Qwen3-vl-8b load frames too slow

[songbaiyang07-star]

When I use transformers code to evaluate videomme, especially for long video, it take long time to load 16 frames(far more than generate), even evaluate hour level videos it took more than 3 minutes just to load frames, the server memory was completely full, and it still hadn’t finished loading. In line

inputs = processor.apply_chat_template(
        messages,
        tokenize=True,
        add_generation_prompt=True,
        return_dict=True,
        return_tensors="pt",
        fps=None,
        num_frames=16,
    )

it takes long time and even the server memory was completely full, and it still hadn’t finished loading.

from transformers import Qwen3VLForConditionalGeneration, AutoProcessor
import os
import torch
import argparse
import math

def run_inference(args):
    model = Qwen3VLForConditionalGeneration.from_pretrained(
    args.model_path,
    dtype=torch.bfloat16,
    attn_implementation="flash_attention_2",
    device_map="auto",
)
    processor = AutoProcessor.from_pretrained(args.model_path,size={"longest_edge": 512*32*32, "shortest_edge": 256*32*32})
    processor = AutoProcessor.from_pretrained(args.model_path)

    prompt = "Select correct option to answer the question. The provided image arranges keyframes from a video in a grid view. \nQuestion: %s\nOptions:\n%s\nSelect the correct answer from the options. \nASSISTANT:\nAnswer:"
    option_prompt, options = get_option_prompt(candidates, version="v4")
    prompt = prompt % (question, option_prompt)
    video_path = os.path.join(args.video_dir, video_name)
    print('video_path', video_path)
    video_path = os.path.join(args.video_dir, video_name)
    print('video_path', video_path)

    messages = [
        {
            "role": "user",
            "content": [
                {
                    "type": "video",
                    "video": video_path,    
                },
                {"type": "text", "text": prompt},
            ],
        }
    ]
    print(messages)
    # 准备输入
    inputs = processor.apply_chat_template(
        messages,
        tokenize=True,
        add_generation_prompt=True,
        return_dict=True,
        return_tensors="pt",
        fps=None,
        num_frames=16,
    )
    # print(f"Shape of input_ids: {inputs['input_ids'].shape}") #torch.Size([1, 3602])
    # print(f"Shape of attention_mask: {inputs['attention_mask'].shape}")#torch.Size([1, 3602])
    for key, value in inputs.items():
        print(key,value.shape)
    '''
    Shape of input_ids: torch.Size([1, 12090])
    Shape of attention_mask: torch.Size([1, 12090])
    input_ids torch.Size([1, 12090])
    attention_mask torch.Size([1, 12090])
    pixel_values_videos torch.Size([45288, 1536])
    video_grid_thw torch.Size([1, 3])
    '''
    # print(f"Shape of pixel_values: {inputs['pixel_values'].shape}")#torch.Size([14308, 1176])

    do_sample = os.getenv("greedy", "false").lower() == "false"  # greedy=false -> 采样
    generation_kwargs = {
        "do_sample": do_sample,
        "top_p": float(os.getenv("top_p", 0.8)),
        "top_k": int(os.getenv("top_k", 20)),
        "temperature": float(os.getenv("temperature", 0.7)),
        "repetition_penalty": float(os.getenv("repetition_penalty", 1.0)),
        "max_new_tokens": int(os.getenv("out_seq_length", 512)),  # 16384 可能过大，这里给个更安全默认
    }

    # 推理生成输出
    generated_ids = model.generate(**inputs, **generation_kwargs)
    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
    )
    output_text = output_text[0].replace("In the image", "In the video")
            # print(output)
    sample_set["pred"] = output_text
    # print(output_text)
    ans_file.write(json.dumps(sample_set) + "\n")
    # exit(0)
    ans_file.close()

[jklj077]

Hi, please report this issue to https://github.com/QwenLM/Qwen3-VL. Thanks for your understanding.