[kahlun]

Open question for reviewers, up front

reduce_avg_allreduce_patch.py patches torch.distributed.all_reduce when op=AVG — the only mechanism available for verl-core's 3 raw dist.all_reduce(op=ReduceOp.AVG) call sites, since there's no per-site wrapper to hook into and this plugin doesn't own those call sites. It's scoped to dist.get_backend(group) == "xccl", so a gloo/nccl group in the same process (e.g. a CPU-only Ray actor's coordination group alongside a GPU worker's xccl group) is untouched and keeps native AVG — this was originally unscoped (any group, any backend) and has been fixed to check the backend per call. What's left: anyone reading verl-core's source still can't see, from that source alone, that an xccl group's ReduceOp.AVG silently becomes SUM+divide, for as long as the process runs (there's no narrower call-boundary available — see Changes below for why). That's a real design smell, not a clean win, and is the one piece here still worth a narrow core-hook ask if that's judged unacceptable — happy to split it into its own PR against #7917 if so.

Summary

Plugin-side replacement for is_reduce_avg_supported() — implemented with zero verl-core changes: no new PlatformBase hook, no dependency on verl-project/verl#7917.

Motivation

verl-hardware-plugin#26 asked maintainers whether Intel XPU should extend PlatformBase in verl-core (the TPU precedent) or stay entirely inside this plugin via monkeypatch (the MLU precedent). The maintainer's answer (2026-09-23): default to the plugin, keep #7917 narrowed to only what truly can't be done from here.

is_reduce_avg_supported/all_reduce_avg has just been dropped from #7917 and from #22 on that basis -- the underlying reduce_avg problem doesn't need a core hook to work around. (Correction from an earlier version of this PR: the reason isn't "xccl doesn't implement ReduceOp.AVG" -- per this review-thread investigation on real 4xB60 hardware, oneCCL implements AVG on its SYCL-kernel path but not its scheduler path, and path selection is internal and not capability-aware, so an AVG request can abort mid-collective rather than reliably succeed or fail. Fix tracked for oneCCL 2022.2 / torch 2.15. Same mechanism below, corrected reasoning.) This PR is the plugin-side replacement.

attention_utils_module() and the profiler hooks passed the maintainer's bar and stay as PlatformBase hooks — implemented for XPU in #22, which this PR does not touch or duplicate. (An earlier draft of this branch also monkeypatched attention as a "how far can pure monkeypatch go" comparison; dropped once it was clear the hook was staying and a plugin-only alternative for it served no purpose.)

Changes

• verl_hardware_plugin/engines/fsdp_xpu.py: extend the existing (non-hook, non-patch) initialize() override to call set_force_sum_reduction_for_comms(True) on the FSDP2 root and every individually-wrapped transformer layer, not just the root — each fully_shard()'d module has independent FSDP state. This is ordinary method-override behavior on the plugin's own EngineRegistry-registered engine, not a monkeypatch, so it's robust regardless of import order. Also fixes an unrelated self.model→self.module attribute bug in the same function (FSDPEngineWithLMHead/WithValueHead expose the wrapped model as self.module, never self.model) — [XPU] feat(xpu): add VTune (ITT) profiler backend #22 had picked up the same fix independently since it lives in the same lines; dropped from [XPU] feat(xpu): add VTune (ITT) profiler backend #22 in favor of this PR's superset version (root + per-layer, vs. root-only).
• verl_hardware_plugin/patches/reduce_avg_allreduce_patch (new): monkeypatches torch.distributed.all_reduce(op=AVG) → SUM + manual divide, scoped to xccl process groups (dist.get_backend(group) == "xccl"), for the three remaining raw dist.all_reduce(op=ReduceOp.AVG) call sites in verl-core (sft_trainer.py, engine_workers.py, profiler/performance.py::reduce_timing). See the open question above. async_op=True call sites are intentionally left unpatched (falls through to original behavior) rather than risk returning a wrong result before the collective completes.
• verl_hardware_plugin/patches/__init__.py (new): plain package docstring, no code — holds reduce_avg_allreduce_patch as the one patch module. No generic apply_all() aggregator: it had exactly one caller and one module, so it wasn't earning its keep (see the platform_xpu.py bullet below).
• verl_hardware_plugin/platforms/platform_xpu.py: PlatformXPU.__init__ calls reduce_avg_allreduce_patch.apply() directly. Deliberately not called from the shared verl_hardware_plugin/__init__.py (that file is common to every vendor this plugin supports) — PlatformXPU() is only constructed once verl.plugin.platform.platform_manager._create_platform() has actually selected "intel" as the platform for this process. Calling it from the shared top-level __init__.py instead would fire on any host where torch.xpu merely happens to be importable, even if a different platform ends up selected for the run — e.g. a mixed host with VERL_PLATFORM=nvidia explicitly set would still get its process-wide all_reduce semantics changed for no reason. (Two earlier drafts of this PR had this wired the wrong way, then routed through an unnecessary same-package apply_all() wrapper — both fixed.)

Testing

• pytest tests/ -v passes (76 passed), including new coverage:
  • PlatformXPU() instantiation triggers reduce_avg_allreduce_patch.apply(); the shared plugin __init__.py no longer binds any patch call directly (regression guard against re-introducing the wrong wiring).
  • reduce_avg_allreduce_patch: op=AVG on an xccl group → SUM + divide by world size; op=AVG on a non-xccl group (e.g. gloo) passes through untouched; other ops pass through unchanged; async_op=True + AVG falls through unpatched; no-op when XPU unavailable; idempotent (calling apply() twice doesn't double-wrap).
• Manually verified on target hardware (B60):
  What's new here is the method-override framing
  • reduce_avg_allreduce_patch.py's actual collective behavior is unit-tested

Acceptance Baseline (for new hardware adaptation PRs)

• Ran scripts/baseline_grpo_gsm8k.sh on target hardware (8 devices)
• Training completed all epochs without error
• critic/rewards/mean shows clear upward trend in first 100 steps
• Curve is consistent with NVIDIA reference

Checklist

• Code follows the project's style and passes pre-commit checks
• Documentation updated — folded into this description rather than a docs/design/ file; this repo has no precedent for keeping RFC/experiment-verdict docs in the merged tree
• No secrets or credentials included

[kahlun]

@wuxibin89 @physics31415926 plugin-only ReduceOp.AVG patch for XPU/xccl, following the direction confirmed in #26 (no core changes).

One open design question is called out in the PR description if you want to weigh in on it specifically; otherwise it should be good to merge as-is.