Root cause: funasr.AutoModel (loaded by SenseVoice encoder) and pyarrow
have a C library initialization conflict causing iter_batches to hang.
Fix:
- lazy-import pyarrow in VAMDataset/VLMDataset.__init__ (after model loads)
- remove unused pyarrow imports from dataset/common.py
- replace librosa.resample with torchaudio.resample (avoids numba+numpy 2.5
incompatibility)
- reduce iter_batches batch_size from 65536 to 4096 (lower memory)
Config updates:
- point Stage 2 (audio_proj) to omni-v Stage 1 checkpoint
- add explicit device: cuda to omni-v config
feat: add VAM T2A config variant using omni-v initialization
- Rename vlm.yaml→vlm_pretrain.yaml, vlm_moe.yaml→vlm_pretrain_moe.yaml
- Add vlm_sft.yaml and vlm_sft_moe.yaml for VLM SFT training
- Add --weight_path, --model_dir, --vision_dir, --tokenizer_dir to VLM trainer
- Update init_vlm_model to support weight_path and model_dir for decoupled loading
- Update VLM pretrain default data_path to dataset/vlm/
- Sync all doc references to new config filenames
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
refactor: reorganize dataset into lm/vlm/vam subdirectories
Move all .jsonl data files to dataset/lm/ and update configs, trainers,
and docs to reflect the new paths.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
reorganize configs into subdirs lm/vlm/vam, remove runs/, add interview docs for pretrain & full_sft
