Commit Graph

13 Commits

Author SHA1 Message Date
Oumnya 96d605b9de fix(mps): align VOXCPM_MPS_DTYPE override set with get_dtype parser
Drop "half" from _VALID_DTYPE_OVERRIDES / _LOW_PRECISION_DTYPES.
get_dtype() has never accepted "half", so VOXCPM_MPS_DTYPE=half would
pass override validation and then crash downstream with
"Unsupported dtype: half". The remaining aliases (bfloat16/bf16,
float16/fp16, float32/fp32) already cover the intended dtype space.

Adds a standalone unit check under scripts/ to guard the invariant
that every accepted override parses through get_dtype().

Addresses review feedback on #263.
2026-04-21 18:24:53 +08:00
Labmem-Zhouyx 68af4fe502 fix: ft log and setting 2026-04-08 18:15:17 +08:00
刘鑫 46cfce0c97 fix VoxCPM2 training sample_rate: 48000 -> 16000 (match AudioVAE encoder)
Made-with: Cursor
2026-04-07 22:59:18 +08:00
刘鑫 d9cf376e16 update voxcpm2 2026-03-31 11:50:37 +08:00
刘鑫 23ed7ffeee fix: fix some bugs in resuming multi-GPU training 2026-03-13 18:43:07 +08:00
symhsym 07e526a231 Update train_voxcpm_finetune.py
修改了issue#185中提到的问题,在训练时进行validate会对原模型执行to(torch.bfloat16)然后to(torch.float32)的操作,这样可能导致模型数值浮动,因此这个修改让validate步骤保留原模型数值
2026-02-11 11:17:47 +08:00
刘鑫 79e75f259e Fix: optimize save ckpt function 2026-01-16 16:22:34 +08:00
刘鑫 e8dd956fc2 Print all log messages to stderr instead of stdout 2026-01-12 15:30:45 +08:00
刘鑫 6499215204 Optimize logging validation set results to TensorBoard 2025-12-27 11:49:04 +08:00
刘鑫 a266c0a88d add lora funetine webUI; optimize lora save and load logic 2025-12-09 21:34:39 +08:00
Ayin1412 58717d7d82 修复lora/ft测试代码处传参错误的内容 2025-12-06 14:49:35 +08:00
刘鑫 400f47a516 Modify lora inference api 2025-12-05 22:22:13 +08:00
Labmem-Zhouyx 3443dbb212 Update: VoxCPM1.5 and fine-tuning supprt 2025-12-05 21:04:51 +08:00