jiangchengchengNLP/qwenvl-2.5-7B-Instruct-AWQ4INT

Load base model

model_name = "Qwen/Qwen2.5-VL-7B-Instruct"
config = AutoConfig.from_pretrained(model_name)
#it cost few minute to load model in float32, although config is to appoint bfloat16
model = Qwen2_5_VLForConditionalGeneration(config)
model.to(torch.bfloat16)
processor = AutoProcessor.from_pretrained(model_name)
# Apply AWQ
quantization_config_path = "./weights/AWQ_config.json"
quantization_weight_path = "./weights/AWQ_weights.pth"
res = apply_AWQ(model, quantization_config_path, quantization_weight_path=quantization_weight_path)

Inference

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image",
                "image": "<url id=\"cuq4ml2misdhuceigs30\" type=\"url\" status=\"failed\" title=\"\" wc=\"0\">https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg</url>",
            },
            {"type": "text", "text": "What does this photo show ?"},
        ],
    }
]

# Preparation for inference
text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
image_inputs, video_inputs = process_vision_info(messages)
inputs = processor(
    text=[text],
    images=image_inputs,
    videos=video_inputs,
    padding=True,
    return_tensors="pt",
)

# Inference: Generation of the output
device = "cuda"
inputs = inputs.to(device)
model.to(device)
model.eval()
generated_ids = model.generate(**inputs, max_new_tokens=128)
generated_ids_trimmed = [
    out_ids[len(in_ids):] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]
output_text = processor.batch_decode(
    generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
)
print(output_text)

Repository Structure AWQ_config.json: Configuration file for AWQ quantization.

quantization.py: Python script for converting the original model to the AWQ quantized version.

model.safetenors: state dict file.

Notes

The quantized model is optimized for memory efficiency but may have a strongly slower (about 3x slower) inference speed compared to the original model, I will try to identify the problem and provide a faster reasoning model at the appropriate time。

Ensure you have sufficient GPU memory (at least 6GB) to load and run the quantized model.