The inference results of the local deployment of lixin4ever/VideoLLaMA2-AV are inconsistent with the web DEMO results

[haloya-0328]

Hi. I downloaded the model you released on huggingface, used the config.json in the path lixin4ever/VideoLLaMA2-AV, sampled 32 frames, and set do_sample=False for both. Why is the inference result inconsistent with the demo?

config details:
{ "_name_or_path": "DAMO-NLP-SG/VideoLLaMA2.1-7B-16F", "architectures": [ "Videollama2Qwen2ForCausalLM" ], "attention_dropout": 0.0, "bos_token_id": 151643, "eos_token_id": 151645, "freeze_mm_mlp_adapter": false, "hidden_act": "silu", "hidden_size": 3584, "hidden_size_a": 3584, "image_aspect_ratio": "pad", "initializer_range": 0.02, "intermediate_size": 18944, "max_position_embeddings": 32768, "max_window_layers": 28, "mm_audio_tower": "work_dirs/VideoLLaMA2.1-7B-A/audio_tower.bin", "mm_hidden_size": 1152, "mm_hidden_size_a": 768, "mm_projector_a_type": "mlp2x_gelu", "mm_projector_lr": null, "mm_projector_type": "stc_connector_v35", "mm_vision_select_feature": "patch", "mm_vision_select_layer": -2, "mm_vision_tower": "google/siglip-so400m-patch14-384", "model_type": "videollama2_qwen2", "num_attention_heads": 28, "num_frames": 8, "num_hidden_layers": 28, "num_key_value_heads": 4, "rms_norm_eps": 1e-06, "rope_theta": 1000000.0, "sliding_window": 131072, "tie_word_embeddings": false, "tokenizer_model_max_length": 2048, "tokenizer_padding_side": "right", "torch_dtype": "bfloat16", "transformers_version": "4.42.3", "tune_mm_mlp_adapter": false, "tune_mm_mlp_adapter_a": true, "use_cache": true, "use_mm_proj": true, "use_sliding_window": false, "vocab_size": 152064 }
To explain this problem, do you need more detailed information?