nm-testing/Mistral-Small-3.1-24B-Instruct-2503-FP8-dynamic

Evaluation

Server command:

vllm serve nm-testing/Mistral-Small-3.1-24B-Instruct-2503-FP8-dynamic

Eval command:

python -m eval.run eval_vllm --model_name nm-testing/Mistral-Small-3.1-24B-Instruct-2503-FP8-dynamic --url http://0.0.0.0:9000 --output_dir output/ --eval_name "chartqa"
Waiting for VLLM server to come online at http://0.0.0.0:9000/health ...
Timeout is 120s
Waiting for server (0s) ...
Server is up!
Loading lmms-lab/ChartQA [test]: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2500/2500 [00:11<00:00, 210.68it/s]
Querying model: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2500/2500 [06:53<00:00,  6.05it/s]
100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2500/2500 [00:00<00:00, 22477.08it/s]
================================================================================
Metrics:
{
    "explicit_prompt_relaxed_correctness": 0.8136,
    "anywhere_in_answer_relaxed_correctness": 0.8144
}
================================================================================

Creation

from transformers import AutoProcessor, AutoModelForImageTextToText

from llmcompressor.modifiers.quantization import QuantizationModifier
from llmcompressor.transformers import oneshot

MODEL_ID = "mistralai/Mistral-Small-3.1-24B-Instruct-2503"

# Load model.
model = AutoModelForImageTextToText.from_pretrained(
    MODEL_ID, device_map="auto", torch_dtype="auto"
)
processor = AutoProcessor.from_pretrained(MODEL_ID)

# Configure the quantization algorithm and scheme.
# In this case, we:
#   * quantize the weights to fp8 with per channel via ptq
#   * quantize the activations to fp8 with dynamic per token
recipe = QuantizationModifier(
    targets="Linear",
    scheme="FP8_DYNAMIC",
    ignore=["re:.*lm_head", "re:multi_modal_projector.*", "re:vision_tower.*"],
)

# Apply quantization and save to disk in compressed-tensors format.
SAVE_DIR = MODEL_ID.split("/")[1] + "-FP8-dynamic"
oneshot(model=model, recipe=recipe, output_dir=SAVE_DIR)
processor.save_pretrained(SAVE_DIR)

# Confirm generations of the quantized model look sane.
print("========== SAMPLE GENERATION ==============")
input_ids = processor(text="Hello my name is", return_tensors="pt").input_ids.to("cuda")
output = model.generate(input_ids, max_new_tokens=20)
print(processor.decode(output[0]))
print("==========================================")