main

文件	最后提交记录	最后更新时间
python	feat(prefill-router): run batched prefill inference (#539) * feat(prefill-router): run batched prefill inference Signed-off-by: nachiketb <nachiketb@nvidia.com> * refactor(prefill-router): clean up checkpoint forward path Signed-off-by: nachiketb <nachiketb@nvidia.com> --------- Signed-off-by: nachiketb <nachiketb@nvidia.com>	1 个月前
src	feat(libsy): Externalize State so that it survives restarts (#1024) Allow the caller to own, persist, and restore the `State`, by passing State to `run_stream`, and returning it in `RoutingOutcome`. Algorithms hold some state, such as `capable_hold_turns` in Composite. Previously this was held in memory. If libsy runs in the agent and the session ends, it would not be restored on `--resume`. The owing process also controls the lifecycle of libsy, and could unload / reload it at will, again losing state. This PR addresses that. Also drop the Rust generic `State` that was passed e.g. `Classifier<State>`. There is only one State type ever used, so it's a nice simplification. The vast majority of the changes are very simple, adding of the state or removing of the generics. Includes some cleanup to make `State` easier for callers to use: - Remove `DECISION_SOURCE_KEY` because nothing reads `decision_source` from `State.extra`. Decision source is passed separately, outside of state. - Remove `turn_count` which was neither read nor written by anything. The Python bindings did not need to change. Fold the VGR work in to general-purpose State. Add a `StateValue::Bool` to support it. Part 1 of https://github.com/NVIDIA-NeMo/Switchyard/issues/1023 Assisted-by: Cursor Agent:Gemini 4 Argon high Reviewed-by: Claude:Opus 5.5 medium Signed-off-by: Graham King <grahamk@nvidia.com>	7 天前
tests	fix(prefill-router): do not score an earlier exchange for a text-less user turn (#843) Signed-off-by: Zeshawn Zahid <zeshawn.zahid@gmail.com>	10 天前
Cargo.toml	fix(release): align package versions for 0.3.0 (#734) Signed-off-by: nachiketb <nachiketb@nvidia.com>	1 个月前
build.rs	feat(prefill-router): add Transformers feature extraction (#506) * feat(prefill-router): add Transformers feature extraction Signed-off-by: nachiketb <nachiketb@nvidia.com> * test(prefill-router): validate Qwen extraction parity Signed-off-by: nachiketb <nachiketb@nvidia.com> * fix(prefill-router): harden feature extraction Signed-off-by: nachiketb <nachiketb@nvidia.com> * fix(prefill-router): use pinned extraction oracle Signed-off-by: nachiketb <nachiketb@nvidia.com> * fix(prefill-router): clarify transformers runtime behavior Signed-off-by: nachiketb <nachiketb@nvidia.com> * fix(prefill-router): harden transformers extraction Signed-off-by: nachiketb <nachiketb@nvidia.com> --------- Signed-off-by: nachiketb <nachiketb@nvidia.com>	1 个月前
pyproject.toml	feat(prefill-router): run batched prefill inference (#539) * feat(prefill-router): run batched prefill inference Signed-off-by: nachiketb <nachiketb@nvidia.com> * refactor(prefill-router): clean up checkpoint forward path Signed-off-by: nachiketb <nachiketb@nvidia.com> --------- Signed-off-by: nachiketb <nachiketb@nvidia.com>	1 个月前
uv.lock	feat(prefill-router): run batched prefill inference (#539) * feat(prefill-router): run batched prefill inference Signed-off-by: nachiketb <nachiketb@nvidia.com> * refactor(prefill-router): clean up checkpoint forward path Signed-off-by: nachiketb <nachiketb@nvidia.com> --------- Signed-off-by: nachiketb <nachiketb@nvidia.com>	1 个月前