[tzhouam]

Purpose

Fourth backport batch for release/v0.32.0, following #9295. Includes the 14 selected upstream PRs tied to v0.32.0 issues or required regression coverage. Base: bb3c0521628cfa13cdb2ace76c525c811c313b33.

Source PR	Release issue / purpose
#8801	#8796: preserve native MiniMax-H3 local snapshot layout
#9289	#9287: register native KV specs in fast-allocation tests
#9244	#6882: repair the NPU full_attn_spans sub-path; does not establish full issue resolution
#9171	#9166: align 1x1 Wan VAE attention queries
#9107	#9072: allow a cold-start timeout budget for the first duplex open
#7880	#7729: re-aim a dropped silence continuation at the opened answer
#9302	Repair the spans-precedence regression from #9248, already included in #9295
#9303	ROCm math-kernel test coverage accompanying #9171
#9041	#6843: place stage-CLI test replicas on separate GPUs
#5784	#5659: resolve missing OmniVoice reference text and align audio preprocessing
#9249	#6179: reject recompute preemption for AR stages that cannot replay safely
#9276	Strengthen the existing #6223 GroupNorm workaround regression; no underlying kernel fix claimed
#8874	#8817: keep diffusion LoRA buffers on the compute device
#8868	#8850: correct trailing VAE tile placement and blending

#9270 and #9273 remain separate release backports (#9274 and #9275); they are not duplicated here.

Backport adaptations

• [Bugfix] MiniMax-H3: keep native layout for local snapshots with modular_model_index.json #8801: adapt the checkpoint-file context without importing the unrelated FastH3 model changes from [Model] Add FastH3 series #9167.
• [Core] Fail recompute preemption on AR stages that cannot replay their decode state #9249: add only the recompute-preemption capability to the release configuration. Preserve its existing MRv2 platform checks; omit unrelated pending-generation/Fish platform changes from main. Add the existing main-side test dictionary annotation required by mypy.
• Validation follow-up: assert the existing non-null spans precondition in the new NPU helper and add missing SPDX headers to the touched OmniVoice client.
• Preserve upstream authors, source SHA provenance, and sign-offs.

No new models or performance optimizations from the intervening main commits are included.

Contributor intent

Outcome: bring the selected v0.32.0 issue fixes and their necessary regressions to the release line. The controlling variable is whether the existing affected release paths contain these fixes. This batch backports the established upstream implementations and tests for users encountering those release failures; it does not broaden the release feature set.

Validation

Head: bbf9bde14189e83469853f396f03e8c060f5e55b.
Local CPU environment: Python 3.12.13, vLLM 0.32.0, PyTorch 2.13.0+cu130, Diffusers 0.40.0, pytest 9.1.1.

• Attention / VAE / LoRA / ROCm guard CPU group: 296 passed, 7 deselected.
• Duplex / scheduler / test harness / OmniVoice CPU group: 278 passed.
• Final integrated configuration, preemption, checkpoint, attention, and speech-adapter group: 620 passed, 1 skipped, 7 deselected. These groups overlap; counts are not a unique-test total.
• [Bugfix] MiniMax-H3: keep native layout for local snapshots with modular_model_index.json #8801: the native-snapshot regression fails before the fix; the adapted checkpoint suite passes all 59 cases.
• [Bugfix][XPU] Fix HunyuanImage3 FLASH_ATTN regression from #9248 #9302: the new automatic/explicit spans-routing cases fail against the pre-backport tree and pass after backport.

Final integrated command (from the backport checkout):

PYTHONPATH=. python -m pytest -o addopts='' -m cpu -q --tb=short \
  tests/core/sched/test_omni_recompute_preemption.py \
  tests/model_executor/models/test_recompute_preemption_declarations.py \
  tests/model_executor/models/fish_omni/test_fish_omni_model_state.py \
  tests/model_executor/models/qwen3_omni/test_qwen3_omni_preempt_replay.py \
  tests/config/test_omni_config.py \
  tests/diffusion/models/minimax_h3/test_fasth3_checkpoint.py \
  tests/diffusion/attention/test_flash_attn.py \
  tests/entrypoints/openai_api/test_tts_adapter.py

Ruff, formatting, documentation lint, SPDX, forbidden imports, test marks, and applicable remaining pre-commit hooks pass. Mypy retains six pre-existing errors, reproduced on the release-base tree: four in attention/layer.py, one in duplex_omni_engine.py, and one in the pre-existing LoRA test. The new NPU helper's typing error was fixed.

Not run locally: CUDA/NPU/XPU/ROCm model E2E or hardware kernels. This environment lacks the native vllm._C extension and has stale installed Omni version metadata; passing CPU tests are not hardware evidence. Current-head project CI is requested with ready.

AI assistance: Codex prepared the backports, resolved release-context conflicts, reviewed patch scope, ran local checks, and drafted this description.

Release CI follow-ups

The original head passed CUDA 18001, AMD 14561, NPU 9742, GitHub Actions, and docs. Intel 10742 hit Level Zero OUT_OF_RESOURCES during unchanged Z-Image text-encoder warmup, before LoRA adapters were loaded. The same test passed on release backports 10739/10740 and main 10736 (already containing #8874). Concurrent builds shared physical devices without an affinity mask; resource interference is a supported hypothesis.

The updated head also includes two minimal CI stabilizations:

• Bootstrap kernels-data>=0.14.0.dev1 before the Intel image's unisolated wheel build. The same missing-plugin-dependency error occurred on the earlier release baseline Intel 10738. An isolated Python 3.12 reproduction failed before this bootstrap and passed after; kernels-data has no Torch/Triton dependencies and the paired-runtime exclusions remain.
• Initialize existing Mammoth attention-comparison weights to prevent a reproduced uninitialized-memory NaN. The test file matched [cherry-pick][release/v0.32.0] cherry-pick #9273: reject requests after repeated stage failures (circuit breaker) #9275 byte-for-byte before the fix; its numerical thresholds and route assertions remain, and the CUDA mark still comes from hardware_marks. Local deterministic GPU regression passed both cases; [cherry-pick][release/v0.32.0] cherry-pick #9273: reject requests after repeated stage failures (circuit breaker) #9275's preceding-head CUDA run also passed the full 310-test model task with this fix.

All applicable checks passed on the changed files. The approved 14 source backports are unchanged. On current head 7ad5f31e3750bbeaf7b0767c9e60d0b31245c6a5, Intel 10749 passed the full four-round suite with the runner's supported ZE_AFFINITY_MASK=4,5,6,7. This retained all selected tests, including the Z-Image LoRA switching test and both three-card Qwen2.5-Omni tests; the original resource error did not recur. All 11/11 current-head checks passed: CUDA 18007, AMD 14567, Intel 10749, NPU 9748, both wheel builds, CodeQL, pre-commit, DCO, and docs. The final NPU MiniCPM-o test reported 5 passed, 12 deselected.

[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	⚠️ Failed 2026-10-24T16:08:46.821386Z We couldn’t complete this review. Comment @codex review to try again.	a4b807d	Draft marked ready

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.

[chatgpt-codex-connector]

Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits.
Credits must be used to enable repository wide code reviews.

[vllm-omni-review-bot]

Omni ReviewBot triage note

Automated triage of commit 7ad5f31e3750 produced:

• Priority: high. Prompt maintainer attention is suggested.

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