ViperEkura 6f49738991 feat: auto-select best available attention backend ...

- Default backend resolves to highest-priority available: flash -> cuda -> torch
- attention() falls back per-call for training/fp32/unsupported head_dim
- Re-apply index_copy_ for CUDA KV cache writes (index_put_ race mitigation)

2026-08-07 13:48:59 +08:00

..

data

fix : correct prefill sampling and record alignment

2026-08-05 22:20:29 +08:00

extension

feat: auto-select best available attention backend

2026-08-07 13:48:59 +08:00

inference

perf: move decode split partials to InferenceWorkspace

2026-08-06 19:12:09 +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

fix: broadcast ref/old model state_dict for FSDP

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

trainer

refactor: emit strategy metrics as floats

2026-08-02 06:38:28 +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