Sync PROGRESS.md: CastError permanent resolution + step <= 10 logging
Browse files- docs/PROGRESS.md +1 -0
docs/PROGRESS.md
CHANGED
|
@@ -404,3 +404,4 @@ Kaggle par manual file upload se version mismatch hota hai (v1 tokenizer wala ca
|
|
| 404 |
- **CUDA Kernel Acceleration**: Enabled torch.backends.cudnn.benchmark = True and allow_tf32 = True for fixed seq_len=1024.
|
| 405 |
- **Live Rich Logging**: Enabled immediate step 1 detailed print, set log_every to 10, and formatted live percentage display for all 8 MoE experts.
|
| 406 |
- **Verification & Push**: Tested via CPU smoke test (exit 0) and pushed live to ViuAI/ViuMini-MoE-242M (commits 58dee8cd and dbaf003f).
|
|
|
|
|
|
| 404 |
- **CUDA Kernel Acceleration**: Enabled torch.backends.cudnn.benchmark = True and allow_tf32 = True for fixed seq_len=1024.
|
| 405 |
- **Live Rich Logging**: Enabled immediate step 1 detailed print, set log_every to 10, and formatted live percentage display for all 8 MoE experts.
|
| 406 |
- **Verification & Push**: Tested via CPU smoke test (exit 0) and pushed live to ViuAI/ViuMini-MoE-242M (commits 58dee8cd and dbaf003f).
|
| 407 |
+
- [x] 2026-09-24 CastError root cause permanently resolved: excluded 'code/' prefix from pretrain streaming so dataset schema resolves to standard 5-col; enabled step <= 10 logging so every initial step prints live feedback. Verified clean test (exit 0) and pushed live (commit eee71567).
|