Qwen2.5-VL-7B-Instruct_gemlite-ao_a16w4_gs_128_pack_32bit

This is an A16W4 HQQ quantized <a href="https://huggingface.co/Qwen/wen2.5-VL-7B-Instruct">Qwen2.5-VL-7B-Instruct</a> model, via <a href="https://github.com/pytorch/ao/">TorchAO</a> and <a href="https://github.com/mobiusml/gemlite/">GemLite</a> as a backend.

Usage

First, install the dependecies:

pip install torchao;
pip install git+https://github.com/mobiusml/gemlite.git;

from transformers import Qwen2_5_VLForConditionalGeneration, AutoTokenizer, AutoProcessor
from qwen_vl_utils import process_vision_info

model_id = "mobiuslabsgmbh/Qwen2.5-VL-7B-Instruct_gemlite-ao_a16w4_gs_128_pack_32bit"
model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    model_id, torch_dtype=torch.float16, device_map="cuda",
    #attn_implementation="flash_attention_2",
)

processor = AutoProcessor.from_pretrained(model_id)

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image",
                "image": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg",
            },
            {"type": "text", "text": "Describe this image."},
        ],
    }
]

# 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",
)
inputs = inputs.to("cuda")

# Inference: Generation of the output
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)

VLLM

import torch
from vllm import LLM
from vllm.sampling_params import SamplingParams

model_id = "mobiuslabsgmbh/Qwen2.5-VL-7B-Instruct_gemlite-ao_a16w4_gs_128_pack_32bit"
processor_args = {
	'limit_mm_per_prompt': {"image": 3}, 
	'mm_processor_kwargs': {"min_pixels": 28 * 28, "max_pixels": 1280 * 28 * 28},
	'disable_mm_preprocessor_cache': False,
}

llm = LLM(model=model_id, gpu_memory_utilization=0.9, dtype=torch.float16, max_model_len=4096, 
			max_num_batched_tokens=4096, **processor_args) 

sampling_params = SamplingParams(max_tokens=1024, temperature=0.5, repetition_penalty=1.1, ignore_eos=False)

messages = [{"content": "You are a helpful assistant", "role":"system"}, {"content":"Solve this equation x^2 + 1 = -1.", "role":"user"}]
outputs = llm.chat(messages, sampling_params, chat_template=llm.get_tokenizer().chat_template)
print(outputs[0].outputs[0].text)