ViperEkura e3ea850dc9 fix: default backend race, raise on explicit fallback ...

- _default_backend lazy init protected with threading.Lock
- attention() raises when explicit backend cannot handle call
- FlashAttnBackend rejects prefill with non-None attn_mask
- training test uses TORCH_NATIVE backend directly

2026-08-08 13:00:58 +08:00

..

data

refactor: break JsonlStore→preprocessing circular dependency

2026-08-07 23:22:32 +08:00

extension

fix: default backend race, raise on explicit fallback

2026-08-08 13:00:58 +08:00

inference

fix: add dtype guard to FlashAttnBackend capability check

2026-08-07 23:08:22 +08:00

module

refactor: stateless MoE routing with grouped dispatch

2026-08-05 18:42:12 +08:00

optim

feat: add Mano manifold optimizer (mano_adamw)

2026-08-01 08:51:08 +08:00

parallel

style: fix ruff lint warnings

2026-08-07 14:17:48 +08:00

trainer

fix: prevent signal handler test from racing with training completion

2026-08-07 18:32:08 +08:00

conftest.py

fix: broadcast ref/old model state_dict for FSDP

2026-07-31 08:32:22 +08:00

helpers.py

fix: broadcast ref/old model state_dict for FSDP

2026-07-31 08:32:22 +08:00

test_train_cli.py

Make Nora+NAdamW the default optimizer

2026-07-31 23:16:39 +08:00