文件最后提交记录最后更新时间
third_party: reduce lerobot eval OOM risk Fold the eval CUDA cache cleanup into 0014 and keep the patch metadata limited to the original training-log profiles. Restore the openEuler filter fixture so the lerobot patch dispatcher does not expect 0014 on that path. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Signed-off-by: liuweihong <liuweihong8@huawei.com> 18 天前
patches: add multi-tag INDEX with double-binding for lerobot stack Introduce third_party/patches/lerobot/INDEX.yaml as the single source of truth for which upstream lerobot tag the setup applier targets, and require each per-tag manifest to re-declare the binding so the resolver can fail closed on drift. Per-tag manifest gains two new top-level fields: * lerobot_tag matches INDEX.yaml.active_tag * lerobot_commit_range {min,max} brackets the upstream sha range this patch stack is valid against The previous duplicate upstream.tag / upstream.commit fields are removed; the URL-only upstream.repo entry is preserved. A new helper scripts/setup/lerobot_resolve_active.py parses INDEX.yaml, cross-checks the active tag's manifest, and emits shell-eval'able LEROBOT_* assignments. lerobot_patches.sh consumes these instead of hard-coding the v0.5.1 directory and branch name, and forwards the current libs/lerobot HEAD sha to lerobot_filter_series.py via the new --lerobot-head-commit flag. The filter rejects any HEAD that falls outside lerobot_commit_range, surfacing a hint to either fast-forward the manifest range or add a new tag directory and flip active_tag. The regression harness gains six new fixtures covering head-commit acceptance and rejection, multi-tag selection, archived-tag rejection, and INDEX vs manifest commit mismatch. All 11 fixtures pass. Documentation in scripts/setup/README.md describes the new layout, the resolver invariants, and the upgrade workflow for adding a new lerobot tag side-by-side with the existing stack. Signed-off-by: XiaoqiangWu <wuxiaoqiang.rtos@huawei.com> 1 个月前