the vocabulary.
-        bos_token_id (`int`, *optional*, defaults to 49406):
-            The id of the beginning-of-sequence token in the vocabulary.
-        eos_token_id (`int`, *optional*, defaults to 49407):
-            The id of the end-of-sequence token in the vocabulary.
-
-    Example:
-
-    ```python
-    >>> from transformers import SiglipTextConfig, SiglipTextModel
-
-    >>> # Initializing a SiglipTextConfig with google/siglip-base-patch16-224 style configuration
-    >>> configuration = SiglipTextConfig()
-
-    >>> # Initializing a SiglipTextModel (with random weights) from the google/siglip-base-patch16-224 style configuration
-    >>> model = SiglipTextModel(configuration)
-
-    >>> # Accessing the model configuration
-    >>> configuration = model.config
-    ```"""
-
-    model_type = "siglip_text_model"
-
-    def __init__(
-        self,
-        vocab_size=32000,
-        hidden_size=768,
-        intermediate_size=3072,
-        num_hidden_layers=12,
-        num_attention_heads=12,
-        max_position_embeddings=64,
-        hidden_act="gelu_pytorch_tanh",
-        layer_norm_eps=1e-6,
-        attention_dropout=0.0,
-        # This differs from `CLIPTokenizer`'s default and from openai/siglip
-        # See https://github.com/huggingface/transformers/pull/24773#issuecomment-1632287538
-        pad_token_id=1,
-        bos_token_id=49406,
-        eos_token_id=49407,
-        **kwargs,
-    ):
-        super().__init__(pad_token_id=pad_token_id, bos_token_id=bos_token_id, eos_token_id=eos_token_id, **kwargs)
-
-        self.vocab_size = vocab_size
-        self.hidden_size = hidden_size
-        self.intermediate_size = intermediate_size
-        self.num_hidden_layers = num_hidden_layers
-        self.num_attention_heads = num_attention_heads
-        self.max_position_embeddings = max_position_embeddings
-        self.layer_norm_eps = layer_norm_eps
-        self.hidden_act = hidden_act
-        self.attention_dropout = attention_dropout
-
-    @classmethod
-    def from_pretrained(cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs) -> "PretrainedConfig":
-        cls._set_token_in_kwargs(kwargs)
-
-        config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path, **kwargs)
-
-        # get the text config dict if we are loading from SiglipConfig
-        if config_dict.get("model_type") == "siglip":
-            config_dict = config_dict["text_config"]
-
-        if "model_type" in config_dict and hasattr(cls, "model_type") and config_dict["model_type"] != cls.model_type:
-            logger.warning(
-                f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "
-                f"{cls.model_type}. This is not supported for all configurations of models and can yield errors."
-            )
-
-        return cls.from_dict(config_dict, **kwargs)
-
-
-class SiglipVisionConfig(PretrainedConfig):
-    r"""
-    This is the configuration class to store the configuration of a [`SiglipVisionModel`]. It is used to instantiate a
-    Siglip vision encoder according to the specified arguments, defining the model architecture. Instantiating a
-    configuration with the defaults will yield a similar configuration to that of the vision encoder of the Siglip
-    [google/siglip-base-patch16-224](https://huggingface.co/google/siglip-base-patch16-224) architecture.
-
-    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
-    documentation from [`PretrainedConfig`] for more information.
-
-    Args:
-        hidden_size (`int`, *optional*, defaults to 768):
-            Dimensionality of the encoder layers and the pooler layer.
-        intermediate_size (`int`, *optional*, defaults to 3072):
-            Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
-        num_hidden_layers (`int`, *optional*, defaults to 12):
-            Number of hidden layers in the Transformer encoder.
-        num_attention_heads (`int`, *optional*, defaults to 12):
-            Number of attention heads for each attention layer in the Transformer encoder.
-        num_channels (`int`, *optional*, defaults to 3):
-            Number of channels in the input images.
-        image_size (`int`, *optional*, defaults to 224):
-            The size (resolution) of each image.
-        patch_size (`int`, *optional*, defaults to 16):
-            The size (resolution) of each patch.
-        hidden_act (`str` or `function`, *optional*, defaults to `"gelu_pytorch_tanh"`):
-            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
-            `"relu"`, `"selu"` and `"gelu_new"` `"quick_gelu"` are supported.
-        layer_norm_eps (`float`, *optional*, defaults to 1e-06):
-            The epsilon used by the layer normalization layers.
-        attention_dropout (`float`, *optional*, defaults to 0.0):
-            The dropout ratio for the attention probabilities.
-
-    Example:
-
-    ```python
-    >>> from transformers import SiglipVisionConfig, SiglipVisionModel
-
-    >>> # Initializing a SiglipVisionConfig with google/siglip-base-patch16-224 style configuration
-    >>> configuration = SiglipVisionConfig()
-
-    >>> # Initializing a SiglipVisionModel (with random weights) from the google/siglip-base-patch16-224 style configuration
-    >>> model = SiglipVisionModel(configuration)
-
-    >>> # Accessing the model configuration
-    >>> configuration = model.config
-    ```"""
-
-    model_type = "siglip_vision_model"
-
-    def __init__(
-        self,
-        hidden_size=768,
-        intermediate_size=3072,
-        num_hidden_layers=12,
-        num_attention_heads=12,
-        num_channels=3,
-        image_size=224,
-        patch_size=16,
-        hidden_act="gelu_pytorch_tanh",
-        layer_norm_eps=1e-6,
-        attention_dropout=0.0,
-        **kwargs,
-    ):
-        super().__init__(**kwargs)
-
-        self.hidden_size = hidden_size
-        self.intermediate_size = intermediate_size
-        self.num_hidden_layers = num_hidden_layers
-        self.num_attention_heads = num_attention_heads
-        self.num_channels = num_channels
-        self.patch_size = patch_size
-        self.image_size = image_size
-        self.attention_dropout = attention_dropout
-        self.layer_norm_eps = layer_norm_eps
-        self.hidden_act = hidden_act
-
-    @classmethod
-    def from_pretrained(cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs) -> "PretrainedConfig":
-        cls._set_token_in_kwargs(kwargs)
-
-        config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path, **kwargs)
-
-        # get the vision config dict if we are loading from SiglipConfig
-        if config_dict.get("model_type") == "siglip":
-            config_dict = config_dict["vision_config"]
-
-        if "model_type" in config_dict and hasattr(cls, "model_type") and config_dict["model_type"] != cls.model_type:
-            logger.warning(
-                f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "
-                f"{cls.model_type}. This is not supported for all configurations of models and can yield errors."
-            )
-
-        return cls.from_dict(config_dict, **kwargs)
-
-
-class SiglipConfig(PretrainedConfig):
-    r"""
-    [`SiglipConfig`] is the configuration class to store the configuration of a [`SiglipModel`]. It is used to
-    instantiate a Siglip model according to the specified arguments, defining the text model and vision model configs.
-    Instantiating a configuration with the defaults will yield a similar configuration to that of the Siglip
-    [google/siglip-base-patch16-224](https://huggingface.co/google/siglip-base-patch16-224) architecture.
-
-    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
-    documentation from [`PretrainedConfig`] for more information.
-
-    Args:
-        text_config (`dict`, *optional*):
-            Dictionary of configuration options used to initialize [`SiglipTextConfig`].
-        vision_config (`dict`, *optional*):
-            Dictionary of configuration options used to initialize [`SiglipVisionConfig`].
-        kwargs (*optional*):
-            Dictionary of keyword arguments.
-
-    Example:
-
-    ```python
-    >>> from transformers import SiglipConfig, SiglipModel
-
-    >>> # Initializing a SiglipConfig with google/siglip-base-patch16-224 style configuration
-    >>> configuration = SiglipConfig()
-
-    >>> # Initializing a SiglipModel (with random weights) from the google/siglip-base-patch16-224 style configuration
-    >>> model = SiglipModel(configuration)
-
-    >>> # Accessing the model configuration
-    >>> configuration = model.config
-
-    >>> # We can also initialize a SiglipConfig from a SiglipTextConfig and a SiglipVisionConfig
-    >>> from transformers import SiglipTextConfig, SiglipVisionConfig
-
-    >>> # Initializing a SiglipText and SiglipVision configuration
-    >>> config_text = SiglipTextConfig()
-    >>> config_vision = SiglipVisionConfig()
-
-    >>> config = SiglipConfig.from_text_vision_configs(config_text, config_vision)
-    ```"""
-
-    model_type = "siglip"
-
-    def __init__(self, text_config=None, vision_config=None, **kwargs):
-        super().__init__(**kwargs)
-
-        if text_config is None:
-            text_config = {}
-            logger.info("`text_config` is `None`. Initializing the `SiglipTextConfig` with default values.")
-
-        if vision_config is None:
-            vision_config = {}
-            logger.info("`vision_config` is `None`. initializing the `SiglipVisionConfig` with default values.")
-
-        self.text_config = SiglipTextConfig(**text_config)
-        self.vision_config = SiglipVisionConfig(**vision_config)
-
-        self.initializer_factor = 1.0
-
-    @classmethod
-    def from_text_vision_configs(cls, text_config: SiglipTextConfig, vision_config: SiglipVisionConfig, **kwargs):
-        r"""
-        Instantiate a [`SiglipConfig`] (or a derived class) from siglip text model configuration and siglip vision
-        model configuration.
-
-        Returns:
-            [`SiglipConfig`]: An instance of a configuration object
-        """
-
-        return cls(text_config=text_config.to_dict(), vision_config=vision_config.to_dict(), **kwargs)
+# Copyright 2024 The HuggingFace Inc. team.
+# SPDX-License-Identifier: Apache-2.0
+
+"""Siglip model configuration"""
+
+import os
+from typing import Union
+
+from transformers.configuration_utils import PretrainedConfig
+from transformers.utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+
+class SiglipTextConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`SiglipTextModel`]. It is used to instantiate a
+    Siglip text encoder according to the specified arguments, defining the model architecture. Instantiating a
+    configuration with the defaults will yield a similar configuration to that of the text encoder of the Siglip
+    [google/siglip-base-patch16-224](https://huggingface.co/google/siglip-base-patch16-224) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 32000):
+            Vocabulary size of the Siglip text model. Defines the number of different tokens that can be represented by
+            the `inputs_ids` passed when calling [`SiglipModel`].
+        hidden_size (`int`, *optional*, defaults to 768):
+            Dimensionality of the encoder layers and the pooler layer.
+        intermediate_size (`int`, *optional*, defaults to 3072):
+            Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
+        num_hidden_layers (`int`, *optional*, defaults to 12):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 12):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        max_position_embeddings (`int`, *optional*, defaults to 64):
+            The maximum sequence length that this model might ever be used with. Typically set this to something large
+            just in case (e.g., 512 or 1024 or 2048).
+        hidden_act (`str` or `function`, *optional*, defaults to `"gelu_pytorch_tanh"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
+            `"relu"`, `"selu"` and `"gelu_new"` `"quick_gelu"` are supported.
+        layer_norm_eps (`float`, *optional*, defaults to 1e-06):
+            The epsilon used by the layer normalization layers.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        pad_token_id (`int`, *optional*, defaults to 1):
+            The id of the padding token in the vocabulary.
+        bos_token_id (`int`, *optional*, defaults to 49406):
+            The id of the beginning-of-sequence token in the vocabulary.
+        eos_token_id (`int`, *optional*, defaults to 49407):
+            The id of the end-of-sequence token in the vocabulary.
+
+    Example:
+
+    ```python
+    >>> from transformers import SiglipTextConfig, SiglipTextModel
+
+    >>> # Initializing a SiglipTextConfig with google/siglip-base-patch16-224 style configuration
+    >>> configuration = SiglipTextConfig()
+
+    >>> # Initializing a SiglipTextModel (with random weights) from the google/siglip-base-patch16-224 style configuration
+    >>> model = SiglipTextModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "siglip_text_model"
+
+    def __init__(
+        self,
+        vocab_size=32000,
+        hidden_size=768,
+        intermediate_size=3072,
+        num_hidden_layers=12,
+        num_attention_heads=12,
+        max_position_embeddings=64,
+        hidden_act="gelu_pytorch_tanh",
+        layer_norm_eps=1e-6,
+        attention_dropout=0.0,
+        # This differs from `CLIPTokenizer`'s default and from openai/siglip
+        # See https://github.com/huggingface/transformers/pull/24773#issuecomment-1632287538
+        pad_token_id=1,
+        bos_token_id=49406,
+        eos_token_id=49407,
+        **kwargs,
+    ):
+        super().__init__(pad_token_id=pad_token_id, bos_token_id=bos_token_id, eos_token_id=eos_token_id, **kwargs)
+
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.max_position_embeddings = max_position_embeddings
+        self.layer_norm_eps = layer_norm_eps
+        self.hidden_act = hidden_act
+        self.attention_dropout = attention_dropout
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs) -> "PretrainedConfig":
+        cls._set_token_in_kwargs(kwargs)
+
+        config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path, **kwargs)
+
+        # get the text config dict if we are loading from SiglipConfig
+        if config_dict.get("model_type") == "siglip":
+            config_dict = config_dict["text_config"]
+
+        if "model_type" in config_dict and hasattr(cls, "model_type") and config_dict["model_type"] != cls.model_type:
+            logger.warning(
+                f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "
+                f"{cls.model_type}. This is not supported for all configurations of models and can yield errors."
+            )
+
+        return cls.from_dict(config_dict, **kwargs)
+
+
+class SiglipVisionConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`SiglipVisionModel`]. It is used to instantiate a
+    Siglip vision encoder according to the specified arguments, defining the model architecture. Instantiating a
+    configuration with the defaults will yield a similar configuration to that of the vision encoder of the Siglip
+    [google/siglip-base-patch16-224](https://huggingface.co/google/siglip-base-patch16-224) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        hidden_size (`int`, *optional*, defaults to 768):
+            Dimensionality of the encoder layers and the pooler layer.
+        intermediate_size (`int`, *optional*, defaults to 3072):
+            Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
+        num_hidden_layers (`int`, *optional*, defaults to 12):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 12):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        num_channels (`int`, *optional*, defaults to 3):
+            Number of channels in the input images.
+        image_size (`int`, *optional*, defaults to 224):
+            The size (resolution) of each image.
+        patch_size (`int`, *optional*, defaults to 16):
+            The size (resolution) of each patch.
+        hidden_act (`str` or `function`, *optional*, defaults to `"gelu_pytorch_tanh"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
+            `"relu"`, `"selu"` and `"gelu_new"` `"quick_gelu"` are supported.
+        layer_norm_eps (`float`, *optional*, defaults to 1e-06):
+            The epsilon used by the layer normalization layers.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+
+    Example:
+
+    ```python
+    >>> from transformers import SiglipVisionConfig, SiglipVisionModel
+
+    >>> # Initializing a SiglipVisionConfig with google/siglip-base-patch16-224 style configuration
+    >>> configuration = SiglipVisionConfig()
+
+    >>> # Initializing a SiglipVisionModel (with random weights) from the google/siglip-base-patch16-224 style configuration
+    >>> model = SiglipVisionModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "siglip_vision_model"
+
+    def __init__(
+        self,
+        hidden_size=768,
+        intermediate_size=3072,
+        num_hidden_layers=12,
+        num_attention_heads=12,
+        num_channels=3,
+        image_size=224,
+        patch_size=16,
+        hidden_act="gelu_pytorch_tanh",
+        layer_norm_eps=1e-6,
+        attention_dropout=0.0,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.num_channels = num_channels
+        self.patch_size = patch_size
+        self.image_size = image_size
+        self.attention_dropout = attention_dropout
+        self.layer_norm_eps = layer_norm_eps
+        self.hidden_act = hidden_act
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs) -> "PretrainedConfig":
+        cls._set_token_in_kwargs(kwargs)
+
+        config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path, **kwargs)
+
+        # get the vision config dict if we are loading from SiglipConfig
+        if config_dict.get("model_type") == "siglip":
+            config_dict = config_dict["vision_config"]
+
+        if "model_type" in config_dict and hasattr(cls, "model_type") and config_dict["model_type"] != cls.model_type:
+            logger.warning(
+                f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "
+                f"{cls.model_type}. This is not supported for all configurations of models and can yield errors."
+            )
+
+        return cls.from_dict(config_dict, **kwargs)
+
+
+class SiglipConfig(PretrainedConfig):
+    r"""
+    [`SiglipConfig`] is the configuration class to store the configuration of a [`SiglipModel`]. It is used to
+    instantiate a Siglip model according to the specified arguments, defining the text model and vision model configs.
+    Instantiating a configuration with the defaults will yield a similar configuration to that of the Siglip
+    [google/siglip-base-patch16-224](https://huggingface.co/google/siglip-base-patch16-224) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        text_config (`dict`, *optional*):
+            Dictionary of configuration options used to initialize [`SiglipTextConfig`].
+        vision_config (`dict`, *optional*):
+            Dictionary of configuration options used to initialize [`SiglipVisionConfig`].
+        kwargs (*optional*):
+            Dictionary of keyword arguments.
+
+    Example:
+
+    ```python
+    >>> from transformers import SiglipConfig, SiglipModel
+
+    >>> # Initializing a SiglipConfig with google/siglip-base-patch16-224 style configuration
+    >>> configuration = SiglipConfig()
+
+    >>> # Initializing a SiglipModel (with random weights) from the google/siglip-base-patch16-224 style configuration
+    >>> model = SiglipModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+
+    >>> # We can also initialize a SiglipConfig from a SiglipTextConfig and a SiglipVisionConfig
+    >>> from transformers import SiglipTextConfig, SiglipVisionConfig
+
+    >>> # Initializing a SiglipText and SiglipVision configuration
+    >>> config_text = SiglipTextConfig()
+    >>> config_vision = SiglipVisionConfig()
+
+    >>> config = SiglipConfig.from_text_vision_configs(config_text, config_vision)
+    ```"""
+
+    model_type = "siglip"
+
+    def __init__(self, text_config=None, vision_config=None, **kwargs):
+        super().__init__(**kwargs)
+
+        if text_config is None:
+            text_config = {}
+            logger.info("`text_config` is `None`. Initializing the `SiglipTextConfig` with default values.")
+
+        if vision_config is None:
+            vision_config = {}
+            logger.info("`vision_config` is `None`. initializing the `SiglipVisionConfig` with default values.")
+
+        self.text_config = SiglipTextConfig(**text_config)
+        self.vision_config = SiglipVisionConfig(**vision_config)
+
+        self.initializer_factor = 1.0
+
+    @classmethod
+    def from_text_vision_configs(cls, text_config: SiglipTextConfig, vision_config: SiglipVisionConfig, **kwargs):
+        r"""
+        Instantiate a [`SiglipConfig`] (or a derived class) from siglip text model configuration and siglip vision
+        model configuration.
+
+        Returns:
+            [`SiglipConfig`]: An instance of a configuration object
+        """
+
+        return cls(text_config=text_config.to_dict(), vision_config=vision_config.to_dict(), **kwargs)

modeling/siglip/convert_siglip_to_hf.py

@@ -1,401 +1,401 @@
-# Copyright 2024 The HuggingFace Inc. team.
-# SPDX-License-Identifier: Apache-2.0
-
-"""Convert SigLIP checkpoints from the original repository.
-
-URL: https://github.com/google-research/big_vision/tree/main
-"""
-
-import argparse
-import collections
-from pathlib import Path
-
-import numpy as np
-import requests
-import torch
-from huggingface_hub import hf_hub_download
-from numpy import load
-from PIL import Image
-
-from transformers import SiglipConfig, SiglipImageProcessor, SiglipModel, SiglipProcessor, SiglipTokenizer
-from transformers.utils import logging
-
-
-logging.set_verbosity_info()
-logger = logging.get_logger(__name__)
-
-
-model_name_to_checkpoint = {
-    # base checkpoints
-    "siglip-base-patch16-224": "/Users/nielsrogge/Documents/SigLIP/webli_en_b16_224_63724782.npz",
-    "siglip-base-patch16-256": "/Users/nielsrogge/Documents/SigLIP/webli_en_b16_256_60500360.npz",
-    "siglip-base-patch16-384": "/Users/nielsrogge/Documents/SigLIP/webli_en_b16_384_68578854.npz",
-    "siglip-base-patch16-512": "/Users/nielsrogge/Documents/SigLIP/webli_en_b16_512_68580893.npz",
-    # large checkpoints
-    "siglip-large-patch16-256": "/Users/nielsrogge/Documents/SigLIP/webli_en_l16_256_60552751.npz",
-    "siglip-large-patch16-384": "/Users/nielsrogge/Documents/SigLIP/webli_en_l16_384_63634585.npz",
-    # multilingual checkpoint
-    "siglip-base-patch16-256-i18n": "/Users/nielsrogge/Documents/SigLIP/webli_i18n_b16_256_66117334.npz",
-    # so400m checkpoints
-    "siglip-so400m-patch14-384": "/Users/nielsrogge/Documents/SigLIP/webli_en_so400m_384_58765454.npz",
-}
-
-model_name_to_image_size = {
-    "siglip-base-patch16-224": 224,
-    "siglip-base-patch16-256": 256,
-    "siglip-base-patch16-384": 384,
-    "siglip-base-patch16-512": 512,
-    "siglip-large-patch16-256": 256,
-    "siglip-large-patch16-384": 384,
-    "siglip-base-patch16-256-i18n": 256,
-    "siglip-so400m-patch14-384": 384,
-}
-
-
-def get_siglip_config(model_name):
-    config = SiglipConfig()
-
-    vocab_size = 250000 if "i18n" in model_name else 32000
-    image_size = model_name_to_image_size[model_name]
-    patch_size = 16 if "patch16" in model_name else 14
-
-    # size of the architecture
-    config.vision_config.image_size = image_size
-    config.vision_config.patch_size = patch_size
-    config.text_config.vocab_size = vocab_size
-
-    if "base" in model_name:
-        pass
-    elif "large" in model_name:
-        config.text_config.hidden_size = 1024
-        config.text_config.intermediate_size = 4096
-        config.text_config.num_hidden_layers = 24
-        config.text_config.num_attention_heads = 16
-        config.vision_config.hidden_size = 1024
-        config.vision_config.intermediate_size = 4096
-        config.vision_config.num_hidden_layers = 24
-        config.vision_config.num_attention_heads = 16
-    elif "so400m" in model_name:
-        config.text_config.hidden_size = 1152
-        config.text_config.intermediate_size = 4304
-        config.text_config.num_hidden_layers = 27
-        config.text_config.num_attention_heads = 16
-        config.vision_config.hidden_size = 1152
-        config.vision_config.intermediate_size = 4304
-        config.vision_config.num_hidden_layers = 27
-        config.vision_config.num_attention_heads = 16
-    else:
-        raise ValueError("Model not supported")
-
-    return config
-
-
-def create_rename_keys(config):
-    rename_keys = []
-    # fmt: off
-
-    # vision encoder
-
-    rename_keys.append(("params/img/embedding/kernel", "vision_model.embeddings.patch_embedding.weight"))
-    rename_keys.append(("params/img/embedding/bias", "vision_model.embeddings.patch_embedding.bias"))
-    rename_keys.append(("params/img/pos_embedding", "vision_model.embeddings.position_embedding.weight"))
-
-    for i in range(config.vision_config.num_hidden_layers):
-        rename_keys.append((f"params/img/Transformer/encoderblock_{i}/LayerNorm_0/scale", f"vision_model.encoder.layers.{i}.layer_norm1.weight"))
-        rename_keys.append((f"params/img/Transformer/encoderblock_{i}/LayerNorm_0/bias", f"vision_model.encoder.layers.{i}.layer_norm1.bias"))
-        rename_keys.append((f"params/img/Transformer/encoderblock_{i}/LayerNorm_1/scale", f"vision_model.encoder.layers.{i}.layer_norm2.weight"))
-        rename_keys.append((f"params/img/Transformer/encoderblock_{i}/LayerNorm_1/bias", f"vision_model.encoder.layers.{i}.layer_norm2.bias"))
-        rename_keys.append((f"params/img/Transformer/encoderblock_{i}/MlpBlock_0/Dense_0/kernel", f"vision_model.encoder.layers.{i}.mlp.fc1.weight"))
-        rename_keys.append((f"params/img/Transformer/encoderblock_{i}/MlpBlock_0/Dense_0/bias", f"vision_model.encoder.layers.{i}.mlp.fc1.bias"))
-        rename_keys.append((f"params/img/Transformer/encoderblock_{i}/MlpBlock_0/Dense_1/kernel", f"vision_model.encoder.layers.{i}.mlp.fc2.weight"))
-        rename_keys.append((f"params/img/Transformer/encoderblock_{i}/MlpBlock_0/Dense_1/bias", f"vision_model.encoder.layers.{i}.mlp.fc2.bias"))
-        rename_keys.append((f"params/img/Transformer/encoderblock_{i}/MultiHeadDotProductAttention_0/key/kernel", f"vision_model.encoder.layers.{i}.self_attn.k_proj.weight"))
-        rename_keys.append((f"params/img/Transformer/encoderblock_{i}/MultiHeadDotProductAttention_0/key/bias", f"vision_model.encoder.layers.{i}.self_attn.k_proj.bias"))
-        rename_keys.append((f"params/img/Transformer/encoderblock_{i}/MultiHeadDotProductAttention_0/value/kernel", f"vision_model.encoder.layers.{i}.self_attn.v_proj.weight"))
-        rename_keys.append((f"params/img/Transformer/encoderblock_{i}/MultiHeadDotProductAttention_0/value/bias", f"vision_model.encoder.layers.{i}.self_attn.v_proj.bias"))
-        rename_keys.append((f"params/img/Transformer/encoderblock_{i}/MultiHeadDotProductAttention_0/query/kernel", f"vision_model.encoder.layers.{i}.self_attn.q_proj.weight"))
-        rename_keys.append((f"params/img/Transformer/encoderblock_{i}/MultiHeadDotProductAttention_0/query/bias", f"vision_model.encoder.layers.{i}.self_attn.q_proj.bias"))
-        rename_keys.append((f"params/img/Transformer/encoderblock_{i}/MultiHeadDotProductAttention_0/out/kernel", f"vision_model.encoder.layers.{i}.self_attn.out_proj.weight"))
-        rename_keys.append((f"params/img/Transformer/encoderblock_{i}/MultiHeadDotProductAttention_0/out/bias", f"vision_model.encoder.layers.{i}.self_attn.out_proj.bias"))
-
-    rename_keys.append(("params/img/Transformer/encoder_norm/scale", "vision_model.post_layernorm.weight"))
-    rename_keys.append(("params/img/Transformer/encoder_norm/bias", "vision_model.post_layernorm.bias"))
-
-    rename_keys.append(("params/img/MAPHead_0/probe", "vision_model.head.probe"))
-    rename_keys.append(("params/img/MAPHead_0/LayerNorm_0/scale", "vision_model.head.layernorm.weight"))
-    rename_keys.append(("params/img/MAPHead_0/LayerNorm_0/bias", "vision_model.head.layernorm.bias"))
-    rename_keys.append(("params/img/MAPHead_0/MlpBlock_0/Dense_0/kernel", "vision_model.head.mlp.fc1.weight"))
-    rename_keys.append(("params/img/MAPHead_0/MlpBlock_0/Dense_0/bias", "vision_model.head.mlp.fc1.bias"))
-    rename_keys.append(("params/img/MAPHead_0/MlpBlock_0/Dense_1/kernel", "vision_model.head.mlp.fc2.weight"))
-    rename_keys.append(("params/img/MAPHead_0/MlpBlock_0/Dense_1/bias", "vision_model.head.mlp.fc2.bias"))
-    rename_keys.append(("params/img/MAPHead_0/MultiHeadDotProductAttention_0/out/kernel", "vision_model.head.attention.out_proj.weight"))
-    rename_keys.append(("params/img/MAPHead_0/MultiHeadDotProductAttention_0/out/bias", "vision_model.head.attention.out_proj.bias"))
-
-    # text encoder
-
-    rename_keys.append(("params/txt/Embed_0/embedding", "text_model.embeddings.token_embedding.weight"))
-    rename_keys.append(("params/txt/pos_embedding", "text_model.embeddings.position_embedding.weight"))
-
-    for i in range(config.text_config.num_hidden_layers):
-        rename_keys.append((f"params/txt/Encoder_0/encoderblock_{i}/LayerNorm_0/scale", f"text_model.encoder.layers.{i}.layer_norm1.weight"))
-        rename_keys.append((f"params/txt/Encoder_0/encoderblock_{i}/LayerNorm_0/bias", f"text_model.encoder.layers.{i}.layer_norm1.bias"))
-        rename_keys.append((f"params/txt/Encoder_0/encoderblock_{i}/LayerNorm_1/scale", f"text_model.encoder.layers.{i}.layer_norm2.weight"))
-        rename_keys.append((f"params/txt/Encoder_0/encoderblock_{i}/LayerNorm_1/bias", f"text_model.encoder.layers.{i}.layer_norm2.bias"))
-        rename_keys.append((f"params/txt/Encoder_0/encoderblock_{i}/MlpBlock_0/Dense_0/kernel", f"text_model.encoder.layers.{i}.mlp.fc1.weight"))
-        rename_keys.append((f"params/txt/Encoder_0/encoderblock_{i}/MlpBlock_0/Dense_0/bias", f"text_model.encoder.layers.{i}.mlp.fc1.bias"))
-        rename_keys.append((f"params/txt/Encoder_0/encoderblock_{i}/MlpBlock_0/Dense_1/kernel", f"text_model.encoder.layers.{i}.mlp.fc2.weight"))
-        rename_keys.append((f"params/txt/Encoder_0/encoderblock_{i}/MlpBlock_0/Dense_1/bias", f"text_model.encoder.layers.{i}.mlp.fc2.bias"))
-        rename_keys.append((f"params/txt/Encoder_0/encoderblock_{i}/MultiHeadDotProductAttention_0/key/kernel", f"text_model.encoder.layers.{i}.self_attn.k_proj.weight"))
-        rename_keys.append((f"params/txt/Encoder_0/encoderblock_{i}/MultiHeadDotProductAttention_0/key/bias", f"text_model.encoder.layers.{i}.self_attn.k_proj.bias"))
-        rename_keys.append((f"params/txt/Encoder_0/encoderblock_{i}/MultiHeadDotProductAttention_0/value/kernel", f"text_model.encoder.layers.{i}.self_attn.v_proj.weight"))
-        rename_keys.append((f"params/txt/Encoder_0/encoderblock_{i}/MultiHeadDotProductAttention_0/value/bias", f"text_model.encoder.layers.{i}.self_attn.v_proj.bias"))
-        rename_keys.append((f"params/txt/Encoder_0/encoderblock_{i}/MultiHeadDotProductAttention_0/query/kernel", f"text_model.encoder.layers.{i}.self_attn.q_proj.weight"))
-        rename_keys.append((f"params/txt/Encoder_0/encoderblock_{i}/MultiHeadDotProductAttention_0/query/bias", f"text_model.encoder.layers.{i}.self_attn.q_proj.bias"))
-        rename_keys.append((f"params/txt/Encoder_0/encoderblock_{i}/MultiHeadDotProductAttention_0/out/kernel", f"text_model.encoder.layers.{i}.self_attn.out_proj.weight"))
-        rename_keys.append((f"params/txt/Encoder_0/encoderblock_{i}/MultiHeadDotProductAttention_0/out/bias", f"text_model.encoder.layers.{i}.self_attn.out_proj.bias"))
-
-    rename_keys.append(("params/txt/Encoder_0/encoder_norm/scale", "text_model.final_layer_norm.weight"))
-    rename_keys.append(("params/txt/Encoder_0/encoder_norm/bias", "text_model.final_layer_norm.bias"))
-    rename_keys.append(("params/txt/head/kernel", "text_model.head.weight"))
-    rename_keys.append(("params/txt/head/bias", "text_model.head.bias"))
-
-    # learned temperature and bias
-    rename_keys.append(("params/t", "logit_scale"))
-    rename_keys.append(("params/b", "logit_bias"))
-
-    # fmt: on
-    return rename_keys
-
-
-def rename_key(dct, old, new, config):
-    val = dct.pop(old)
-
-    if ("out_proj" in new or "v_proj" in new or "k_proj" in new or "q_proj" in new) and "vision" in new:
-        val = val.reshape(-1, config.vision_config.hidden_size)
-    if ("out_proj" in new or "v_proj" in new or "k_proj" in new or "q_proj" in new) and "text" in new:
-        val = val.reshape(-1, config.text_config.hidden_size)
-
-    if "patch_embedding.weight" in new:
-        val = val.transpose(3, 2, 0, 1)
-    elif new.endswith("weight") and "position_embedding" not in new and "token_embedding" not in new:
-        val = val.T
-
-    if "position_embedding" in new and "vision" in new:
-        val = val.reshape(-1, config.vision_config.hidden_size)
-    if "position_embedding" in new and "text" in new:
-        val = val.reshape(-1, config.text_config.hidden_size)
-
-    if new.endswith("bias"):
-        val = val.reshape(-1)
-
-    dct[new] = torch.from_numpy(val)
-
-
-def read_in_q_k_v_head(state_dict, config):
-    # read in individual input projection layers
-    key_proj_weight = (
-        state_dict.pop("params/img/MAPHead_0/MultiHeadDotProductAttention_0/key/kernel")
-        .reshape(-1, config.vision_config.hidden_size)
-        .T
-    )
-    key_proj_bias = state_dict.pop("params/img/MAPHead_0/MultiHeadDotProductAttention_0/key/bias").reshape(-1)
-    value_proj_weight = (
-        state_dict.pop("params/img/MAPHead_0/MultiHeadDotProductAttention_0/value/kernel")
-        .reshape(-1, config.vision_config.hidden_size)
-        .T
-    )
-    value_proj_bias = state_dict.pop("params/img/MAPHead_0/MultiHeadDotProductAttention_0/value/bias").reshape(-1)
-    query_proj_weight = (
-        state_dict.pop("params/img/MAPHead_0/MultiHeadDotProductAttention_0/query/kernel")
-        .reshape(-1, config.vision_config.hidden_size)
-        .T
-    )
-    query_proj_bias = state_dict.pop("params/img/MAPHead_0/MultiHeadDotProductAttention_0/query/bias").reshape(-1)
-
-    # next, add them to the state dict as a single matrix + vector
-    state_dict["vision_model.head.attention.in_proj_weight"] = torch.from_numpy(
-        np.concatenate([query_proj_weight, key_proj_weight, value_proj_weight], axis=0)
-    )
-    state_dict["vision_model.head.attention.in_proj_bias"] = torch.from_numpy(
-        np.concatenate([query_proj_bias, key_proj_bias, value_proj_bias], axis=0)
-    )
-
-
-# We will verify our results on an image of cute cats
-def prepare_img():
-    url = "http://images.cocodataset.org/val2017/000000039769.jpg"
-    image = Image.open(requests.get(url, stream=True).raw)
-    return image
-
-
-def flatten_nested_dict(params, parent_key="", sep="/"):
-    items = []
-
-    for k, v in params.items():
-        new_key = parent_key + sep + k if parent_key else k
-
-        if isinstance(v, collections.abc.MutableMapping):
-            items.extend(flatten_nested_dict(v, new_key, sep=sep).items())
-        else:
-            items.append((new_key, v))
-    return dict(items)
-
-
-@torch.no_grad()
-def convert_siglip_checkpoint(model_name, pytorch_dump_folder_path, verify_logits=True, push_to_hub=False):
-    """
-    Copy/paste/tweak model's weights to our SigLIP structure.
-    """
-
-    # define default SigLIP configuration
-    config = get_siglip_config(model_name)
-
-    # get checkpoint
-    checkpoint = model_name_to_checkpoint[model_name]
-
-    # get vocab file
-    if "i18n" in model_name:
-        vocab_file = "/Users/nielsrogge/Documents/SigLIP/multilingual_vocab/sentencepiece.model"
-    else:
-        vocab_file = "/Users/nielsrogge/Documents/SigLIP/english_vocab/sentencepiece.model"
-
-    # load original state dict
-    data = load(checkpoint)
-    state_dict = flatten_nested_dict(data)
-
-    # remove and rename some keys
-    rename_keys = create_rename_keys(config)
-    for src, dest in rename_keys:
-        rename_key(state_dict, src, dest, config)
-
-    # qkv matrices of attention pooling head need special treatment
-    read_in_q_k_v_head(state_dict, config)
-
-    # load HuggingFace model
-    model = SiglipModel(config).eval()
-    model.load_state_dict(state_dict)
-
-    # create processor
-    # important: make tokenizer not return attention_mask since original one doesn't require it
-    image_size = config.vision_config.image_size
-    size = {"height": image_size, "width": image_size}
-    image_processor = SiglipImageProcessor(size=size)
-    tokenizer = SiglipTokenizer(vocab_file=vocab_file, model_input_names=["input_ids"])
-    processor = SiglipProcessor(image_processor=image_processor, tokenizer=tokenizer)
-
-    # verify on dummy images and texts
-    url_1 = "https://cdn.openai.com/multimodal-neurons/assets/apple/apple-ipod.jpg"
-    image_1 = Image.open(requests.get(url_1, stream=True).raw).convert("RGB")
-    url_2 = "https://cdn.openai.com/multimodal-neurons/assets/apple/apple-blank.jpg"
-    image_2 = Image.open(requests.get(url_2, stream=True).raw).convert("RGB")
-    texts = ["an apple", "a picture of an apple"]
-
-    inputs = processor(images=[image_1, image_2], text=texts, return_tensors="pt", padding="max_length")
-
-    # verify input_ids against original ones
-    if image_size == 224:
-        filename = "siglip_pixel_values.pt"
-    elif image_size == 256:
-        filename = "siglip_pixel_values_256.pt"
-    elif image_size == 384:
-        filename = "siglip_pixel_values_384.pt"
-    elif image_size == 512:
-        filename = "siglip_pixel_values_512.pt"
-    else:
-        raise ValueError("Image size not supported")
-
-    filepath = hf_hub_download(repo_id="nielsr/test-image", filename=filename, repo_type="dataset")
-    original_pixel_values = torch.load(filepath)
-    filepath = hf_hub_download(repo_id="nielsr/test-image", filename="siglip_input_ids.pt", repo_type="dataset")
-    original_input_ids = torch.load(filepath)
-
-    if "i18n" not in model_name:
-        assert inputs.input_ids.tolist() == original_input_ids.tolist()
-
-    print("Mean of original pixel values:", original_pixel_values.mean())
-    print("Mean of new pixel values:", inputs.pixel_values.mean())
-
-    # note: we're testing with original pixel values here since we don't have exact pixel values
-    with torch.no_grad():
-        outputs = model(input_ids=inputs.input_ids, pixel_values=original_pixel_values)
-
-    # with torch.no_grad():
-    #     outputs = model(input_ids=inputs.input_ids, pixel_values=inputs.pixel_values)
-
-    print(outputs.logits_per_image[:3, :3])
-
-    probs = torch.sigmoid(outputs.logits_per_image)  # these are the probabilities
-    print(f"{probs[0][0]:.1%} that image 0 is '{texts[0]}'")
-    print(f"{probs[0][1]:.1%} that image 0 is '{texts[1]}'")
-
-    if verify_logits:
-        if model_name == "siglip-base-patch16-224":
-            expected_slice = torch.tensor(
-                [[-2.9621, -2.1672], [-0.2713, 0.2910]],
-            )
-        elif model_name == "siglip-base-patch16-256":
-            expected_slice = torch.tensor(
-                [[-3.1146, -1.9894], [-0.7312, 0.6387]],
-            )
-        elif model_name == "siglip-base-patch16-384":
-            expected_slice = torch.tensor(
-                [[-2.8098, -2.1891], [-0.4242, 0.4102]],
-            )
-        elif model_name == "siglip-base-patch16-512":
-            expected_slice = torch.tensor(
-                [[-2.7899, -2.2668], [-0.4295, -0.0735]],
-            )
-        elif model_name == "siglip-large-patch16-256":
-            expected_slice = torch.tensor(
-                [[-1.5827, -0.5801], [-0.9153, 0.1363]],
-            )
-        elif model_name == "siglip-large-patch16-384":
-            expected_slice = torch.tensor(
-                [[-2.1523, -0.2899], [-0.2959, 0.7884]],
-            )
-        elif model_name == "siglip-so400m-patch14-384":
-            expected_slice = torch.tensor([[-1.2441, -0.6649], [-0.7060, 0.7374]])
-        elif model_name == "siglip-base-patch16-256-i18n":
-            expected_slice = torch.tensor(
-                [[-0.9064, 0.1073], [-0.0299, 0.5304]],
-            )
-
-        assert torch.allclose(outputs.logits_per_image[:3, :3], expected_slice, atol=1e-4)
-        print("Looks ok!")
-
-    if pytorch_dump_folder_path is not None:
-        Path(pytorch_dump_folder_path).mkdir(exist_ok=True)
-        print(f"Saving model {model_name} to {pytorch_dump_folder_path}")
-        model.save_pretrained(pytorch_dump_folder_path)
-        print(f"Saving processor to {pytorch_dump_folder_path}")
-        processor.save_pretrained(pytorch_dump_folder_path)
-
-    if push_to_hub:
-        model.push_to_hub(f"nielsr/{model_name}")
-        processor.push_to_hub(f"nielsr/{model_name}")
-
-
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser()
-    # Required parameters
-    parser.add_argument(
-        "--model_name",
-        default="siglip-base-patch16-224",
-        type=str,
-        choices=model_name_to_checkpoint.keys(),
-        help="Name of the model you'd like to convert.",
-    )
-    parser.add_argument(
-        "--pytorch_dump_folder_path", default=None, type=str, help="Path to the output PyTorch model directory."
-    )
-    parser.add_argument(
-        "--verify_logits",
-        action="store_false",
-        help="Whether to verify logits against the original implementation.",
-    )
-    parser.add_argument(
-        "--push_to_hub", action="store_true", help="Whether or not to push the converted model to the 🤗 hub."
-    )
-
-    args = parser.parse_args()
-    convert_siglip_checkpoint(args.model_name, args.pytorch_dump_folder_path, args.verify_logits, args.push_to_hub)
+# Copyright 2024 The HuggingFace Inc. team.
+# SPDX-License-Identifier: Apache-2.0
+
+"""Convert SigLIP checkpoints from the original repository.
+
+URL: https://github.com/google-research/big_vision/tree/main
+"""
+
+import argparse
+import collections
+from pathlib import Path
+
+import numpy as np
+import requests
+import torch
+from huggingface_hub import hf_hub_download
+from numpy import load
+from PIL import Image
+
+from transformers import SiglipConfig, SiglipImageProcessor, SiglipModel, SiglipProcessor, SiglipTokenizer
+from transformers.utils import logging
+
+
+logging.set_verbosity_info()
+logger = logging.get_logger(__name__)
+
+
+model_name_to_checkpoint = {
+    # base checkpoints
+    "siglip-base-patch16-224": "/Users/nielsrogge/Documents/SigLIP/webli_en_b16_224_63724782.npz",
+    "siglip-base-patch16-256": "/Users/nielsrogge/Documents/SigLIP/webli_en_b16_256_60500360.npz",
+    "siglip-base-patch16-384": "/Users/nielsrogge/Documents/SigLIP/webli_en_b16_384_68578854.npz",
+    "siglip-base-patch16-512": "/Users/nielsrogge/Documents/SigLIP/webli_en_b16_512_68580893.npz",
+    # large checkpoints
+    "siglip-large-patch16-256": "/Users/nielsrogge/Documents/SigLIP/webli_en_l16_256_60552751.npz",
+    "siglip-large-patch16-384": "/Users/nielsrogge/Documents/SigLIP/webli_en_l16_384_63634585.npz",
+    # multilingual checkpoint
+    "siglip-base-patch16-256-i18n": "/Users/nielsrogge/Documents/SigLIP/webli_i18n_b16_256_66117334.npz",
+    # so400m checkpoints
+    "siglip-so400m-patch14-384": "/Users/nielsrogge/Documents/SigLIP/webli_en_so400m_384_58765454.npz",
+}
+
+model_name_to_image_size = {
+    "siglip-base-patch16-224": 224,
+    "siglip-base-patch16-256": 256,
+    "siglip-base-patch16-384": 384,
+    "siglip-base-patch16-512": 512,
+    "siglip-large-patch16-256": 256,
+    "siglip-large-patch16-384": 384,
+    "siglip-base-patch16-256-i18n": 256,
+    "siglip-so400m-patch14-384": 384,
+}
+
+
+def get_siglip_config(model_name):
+    config = SiglipConfig()
+
+    vocab_size = 250000 if "i18n" in model_name else 32000
+    image_size = model_name_to_image_size[model_name]
+    patch_size = 16 if "patch16" in model_name else 14
+
+    # size of the architecture
+    config.vision_config.image_size = image_size
+    config.vision_config.patch_size = patch_size
+    config.text_config.vocab_size = vocab_size
+
+    if "base" in model_name:
+        pass
+    elif "large" in model_name:
+        config.text_config.hidden_size = 1024
+        config.text_config.intermediate_size = 4096
+        config.text_config.num_hidden_layers = 24
+        config.text_config.num_attention_heads = 16
+        config.vision_config.hidden_size = 1024
+        config.vision_config.intermediate_size = 4096
+        config.vision_config.num_hidden_layers = 24
+        config.vision_config.num_attention_heads = 16
+    elif "so400m" in model_name:
+        config.text_config.hidden_size = 1152
+        config.text_config.intermediate_size = 4304
+        config.text_config.num_hidden_layers = 27
+        config.text_config.num_attention_heads = 16
+        config.vision_config.hidden_size = 1152
+        config.vision_config.intermediate_size = 4304
+        config.vision_config.num_hidden_layers = 27
+        config.vision_config.num_attention_heads = 16
+    else:
+        raise ValueError("Model not supported")
+
+    return config
+
+
+def create_rename_keys(config):
+    rename_keys = []
+    # fmt: off
+
+    # vision encoder
+
+    rename_keys.append(("params/img/embedding/kernel", "vision_model.embeddings.patch_embedding.weight"))
+    rename_keys.append(("params/img/embedding/bias", "vision_model.embeddings.patch_embedding.bias"))
+    rename_keys.append(("params/img/pos_embedding", "vision_model.embeddings.position_embedding.weight"))
+
+    for i in range(config.vision_config.num_hidden_layers):
+        rename_keys.append((f"params/img/Transformer/encoderblock_{i}/LayerNorm_0/scale", f"vision_model.encoder.layers.{i}.layer_norm1.weight"))
+        rename_keys.append((f"params/img/Transformer/encoderblock_{i}/LayerNorm_0/bias", f"vision_model.encoder.layers.{i}.layer_norm1.bias"))
+        rename_keys.append((f"params/img/Transformer/encoderblock_{i}/LayerNorm_1/scale", f"vision_model.encoder.layers.{i}.layer_norm2.weight"))
+        rename_keys.append((f"params/img/Transformer/encoderblock_{i}/LayerNorm_1/bias", f"vision_model.encoder.layers.{i}.layer_norm2.bias"))
+        rename_keys.append((f"params/img/Transformer/encoderblock_{i}/MlpBlock_0/Dense_0/kernel", f"vision_model.encoder.layers.{i}.mlp.fc1.weight"))
+        rename_keys.append((f"params/img/Transformer/encoderblock_{i}/MlpBlock_0/Dense_0/bias", f"vision_model.encoder.layers.{i}.mlp.fc1.bias"))
+        rename_keys.append((f"params/img/Transformer/encoderblock_{i}/MlpBlock_0/Dense_1/kernel", f"vision_model.encoder.layers.{i}.mlp.fc2.weight"))
+        rename_keys.append((f"params/img/Transformer/encoderblock_{i}/MlpBlock_0/Dense_1/bias", f"vision_model.encoder.layers.{i}.mlp.fc2.bias"))
+        rename_keys.append((f"params/img/Transformer/encoderblock_{i}/MultiHeadDotProductAttention_0/key/kernel", f"vision_model.encoder.layers.{i}.self_attn.k_proj.weight"))
+        rename_keys.append((f"params/img/Transformer/encoderblock_{i}/MultiHeadDotProductAttention_0/key/bias", f"vision_model.encoder.layers.{i}.self_attn.k_proj.bias"))
+        rename_keys.append((f"params/img/Transformer/encoderblock_{i}/MultiHeadDotProductAttention_0/value/kernel", f"vision_model.encoder.layers.{i}.self_attn.v_proj.weight"))
+        rename_keys.append((f"params/img/Transformer/encoderblock_{i}/MultiHeadDotProductAttention_0/value/bias", f"vision_model.encoder.layers.{i}.self_attn.v_proj.bias"))
+        rename_keys.append((f"params/img/Transformer/encoderblock_{i}/MultiHeadDotProductAttention_0/query/kernel", f"vision_model.encoder.layers.{i}.self_attn.q_proj.weight"))
+        rename_keys.append((f"params/img/Transformer/encoderblock_{i}/MultiHeadDotProductAttention_0/query/bias", f"vision_model.encoder.layers.{i}.self_attn.q_proj.bias"))
+        rename_keys.append((f"params/img/Transformer/encoderblock_{i}/MultiHeadDotProductAttention_0/out/kernel", f"vision_model.encoder.layers.{i}.self_attn.out_proj.weight"))
+        rename_keys.append((f"params/img/Transformer/encoderblock_{i}/MultiHeadDotProductAttention_0/out/bias", f"vision_model.encoder.layers.{i}.self_attn.out_proj.bias"))
+
+    rename_keys.append(("params/img/Transformer/encoder_norm/scale", "vision_model.post_layernorm.weight"))
+    rename_keys.append(("params/img/Transformer/encoder_norm/bias", "vision_model.post_layernorm.bias"))
+
+    rename_keys.append(("params/img/MAPHead_0/probe", "vision_model.head.probe"))
+    rename_keys.append(("params/img/MAPHead_0/LayerNorm_0/scale", "vision_model.head.layernorm.weight"))
+    rename_keys.append(("params/img/MAPHead_0/LayerNorm_0/bias", "vision_model.head.layernorm.bias"))
+    rename_keys.append(("params/img/MAPHead_0/MlpBlock_0/Dense_0/kernel", "vision_model.head.mlp.fc1.weight"))
+    rename_keys.append(("params/img/MAPHead_0/MlpBlock_0/Dense_0/bias", "vision_model.head.mlp.fc1.bias"))
+    rename_keys.append(("params/img/MAPHead_0/MlpBlock_0/Dense_1/kernel", "vision_model.head.mlp.fc2.weight"))
+    rename_keys.append(("params/img/MAPHead_0/MlpBlock_0/Dense_1/bias", "vision_model.head.mlp.fc2.bias"))
+    rename_keys.append(("params/img/MAPHead_0/MultiHeadDotProductAttention_0/out/kernel", "vision_model.head.attention.out_proj.weight"))
+    rename_keys.append(("params/img/MAPHead_0/MultiHeadDotProductAttention_0/out/bias", "vision_model.head.attention.out_proj.bias"))
+
+    # text encoder
+
+    rename_keys.append(("params/txt/Embed_0/embedding", "text_model.embeddings.token_embedding.weight"))
+    rename_keys.append(("params/txt/pos_embedding", "text_model.embeddings.position_embedding.weight"))
+
+    for i in range(config.text_config.num_hidden_layers):
+        rename_keys.append((f"params/txt/Encoder_0/encoderblock_{i}/LayerNorm_0/scale", f"text_model.encoder.layers.{i}.layer_norm1.weight"))
+        rename_keys.append((f"params/txt/Encoder_0/encoderblock_{i}/LayerNorm_0/bias", f"text_model.encoder.layers.{i}.layer_norm1.bias"))
+        rename_keys.append((f"params/txt/Encoder_0/encoderblock_{i}/LayerNorm_1/scale", f"text_model.encoder.layers.{i}.layer_norm2.weight"))
+        rename_keys.append((f"params/txt/Encoder_0/encoderblock_{i}/LayerNorm_1/bias", f"text_model.encoder.layers.{i}.layer_norm2.bias"))
+        rename_keys.append((f"params/txt/Encoder_0/encoderblock_{i}/MlpBlock_0/Dense_0/kernel", f"text_model.encoder.layers.{i}.mlp.fc1.weight"))
+        rename_keys.append((f"params/txt/Encoder_0/encoderblock_{i}/MlpBlock_0/Dense_0/bias", f"text_model.encoder.layers.{i}.mlp.fc1.bias"))
+        rename_keys.append((f"params/txt/Encoder_0/encoderblock_{i}/MlpBlock_0/Dense_1/kernel", f"text_model.encoder.layers.{i}.mlp.fc2.weight"))
+        rename_keys.append((f"params/txt/Encoder_0/encoderblock_{i}/MlpBlock_0/Dense_1/bias", f"text_model.encoder.layers.{i}.mlp.fc2.bias"))
+        rename_keys.append((f"params/txt/Encoder_0/encoderblock_{i}/MultiHeadDotProductAttention_0/key/kernel", f"text_model.encoder.layers.{i}.self_attn.k_proj.weight"))
+        rename_keys.append((f"params/txt/Encoder_0/encoderblock_{i}/MultiHeadDotProductAttention_0/key/bias", f"text_model.encoder.layers.{i}.self_attn.k_proj.bias"))
+        rename_keys.append((f"params/txt/Encoder_0/encoderblock_{i}/MultiHeadDotProductAttention_0/value/kernel", f"text_model.encoder.layers.{i}.self_attn.v_proj.weight"))
+        rename_keys.append((f"params/txt/Encoder_0/encoderblock_{i}/MultiHeadDotProductAttention_0/value/bias", f"text_model.encoder.layers.{i}.self_attn.v_proj.bias"))
+        rename_keys.append((f"params/txt/Encoder_0/encoderblock_{i}/MultiHeadDotProductAttention_0/query/kernel", f"text_model.encoder.layers.{i}.self_attn.q_proj.weight"))
+        rename_keys.append((f"params/txt/Encoder_0/encoderblock_{i}/MultiHeadDotProductAttention_0/query/bias", f"text_model.encoder.layers.{i}.self_attn.q_proj.bias"))
+        rename_keys.append((f"params/txt/Encoder_0/encoderblock_{i}/MultiHeadDotProductAttention_0/out/kernel", f"text_model.encoder.layers.{i}.self_attn.out_proj.weight"))
+        rename_keys.append((f"params/txt/Encoder_0/encoderblock_{i}/MultiHeadDotProductAttention_0/out/bias", f"text_model.encoder.layers.{i}.self_attn.out_proj.bias"))
+
+    rename_keys.append(("params/txt/Encoder_0/encoder_norm/scale", "text_model.final_layer_norm.weight"))
+    rename_keys.append(("params/txt/Encoder_0/encoder_norm/bias", "text_model.final_layer_norm.bias"))
+    rename_keys.append(("params/txt/head/kernel", "text_model.head.weight"))
+    rename_keys.append(("params/txt/head/bias", "text_model.head.bias"))
+
+    # learned temperature and bias
+    rename_keys.append(("params/t", "logit_scale"))
+    rename_keys.append(("params/b", "logit_bias"))
+
+    # fmt: on
+    return rename_keys
+
+
+def rename_key(dct, old, new, config):
+    val = dct.pop(old)
+
+    if ("out_proj" in new or "v_proj" in new or "k_proj" in new or "q_proj" in new) and "vision" in new:
+        val = val.reshape(-1, config.vision_config.hidden_size)
+    if ("out_proj" in new or "v_proj" in new or "k_proj" in new or "q_proj" in new) and "text" in new:
+        val = val.reshape(-1, config.text_config.hidden_size)
+
+    if "patch_embedding.weight" in new:
+        val = val.transpose(3, 2, 0, 1)
+    elif new.endswith("weight") and "position_embedding" not in new and "token_embedding" not in new:
+        val = val.T
+
+    if "position_embedding" in new and "vision" in new:
+        val = val.reshape(-1, config.vision_config.hidden_size)
+    if "position_embedding" in new and "text" in new:
+        val = val.reshape(-1, config.text_config.hidden_size)
+
+    if new.endswith("bias"):
+        val = val.reshape(-1)
+
+    dct[new] = torch.from_numpy(val)
+
+
+def read_in_q_k_v_head(state_dict, config):
+    # read in individual input projection layers
+    key_proj_weight = (
+        state_dict.pop("params/img/MAPHead_0/MultiHeadDotProductAttention_0/key/kernel")
+        .reshape(-1, config.vision_config.hidden_size)
+        .T
+    )
+    key_proj_bias = state_dict.pop("params/img/MAPHead_0/MultiHeadDotProductAttention_0/key/bias").reshape(-1)
+    value_proj_weight = (
+        state_dict.pop("params/img/MAPHead_0/MultiHeadDotProductAttention_0/value/kernel")
+        .reshape(-1, config.vision_config.hidden_size)
+        .T
+    )
+    value_proj_bias = state_dict.pop("params/img/MAPHead_0/MultiHeadDotProductAttention_0/value/bias").reshape(-1)
+    query_proj_weight = (
+        state_dict.pop("params/img/MAPHead_0/MultiHeadDotProductAttention_0/query/kernel")
+        .reshape(-1, config.vision_config.hidden_size)
+        .T
+    )
+    query_proj_bias = state_dict.pop("params/img/MAPHead_0/MultiHeadDotProductAttention_0/query/bias").reshape(-1)
+
+    # next, add them to the state dict as a single matrix + vector
+    state_dict["vision_model.head.attention.in_proj_weight"] = torch.from_numpy(
+        np.concatenate([query_proj_weight, key_proj_weight, value_proj_weight], axis=0)
+    )
+    state_dict["vision_model.head.attention.in_proj_bias"] = torch.from_numpy(
+        np.concatenate([query_proj_bias, key_proj_bias, value_proj_bias], axis=0)
+    )
+
+
+# We will verify our results on an image of cute cats
+def prepare_img():
+    url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+    image = Image.open(requests.get(url, stream=True).raw)
+    return image
+
+
+def flatten_nested_dict(params, parent_key="", sep="/"):
+    items = []
+
+    for k, v in params.items():
+        new_key = parent_key + sep + k if parent_key else k
+
+        if isinstance(v, collections.abc.MutableMapping):
+            items.extend(flatten_nested_dict(v, new_key, sep=sep).items())
+        else:
+            items.append((new_key, v))
+    return dict(items)
+
+
+@torch.no_grad()
+def convert_siglip_checkpoint(model_name, pytorch_dump_folder_path, verify_logits=True, push_to_hub=False):
+    """
+    Copy/paste/tweak model's weights to our SigLIP structure.
+    """
+
+    # define default SigLIP configuration
+    config = get_siglip_config(model_name)
+
+    # get checkpoint
+    checkpoint = model_name_to_checkpoint[model_name]
+
+    # get vocab file
+    if "i18n" in model_name:
+        vocab_file = "/Users/nielsrogge/Documents/SigLIP/multilingual_vocab/sentencepiece.model"
+    else:
+        vocab_file = "/Users/nielsrogge/Documents/SigLIP/english_vocab/sentencepiece.model"
+
+    # load original state dict
+    data = load(checkpoint)
+    state_dict = flatten_nested_dict(data)
+
+    # remove and rename some keys
+    rename_keys = create_rename_keys(config)
+    for src, dest in rename_keys:
+        rename_key(state_dict, src, dest, config)
+
+    # qkv matrices of attention pooling head need special treatment
+    read_in_q_k_v_head(state_dict, config)
+
+    # load HuggingFace model
+    model = SiglipModel(config).eval()
+    model.load_state_dict(state_dict)
+
+    # create processor
+    # important: make tokenizer not return attention_mask since original one doesn't require it
+    image_size = config.vision_config.image_size
+    size = {"height": image_size, "width": image_size}
+    image_processor = SiglipImageProcessor(size=size)
+    tokenizer = SiglipTokenizer(vocab_file=vocab_file, model_input_names=["input_ids"])
+    processor = SiglipProcessor(image_processor=image_processor, tokenizer=tokenizer)
+
+    # verify on dummy images and texts
+    url_1 = "https://cdn.openai.com/multimodal-neurons/assets/apple/apple-ipod.jpg"
+    image_1 = Image.open(requests.get(url_1, stream=True).raw).convert("RGB")
+    url_2 = "https://cdn.openai.com/multimodal-neurons/assets/apple/apple-blank.jpg"
+    image_2 = Image.open(requests.get(url_2, stream=True).raw).convert("RGB")
+    texts = ["an apple", "a picture of an apple"]
+
+    inputs = processor(images=[image_1, image_2], text=texts, return_tensors="pt", padding="max_length")
+
+    # verify input_ids against original ones
+    if image_size == 224:
+        filename = "siglip_pixel_values.pt"
+    elif image_size == 256:
+        filename = "siglip_pixel_values_256.pt"
+    elif image_size == 384:
+        filename = "siglip_pixel_values_384.pt"
+    elif image_size == 512:
+        filename = "siglip_pixel_values_512.pt"
+    else:
+        raise ValueError("Image size not supported")
+
+    filepath = hf_hub_download(repo_id="nielsr/test-image", filename=filename, repo_type="dataset")
+    original_pixel_values = torch.load(filepath)
+    filepath = hf_hub_download(repo_id="nielsr/test-image", filename="siglip_input_ids.pt", repo_type="dataset")
+    original_input_ids = torch.load(filepath)
+
+    if "i18n" not in model_name:
+        assert inputs.input_ids.tolist() == original_input_ids.tolist()
+
+    print("Mean of original pixel values:", original_pixel_values.mean())
+    print("Mean of new pixel values:", inputs.pixel_values.mean())
+
+    # note: we're testing with original pixel values here since we don't have exact pixel values
+    with torch.no_grad():
+        outputs = model(input_ids=inputs.input_ids, pixel_values=original_pixel_values)
+
+    # with torch.no_grad():
+    #     outputs = model(input_ids=inputs.input_ids, pixel_values=inputs.pixel_values)
+
+    print(outputs.logits_per_image[:3, :3])
+
+    probs = torch.sigmoid(outputs.logits_per_image)  # these are the probabilities
+    print(f"{probs[0][0]:.1%} that image 0 is '{texts[0]}'")
+    print(f"{probs[0][1]:.1%} that image 0 is '{texts[1]}'")
+
+    if verify_logits:
+        if model_name == "siglip-base-patch16-224":
+            expected_slice = torch.tensor(
+                [[-2.9621, -2.1672], [-0.2713, 0.2910]],
+            )
+        elif model_name == "siglip-base-patch16-256":
+            expected_slice = torch.tensor(
+                [[-3.1146, -1.9894], [-0.7312, 0.6387]],
+            )
+        elif model_name == "siglip-base-patch16-384":
+            expected_slice = torch.tensor(
+                [[-2.8098, -2.1891], [-0.4242, 0.4102]],
+            )
+        elif model_name == "siglip-base-patch16-512":
+            expected_slice = torch.tensor(
+                [[-2.7899, -2.2668], [-0.4295, -0.0735]],
+            )
+        elif model_name == "siglip-large-patch16-256":
+            expected_slice = torch.tensor(
+                [[-1.5827, -0.5801], [-0.9153, 0.1363]],
+            )
+        elif model_name == "siglip-large-patch16-384":
+            expected_slice = torch.tensor(
+                [[-2.1523, -0.2899], [-0.2959, 0.7884]],
+            )
+        elif model_name == "siglip-so400m-patch14-384":
+            expected_slice = torch.tensor([[-1.2441, -0.6649], [-0.7060, 0.7374]])
+        elif model_name == "siglip-base-patch16-256-i18n":
+            expected_slice = torch.tensor(
+                [[-0.9064, 0.1073], [-0.0299, 0.5304]],
+            )
+
+        assert torch.allclose(outputs.logits_per_image[:3, :3], expected_slice, atol=1e-4)
+        print("Looks ok!")
+
+    if pytorch_dump_folder_path is not None:
+        Path(pytorch_dump_folder_path).mkdir(exist_ok=True)
+        print(f"Saving model {model_name} to {pytorch_dump_folder_path}")
+        model.save_pretrained(pytorch_dump_folder_path)
+        print(f"Saving processor to {pytorch_dump_folder_path}")
+        processor.save_pretrained(pytorch_dump_folder_path)
+
+    if push_to_hub:
+        model.push_to_hub(f"nielsr/{model_name}")
+        processor.push_to_hub(f"nielsr/{model_name}")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    # Required parameters
+    parser.add_argument(
+        "--model_name",
+        default="siglip-base-patch16-224",
+        type=str,
+        choices=model_name_to_checkpoint.keys(),
+        help="Name of the model you'd like to convert.",
+    )
+    parser.add_argument(
+        "--pytorch_dump_folder_path", default=None, type=str, help="Path to the output PyTorch model directory."
+    )
+    parser.add_argument(
+        "--verify_logits",
+        action="store_false",
+        help="Whether to verify logits against the original implementation.",
+    )
+    parser.add_argument(
+        "--push_to_hub", action="store_true", help="Whether or not to push the converted model to the 🤗 hub."
+    )
+
+    args = parser.parse_args()
+    convert_siglip_checkpoint(args.model_name, args.pytorch_dump_folder_path, args.verify_logits, args.push_to_hub)

modeling/siglip/image_processing_siglip.py

@@ -1,230 +1,230 @@
-# Copyright 2024 The HuggingFace Inc. team.
-# SPDX-License-Identifier: Apache-2.0
-
-"""Image processor class for SigLIP."""
-
-from typing import Dict, List, Optional, Union
-
-from transformers.image_processing_utils import BaseImageProcessor, BatchFeature, get_size_dict
-from transformers.image_transforms import (
-    convert_to_rgb,
-    resize,
-    to_channel_dimension_format,
-)
-from transformers.image_utils import (
-    IMAGENET_STANDARD_MEAN,
-    IMAGENET_STANDARD_STD,
-    ChannelDimension,
-    ImageInput,
-    PILImageResampling,
-    infer_channel_dimension_format,
-    is_scaled_image,
-    make_list_of_images,
-    to_numpy_array,
-    valid_images,
-    validate_preprocess_arguments,
-)
-from transformers.utils import TensorType, filter_out_non_signature_kwargs, is_vision_available, logging
-
-
-logger = logging.get_logger(__name__)
-
-
-if is_vision_available():
-    import PIL
-
-
-class SiglipImageProcessor(BaseImageProcessor):
-    r"""
-    Constructs a SigLIP image processor.
-
-    Args:
-        do_resize (`bool`, *optional*, defaults to `True`):
-            Whether to resize the image's (height, width) dimensions to the specified `size`. Can be overridden by
-            `do_resize` in the `preprocess` method.
-        size (`Dict[str, int]` *optional*, defaults to `{"height": 224, "width": 224}`):
-            Size of the image after resizing. Can be overridden by `size` in the `preprocess` method.
-        resample (`PILImageResampling`, *optional*, defaults to `Resampling.BICUBIC`):
-            Resampling filter to use if resizing the image. Can be overridden by `resample` in the `preprocess` method.
-        do_rescale (`bool`, *optional*, defaults to `True`):
-            Whether to rescale the image by the specified scale `rescale_factor`. Can be overridden by `do_rescale` in
-            the `preprocess` method.
-        rescale_factor (`int` or `float`, *optional*, defaults to `1/255`):
-            Scale factor to use if rescaling the image. Can be overridden by `rescale_factor` in the `preprocess`
-            method.
-        do_normalize (`bool`, *optional*, defaults to `True`):
-            Whether to normalize the image by the specified mean and standard deviation. Can be overridden by
-            `do_normalize` in the `preprocess` method.
-        image_mean (`float` or `List[float]`, *optional*, defaults to `[0.5, 0.5, 0.5]`):
-            Mean to use if normalizing the image. This is a float or list of floats the length of the number of
-            channels in the image. Can be overridden by the `image_mean` parameter in the `preprocess` method.
-        image_std (`float` or `List[float]`, *optional*, defaults to `[0.5, 0.5, 0.5]`):
-            Standard deviation to use if normalizing the image. This is a float or list of floats the length of the
-            number of channels in the image. Can be overridden by the `image_std` parameter in the `preprocess` method.
-            Can be overridden by the `image_std` parameter in the `preprocess` method.
-        do_convert_rgb (`bool`, *optional*, defaults to `True`):
-            Whether to convert the image to RGB.
-    """
-
-    model_input_names = ["pixel_values"]
-
-    def __init__(
-        self,
-        do_resize: bool = True,
-        size: Dict[str, int] = None,
-        resample: PILImageResampling = PILImageResampling.BICUBIC,
-        do_rescale: bool = True,
-        rescale_factor: Union[int, float] = 1 / 255,
-        do_normalize: bool = True,
-        image_mean: Optional[Union[float, List[float]]] = None,
-        image_std: Optional[Union[float, List[float]]] = None,
-        do_convert_rgb: bool = None,
-        **kwargs,
-    ) -> None:
-        super().__init__(**kwargs)
-        size = size if size is not None else {"height": 224, "width": 224}
-        image_mean = image_mean if image_mean is not None else IMAGENET_STANDARD_MEAN
-        image_std = image_std if image_std is not None else IMAGENET_STANDARD_STD
-
-        self.do_resize = do_resize
-        self.size = size
-        self.resample = resample
-        self.do_rescale = do_rescale
-        self.rescale_factor = rescale_factor
-        self.do_normalize = do_normalize
-        self.image_mean = image_mean
-        self.image_std = image_std
-        self.do_convert_rgb = do_convert_rgb
-
-    @filter_out_non_signature_kwargs()
-    def preprocess(
-        self,
-        images: ImageInput,
-        do_resize: bool = None,
-        size: Dict[str, int] = None,
-        resample: PILImageResampling = None,
-        do_rescale: bool = None,
-        rescale_factor: float = None,
-        do_normalize: bool = None,
-        image_mean: Optional[Union[float, List[float]]] = None,
-        image_std: Optional[Union[float, List[float]]] = None,
-        return_tensors: Optional[Union[str, TensorType]] = None,
-        data_format: Optional[ChannelDimension] = ChannelDimension.FIRST,
-        input_data_format: Optional[Union[str, ChannelDimension]] = None,
-        do_convert_rgb: bool = None,
-    ) -> PIL.Image.Image:
-        """
-        Preprocess an image or batch of images.
-
-        Args:
-            images (`ImageInput`):
-                Image to preprocess. Expects a single or batch of images with pixel values ranging from 0 to 255. If
-                passing in images with pixel values between 0 and 1, set `do_rescale=False`.
-            do_resize (`bool`, *optional*, defaults to `self.do_resize`):
-                Whether to resize the image.
-            size (`Dict[str, int]`, *optional*, defaults to `self.size`):
-                Size of the image after resizing.
-            resample (`int`, *optional*, defaults to `self.resample`):
-                Resampling filter to use if resizing the image. This can be one of the enum `PILImageResampling`. Only
-                has an effect if `do_resize` is set to `True`.
-            do_rescale (`bool`, *optional*, defaults to `self.do_rescale`):
-                Whether to rescale the image.
-            rescale_factor (`float`, *optional*, defaults to `self.rescale_factor`):
-                Rescale factor to rescale the image by if `do_rescale` is set to `True`.
-            do_normalize (`bool`, *optional*, defaults to `self.do_normalize`):
-                Whether to normalize the image.
-            image_mean (`float` or `List[float]`, *optional*, defaults to `self.image_mean`):
-                Image mean to use for normalization. Only has an effect if `do_normalize` is set to `True`.
-            image_std (`float` or `List[float]`, *optional*, defaults to `self.image_std`):
-                Image standard deviation to use for normalization. Only has an effect if `do_normalize` is set to
-                `True`.
-            return_tensors (`str` or `TensorType`, *optional*):
-                The type of tensors to return. Can be one of:
-                - Unset: Return a list of `np.ndarray`.
-                - `TensorType.TENSORFLOW` or `'tf'`: Return a batch of type `tf.Tensor`.
-                - `TensorType.PYTORCH` or `'pt'`: Return a batch of type `torch.Tensor`.
-                - `TensorType.NUMPY` or `'np'`: Return a batch of type `np.ndarray`.
-                - `TensorType.JAX` or `'jax'`: Return a batch of type `jax.numpy.ndarray`.
-            data_format (`ChannelDimension` or `str`, *optional*, defaults to `ChannelDimension.FIRST`):
-                The channel dimension format for the output image. Can be one of:
-                - `"channels_first"` or `ChannelDimension.FIRST`: image in (num_channels, height, width) format.
-                - `"channels_last"` or `ChannelDimension.LAST`: image in (height, width, num_channels) format.
-                - Unset: Use the channel dimension format of the input image.
-            input_data_format (`ChannelDimension` or `str`, *optional*):
-                The channel dimension format for the input image. If unset, the channel dimension format is inferred
-                from the input image. Can be one of:
-                - `"channels_first"` or `ChannelDimension.FIRST`: image in (num_channels, height, width) format.
-                - `"channels_last"` or `ChannelDimension.LAST`: image in (height, width, num_channels) format.
-                - `"none"` or `ChannelDimension.NONE`: image in (height, width) format.
-            do_convert_rgb (`bool`, *optional*, defaults to `self.do_convert_rgb`):
-                Whether to convert the image to RGB.
-        """
-        do_resize = do_resize if do_resize is not None else self.do_resize
-        size = size if size is not None else self.size
-        size = get_size_dict(size, param_name="size", default_to_square=False)
-        resample = resample if resample is not None else self.resample
-        do_rescale = do_rescale if do_rescale is not None else self.do_rescale
-        rescale_factor = rescale_factor if rescale_factor is not None else self.rescale_factor
-        do_normalize = do_normalize if do_normalize is not None else self.do_normalize
-        image_mean = image_mean if image_mean is not None else self.image_mean
-        image_std = image_std if image_std is not None else self.image_std
-        do_convert_rgb = do_convert_rgb if do_convert_rgb is not None else self.do_convert_rgb
-
-        images = make_list_of_images(images)
-
-        if not valid_images(images):
-            raise ValueError(
-                "Invalid image type. Must be of type PIL.Image.Image, numpy.ndarray, "
-                "torch.Tensor, tf.Tensor or jax.ndarray."
-            )
-        validate_preprocess_arguments(
-            do_rescale=do_rescale,
-            rescale_factor=rescale_factor,
-            do_normalize=do_normalize,
-            image_mean=image_mean,
-            image_std=image_std,
-            do_resize=do_resize,
-            size=size,
-            resample=resample,
-        )
-        # All transformations expect numpy arrays.
-        images = [to_numpy_array(image) for image in images]
-
-        if do_convert_rgb:
-            images = [convert_to_rgb(image) for image in images]
-
-        if is_scaled_image(images[0]) and do_rescale:
-            logger.warning_once(
-                "It looks like you are trying to rescale already rescaled images. If the input"
-                " images have pixel values between 0 and 1, set `do_rescale=False` to avoid rescaling them again."
-            )
-
-        if input_data_format is None:
-            # We assume that all images have the same channel dimension format.
-            input_data_format = infer_channel_dimension_format(images[0])
-
-        if do_resize:
-            height, width = size["height"], size["width"]
-            images = [
-                resize(image=image, size=(height, width), resample=resample, input_data_format=input_data_format)
-                for image in images
-            ]
-
-        if do_rescale:
-            images = [
-                self.rescale(image=image, scale=rescale_factor, input_data_format=input_data_format)
-                for image in images
-            ]
-
-        if do_normalize:
-            images = [
-                self.normalize(image=image, mean=image_mean, std=image_std, input_data_format=input_data_format)
-                for image in images
-            ]
-
-        images = [
-            to_channel_dimension_format(image, data_format, input_channel_dim=input_data_format) for image in images
-        ]
-
-        data = {"pixel_values": images}
-        return BatchFeature(data=data, tensor_type=return_tensors)
+# Copyright 2024 The HuggingFace Inc. team.
+# SPDX-License-Identifier: Apache-2.0
+
+"""Image processor class for SigLIP."""
+
+from typing import Dict, List, Optional, Union
+
+from transformers.image_processing_utils import BaseImageProcessor, BatchFeature, get_size_dict
+from transformers.image_transforms import (
+    convert_to_rgb,
+    resize,
+    to_channel_dimension_format,
+)
+from transformers.image_utils import (
+    IMAGENET_STANDARD_MEAN,
+    IMAGENET_STANDARD_STD,
+    ChannelDimension,
+    ImageInput,
+    PILImageResampling,
+    infer_channel_dimension_format,
+    is_scaled_image,
+    make_list_of_images,
+    to_numpy_array,
+    valid_images,
+    validate_preprocess_arguments,
+)
+from transformers.utils import TensorType, filter_out_non_signature_kwargs, is_vision_available, logging
+
+
+logger = logging.get_logger(__name__)
+
+
+if is_vision_available():
+    import PIL
+
+
+class SiglipImageProcessor(BaseImageProcessor):
+    r"""
+    Constructs a SigLIP image processor.
+
+    Args:
+        do_resize (`bool`, *optional*, defaults to `True`):
+            Whether to resize the image's (height, width) dimensions to the specified `size`. Can be overridden by
+            `do_resize` in the `preprocess` method.
+        size (`Dict[str, int]` *optional*, defaults to `{"height": 224, "width": 224}`):
+            Size of the image after resizing. Can be overridden by `size` in the `preprocess` method.
+        resample (`PILImageResampling`, *optional*, defaults to `Resampling.BICUBIC`):
+            Resampling filter to use if resizing the image. Can be overridden by `resample` in the `preprocess` method.
+        do_rescale (`bool`, *optional*, defaults to `True`):
+            Whether to rescale the image by the specified scale `rescale_factor`. Can be overridden by `do_rescale` in
+            the `preprocess` method.
+        rescale_factor (`int` or `float`, *optional*, defaults to `1/255`):
+            Scale factor to use if rescaling the image. Can be overridden by `rescale_factor` in the `preprocess`
+            method.
+        do_normalize (`bool`, *optional*, defaults to `True`):
+            Whether to normalize the image by the specified mean and standard deviation. Can be overridden by
+            `do_normalize` in the `preprocess` method.
+        image_mean (`float` or `List[float]`, *optional*, defaults to `[0.5, 0.5, 0.5]`):
+            Mean to use if normalizing the image. This is a float or list of floats the length of the number of
+            channels in the image. Can be overridden by the `image_mean` parameter in the `preprocess` method.
+        image_std (`float` or `List[float]`, *optional*, defaults to `[0.5, 0.5, 0.5]`):
+            Standard deviation to use if normalizing the image. This is a float or list of floats the length of the
+            number of channels in the image. Can be overridden by the `image_std` parameter in the `preprocess` method.
+            Can be overridden by the `image_std` parameter in the `preprocess` method.
+        do_convert_rgb (`bool`, *optional*, defaults to `True`):
+            Whether to convert the image to RGB.
+    """
+
+    model_input_names = ["pixel_values"]
+
+    def __init__(
+        self,
+        do_resize: bool = True,
+        size: Dict[str, int] = None,
+        resample: PILImageResampling = PILImageResampling.BICUBIC,
+        do_rescale: bool = True,
+        rescale_factor: Union[int, float] = 1 / 255,
+        do_normalize: bool = True,
+        image_mean: Optional[Union[float, List[float]]] = None,
+        image_std: Optional[Union[float, List[float]]] = None,
+        do_convert_rgb: bool = None,
+        **kwargs,
+    ) -> None:
+        super().__init__(**kwargs)
+        size = size if size is not None else {"height": 224, "width": 224}
+        image_mean = image_mean if image_mean is not None else IMAGENET_STANDARD_MEAN
+        image_std = image_std if image_std is not None else IMAGENET_STANDARD_STD
+
+        self.do_resize = do_resize
+        self.size = size
+        self.resample = resample
+        self.do_rescale = do_rescale
+        self.rescale_factor = rescale_factor
+        self.do_normalize = do_normalize
+        self.image_mean = image_mean
+        self.image_std = image_std
+        self.do_convert_rgb = do_convert_rgb
+
+    @filter_out_non_signature_kwargs()
+    def preprocess(
+        self,
+        images: ImageInput,
+        do_resize: bool = None,
+        size: Dict[str, int] = None,
+        resample: PILImageResampling = None,
+        do_rescale: bool = None,
+        rescale_factor: float = None,
+        do_normalize: bool = None,
+        image_mean: Optional[Union[float, List[float]]] = None,
+        image_std: Optional[Union[float, List[float]]] = None,
+        return_tensors: Optional[Union[str, TensorType]] = None,
+        data_format: Optional[ChannelDimension] = ChannelDimension.FIRST,
+        input_data_format: Optional[Union[str, ChannelDimension]] = None,
+        do_convert_rgb: bool = None,
+    ) -> PIL.Image.Image:
+        """
+        Preprocess an image or batch of images.
+
+        Args:
+            images (`ImageInput`):
+                Image to preprocess. Expects a single or batch of images with pixel values ranging from 0 to 255. If
+                passing in images with pixel values between 0 and 1, set `do_rescale=False`.
+            do_resize (`bool`, *optional*, defaults to `self.do_resize`):
+                Whether to resize the image.
+            size (`Dict[str, int]`, *optional*, defaults to `self.size`):
+                Size of the image after resizing.
+            resample (`int`, *optional*, defaults to `self.resample`):
+                Resampling filter to use if resizing the image. This can be one of the enum `PILImageResampling`. Only
+                has an effect if `do_resize` is set to `True`.
+            do_rescale (`bool`, *optional*, defaults to `self.do_rescale`):
+                Whether to rescale the image.
+            rescale_factor (`float`, *optional*, defaults to `self.rescale_factor`):
+                Rescale factor to rescale the image by if `do_rescale` is set to `True`.
+            do_normalize (`bool`, *optional*, defaults to `self.do_normalize`):
+                Whether to normalize the image.
+            image_mean (`float` or `List[float]`, *optional*, defaults to `self.image_mean`):
+                Image mean to use for normalization. Only has an effect if `do_normalize` is set to `True`.
+            image_std (`float` or `List[float]`, *optional*, defaults to `self.image_std`):
+                Image standard deviation to use for normalization. Only has an effect if `do_normalize` is set to
+                `True`.
+            return_tensors (`str` or `TensorType`, *optional*):
+                The type of tensors to return. Can be one of:
+                - Unset: Return a list of `np.ndarray`.
+                - `TensorType.TENSORFLOW` or `'tf'`: Return a batch of type `tf.Tensor`.
+                - `TensorType.PYTORCH` or `'pt'`: Return a batch of type `torch.Tensor`.
+                - `TensorType.NUMPY` or `'np'`: Return a batch of type `np.ndarray`.
+                - `TensorType.JAX` or `'jax'`: Return a batch of type `jax.numpy.ndarray`.
+            data_format (`ChannelDimension` or `str`, *optional*, defaults to `ChannelDimension.FIRST`):
+                The channel dimension format for the output image. Can be one of:
+                - `"channels_first"` or `ChannelDimension.FIRST`: image in (num_channels, height, width) format.
+                - `"channels_last"` or `ChannelDimension.LAST`: image in (height, width, num_channels) format.
+                - Unset: Use the channel dimension format of the input image.
+            input_data_format (`ChannelDimension` or `str`, *optional*):
+                The channel dimension format for the input image. If unset, the channel dimension format is inferred
+                from the input image. Can be one of:
+                - `"channels_first"` or `ChannelDimension.FIRST`: image in (num_channels, height, width) format.
+                - `"channels_last"` or `ChannelDimension.LAST`: image in (height, width, num_channels) format.
+                - `"none"` or `ChannelDimension.NONE`: image in (height, width) format.
+            do_convert_rgb (`bool`, *optional*, defaults to `self.do_convert_rgb`):
+                Whether to convert the image to RGB.
+        """
+        do_resize = do_resize if do_resize is not None else self.do_resize
+        size = size if size is not None else self.size
+        size = get_size_dict(size, param_name="size", default_to_square=False)
+        resample = resample if resample is not None else self.resample
+        do_rescale = do_rescale if do_rescale is not None else self.do_rescale
+        rescale_factor = rescale_factor if rescale_factor is not None else self.rescale_factor
+        do_normalize = do_normalize if do_normalize is not None else self.do_normalize
+        image_mean = image_mean if image_mean is not None else self.image_mean
+        image_std = image_std if image_std is not None else self.image_std
+        do_convert_rgb = do_convert_rgb if do_convert_rgb is not None else self.do_convert_rgb
+
+        images = make_list_of_images(images)
+
+        if not valid_images(images):
+            raise ValueError(
+                "Invalid image type. Must be of type PIL.Image.Image, numpy.ndarray, "
+                "torch.Tensor, tf.Tensor or jax.ndarray."
+            )
+        validate_preprocess_arguments(
+            do_rescale=do_rescale,
+            rescale_factor=rescale_factor,
+            do_normalize=do_normalize,
+            image_mean=image_mean,
+            image_std=image_std,
+            do_resize=do_resize,
+            size=size,
+            resample=resample,
+        )
+        # All transformations expect numpy arrays.
+        images = [to_numpy_array(image) for image in images]
+
+        if do_convert_rgb:
+            images = [convert_to_rgb(image) for image in images]
+
+        if is_scaled_image(images[0]) and do_rescale:
+            logger.warning_once(
+                "It looks like you are trying to rescale already rescaled images. If the input"
+                " images have pixel values between 0 and 1, set `do_rescale=False` to avoid rescaling them again."
+            )
+
+        if input_data_format is None:
+            # We assume that all images have the same channel dimension format.
+            input_data_format = infer_channel_dimension_format(images[0])
+
+        if do_resize:
+            height, width = size["height"], size["width"]
+            images = [
+                resize(image=image, size=(height, width), resample=resample, input_data_format=input_data_format)
+                for image in images
+            ]
+
+        if do_rescale:
+            images = [
+                self.rescale(image=image, scale=rescale_factor, input_data_format=input_data_format)
+                for image in images
+            ]
+
+        if do_normalize:
+            images = [
+                self.normalize(image=image, mean=image_mean, std=image_std, input_data_format=input_data_format)
+                for image in images
+            ]
+
+        images = [
+            to_channel_dimension_format(image, data_format, input_channel_dim=input_data_format) for image in images
+        ]
+
+        data = {"pixel_values": images}
+        return BatchFeature(data=data, tensor_type=return_tensors)

modeling/siglip/processing_siglip.py

@@ -1,131 +1,131 @@
-# Copyright 2024 The HuggingFace Inc. team.
-# SPDX-License-Identifier: Apache-2.0
-
-"""
-Image/Text processor class for SigLIP.
-"""
-
-from typing import List, Optional, Union
-
-from transformers.feature_extraction_utils import BatchFeature
-from transformers.image_utils import ImageInput
-from transformers.processing_utils import ProcessorMixin
-from transformers.tokenization_utils_base import PaddingStrategy, PreTokenizedInput, TextInput, TruncationStrategy
-from transformers.utils import TensorType
-
-
-class SiglipProcessor(ProcessorMixin):
-    r"""
-    Constructs a Siglip processor which wraps a Siglip image processor and a Siglip tokenizer into a single processor.
-
-    [`SiglipProcessor`] offers all the functionalities of [`SiglipImageProcessor`] and [`SiglipTokenizer`]. See the
-    [`~SiglipProcessor.__call__`] and [`~SiglipProcessor.decode`] for more information.
-
-    Args:
-        image_processor ([`SiglipImageProcessor`]):
-            The image processor is a required input.
-        tokenizer ([`SiglipTokenizer`]):
-            The tokenizer is a required input.
-    """
-
-    attributes = ["image_processor", "tokenizer"]
-    image_processor_class = "SiglipImageProcessor"
-    tokenizer_class = "SiglipTokenizer"
-
-    def __init__(self, image_processor, tokenizer):
-        super().__init__(image_processor, tokenizer)
-
-    def __call__(
-        self,
-        text: Union[TextInput, PreTokenizedInput, List[TextInput], List[PreTokenizedInput]] = None,
-        images: ImageInput = None,
-        padding: Union[bool, str, PaddingStrategy] = False,
-        truncation: Union[bool, str, TruncationStrategy] = None,
-        max_length: int = None,
-        return_tensors: Optional[Union[str, TensorType]] = TensorType.PYTORCH,
-    ) -> BatchFeature:
-        """
-        Main method to prepare for the model one or several sequences(s) and image(s). This method forwards the `text`
-        and `kwargs` arguments to SiglipTokenizer's [`~SiglipTokenizer.__call__`] if `text` is not `None` to encode
-        the text. To prepare the image(s), this method forwards the `images` argument to
-        SiglipImageProcessor's [`~SiglipImageProcessor.__call__`] if `images` is not `None`. Please refer to the doctsring
-        of the above two methods for more information.
-
-        Args:
-            text (`str`, `List[str]`, `List[List[str]]`):
-                The sequence or batch of sequences to be encoded. Each sequence can be a string or a list of strings
-                (pretokenized string). If the sequences are provided as list of strings (pretokenized), you must set
-                `is_split_into_words=True` (to lift the ambiguity with a batch of sequences).
-            images (`PIL.Image.Image`, `np.ndarray`, `torch.Tensor`, `List[PIL.Image.Image]`, `List[np.ndarray]`, `List[torch.Tensor]`):
-                The image or batch of images to be prepared. Each image can be a PIL image, NumPy array or PyTorch
-                tensor. Both channels-first and channels-last formats are supported.
-            padding (`bool`, `str` or [`~utils.PaddingStrategy`], *optional*, defaults to `False`):
-                Select a strategy to pad the returned sequences (according to the model's padding side and padding
-                index) among:
-                - `True` or `'longest'`: Pad to the longest sequence in the batch (or no padding if only a single
-                  sequence if provided).
-                - `'max_length'`: Pad to a maximum length specified with the argument `max_length` or to the maximum
-                  acceptable input length for the model if that argument is not provided.
-                - `False` or `'do_not_pad'` (default): No padding (i.e., can output a batch with sequences of different
-                  lengths).
-            max_length (`int`, *optional*):
-                Maximum length of the returned list and optionally padding length (see above).
-            truncation (`bool`, *optional*):
-                Activates truncation to cut input sequences longer than `max_length` to `max_length`.
-            return_tensors (`str` or [`~utils.TensorType`], *optional*):
-                If set, will return tensors of a particular framework. Acceptable values are:
-
-                - `'tf'`: Return TensorFlow `tf.constant` objects.
-                - `'pt'`: Return PyTorch `torch.Tensor` objects.
-                - `'np'`: Return NumPy `np.ndarray` objects.
-                - `'jax'`: Return JAX `jnp.ndarray` objects.
-
-        Returns:
-            [`BatchFeature`]: A [`BatchFeature`] with the following fields:
-
-            - **input_ids** -- List of token ids to be fed to a model. Returned when `text` is not `None`.
-            - **attention_mask** -- List of indices specifying which tokens should be attended to by the model (when
-              `return_attention_mask=True` or if *"attention_mask"* is in `self.model_input_names` and if `text` is not
-              `None`).
-            - **pixel_values** -- Pixel values to be fed to a model. Returned when `images` is not `None`.
-        """
-
-        if text is None and images is None:
-            raise ValueError("You have to specify either text or images. Both cannot be none.")
-
-        if text is not None:
-            encoding = self.tokenizer(
-                text, return_tensors=return_tensors, padding=padding, truncation=truncation, max_length=max_length
-            )
-
-        if images is not None:
-            image_features = self.image_processor(images, return_tensors=return_tensors)
-
-        if text is not None and images is not None:
-            encoding["pixel_values"] = image_features.pixel_values
-            return encoding
-        elif text is not None:
-            return encoding
-        else:
-            return BatchFeature(data=dict(**image_features), tensor_type=return_tensors)
-
-    def decode(self, *args, **kwargs):
-        """
-        This method forwards all its arguments to SiglipTokenizer's [`~PreTrainedTokenizer.decode`]. Please refer to
-        the docstring of this method for more information.
-        """
-        return self.tokenizer.decode(*args, **kwargs)
-
-    def batch_decode(self, *args, **kwargs):
-        """
-        This method forwards all its arguments to SiglipTokenizer's [`~PreTrainedTokenizer.batch_decode`]. Please
-        refer to the docstring of this method for more information.
-        """
-        return self.tokenizer.batch_decode(*args, **kwargs)
-
-    @property
-    # Copied from transformers.models.clip.processing_clip.CLIPProcessor.model_input_names with CLIP->Siglip, T5->Siglip
-    def model_input_names(self):
-        tokenizer_input_names = self.tokenizer.model_input_names
-        image_processor_input_names = self.image_processor.model_input_names
-        return list(dict.fromkeys(tokenizer_input_names + image_processor_input_names))
+# Copyright 2024 The HuggingFace Inc. team.
+# SPDX-License-Identifier: Apache-2.0
+
+"""
+Image/Text processor class for SigLIP.
+"""
+
+from typing import List, Optional, Union
+
+from transformers.feature_extraction_utils import BatchFeature
+from transformers.image_utils import ImageInput
+from transformers.processing_utils import ProcessorMixin
+from transformers.tokenization_utils_base import PaddingStrategy, PreTokenizedInput, TextInput, TruncationStrategy
+from transformers.utils import TensorType
+
+
+class SiglipProcessor(ProcessorMixin):
+    r"""
+    Constructs a Siglip processor which wraps a Siglip image processor and a Siglip tokenizer into a single processor.
+
+    [`SiglipProcessor`] offers all the functionalities of [`SiglipImageProcessor`] and [`SiglipTokenizer`]. See the
+    [`~SiglipProcessor.__call__`] and [`~SiglipProcessor.decode`] for more information.
+
+    Args:
+        image_processor ([`SiglipImageProcessor`]):
+            The image processor is a required input.
+        tokenizer ([`SiglipTokenizer`]):
+            The tokenizer is a required input.
+    """
+
+    attributes = ["image_processor", "tokenizer"]
+    image_processor_class = "SiglipImageProcessor"
+    tokenizer_class = "SiglipTokenizer"
+
+    def __init__(self, image_processor, tokenizer):
+        super().__init__(image_processor, tokenizer)
+
+    def __call__(
+        self,
+        text: Union[TextInput, PreTokenizedInput, List[TextInput], List[PreTokenizedInput]] = None,
+        images: ImageInput = None,
+        padding: Union[bool, str, PaddingStrategy] = False,
+        truncation: Union[bool, str, TruncationStrategy] = None,
+        max_length: int = None,
+        return_tensors: Optional[Union[str, TensorType]] = TensorType.PYTORCH,
+    ) -> BatchFeature:
+        """
+        Main method to prepare for the model one or several sequences(s) and image(s). This method forwards the `text`
+        and `kwargs` arguments to SiglipTokenizer's [`~SiglipTokenizer.__call__`] if `text` is not `None` to encode
+        the text. To prepare the image(s), this method forwards the `images` argument to
+        SiglipImageProcessor's [`~SiglipImageProcessor.__call__`] if `images` is not `None`. Please refer to the doctsring
+        of the above two methods for more information.
+
+        Args:
+            text (`str`, `List[str]`, `List[List[str]]`):
+                The sequence or batch of sequences to be encoded. Each sequence can be a string or a list of strings
+                (pretokenized string). If the sequences are provided as list of strings (pretokenized), you must set
+                `is_split_into_words=True` (to lift the ambiguity with a batch of sequences).
+            images (`PIL.Image.Image`, `np.ndarray`, `torch.Tensor`, `List[PIL.Image.Image]`, `List[np.ndarray]`, `List[torch.Tensor]`):
+                The image or batch of images to be prepared. Each image can be a PIL image, NumPy array or PyTorch
+                tensor. Both channels-first and channels-last formats are supported.
+            padding (`bool`, `str` or [`~utils.PaddingStrategy`], *optional*, defaults to `False`):
+                Select a strategy to pad the returned sequences (according to the model's padding side and padding
+                index) among:
+                - `True` or `'longest'`: Pad to the longest sequence in the batch (or no padding if only a single
+                  sequence if provided).
+                - `'max_length'`: Pad to a maximum length specified with the argument `max_length` or to the maximum
+                  acceptable input length for the model if that argument is not provided.
+                - `False` or `'do_not_pad'` (default): No padding (i.e., can output a batch with sequences of different
+                  lengths).
+            max_length (`int`, *optional*):
+                Maximum length of the returned list and optionally padding length (see above).
+            truncation (`bool`, *optional*):
+                Activates truncation to cut input sequences longer than `max_length` to `max_length`.
+            return_tensors (`str` or [`~utils.TensorType`], *optional*):
+                If set, will return tensors of a particular framework. Acceptable values are:
+
+                - `'tf'`: Return TensorFlow `tf.constant` objects.
+                - `'pt'`: Return PyTorch `torch.Tensor` objects.
+                - `'np'`: Return NumPy `np.ndarray` objects.
+                - `'jax'`: Return JAX `jnp.ndarray` objects.
+
+        Returns:
+            [`BatchFeature`]: A [`BatchFeature`] with the following fields:
+
+            - **input_ids** -- List of token ids to be fed to a model. Returned when `text` is not `None`.
+            - **attention_mask** -- List of indices specifying which tokens should be attended to by the model (when
+              `return_attention_mask=True` or if *"attention_mask"* is in `self.model_input_names` and if `text` is not
+              `None`).
+            - **pixel_values** -- Pixel values to be fed to a model. Returned when `images` is not `None`.
+        """
+
+        if text is None and images is None:
+            raise ValueError("You have to specify either text or images. Both cannot be none.")
+
+        if text is not None:
+            encoding = self.tokenizer(
+                text, return_tensors=return_tensors, padding=padding, truncation=truncation, max_length=max_length
+            )
+
+        if images is not None:
+            image_features = self.image_processor(images, return_tensors=return_tensors)
+
+        if text is not None and images is not None:
+            encoding["pixel_values"] = image_features.pixel_values
+            return encoding
+        elif text is not None:
+            return encoding
+        else:
+            return BatchFeature(data=dict(**image_features), tensor_type=return_tensors)
+
+    def decode(self, *args, **kwargs):
+        """
+        This method forwards all its arguments to SiglipTokenizer's [`~PreTrainedTokenizer.decode`]. Please refer to
+        the docstring of this method for more information.
+        """
+        return self.tokenizer.decode(*args, **kwargs)
+
+    def batch_decode(self, *args, **kwargs):
+        """
+        This method forwards all its arguments to SiglipTokenizer's [`~PreTrainedTokenizer.batch_decode`]. Please
+        refer to the docstring of this method for more information.
+        """
+        return self.tokenizer.batch_decode(*args, **kwargs)
+
+    @property
+    # Copied from transformers.models.clip.processing_clip.CLIPProcessor.model_input_names with CLIP->Siglip, T5->Siglip
+    def model_input_names(self):
+        tokenizer_input_names = self.tokenizer.model_input_names
+        image_processor_input_names = self.image_processor.model_input_names
+        return list(dict.fromkeys(tokenizer_input_names + image_processor_input_names))

modeling/siglip/tokenization_siglip.py

@@ -1,364 +1,364 @@
-# Copyright 2024 The HuggingFace Inc. team.
-# SPDX-License-Identifier: Apache-2.0
-
-"""Tokenization class for SigLIP model."""
-
-import os
-import re
-import string
-import warnings
-from shutil import copyfile
-from typing import TYPE_CHECKING, Any, Dict, List, Optional, Tuple
-
-import sentencepiece as spm
-
-from transformers.convert_slow_tokenizer import import_protobuf
-from transformers.tokenization_utils import PreTrainedTokenizer
-from transformers.tokenization_utils_base import AddedToken
-
-
-if TYPE_CHECKING:
-    from transformers.tokenization_utils_base import TextInput
-from transformers.utils import logging, requires_backends
-
-
-logger = logging.get_logger(__name__)
-
-VOCAB_FILES_NAMES = {"vocab_file": "spiece.model"}
-
-
-SPIECE_UNDERLINE = "▁"
-
-
-class SiglipTokenizer(PreTrainedTokenizer):
-    """
-    Construct a Siglip tokenizer. Based on [SentencePiece](https://github.com/google/sentencepiece).
-
-    This tokenizer inherits from [`PreTrainedTokenizer`] which contains most of the main methods. Users should refer to
-    this superclass for more information regarding those methods.
-
-    Args:
-        vocab_file (`str`):
-            [SentencePiece](https://github.com/google/sentencepiece) file (generally has a *.spm* extension) that
-            contains the vocabulary necessary to instantiate a tokenizer.
-        eos_token (`str`, *optional*, defaults to `"</s>"`):
-            The end of sequence token.
-        unk_token (`str`, *optional*, defaults to `"<unk>"`):
-            The unknown token. A token that is not in the vocabulary cannot be converted to an ID and is set to be this
-            token instead.
-        pad_token (`str`, *optional*, defaults to `"</s>"`):
-            The token used for padding, for example when batching sequences of different lengths.
-        additional_special_tokens (`List[str]`, *optional*):
-            Additional special tokens used by the tokenizer.
-        sp_model_kwargs (`dict`, *optional*):
-            Will be passed to the `SentencePieceProcessor.__init__()` method. The [Python wrapper for
-            SentencePiece](https://github.com/google/sentencepiece/tree/master/python) can be used, among other things,
-            to set:
-
-            - `enable_sampling`: Enable subword regularization.
-            - `nbest_size`: Sampling parameters for unigram. Invalid for BPE-Dropout.
-
-              - `nbest_size = {0,1}`: No sampling is performed.
-              - `nbest_size > 1`: samples from the nbest_size results.
-              - `nbest_size < 0`: assuming that nbest_size is infinite and samples from the all hypothesis (lattice)
-                using forward-filtering-and-backward-sampling algorithm.
-
-            - `alpha`: Smoothing parameter for unigram sampling, and dropout probability of merge operations for
-              BPE-dropout.
-        model_max_length (`int`, *optional*, defaults to 64):
-            The maximum length (in number of tokens) for model inputs.
-        do_lower_case (`bool`, *optional*, defaults to `True`):
-            Whether or not to lowercase the input when tokenizing.
-    """
-
-    vocab_files_names = VOCAB_FILES_NAMES
-    model_input_names = ["input_ids", "attention_mask"]
-
-    def __init__(
-        self,
-        vocab_file,
-        eos_token="</s>",
-        unk_token="<unk>",
-        pad_token="</s>",
-        additional_special_tokens=None,
-        sp_model_kwargs: Optional[Dict[str, Any]] = None,
-        model_max_length=64,
-        do_lower_case=True,
-        **kwargs,
-    ) -> None:
-        requires_backends(self, "protobuf")
-
-        pad_token = (
-            AddedToken(pad_token, rstrip=True, lstrip=True, normalized=False, special=True)
-            if isinstance(pad_token, str)
-            else pad_token
-        )
-        unk_token = (
-            AddedToken(unk_token, rstrip=True, lstrip=True, normalized=False, special=True)
-            if isinstance(unk_token, str)
-            else unk_token
-        )
-        eos_token = (
-            AddedToken(eos_token, rstrip=True, lstrip=True, normalized=False, special=True)
-            if isinstance(eos_token, str)
-            else eos_token
-        )
-
-        self.sp_model_kwargs = {} if sp_model_kwargs is None else sp_model_kwargs
-
-        self.do_lower_case = do_lower_case
-        self.vocab_file = vocab_file
-
-        self.sp_model = self.get_spm_processor()
-        self.vocab_file = vocab_file
-
-        super().__init__(
-            eos_token=eos_token,
-            unk_token=unk_token,
-            pad_token=pad_token,
-            additional_special_tokens=additional_special_tokens,
-            sp_model_kwargs=self.sp_model_kwargs,
-            model_max_length=model_max_length,
-            do_lower_case=do_lower_case,
-            **kwargs,
-        )
-
-    def get_spm_processor(self):
-        tokenizer = spm.SentencePieceProcessor(**self.sp_model_kwargs)
-        with open(self.vocab_file, "rb") as f:
-            sp_model = f.read()
-            model_pb2 = import_protobuf()
-            model = model_pb2.ModelProto.FromString(sp_model)
-            normalizer_spec = model_pb2.NormalizerSpec()
-            normalizer_spec.add_dummy_prefix = False
-            model.normalizer_spec.MergeFrom(normalizer_spec)
-            sp_model = model.SerializeToString()
-            tokenizer.LoadFromSerializedProto(sp_model)
-        return tokenizer
-
-    @property
-    # Copied from transformers.models.t5.tokenization_t5.T5Tokenizer.vocab_size
-    def vocab_size(self):
-        return self.sp_model.get_piece_size()
-
-    # Copied from transformers.models.t5.tokenization_t5.T5Tokenizer.get_vocab
-    def get_vocab(self):
-        vocab = {self.convert_ids_to_tokens(i): i for i in range(self.vocab_size)}
-        vocab.update(self.added_tokens_encoder)
-        return vocab
-
-    # Copied from transformers.models.t5.tokenization_t5.T5Tokenizer.get_special_tokens_mask
-    def get_special_tokens_mask(
-        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None, already_has_special_tokens: bool = False
-    ) -> List[int]:
-        """
-        Retrieve sequence ids from a token list that has no special tokens added. This method is called when adding
-        special tokens using the tokenizer `prepare_for_model` method.
-
-        Args:
-            token_ids_0 (`List[int]`):
-                List of IDs.
-            token_ids_1 (`List[int]`, *optional*):
-                Optional second list of IDs for sequence pairs.
-            already_has_special_tokens (`bool`, *optional*, defaults to `False`):
-                Whether or not the token list is already formatted with special tokens for the model.
-
-        Returns:
-            `List[int]`: A list of integers in the range [0, 1]: 1 for a special token, 0 for a sequence token.
-        """
-        if already_has_special_tokens:
-            return super().get_special_tokens_mask(
-                token_ids_0=token_ids_0, token_ids_1=token_ids_1, already_has_special_tokens=True
-            )
-
-        # normal case: some special tokens
-        if token_ids_1 is None:
-            return ([0] * len(token_ids_0)) + [1]
-        return ([0] * len(token_ids_0)) + [1] + ([0] * len(token_ids_1)) + [1]
-
-    # Copied from transformers.models.t5.tokenization_t5.T5Tokenizer._add_eos_if_not_present
-    def _add_eos_if_not_present(self, token_ids: List[int]) -> List[int]:
-        """Do not add eos again if user already added it."""
-        if len(token_ids) > 0 and token_ids[-1] == self.eos_token_id:
-            warnings.warn(
-                f"This sequence already has {self.eos_token}. In future versions this behavior may lead to duplicated"
-                " eos tokens being added."
-            )
-            return token_ids
-        else:
-            return token_ids + [self.eos_token_id]
-
-    # Copied from transformers.models.t5.tokenization_t5.T5Tokenizer.create_token_type_ids_from_sequences
-    def create_token_type_ids_from_sequences(
-        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
-    ) -> List[int]:
-        """
-        Create a mask from the two sequences passed to be used in a sequence-pair classification task. T5 does not make
-        use of token type ids, therefore a list of zeros is returned.
-
-        Args:
-            token_ids_0 (`List[int]`):
-                List of IDs.
-            token_ids_1 (`List[int]`, *optional*):
-                Optional second list of IDs for sequence pairs.
-
-        Returns:
-            `List[int]`: List of zeros.
-        """
-        eos = [self.eos_token_id]
-
-        if token_ids_1 is None:
-            return len(token_ids_0 + eos) * [0]
-        return len(token_ids_0 + eos + token_ids_1 + eos) * [0]
-
-    # Copied from transformers.models.t5.tokenization_t5.T5Tokenizer.build_inputs_with_special_tokens
-    def build_inputs_with_special_tokens(
-        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
-    ) -> List[int]:
-        """
-        Build model inputs from a sequence or a pair of sequence for sequence classification tasks by concatenating and
-        adding special tokens. A sequence has the following format:
-
-        - single sequence: `X </s>`
-        - pair of sequences: `A </s> B </s>`
-
-        Args:
-            token_ids_0 (`List[int]`):
-                List of IDs to which the special tokens will be added.
-            token_ids_1 (`List[int]`, *optional*):
-                Optional second list of IDs for sequence pairs.
-
-        Returns:
-            `List[int]`: List of [input IDs](../glossary#input-ids) with the appropriate special tokens.
-        """
-        token_ids_0 = self._add_eos_if_not_present(token_ids_0)
-        if token_ids_1 is None:
-            return token_ids_0
-        else:
-            token_ids_1 = self._add_eos_if_not_present(token_ids_1)
-            return token_ids_0 + token_ids_1
-
-    # Copied from transformers.models.t5.tokenization_t5.T5Tokenizer.__getstate__
-    def __getstate__(self):
-        state = self.__dict__.copy()
-        state["sp_model"] = None
-        return state
-
-    # Copied from transformers.models.t5.tokenization_t5.T5Tokenizer.__setstate__
-    def __setstate__(self, d):
-        self.__dict__ = d
-
-        # for backward compatibility
-        if not hasattr(self, "sp_model_kwargs"):
-            self.sp_model_kwargs = {}
-
-        self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
-        self.sp_model.Load(self.vocab_file)
-
-    def remove_punctuation(self, text: str) -> str:
-        return text.translate(str.maketrans("", "", string.punctuation))
-
-    # source: https://github.com/google-research/big_vision/blob/3b8e5ab6ad4f96e32b32826f9e1b8fd277914f9c/big_vision/evaluators/proj/image_text/prompt_engineering.py#L94
-    def canonicalize_text(self, text, *, keep_punctuation_exact_string=None):
-        """Returns canonicalized `text` (puncuation removed).
-
-        Args:
-            text (`str`):
-                String to be canonicalized.
-            keep_punctuation_exact_string (`str`, *optional*):
-                If provided, then this exact string is kept. For example providing '{}' will keep any occurrences of '{}'
-                (but will still remove '{' and '}' that appear separately).
-        """
-        if keep_punctuation_exact_string:
-            text = keep_punctuation_exact_string.join(
-                self.remove_punctuation(part) for part in text.split(keep_punctuation_exact_string)
-            )
-        else:
-            text = self.remove_punctuation(text)
-        text = re.sub(r"\s+", " ", text)
-        text = text.strip()
-
-        return text
-
-    def tokenize(self, text: "TextInput", add_special_tokens=False, **kwargs) -> List[str]:
-        """
-        Converts a string to a list of tokens.
-        """
-        tokens = super().tokenize(SPIECE_UNDERLINE + text.replace(SPIECE_UNDERLINE, " "), **kwargs)
-
-        if len(tokens) > 1 and tokens[0] == SPIECE_UNDERLINE and tokens[1] in self.all_special_tokens:
-            tokens = tokens[1:]
-        return tokens
-
-    @property
-    # Copied from transformers.models.t5.tokenization_t5.T5Tokenizer.unk_token_length
-    def unk_token_length(self):
-        return len(self.sp_model.encode(str(self.unk_token)))
-
-    def _tokenize(self, text, **kwargs):
-        """
-        Returns a tokenized string.
-
-        We de-activated the `add_dummy_prefix` option, thus the sentencepiece internals will always strip any
-        SPIECE_UNDERLINE.
-
-        For example: `self.sp_model.encode(f"{SPIECE_UNDERLINE}Hey", out_type = str)` will give `['H', 'e', 'y']` instead of `['▁He', 'y']`.
-
-        Thus we always encode `f"{unk_token}text"` and strip the `unk_token`. Here is an example with `unk_token = "<unk>"` and `unk_token_length = 4`.
-        `self.tokenizer.sp_model.encode("<unk> Hey", out_type = str)[4:]`.
-        """
-        text = self.canonicalize_text(text, keep_punctuation_exact_string=None)
-        tokens = self.sp_model.encode(text, out_type=str)
-
-        # 1. Encode string + prefix ex: "<unk> Hey"
-        tokens = self.sp_model.encode(self.unk_token + text, out_type=str)
-        # 2. Remove self.unk_token from ['<','unk','>', '▁Hey']
-        return tokens[self.unk_token_length :] if len(tokens) >= self.unk_token_length else tokens
-
-    # Copied from transformers.models.t5.tokenization_t5.T5Tokenizer._convert_token_to_id
-    def _convert_token_to_id(self, token):
-        """Converts a token (str) in an id using the vocab."""
-        return self.sp_model.piece_to_id(token)
-
-    # Copied from transformers.models.t5.tokenization_t5.T5Tokenizer._convert_id_to_token
-    def _convert_id_to_token(self, index):
-        """Converts an index (integer) in a token (str) using the vocab."""
-        token = self.sp_model.IdToPiece(index)
-        return token
-
-    def convert_tokens_to_string(self, tokens):
-        """Converts a sequence of tokens (string) in a single string."""
-        current_sub_tokens = []
-        out_string = ""
-        prev_is_special = False
-        for token in tokens:
-            # make sure that special tokens are not decoded using sentencepiece model
-            if token in self.all_special_tokens:
-                if not prev_is_special:
-                    out_string += " "
-                out_string += self.sp_model.decode(current_sub_tokens) + token
-                prev_is_special = True
-                current_sub_tokens = []
-            else:
-                current_sub_tokens.append(token)
-                prev_is_special = False
-        out_string += self.sp_model.decode(current_sub_tokens)
-        return out_string.strip()
-
-    # Copied from transformers.models.t5.tokenization_t5.T5Tokenizer.save_vocabulary
-    def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> Tuple[str]:
-        if not os.path.isdir(save_directory):
-            logger.error(f"Vocabulary path ({save_directory}) should be a directory")
-            return
-        out_vocab_file = os.path.join(
-            save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["vocab_file"]
-        )
-
-        if os.path.abspath(self.vocab_file) != os.path.abspath(out_vocab_file) and os.path.isfile(self.vocab_file):
-            copyfile(self.vocab_file, out_vocab_file)
-        elif not os.path.isfile(self.vocab_file):
-            with open(out_vocab_file, "wb") as fi:
-                content_spiece_model = self.sp_model.serialized_model_proto()
-                fi.write(content_spiece_model)
-
-        return (out_vocab_file,)
+# Copyright 2024 The HuggingFace Inc. team.
+# SPDX-License-Identifier: Apache-2.0
+
+"""Tokenization class for SigLIP model."""
+
+import os
+import re
+import string
+import warnings
+from shutil import copyfile
+from typing import TYPE_CHECKING, Any, Dict, List, Optional, Tuple
+
+import sentencepiece as spm
+
+from transformers.convert_slow_tokenizer import import_protobuf
+from transformers.tokenization_utils import PreTrainedTokenizer
+from transformers.tokenization_utils_base import AddedToken
+
+
+if TYPE_CHECKING:
+    from transformers.tokenization_utils_base import TextInput
+from transformers.utils import logging, requires_backends
+
+
+logger = logging.get_logger(__name__)
+
+VOCAB_FILES_NAMES = {"vocab_file": "spiece.model"}
+
+
+SPIECE_UNDERLINE = "▁"
+
+
+class SiglipTokenizer(PreTrainedTokenizer):
+    """
+    Construct a Siglip tokenizer. Based on [SentencePiece](https://github.com/google/sentencepiece).
+
+    This tokenizer inherits from [`PreTrainedTokenizer`] which contains most of the main methods. Users should refer to
+    this superclass for more information regarding those methods.
+
+    Args:
+        vocab_file (`str`):
+            [SentencePiece](https://github.com/google/sentencepiece) file (generally has a *.spm* extension) that
+            contains the vocabulary necessary to instantiate a tokenizer.
+        eos_token (`str`, *optional*, defaults to `"</s>"`):
+            The end of sequence token.
+        unk_token (`str`, *optional*, defaults to `"<unk>"`):
+            The unknown token. A token that is not in the vocabulary cannot be converted to an ID and is set to be this
+            token instead.
+        pad_token (`str`, *optional*, defaults to `"</s>"`):
+            The token used for padding, for example when batching sequences of different lengths.
+        additional_special_tokens (`List[str]`, *optional*):
+            Additional special tokens used by the tokenizer.
+        sp_model_kwargs (`dict`, *optional*):
+            Will be passed to the `SentencePieceProcessor.__init__()` method. The [Python wrapper for
+            SentencePiece](https://github.com/google/sentencepiece/tree/master/python) can be used, among other things,
+            to set:
+
+            - `enable_sampling`: Enable subword regularization.
+            - `nbest_size`: Sampling parameters for unigram. Invalid for BPE-Dropout.
+
+              - `nbest_size = {0,1}`: No sampling is performed.
+              - `nbest_size > 1`: samples from the nbest_size results.
+              - `nbest_size < 0`: assuming that nbest_size is infinite and samples from the all hypothesis (lattice)
+                using forward-filtering-and-backward-sampling algorithm.
+
+            - `alpha`: Smoothing parameter for unigram sampling, and dropout probability of merge operations for
+              BPE-dropout.
+        model_max_length (`int`, *optional*, defaults to 64):
+            The maximum length (in number of tokens) for model inputs.
+        do_lower_case (`bool`, *optional*, defaults to `True`):
+            Whether or not to lowercase the input when tokenizing.
+    """
+
+    vocab_files_names = VOCAB_FILES_NAMES
+    model_input_names = ["input_ids", "attention_mask"]
+
+    def __init__(
+        self,
+        vocab_file,
+        eos_token="</s>",
+        unk_token="<unk>",
+        pad_token="</s>",
+        additional_special_tokens=None,
+        sp_model_kwargs: Optional[Dict[str, Any]] = None,
+        model_max_length=64,
+        do_lower_case=True,
+        **kwargs,
+    ) -> None:
+        requires_backends(self, "protobuf")
+
+        pad_token = (
+            AddedToken(pad_token, rstrip=True, lstrip=True, normalized=False, special=True)
+            if isinstance(pad_token, str)
+            else pad_token
+        )
+        unk_token = (
+            AddedToken(unk_token, rstrip=True, lstrip=True, normalized=False, special=True)
+            if isinstance(unk_token, str)
+            else unk_token
+        )
+        eos_token = (
+            AddedToken(eos_token, rstrip=True, lstrip=True, normalized=False, special=True)
+            if isinstance(eos_token, str)
+            else eos_token
+        )
+
+        self.sp_model_kwargs = {} if sp_model_kwargs is None else sp_model_kwargs
+
+        self.do_lower_case = do_lower_case
+        self.vocab_file = vocab_file
+
+        self.sp_model = self.get_spm_processor()
+        self.vocab_file = vocab_file
+
+        super().__init__(
+            eos_token=eos_token,
+            unk_token=unk_token,
+            pad_token=pad_token,
+            additional_special_tokens=additional_special_tokens,
+            sp_model_kwargs=self.sp_model_kwargs,
+            model_max_length=model_max_length,
+            do_lower_case=do_lower_case,
+            **kwargs,
+        )
+
+    def get_spm_processor(self):
+        tokenizer = spm.SentencePieceProcessor(**self.sp_model_kwargs)
+        with open(self.vocab_file, "rb") as f:
+            sp_model = f.read()
+            model_pb2 = import_protobuf()
+            model = model_pb2.ModelProto.FromString(sp_model)
+            normalizer_spec = model_pb2.NormalizerSpec()
+            normalizer_spec.add_dummy_prefix = False
+            model.normalizer_spec.MergeFrom(normalizer_spec)
+            sp_model = model.SerializeToString()
+            tokenizer.LoadFromSerializedProto(sp_model)
+        return tokenizer
+
+    @property
+    # Copied from transformers.models.t5.tokenization_t5.T5Tokenizer.vocab_size
+    def vocab_size(self):
+        return self.sp_model.get_piece_size()
+
+    # Copied from transformers.models.t5.tokenization_t5.T5Tokenizer.get_vocab
+    def get_vocab(self):
+        vocab = {self.convert_ids_to_tokens(i): i for i in range(self.vocab_size)}
+        vocab.update(self.added_tokens_encoder)
+        return vocab
+
+    # Copied from transformers.models.t5.tokenization_t5.T5Tokenizer.get_special_tokens_mask
+    def get_special_tokens_mask(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None, already_has_special_tokens: bool = False
+    ) -> List[int]:
+        """
+        Retrieve sequence ids from a token list that has no special tokens added. This method is called when adding
+        special tokens using the tokenizer `prepare_for_model` method.
+
+        Args:
+            token_ids_0 (`List[int]`):
+                List of IDs.
+            token_ids_1 (`List[int]`, *optional*):
+                Optional second list of IDs for sequence pairs.
+            already_has_special_tokens (`bool`, *optional*, defaults to `False`):
+                Whether or not the token list is already formatted with special tokens for the model.
+
+        Returns:
+            `List[int]`: A list of integers in the range [0, 1]: 1 for a special token, 0 for a sequence token.
+        """
+        if already_has_special_tokens:
+            return super().get_special_tokens_mask(
+                token_ids_0=token_ids_0, token_ids_1=token_ids_1, already_has_special_tokens=True
+            )
+
+        # normal case: some special tokens
+        if token_ids_1 is None:
+            return ([0] * len(token_ids_0)) + [1]
+        return ([0] * len(token_ids_0)) + [1] + ([0] * len(token_ids_1)) + [1]
+
+    # Copied from transformers.models.t5.tokenization_t5.T5Tokenizer._add_eos_if_not_present
+    def _add_eos_if_not_present(self, token_ids: List[int]) -> List[int]:
+        """Do not add eos again if user already added it."""
+        if len(token_ids) > 0 and token_ids[-1] == self.eos_token_id:
+            warnings.warn(
+                f"This sequence already has {self.eos_token}. In future versions this behavior may lead to duplicated"
+                " eos tokens being added."
+            )
+            return token_ids
+        else:
+            return token_ids + [self.eos_token_id]
+
+    # Copied from transformers.models.t5.tokenization_t5.T5Tokenizer.create_token_type_ids_from_sequences
+    def create_token_type_ids_from_sequences(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
+    ) -> List[int]:
+        """
+        Create a mask from the two sequences passed to be used in a sequence-pair classification task. T5 does not make
+        use of token type ids, therefore a list of zeros is returned.
+
+        Args:
+            token_ids_0 (`List[int]`):
+                List of IDs.
+            token_ids_1 (`List[int]`, *optional*):
+                Optional second list of IDs for sequence pairs.
+
+        Returns:
+            `List[int]`: List of zeros.
+        """
+        eos = [self.eos_token_id]
+
+        if token_ids_1 is None:
+            return len(token_ids_0 + eos) * [0]
+        return len(token_ids_0 + eos + token_ids_1 + eos) * [0]
+
+    # Copied from transformers.models.t5.tokenization_t5.T5Tokenizer.build_inputs_with_special_tokens
+    def build_inputs_with_special_tokens(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
+    ) -> List[int]:
+        """
+        Build model inputs from a sequence or a pair of sequence for sequence classification tasks by concatenating and
+        adding special tokens. A sequence has the following format:
+
+        - single sequence: `X </s>`
+        - pair of sequences: `A </s> B </s>`
+
+        Args:
+            token_ids_0 (`List[int]`):
+                List of IDs to which the special tokens will be added.
+            token_ids_1 (`List[int]`, *optional*):
+                Optional second list of IDs for sequence pairs.
+
+        Returns:
+            `List[int]`: List of [input IDs](../glossary#input-ids) with the appropriate special tokens.
+        """
+        token_ids_0 = self._add_eos_if_not_present(token_ids_0)
+        if token_ids_1 is None:
+            return token_ids_0
+        else:
+            token_ids_1 = self._add_eos_if_not_present(token_ids_1)
+            return token_ids_0 + token_ids_1
+
+    # Copied from transformers.models.t5.tokenization_t5.T5Tokenizer.__getstate__
+    def __getstate__(self):
+        state = self.__dict__.copy()
+        state["sp_model"] = None
+        return state
+
+    # Copied from transformers.models.t5.tokenization_t5.T5Tokenizer.__setstate__
+    def __setstate__(self, d):
+        self.__dict__ = d
+
+        # for backward compatibility
+        if not hasattr(self, "sp_model_kwargs"):
+            self.sp_model_kwargs = {}
+
+        self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
+        self.sp_model.Load(self.vocab_file)
+
+    def remove_punctuation(self, text: str) -> str:
+        return text.translate(str.maketrans("", "", string.punctuation))
+
+    # source: https://github.com/google-research/big_vision/blob/3b8e5ab6ad4f96e32b32826f9e1b8fd277914f9c/big_vision/evaluators/proj/image_text/prompt_engineering.py#L94
+    def canonicalize_text(self, text, *, keep_punctuation_exact_string=None):
+        """Returns canonicalized `text` (puncuation removed).
+
+        Args:
+            text (`str`):
+                String to be canonicalized.
+            keep_punctuation_exact_string (`str`, *optional*):
+                If provided, then this exact string is kept. For example providing '{}' will keep any occurrences of '{}'
+                (but will still remove '{' and '}' that appear separately).
+        """
+        if keep_punctuation_exact_string:
+            text = keep_punctuation_exact_string.join(
+                self.remove_punctuation(part) for part in text.split(keep_punctuation_exact_string)
+            )
+        else:
+            text = self.remove_punctuation(text)
+        text = re.sub(r"\s+", " ", text)
+        text = text.strip()
+
+        return text
+
+    def tokenize(self, text: "TextInput", add_special_tokens=False, **kwargs) -> List[str]:
+        """
+        Converts a string to a list of tokens.
+        """
+        tokens = super().tokenize(SPIECE_UNDERLINE + text.replace(SPIECE_UNDERLINE, " "), **kwargs)
+
+        if len(tokens) > 1 and tokens[0] == SPIECE_UNDERLINE and tokens[1] in self.all_special_tokens:
+            tokens = tokens[1:]
+        return tokens
+
+    @property
+    # Copied from transformers.models.t5.tokenization_t5.T5Tokenizer.unk_token_length
+    def unk_token_length(self):
+        return len(self.sp_model.encode(str(self.unk_token)))
+
+    def _tokenize(self, text, **kwargs):
+        """
+        Returns a tokenized string.
+
+        We de-activated the `add_dummy_prefix` option, thus the sentencepiece internals will always strip any
+        SPIECE_UNDERLINE.
+
+        For example: `self.sp_model.encode(f"{SPIECE_UNDERLINE}Hey", out_type = str)` will give `['H', 'e', 'y']` instead of `['▁He', 'y']`.
+
+        Thus we always encode `f"{unk_token}text"` and strip the `unk_token`. Here is an example with `unk_token = "<unk>"` and `unk_token_length = 4`.
+        `self.tokenizer.sp_model.encode("<unk> Hey", out_type = str)[4:]`.
+        """
+        text = self.canonicalize_text(text, keep_punctuation_exact_string=None)
+        tokens = self.sp_model.encode(text, out_type=str)
+
+        # 1. Encode string + prefix ex: "<unk> Hey"
+        tokens = self.sp_model.encode(self.unk_token + text, out_type=str)
+        # 2. Remove self.unk_token from ['<','unk','>', '▁Hey']
+        return tokens[self.unk_token_length :] if len(tokens) >= self.unk_token_length else tokens
+
+    # Copied from transformers.models.t5.tokenization_t5.T5Tokenizer._convert_token_to_id
+    def _convert_token_to_id(self, token):
+        """Converts a token (str) in an id using the vocab."""
+        return self.sp_model.piece_to_id(token)
+
+    # Copied from transformers.models.t5.tokenization_t5.T5Tokenizer._convert_id_to_token
+    def _convert_id_to_token(self, index):
+        """Converts an index (integer) in a token (str) using the vocab."""
+        token = self.sp_model.IdToPiece(index)
+        return token
+
+    def convert_tokens_to_string(self, tokens):
+        """Converts a sequence of tokens (string) in a single string."""
+        current_sub_tokens = []
+        out_string = ""
+        prev_is_special = False
+        for token in tokens:
+            # make sure that special tokens are not decoded using sentencepiece model
+            if token in self.all_special_tokens:
+                if not prev_is_special:
+                    out_string += " "
+                out_string += self.sp_model.decode(current_sub_tokens) + token
+                prev_is_special = True
+                current_sub_tokens = []
+            else:
+                current_sub_tokens.append(token)
+                prev_is_special = False
+        out_string += self.sp_model.decode(current_sub_tokens)
+        return out_string.strip()
+
+    # Copied from transformers.models.t5.tokenization_t5.T5Tokenizer.save_vocabulary
+    def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> Tuple[str]:
+        if not os.path.isdir(save_directory):
+            logger.error(f"Vocabulary path ({save_directory}) should be a directory")
+            return
+        out_vocab_file = os.path.join(
+            save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["vocab_file"]
+        )
+
+        if os.path.abspath(self.vocab_file) != os.path.abspath(out_vocab_file) and os.path.isfile(self.vocab_file):
+            copyfile(self.vocab_file, out_vocab_file)
+        elif not os.path.isfile(self.vocab_file):
+            with open(out_vocab_file, "wb") as fi:
+                content_spiece_model = self.sp_model.serialized_model_proto()
+                fi.write(content_spiece_model)
+
+        return (out_vocab_file,)