TypeError: argument of type 'NoneType' is not iterable

[OgasTm]

Describe the bug
What the bug is, and how to reproduce, better with screenshots(描述bug以及复现过程，最好有截图)
I attempted to perform inference with Qwen 2.5 Omni 7B, and it appears that there is an EOS issue related to this model. My colleague uses the same script to do inference with Qwen2.5-VL-7B-Instruct and does not report this issue.

Your hardware and system info
Write your system info like CUDA version/system/GPU/torch version here(在这里给出硬件信息和系统信息，如CUDA版本，系统，GPU型号和torch版本等)

Additional context
Add any other context about the problem here(在这里补充其他信息)

[OgasTm]

torch : 2.7.1

[Jintao-Huang]

Please send me the version of ms-swift and the inference script.

[OgasTm]

ms-swift: 3.6.0

import os
import torch
import numpy as np

# set the directories/paths here!
################################################################################################
root_dir = '/mnt/.../'
best_chkpt_path = "/mnt/.../saved_models/Qwen2.5-Omni-7B/v0-20250704-094104/checkpoint-620"
model_name = "Qwen/Qwen2.5-Omni-7B"
seizure_feature_gt_fp = '/mnt/.../SeizureDataset_Annotation.csv'     # CSV file to read

inference_dir = root_dir + 'inference/'
inference_result_dir = inference_dir + f'result/'
inference_datasets_dir = inference_dir + f'datasets/'
os.makedirs(inference_result_dir, exist_ok=True)
os.makedirs(inference_datasets_dir, exist_ok=True)

inf_result_csv_fp = inference_dir + 'inference_result.csv'
model_cache_dir = root_dir + 'model_cache'
hf_cache_dir = os.path.join(model_cache_dir, 'huggingface')
modelscope_cache_dir = os.path.join(model_cache_dir, 'modelscope')
################################################################################################

# to do list:

try:
    from numpy.core.multiarray import _reconstruct
    torch.serialization.add_safe_globals([_reconstruct])
    print("Added numpy safe globals for checkpoint loading")
except ImportError:
    print("Warning: Could not import numpy _reconstruct. Checkpoint loading might fail.")

os.chdir(os.path.dirname(os.path.abspath(__file__)))

os.makedirs(hf_cache_dir, exist_ok=True)
os.makedirs(modelscope_cache_dir, exist_ok=True)

# set environment variables
os.environ['MODELSCOPE_CACHE'] = modelscope_cache_dir
os.environ['HF_HOME'] = hf_cache_dir
os.environ['PYTORCH_DISABLE_WEIGHTS_ONLY_LOAD'] = '1'
os.environ['CUDA_VISIBLE_DEVICES'] = '2,3'  # Set visible GPUs
os.environ['CUDA_LAUNCH_BLOCKING'] = '1'  # Enable synchronous CUDA operations for debugging
# os.environ['TORCH_CUDA_ALLOC_CONF'] = 'max_split_size_mb:128'
os.environ['TORCH_USE_CUDA_DSA'] = '1'  # Enable CUDA Dynamic Parallelism for better memory management

max_length = 9600
os.environ['MAX_PIXELS'] = f'{1920 * 1080}'
os.environ['MIN_PIXELS'] = f'{64 * 28 * 28}' 
os.environ['FPS'] = '1'
os.environ['FPS_MIN_FRAMES'] = '2' 
os.environ['FPS_MAX_FRAMES'] = '60' 


def run_inf(val_dataset_json_fp, save_fp):
    cmd = [
            "swift infer",
            f"--adapters {best_chkpt_path}",
            "--infer_backend pt",
            "--max_new_tokens 2048",
            f"--val_dataset {val_dataset_json_fp}",
            f"--result_path {save_fp}",
            "--temperature 0",
            # "--load_data_args true",
            "--max_batch_size 1",
        ]

    result = os.system(" ".join(cmd))
    return result

import pandas as pd 
def get_prompts():
    if os.path.exists(seizure_feature_gt_fp):
        df = pd.read_csv(seizure_feature_gt_fp)
    columns = df.columns.tolist()
    feature_names = columns.copy() 
    
    feature_names.remove('file_name')  
    feature_names.remove('label')
    feature_names.remove('start_time')
    feature_names.remove('end_time')

    prompts = {}
    # feature with missing prompt: ['continuous_intensity'] 
    feature_names.remove('continuous_intensity')
    #  "Does the patient in the video experience jerking of the entire body, including the arms, legs, and torso? Answer with 'yes' or 'no'. Do not include extra text in your output—only the answer."
   
    for feature in feature_names:
        if feature == 'blank_stare':
            prompts[feature] = "Does the patient exhibit a blank stare (vacant or unfocused gaze) at any time during the event? Answer with 'yes' or 'no'. Do not include extra text in your output—only the answer. "
        if feature == 'arm_flexion':
            prompts[feature] = "Does the patient maintain a sustained flexion of the arms at the elbows (i.e., holding them in a flexed position for a noticeable duration) during the seizure? Answer with 'yes' or 'no'. Do not include extra text in your output—only the answer."
        if feature == 'limb_movement_pattern':
            prompts[feature] = "Analyze the patient's upper limb movements in this video. Which of the following best describes the movements? \
                                    1.Thrashing/Flailing: Significant thrashing and flailing, somewhat asynchronous and variable. \
                                    2.Rhythmic Jerking: Stereotyped, rhythmic clonic jerking, potentially following a tonic phase. \
                                    3.Neither: The movements do not fit descriptions 1 or 2. \
                                    Please answer with 1, 2, or 3. Do not include extra text in your output—only the answer."

        if feature == 'hands_move_simultaneously':
            prompts[feature] = "Do the patient's hands start moving simultaneously? Answer with 'yes' or 'no'. "
        if feature == 'gender':
            prompts[feature] = "Please identify the gender of the patient in the video. Please answer with \"female\" or \"male\". Do not include extra text in your output—only the answer. "
        if feature == 'occur_during_sleep':
            prompts[feature] = "Please determine if this seizure event occurs while the patient is asleep. Answer with 'yes' or 'no'. Do not include extra text in your output—only the answer."
        if feature == 'ictal_vocalization':
            prompts[feature] = "Please check if the patient produces any vocalization (such as groaning, moaning, or screaming) during the event. Answer with 'yes' or 'no'. Do not include extra text in your output—only the answer. "
        if feature == 'close_eyes':
            prompts[feature] = "Do the patient's eyes remain consistently closed or mostly closed throughout the event? Answer with 'yes' or 'no'. Do not include extra text in your output—only the answer.  "
        if feature == 'eye_blinking':
            prompts[feature] = "Does the patient show repeated or rapid blinking of the eyes during the event? Answer with 'yes' or 'no'. Do not include extra text in your output—only the answer."
        if feature == 'tonic':
            prompts[feature] = "Please observe whether the patient has a prolonged, sustained muscle contraction (tonic phase). Answer with 'yes' or 'no'. Do not include extra text in your output—only the answer. "
        if feature == 'clonic':
            prompts[feature] = "Clonic movements typically involve repetitive, rhythmic jerking of muscles—marked by a contraction phase followed by relaxation, then repeating in a clear pattern. Please determine if the patient exhibits these repetitive, rhythmic jerking (clonic) movements, distinct from small or continuous trembling (tremor), at any point during the event. Answer with 'yes' or 'no'. Do not include extra text in your output—only the answer."
        if feature == 'arm_straightening':
            prompts[feature] = "Does the patient straighten or stiffen their arms (extended position)? Answer with 'yes' or 'no'. Do not include extra text in your output—only the answer."
        if feature == 'figure4':
            prompts[feature] = "Figure 4 refers to a specific posture or movement observed in a patient during a seizure, where one upper limb is extended (typically in a tonic stretch) while the other upper limb is flexed, forming a shape resembling the number 4. Please check if there is a 'figure 4' posture of the arms at any point. Answer with 'yes' or 'no'. Do not include extra text in your output—only the answer."
        if feature == 'oral_automatisms':
            prompts[feature] = "Does the patient exhibit repetitive mouth or tongue movements such as chewing, lip-smacking, or swallowing? Answer with 'yes' or 'no'. Do not include extra text in your output—only the answer."
        if feature == 'limb_automatisms':
            prompts[feature] = "Are there repetitive, purposeless limb movements (e.g., fumbling, picking, patting, cycling) observed? Answer with 'yes' or 'no'. Do not include extra text in your output—only the answer."
        if feature == 'face_pulling':
            prompts[feature] = "Does the patient's facial expression indicate grimacing or face-pulling movements? Answer with 'yes' or 'no'. Do not include extra text in your output—only the answer."
        if feature == 'face_twitching':
            prompts[feature] = "Are there small, involuntary twitches or jerks observed on the patient's face? Answer with 'yes' or 'no'. Do not include extra text in your output—only the answer."
        if feature == 'head_turning':
            prompts[feature] = "Does the patient forcibly or stiffly rotate their head to one side during the event?  Answer with 'yes' or 'no'. Do not include extra text in your output—only the answer."
        if feature == 'tremor':
            prompts[feature] = "Do you observe any rhythmic, trembling movement (tremor) in the patient's limbs or body? Answer with 'yes' or 'no'. Do not include extra text in your output—only the answer."
        if feature == 'pelvic thrusting':
            prompts[feature] = "Does the patient display any pelvic thrusting movements during the event? Answer with 'yes' or 'no'. Do not include extra text in your output—only the answer. "
        if feature == 'verbal_responsiveness':
            prompts[feature] = "Please determine if the patient is able to respond verbally or demonstrate any comprehension during the event. Answer with 'yes' or 'no', if no one asks the patient any questions, answer 'NA.' Do not include extra text in your output—only the answer."
        # if feature == 'start_time':
        #     prompts[feature] = "At what time does the seizure start in the video? provide the timestamp in the format MM:SS (minutes and seconds).Do not include extra text in your output—only the timestamps."
        # if feature == 'end_time':
        #     prompts[feature] = "At what time does the seizure end in the video? provide the timestamp in the format MM:SS (minutes and seconds).Do not include extra text in your output—only the timestamps."

    assert len(feature_names) == len(prompts), f"feature_names length {len(feature_names)} and prompt_list lengths {len(prompts)} does not match."
    return prompts

def get_features(prompt_dict):
    # swap the keys and values in the prompt_dict
    features = {}
    for feature, prompt in prompt_dict.items():
        features[prompt] = feature
    return features

# ================== Utility functions ==================
import csv
def append_to_csv(csv_fp, data, header=None):
    """
    Append the list 'data' to 'csv_file'. If 'csv_file' does not exist, write the 'header' row first.
    """
    file_exists = os.path.isfile(csv_fp)
    with open(csv_fp, mode='a', newline='', encoding='utf-8') as file:
        writer = csv.writer(file)
        if (not file_exists) and (header is not None):
            writer.writerow(header)
        writer.writerow(data)

def generate_message(prompt, video_path):
    entry = {
                "messages": [
                                {"role": "user", "content": f"<video>{prompt}"}, 
                            ],
                "videos":   video_path,
            }
    return entry

import json
import traceback
from tqdm import tqdm
def ExtractFeatureByVLM(video_paths, inf_result_csv_fp):
    """
    Extract features from one video by the VLM for each prompt in prompt_list,
    Return a list of extracted features in the same order as prompt_list.
    """
    
    infer_dataset = []
    # generate infer dataset
    for video_path in video_paths:
        if not os.path.exists(video_path):
            print(f"Video file {video_path} does not exist.")
            continue
        prompt_dict = get_prompts()  # Get the prompts for each feature
        feature_dict = get_features(prompt_dict)
        for feature in prompt_dict:
            prompt = prompt_dict[feature]
            entry = generate_message(prompt, video_path)
            infer_dataset.append(entry)
    # Save the infer dataset to a JSON file
    inf_dataset_fp = f'{inference_datasets_dir}/infer_ds.jsonl'
    if not os.path.exists(inf_dataset_fp):
        with open(inf_dataset_fp, 'w') as f:
            for item in infer_dataset:
                f.write(json.dumps(item) + '\n')

    max_retries = 3
    inference_result_fp = f"{inference_result_dir}/inf_result.jsonl" 
    if not os.path.exists(inference_result_fp):
        for retry_count in range(max_retries):
            try:
                run_inf(inf_dataset_fp, inference_result_fp)  # Run inference
                break
            except Exception as e:
                print(f"Error in prompt: {prompt}")
                print(f"Exception: {str(e)}. Retrying ({retry_count + 1}/{max_retries})...")
                traceback.print_exc()
                #time.sleep(10 * (retry_count + 1))
        
    # read the inference result (JSONL format)
    inf_results = []
    with open(inference_result_fp, 'r') as f:
        for line in f:
            if line.strip():  # Skip empty lines
                inf_results.append(json.loads(line.strip()))

    results_dict = {}
    for line in inf_results:
        file_name = line['videos'].split('/')[-1].split('.')[0]  # Get the video name without extension
        if file_name not in results_dict:
            results_dict[file_name] = {}
        prompt = line['messages'][0]['content']  # Get the prompt from the messages
        answer = line['response'].strip()  # Get the answer from the response

        feature = feature_dict[prompt.replace("<video>", "").strip()]
        results_dict[file_name][feature] = answer  # Store the answer in the results dictionary
            
    # write the results to the CSV file
    output_list = []
    for video_name in results_dict.keys():
        feature_answers = results_dict[video_name]
        output_list.append({'file_name': video_name, **feature_answers})

    # create the output CSV file
    columns = output_list[0].keys()
    if not os.path.exists(inf_result_csv_fp):
        result_csv = open(inf_result_csv_fp, 'w')
        # add header to the CSV file
        writer = csv.writer(result_csv)
        writer.writerow(columns)
    else:
        result_csv = open(inf_result_csv_fp, 'a')
    
    for row in output_list:
        writer.writerow(row)
    return

if __name__ == "__main__":
    # # get all video paths
    # seizure_videos_dir = '/mnt/.../trimmed_converted_videos'
    # df = pd.read_csv(seizure_feature_gt_fp)
    # video_files = df['file_name'].tolist()
    # video_fps = []
    
    # for video_file in video_files:
    #     video_fp = os.path.join(seizure_videos_dir, video_file)
    #     if os.path.exists(video_fp):
    #         video_fps.append(video_fp)
    #     else:
    #         print(f"Video file {video_fp} does not exist.")
    #     break
    
    
    video_fps = ['/mnt/.../A0002@5-13-2021@UA6693LK@sz_v1_1.mp4',
                 '/mnt/.../A0002@5-13-2021@UA6693LK@sz_v2_1.mp4']
    # run feature extraction
    ExtractFeatureByVLM(video_fps, inf_result_csv_fp=inf_result_csv_fp)