[Model] MAGI-2: head-sharded expert parallelism with an independent EP degree

[yeahdongcn]

Purpose

MAGI-2 shards its MoE experts along the head axis over the Ulysses ranks, so today the expert-parallel degree always equals the sequence-parallel degree. This PR lets a model declare head-sharded expert parallelism and set expert_parallel_size to a divisor of the SP degree, as proposed in #7085. It consolidates and supersedes #7266 and #7270. Five commits:

1. [Refactor] Share packed head-parallel exchanges for diffusion models: a helper in vllm_omni/diffusion/distributed/head_parallel.py for the packed head all-to-all, used by MAGI-2's parallel.py (from [Refactor] Share packed head-parallel exchanges for diffusion models #7266).
2. [Feature] Support model-declared head-sharded expert parallel groups: DiffusionModelMetadata.expert_parallel_style="head" makes parallel_state build the EP group as SP subgroups of size expert_parallel_size (from [Feature] Support model-declared head-sharded expert parallel groups #7270).
3. [Model] MAGI-2: enable head-sharded expert parallelism. MAGI-2 declares the head style; its MoE uses the head-EP subgroup, and the transformer passes each layer the token counts in EP rank order instead of gathering them. expert_parallel_size is declared in the structured stage config, so the engine's kwargs path accepts it, and is exposed as --expert-parallel-size in the text-to-video example. Head-EP requires tensor_parallel_size=1.
4. [Perf] MAGI-2: exchange head-EP tokens over the whole SP group on MUSA. When EP is smaller than SP (SP8xCFG1 with EP4), ranks at the same position of every EP group own the same MoE head shard with the same weights. On MUSA the MoE dispatch and combine then run over the SP group: each rank splits its tokens into one contiguous part per EP group and sends each part to the owners in that group, so the all-to-all spans all 8 SP ranks instead of the 4 of one EP group. HeadParallelLayout gains replica_counts for this layout. The output stays bit-exact: each part goes to an owner that holds the same head shard with the same weights, and each owner receives exactly as many tokens as the exchange inside its own EP group gives it, so every MoE batch keeps its size. Only which rows share an owner's batch, and their order, change, and the MoE computes each (token, head) row independently of the rest of its batch; test_owner_moe_rows_do_not_depend_on_their_batch checks this on the device at the production shape. When the EP groups hold different token totals, keeping those sizes would move tokens into batches of another size, so that call falls back to the exchange inside each EP group. With EP equal to SP, and on other platforms, the exchange is unchanged.
5. [Misc] Document expert_parallel_size on DiffusionParallelConfig: enable_expert_parallel and expert_parallel_size each carry their own docstring. No behaviour change.

Without --enable-expert-parallel nothing changes: at SP4xCFG2 the generated video and audio are byte-identical to the same stack without this PR.

Part of the MAGI-2 work split from #7156. MAGI-2's MUSA support is in main since #8498.

Test Plan

python -m pytest -q -o addopts='' \
  tests/diffusion/distributed/test_head_parallel.py \
  tests/diffusion/distributed/test_head_expert_parallel.py \
  tests/diffusion/distributed/test_expert_parallel_size_plumbing.py \
  tests/diffusion/distributed/test_expert_parallel_layout.py \
  tests/diffusion/distributed/test_expert_parallel.py \
  tests/diffusion/models/magi2/test_ep_replicas.py \
  tests/diffusion/models/magi2/test_head_ep_activation.py \
  tests/diffusion/models/magi2/test_pipeline_magi2.py \
  tests/config/test_omni_config.py tests/config/test_config_factory.py \
  tests/config/composable_parallel/test_translator.py

test_ep_replicas.py and the new cases in test_head_parallel.py cover the replicated layout and its split sizes, the per-call fallback, a four-rank Gloo MoE forward over the SP group against the EP-group exchange, and on MUSA the production-shape row invariance and an MCCL MoE forward over the SP group against the EP groups, compared with torch.equal.

Functional run through the shared entrypoint, SP8xCFG1 with EP4:

python examples/offline_inference/text_to_video/text_to_video.py \
  --model "$MAGI2_CKPT_ROOT" --model-class-name Magi2Pipeline \
  --prompt "A golden retriever running through a sunlit meadow, cinematic camera movement" \
  --height 256 --width 448 --num-frames 125 --fps 12.5 --num-inference-steps 100 --seed 42 \
  --ulysses-degree 8 --cfg-parallel-size 1 --enable-expert-parallel --expert-parallel-size 4 \
  --extra-body '{"seconds":10,"resolution":"272p"}' --output magi2_272p.mp4

End to end: MAGI-2 Preview T2VA, the prompt above, 272p (448x256), 125 frames, 100 steps, seed 42, diffusion_compile_dynamic=False, on 8x MTT S5000. The per-step figures come from an out-of-tree timing harness around AsyncOmni with the per-step boundary from #7274; the metric is the per-step sampler.diffuse time, maximum over the 8 ranks, mean over steps 2-99.

Byte identity of commit 4: Inductor benchmarks reduction configs on every fresh compile, so the output of one code base differs across fresh caches, and TORCHINDUCTOR_DETERMINISTIC=1 does not reach the compiled regions: the MAGI-2 pipeline calls torch.use_deterministic_algorithms at construction, which resets Inductor's deterministic flag. The comparison therefore ran with TORCHINDUCTOR_DETERMINISTIC=1 TORCHINDUCTOR_FORCE_FILTER_REDUCTION_CONFIGS=1 and a fresh TORCHINDUCTOR_CACHE_DIR per run, and compares the sha256 of the generated video and audio. GEMM shape padding is still chosen by benchmarking under this setting, so differing hashes alone would not show a change in numerics; identical hashes do show identical output.

vLLM Version: 0.28.0 (MUSA image vllm:v0.28.0-ph1-5.2.0-torch2.11.0.post2-20261001), with a local test-only shim for the vLLM 0.30 and 0.31 names that c1e84ce and 198c210 import; the CPU run on the rebased branch used vLLM 0.31.0. Commit 4's A/B ran with torchada MooreThreads/torchada#120; the fixed-config byte-identity runs and the test suites also had torchada's cpp_extension path-signature fix MooreThreads/torchada#121 (host-side build paths only), without which cold-cache CPU Inductor compiles fail on MUSA. Both are merged.

vLLM-Omni Commit: 4fbea97 on top of main 198c210. The device results below were taken on c1e84ce (head 015ce08), except the second Test Result bullet, which is at 4fbea97. The changes of 4fbea97 differ from those of 362c219c0 only in parallel_state.py, where main now annotates the _PP and _SP assignments itself, so this PR no longer wraps them in cast. Against c1e84ce the branch also restores Field in the pydantic import of vllm_omni/diffusion/data.py, which main dropped as unused, and commit 5 changes only documentation. Commits 1-3 were tested at 5763f2c, which differs only in one test comment and two unreachable lines in an overridden engine method. Commit 4 was tested in an integration tree with the other pending MAGI-2 changes, where the code of this commit is identical; only the current_omni_platform import in modeling_magi2.py comes from another pending change there.

Test Result

• Rebased on main 198c210, Test Plan on CPU: 896 passed, 7 skipped (MUSA, 8-GPU and HF-config cases), 3 failed; main alone gives 801 passed, 2 skipped and the same 3 failures (test_frontend_log_stats_flag_is_not_an_unowned_stage_argument[True]/[False] and test_mps_stays_in_runtime_instead_of_engine_arguments, which need an accelerator platform).
• At 4fbea97 on 8x MTT S5000, Test Plan run offline: 904 passed, 1 skipped, 1 failed. The failure is test_ep[tencent/HunyuanImage-3.0] (downloads a config), which fails the same way on main. test_two_rank_musa_exchange, test_real_four_rank_musa_head_ep_reinitialization, test_moe_forward_over_the_sp_group_is_bit_exact_on_musa and test_owner_moe_rows_do_not_depend_on_their_batch (both token layouts) pass. SP8xCFG1 with EP4 under static compile, run interleaved with 362c219c0 (two runs each): byte-identical video and audio, and the same per-step time within run-to-run noise.
• Tests on MTT S5000 at 5763f2c, one visible device: 845 passed, 4 skipped, 8 failed. The 8 failures are tests/config/test_omni_config.py::test_diffusion_stage_payload_keys_roundtrip and fail identically on c1e84ce without this PR. With all devices visible, the two head-parallel files pass in full (44 passed), including the real MCCL tests test_two_rank_musa_exchange and test_real_four_rank_musa_head_ep_reinitialization.
• Example command above, with 4 steps: runs to completion and writes the video. test_expert_parallel_size_plumbing.py covers that --expert-parallel-size reaches the parallel config through this kwargs path.
• E2E on the stack described in [Model][Hardware][MUSA] Add MAGI-2 Preview MUSA support #8498 plus [Perf] MAGI-2: use a 64-wide K tile for the BF16 MoE GEMMs on MUSA #8506 and [Perf] MAGI-2: use FP32 mHC stream contractions on MUSA #8510 (all since merged; [Perf] MAGI-2: use FP32 mHC stream contractions on MUSA #8510 at its first commit), without the split-size passthrough of [Perf][Diffusion] MAGI-2: reuse SP MoE sequence-size metadata #7120, two interleaved runs per arm:
  • SP4xCFG2, EP off: 775.3 / 773.6 ms per step with this PR and 774.0 / 777.6 ms without it; the outputs are byte-identical.
  • SP8xCFG1, EP off (EP = SP = 8): 866.7 / 862.6 ms.
  • SP8xCFG1 with EP4: 718.3 / 716.8 ms, 17% faster than EP8 and 7% faster than SP4xCFG2. EP4 and EP8 outputs are byte-identical.
• Commit 4, tests on 8x MTT S5000 with all devices visible and a fresh Inductor cache: test_ep_replicas.py, test_head_parallel.py and test_head_ep_activation.py 75 passed, 0 skipped, including test_owner_moe_rows_do_not_depend_on_their_batch (both token layouts) and test_moe_forward_over_the_sp_group_is_bit_exact_on_musa. The MAGI-2, head-parallel, FA3-adapter and mHC suites on that tree: 818 passed, 4 failed; the 4 are test_native_compile*.py with NameError: name 'device' is not defined, which fail identically on plain c1e84ce in this image (fix(device): register the device factory as the FX device builtin MooreThreads/torchada#124 (merged) fixes the NameError; the distributed test then stops at torch_musa, which does not support FSDP2 on a CPU mesh).
• Commit 4, A/B at SP8xCFG1 with EP4, this change alone on the stack above plus commits 1-3 and the other pending MAGI-2 changes (MoE routing and route layout with per-projection shared-expert activation, [Perf] MAGI-2: bit-exact MoE route build, token-major routes and separate shared-expert activation #8592; per-request rope and modality metadata with cached sink LSE, [Perf] MAGI-2: reuse per-request rope, modality and sink-LSE metadata and trim the FA3 sink correction #8593; and an experimental step-input cache with a reused zero-padded token buffer that is not proposed upstream), one run between the two base runs, with TORCHINDUCTOR_DETERMINISTIC=1 and a fresh Inductor cache per run: 688.80 ms per step with it vs 703.14 / 700.34 without it (-12.9 ms, -1.8%). The output hashes matched the base runs; byte identity rests on the fixed-config check below, since that setting alone does not fix the reduction configs.
• Commit 4, byte identity under the fixed-config setting above: the tree with this change and the other pending MAGI-2 changes ([Perf] MAGI-2: bit-exact MoE route build, token-major routes and separate shared-expert activation #8592, [Perf] MAGI-2: reuse per-request rope, modality and sink-LSE metadata and trim the FA3 sink correction #8593, [Perf] MAGI-2: run the MUSA mHC Sinkhorn iterations in one Triton kernel #8594, [Perf] MAGI-2: drop the Ulysses send/receive buffer copies #8595 and [Perf] MAGI-2: use FP32 mHC stream contractions on MUSA #8510's second commit, now in main) gives the same video and audio sha256 as the tree without them, at SP8xCFG1 with EP4 and at SP4xCFG2 with EP4 (EP equals SP there, so the exchange path is unchanged).
• Final integrated validation on 8x MTT S5000 (100 steps, fixed reduction configs, fresh Inductor cache per run). Integrated tree: c1e84ce with [Perf] MAGI-2: count BF16 MoE routes in int32 #8496, [Bugfix] MAGI-2: keep a single W13 bank for the fused BF16 MoE path #8497, [Model][Hardware][MUSA] Add MAGI-2 Preview MUSA support #8498, [Perf] MAGI-2: use a 64-wide K tile for the BF16 MoE GEMMs on MUSA #8506, [Perf] MAGI-2: use FP32 mHC stream contractions on MUSA #8510, [Model] MAGI-2: head-sharded expert parallelism with an independent EP degree #8511, [Perf] MAGI-2: bit-exact MoE route build, token-major routes and separate shared-expert activation #8592, [Perf] MAGI-2: reuse per-request rope, modality and sink-LSE metadata and trim the FA3 sink correction #8593, [Perf] MAGI-2: run the MUSA mHC Sinkhorn iterations in one Triton kernel #8594 and [Perf] MAGI-2: drop the Ulysses send/receive buffer copies #8595 (before its dynamic-compile fix) and no other change; control: c1e84ce with [Perf] MAGI-2: count BF16 MoE routes in int32 #8496, [Bugfix] MAGI-2: keep a single W13 bank for the fused BF16 MoE path #8497, [Model][Hardware][MUSA] Add MAGI-2 Preview MUSA support #8498, [Perf] MAGI-2: use a 64-wide K tile for the BF16 MoE GEMMs on MUSA #8506, the first commit of [Perf] MAGI-2: use FP32 mHC stream contractions on MUSA #8510 and the first three commits of [Model] MAGI-2: head-sharded expert parallelism with an independent EP degree #8511. [Perf] MAGI-2: count BF16 MoE routes in int32 #8496, [Bugfix] MAGI-2: keep a single W13 bank for the fused BF16 MoE path #8497, [Model][Hardware][MUSA] Add MAGI-2 Preview MUSA support #8498, [Perf] MAGI-2: use a 64-wide K tile for the BF16 MoE GEMMs on MUSA #8506 and [Perf] MAGI-2: use FP32 mHC stream contractions on MUSA #8510 have since been merged unchanged.
  • Static compile: SP8xCFG1+EP4 714.14 / 714.86 -> 669.93 / 665.65 ms (-46.7 ms, -6.5%); SP4xCFG2 783.05 -> 729.93 ms; SP4xCFG2 with EP4 732.59 -> 690.81 ms. Video and audio sha256 identical to the control in every configuration.
  • Dynamic compile (diffusion_compile_dynamic=True, the default), SP8xCFG1+EP4: control 761.00 ms; integrated 785.71 ms; integrated without [Perf] MAGI-2: drop the Ulysses send/receive buffer copies #8595 741.57 ms; without the second commit of [Perf] MAGI-2: use FP32 mHC stream contractions on MUSA #8510 807.50 ms; without [Perf] MAGI-2: run the MUSA mHC Sinkhorn iterations in one Triton kernel #8594 759.46 ms. Without [Perf] MAGI-2: run the MUSA mHC Sinkhorn iterations in one Triton kernel #8594 the output is identical to the dynamic control, so every other change is byte-identical under dynamic compile too; with [Perf] MAGI-2: run the MUSA mHC Sinkhorn iterations in one Triton kernel #8594 the output equals the static-compile output (see [Perf] MAGI-2: run the MUSA mHC Sinkhorn iterations in one Triton kernel #8594). Only runs with identical outputs are compared for time: the eager MoE GEMM time follows the routing, which changes with the output bits.
  • Device suites on the integrated tree: 1574 passed, 55 skipped, 22 failed, none from these PRs: 9 need network access that the test container did not have (7 in test_musa_pipeline.py, 2 setup_compile tests in test_native_compile.py), 4 are the test_native_compile*.py FX NameError from torchada's torch.device replacement (fix(device): register the device factory as the FX device builtin MooreThreads/torchada#124 (merged) fixes the NameError; the distributed test then stops at torch_musa, which does not support FSDP2 on a CPU mesh), 1 is test_expert_parallel.py::test_ep[tencent/HunyuanImage-3.0] (downloads a config), and 8 are test_omni_config.py::test_diffusion_stage_payload_keys_roundtrip, which fail the same way on c1e84ce.
  • Commit 4 under dynamic compile: byte-identical to the dynamic control.
• Not run: CUDA. Commit 4 only changes the exchange on MUSA; other platforms keep the exchange inside the EP group.

AI assistance: Claude Code drafted the change, the tests and this description and ran the MUSA validation listed above. I reviewed every change in this PR and re-ran the validation listed above.

[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]

This PR appears to belong to: docs/design/module/diffusion/index.md, docs/design/module/diffusion/parallelism.md.

Module owners: @princepride @wtomin @david6666666

Routing: @princepride via module of the changed files, module named in the PR description, CODEOWNERS; @wtomin via module of the changed files, module named in the PR description, CODEOWNERS; @david6666666 via module of the changed files, module named in the PR description

@yeahdongcn, 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.

[vllm-omni-review-bot]

Omni ReviewBot routing record

Assigned Strict on zcode (GLM-5.3-Flash) under experiment fleet-strict-cursor-grok46-zcode-glm53flash-5050-c5-z10-20261002.

[vllm-omni-review-bot]

Omni ReviewBot: three questions on the performance claim

@yeahdongcn this PR reads as a performance or value claim:

• claim: - SP8xCFG1 with EP4: 718.3 / 716.8 ms, 17% faster than EP8 and 7% faster than SP4xCFG2.

Before the full evidence checklist, three short questions:

1. Bottleneck — what is the current bottleneck, and which profile, trace or per-stage measurement shows it?
2. Value — what does the change buy the user or the system (latency, throughput, memory, cost), and at which workload?
3. A/B or ablation — is there a same-workload, same-head/config comparison that isolates each main claim on its own? For stacked optimizations, one number per item rather than a blended delta.

When you answer, the evidence that settles it is: base and head SHA, hardware, model, workload, warm-up and repeat count, mean or percentiles with their spread, and a correctness/quality-equivalence signal; an end-to-end claim also needs stage attribution.

[vllm-omni-review-bot]

Omni ReviewBot attempt record

Review attempt ended as failed (step 'review' (agent.review_diff): unhandled error: RuntimeError: zcode exited 1 without a result event: statusCode: undefined } Error: Turn execution failed (traceId: c6f0781c-cfb6-4e11-8f41-e4eb08e17628) — check zcode login and the CLI version; retrying strict/zcode/GLM-5.3-Flash in 120s (try ).

[vllm-omni-review-bot]

Omni ReviewBot attempt record

Review attempt ended as failed (step 'review' (agent.review_diff): unhandled error: RuntimeError: zcode exited 1 without a result event: statusCode: undefined } Error: Turn execution failed (traceId: af45b65f-dedc-4f79-a021-72dd13a9b1da) — check zcode login and the CLI version; retrying strict/zcode/GLM-5.3-Flash in 600s (try ).

[vllm-omni-review-bot]

Omni ReviewBot attempt record

Review attempt ended as failed (step 'review' (agent.review_diff): unhandled error: RuntimeError: zcode exited 1 without a result event: statusCode: undefined } Error: Turn execution failed (traceId: 40085e0e-1c75-4057-9d85-56ee12485783) — check zcode login and the CLI version; falling back to direct/cursor/auto).

[vllm-omni-review-bot]

Omni ReviewBot review

1 actionable finding(s).

CI at 5fe48b548484 (2026-10-10T03:37:58.612346+00:00): required check(s) blocking: buildkite/vllm-omni (missing).

Note: The assigned review arm strict/zcode/GLM-5.3-Flash 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

MAGI-2 can now declare head-sharded expert parallelism and set expert_parallel_size to a divisor of the sequence-parallel degree, instead of always using one expert group per SP group. The diffusion parallel state builds those EP groups as contiguous SP subgroups, and the MAGI-2 MoE dispatches on that subgroup while the transformer passes host token counts in EP rank order. On MUSA, when several of those subgroups tile one SP group and each subgroup has the same token total, the same head shard is exchanged over the whole SP group; other platforms, and any call whose EP groups have different token totals, keep the exchange inside each EP group. With expert parallelism left off, group construction and the MoE exchange stay on the previous path.

Change flow

flowchart LR
  cli["[CHANGED] expert_parallel_size on stage config and T2V CLI"]:::changed
  groups["[CHANGED] head-EP groups as SP subgroups"]:::groups
  layout["[NEW] HeadParallelLayout packed head exchange"]:::new
  moe["[CHANGED] MAGI-2 MoE dispatch and combine"]:::changed
  out["[EXISTING] per-token MoE head rows"]:::existing
  cli --> groups --> moe
  groups --> layout --> moe --> out
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
class groups changed

Loading

See inline comments below.

---

_{🤖 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!}

[yeahdongcn]

MAGI-2's MoE has 12 heads, sharded over the EP ranks. With EP equal to SP=8 they are padded to 16, so six ranks route two heads over all of the forward's tokens and two ranks hold only padding; with EP4 every rank routes three heads over the tokens of its four-rank group, which gives the busiest rank a quarter fewer (head, token) rows. The MoE grouped GEMMs are the largest item in a profiled EP4 step, about 300 of 740 ms. Against SP4xCFG2 the profiled steps attribute the gap to the 4-way Ulysses all-to-all (about 15 ms more per step than 8-way) and to the two CFG branches carrying different token counts (about 10 ms of extra compute on half the ranks).

What it buys is per-step latency for one MAGI-2 T2VA request on 8x MTT S5000 (272p, 125 frames, 100 steps): SP8xCFG1 with EP4 runs at 717.5 ms per step, against 864.6 at SP8xCFG1 without EP and 774.4 at SP4xCFG2, the faster layout without EP. The cost is memory: each rank holds the experts of three MoE heads instead of two at EP8. With EP off nothing changes, and the output is byte-identical.

Both deltas blend two mechanisms. With EP on, the transformer passes the MoE its token counts from the host, so the EP arms skip the per-layer size all-gather and host reads that the EP-off arms still do. In the same session, passing the counts with EP off at SP4xCFG2 takes 775.8 to 733.6 ms (this PR passes them only when EP is on); against that, EP4 is 2.2% faster (717.5 vs 733.6). The head layout alone, with the counts passed in both arms: EP8 takes 12.6% longer per step than EP4 (two runs each), with byte-identical video and audio. Commit 4 has its own A/B in the body (688.80 vs 703.14 / 700.34 ms).

[hsliuustc0106]

@yeahdongcn GitHub currently reports merge conflicts for head 4fbea97c9002358450d9379e00376ab35d505282 against main. Please update the branch against current main and resolve the conflicts; then attach the new head SHA and fresh results for the affected MAGI-2 expert-parallel tests. The existing MTT S5000 performance explanation is noted; this request is specifically for the conflicting branch and refreshed evidence after reconciliation. Project CI admission and platform L2 remain maintainer-owned.

[yeahdongcn]

@yeahdongcn GitHub currently reports merge conflicts for head 4fbea97c9002358450d9379e00376ab35d505282 against main. Please update the branch against current main and resolve the conflicts; then attach the new head SHA and fresh results for the affected MAGI-2 expert-parallel tests. The existing MTT S5000 performance explanation is noted; this request is specifically for the conflicting branch and refreshed evidence after reconciliation. Project CI admission and platform L2 remain maintainer-owned.

No problem. Let me do a rebase. Thanks!

[hsliuustc0106]

Reviewed head b042caddda0dfe707b0342cbb6c2f6d2bb4d3d69.

[P2] Document the new head-EP configuration — vllm_omni/diffusion/data.py:270

expert_parallel_size changes the documented EP rule, but docs/design/feature/expert_parallel.md still universally states EP=TP×SP×CFG×DP. Please document MAGI-2's SP-local head-EP groups, default degree, divisibility constraint and TP=1 requirement.

No high-confidence runtime blocker found in the source review. The refreshed affected-test results after the latest rebase remain outstanding; CodeQL and the docs check are still pending at this check.

Validation: Python syntax parsing and git diff --check passed. This was a static review of the fork snapshot; no imports, tests or GPU benchmarks were executed.