Upload app2.py
Browse files
app2.py
ADDED
@@ -0,0 +1,194 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import gradio as gr
|
3 |
+
import numpy as np
|
4 |
+
import spaces
|
5 |
+
import torch
|
6 |
+
import torchaudio
|
7 |
+
from generator import Segment, load_csm_1b
|
8 |
+
from huggingface_hub import hf_hub_download, login
|
9 |
+
from watermarking import watermark
|
10 |
+
import whisper
|
11 |
+
from transformers import AutoTokenizer, AutoModelForCausalLM
|
12 |
+
import logging
|
13 |
+
from transformers import GenerationConfig
|
14 |
+
|
15 |
+
# Configure logging
|
16 |
+
logging.basicConfig(level=logging.DEBUG, format='%(asctime)s - %(levelname)s - %(message)s')
|
17 |
+
|
18 |
+
# --- Authentication and Configuration --- (Moved BEFORE model loading)
|
19 |
+
try:
|
20 |
+
api_key = os.getenv("HF_TOKEN")
|
21 |
+
if not api_key:
|
22 |
+
raise ValueError("HF_TOKEN not found in environment variables.")
|
23 |
+
login(token=api_key)
|
24 |
+
|
25 |
+
CSM_1B_HF_WATERMARK = list(map(int, os.getenv("WATERMARK_KEY").split(" ")))
|
26 |
+
if not CSM_1B_HF_WATERMARK:
|
27 |
+
raise ValueError("WATERMARK_KEY not found or invalid in environment variables.")
|
28 |
+
|
29 |
+
gpu_timeout = int(os.getenv("GPU_TIMEOUT", 120))
|
30 |
+
except (ValueError, TypeError) as e:
|
31 |
+
logging.error(f"Configuration error: {e}")
|
32 |
+
raise
|
33 |
+
|
34 |
+
SPACE_INTRO_TEXT = """
|
35 |
+
# Sesame CSM 1B - Conversational Demo
|
36 |
+
|
37 |
+
This demo allows you to have a conversation with Sesame CSM 1B, leveraging Whisper for speech-to-text and Gemma for generating responses. This is an experimental integration and may require significant resources.
|
38 |
+
|
39 |
+
*Disclaimer: This demo relies on several large models. Expect longer processing times, and potential resource limitations.*
|
40 |
+
"""
|
41 |
+
|
42 |
+
# --- Model Loading --- (Moved INSIDE infer function)
|
43 |
+
|
44 |
+
# --- Constants --- (Constants can stay outside)
|
45 |
+
SPEAKER_ID = 0
|
46 |
+
MAX_CONTEXT_SEGMENTS = 1
|
47 |
+
MAX_GEMMA_LENGTH = 150
|
48 |
+
|
49 |
+
# --- Global Conversation History ---
|
50 |
+
conversation_history = []
|
51 |
+
|
52 |
+
# --- Helper Functions ---
|
53 |
+
|
54 |
+
def transcribe_audio(audio_path: str, whisper_model) -> str: # Pass whisper_model
|
55 |
+
try:
|
56 |
+
audio = whisper.load_audio(audio_path)
|
57 |
+
audio = whisper.pad_or_trim(audio)
|
58 |
+
result = whisper_model.transcribe(audio)
|
59 |
+
return result["text"]
|
60 |
+
except Exception as e:
|
61 |
+
logging.error(f"Whisper transcription error: {e}")
|
62 |
+
return "Error: Could not transcribe audio."
|
63 |
+
|
64 |
+
def generate_response(text: str, model_gemma, tokenizer_gemma, device) -> str: # Pass model and tokenizer
|
65 |
+
try:
|
66 |
+
# Gemma 3 chat template format
|
67 |
+
messages = [{"role": "user", "content": text}]
|
68 |
+
input = tokenizer_gemma.apply_chat_template(messages, return_tensors="pt").to(device)
|
69 |
+
generation_config = GenerationConfig(
|
70 |
+
max_new_tokens=MAX_GEMMA_LENGTH,
|
71 |
+
early_stopping=True,
|
72 |
+
)
|
73 |
+
|
74 |
+
generated_output = model_gemma.generate(input, generation_config=generation_config)
|
75 |
+
return tokenizer_gemma.decode(generated_output[0], skip_special_tokens=True)
|
76 |
+
|
77 |
+
#input_text = "Reapond to the users prompt: " + text
|
78 |
+
#input = tokenizer_gemma(input_text, return_tensors="pt").to(device)
|
79 |
+
#generated_output = model_gemma.generate(**input, max_length=MAX_GEMMA_LENGTH, early_stopping=True)
|
80 |
+
#return tokenizer_gemma.decode(generated_output[0], skip_special_tokens=True)
|
81 |
+
except Exception as e:
|
82 |
+
logging.error(f"Gemma response generation error: {e}")
|
83 |
+
return "I'm sorry, I encountered an error generating a response."
|
84 |
+
|
85 |
+
def load_audio(audio_path: str, generator) -> torch.Tensor: #Pass generator
|
86 |
+
try:
|
87 |
+
audio_tensor, sample_rate = torchaudio.load(audio_path)
|
88 |
+
audio_tensor = audio_tensor.mean(dim=0)
|
89 |
+
if sample_rate != generator.sample_rate:
|
90 |
+
audio_tensor = torchaudio.functional.resample(audio_tensor, orig_freq=sample_rate, new_freq=generator.sample_rate)
|
91 |
+
return audio_tensor
|
92 |
+
except Exception as e:
|
93 |
+
logging.error(f"Audio loading error: {e}")
|
94 |
+
raise gr.Error("Could not load or process the audio file.") from e
|
95 |
+
|
96 |
+
def clear_history():
|
97 |
+
global conversation_history
|
98 |
+
conversation_history = []
|
99 |
+
logging.info("Conversation history cleared.")
|
100 |
+
return "Conversation history cleared."
|
101 |
+
|
102 |
+
# --- Main Inference Function ---
|
103 |
+
|
104 |
+
@spaces.GPU(duration=gpu_timeout) # Decorator FIRST
|
105 |
+
def infer(user_audio) -> tuple[int, np.ndarray]:
|
106 |
+
# --- CUDA Availability Check (INSIDE infer) ---
|
107 |
+
if torch.cuda.is_available():
|
108 |
+
print(f"CUDA is available! Device count: {torch.cuda.device_count()}")
|
109 |
+
print(f"CUDA device name: {torch.cuda.get_device_name(0)}")
|
110 |
+
print(f"CUDA version: {torch.version.cuda}")
|
111 |
+
device = "cuda"
|
112 |
+
else:
|
113 |
+
print("CUDA is NOT available. Using CPU.") # Use CPU, don't raise
|
114 |
+
device = "cpu"
|
115 |
+
|
116 |
+
try:
|
117 |
+
# --- Model Loading (INSIDE infer, after device is set) ---
|
118 |
+
model_path = hf_hub_download(repo_id="sesame/csm-1b", filename="ckpt.pt")
|
119 |
+
generator = load_csm_1b(model_path, device)
|
120 |
+
logging.info("Sesame CSM 1B loaded successfully.")
|
121 |
+
|
122 |
+
whisper_model = whisper.load_model("small.en", device=device)
|
123 |
+
logging.info("Whisper model loaded successfully.")
|
124 |
+
|
125 |
+
tokenizer_gemma = AutoTokenizer.from_pretrained("google/gemma-3-1b-it")
|
126 |
+
model_gemma = AutoModelForCausalLM.from_pretrained("google/gemma-3-1b-it").to(device)
|
127 |
+
logging.info("Gemma 3 1B pt model loaded successfully.")
|
128 |
+
|
129 |
+
if not user_audio:
|
130 |
+
raise ValueError("No audio input received.")
|
131 |
+
return _infer(user_audio, generator, whisper_model, tokenizer_gemma, model_gemma, device) #Pass all models
|
132 |
+
except Exception as e:
|
133 |
+
logging.exception(f"Inference error: {e}")
|
134 |
+
raise gr.Error(f"An error occurred during processing: {e}")
|
135 |
+
|
136 |
+
def _infer(user_audio, generator, whisper_model, tokenizer_gemma, model_gemma, device) -> tuple[int, np.ndarray]:
|
137 |
+
global conversation_history
|
138 |
+
|
139 |
+
try:
|
140 |
+
user_text = transcribe_audio(user_audio, whisper_model) # Pass whisper_model
|
141 |
+
logging.info(f"User: {user_text}")
|
142 |
+
|
143 |
+
ai_text = generate_response(user_text, model_gemma, tokenizer_gemma, device) # Pass model and tokenizer
|
144 |
+
logging.info(f"AI: {ai_text}")
|
145 |
+
|
146 |
+
try:
|
147 |
+
ai_audio = generator.generate(
|
148 |
+
text=ai_text,
|
149 |
+
speaker=SPEAKER_ID,
|
150 |
+
context=conversation_history,
|
151 |
+
max_audio_length_ms=10_000,
|
152 |
+
)
|
153 |
+
logging.info("Audio generated successfully.")
|
154 |
+
except Exception as e:
|
155 |
+
logging.error(f"Sesame response generation error: {e}")
|
156 |
+
raise gr.Error(f"Sesame response generation error: {e}")
|
157 |
+
|
158 |
+
|
159 |
+
user_segment = Segment(speaker = 1, text = user_text, audio = load_audio(user_audio, generator)) #Pass Generator
|
160 |
+
ai_segment = Segment(speaker = SPEAKER_ID, text = ai_text, audio = ai_audio)
|
161 |
+
conversation_history.append(user_segment)
|
162 |
+
conversation_history.append(ai_segment)
|
163 |
+
|
164 |
+
if len(conversation_history) > MAX_CONTEXT_SEGMENTS:
|
165 |
+
conversation_history.pop(0)
|
166 |
+
|
167 |
+
audio_tensor, wm_sample_rate = watermark(
|
168 |
+
generator._watermarker, ai_audio, generator.sample_rate, CSM_1B_HF_WATERMARK
|
169 |
+
)
|
170 |
+
audio_tensor = torchaudio.functional.resample(
|
171 |
+
audio_tensor, orig_freq=wm_sample_rate, new_freq=generator.sample_rate
|
172 |
+
)
|
173 |
+
|
174 |
+
ai_audio_array = (audio_tensor * 32768).to(torch.int16).cpu().numpy()
|
175 |
+
return generator.sample_rate, ai_audio_array
|
176 |
+
|
177 |
+
except Exception as e:
|
178 |
+
logging.exception(f"Error in _infer: {e}")
|
179 |
+
raise gr.Error(f"An error occurred during processing: {e}")
|
180 |
+
|
181 |
+
# --- Gradio Interface ---
|
182 |
+
|
183 |
+
with gr.Blocks() as app:
|
184 |
+
gr.Markdown(SPACE_INTRO_TEXT)
|
185 |
+
audio_input = gr.Audio(label="Your Input", type="filepath")
|
186 |
+
audio_output = gr.Audio(label="AI Response")
|
187 |
+
clear_button = gr.Button("Clear Conversation History")
|
188 |
+
status_display = gr.Textbox(label="Status", visible=False)
|
189 |
+
|
190 |
+
btn = gr.Button("Generate Response")
|
191 |
+
btn.click(infer, inputs=[audio_input], outputs=[audio_output])
|
192 |
+
clear_button.click(clear_history, outputs=[status_display])
|
193 |
+
|
194 |
+
app.launch(ssr_mode=False, share=True)
|