[Misc][Main2Main] Main2Main 0911

[zhao-stack]

Fixes #17755

What this PR does / why we need it?

Advances the verified vLLM main adaptation point from September 9 to the frozen September 11, 2026, 09:00 HKT cutoff, retaining v0.28.0 support.

• Old vLLM: a97dacb7106ee49f39f3d1fc6ae1800ff724e01d.
• New vLLM: 84030bbe3d74d99bad477a3d2e37a973ccd8865c, committed 2026-09-11 08:40:10 HKT, latest available at the fixed 09:00 HKT cutoff.
• Release unchanged: v0.28.0 / 2cf0a6915ce544dc493a0990f2ea38d81601128a.
• Dependency [Misc][Main2Main] Main2main 0909 #16009 merged on 2026-09-12 at 10:38:14 HKT as 15b0cd4fa8cd62bb72120cda67319185aaf6f521.
• Rebased onto verified upstream main 3b7cf0fad5f37fb9e55d2622143819081a91e91a. Current PR head: 18cc17834a727e24cd7a5e68c0019fae53171161.
• Full PR diff: 20 files, +106/-93, now entirely the 0909-to-0911 increment. Parent commits are part of upstream main and no longer appear in this PR's commit stack.

Only three signed-off upgrade commits remain:

Commit	Increment
dcd0cbe956e81826646ffe6e47e9505fb8a0dd27	Pin, scheduler block-state, RoPE, offload and DCP adaptations, with existing dual-version UT adjustments
11fcbc61e14141b1ef3cf6f0cc9b56bfbba0ddbd	Existing compressed-prefix-cache UT call contract
18cc17834a727e24cd7a5e68c0019fae53171161	Encoder graph axis-key call compatibility and existing UT

No workflow changes, release-marker changes, new test functions, added skips, golden/threshold changes or port changes. The parent DSpark guard restoration, slot-mapping UT and subsequent upstream fixes are inherited from main. Existing upstream exclusions, including the parent's release extraction skip, are baseline content and are neither added by this PR nor counted as passing tests.

The former head 9f862c26f7f00c4974e58f20ff16588882addb68 is backed up in codex/backup-16216-parent-merged-20260912. The 36 parent replay commits and the later slot-UT replay commit were removed from this PR's stack. All 20 retained per-file added/deleted patches match the reviewed upgrade increment. Push used an exact force-with-lease against the former head; no other branch or PR was changed.

No expensive interface analyzer was run, following the user's explicit override. Exact upstream source and consumers were reviewed; the frozen target was not advanced.

Ascend file / symbol	Exact upstream cause and PR diff	Affected lane	Verification
core/recompute_scheduler.py::schedule, core/dyntra_lb_scheduler.py::schedule (including async inheritance), core/scheduler_profiling_chunk.py::schedule, patch/platform/patch_balance_schedule.py::schedule	#54853 block-state dataclass and scheduler, 0b066293f3: removes constructor block_ids; requires req_ids and lazy resolve_block_ids for every scheduled request.	Main uses scheduled IDs plus live boundary IDs and lazy lookup. Release retains partial-tail behavior. Latest-parent common kv_cache_block_copies / ec_manager_metadata handling is preserved.	Four exact producer bodies executed with the target upstream dataclass, including scheduled requests without allocation. Existing scheduler fixture/lookup assertions updated. Local exact-source checks passed again after parent-baseline reconciliation; current-head main CPU passed; see the current NPU evidence below. Release CPU remains unexecuted.
worker/model_runner_v1.py::_prepare_inputs/_dummy_run, _310p/model_runner_310p.py::_prepare_inputs	#56078 runner, 719284fe15: XD-RoPE is unified into M-RoPE; legacy XD fields are removed.	Legacy XD accesses are release-only via vllm_version_is("0.28.0"); main uses the inherited unified state.	Exact branch execution against main objects without XD fields and release objects with them. Local exact-source checks passed again after parent-baseline reconciliation; current-head main CPU passed; see the current NPU evidence below. Release CPU remains unexecuted.
spec_decode/llm_base_proposer.py::__init__/_run_merged_draft, _310p/spec_decode/llm_base_proposer_310.py::_run_merged_draft, _310p/worker/v2/rope.py::get_310p_rope_state	#56078 proposer and RoPE factory: unified M-RoPE dimensions come from model config.	Main uses mrope_num_dims; release keeps three-channel M-RoPE and legacy XD conversion.	Eight real CPU-tensor allocation/factory cases passed for main 3/4 dimensions and release 3; version guards also exercised. Local exact-source checks passed again after parent-baseline reconciliation; current-head main CPU passed; see the current NPU evidence below. Release CPU remains unexecuted.
distributed/kv_transfer/kv_pool/kv_offload/native/npu.py::NPUOffloadingSpec.create_worker/NPUTieringOffloadingSpec.create_worker	#52615 CPU spec and shared region, 138d137b5b: inherited num_blocks becomes num_chunks; shared-region kwargs become num_chunks / kv_bytes_per_chunk.	Explicit release/main fields and kwargs; local NPU worker's num_cpu_blocks contract remains unchanged.	Eight exact production-body/upstream-signature cases across both versions and replicated/sharded layouts passed. Existing offload UT fixtures and assertions updated. Local exact-source checks passed again after parent-baseline reconciliation; current-head main CPU passed; see the current NPU evidence below. Release CPU remains unexecuted.
worker/v2/model_runner.py import and NPUModelRunner.prepare_inputs	#55212 DCP helper and runner, 83990f5bcc: deletes prepare_dcp_local_seq_lens; main now computes DCP lengths after PCP/microbatch partitioning. The unconditional Ascend import would fail.	Import/call the old helper only on v0.28.0. Main leaves the field unset during preparation and uses inherited execute_model to compute it from the final batch and persistent buffer.	Eight isolated exact-source cases passed: release/main import and call decisions, PCP/global storage selection, padded extent, and dummy-skip behavior. Local exact-source checks passed again after parent-baseline reconciliation; current-head main CPU passed; see the current NPU evidence below. Release CPU remains unexecuted.
Existing test_dyntra_lb_scheduler, test_scheduler_connector_block_state, test_native_cpu_offload, test_npu_offload_spec, test_eagle_proposer	The same #54853, #52615, and #56078 ModelConfig contracts.	Fixture constructor names, lazy lookup assertions and exact version properties only; no new test functions.	Python parsing and an AST comparison against the new upstream base confirm no added test functions. Current-head main CPU passed; see the current NPU evidence below. Release CPU remains unexecuted.
tests/ut/test_compressed_prefix_cache.py: the three existing logical-hash/cache-hit/hybrid-prefix cases	#54713 cache manager diff, b28c3e1568: cache_blocks replaces scalar replay_boundary with sequence replay_boundaries.	Main passes (logical_block_size,); v0.28.0 continues omitting the keyword. Production inherits the upstream implementation unchanged.	Previous CPU run exposed exactly these 3 stale calls. Six exact release/main signature bindings passed after the fix; no new UT functions. Local exact-source checks passed again after parent-baseline reconciliation; current-head main CPU passed; see the current NPU evidence below. Release CPU remains unexecuted.
worker/encoder_acl_graph.py::EncoderAclGraphManager._capture_budget_graph/_run_budget_graph and existing test_capture_budget_graph_npu	#42785 encoder manager diff, 442d36031c: inherited capture/replay always pass axis_keys=(), even for models without extra capture axes. Ascend overrides rejected the keyword.	Both overrides accept the optional empty tuple and retain their common existing execution. Release callers omit it; the existing UT selects the exact call shape with vllm_version_is. Non-empty capture axes are explicitly unsupported because FIA bookkeeping remains keyed by token budget; this optional new feature is not silently enabled.	Exact upstream capture body and replay call executed both Ascend overrides against release/main metadata and model signatures, with real CPU tensors and mocked NPU primitives; passed. Two non-empty-axis rejection checks passed. Existing UT updated, no new test function. Current-head main CPU passed; see the current NPU evidence below. Release CPU remains unexecuted.

Does this PR introduce any user-facing change?

Updates main compatibility while retaining release v0.28.0. Optional upstream features without a demonstrated Ascend contract break are not separately ported.

How was this patch tested?

CI run 34668925819, head 18cc17834:

• vLLM main (84030bbe3): CPU 4292 passed / 38 skipped; NPU 539 passed / 32 skipped.

• vLLM v0.28.0 (2cf0a6915): NPU 535 passed / 36 skipped; CPU not run.

• Pre-commit and all 32 NPU shards passed.

• vLLM main: vllm-project/vllm@a97dacb

[gemini-code-assist]

Summary of Changes

Hello, I'm Gemini Code Assist¹! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!

This pull request updates the Ascend-specific vLLM adaptations to align with the latest upstream main branch as of September 10, 2026. The changes focus on reconciling breaking API shifts in schedulers, KV cache management, and RoPE implementation while ensuring that the v0.28.0 release baseline remains functional. No new features were added; the focus is strictly on maintaining parity with upstream main while preserving Ascend-specific performance optimizations and hardware contracts.

Highlights

• Main Pin Update: Advanced the upstream main pin to commit 6e1d051fc942cc83b72ee5a173786baf6666700a while maintaining compatibility with v0.28.0 and existing parent adaptations.
• Scheduler and Dataclass Adaptations: Updated scheduler logic and block-state handling to accommodate upstream changes in block-id resolution and dataclass structures, ensuring lazy resolver contracts are preserved.
• RoPE and Model Runner Refactoring: Adapted RoPE factory and model runner inputs to use model-derived M-RoPE dimensions, removing legacy XD-RoPE fields as per upstream main changes.
• KV Offloading and Pool Management: Updated NPU offloading specifications to handle renamed fields (num_blocks to num_chunks) and adjusted mmap kwargs to maintain consistency across versions.
• Test Infrastructure: Added new regression tests for circular buffer slot mapping and expanded test coverage for RoPE and speculative decoding contracts.

New Features

🧠 You can now enable Memory (public preview) to help Gemini Code Assist learn from your team's feedback. This makes future code reviews more consistent and personalized to your project's style. Click here to enable Memory in your admin console.

Ignored Files

• Ignored by pattern: .github/workflows/** (2)
  • .github/workflows/pr_test.yaml
  • .github/workflows/scripts/test_config.yaml

Using Gemini Code Assist

The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.

Invoking Gemini

You can request assistance from Gemini at any point by creating a comment using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands on the current page.

Feature	Command	Description
Code Review	/gemini review	Performs a code review for the current pull request in its current state.
Pull Request Summary	/gemini summary	Provides a summary of the current pull request in its current state.
Comment	@gemini-code-assist	Responds in comments when explicitly tagged, both in pull request comments and review comments.
Help	/gemini help	Displays a list of available commands.

Customization

To customize the Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.

Limitations & Feedback

Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counterproductive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for GitHub and other Google products, sign up here.

Footnotes

1. Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

[github-actions]

👋 Hi! Thank you for contributing to the vLLM Ascend project. The following points will speed up your PR merge:‌‌

• A PR should do only one thing, smaller PRs enable faster reviews.
• Every PR should include unit tests and end-to-end tests ‌to ensure it works and is not broken by other future PRs.
• Write the commit message by fulfilling the PR description to help reviewer and future developers understand.

If CI fails, you can run linting and testing checks locally according Contributing and Testing.

---

Tip

💡 Consider Linking a Related Issue or RFC

Your PR title contains the [BugFix] tag, indicating a bug fix or new feature.

Linking a related issue or RFC in the PR description is strongly encouraged — it gives reviewers helpful context and speeds up the review. You can use any of these keywords:

• Fixes #<issue_number>
• Closes #<issue_number>
• Resolves #<issue_number>
• Refs #<rfc_or_issue_number> (for RFCs)

🙏 Thanks for helping us keep the project well-organized!

[gemini-code-assist]

Code Review

This pull request introduces support for vLLM v0.28.0 by adding version-gating logic across the codebase, particularly in the Ascend NPU worker, scheduler, and attention implementations. It refactors internal contracts to handle differences between the release version and the main branch, including adjustments to KV cache management, block table layouts, and speculative decoding. Several new tests were added to verify these version-specific behaviors, and I have no further feedback to provide.

[zhao-stack]

/rerun
[Bot]: rerun completed.

Rerun (failed jobs only):

• E2E

[Ronald1995]

LGTM