Who contributed to Alignment & Post-training on October 9, 2026?

3 developers shipped this update, including kahlun, fangweii1, and lxb007981.

What were the notable Alignment & Post-training updates?

[hardware, perf, training_utils] feat: add a profiler_markers() platform hook and share the attention padding helpers (#7917), [vllm, hardware] fix: vLLM-Ascend 0.23 sched_yield on ARM hosts and FIA causal mask for genuine decode (#8144), and [ci] fix: move nightly PPO CI back to Ascend A2 (#8148).

Oct 9, 2026

$ the-wire · showcase

verl opens the profiler and attention paths to non-CUDA backends

By RepoJournal · Filed 17:30 UTC on October 9, 2026 · About Alignment & Post-training · Composed from the cited sources · methodology

3 people shipped this

lxb007981 @lxb007981 3 cited

kahlun @kahlun 1 cited

fangweii1 @fangweii1 1 cited

verl's platform layer gained a profiler_markers() hook and shared attention padding helpers, while the vLLM-Ascend 0.23 compatibility fixes and a set of Ascend CI corrections landed alongside.

[hardware, perf, training_utils] feat: add a profiler_markers() platform hook and share the attention padding helpers (#7917) verl-project/verl

by kahlun

PlatformBase now exposes a profiler_markers() hook returning mark_start_range, mark_end_range, mark_annotate and marked_timer, so a hardware backend can supply its own tracing markers instead of nvtx or mstx. The default is None, so existing platforms keep their current behavior untouched, and the attention padding helpers are now shared rather than CUDA-specific.

[vllm, hardware] fix: vLLM-Ascend 0.23 sched_yield on ARM hosts and FIA causal mask for genuine decode (#8144) verl-project/verl

by fangweii1

Two vLLM-Ascend 0.23 NPU fixes were forward-ported to main and adapted to the refactored apply_npu_vllm_patches() structure: sched_yield is disabled on ARM NPU hosts such as Atlas A2, where the vLLM 0.23 scheduler busy-wait degrades, and the FIA causal mask is corrected for genuine decode. If you run vLLM-Ascend on ARM hosts, this is the patch to pick up.

[ci] fix: move nightly PPO CI back to Ascend A2 (#8148) verl-project/verl

by lxb007981

The nightly Qwen3-8B FSDP PPO job moves back to A2 runners with the matching 910B container image, after host OOM failures on A3. Model-cache symlinks in the affected nightly jobs now point at /root/.cache/models.

[ci] chore: reduce redundant Ascend workflow runs (#8142) verl-project/verl

by lxb007981

Ascend workflows no longer trigger on push, retaining only pull request triggers for main and v0.* branches, and the e2e_ascend path filters narrowed to source code, training configs, and the entrypoints and fixtures its jobs use. Expect fewer redundant pipeline runs queued behind your PRs.

[ci] chore: use cluster PyPI cache for Ascend SFT and VeOmni PPO (#8149) verl-project/verl

by lxb007981

The Ascend SFT and VeOmni PPO jobs now set PIP_INDEX_URL and PIP_TRUSTED_HOST to the cluster's existing PyPI cache, with the explicit public index overrides removed.