[Test] Cover CustomOp platform dispatch

[LinzeShi]

Purpose

Add the focused CustomOp test coverage requested in #7382: constructor dispatch, argument forwarding, backend overrides and inherited fallbacks, and missing backend implementations. The MoTRMSNorm regression exercises the normal module call path.

An unmocked native-platform smoke test also checks op(x) against x * 2, covering real dispatch and bound-method invocation. It skips when an accelerator backend is selected.

This PR changes tests only; it does not qualify device kernels.

Test Plan

pytest -q tests/diffusion/layers/test_custom_op.py

Test Result

Python 3.12.4 in both environments:

Environment	vLLM	PyTorch	Result
CPU	0.29.0+cpu	2.13.0+cpu	20 passed
CUDA	0.29.0	2.13.0+cu130	19 passed, 1 skipped

The CUDA run skips the native-platform smoke test as intended; the remaining tests use CPU tensors. Both runs report 14 existing torch.jit.script_method deprecation warnings. Changed-file pre-commit checks passed.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-09-15T08:50:08.251913Z	0771468	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[vllm-omni-review-bot]

This PR appears to belong to: docs/design/module/diffusion/diffusion_model_integration.md.

Module owners: @Bounty-hunter @fhfuih @RuixiangMa

Routing: @Bounty-hunter via semantic router; @fhfuih via semantic router; @RuixiangMa via semantic router

@LinzeShi, please review your own changes and leave a short self-review comment describing what you checked. PRs without author self-review may not be assigned a reviewer.

Please take a look when you have a chance. If you would like an automated review, mention @vllm-omni-review-bot in a comment.

[LinzeShi]

@vllm-omni-review-bot I would like an automated review

[vllm-omni-review-bot]

Omni ReviewBot triage note

Automated triage of commit 4fe204a1ad40 produced:

• Priority: high. Prompt maintainer attention is suggested.

These are automated triage suggestions only — the final decision belongs to the maintainers.

[Dong1017]

Reviewed 0771468e3633 (single commit, unchanged since opening). Test-only (+104/−4); custom_op.py is byte-identical to main.

Closes the systematic-coverage gap flagged in #7382: 6-way branch selection (native/cuda/rocm/npu/xpu/musa) with argument-identity forwarding via op(...), inherited fallbacks (xpu→native, rocm→cuda, musa→cuda), override precedence, missing-backend NotImplementedError, and a dispatch-at-construction assertion (predicates poisoned with side_effect=AssertionError post-__init__). The MoTRMSNorm regression now drives the real norm(x) path. 6+3+3+4+3 = 19.

Traced all cases against the production custom_op.py; the mock targets the module global production actually reads, so the dispatch logic under test is real. Independently reproduced the author's run (H200 host, vllm 0.29.0 / torch 2.13.0+cu130, 2026-09-20): 19 passed in 15.25s from an archive of this exact head — machine-local evidence, since this repo's CI runs no unit-test lane here. The SPDX edit (vLLM → vLLM-Omni) matches this repo's own check_spdx_header.py.

Boundary: every new case mocks the platform — nothing here exercises real CUDA/NPU kernels or qualifies a device path, consistent with the PR description.

Two non-blocking notes inline.

[vllm-omni-review-bot]

Omni ReviewBot: no human activity for 7 days

@LinzeShi this pull request has had no human commit, comment or review since 2026-09-23. Please confirm the current plan and next step. The author or a maintainer decides whether to change the PR state.

To keep it moving, any one of these is enough: push an update, reply to the open blocker, or post the current plan and timeline.

[LinzeShi]

@Dong1017 I addressed both notes and added the unmocked smoke test. Could you take another look? Thanks!

[Dong1017]

@LinzeShi Thanks for the update. At 4fe204a1ad4051ae7c4de75d100bfc0dcac24456, the new unmocked op(x) smoke addresses the remaining native-dispatch/bound-method coverage note. The existing MoTRMSNorm norm(...) regression is retained with its updated BAGEL import path.

I independently ran the focused file on H200 with vLLM 0.30/PyTorch 2.13, using a scoped pytest config: 19 passed, 1 skipped. The tensor cases ran on CPU; the new native-platform smoke correctly skipped on the real CUDA platform. Your reported CPU-only 20-pass run therefore remains author evidence for that smoke. No further change requested from this focused follow-up.

[LinzeShi]

@hsliuustc0106 Dong1017 has reviewed the update and confirmed no further changes are needed. Could you take a look or suggest a maintainer for the final review?

[vllm-omni-review-bot]

Omni ReviewBot routing record

Assigned Strict on cursor (cursor-grok-4.6-high) under experiment fleet-strict-cursor-grok46-zcode-glm53flash-5050-c5-z10-20261002.

[vllm-omni-review-bot]

Omni ReviewBot attempt record

Review attempt ended as failed (failed; retrying strict/cursor/cursor-grok-4.6-high in 120s (try 2 of 3)).

[vllm-omni-review-bot]

Omni ReviewBot attempt record

Review attempt ended as failed (failed; retrying strict/cursor/cursor-grok-4.6-high in 600s (try 3 of 3)).

[vllm-omni-review-bot]

Omni ReviewBot attempt record

Review attempt ended as failed (failed; falling back to direct/cursor/auto).

[vllm-omni-review-bot]

Omni ReviewBot review

Changes since the previous review

• 0 new inline finding(s); 0 finding(s) below.

CI at 4fe204a1ad40 (2026-10-10T14:59:42.917937+00:00): required check(s) blocking: buildkite/vllm-omni (missing).

Note: The assigned review arm strict/cursor/cursor-grok-4.6-high could not complete this review, so it was produced by the fallback arm direct/cursor/auto. It is excluded from the routing experiment.

Full review analysis

PR description

This pull request only updates tests/diffusion/layers/test_custom_op.py. A new mock_platform fixture replaces current_omni_platform with a five-predicate mock, and new cases check constructor dispatch, argument identity, inherited xpu/rocm/musa fallbacks, override precedence, and NotImplementedError when a backend method is missing. An unmocked smoke test constructs _NativeOnlyOp and expects op(x) to equal x * 2 when no accelerator predicate is selected; that case skips on an accelerator platform. The existing MoTRMSNorm regression now calls norm(...) after setting gen_weight to 2, instead of calling forward_xpu directly.

Change flow

flowchart LR
  platform["[EXISTING] CustomOp platform predicates"]:::existing
  fixture["[NEW] mock_platform fixture"]:::new
  cases["[NEW] Dispatch, fallback, override, and missing-backend cases"]:::new
  smoke["[NEW] Unmocked native op(x) smoke"]:::new
  mot["[CHANGED] MoTRMSNorm norm(x) regression"]:::changed
  platform --> fixture --> cases
  platform --> smoke
  fixture --> mot
  classDef existing fill:#e5e7eb,stroke:#6b7280,color:#111827
  classDef changed fill:#fef3c7,stroke:#d97706,color:#451a03,stroke-width:2px
  classDef new fill:#dcfce7,stroke:#16a34a,color:#052e16,stroke-width:2px
  classDef removed fill:#fee2e2,stroke:#dc2626,color:#450a0a,stroke-width:2px

Loading

---

_{🤖 This review was generated by InferMatrix Copilot, an open-source repo-maintenance agent for PR review, CI debugging and issue triage. Try it on your own repo, and ⭐ star it if it helped!}