[kahlun]

Open question for reviewers, up front

reduce_avg_allreduce_patch_xpu.py patches torch.distributed.all_reduce process-wide 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. This changes ReduceOp.AVG's process-wide meaning for every caller in the process, XPU-related or not, for as long as the process runs (there's no narrower call-boundary to scope it to — 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 the process-wide semantics change is 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 (xccl not implementing ReduceOp.AVG doesn't need a core hook). 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) — [rebase to main] 13-10-2026 #22 had picked up the same fix independently since it lives in the same lines; dropped from [rebase to main] 13-10-2026 #22 in favor of this PR's superset version (root + per-layer, vs. root-only).
• verl_hardware_plugin/patches/reduce_avg_allreduce_patch_xpu.py (new): monkeypatches torch.distributed.all_reduce(op=AVG) → SUM + manual divide, process-wide, 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_xpu 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_xpu.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 (75 passed), including new coverage:
  • PlatformXPU() instantiation triggers reduce_avg_allreduce_patch_xpu.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_xpu: op=AVG → SUM + divide by world size; 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) — partial:
  • The FSDP2 per-layer fix itself (root + every wrapped layer) is hardware-validated: the underlying mechanism matches the per-layer+root fix already confirmed via a 12-run controlled A/B on run_grpo_fsdp2_intel_gpu.sh (6/6 PASS with the fix, 6/6 FAIL without), independently re-confirmed on run_grpo_colocate_rm_fsdp2_intel_gpu.sh. What's new here is the method-override framing — not yet re-run standalone in this exact form.
  • reduce_avg_allreduce_patch_xpu.py's actual collective behavior is unit-tested (above) but not yet run on a B60 end-to-end. Opening as draft for that reason.

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

Not yet run standalone on this branch — will complete before requesting review to move this out of draft.

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