_calculate_timestamps参数使用
Author: UCTCCreated Jul 23, 2026Updated Sep 8, 2026
Labelsbugpending
Reminder
- I have read the above rules and searched the existing issues.
System Info
llamafactoryversion: 0.9.6.dev0- Platform: Linux-6.1.52-9-x86_64-with-glibc2.35
- Python version: 3.12.13
- PyTorch version: 2.11.0+cu130 (GPU)
- Transformers version: 5.6.0
- Datasets version: 3.5.0
- Accelerate version: 1.11.0
- PEFT version: 0.18.1
- GPU type: NVIDIA L20
- GPU number: 2
- GPU memory: 44.53GB
- TRL version: 0.24.0
- DeepSpeed version: 0.19.2
- Bitsandbytes version: 0.49.2
- vLLM version: 0.20.2
- Git commit: 8a43588d3ea6cafe755713bb9fcc3bd04604a232
- Default data directory: detected
Reproduction
我正在尝试用
"temporal_patch_size": 1,
"merge_size": 2,
的参数不使用预训练权重对 Qwen3-VL 模型做debug来理解源码,我发现在/src/llamafactory/data/mm_plugin.py的_calculate_timestamps为什么使用的是merge_size传入而不是temporal_patch_size, 这里做的应该是时间维度的压缩,我认为应该使用temporal_patch_size,因为merge_size是用于Patch Merger当中的空间token的下采样吧?
Others
No response
Source: hiyouga/LlamaFactory