ViperEkura 4b25664c79 perf: precompute kv_indptr once per decode step ...

- bind_tasks builds kv_indptr (prefix sum of seq_lens) a single time
- fwd_decode/fwd_prefill reuse it instead of rebuilding per layer
- Removes 24 cumsum launches per decode step (was ~1ms/step at B=4)
- Decode B=4: 9.60 -> 7.82 ms/step (-18.5%), +22.8% tok/s

2026-08-01 16:09:26 +08:00

..

config

Make Nora+NAdamW the default optimizer

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

dataset

refactor: remove H5 storage backend in favor of mmap bin

2026-07-29 12:50:27 +08:00

extension

perf: precompute kv_indptr once per decode step

2026-08-01 16:09:26 +08:00

inference

perf: precompute kv_indptr once per decode step

2026-08-01 16:09:26 +08:00

model

refactor: unify rotary embedding interface and update docs

2026-07-31 16:52:25 +08:00

optim

fix: use torch.optim.AdamW in ManoAdamW instead of NAdamW

2026-08-01 09:20:54 +08:00

parallel

fix: broadcast ref/old model state_dict for FSDP

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

preprocessing

refactor: remove H5 storage backend in favor of mmap bin

2026-07-29 12:50:27 +08:00

serialization

refactor: remove H5 storage backend in favor of mmap bin

2026-07-29 12:50:27 +08:00

tokenize

refactor: simplify BaseFactory and separate ModelFactory from AutoModel

2026-07-30 09:38:20 +08:00

trainer

feat: add grad_snr metric with EMA-based gradient SNR tracking

2026-08-01 08:54:44 +08:00

__init__.py

chore: bump version to 1.3.12

2026-08-01 09:22:16 +08:00

factory.py

refactor: simplify BaseFactory and separate ModelFactory from AutoModel

2026-07-30 09:38:20 +08:00

protocols.py

refactor: 重构训练后端为 Executor 模式

2026-05-24 20:35:44 +08:00

signal_handler.py

refactor: move signal_handler from parallel/ to top-level for broader reuse

2026-07-28 10:36:17 +08:00