Minicpm_o-4-5进行视频推理和多模混合推理时报错

File "/workdir/user_repository/llm-eval-public/inference/local_deploy_minicpm_o_4_5.py", line 115, in generate
answer = self.model.chat(msgs=messages, max_new_tokens=32768,max_inp_length=32000,max_slice_nums=1, omni_mode=True, use_tts_template=False,
File "/usr/local/conda/lib/python3.10/site-packages/torch/utils/_contextlib.py", line 116, in decorate_context
return func(*args, **kwargs)
File "/home/hadoop-aipnlp/.cache/huggingface/modules/transformers_modules/main/modeling_minicpmo.py", line 1202, in chat
res, outputs = self.generate(
File "/usr/local/conda/lib/python3.10/site-packages/torch/utils/_contextlib.py", line 116, in decorate_context
return func(*args, **kwargs)
File "/home/hadoop-aipnlp/.cache/huggingface/modules/transformers_modules/main/modeling_minicpmo.py", line 892, in generate
model_inputs["inputs_embeds"] = self.get_omni_embedding(
File "/home/hadoop-aipnlp/.cache/huggingface/modules/transformers_modules/main/modeling_minicpmo.py", line 771, in get_omni_embedding
input_embeddings[i, bound[0] : bound[1]] = audio_embs[
RuntimeError: The expanded size of the tensor (21) must match the existing size (20) at non-singleton dimension 0. Target sizes: [21, 4096]. Tensor sizes: [20, 4096]

视频推理的报错：
File "/workdir/user_repository/llm-eval-public/inference/local_deploy_minicpm_o_4_5.py", line 115, in generate
answer = self.model.chat(msgs=messages, max_new_tokens=32768,max_inp_length=32000,max_slice_nums=1, omni_mode=True, use_tts_template=False,
File "/usr/local/conda/lib/python3.10/site-packages/torch/utils/_contextlib.py", line 116, in decorate_context
return func(*args, **kwargs)
File "/home/hadoop-aipnlp/.cache/huggingface/modules/transformers_modules/main/modeling_minicpmo.py", line 1146, in chat
content = normalize_content(content)
File "/home/hadoop-aipnlp/.cache/huggingface/modules/transformers_modules/main/utils.py", line 2400, in normalize_content
normalized = normalize_content_item(item)
File "/home/hadoop-aipnlp/.cache/huggingface/modules/transformers_modules/main/utils.py", line 2338, in normalize_content_item
video_frames, audio_segments, stacked_frames = get_video_frame_audio_segments(
File "/home/hadoop-aipnlp/.local/lib/python3.10/site-packages/minicpmo/utils.py", line 392, in get_video_frame_audio_segments
video_segments, timestamps = _extract_frames_by_timestamps(timestamps, is_long_video)
File "/home/hadoop-aipnlp/.local/lib/python3.10/site-packages/minicpmo/utils.py", line 357, in _extract_frames_by_timestamps
video = _vr.get_batch(frame_idx).asnumpy()
File "/usr/local/conda/lib/python3.10/site-packages/decord/video_reader.py", line 175, in get_batch
arr = _CAPI_VideoReaderGetBatch(self._handle, indices)
File "/usr/local/conda/lib/python3.10/site-packages/decord/_ffi/_ctypes/function.py", line 173, in call
check_call(_LIB.DECORDFuncCall(
File "/usr/local/conda/lib/python3.10/site-packages/decord/_ffi/base.py", line 78, in check_call
raise DECORDError(err_str)
decord.ffi.base.DECORDError: [11:20:58] /github/workspace/src/video/ffmpeg/threaded_decoder.cc:104: Check failed: run.load()
请问下这两个有什么好的修复建议吗

Start with inference/local_deploy_minicpm_o_4_5.py at line 115 and inspect the two failing paths in modeling_minicpmo.py, especially get_omni_embedding. For video, trace minicpmo/utils.py through get_video_frame_audio_segments and _extract_frames_by_timestamps into decord.get_batch. Done means both multimodal and video inference complete without either reported error.