[BugFix][MoE] Align hash routing input IDs for PCP - #16834
Merged
yiz-liu merged 1 commit intoSep 20, 2026
Merged
Conversation
Contributor
|
👋 Hi! Thank you for contributing to the vLLM Ascend project. The following points will speed up your PR merge:
If CI fails, you can run linting and testing checks locally according Contributing and Testing. Tip 💡 Consider Linking a Related Issue or RFCYour 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:
🙏 Thanks for helping us keep the project well-organized! |
There was a problem hiding this comment.
🟡 Changes recommended
The PCP gather path can fail for SP and duplicate collectives for SP+DP/PCP, preventing reliable ID alignment.
Get a fresh assessment by requesting another Copilot review.
Pull request overview
This PR aligns hash-routing input_ids with AllGather MoE layouts across DP and PCP parallelism.
Changes:
- Renames and extends the input-ID gathering helper.
- Updates hash routing to use DP-then-PCP gathering.
- Expands regression coverage across parallel layouts.
File summaries
| File | Summary |
|---|---|
vllm_ascend/ops/fused_moe/router/fused_topk_router.py |
Uses the renamed input-ID gather helper. |
vllm_ascend/ops/fused_moe/prepare_finalize.py |
Adds DP/PCP-aligned input-ID gathering; the SP+PCP path has a critical unresolved issue. |
tests/ut/ops/test_prepare_finalize.py |
Tests input-ID alignment across parallel layouts. |
tests/ut/ops/test_fused_moe.py |
Updates hash-router regression coverage. |
Review details
- Files reviewed: 4/4 changed files
- Comments generated: 1
- Review effort level: Lite
💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.
Contributor
Summary of ChangesHello, 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 addresses a bug in the MoE hash routing mechanism where input IDs were not correctly aligned with gathered router logits when using PCP (Parallel Column Partitioning). By synchronizing the communication layout for input IDs with the existing AllGather prepare path, the fix ensures that hash-based routing functions correctly across various parallel configurations. Highlights
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. Using Gemini Code AssistThe 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
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 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
|
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
Contributor
There was a problem hiding this comment.
Code Review
Suggested PR Title:
[Ops][Feature] Support PCP gathering for input IDs in fused MoESuggested PR Summary:
### What this PR does / why we need it?
This pull request updates the fused MoE routing logic to support PCP (Pipeline Cell Parallelism) gathering for input IDs. Specifically, it renames `all_gather_input_id_with_dp_group` to `all_gather_input_ids` in `PrepareAndFinalizeWithAllGather` and adds support for gathering input IDs across the PCP group when `pcp_size > 1`.
However, there is a critical issue when sequence parallelism is enabled (`self._use_ep_sequence_parallel()` is `True`). In this case, `self.num_tokens_pcp` is undefined, and `self.num_tokens` is set to the post-gather token count, which will cause incorrect padding or an `AttributeError`. We should handle sequence parallelism by using `torch.ops.vllm.maybe_all_gather_and_maybe_unpad` to gather `input_ids` across the EP group.
### Does this PR introduce _any_ user-facing change?
No.
### How was this patch tested?
The changes are tested by updating existing unit tests in `tests/ut/ops/test_fused_moe.py` and expanding `test_allgather_prepare_finalize` in `tests/ut/ops/test_prepare_finalize.py` to cover various combinations of DP and PCP sizes.
Contributor
|
This pull request has conflicts, please resolve those before we can evaluate the pull request. |
1 similar comment
Contributor
|
This pull request has conflicts, please resolve those before we can evaluate the pull request. |
li1how
force-pushed
the
bugfix/pcp-hash-router-input-ids
branch
from
September 18, 2026 07:09
8add7b2 to
528573c
Compare
nwpu-zxr
approved these changes
Sep 18, 2026
Contributor
|
This pull request has conflicts, please resolve those before we can evaluate the pull request. |
Signed-off-by: leolee <yihao.li@huawei.com>
li1how
force-pushed
the
bugfix/pcp-hash-router-input-ids
branch
from
September 19, 2026 06:06
528573c to
4c7eca1
Compare
Contributor
Author
|
/rerun Rerun (failed jobs only):
|
yiz-liu
approved these changes
Sep 20, 2026
ZT-AIA
pushed a commit
that referenced
this pull request
Sep 20, 2026
) ### What this PR does / why we need it? Revert #16393 at the author's request. This reverses squash commit `af277b2681443089e4d29684b1101707e2eb6426` on upstream main `8038f64129a1f8751ac6684f47c8d058208e4ff3`, restoring the previous fixed-main + v0.28.0 support contract. The revert applies without conflicts. Subsequent upstream patches are retained, including #16775 (PCP/DCP slots), #16923 (A3 SFA prefill all-to-all), and #16834 (PCP hash-routing input IDs). No unrelated fixes are included. | Reverted change | Reason and source | Lane | | --- | --- | --- | | Release marker | Reverse [#16393](https://github.com/vllm-project/vllm-ascend/pull/16393/files): v0.29.0 (`98dff2a81d747d1dba01a47f939f48c3526d4206`) to v0.28.0 (`2cf0a6915ce544dc493a0990f2ea38d81601128a`) | Release | | Version branches and interface adapters across scheduler, KV cache, attention, model runners and speculative decoding | Restore the pre-#16393 v0.28.0 contracts by reversing the exact merged diff; keep main implementations behind their original version selection | Main + release | | v0.29.0-only parallel-config patch, registration and documentation | Remove the patch introduced by #16393; retain the independent upstream v0.28.0 PCP+DP workaround | Release | | Existing unit/e2e tests and release-only skips | Restore the tests and conditions changed by #16393, without introducing additional skip scope beyond the reverted baseline | Main + release | The vLLM main pin remains `84030bbe3d74d99bad477a3d2e37a973ccd8865c`. No main interface scan is needed for this release-only revert. The only changes beyond the inverse patch are formatting in `tests/ut/test_utils.py` and line-ending normalization in `categorical_sample.py` needed for formatting/whitespace checks. Workflows are unchanged. ### Does this PR introduce _any_ user-facing change? Yes. The supported release returns to vLLM v0.28.0; v0.29.0 support and its specific compatibility changes from #16393 are reverted. The vllm-ascend package version and fixed vLLM main pin are unchanged. ### How was this patch tested? - `git revert --no-commit af277b2` applied cleanly on the latest fetched main. - Verified that the complete post-#16393 upstream patch can still reverse-apply to the staged result (`git apply --reverse --check --cached`), confirming subsequent changes are preserved. - `git diff --cached --check`, Python AST parsing, Ruff lint and formatting checks passed for all 94 remaining changed Python files. - Ran the equivalent of `format.sh ci`: `pre-commit run --all-files --hook-stage manual`. Ruff, codespell, typos, clang-format, markdownlint, actionlint, package-init, forbidden-import and boolean-context checks passed. Bash-based hooks could not execute on this Windows host; two Python launcher hooks exited 9009. Full pre-commit validation remains for CI. - Actual CPU/NPU tests have not been run locally because the required vLLM/NPU runtime is unavailable. New PR CI is pending; the original upgrade PR's green run is not validation of this revert. - vLLM main: vllm-project/vllm@84030bb Signed-off-by: shenzhao <shenzhao9@huawei.com> Co-authored-by: shenzhao <shenzhao9@huawei.com>
ningjingbengxiaohai
pushed a commit
that referenced
this pull request
Sep 20, 2026
…#17004) ### What this PR does / why we need it? Reintroduce the vLLM v0.29.0 release upgrade from #16393, reverted by #16949, with container defaults aligned to the supported release. Based on upstream main `d3f6b4b59ab55dc42af1f30f425c893b95af5ee2`; preserves later merged changes. The image/source mismatch is confirmed in [nightly job 106004497020](https://github.com/vllm-project/vllm-ascend/actions/runs/35482881089/job/106004497020): the image installed `vllm 0.28.0+empty` (tag `v0.28.0`) while upgraded Ascend failed importing `_get_packed_kv_cache_groups`. All eight root Dockerfiles still defaulted to v0.28.0. The reusable image workflow passes no `VLLM_TAG` override, so those defaults govern release builds. Changing only the release marker does not update those images. - Fixed vLLM main remains `84030bbe3d74d99bad477a3d2e37a973ccd8865c`. - Release changes from v0.28.0 (`2cf0a6915ce544dc493a0990f2ea38d81601128a`) to official v0.29.0 (`98dff2a81d747d1dba01a47f939f48c3526d4206`, rechecked against the remote tag). - Restore #16393's source compatibility and existing UT changes by reversing #16949, then reconcile current upstream changes. No main interface scan is rerun for this release-only upgrade. - Update `VLLM_TAG` in all eight Dockerfiles: default/A3/A5/310P, Ubuntu/openEuler. Preserve existing exact-commit build overrides. No workflow changes. | Current-base adaptation | Exact cause and evidence | Lane | Validation | | --- | --- | --- | --- | | Eight Dockerfile `VLLM_TAG` defaults and release marker | #16393 changed the release contract but omitted image defaults; the linked nightly log proves installation of 0.28.0 and missing `_get_packed_kv_cache_groups`. | Release images | All eight defaults match marker; image-build CI requested, pending | | `worker/v2/attn_utils.py::get_kv_cache_spec` and existing `_make_mla_layer` UT fixture | Ascend [#16755](https://github.com/vllm-project/vllm-ascend/pull/16755/files), `df3e755e98fba8c6a18f200c645e0c2050469bc3`, added an old-release field selector. vLLM [#51718](https://github.com/vllm-project/vllm/pull/51718/files), `8bdc70ec7b379279ec0152343239c2d50aced687`, replaced `compress_ratio` with `tokens_per_state`; both exact supported pins use the latter. Use the common field, retaining metadata and cache-view assertions. | Both | Source inspection and static checks passed; actual CI pending | | `attention/attention_v1.py` import conflict | Preserve `attention_transfer_window` from Ascend [#15636](https://github.com/vllm-project/vllm-ascend/pull/15636/files), `5c80630f28f8529aa82716e58b981a78819ec429`, and graph changes from #16798; keep the common relocated PCP import from #16393. Do not restore the old compute-start import or unused weak-reference import. | Both | Conflict resolved; static checks passed | | `worker/v2/aclgraph_utils.py` import conflict | Preserve `ContextSource`, `UpdatableGraph`, and `use_updatable_graph` from Ascend [#16798](https://github.com/vllm-project/vllm-ascend/pull/16798/files), `34bb51f93724c565362f5108f5226303e1b56cad`, while removing the obsolete 0.28 version selector restored by the revert. | Both | Conflict resolved; static checks passed | | Existing 310P and Mamba model-runner UT imports | Preserve hardware-profile imports and mocks from Ascend [#16803](https://github.com/vllm-project/vllm-ascend/pull/16803/files), `d3f6b4b59ab55dc42af1f30f425c893b95af5ee2`; remove only the unused old-release selector import. Hardware capability routing remains unchanged. | Both | Real rebase conflicts resolved; AST/Ruff/format passed; fresh CI pending | Newly merged changes were reviewed for version-contract impact: #16775/#16834 PCP metadata and routing, #16923 A3 SFA, #16426/#16924/#16955 custom ops, #16081 MTP/SP, #16669 xlite, #15636/#16747 transfer, #16913 A5 pages, #16798 graph updates, #16755 GLM PD, #16952 C8 config, #16673 operator removal and #16320 Kimi-K3 KV pool. Preserve these changes; no additional source-proven version branch was identified beyond the entries above. In particular, #16747's `UniformTypeKVCacheSpecs.kv_cache_specs` and per-layer `block_size` exist in both exact pins; #16320 adds Ascend connector hooks. CI remains necessary to validate runtime interactions. CI/documentation-only PRs are retained unchanged. <details> <summary>Inherited per-file compatibility evidence from #16393</summary> The following source-contract ledger is inherited from #16393. Any historical verification wording refers only to that earlier PR; it does not certify this new head. New-head validation is listed below. | Reference | Commit | |---|---| | vllm-ascend rebase base | `9dc6704559ebe2809b5cb7b7fc44184bd1338b3c` | | PR head | `8c5d785931c495701bf1da8b5bc80b321ebd3cd0` | | Fixed vLLM main | `84030bbe3d74d99bad477a3d2e37a973ccd8865c` | | Previous release: v0.28.0 | `2cf0a6915ce544dc493a0990f2ea38d81601128a` | | Target release: v0.29.0 | `98dff2a81d747d1dba01a47f939f48c3526d4206` | - Use common implementations where v0.29.0 and the fixed main share KV-cache layouts, Mamba copy/group APIs, PCP handling and speculative-decoding contracts. - Retain explicit `vllm_version_is("0.29.0")` branches for contracts that still differ, including RoPE, scheduler block snapshots, InputBatch, ReplaySSM, KV zeroing and DSpark PP handling. - Remove obsolete v0.28.0 compatibility and adapt existing test fixtures. Version detection uses package versions and the explicit `VLLM_VERSION` override, with local-version suffix handling and UT environment isolation retained; no hard-coded release-SHA inference remains. - Apply the Ascend PCP+DP validator patch only to v0.29.0, preserving other validations and EPLB platform binding. Rebuild dependent Pydantic schemas so nested configuration validation uses the patched validator. The global patch documentation records its rationale and removal criteria. - Preserve #15747's Spec+PP protocol/partition handling after rebase; use the common exact-release selector and the real function-local DSpark sharing import. This release-only upgrade does not require a new main old-to-new interface scan. The latest rebase incorporates [#16905](https://github.com/vllm-project/vllm-ascend/pull/16905/files), which reverted #16544 at `9dc6704559ebe2809b5cb7b7fc44184bd1338b3c`. The now-unnecessary V4.1 drafter import gate and tuple annotation have been removed. Other release adaptations, including #15747 Spec+PP handling, remain. Detailed contract evidence is retained below for review. <details> <summary>Per-file adaptations and exact upstream evidence</summary> #### Per-file adaptation ledger Evidence IDs refer to the exact source contract and upstream diff table below. Every row is syntax checked; branch-normalized AST comparison confirms unchanged main function bodies except the version identity helper and the explicitly retired propose argument/type annotations. Both supported versions completed CPU and NPU execution as recorded below. | Ascend file / symbols | Disposition and upstream evidence | Verification | |---|---|---| | `vllm_ascend/patch/worker/patch_v2/patch_dspark.py` / `_load_dspark_model_with_target_quant`; `tests/ut/patch/worker/test_patch_dspark_pp.py` | Preserve rebased [Ascend #15747](https://github.com/vllm-project/vllm-ascend/pull/15747/files#diff-a170a42fe1e9275c999642b05c4a437d0c9104dd8ea19151796142909b00ab2a) (`82df9d871`), including manual PP partition masking. vLLM [#52809 diff](https://github.com/vllm-project/vllm/pull/52809/files#diff-2223c5e6b44b302db9bff4023777cc88b4427f4fb9d4a04476f6d6e4bd687393) (`91a893de64722019ea2faf852e06cabe143b3490`) moved `_should_share` inside the loader on both supported pins, so retain the earlier release fix: patch `eagle_utils`, never read/patch a nonexistent `dspark_utils._should_share`. v0.29 keeps its global PP guard; main has native PP via [#50514](https://github.com/vllm-project/vllm/pull/50514/files#diff-2223c5e6b44b302db9bff4023777cc88b4427f4fb9d4a04476f6d6e4bd687393) (`d87a440f88e28e5b37f9b1e22ce214d0426d5352`). | Existing UT now checks the real function-local import, absent module alias, and restoration on success/failure for both lanes; partition and PP assertions retained. Ruff/syntax pass; actual CPU/NPU pending. | | `vllm_ascend/worker/v2/pp_utils.py` / `use_legacy_spec_pp`; `tests/ut/worker/v2/test_pp_utils.py` | #15747 added broad 0.28/0.29 routing and an obsolete 0.28 dev-build recognition path. For the two supported pins, #50514 exists only on fixed main. Route through `vllm_version_is("0.29.0")`; retain package/local-suffix and explicit environment-override semantics. No release-SHA inference or third release lane. | Existing UTs exercise the real uncached version helper with monkeypatch isolation, local suffix, fixed-main dev string, explicit override, and non-target versions. Isolated routing checks and Ruff/syntax pass; full CPU UT pending. | | `vllm_ascend/_310p/model_runner_310p.py`<br>`_prepare_inputs`, `_allocate_kv_cache_tensors` | #51718, #56078; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/_310p/spec_decode/llm_base_proposer_310.py`<br>`set_inputs_first_pass` | #56078; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/_310p/worker/v2/model_runner.py`<br>`initialize_kv_cache`, `_allocate_kv_cache_tensors`, `_prepare_inputs_310p` | #51718/#54436; use shared standardized layouts and retain the v0.29.0 InputBatch gate. The rebase preserves vllm-ascend #16043's MTP copy tracking while removing only legacy v0.28.0 allocation paths. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/_310p/worker/v2/rope.py`<br>`get_310p_rope_state` | #56078; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/attention/attention_v1.py`<br>`module imports/dispatch` | #52839; both lanes use the common PCP import. Rebase keeps current upstream graph code and removes only the v0.28.0 import branch. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/attention/context_parallel/sfa_cp.py`<br>`module imports/dispatch` | #52839; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/attention/indexer.py`<br>`module imports/dispatch` | #52839; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/attention/mla_v1.py`<br>`module imports/dispatch` | #52839; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/core/dyntra_lb_scheduler.py`<br>`module imports/dispatch` | #51358, #54853; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/core/recompute_scheduler.py`<br>`module imports/dispatch`, `schedule` | #51358, #54853; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/core/scheduler_profiling_chunk.py`<br>`module imports/dispatch` | #51358, #54853; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/core/kv_cache_interface.py`<br>`get_kv_cache_compression_ratio`, `AscendMLAAttentionSpec`, `get_storage_block_size` | #51718, #53906; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/distributed/kv_transfer/kv_pool/ascend_store/layerwise_cache_layout.py`<br>`_merge_specs` | #51718; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/distributed/kv_transfer/kv_pool/kv_offload/preempt_offload/manager.py`<br>`_derive_cpu_config` | #51718; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/distributed/kv_transfer/kv_pool/kv_offload/native/npu.py`<br>`create_worker` | #52615; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/models/deepseek_mtp.py`<br>`load_weights` | #53106; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/models/deepseek_v4/indexer.py`<br>`get_kv_cache_spec` | #51718; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/models/glm5next/cache_config.py`<br>`make_tensor` | #51718; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/models/glm5next/kv_cache.py`<br>`get_kv_cache_spec` | #51718; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/models/kimi_k3_dspark.py`<br>`load_weights` | #53106; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/models/layer/attention/layer.py`<br>`get_kv_cache_spec` | #51718; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/patch/platform/patch_balance_schedule.py`<br>`module imports/dispatch` | #51358, #54853; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/patch/platform/patch_kv_cache_coordinator.py`<br>`__init__` | #53614; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/patch/platform/patch_kv_cache_utils.py`<br>`module imports/dispatch`, `_get_max_layers_per_page_size`, `_ascend_max_memory_usage_bytes_from_groups`, `_ascend_get_kv_cache_config_from_groups` | #51718; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/patch/platform/patch_use_v2_model_runner.py`<br>`module imports/dispatch`, `_patched_get_unsupported_features` | #53853, #53183; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/patch/worker/patch_bind_kv_cache.py`<br>`bind_kv_cache` | #51718; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/patch/worker/patch_bind_kv_cache.py::bind_kv_cache` | v0.29.0 has no ReplaySSM ring-tracker helper; fixed main requires it after KV binding. Evidence: [#52506, `adebc41b7e9f1085d3f73434e23beb76883b9eb4`, worker-utils diff](https://github.com/vllm-project/vllm/pull/52506/files#diff-645d58630d5acf3a0b07226bfef1e890a584c32502ab97c3d4642070f39a783c). Keep an explicit v0.29.0 exclusion and preserve the fixed-main call. | Failure reproduced on v0.29.0; Historical release/main NPU validation passed; current-head CI pending | | `vllm_ascend/patch/worker/patch_mamba_utils.py`<br>`module imports/dispatch`, `_get_state_copy_funcs_for_layer` | #53896; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/patch/worker/patch_v2/patch_attn_utils.py`<br>`module imports/dispatch` | #51718; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/patch/worker/patch_v2/patch_dspark.py`<br>`_load_dspark_model_with_target_quant` | #50514, #52809; v0.29.0 keeps `get_pp_group` module-bound but imports `_should_share` locally from Eagle utilities; fixed main removes the PP guard. Keep the release `get_pp_group` patch, share through the common Eagle utility, and delete the obsolete v0.28.0 `dspark_utils._should_share` patch. | Exact release failure reproduced; Historical release/main CPU/NPU validation passed; current-head CI pending | | `vllm_ascend/spec_decode/llm_base_proposer.py`<br>`model_returns_tuple`, `__init__`, `set_inputs_first_pass` | #56078, #52861; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/utils.py`<br>`get_kv_cache_tensor_layers`, `register_ascend_customop`, `vllm_version_is` | #52839, #51718, #52494; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/worker/model_runner_v1.py`<br>`_get_ascend_mamba_state_copy_funcs`, `_allocate_kv_cache_tensors`, `_prepare_inputs`, `_dummy_run`, `_reshape_kv_cache_tensors`, `get_kv_cache_spec`, NaN helpers | #51718/#53896/#56078 plus #50323; remove v0.28.0 layouts/copy APIs and import the NaN helpers directly because v0.29.0 and fixed main expose the same contract. XD-RoPE remains explicitly gated to v0.29.0. | Static checked; historical dual-version CPU/NPU validation passed; current-head CI pending | | `vllm_ascend/worker/v2/aclgraph_utils.py`<br>`capture` | #53869 makes `pcp_manager` common to both lanes. The rebase preserves vllm-ascend #16409's host-parameter-update revert and removes only the obsolete v0.28.0 capture branch. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/worker/v2/attn_utils.py`<br>`get_kv_cache_spec`, `_allocate_kv_cache`, `_reshape_kv_cache_v2` | #51718; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/worker/v2/block_table.py`<br>`__init__`, `init_block_table_layout_tensors`, `compute_slot_mappings` | #51718, #51031; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/worker/v2/model_runner.py`<br>`module imports/dispatch`, `prepare_inputs`, `__init__`, `sample_tokens`, `prepare_dummy_attn`, `execute_model` | #50514, #54436, #52506, #55212, #53515; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/worker/v2/pcp_manager.py`<br>`partition_batch` | #53515; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/worker/v2/sample/gumbel.py`<br>`module imports/dispatch` | #54282; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/worker/v2/spec_decode/__init__.py`<br>`init_speculator` | #49811; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/worker/v2/spec_decode/autoregressive/speculator.py`<br>`propose` | #53694; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/worker/v2/spec_decode/dflash/speculator.py`<br>`module imports/dispatch`, `propose` | #53694, #52188; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/worker/v2/spec_decode/dspark/speculator.py`<br>`propose` | #53694; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/worker/worker.py`<br>`_scale_kv_cache_memory_for_multi_group`, `wake_up` | #51718, #53508; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | #### Exact upstream evidence | Upstream change | Full commit SHA / direct diff | Actual supported contracts and branch decision | |---|---|---| | #51718 | `8bdc70ec7b379279ec0152343239c2d50aced687`<br>[vllm/v1/kv_cache_interface.py](https://github.com/vllm-project/vllm/pull/51718/files#diff-f76cdfbf02dacd9dffbfbd0d9ad68a7a6ac0d8aed70834f95a2ae8ccd2e333cb) | Both use layers/layer_stride/block_stride/offset, tokens_per_state, CircularBufferSpec and standardized backing; retire shared_by/compress_ratio allocation branches. | | #52839 | `58e5ee0158b6a264c3506f00480e108a34b33ee3`<br>[vllm/v1/attention/ops/pcp.py](https://github.com/vllm-project/vllm/pull/52839/files#diff-23d7e7ab01a7f44414796e1e1b08cfa5277a05f54143ee2dd3a1041ca47772a8) | Both import PCP operations from vllm.v1.attention.ops.pcp. | | #53896 | `e126687a9a828d513c01a07cd69f025f27d63280`<br>[vllm/v1/worker/mamba_utils.py](https://github.com/vllm-project/vllm/pull/53896/files#diff-8ab86225fa08e9a6700851191abcb3f4f8b0bba40e97a82aad700d7a23a4d7ec) | Both use Mamba copy-function dictionaries and unwrap UniformTypeKVCacheSpecs. | | #53106 | `1fe3a1571ac67581478a11743e55a306de1d136f`<br>[vllm/model_executor/models/utils.py](https://github.com/vllm-project/vllm/pull/53106/files#diff-b0ba1095e9881e5c87e33dfd20958d1e1ceafe8a4433aa692f468e61be130b21) | Both use WeightsMapper instead of AutoWeightsLoader skip_prefixes/skip_substrs. | | #53906 | `98ed0856f31fa3aaf5e27464e2b4ef5a8ee6b2f5`<br>[vllm/v1/kv_cache_interface.py](https://github.com/vllm-project/vllm/pull/53906/files#diff-f76cdfbf02dacd9dffbfbd0d9ad68a7a6ac0d8aed70834f95a2ae8ccd2e333cb) | Only pinned main has the optional MLA storage_block_size dataclass field; release keeps the Ascend derived property, using tokens_per_state. | | #56078 | `719284fe158f1be8a9dd92953295fc9d49015730`<br>[vllm/config/model.py](https://github.com/vllm-project/vllm/pull/56078/files#diff-998c640befaf137b9af825f29f4e6e47d273caab1fd04093c97df24b18f5c417) | Release retains uses_xdrope_dim and three M-RoPE dimensions; pinned main uses mrope_num_dims and unified RoPE. | | #52615 | `138d137b5b955a5ebee98dfc946ecdd65d7b87ce`<br>[vllm/v1/kv_offload/cpu/spec.py](https://github.com/vllm-project/vllm/pull/52615/files#diff-7d073c26c9b22f74e9ff9e4c8da733d1807233edb83439fc64fbd48c43ae6045) | Release uses num_blocks/kv_bytes_per_block; main uses num_chunks/kv_bytes_per_chunk. | | #51358 | `6b110badbb22d3f66c7218b71138f13b7a6b3419`<br>[vllm/v1/core/sched/output.py](https://github.com/vllm-project/vllm/pull/51358/files#diff-cafd89ce8a698a56acb24ada62831cbc7a980782f78a52d1742ba238031f296c) | Release now has boundary_state_offloads and KVConnectorBlockState; remove partial_tail_offloads plumbing. | | #54853 | `0b066293f3c738a0cbd3a087bf893f2f4dcd61f2`<br>[vllm/v1/core/sched/output.py](https://github.com/vllm-project/vllm/pull/54853/files#diff-cafd89ce8a698a56acb24ada62831cbc7a980782f78a52d1742ba238031f296c) | Release constructor takes block_ids snapshots; main takes req_ids and resolve_block_ids. Keep exact release snapshot membership and main lazy-resolution membership. | | #53614 | `144e79c8106da23141ac010394b782f730cc7fe8`<br>[vllm/v1/core/kv_cache_coordinator.py](https://github.com/vllm-project/vllm/pull/53614/files#diff-43875c71daa893ef7567e21633d9988c2baf95bef61e3a334a6d584d6444d725) | Only main configures drop_eagle_checkpoint_block for replay-aligned Mamba checkpoints. | | #50514 | `d87a440f88e28e5b37f9b1e22ce214d0426d5352`<br>[vllm/v1/worker/gpu/spec_decode/dspark/utils.py](https://github.com/vllm-project/vllm/pull/50514/files#diff-2223c5e6b44b302db9bff4023777cc88b4427f4fb9d4a04476f6d6e4bd687393) | Release retains module-level PP/share symbols and Ascend PP workaround; main has the subsequent PP integration. | | #52809 | `91a893de64722019ea2faf852e06cabe143b3490`<br>[vllm/v1/worker/gpu/spec_decode/dspark/utils.py](https://github.com/vllm-project/vllm/pull/52809/files#diff-2223c5e6b44b302db9bff4023777cc88b4427f4fb9d4a04476f6d6e4bd687393) | Between v0.28.0 and v0.29.0, `_should_share` moved from a DSpark module binding to a function-local import from Eagle utilities. The shared Eagle patch remains effective; the old DSpark-module read/write must be removed. | | #54436 | `6bafc049aae6c26e210162630914ee9177a4b586`<br>[vllm/v1/worker/gpu/input_batch.py](https://github.com/vllm-project/vllm/pull/54436/files#diff-106f39c08266f186830bb8fcd7fb1df35c0aaa5fd0ac5c17aac64aeddee48902) | Release InputBatch requires max_seq_len_np; main removed it. | | #52506 | `adebc41b7e9f1085d3f73434e23beb76883b9eb4`<br>[vllm/v1/worker/gpu/model_runner.py](https://github.com/vllm-project/vllm/pull/52506/files#diff-5823f988fc0264681a80db24ccaba4d364f14394815d83ec1d90944c09f571f0) | Only main accepts valid_dummy_state_slots/valid_state_slots capture arguments. | | #55212 | `83990f5bcc0b5eb08b2f1fd1b109fa8a31b122f3`<br>[vllm/v1/worker/gpu/model_runner.py](https://github.com/vllm-project/vllm/pull/55212/files#diff-5823f988fc0264681a80db24ccaba4d364f14394815d83ec1d90944c09f571f0) | Release prepares DCP local sequence lengths before partitioning; main initializes DCP metadata afterwards. | | #53515 | `b1fbbc2ade51e3826bc92e4733c9c692ee21d42d`<br>[vllm/v1/worker/gpu/pcp_manager.py](https://github.com/vllm-project/vllm/pull/53515/files#diff-45dad5fe975c2ec8d02124fa65579cf11bfe9169a94217ceae1de0ea29518127) | Both accept padded_num_tokens for persistent PCP input buffers. | | #53869 | `b3af042abd8fe5a297ec3ec72db276fd661a67b3`<br>[vllm/v1/worker/gpu/cudagraph_utils.py](https://github.com/vllm-project/vllm/pull/53869/files#diff-fc699ff1c69fd17adb60116b3a7a20aa88b5187205f589952d0f3792475ec212) | Both accept pcp_manager during graph capture. | | #51031 | `0ecc284790e5403f74b899524ef82ecb69f83cb3`<br>[vllm/v1/worker/gpu/block_table.py](https://github.com/vllm-project/vllm/pull/51031/files#diff-21052649468f36e592c9ed378a9cd7b4615c558fb4d10f034f7b4268d3b6f9e1) | Both distinguish KV and kernel block sizes during DCP slot mapping. | | #54282 | `fe755c88995ad468882517b6c4bdd60138d46a3a`<br>[vllm/v1/worker/gpu/sample/gumbel.py](https://github.com/vllm-project/vllm/pull/54282/files#diff-02f7c5a06a2d23d544a07d57af16c7ffb74886469c8844385f51a719e329e6b8) | Both gumbel sampling APIs include is_drafting. | | #52188 | `d1e3eee6fb8ed3623241ef5c8e3ac533f775bff9`<br>[vllm/v1/worker/gpu/spec_decode/dflash/speculator.py](https://github.com/vllm-project/vllm/pull/52188/files#diff-0220b514682125dea26026db2d7caa1a0d9c772d84452e9e124434d1e11be5f0) | Both DFlash input kernels take cp_rank/CP_SIZE/CP_INTERLEAVE. | | #53694 | `5acc1c4e4b8730298cbff7a4a7c68c814dc24fd7`<br>[vllm/v1/worker/gpu/spec_decode/autoregressive/speculator.py](https://github.com/vllm-project/vllm/pull/53694/files#diff-3e7e2ba21d64e309503b7c5fe537363a7669fa1baf84d3c2e4891d9aed64fbe6) | Both propose APIs take DPSyncState; remove the obsolete token-count argument and retain replicated-PCP synchronization. | | #49811 | `01af92e175407231b1433b0aef01a1b9c983d955`<br>[vllm/v1/worker/gpu/spec_decode/extract_hidden_states.py](https://github.com/vllm-project/vllm/pull/49811/files#diff-b13061cd1be0bc80aaa9fce208fe9128aaaad25b3832c5f1693da1d4e8263f5e) | Both support extract_hidden_states on MRV2; remove old unsupported dispatch/skip. | | #53508 | `479eeb32d2b432dbb4e442fbd3f94ca2eca35d67`<br>[vllm/v1/worker/gpu_model_runner.py](https://github.com/vllm-project/vllm/pull/53508/files#diff-80ee7e2a62f9dcfbb8a312dc4e3948557e97ef187290daebbcae1e28596bda29) | Both remove post_kv_cache_wake_up; retire release-only call. | | #52494 | `3ff4f02dfe69abc1a0375d1ea8d8d5cb25609fcc`<br>[vllm/models/kimi_k3/amd/mla.py](https://github.com/vllm-project/vllm/pull/52494/files#diff-aad295e6de3607a46cea22f9f1cca768d4937dc5e1d3ca58105513358ed4e4c1) | Only main provides KimiK3MultiHeadLatentAttentionWrapper; retain release exclusion. | | #52861 | `b09bd69b5bf14911abef9a0e8e493b83c8a38fa6`<br>[vllm/v1/spec_decode/llm_base_proposer.py](https://github.com/vllm-project/vllm/pull/52861/files#diff-56fcad87dae9192cb0ab1643473765096cc44e1e744dccf3b99c4f2d85360d28) | Both include DeepseekV32MTPModel in the two-hidden-state architecture set. | | #54713 | `b28c3e1568bfae930f61d4b24940e47528c85d4a`<br>[vllm/v1/core/single_type_kv_cache_manager.py](https://github.com/vllm-project/vllm/pull/54713/files#diff-97c184a680b7a4bd7d58b11aa0073706533cc887d990eddb98469e7374025ab3) | Only main takes replay_boundaries in compressed-prefix hit lookup; preserve release calls without that keyword. | | #42785 | `442d36031ce710aa2a777c353becb660c57ab2bd`<br>[vllm/v1/worker/encoder_cudagraph.py](https://github.com/vllm-project/vllm/pull/42785/files#diff-287435cc78f753ae42acda542bca01c094557cd17db6bc2991d968fd012c1f18) | Only main capture callers pass axis_keys; preserve the existing Ascend rejection of nonempty axes. | | #52358 | `8f816a3f665489d7f0d222115d4f72ebab01076b`<br>[vllm/v1/worker/gpu/model_runner.py](https://github.com/vllm-project/vllm/pull/52358/files#diff-5823f988fc0264681a80db24ccaba4d364f14394815d83ec1d90944c09f571f0) | Both ExecuteModelState have dp_sync; only main has cudagraph_stats. | | #52789 | `9eb9d9d3953959695108600c8ed33d36bc6a1e5f`<br>[vllm/v1/core/sched/scheduler.py](https://github.com/vllm-project/vllm/pull/52789/files#diff-9eeca590fd99f15621897e559dba39b3ec4e7c2c65ec3c3229711689e008b5f4) | Release already has mamba_has_prefill_checkpoint_blocks; later main also has fine-grained prefix-cache state. | | #51251 | `7bbbf7c8e5040f7ebd374e8cbc657e01af1136dd`<br>[vllm/config/vllm.py](https://github.com/vllm-project/vllm/pull/51251/files#diff-bee6813076031d3ca1edc903c1b02b81e4676519afc562ce3fefe37f20c7b650) | v0.29.0 and main expose ec_manager_config; retire the old release-only ScoreEncoder configuration skip. | | #53240 | `b2db227a7c4c5e55f85524a094b607e5d27408b4`<br>[vllm/model_executor/layers/fused_moe/routed_experts_capturer.py](https://github.com/vllm-project/vllm/pull/53240/files#diff-0bbfc19dd02619bbcc948fa0281d4ab9a5b77b7d5062cde9911319c93089535c) | v0.29.0 supports MRV2 routed-expert capture and unwraps uniform cache groups; retire the old release-only replay skip. | | #53853 | `e376d45e82cb7e220da430e3179e81eb0922cf56`<br>[vllm/config/vllm.py](https://github.com/vllm-project/vllm/pull/53853/files#diff-bee6813076031d3ca1edc903c1b02b81e4676519afc562ce3fefe37f20c7b650) | Both delegate PCP compatibility validation to the PCP manager. | | #53183 | `4aab2b0ebed20343efe543c633f71b3c1336d5b8`<br>[vllm/config/vllm.py](https://github.com/vllm-project/vllm/pull/53183/files#diff-bee6813076031d3ca1edc903c1b02b81e4676519afc562ce3fefe37f20c7b650) | Both expose the V1 unsupported-feature helper used by the existing Ascend MRV1 feature filter. | #### Additional inherited contracts | vllm-ascend change | Why it is required | Upstream cause and direct link | Lane | Verification | |---|---|---|---|---| | `_310p/worker/v2/model_state.py`: remove tuple-return and RecoverSSM v0.28 fallbacks | v0.29.0 and fixed main both return `dict[MambaSpec, list[int]]` from `get_mamba_groups` and both initialize `recoverssm`; keeping the old fallback would preserve an unsupported third contract | [v0.29.0 mamba groups](https://github.com/vllm-project/vllm/blob/98dff2a81d747d1dba01a47f939f48c3526d4206/vllm/v1/worker/mamba_utils.py#L691-L704), [fixed-main mamba groups](https://github.com/vllm-project/vllm/blob/84030bbe3d74d99bad477a3d2e37a973ccd8865c/vllm/v1/worker/mamba_utils.py#L691-L704), [v0.29.0 RecoverSSM](https://github.com/vllm-project/vllm/blob/98dff2a81d747d1dba01a47f939f48c3526d4206/vllm/v1/worker/gpu/model_states/mamba_hybrid.py#L80-L103), [fixed-main RecoverSSM](https://github.com/vllm-project/vllm/blob/84030bbe3d74d99bad477a3d2e37a973ccd8865c/vllm/v1/worker/gpu/model_states/mamba_hybrid.py#L80-L103) | both, common implementation | Existing constructor UT now asserts the parent-created RecoverSSM value is retained; Ruff and compileall pass | | `worker/v2/model_runner.py`: always forward `kv_cache_allocation_context` | v0.29.0 and fixed main both accept this keyword; the new-base gate was specifically for v0.28.0 | [v0.29.0 signature](https://github.com/vllm-project/vllm/blob/98dff2a81d747d1dba01a47f939f48c3526d4206/vllm/v1/worker/gpu/model_runner.py#L533-L538), [fixed-main signature](https://github.com/vllm-project/vllm/blob/84030bbe3d74d99bad477a3d2e37a973ccd8865c/vllm/v1/worker/gpu/model_runner.py#L561-L566), [vllm-ascend #16791](https://github.com/vllm-project/vllm-ascend/pull/16791/files) | both, common implementation | Existing UT continues to assert the exact context object reaches the parent; Ruff and compileall pass | | `_310p/worker/v2/model_runner.py`: select the release KV-zeroing contract with `vllm_version_is("0.29.0")` | v0.29.0 exposes `KVCacheGroupSpec.is_eagle_group` but lacks `SpeculativeConfig.use_eagle_block_drop`; fixed main added the method | vLLM [#53388 diff](https://github.com/vllm-project/vllm/pull/53388/files), commit [`481839ad9e5ebf87aecb54fa5c9d986bd5ea4b81`](vllm-project/vllm@481839a); [v0.29.0 group field](https://github.com/vllm-project/vllm/blob/98dff2a81d747d1dba01a47f939f48c3526d4206/vllm/v1/kv_cache_interface.py#L1200), [fixed-main helper](https://github.com/vllm-project/vllm/blob/84030bbe3d74d99bad477a3d2e37a973ccd8865c/vllm/config/speculative.py#L1874-L1876) | release differs from main | Existing two-path KV-zeroing UT retained and renamed for v0.29.0; Ruff and compileall pass | | existing DFlash kernel UT: remove v0.28-only kwarg omission | the current Ascend kernel accepts the CP arguments and the only excluded lane was v0.28.0, which this PR replaces | [vllm-ascend #15098](https://github.com/vllm-project/vllm-ascend/pull/15098/files) | both, common invocation | Existing NPU test remains enabled with all assertions; Current-head CI pending | | Ascend change | Why / upstream cause | Lane | Verification | |---|---|---|---| | `patch/platform/patch_parallel_config.py`, registration, and global patch documentation | Allow Ascend PCP+DP by removing the generic GPU restriction, following vLLM [#54523](https://github.com/vllm-project/vllm/pull/54523/files), commit `7c2f1ff4958eaf0818405e9192c71608fe4a16b1`. Preserve dynamic `parallel_config.current_platform` lookup and rebuild ParallelConfig → SpeculativeConfig → VllmConfig. | v0.29.0 only | Existing configuration/EPLB UTs and historical PCP+DP NPU execution passed; current-head CI pending. | | `ops/triton/v2/sample/categorical_sample.py`, existing Gumbel UT | Both supported contracts require `is_drafting`, from [#54282](https://github.com/vllm-project/vllm/pull/54282/files), `fe755c88995ad468882517b6c4bdd60138d46a3a`; preserve the inherited optimized kernel and use a common wrapper. | Both | Existing positive drafting assertion retained; current-head CI pending. | | `patch/platform/patch_kv_cache_coordinator.py`, existing prefix-cache UT | Both support `cache_hit_alignment_tokens`, introduced by [#53598](https://github.com/vllm-project/vllm/pull/53598/files), `2ba984a5d06db414f3b2474fe9338faf6cd80a1c`; retire the v0.28-only write-mask branch. | Both | Existing assertions retained; current-head CI pending. | #### Rebase and retired-fallback evidence | Change | Exact source evidence | Decision | |---|---|---| | `tests/e2e/conftest.py::PROMPT_CONFIGS[HunyuanOCR]` | vLLM [#53272](https://github.com/vllm-project/vllm/pull/53272/files#diff-0852f1e9753819abe4f85380abdc0660fd4a998f44bd7bb73d68f462bd776d95), `d53b1c2efc0ca7161c3844f51ca9acbcdb7129d5` removes native Hunyuan V1/VL before v0.29.0; both exact supported sources lack it. The old conditional came from vllm-ascend [#14898](https://github.com/vllm-project/vllm-ascend/pull/14898/files#diff-824bc3e5c28226a3c0f4581750fa93a7da7001aff9e0ff4aee4131e3edd7b144), `e5118d151314ae18e56c0b63aa8dd00d294adc22`. | Remove the v0.28.0 selector and use the existing exclusion for both supported lanes. This does not migrate the protected SFA PCP skip or add a v0.29.0 accuracy skip. | | `vllm_ascend/worker/model_runner_v1.py` NaN helper imports and `tests/ut/worker/test_model_runner_v1_nan_detection.py` | vLLM [#50323](https://github.com/vllm-project/vllm/pull/50323/files#diff-80ee7e2a62f9dcfbb8a312dc4e3948557e97ef187290daebbcae1e28596bda29), `12292d94b25869be2af6b6d4f8eea6c2445e935f`; both exact sources export `nans_to_dict`, `gpu_sync_allowed`, and `raise_if_nan_logits`. The v0.28.0 fallback originated in vllm-ascend [#14898](https://github.com/vllm-project/vllm-ascend/pull/14898/files#diff-c49594855b615477bbc34f06d2d423a7dd84c021a7925cd1f61fdb79cb814c08), `e5118d151314ae18e56c0b63aa8dd00d294adc22`. | Delete import-failure/`None` fallbacks and the existing UT's obsolete availability skip; assertions remain unchanged and execute on both lanes. | | `vllm_ascend/worker/v2/aclgraph_utils.py` | vllm-ascend [#16409](https://github.com/vllm-project/vllm-ascend/pull/16409/files#diff-dc1c8a80dadf23627467bf04d6bc59fe13c6d075b7d2564d0acc07dc97d921f3), `799801feef347469d5e9b39374b9210e0d4f7431`. | Preserve the upstream revert while resolving the real rebase conflict; do not reintroduce the reverted graph-update behavior. | | `vllm_ascend/_310p/worker/v2/model_runner.py` | vllm-ascend [#16043](https://github.com/vllm-project/vllm-ascend/pull/16043/files#diff-2f27611a7cfbbdb780b9e88157f05d14d6dfe65406c8d66a5ba1f6949124f5d2), `d4d2957e5208c2f464d4625c05920bd29ea233cb`. | Preserve the newly merged 310P MTP copy tracking and remove only the v0.28.0 descriptor/allocation branch. | </details> </details> ### Does this PR introduce _any_ user-facing change? Yes. The supported vLLM release and default container builds move to v0.29.0, while the fixed main remains supported. The vllm-ascend package version does not change. Existing 0.29-only PCP+DP compatibility is restored. ### How was this patch tested? - Final E2E result for head `a3b76b073a201851454e873aa89fc2254992fc06`: [run 35504646654](https://github.com/vllm-project/vllm-ascend/actions/runs/35504646654) succeeded (39 successful jobs, 6 skipped). Raw logs from all 32 selected NPU jobs confirm the requested Ascend head, integration base `2de71b594319bde52c8bded69eba154f50be8e75`, and the actual vLLM pins/installations: fixed main `84030bbe3d74d99bad477a3d2e37a973ccd8865c` / `0.1.dev1+g84030bbe3.empty`, release `98dff2a81d747d1dba01a47f939f48c3526d4206` / `0.29.0+empty`. Each lane totals **563 passed, 35 skipped, 1 xfailed** across its selected pytest invocations. Skips/xfails are not passes. The resulting rebased integration commit is not printed and is not inferred. Actual release CPU and failed/cancelled image variants remain gaps. - Current-head CI update (2026-09-20): [main CPU UT](https://github.com/vllm-project/vllm-ascend/actions/runs/35504646654/job/106062951955) passed **5157 tests**, with **67 skipped**; actual installed vLLM was `0.1.dev1+g84030bbe3.empty`. Ascend checkout was the current PR head; the log does not print the full resulting integration head/base. Pre-commit and mypy passed. NPU/E2E results are recorded above; actual release CPU remains unverified. - Image build is partially blocked by infrastructure: A5 amd64 [Ubuntu](https://github.com/vllm-project/vllm-ascend/actions/runs/35504646536/job/106062362005) and [openEuler](https://github.com/vllm-project/vllm-ascend/actions/runs/35504646536/job/106062362068) failed before reading the Dockerfile because BuildKit could not create a snapshot temporary directory (`no space left on device`). No release compatibility code change is justified by this failure; cancelled variants remain unverified. The successful [310P openEuler arm64 build](https://github.com/vllm-project/vllm-ascend/actions/runs/35504646536/job/106062361987) explicitly checked out release `98dff2a81d747d1dba01a47f939f48c3526d4206` and installed `0.29.0+empty`; this is build evidence, not runtime UT coverage. - Current head: `a3b76b073a201851454e873aa89fc2254992fc06`, rebased onto the baseline above. Previous-head CPU [job 106060931458](https://github.com/vllm-project/vllm-ascend/actions/runs/35503843936/job/106060931458) stopped during Ascend integration rebase after #16803 merged, before any UT ran. Its vLLM checkout was the fixed main and installation reported `0.1.dev1+g84030bbe3.empty`. The two import conflicts are resolved; fresh-head CI was retriggered by the push. - Remote official v0.29.0 tag resolved to the exact SHA above; both source pins inspected. - All 95 changed Python files pass Ruff lint, Ruff format and AST parsing; `git diff --check` passes. Dockerfile tag/marker consistency checked across all eight variants. - Full pre-commit invocation: Ruff, codespell, typos, clang-format, markdownlint, actionlint, package-init, forbidden-import and boolean-context checks passed. Bash-dependent hooks cannot run on this Windows host; Python launcher hooks exit 9009. Full CI lint remains pending. - Actual main/release CPU/NPU and image builds are not claimed locally: the required Linux/NPU/container environment is unavailable. New PR E2E and image-build CI are requested. Existing CPU workflow runs fixed main only, so actual release CPU remains a validation gap. - #16393's historical green CI is not a substitute for this new head. No new test functions, workflow changes, golden/threshold changes or additional skips are introduced beyond restoring #16393. Inherited 0.28-only tests/skips are not counted as passes. - vLLM main: vllm-project/vllm@84030bb Signed-off-by: shenzhao <shenzhao9@huawei.com> Co-authored-by: shenzhao <shenzhao9@huawei.com>
xqchen7
pushed a commit
to xqchen7/vllm-ascend
that referenced
this pull request
Sep 22, 2026
### What this PR does / why we need it? The AllGather MoE path gathers hidden states and router logits in DP-then-PCP order, while hash routing previously gathered `input_ids` only across DP. With PCP enabled, this leaves fewer token IDs than router-logit rows and causes hash routing to fail. This PR: - gathers hash-routing input IDs with the same DP-then-PCP padding and communication layout used by the AllGather prepare path; - renames the helper to reflect that it covers every required parallel dimension; - extends the existing prepare/finalize and hash-router regression coverage for no parallelism, DP, PCP, and combined DP+PCP layouts. ### Does this PR introduce _any_ user-facing change? No API or configuration change. Hash-based MoE routing now works with the existing PCP configuration. ### How was this patch tested? - A5 PCP4 (`DP1 × TP1 × PCP4`): service started and ran normally. Full GPQA Diamond (`enable_thinking=false`) achieved 76.26% accuracy (151/198), with 198/198 requests successful and a 56.68% DSpark acceptance rate. - vLLM main: vllm-project/vllm@84030bb Signed-off-by: leolee <yihao.li@huawei.com>
xqchen7
pushed a commit
to xqchen7/vllm-ascend
that referenced
this pull request
Sep 22, 2026
…#16393) (vllm-project#16949) ### What this PR does / why we need it? Revert vllm-project#16393 at the author's request. This reverses squash commit `af277b2681443089e4d29684b1101707e2eb6426` on upstream main `8038f64129a1f8751ac6684f47c8d058208e4ff3`, restoring the previous fixed-main + v0.28.0 support contract. The revert applies without conflicts. Subsequent upstream patches are retained, including vllm-project#16775 (PCP/DCP slots), vllm-project#16923 (A3 SFA prefill all-to-all), and vllm-project#16834 (PCP hash-routing input IDs). No unrelated fixes are included. | Reverted change | Reason and source | Lane | | --- | --- | --- | | Release marker | Reverse [vllm-project#16393](https://github.com/vllm-project/vllm-ascend/pull/16393/files): v0.29.0 (`98dff2a81d747d1dba01a47f939f48c3526d4206`) to v0.28.0 (`2cf0a6915ce544dc493a0990f2ea38d81601128a`) | Release | | Version branches and interface adapters across scheduler, KV cache, attention, model runners and speculative decoding | Restore the pre-vllm-project#16393 v0.28.0 contracts by reversing the exact merged diff; keep main implementations behind their original version selection | Main + release | | v0.29.0-only parallel-config patch, registration and documentation | Remove the patch introduced by vllm-project#16393; retain the independent upstream v0.28.0 PCP+DP workaround | Release | | Existing unit/e2e tests and release-only skips | Restore the tests and conditions changed by vllm-project#16393, without introducing additional skip scope beyond the reverted baseline | Main + release | The vLLM main pin remains `84030bbe3d74d99bad477a3d2e37a973ccd8865c`. No main interface scan is needed for this release-only revert. The only changes beyond the inverse patch are formatting in `tests/ut/test_utils.py` and line-ending normalization in `categorical_sample.py` needed for formatting/whitespace checks. Workflows are unchanged. ### Does this PR introduce _any_ user-facing change? Yes. The supported release returns to vLLM v0.28.0; v0.29.0 support and its specific compatibility changes from vllm-project#16393 are reverted. The vllm-ascend package version and fixed vLLM main pin are unchanged. ### How was this patch tested? - `git revert --no-commit af277b2` applied cleanly on the latest fetched main. - Verified that the complete post-vllm-project#16393 upstream patch can still reverse-apply to the staged result (`git apply --reverse --check --cached`), confirming subsequent changes are preserved. - `git diff --cached --check`, Python AST parsing, Ruff lint and formatting checks passed for all 94 remaining changed Python files. - Ran the equivalent of `format.sh ci`: `pre-commit run --all-files --hook-stage manual`. Ruff, codespell, typos, clang-format, markdownlint, actionlint, package-init, forbidden-import and boolean-context checks passed. Bash-based hooks could not execute on this Windows host; two Python launcher hooks exited 9009. Full pre-commit validation remains for CI. - Actual CPU/NPU tests have not been run locally because the required vLLM/NPU runtime is unavailable. New PR CI is pending; the original upgrade PR's green run is not validation of this revert. - vLLM main: vllm-project/vllm@84030bb Signed-off-by: shenzhao <shenzhao9@huawei.com> Co-authored-by: shenzhao <shenzhao9@huawei.com>
xqchen7
pushed a commit
to xqchen7/vllm-ascend
that referenced
this pull request
Sep 22, 2026
…vllm-project#17004) ### What this PR does / why we need it? Reintroduce the vLLM v0.29.0 release upgrade from vllm-project#16393, reverted by vllm-project#16949, with container defaults aligned to the supported release. Based on upstream main `d3f6b4b59ab55dc42af1f30f425c893b95af5ee2`; preserves later merged changes. The image/source mismatch is confirmed in [nightly job 106004497020](https://github.com/vllm-project/vllm-ascend/actions/runs/35482881089/job/106004497020): the image installed `vllm 0.28.0+empty` (tag `v0.28.0`) while upgraded Ascend failed importing `_get_packed_kv_cache_groups`. All eight root Dockerfiles still defaulted to v0.28.0. The reusable image workflow passes no `VLLM_TAG` override, so those defaults govern release builds. Changing only the release marker does not update those images. - Fixed vLLM main remains `84030bbe3d74d99bad477a3d2e37a973ccd8865c`. - Release changes from v0.28.0 (`2cf0a6915ce544dc493a0990f2ea38d81601128a`) to official v0.29.0 (`98dff2a81d747d1dba01a47f939f48c3526d4206`, rechecked against the remote tag). - Restore vllm-project#16393's source compatibility and existing UT changes by reversing vllm-project#16949, then reconcile current upstream changes. No main interface scan is rerun for this release-only upgrade. - Update `VLLM_TAG` in all eight Dockerfiles: default/A3/A5/310P, Ubuntu/openEuler. Preserve existing exact-commit build overrides. No workflow changes. | Current-base adaptation | Exact cause and evidence | Lane | Validation | | --- | --- | --- | --- | | Eight Dockerfile `VLLM_TAG` defaults and release marker | vllm-project#16393 changed the release contract but omitted image defaults; the linked nightly log proves installation of 0.28.0 and missing `_get_packed_kv_cache_groups`. | Release images | All eight defaults match marker; image-build CI requested, pending | | `worker/v2/attn_utils.py::get_kv_cache_spec` and existing `_make_mla_layer` UT fixture | Ascend [vllm-project#16755](https://github.com/vllm-project/vllm-ascend/pull/16755/files), `df3e755e98fba8c6a18f200c645e0c2050469bc3`, added an old-release field selector. vLLM [#51718](https://github.com/vllm-project/vllm/pull/51718/files), `8bdc70ec7b379279ec0152343239c2d50aced687`, replaced `compress_ratio` with `tokens_per_state`; both exact supported pins use the latter. Use the common field, retaining metadata and cache-view assertions. | Both | Source inspection and static checks passed; actual CI pending | | `attention/attention_v1.py` import conflict | Preserve `attention_transfer_window` from Ascend [vllm-project#15636](https://github.com/vllm-project/vllm-ascend/pull/15636/files), `5c80630f28f8529aa82716e58b981a78819ec429`, and graph changes from vllm-project#16798; keep the common relocated PCP import from vllm-project#16393. Do not restore the old compute-start import or unused weak-reference import. | Both | Conflict resolved; static checks passed | | `worker/v2/aclgraph_utils.py` import conflict | Preserve `ContextSource`, `UpdatableGraph`, and `use_updatable_graph` from Ascend [vllm-project#16798](https://github.com/vllm-project/vllm-ascend/pull/16798/files), `34bb51f93724c565362f5108f5226303e1b56cad`, while removing the obsolete 0.28 version selector restored by the revert. | Both | Conflict resolved; static checks passed | | Existing 310P and Mamba model-runner UT imports | Preserve hardware-profile imports and mocks from Ascend [vllm-project#16803](https://github.com/vllm-project/vllm-ascend/pull/16803/files), `d3f6b4b59ab55dc42af1f30f425c893b95af5ee2`; remove only the unused old-release selector import. Hardware capability routing remains unchanged. | Both | Real rebase conflicts resolved; AST/Ruff/format passed; fresh CI pending | Newly merged changes were reviewed for version-contract impact: vllm-project#16775/vllm-project#16834 PCP metadata and routing, vllm-project#16923 A3 SFA, vllm-project#16426/vllm-project#16924/vllm-project#16955 custom ops, vllm-project#16081 MTP/SP, vllm-project#16669 xlite, vllm-project#15636/vllm-project#16747 transfer, vllm-project#16913 A5 pages, vllm-project#16798 graph updates, vllm-project#16755 GLM PD, vllm-project#16952 C8 config, vllm-project#16673 operator removal and vllm-project#16320 Kimi-K3 KV pool. Preserve these changes; no additional source-proven version branch was identified beyond the entries above. In particular, vllm-project#16747's `UniformTypeKVCacheSpecs.kv_cache_specs` and per-layer `block_size` exist in both exact pins; vllm-project#16320 adds Ascend connector hooks. CI remains necessary to validate runtime interactions. CI/documentation-only PRs are retained unchanged. <details> <summary>Inherited per-file compatibility evidence from vllm-project#16393</summary> The following source-contract ledger is inherited from vllm-project#16393. Any historical verification wording refers only to that earlier PR; it does not certify this new head. New-head validation is listed below. | Reference | Commit | |---|---| | vllm-ascend rebase base | `9dc6704559ebe2809b5cb7b7fc44184bd1338b3c` | | PR head | `8c5d785931c495701bf1da8b5bc80b321ebd3cd0` | | Fixed vLLM main | `84030bbe3d74d99bad477a3d2e37a973ccd8865c` | | Previous release: v0.28.0 | `2cf0a6915ce544dc493a0990f2ea38d81601128a` | | Target release: v0.29.0 | `98dff2a81d747d1dba01a47f939f48c3526d4206` | - Use common implementations where v0.29.0 and the fixed main share KV-cache layouts, Mamba copy/group APIs, PCP handling and speculative-decoding contracts. - Retain explicit `vllm_version_is("0.29.0")` branches for contracts that still differ, including RoPE, scheduler block snapshots, InputBatch, ReplaySSM, KV zeroing and DSpark PP handling. - Remove obsolete v0.28.0 compatibility and adapt existing test fixtures. Version detection uses package versions and the explicit `VLLM_VERSION` override, with local-version suffix handling and UT environment isolation retained; no hard-coded release-SHA inference remains. - Apply the Ascend PCP+DP validator patch only to v0.29.0, preserving other validations and EPLB platform binding. Rebuild dependent Pydantic schemas so nested configuration validation uses the patched validator. The global patch documentation records its rationale and removal criteria. - Preserve vllm-project#15747's Spec+PP protocol/partition handling after rebase; use the common exact-release selector and the real function-local DSpark sharing import. This release-only upgrade does not require a new main old-to-new interface scan. The latest rebase incorporates [vllm-project#16905](https://github.com/vllm-project/vllm-ascend/pull/16905/files), which reverted vllm-project#16544 at `9dc6704559ebe2809b5cb7b7fc44184bd1338b3c`. The now-unnecessary V4.1 drafter import gate and tuple annotation have been removed. Other release adaptations, including vllm-project#15747 Spec+PP handling, remain. Detailed contract evidence is retained below for review. <details> <summary>Per-file adaptations and exact upstream evidence</summary> #### Per-file adaptation ledger Evidence IDs refer to the exact source contract and upstream diff table below. Every row is syntax checked; branch-normalized AST comparison confirms unchanged main function bodies except the version identity helper and the explicitly retired propose argument/type annotations. Both supported versions completed CPU and NPU execution as recorded below. | Ascend file / symbols | Disposition and upstream evidence | Verification | |---|---|---| | `vllm_ascend/patch/worker/patch_v2/patch_dspark.py` / `_load_dspark_model_with_target_quant`; `tests/ut/patch/worker/test_patch_dspark_pp.py` | Preserve rebased [Ascend vllm-project#15747](https://github.com/vllm-project/vllm-ascend/pull/15747/files#diff-a170a42fe1e9275c999642b05c4a437d0c9104dd8ea19151796142909b00ab2a) (`82df9d871`), including manual PP partition masking. vLLM [#52809 diff](https://github.com/vllm-project/vllm/pull/52809/files#diff-2223c5e6b44b302db9bff4023777cc88b4427f4fb9d4a04476f6d6e4bd687393) (`91a893de64722019ea2faf852e06cabe143b3490`) moved `_should_share` inside the loader on both supported pins, so retain the earlier release fix: patch `eagle_utils`, never read/patch a nonexistent `dspark_utils._should_share`. v0.29 keeps its global PP guard; main has native PP via [#50514](https://github.com/vllm-project/vllm/pull/50514/files#diff-2223c5e6b44b302db9bff4023777cc88b4427f4fb9d4a04476f6d6e4bd687393) (`d87a440f88e28e5b37f9b1e22ce214d0426d5352`). | Existing UT now checks the real function-local import, absent module alias, and restoration on success/failure for both lanes; partition and PP assertions retained. Ruff/syntax pass; actual CPU/NPU pending. | | `vllm_ascend/worker/v2/pp_utils.py` / `use_legacy_spec_pp`; `tests/ut/worker/v2/test_pp_utils.py` | vllm-project#15747 added broad 0.28/0.29 routing and an obsolete 0.28 dev-build recognition path. For the two supported pins, #50514 exists only on fixed main. Route through `vllm_version_is("0.29.0")`; retain package/local-suffix and explicit environment-override semantics. No release-SHA inference or third release lane. | Existing UTs exercise the real uncached version helper with monkeypatch isolation, local suffix, fixed-main dev string, explicit override, and non-target versions. Isolated routing checks and Ruff/syntax pass; full CPU UT pending. | | `vllm_ascend/_310p/model_runner_310p.py`<br>`_prepare_inputs`, `_allocate_kv_cache_tensors` | #51718, #56078; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/_310p/spec_decode/llm_base_proposer_310.py`<br>`set_inputs_first_pass` | #56078; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/_310p/worker/v2/model_runner.py`<br>`initialize_kv_cache`, `_allocate_kv_cache_tensors`, `_prepare_inputs_310p` | #51718/#54436; use shared standardized layouts and retain the v0.29.0 InputBatch gate. The rebase preserves vllm-ascend vllm-project#16043's MTP copy tracking while removing only legacy v0.28.0 allocation paths. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/_310p/worker/v2/rope.py`<br>`get_310p_rope_state` | #56078; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/attention/attention_v1.py`<br>`module imports/dispatch` | #52839; both lanes use the common PCP import. Rebase keeps current upstream graph code and removes only the v0.28.0 import branch. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/attention/context_parallel/sfa_cp.py`<br>`module imports/dispatch` | #52839; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/attention/indexer.py`<br>`module imports/dispatch` | #52839; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/attention/mla_v1.py`<br>`module imports/dispatch` | #52839; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/core/dyntra_lb_scheduler.py`<br>`module imports/dispatch` | #51358, #54853; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/core/recompute_scheduler.py`<br>`module imports/dispatch`, `schedule` | #51358, #54853; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/core/scheduler_profiling_chunk.py`<br>`module imports/dispatch` | #51358, #54853; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/core/kv_cache_interface.py`<br>`get_kv_cache_compression_ratio`, `AscendMLAAttentionSpec`, `get_storage_block_size` | #51718, #53906; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/distributed/kv_transfer/kv_pool/ascend_store/layerwise_cache_layout.py`<br>`_merge_specs` | #51718; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/distributed/kv_transfer/kv_pool/kv_offload/preempt_offload/manager.py`<br>`_derive_cpu_config` | #51718; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/distributed/kv_transfer/kv_pool/kv_offload/native/npu.py`<br>`create_worker` | #52615; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/models/deepseek_mtp.py`<br>`load_weights` | #53106; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/models/deepseek_v4/indexer.py`<br>`get_kv_cache_spec` | #51718; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/models/glm5next/cache_config.py`<br>`make_tensor` | #51718; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/models/glm5next/kv_cache.py`<br>`get_kv_cache_spec` | #51718; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/models/kimi_k3_dspark.py`<br>`load_weights` | #53106; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/models/layer/attention/layer.py`<br>`get_kv_cache_spec` | #51718; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/patch/platform/patch_balance_schedule.py`<br>`module imports/dispatch` | #51358, #54853; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/patch/platform/patch_kv_cache_coordinator.py`<br>`__init__` | #53614; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/patch/platform/patch_kv_cache_utils.py`<br>`module imports/dispatch`, `_get_max_layers_per_page_size`, `_ascend_max_memory_usage_bytes_from_groups`, `_ascend_get_kv_cache_config_from_groups` | #51718; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/patch/platform/patch_use_v2_model_runner.py`<br>`module imports/dispatch`, `_patched_get_unsupported_features` | #53853, #53183; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/patch/worker/patch_bind_kv_cache.py`<br>`bind_kv_cache` | #51718; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/patch/worker/patch_bind_kv_cache.py::bind_kv_cache` | v0.29.0 has no ReplaySSM ring-tracker helper; fixed main requires it after KV binding. Evidence: [#52506, `adebc41b7e9f1085d3f73434e23beb76883b9eb4`, worker-utils diff](https://github.com/vllm-project/vllm/pull/52506/files#diff-645d58630d5acf3a0b07226bfef1e890a584c32502ab97c3d4642070f39a783c). Keep an explicit v0.29.0 exclusion and preserve the fixed-main call. | Failure reproduced on v0.29.0; Historical release/main NPU validation passed; current-head CI pending | | `vllm_ascend/patch/worker/patch_mamba_utils.py`<br>`module imports/dispatch`, `_get_state_copy_funcs_for_layer` | #53896; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/patch/worker/patch_v2/patch_attn_utils.py`<br>`module imports/dispatch` | #51718; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/patch/worker/patch_v2/patch_dspark.py`<br>`_load_dspark_model_with_target_quant` | #50514, #52809; v0.29.0 keeps `get_pp_group` module-bound but imports `_should_share` locally from Eagle utilities; fixed main removes the PP guard. Keep the release `get_pp_group` patch, share through the common Eagle utility, and delete the obsolete v0.28.0 `dspark_utils._should_share` patch. | Exact release failure reproduced; Historical release/main CPU/NPU validation passed; current-head CI pending | | `vllm_ascend/spec_decode/llm_base_proposer.py`<br>`model_returns_tuple`, `__init__`, `set_inputs_first_pass` | #56078, #52861; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/utils.py`<br>`get_kv_cache_tensor_layers`, `register_ascend_customop`, `vllm_version_is` | #52839, #51718, #52494; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/worker/model_runner_v1.py`<br>`_get_ascend_mamba_state_copy_funcs`, `_allocate_kv_cache_tensors`, `_prepare_inputs`, `_dummy_run`, `_reshape_kv_cache_tensors`, `get_kv_cache_spec`, NaN helpers | #51718/#53896/#56078 plus #50323; remove v0.28.0 layouts/copy APIs and import the NaN helpers directly because v0.29.0 and fixed main expose the same contract. XD-RoPE remains explicitly gated to v0.29.0. | Static checked; historical dual-version CPU/NPU validation passed; current-head CI pending | | `vllm_ascend/worker/v2/aclgraph_utils.py`<br>`capture` | #53869 makes `pcp_manager` common to both lanes. The rebase preserves vllm-ascend vllm-project#16409's host-parameter-update revert and removes only the obsolete v0.28.0 capture branch. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/worker/v2/attn_utils.py`<br>`get_kv_cache_spec`, `_allocate_kv_cache`, `_reshape_kv_cache_v2` | #51718; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/worker/v2/block_table.py`<br>`__init__`, `init_block_table_layout_tensors`, `compute_slot_mappings` | #51718, #51031; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/worker/v2/model_runner.py`<br>`module imports/dispatch`, `prepare_inputs`, `__init__`, `sample_tokens`, `prepare_dummy_attn`, `execute_model` | #50514, #54436, #52506, #55212, #53515; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/worker/v2/pcp_manager.py`<br>`partition_batch` | #53515; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/worker/v2/sample/gumbel.py`<br>`module imports/dispatch` | #54282; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/worker/v2/spec_decode/__init__.py`<br>`init_speculator` | #49811; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/worker/v2/spec_decode/autoregressive/speculator.py`<br>`propose` | #53694; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/worker/v2/spec_decode/dflash/speculator.py`<br>`module imports/dispatch`, `propose` | #53694, #52188; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/worker/v2/spec_decode/dspark/speculator.py`<br>`propose` | #53694; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/worker/worker.py`<br>`_scale_kv_cache_memory_for_multi_group`, `wake_up` | #51718, #53508; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | #### Exact upstream evidence | Upstream change | Full commit SHA / direct diff | Actual supported contracts and branch decision | |---|---|---| | #51718 | `8bdc70ec7b379279ec0152343239c2d50aced687`<br>[vllm/v1/kv_cache_interface.py](https://github.com/vllm-project/vllm/pull/51718/files#diff-f76cdfbf02dacd9dffbfbd0d9ad68a7a6ac0d8aed70834f95a2ae8ccd2e333cb) | Both use layers/layer_stride/block_stride/offset, tokens_per_state, CircularBufferSpec and standardized backing; retire shared_by/compress_ratio allocation branches. | | #52839 | `58e5ee0158b6a264c3506f00480e108a34b33ee3`<br>[vllm/v1/attention/ops/pcp.py](https://github.com/vllm-project/vllm/pull/52839/files#diff-23d7e7ab01a7f44414796e1e1b08cfa5277a05f54143ee2dd3a1041ca47772a8) | Both import PCP operations from vllm.v1.attention.ops.pcp. | | #53896 | `e126687a9a828d513c01a07cd69f025f27d63280`<br>[vllm/v1/worker/mamba_utils.py](https://github.com/vllm-project/vllm/pull/53896/files#diff-8ab86225fa08e9a6700851191abcb3f4f8b0bba40e97a82aad700d7a23a4d7ec) | Both use Mamba copy-function dictionaries and unwrap UniformTypeKVCacheSpecs. | | #53106 | `1fe3a1571ac67581478a11743e55a306de1d136f`<br>[vllm/model_executor/models/utils.py](https://github.com/vllm-project/vllm/pull/53106/files#diff-b0ba1095e9881e5c87e33dfd20958d1e1ceafe8a4433aa692f468e61be130b21) | Both use WeightsMapper instead of AutoWeightsLoader skip_prefixes/skip_substrs. | | #53906 | `98ed0856f31fa3aaf5e27464e2b4ef5a8ee6b2f5`<br>[vllm/v1/kv_cache_interface.py](https://github.com/vllm-project/vllm/pull/53906/files#diff-f76cdfbf02dacd9dffbfbd0d9ad68a7a6ac0d8aed70834f95a2ae8ccd2e333cb) | Only pinned main has the optional MLA storage_block_size dataclass field; release keeps the Ascend derived property, using tokens_per_state. | | #56078 | `719284fe158f1be8a9dd92953295fc9d49015730`<br>[vllm/config/model.py](https://github.com/vllm-project/vllm/pull/56078/files#diff-998c640befaf137b9af825f29f4e6e47d273caab1fd04093c97df24b18f5c417) | Release retains uses_xdrope_dim and three M-RoPE dimensions; pinned main uses mrope_num_dims and unified RoPE. | | #52615 | `138d137b5b955a5ebee98dfc946ecdd65d7b87ce`<br>[vllm/v1/kv_offload/cpu/spec.py](https://github.com/vllm-project/vllm/pull/52615/files#diff-7d073c26c9b22f74e9ff9e4c8da733d1807233edb83439fc64fbd48c43ae6045) | Release uses num_blocks/kv_bytes_per_block; main uses num_chunks/kv_bytes_per_chunk. | | #51358 | `6b110badbb22d3f66c7218b71138f13b7a6b3419`<br>[vllm/v1/core/sched/output.py](https://github.com/vllm-project/vllm/pull/51358/files#diff-cafd89ce8a698a56acb24ada62831cbc7a980782f78a52d1742ba238031f296c) | Release now has boundary_state_offloads and KVConnectorBlockState; remove partial_tail_offloads plumbing. | | #54853 | `0b066293f3c738a0cbd3a087bf893f2f4dcd61f2`<br>[vllm/v1/core/sched/output.py](https://github.com/vllm-project/vllm/pull/54853/files#diff-cafd89ce8a698a56acb24ada62831cbc7a980782f78a52d1742ba238031f296c) | Release constructor takes block_ids snapshots; main takes req_ids and resolve_block_ids. Keep exact release snapshot membership and main lazy-resolution membership. | | #53614 | `144e79c8106da23141ac010394b782f730cc7fe8`<br>[vllm/v1/core/kv_cache_coordinator.py](https://github.com/vllm-project/vllm/pull/53614/files#diff-43875c71daa893ef7567e21633d9988c2baf95bef61e3a334a6d584d6444d725) | Only main configures drop_eagle_checkpoint_block for replay-aligned Mamba checkpoints. | | #50514 | `d87a440f88e28e5b37f9b1e22ce214d0426d5352`<br>[vllm/v1/worker/gpu/spec_decode/dspark/utils.py](https://github.com/vllm-project/vllm/pull/50514/files#diff-2223c5e6b44b302db9bff4023777cc88b4427f4fb9d4a04476f6d6e4bd687393) | Release retains module-level PP/share symbols and Ascend PP workaround; main has the subsequent PP integration. | | #52809 | `91a893de64722019ea2faf852e06cabe143b3490`<br>[vllm/v1/worker/gpu/spec_decode/dspark/utils.py](https://github.com/vllm-project/vllm/pull/52809/files#diff-2223c5e6b44b302db9bff4023777cc88b4427f4fb9d4a04476f6d6e4bd687393) | Between v0.28.0 and v0.29.0, `_should_share` moved from a DSpark module binding to a function-local import from Eagle utilities. The shared Eagle patch remains effective; the old DSpark-module read/write must be removed. | | #54436 | `6bafc049aae6c26e210162630914ee9177a4b586`<br>[vllm/v1/worker/gpu/input_batch.py](https://github.com/vllm-project/vllm/pull/54436/files#diff-106f39c08266f186830bb8fcd7fb1df35c0aaa5fd0ac5c17aac64aeddee48902) | Release InputBatch requires max_seq_len_np; main removed it. | | #52506 | `adebc41b7e9f1085d3f73434e23beb76883b9eb4`<br>[vllm/v1/worker/gpu/model_runner.py](https://github.com/vllm-project/vllm/pull/52506/files#diff-5823f988fc0264681a80db24ccaba4d364f14394815d83ec1d90944c09f571f0) | Only main accepts valid_dummy_state_slots/valid_state_slots capture arguments. | | #55212 | `83990f5bcc0b5eb08b2f1fd1b109fa8a31b122f3`<br>[vllm/v1/worker/gpu/model_runner.py](https://github.com/vllm-project/vllm/pull/55212/files#diff-5823f988fc0264681a80db24ccaba4d364f14394815d83ec1d90944c09f571f0) | Release prepares DCP local sequence lengths before partitioning; main initializes DCP metadata afterwards. | | #53515 | `b1fbbc2ade51e3826bc92e4733c9c692ee21d42d`<br>[vllm/v1/worker/gpu/pcp_manager.py](https://github.com/vllm-project/vllm/pull/53515/files#diff-45dad5fe975c2ec8d02124fa65579cf11bfe9169a94217ceae1de0ea29518127) | Both accept padded_num_tokens for persistent PCP input buffers. | | #53869 | `b3af042abd8fe5a297ec3ec72db276fd661a67b3`<br>[vllm/v1/worker/gpu/cudagraph_utils.py](https://github.com/vllm-project/vllm/pull/53869/files#diff-fc699ff1c69fd17adb60116b3a7a20aa88b5187205f589952d0f3792475ec212) | Both accept pcp_manager during graph capture. | | #51031 | `0ecc284790e5403f74b899524ef82ecb69f83cb3`<br>[vllm/v1/worker/gpu/block_table.py](https://github.com/vllm-project/vllm/pull/51031/files#diff-21052649468f36e592c9ed378a9cd7b4615c558fb4d10f034f7b4268d3b6f9e1) | Both distinguish KV and kernel block sizes during DCP slot mapping. | | #54282 | `fe755c88995ad468882517b6c4bdd60138d46a3a`<br>[vllm/v1/worker/gpu/sample/gumbel.py](https://github.com/vllm-project/vllm/pull/54282/files#diff-02f7c5a06a2d23d544a07d57af16c7ffb74886469c8844385f51a719e329e6b8) | Both gumbel sampling APIs include is_drafting. | | #52188 | `d1e3eee6fb8ed3623241ef5c8e3ac533f775bff9`<br>[vllm/v1/worker/gpu/spec_decode/dflash/speculator.py](https://github.com/vllm-project/vllm/pull/52188/files#diff-0220b514682125dea26026db2d7caa1a0d9c772d84452e9e124434d1e11be5f0) | Both DFlash input kernels take cp_rank/CP_SIZE/CP_INTERLEAVE. | | #53694 | `5acc1c4e4b8730298cbff7a4a7c68c814dc24fd7`<br>[vllm/v1/worker/gpu/spec_decode/autoregressive/speculator.py](https://github.com/vllm-project/vllm/pull/53694/files#diff-3e7e2ba21d64e309503b7c5fe537363a7669fa1baf84d3c2e4891d9aed64fbe6) | Both propose APIs take DPSyncState; remove the obsolete token-count argument and retain replicated-PCP synchronization. | | #49811 | `01af92e175407231b1433b0aef01a1b9c983d955`<br>[vllm/v1/worker/gpu/spec_decode/extract_hidden_states.py](https://github.com/vllm-project/vllm/pull/49811/files#diff-b13061cd1be0bc80aaa9fce208fe9128aaaad25b3832c5f1693da1d4e8263f5e) | Both support extract_hidden_states on MRV2; remove old unsupported dispatch/skip. | | #53508 | `479eeb32d2b432dbb4e442fbd3f94ca2eca35d67`<br>[vllm/v1/worker/gpu_model_runner.py](https://github.com/vllm-project/vllm/pull/53508/files#diff-80ee7e2a62f9dcfbb8a312dc4e3948557e97ef187290daebbcae1e28596bda29) | Both remove post_kv_cache_wake_up; retire release-only call. | | #52494 | `3ff4f02dfe69abc1a0375d1ea8d8d5cb25609fcc`<br>[vllm/models/kimi_k3/amd/mla.py](https://github.com/vllm-project/vllm/pull/52494/files#diff-aad295e6de3607a46cea22f9f1cca768d4937dc5e1d3ca58105513358ed4e4c1) | Only main provides KimiK3MultiHeadLatentAttentionWrapper; retain release exclusion. | | #52861 | `b09bd69b5bf14911abef9a0e8e493b83c8a38fa6`<br>[vllm/v1/spec_decode/llm_base_proposer.py](https://github.com/vllm-project/vllm/pull/52861/files#diff-56fcad87dae9192cb0ab1643473765096cc44e1e744dccf3b99c4f2d85360d28) | Both include DeepseekV32MTPModel in the two-hidden-state architecture set. | | #54713 | `b28c3e1568bfae930f61d4b24940e47528c85d4a`<br>[vllm/v1/core/single_type_kv_cache_manager.py](https://github.com/vllm-project/vllm/pull/54713/files#diff-97c184a680b7a4bd7d58b11aa0073706533cc887d990eddb98469e7374025ab3) | Only main takes replay_boundaries in compressed-prefix hit lookup; preserve release calls without that keyword. | | #42785 | `442d36031ce710aa2a777c353becb660c57ab2bd`<br>[vllm/v1/worker/encoder_cudagraph.py](https://github.com/vllm-project/vllm/pull/42785/files#diff-287435cc78f753ae42acda542bca01c094557cd17db6bc2991d968fd012c1f18) | Only main capture callers pass axis_keys; preserve the existing Ascend rejection of nonempty axes. | | #52358 | `8f816a3f665489d7f0d222115d4f72ebab01076b`<br>[vllm/v1/worker/gpu/model_runner.py](https://github.com/vllm-project/vllm/pull/52358/files#diff-5823f988fc0264681a80db24ccaba4d364f14394815d83ec1d90944c09f571f0) | Both ExecuteModelState have dp_sync; only main has cudagraph_stats. | | #52789 | `9eb9d9d3953959695108600c8ed33d36bc6a1e5f`<br>[vllm/v1/core/sched/scheduler.py](https://github.com/vllm-project/vllm/pull/52789/files#diff-9eeca590fd99f15621897e559dba39b3ec4e7c2c65ec3c3229711689e008b5f4) | Release already has mamba_has_prefill_checkpoint_blocks; later main also has fine-grained prefix-cache state. | | #51251 | `7bbbf7c8e5040f7ebd374e8cbc657e01af1136dd`<br>[vllm/config/vllm.py](https://github.com/vllm-project/vllm/pull/51251/files#diff-bee6813076031d3ca1edc903c1b02b81e4676519afc562ce3fefe37f20c7b650) | v0.29.0 and main expose ec_manager_config; retire the old release-only ScoreEncoder configuration skip. | | #53240 | `b2db227a7c4c5e55f85524a094b607e5d27408b4`<br>[vllm/model_executor/layers/fused_moe/routed_experts_capturer.py](https://github.com/vllm-project/vllm/pull/53240/files#diff-0bbfc19dd02619bbcc948fa0281d4ab9a5b77b7d5062cde9911319c93089535c) | v0.29.0 supports MRV2 routed-expert capture and unwraps uniform cache groups; retire the old release-only replay skip. | | #53853 | `e376d45e82cb7e220da430e3179e81eb0922cf56`<br>[vllm/config/vllm.py](https://github.com/vllm-project/vllm/pull/53853/files#diff-bee6813076031d3ca1edc903c1b02b81e4676519afc562ce3fefe37f20c7b650) | Both delegate PCP compatibility validation to the PCP manager. | | #53183 | `4aab2b0ebed20343efe543c633f71b3c1336d5b8`<br>[vllm/config/vllm.py](https://github.com/vllm-project/vllm/pull/53183/files#diff-bee6813076031d3ca1edc903c1b02b81e4676519afc562ce3fefe37f20c7b650) | Both expose the V1 unsupported-feature helper used by the existing Ascend MRV1 feature filter. | #### Additional inherited contracts | vllm-ascend change | Why it is required | Upstream cause and direct link | Lane | Verification | |---|---|---|---|---| | `_310p/worker/v2/model_state.py`: remove tuple-return and RecoverSSM v0.28 fallbacks | v0.29.0 and fixed main both return `dict[MambaSpec, list[int]]` from `get_mamba_groups` and both initialize `recoverssm`; keeping the old fallback would preserve an unsupported third contract | [v0.29.0 mamba groups](https://github.com/vllm-project/vllm/blob/98dff2a81d747d1dba01a47f939f48c3526d4206/vllm/v1/worker/mamba_utils.py#L691-L704), [fixed-main mamba groups](https://github.com/vllm-project/vllm/blob/84030bbe3d74d99bad477a3d2e37a973ccd8865c/vllm/v1/worker/mamba_utils.py#L691-L704), [v0.29.0 RecoverSSM](https://github.com/vllm-project/vllm/blob/98dff2a81d747d1dba01a47f939f48c3526d4206/vllm/v1/worker/gpu/model_states/mamba_hybrid.py#L80-L103), [fixed-main RecoverSSM](https://github.com/vllm-project/vllm/blob/84030bbe3d74d99bad477a3d2e37a973ccd8865c/vllm/v1/worker/gpu/model_states/mamba_hybrid.py#L80-L103) | both, common implementation | Existing constructor UT now asserts the parent-created RecoverSSM value is retained; Ruff and compileall pass | | `worker/v2/model_runner.py`: always forward `kv_cache_allocation_context` | v0.29.0 and fixed main both accept this keyword; the new-base gate was specifically for v0.28.0 | [v0.29.0 signature](https://github.com/vllm-project/vllm/blob/98dff2a81d747d1dba01a47f939f48c3526d4206/vllm/v1/worker/gpu/model_runner.py#L533-L538), [fixed-main signature](https://github.com/vllm-project/vllm/blob/84030bbe3d74d99bad477a3d2e37a973ccd8865c/vllm/v1/worker/gpu/model_runner.py#L561-L566), [vllm-ascend vllm-project#16791](https://github.com/vllm-project/vllm-ascend/pull/16791/files) | both, common implementation | Existing UT continues to assert the exact context object reaches the parent; Ruff and compileall pass | | `_310p/worker/v2/model_runner.py`: select the release KV-zeroing contract with `vllm_version_is("0.29.0")` | v0.29.0 exposes `KVCacheGroupSpec.is_eagle_group` but lacks `SpeculativeConfig.use_eagle_block_drop`; fixed main added the method | vLLM [#53388 diff](https://github.com/vllm-project/vllm/pull/53388/files), commit [`481839ad9e5ebf87aecb54fa5c9d986bd5ea4b81`](vllm-project/vllm@481839a); [v0.29.0 group field](https://github.com/vllm-project/vllm/blob/98dff2a81d747d1dba01a47f939f48c3526d4206/vllm/v1/kv_cache_interface.py#L1200), [fixed-main helper](https://github.com/vllm-project/vllm/blob/84030bbe3d74d99bad477a3d2e37a973ccd8865c/vllm/config/speculative.py#L1874-L1876) | release differs from main | Existing two-path KV-zeroing UT retained and renamed for v0.29.0; Ruff and compileall pass | | existing DFlash kernel UT: remove v0.28-only kwarg omission | the current Ascend kernel accepts the CP arguments and the only excluded lane was v0.28.0, which this PR replaces | [vllm-ascend vllm-project#15098](https://github.com/vllm-project/vllm-ascend/pull/15098/files) | both, common invocation | Existing NPU test remains enabled with all assertions; Current-head CI pending | | Ascend change | Why / upstream cause | Lane | Verification | |---|---|---|---| | `patch/platform/patch_parallel_config.py`, registration, and global patch documentation | Allow Ascend PCP+DP by removing the generic GPU restriction, following vLLM [#54523](https://github.com/vllm-project/vllm/pull/54523/files), commit `7c2f1ff4958eaf0818405e9192c71608fe4a16b1`. Preserve dynamic `parallel_config.current_platform` lookup and rebuild ParallelConfig → SpeculativeConfig → VllmConfig. | v0.29.0 only | Existing configuration/EPLB UTs and historical PCP+DP NPU execution passed; current-head CI pending. | | `ops/triton/v2/sample/categorical_sample.py`, existing Gumbel UT | Both supported contracts require `is_drafting`, from [#54282](https://github.com/vllm-project/vllm/pull/54282/files), `fe755c88995ad468882517b6c4bdd60138d46a3a`; preserve the inherited optimized kernel and use a common wrapper. | Both | Existing positive drafting assertion retained; current-head CI pending. | | `patch/platform/patch_kv_cache_coordinator.py`, existing prefix-cache UT | Both support `cache_hit_alignment_tokens`, introduced by [#53598](https://github.com/vllm-project/vllm/pull/53598/files), `2ba984a5d06db414f3b2474fe9338faf6cd80a1c`; retire the v0.28-only write-mask branch. | Both | Existing assertions retained; current-head CI pending. | #### Rebase and retired-fallback evidence | Change | Exact source evidence | Decision | |---|---|---| | `tests/e2e/conftest.py::PROMPT_CONFIGS[HunyuanOCR]` | vLLM [#53272](https://github.com/vllm-project/vllm/pull/53272/files#diff-0852f1e9753819abe4f85380abdc0660fd4a998f44bd7bb73d68f462bd776d95), `d53b1c2efc0ca7161c3844f51ca9acbcdb7129d5` removes native Hunyuan V1/VL before v0.29.0; both exact supported sources lack it. The old conditional came from vllm-ascend [vllm-project#14898](https://github.com/vllm-project/vllm-ascend/pull/14898/files#diff-824bc3e5c28226a3c0f4581750fa93a7da7001aff9e0ff4aee4131e3edd7b144), `e5118d151314ae18e56c0b63aa8dd00d294adc22`. | Remove the v0.28.0 selector and use the existing exclusion for both supported lanes. This does not migrate the protected SFA PCP skip or add a v0.29.0 accuracy skip. | | `vllm_ascend/worker/model_runner_v1.py` NaN helper imports and `tests/ut/worker/test_model_runner_v1_nan_detection.py` | vLLM [#50323](https://github.com/vllm-project/vllm/pull/50323/files#diff-80ee7e2a62f9dcfbb8a312dc4e3948557e97ef187290daebbcae1e28596bda29), `12292d94b25869be2af6b6d4f8eea6c2445e935f`; both exact sources export `nans_to_dict`, `gpu_sync_allowed`, and `raise_if_nan_logits`. The v0.28.0 fallback originated in vllm-ascend [vllm-project#14898](https://github.com/vllm-project/vllm-ascend/pull/14898/files#diff-c49594855b615477bbc34f06d2d423a7dd84c021a7925cd1f61fdb79cb814c08), `e5118d151314ae18e56c0b63aa8dd00d294adc22`. | Delete import-failure/`None` fallbacks and the existing UT's obsolete availability skip; assertions remain unchanged and execute on both lanes. | | `vllm_ascend/worker/v2/aclgraph_utils.py` | vllm-ascend [vllm-project#16409](https://github.com/vllm-project/vllm-ascend/pull/16409/files#diff-dc1c8a80dadf23627467bf04d6bc59fe13c6d075b7d2564d0acc07dc97d921f3), `799801feef347469d5e9b39374b9210e0d4f7431`. | Preserve the upstream revert while resolving the real rebase conflict; do not reintroduce the reverted graph-update behavior. | | `vllm_ascend/_310p/worker/v2/model_runner.py` | vllm-ascend [vllm-project#16043](https://github.com/vllm-project/vllm-ascend/pull/16043/files#diff-2f27611a7cfbbdb780b9e88157f05d14d6dfe65406c8d66a5ba1f6949124f5d2), `d4d2957e5208c2f464d4625c05920bd29ea233cb`. | Preserve the newly merged 310P MTP copy tracking and remove only the v0.28.0 descriptor/allocation branch. | </details> </details> ### Does this PR introduce _any_ user-facing change? Yes. The supported vLLM release and default container builds move to v0.29.0, while the fixed main remains supported. The vllm-ascend package version does not change. Existing 0.29-only PCP+DP compatibility is restored. ### How was this patch tested? - Final E2E result for head `a3b76b073a201851454e873aa89fc2254992fc06`: [run 35504646654](https://github.com/vllm-project/vllm-ascend/actions/runs/35504646654) succeeded (39 successful jobs, 6 skipped). Raw logs from all 32 selected NPU jobs confirm the requested Ascend head, integration base `2de71b594319bde52c8bded69eba154f50be8e75`, and the actual vLLM pins/installations: fixed main `84030bbe3d74d99bad477a3d2e37a973ccd8865c` / `0.1.dev1+g84030bbe3.empty`, release `98dff2a81d747d1dba01a47f939f48c3526d4206` / `0.29.0+empty`. Each lane totals **563 passed, 35 skipped, 1 xfailed** across its selected pytest invocations. Skips/xfails are not passes. The resulting rebased integration commit is not printed and is not inferred. Actual release CPU and failed/cancelled image variants remain gaps. - Current-head CI update (2026-09-20): [main CPU UT](https://github.com/vllm-project/vllm-ascend/actions/runs/35504646654/job/106062951955) passed **5157 tests**, with **67 skipped**; actual installed vLLM was `0.1.dev1+g84030bbe3.empty`. Ascend checkout was the current PR head; the log does not print the full resulting integration head/base. Pre-commit and mypy passed. NPU/E2E results are recorded above; actual release CPU remains unverified. - Image build is partially blocked by infrastructure: A5 amd64 [Ubuntu](https://github.com/vllm-project/vllm-ascend/actions/runs/35504646536/job/106062362005) and [openEuler](https://github.com/vllm-project/vllm-ascend/actions/runs/35504646536/job/106062362068) failed before reading the Dockerfile because BuildKit could not create a snapshot temporary directory (`no space left on device`). No release compatibility code change is justified by this failure; cancelled variants remain unverified. The successful [310P openEuler arm64 build](https://github.com/vllm-project/vllm-ascend/actions/runs/35504646536/job/106062361987) explicitly checked out release `98dff2a81d747d1dba01a47f939f48c3526d4206` and installed `0.29.0+empty`; this is build evidence, not runtime UT coverage. - Current head: `a3b76b073a201851454e873aa89fc2254992fc06`, rebased onto the baseline above. Previous-head CPU [job 106060931458](https://github.com/vllm-project/vllm-ascend/actions/runs/35503843936/job/106060931458) stopped during Ascend integration rebase after vllm-project#16803 merged, before any UT ran. Its vLLM checkout was the fixed main and installation reported `0.1.dev1+g84030bbe3.empty`. The two import conflicts are resolved; fresh-head CI was retriggered by the push. - Remote official v0.29.0 tag resolved to the exact SHA above; both source pins inspected. - All 95 changed Python files pass Ruff lint, Ruff format and AST parsing; `git diff --check` passes. Dockerfile tag/marker consistency checked across all eight variants. - Full pre-commit invocation: Ruff, codespell, typos, clang-format, markdownlint, actionlint, package-init, forbidden-import and boolean-context checks passed. Bash-dependent hooks cannot run on this Windows host; Python launcher hooks exit 9009. Full CI lint remains pending. - Actual main/release CPU/NPU and image builds are not claimed locally: the required Linux/NPU/container environment is unavailable. New PR E2E and image-build CI are requested. Existing CPU workflow runs fixed main only, so actual release CPU remains a validation gap. - vllm-project#16393's historical green CI is not a substitute for this new head. No new test functions, workflow changes, golden/threshold changes or additional skips are introduced beyond restoring vllm-project#16393. Inherited 0.28-only tests/skips are not counted as passes. - vLLM main: vllm-project/vllm@84030bb Signed-off-by: shenzhao <shenzhao9@huawei.com> Co-authored-by: shenzhao <shenzhao9@huawei.com>
zhaochuang001
pushed a commit
to zhaochuang001/vllm-ascend
that referenced
this pull request
Sep 22, 2026
### What this PR does / why we need it? The AllGather MoE path gathers hidden states and router logits in DP-then-PCP order, while hash routing previously gathered `input_ids` only across DP. With PCP enabled, this leaves fewer token IDs than router-logit rows and causes hash routing to fail. This PR: - gathers hash-routing input IDs with the same DP-then-PCP padding and communication layout used by the AllGather prepare path; - renames the helper to reflect that it covers every required parallel dimension; - extends the existing prepare/finalize and hash-router regression coverage for no parallelism, DP, PCP, and combined DP+PCP layouts. ### Does this PR introduce _any_ user-facing change? No API or configuration change. Hash-based MoE routing now works with the existing PCP configuration. ### How was this patch tested? - A5 PCP4 (`DP1 × TP1 × PCP4`): service started and ran normally. Full GPQA Diamond (`enable_thinking=false`) achieved 76.26% accuracy (151/198), with 198/198 requests successful and a 56.68% DSpark acceptance rate. - vLLM main: vllm-project/vllm@84030bb Signed-off-by: leolee <yihao.li@huawei.com>
zhaochuang001
pushed a commit
to zhaochuang001/vllm-ascend
that referenced
this pull request
Sep 22, 2026
…#16393) (vllm-project#16949) Revert vllm-project#16393 at the author's request. This reverses squash commit `af277b2681443089e4d29684b1101707e2eb6426` on upstream main `8038f64129a1f8751ac6684f47c8d058208e4ff3`, restoring the previous fixed-main + v0.28.0 support contract. The revert applies without conflicts. Subsequent upstream patches are retained, including vllm-project#16775 (PCP/DCP slots), vllm-project#16923 (A3 SFA prefill all-to-all), and vllm-project#16834 (PCP hash-routing input IDs). No unrelated fixes are included. | Reverted change | Reason and source | Lane | | --- | --- | --- | | Release marker | Reverse [vllm-project#16393](https://github.com/vllm-project/vllm-ascend/pull/16393/files): v0.29.0 (`98dff2a81d747d1dba01a47f939f48c3526d4206`) to v0.28.0 (`2cf0a6915ce544dc493a0990f2ea38d81601128a`) | Release | | Version branches and interface adapters across scheduler, KV cache, attention, model runners and speculative decoding | Restore the pre-vllm-project#16393 v0.28.0 contracts by reversing the exact merged diff; keep main implementations behind their original version selection | Main + release | | v0.29.0-only parallel-config patch, registration and documentation | Remove the patch introduced by vllm-project#16393; retain the independent upstream v0.28.0 PCP+DP workaround | Release | | Existing unit/e2e tests and release-only skips | Restore the tests and conditions changed by vllm-project#16393, without introducing additional skip scope beyond the reverted baseline | Main + release | The vLLM main pin remains `84030bbe3d74d99bad477a3d2e37a973ccd8865c`. No main interface scan is needed for this release-only revert. The only changes beyond the inverse patch are formatting in `tests/ut/test_utils.py` and line-ending normalization in `categorical_sample.py` needed for formatting/whitespace checks. Workflows are unchanged. Yes. The supported release returns to vLLM v0.28.0; v0.29.0 support and its specific compatibility changes from vllm-project#16393 are reverted. The vllm-ascend package version and fixed vLLM main pin are unchanged. - `git revert --no-commit af277b2` applied cleanly on the latest fetched main. - Verified that the complete post-vllm-project#16393 upstream patch can still reverse-apply to the staged result (`git apply --reverse --check --cached`), confirming subsequent changes are preserved. - `git diff --cached --check`, Python AST parsing, Ruff lint and formatting checks passed for all 94 remaining changed Python files. - Ran the equivalent of `format.sh ci`: `pre-commit run --all-files --hook-stage manual`. Ruff, codespell, typos, clang-format, markdownlint, actionlint, package-init, forbidden-import and boolean-context checks passed. Bash-based hooks could not execute on this Windows host; two Python launcher hooks exited 9009. Full pre-commit validation remains for CI. - Actual CPU/NPU tests have not been run locally because the required vLLM/NPU runtime is unavailable. New PR CI is pending; the original upgrade PR's green run is not validation of this revert. - vLLM main: vllm-project/vllm@84030bb Signed-off-by: shenzhao <shenzhao9@huawei.com> Co-authored-by: shenzhao <shenzhao9@huawei.com>
tangdafu
pushed a commit
to tangdafu/vllm-ascend
that referenced
this pull request
Sep 23, 2026
…vllm-project#17004) ### What this PR does / why we need it? Reintroduce the vLLM v0.29.0 release upgrade from vllm-project#16393, reverted by vllm-project#16949, with container defaults aligned to the supported release. Based on upstream main `d3f6b4b59ab55dc42af1f30f425c893b95af5ee2`; preserves later merged changes. The image/source mismatch is confirmed in [nightly job 106004497020](https://github.com/vllm-project/vllm-ascend/actions/runs/35482881089/job/106004497020): the image installed `vllm 0.28.0+empty` (tag `v0.28.0`) while upgraded Ascend failed importing `_get_packed_kv_cache_groups`. All eight root Dockerfiles still defaulted to v0.28.0. The reusable image workflow passes no `VLLM_TAG` override, so those defaults govern release builds. Changing only the release marker does not update those images. - Fixed vLLM main remains `84030bbe3d74d99bad477a3d2e37a973ccd8865c`. - Release changes from v0.28.0 (`2cf0a6915ce544dc493a0990f2ea38d81601128a`) to official v0.29.0 (`98dff2a81d747d1dba01a47f939f48c3526d4206`, rechecked against the remote tag). - Restore vllm-project#16393's source compatibility and existing UT changes by reversing vllm-project#16949, then reconcile current upstream changes. No main interface scan is rerun for this release-only upgrade. - Update `VLLM_TAG` in all eight Dockerfiles: default/A3/A5/310P, Ubuntu/openEuler. Preserve existing exact-commit build overrides. No workflow changes. | Current-base adaptation | Exact cause and evidence | Lane | Validation | | --- | --- | --- | --- | | Eight Dockerfile `VLLM_TAG` defaults and release marker | vllm-project#16393 changed the release contract but omitted image defaults; the linked nightly log proves installation of 0.28.0 and missing `_get_packed_kv_cache_groups`. | Release images | All eight defaults match marker; image-build CI requested, pending | | `worker/v2/attn_utils.py::get_kv_cache_spec` and existing `_make_mla_layer` UT fixture | Ascend [vllm-project#16755](https://github.com/vllm-project/vllm-ascend/pull/16755/files), `df3e755e98fba8c6a18f200c645e0c2050469bc3`, added an old-release field selector. vLLM [#51718](https://github.com/vllm-project/vllm/pull/51718/files), `8bdc70ec7b379279ec0152343239c2d50aced687`, replaced `compress_ratio` with `tokens_per_state`; both exact supported pins use the latter. Use the common field, retaining metadata and cache-view assertions. | Both | Source inspection and static checks passed; actual CI pending | | `attention/attention_v1.py` import conflict | Preserve `attention_transfer_window` from Ascend [vllm-project#15636](https://github.com/vllm-project/vllm-ascend/pull/15636/files), `5c80630f28f8529aa82716e58b981a78819ec429`, and graph changes from vllm-project#16798; keep the common relocated PCP import from vllm-project#16393. Do not restore the old compute-start import or unused weak-reference import. | Both | Conflict resolved; static checks passed | | `worker/v2/aclgraph_utils.py` import conflict | Preserve `ContextSource`, `UpdatableGraph`, and `use_updatable_graph` from Ascend [vllm-project#16798](https://github.com/vllm-project/vllm-ascend/pull/16798/files), `34bb51f93724c565362f5108f5226303e1b56cad`, while removing the obsolete 0.28 version selector restored by the revert. | Both | Conflict resolved; static checks passed | | Existing 310P and Mamba model-runner UT imports | Preserve hardware-profile imports and mocks from Ascend [vllm-project#16803](https://github.com/vllm-project/vllm-ascend/pull/16803/files), `d3f6b4b59ab55dc42af1f30f425c893b95af5ee2`; remove only the unused old-release selector import. Hardware capability routing remains unchanged. | Both | Real rebase conflicts resolved; AST/Ruff/format passed; fresh CI pending | Newly merged changes were reviewed for version-contract impact: vllm-project#16775/vllm-project#16834 PCP metadata and routing, vllm-project#16923 A3 SFA, vllm-project#16426/vllm-project#16924/vllm-project#16955 custom ops, vllm-project#16081 MTP/SP, vllm-project#16669 xlite, vllm-project#15636/vllm-project#16747 transfer, vllm-project#16913 A5 pages, vllm-project#16798 graph updates, vllm-project#16755 GLM PD, vllm-project#16952 C8 config, vllm-project#16673 operator removal and vllm-project#16320 Kimi-K3 KV pool. Preserve these changes; no additional source-proven version branch was identified beyond the entries above. In particular, vllm-project#16747's `UniformTypeKVCacheSpecs.kv_cache_specs` and per-layer `block_size` exist in both exact pins; vllm-project#16320 adds Ascend connector hooks. CI remains necessary to validate runtime interactions. CI/documentation-only PRs are retained unchanged. <details> <summary>Inherited per-file compatibility evidence from vllm-project#16393</summary> The following source-contract ledger is inherited from vllm-project#16393. Any historical verification wording refers only to that earlier PR; it does not certify this new head. New-head validation is listed below. | Reference | Commit | |---|---| | vllm-ascend rebase base | `9dc6704559ebe2809b5cb7b7fc44184bd1338b3c` | | PR head | `8c5d785931c495701bf1da8b5bc80b321ebd3cd0` | | Fixed vLLM main | `84030bbe3d74d99bad477a3d2e37a973ccd8865c` | | Previous release: v0.28.0 | `2cf0a6915ce544dc493a0990f2ea38d81601128a` | | Target release: v0.29.0 | `98dff2a81d747d1dba01a47f939f48c3526d4206` | - Use common implementations where v0.29.0 and the fixed main share KV-cache layouts, Mamba copy/group APIs, PCP handling and speculative-decoding contracts. - Retain explicit `vllm_version_is("0.29.0")` branches for contracts that still differ, including RoPE, scheduler block snapshots, InputBatch, ReplaySSM, KV zeroing and DSpark PP handling. - Remove obsolete v0.28.0 compatibility and adapt existing test fixtures. Version detection uses package versions and the explicit `VLLM_VERSION` override, with local-version suffix handling and UT environment isolation retained; no hard-coded release-SHA inference remains. - Apply the Ascend PCP+DP validator patch only to v0.29.0, preserving other validations and EPLB platform binding. Rebuild dependent Pydantic schemas so nested configuration validation uses the patched validator. The global patch documentation records its rationale and removal criteria. - Preserve vllm-project#15747's Spec+PP protocol/partition handling after rebase; use the common exact-release selector and the real function-local DSpark sharing import. This release-only upgrade does not require a new main old-to-new interface scan. The latest rebase incorporates [vllm-project#16905](https://github.com/vllm-project/vllm-ascend/pull/16905/files), which reverted vllm-project#16544 at `9dc6704559ebe2809b5cb7b7fc44184bd1338b3c`. The now-unnecessary V4.1 drafter import gate and tuple annotation have been removed. Other release adaptations, including vllm-project#15747 Spec+PP handling, remain. Detailed contract evidence is retained below for review. <details> <summary>Per-file adaptations and exact upstream evidence</summary> #### Per-file adaptation ledger Evidence IDs refer to the exact source contract and upstream diff table below. Every row is syntax checked; branch-normalized AST comparison confirms unchanged main function bodies except the version identity helper and the explicitly retired propose argument/type annotations. Both supported versions completed CPU and NPU execution as recorded below. | Ascend file / symbols | Disposition and upstream evidence | Verification | |---|---|---| | `vllm_ascend/patch/worker/patch_v2/patch_dspark.py` / `_load_dspark_model_with_target_quant`; `tests/ut/patch/worker/test_patch_dspark_pp.py` | Preserve rebased [Ascend vllm-project#15747](https://github.com/vllm-project/vllm-ascend/pull/15747/files#diff-a170a42fe1e9275c999642b05c4a437d0c9104dd8ea19151796142909b00ab2a) (`82df9d871`), including manual PP partition masking. vLLM [#52809 diff](https://github.com/vllm-project/vllm/pull/52809/files#diff-2223c5e6b44b302db9bff4023777cc88b4427f4fb9d4a04476f6d6e4bd687393) (`91a893de64722019ea2faf852e06cabe143b3490`) moved `_should_share` inside the loader on both supported pins, so retain the earlier release fix: patch `eagle_utils`, never read/patch a nonexistent `dspark_utils._should_share`. v0.29 keeps its global PP guard; main has native PP via [#50514](https://github.com/vllm-project/vllm/pull/50514/files#diff-2223c5e6b44b302db9bff4023777cc88b4427f4fb9d4a04476f6d6e4bd687393) (`d87a440f88e28e5b37f9b1e22ce214d0426d5352`). | Existing UT now checks the real function-local import, absent module alias, and restoration on success/failure for both lanes; partition and PP assertions retained. Ruff/syntax pass; actual CPU/NPU pending. | | `vllm_ascend/worker/v2/pp_utils.py` / `use_legacy_spec_pp`; `tests/ut/worker/v2/test_pp_utils.py` | vllm-project#15747 added broad 0.28/0.29 routing and an obsolete 0.28 dev-build recognition path. For the two supported pins, #50514 exists only on fixed main. Route through `vllm_version_is("0.29.0")`; retain package/local-suffix and explicit environment-override semantics. No release-SHA inference or third release lane. | Existing UTs exercise the real uncached version helper with monkeypatch isolation, local suffix, fixed-main dev string, explicit override, and non-target versions. Isolated routing checks and Ruff/syntax pass; full CPU UT pending. | | `vllm_ascend/_310p/model_runner_310p.py`<br>`_prepare_inputs`, `_allocate_kv_cache_tensors` | #51718, #56078; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/_310p/spec_decode/llm_base_proposer_310.py`<br>`set_inputs_first_pass` | #56078; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/_310p/worker/v2/model_runner.py`<br>`initialize_kv_cache`, `_allocate_kv_cache_tensors`, `_prepare_inputs_310p` | #51718/#54436; use shared standardized layouts and retain the v0.29.0 InputBatch gate. The rebase preserves vllm-ascend vllm-project#16043's MTP copy tracking while removing only legacy v0.28.0 allocation paths. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/_310p/worker/v2/rope.py`<br>`get_310p_rope_state` | #56078; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/attention/attention_v1.py`<br>`module imports/dispatch` | #52839; both lanes use the common PCP import. Rebase keeps current upstream graph code and removes only the v0.28.0 import branch. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/attention/context_parallel/sfa_cp.py`<br>`module imports/dispatch` | #52839; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/attention/indexer.py`<br>`module imports/dispatch` | #52839; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/attention/mla_v1.py`<br>`module imports/dispatch` | #52839; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/core/dyntra_lb_scheduler.py`<br>`module imports/dispatch` | #51358, #54853; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/core/recompute_scheduler.py`<br>`module imports/dispatch`, `schedule` | #51358, #54853; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/core/scheduler_profiling_chunk.py`<br>`module imports/dispatch` | #51358, #54853; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/core/kv_cache_interface.py`<br>`get_kv_cache_compression_ratio`, `AscendMLAAttentionSpec`, `get_storage_block_size` | #51718, #53906; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/distributed/kv_transfer/kv_pool/ascend_store/layerwise_cache_layout.py`<br>`_merge_specs` | #51718; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/distributed/kv_transfer/kv_pool/kv_offload/preempt_offload/manager.py`<br>`_derive_cpu_config` | #51718; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/distributed/kv_transfer/kv_pool/kv_offload/native/npu.py`<br>`create_worker` | #52615; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/models/deepseek_mtp.py`<br>`load_weights` | #53106; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/models/deepseek_v4/indexer.py`<br>`get_kv_cache_spec` | #51718; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/models/glm5next/cache_config.py`<br>`make_tensor` | #51718; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/models/glm5next/kv_cache.py`<br>`get_kv_cache_spec` | #51718; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/models/kimi_k3_dspark.py`<br>`load_weights` | #53106; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/models/layer/attention/layer.py`<br>`get_kv_cache_spec` | #51718; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/patch/platform/patch_balance_schedule.py`<br>`module imports/dispatch` | #51358, #54853; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/patch/platform/patch_kv_cache_coordinator.py`<br>`__init__` | #53614; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/patch/platform/patch_kv_cache_utils.py`<br>`module imports/dispatch`, `_get_max_layers_per_page_size`, `_ascend_max_memory_usage_bytes_from_groups`, `_ascend_get_kv_cache_config_from_groups` | #51718; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/patch/platform/patch_use_v2_model_runner.py`<br>`module imports/dispatch`, `_patched_get_unsupported_features` | #53853, #53183; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/patch/worker/patch_bind_kv_cache.py`<br>`bind_kv_cache` | #51718; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/patch/worker/patch_bind_kv_cache.py::bind_kv_cache` | v0.29.0 has no ReplaySSM ring-tracker helper; fixed main requires it after KV binding. Evidence: [#52506, `adebc41b7e9f1085d3f73434e23beb76883b9eb4`, worker-utils diff](https://github.com/vllm-project/vllm/pull/52506/files#diff-645d58630d5acf3a0b07226bfef1e890a584c32502ab97c3d4642070f39a783c). Keep an explicit v0.29.0 exclusion and preserve the fixed-main call. | Failure reproduced on v0.29.0; Historical release/main NPU validation passed; current-head CI pending | | `vllm_ascend/patch/worker/patch_mamba_utils.py`<br>`module imports/dispatch`, `_get_state_copy_funcs_for_layer` | #53896; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/patch/worker/patch_v2/patch_attn_utils.py`<br>`module imports/dispatch` | #51718; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/patch/worker/patch_v2/patch_dspark.py`<br>`_load_dspark_model_with_target_quant` | #50514, #52809; v0.29.0 keeps `get_pp_group` module-bound but imports `_should_share` locally from Eagle utilities; fixed main removes the PP guard. Keep the release `get_pp_group` patch, share through the common Eagle utility, and delete the obsolete v0.28.0 `dspark_utils._should_share` patch. | Exact release failure reproduced; Historical release/main CPU/NPU validation passed; current-head CI pending | | `vllm_ascend/spec_decode/llm_base_proposer.py`<br>`model_returns_tuple`, `__init__`, `set_inputs_first_pass` | #56078, #52861; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/utils.py`<br>`get_kv_cache_tensor_layers`, `register_ascend_customop`, `vllm_version_is` | #52839, #51718, #52494; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/worker/model_runner_v1.py`<br>`_get_ascend_mamba_state_copy_funcs`, `_allocate_kv_cache_tensors`, `_prepare_inputs`, `_dummy_run`, `_reshape_kv_cache_tensors`, `get_kv_cache_spec`, NaN helpers | #51718/#53896/#56078 plus #50323; remove v0.28.0 layouts/copy APIs and import the NaN helpers directly because v0.29.0 and fixed main expose the same contract. XD-RoPE remains explicitly gated to v0.29.0. | Static checked; historical dual-version CPU/NPU validation passed; current-head CI pending | | `vllm_ascend/worker/v2/aclgraph_utils.py`<br>`capture` | #53869 makes `pcp_manager` common to both lanes. The rebase preserves vllm-ascend vllm-project#16409's host-parameter-update revert and removes only the obsolete v0.28.0 capture branch. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/worker/v2/attn_utils.py`<br>`get_kv_cache_spec`, `_allocate_kv_cache`, `_reshape_kv_cache_v2` | #51718; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/worker/v2/block_table.py`<br>`__init__`, `init_block_table_layout_tensors`, `compute_slot_mappings` | #51718, #51031; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/worker/v2/model_runner.py`<br>`module imports/dispatch`, `prepare_inputs`, `__init__`, `sample_tokens`, `prepare_dummy_attn`, `execute_model` | #50514, #54436, #52506, #55212, #53515; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/worker/v2/pcp_manager.py`<br>`partition_batch` | #53515; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/worker/v2/sample/gumbel.py`<br>`module imports/dispatch` | #54282; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/worker/v2/spec_decode/__init__.py`<br>`init_speculator` | #49811; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/worker/v2/spec_decode/autoregressive/speculator.py`<br>`propose` | #53694; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/worker/v2/spec_decode/dflash/speculator.py`<br>`module imports/dispatch`, `propose` | #53694, #52188; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/worker/v2/spec_decode/dspark/speculator.py`<br>`propose` | #53694; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/worker/worker.py`<br>`_scale_kv_cache_memory_for_multi_group`, `wake_up` | #51718, #53508; common contracts merged, differing release contracts explicitly gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | #### Exact upstream evidence | Upstream change | Full commit SHA / direct diff | Actual supported contracts and branch decision | |---|---|---| | #51718 | `8bdc70ec7b379279ec0152343239c2d50aced687`<br>[vllm/v1/kv_cache_interface.py](https://github.com/vllm-project/vllm/pull/51718/files#diff-f76cdfbf02dacd9dffbfbd0d9ad68a7a6ac0d8aed70834f95a2ae8ccd2e333cb) | Both use layers/layer_stride/block_stride/offset, tokens_per_state, CircularBufferSpec and standardized backing; retire shared_by/compress_ratio allocation branches. | | #52839 | `58e5ee0158b6a264c3506f00480e108a34b33ee3`<br>[vllm/v1/attention/ops/pcp.py](https://github.com/vllm-project/vllm/pull/52839/files#diff-23d7e7ab01a7f44414796e1e1b08cfa5277a05f54143ee2dd3a1041ca47772a8) | Both import PCP operations from vllm.v1.attention.ops.pcp. | | #53896 | `e126687a9a828d513c01a07cd69f025f27d63280`<br>[vllm/v1/worker/mamba_utils.py](https://github.com/vllm-project/vllm/pull/53896/files#diff-8ab86225fa08e9a6700851191abcb3f4f8b0bba40e97a82aad700d7a23a4d7ec) | Both use Mamba copy-function dictionaries and unwrap UniformTypeKVCacheSpecs. | | #53106 | `1fe3a1571ac67581478a11743e55a306de1d136f`<br>[vllm/model_executor/models/utils.py](https://github.com/vllm-project/vllm/pull/53106/files#diff-b0ba1095e9881e5c87e33dfd20958d1e1ceafe8a4433aa692f468e61be130b21) | Both use WeightsMapper instead of AutoWeightsLoader skip_prefixes/skip_substrs. | | #53906 | `98ed0856f31fa3aaf5e27464e2b4ef5a8ee6b2f5`<br>[vllm/v1/kv_cache_interface.py](https://github.com/vllm-project/vllm/pull/53906/files#diff-f76cdfbf02dacd9dffbfbd0d9ad68a7a6ac0d8aed70834f95a2ae8ccd2e333cb) | Only pinned main has the optional MLA storage_block_size dataclass field; release keeps the Ascend derived property, using tokens_per_state. | | #56078 | `719284fe158f1be8a9dd92953295fc9d49015730`<br>[vllm/config/model.py](https://github.com/vllm-project/vllm/pull/56078/files#diff-998c640befaf137b9af825f29f4e6e47d273caab1fd04093c97df24b18f5c417) | Release retains uses_xdrope_dim and three M-RoPE dimensions; pinned main uses mrope_num_dims and unified RoPE. | | #52615 | `138d137b5b955a5ebee98dfc946ecdd65d7b87ce`<br>[vllm/v1/kv_offload/cpu/spec.py](https://github.com/vllm-project/vllm/pull/52615/files#diff-7d073c26c9b22f74e9ff9e4c8da733d1807233edb83439fc64fbd48c43ae6045) | Release uses num_blocks/kv_bytes_per_block; main uses num_chunks/kv_bytes_per_chunk. | | #51358 | `6b110badbb22d3f66c7218b71138f13b7a6b3419`<br>[vllm/v1/core/sched/output.py](https://github.com/vllm-project/vllm/pull/51358/files#diff-cafd89ce8a698a56acb24ada62831cbc7a980782f78a52d1742ba238031f296c) | Release now has boundary_state_offloads and KVConnectorBlockState; remove partial_tail_offloads plumbing. | | #54853 | `0b066293f3c738a0cbd3a087bf893f2f4dcd61f2`<br>[vllm/v1/core/sched/output.py](https://github.com/vllm-project/vllm/pull/54853/files#diff-cafd89ce8a698a56acb24ada62831cbc7a980782f78a52d1742ba238031f296c) | Release constructor takes block_ids snapshots; main takes req_ids and resolve_block_ids. Keep exact release snapshot membership and main lazy-resolution membership. | | #53614 | `144e79c8106da23141ac010394b782f730cc7fe8`<br>[vllm/v1/core/kv_cache_coordinator.py](https://github.com/vllm-project/vllm/pull/53614/files#diff-43875c71daa893ef7567e21633d9988c2baf95bef61e3a334a6d584d6444d725) | Only main configures drop_eagle_checkpoint_block for replay-aligned Mamba checkpoints. | | #50514 | `d87a440f88e28e5b37f9b1e22ce214d0426d5352`<br>[vllm/v1/worker/gpu/spec_decode/dspark/utils.py](https://github.com/vllm-project/vllm/pull/50514/files#diff-2223c5e6b44b302db9bff4023777cc88b4427f4fb9d4a04476f6d6e4bd687393) | Release retains module-level PP/share symbols and Ascend PP workaround; main has the subsequent PP integration. | | #52809 | `91a893de64722019ea2faf852e06cabe143b3490`<br>[vllm/v1/worker/gpu/spec_decode/dspark/utils.py](https://github.com/vllm-project/vllm/pull/52809/files#diff-2223c5e6b44b302db9bff4023777cc88b4427f4fb9d4a04476f6d6e4bd687393) | Between v0.28.0 and v0.29.0, `_should_share` moved from a DSpark module binding to a function-local import from Eagle utilities. The shared Eagle patch remains effective; the old DSpark-module read/write must be removed. | | #54436 | `6bafc049aae6c26e210162630914ee9177a4b586`<br>[vllm/v1/worker/gpu/input_batch.py](https://github.com/vllm-project/vllm/pull/54436/files#diff-106f39c08266f186830bb8fcd7fb1df35c0aaa5fd0ac5c17aac64aeddee48902) | Release InputBatch requires max_seq_len_np; main removed it. | | #52506 | `adebc41b7e9f1085d3f73434e23beb76883b9eb4`<br>[vllm/v1/worker/gpu/model_runner.py](https://github.com/vllm-project/vllm/pull/52506/files#diff-5823f988fc0264681a80db24ccaba4d364f14394815d83ec1d90944c09f571f0) | Only main accepts valid_dummy_state_slots/valid_state_slots capture arguments. | | #55212 | `83990f5bcc0b5eb08b2f1fd1b109fa8a31b122f3`<br>[vllm/v1/worker/gpu/model_runner.py](https://github.com/vllm-project/vllm/pull/55212/files#diff-5823f988fc0264681a80db24ccaba4d364f14394815d83ec1d90944c09f571f0) | Release prepares DCP local sequence lengths before partitioning; main initializes DCP metadata afterwards. | | #53515 | `b1fbbc2ade51e3826bc92e4733c9c692ee21d42d`<br>[vllm/v1/worker/gpu/pcp_manager.py](https://github.com/vllm-project/vllm/pull/53515/files#diff-45dad5fe975c2ec8d02124fa65579cf11bfe9169a94217ceae1de0ea29518127) | Both accept padded_num_tokens for persistent PCP input buffers. | | #53869 | `b3af042abd8fe5a297ec3ec72db276fd661a67b3`<br>[vllm/v1/worker/gpu/cudagraph_utils.py](https://github.com/vllm-project/vllm/pull/53869/files#diff-fc699ff1c69fd17adb60116b3a7a20aa88b5187205f589952d0f3792475ec212) | Both accept pcp_manager during graph capture. | | #51031 | `0ecc284790e5403f74b899524ef82ecb69f83cb3`<br>[vllm/v1/worker/gpu/block_table.py](https://github.com/vllm-project/vllm/pull/51031/files#diff-21052649468f36e592c9ed378a9cd7b4615c558fb4d10f034f7b4268d3b6f9e1) | Both distinguish KV and kernel block sizes during DCP slot mapping. | | #54282 | `fe755c88995ad468882517b6c4bdd60138d46a3a`<br>[vllm/v1/worker/gpu/sample/gumbel.py](https://github.com/vllm-project/vllm/pull/54282/files#diff-02f7c5a06a2d23d544a07d57af16c7ffb74886469c8844385f51a719e329e6b8) | Both gumbel sampling APIs include is_drafting. | | #52188 | `d1e3eee6fb8ed3623241ef5c8e3ac533f775bff9`<br>[vllm/v1/worker/gpu/spec_decode/dflash/speculator.py](https://github.com/vllm-project/vllm/pull/52188/files#diff-0220b514682125dea26026db2d7caa1a0d9c772d84452e9e124434d1e11be5f0) | Both DFlash input kernels take cp_rank/CP_SIZE/CP_INTERLEAVE. | | #53694 | `5acc1c4e4b8730298cbff7a4a7c68c814dc24fd7`<br>[vllm/v1/worker/gpu/spec_decode/autoregressive/speculator.py](https://github.com/vllm-project/vllm/pull/53694/files#diff-3e7e2ba21d64e309503b7c5fe537363a7669fa1baf84d3c2e4891d9aed64fbe6) | Both propose APIs take DPSyncState; remove the obsolete token-count argument and retain replicated-PCP synchronization. | | #49811 | `01af92e175407231b1433b0aef01a1b9c983d955`<br>[vllm/v1/worker/gpu/spec_decode/extract_hidden_states.py](https://github.com/vllm-project/vllm/pull/49811/files#diff-b13061cd1be0bc80aaa9fce208fe9128aaaad25b3832c5f1693da1d4e8263f5e) | Both support extract_hidden_states on MRV2; remove old unsupported dispatch/skip. | | #53508 | `479eeb32d2b432dbb4e442fbd3f94ca2eca35d67`<br>[vllm/v1/worker/gpu_model_runner.py](https://github.com/vllm-project/vllm/pull/53508/files#diff-80ee7e2a62f9dcfbb8a312dc4e3948557e97ef187290daebbcae1e28596bda29) | Both remove post_kv_cache_wake_up; retire release-only call. | | #52494 | `3ff4f02dfe69abc1a0375d1ea8d8d5cb25609fcc`<br>[vllm/models/kimi_k3/amd/mla.py](https://github.com/vllm-project/vllm/pull/52494/files#diff-aad295e6de3607a46cea22f9f1cca768d4937dc5e1d3ca58105513358ed4e4c1) | Only main provides KimiK3MultiHeadLatentAttentionWrapper; retain release exclusion. | | #52861 | `b09bd69b5bf14911abef9a0e8e493b83c8a38fa6`<br>[vllm/v1/spec_decode/llm_base_proposer.py](https://github.com/vllm-project/vllm/pull/52861/files#diff-56fcad87dae9192cb0ab1643473765096cc44e1e744dccf3b99c4f2d85360d28) | Both include DeepseekV32MTPModel in the two-hidden-state architecture set. | | #54713 | `b28c3e1568bfae930f61d4b24940e47528c85d4a`<br>[vllm/v1/core/single_type_kv_cache_manager.py](https://github.com/vllm-project/vllm/pull/54713/files#diff-97c184a680b7a4bd7d58b11aa0073706533cc887d990eddb98469e7374025ab3) | Only main takes replay_boundaries in compressed-prefix hit lookup; preserve release calls without that keyword. | | #42785 | `442d36031ce710aa2a777c353becb660c57ab2bd`<br>[vllm/v1/worker/encoder_cudagraph.py](https://github.com/vllm-project/vllm/pull/42785/files#diff-287435cc78f753ae42acda542bca01c094557cd17db6bc2991d968fd012c1f18) | Only main capture callers pass axis_keys; preserve the existing Ascend rejection of nonempty axes. | | #52358 | `8f816a3f665489d7f0d222115d4f72ebab01076b`<br>[vllm/v1/worker/gpu/model_runner.py](https://github.com/vllm-project/vllm/pull/52358/files#diff-5823f988fc0264681a80db24ccaba4d364f14394815d83ec1d90944c09f571f0) | Both ExecuteModelState have dp_sync; only main has cudagraph_stats. | | #52789 | `9eb9d9d3953959695108600c8ed33d36bc6a1e5f`<br>[vllm/v1/core/sched/scheduler.py](https://github.com/vllm-project/vllm/pull/52789/files#diff-9eeca590fd99f15621897e559dba39b3ec4e7c2c65ec3c3229711689e008b5f4) | Release already has mamba_has_prefill_checkpoint_blocks; later main also has fine-grained prefix-cache state. | | #51251 | `7bbbf7c8e5040f7ebd374e8cbc657e01af1136dd`<br>[vllm/config/vllm.py](https://github.com/vllm-project/vllm/pull/51251/files#diff-bee6813076031d3ca1edc903c1b02b81e4676519afc562ce3fefe37f20c7b650) | v0.29.0 and main expose ec_manager_config; retire the old release-only ScoreEncoder configuration skip. | | #53240 | `b2db227a7c4c5e55f85524a094b607e5d27408b4`<br>[vllm/model_executor/layers/fused_moe/routed_experts_capturer.py](https://github.com/vllm-project/vllm/pull/53240/files#diff-0bbfc19dd02619bbcc948fa0281d4ab9a5b77b7d5062cde9911319c93089535c) | v0.29.0 supports MRV2 routed-expert capture and unwraps uniform cache groups; retire the old release-only replay skip. | | #53853 | `e376d45e82cb7e220da430e3179e81eb0922cf56`<br>[vllm/config/vllm.py](https://github.com/vllm-project/vllm/pull/53853/files#diff-bee6813076031d3ca1edc903c1b02b81e4676519afc562ce3fefe37f20c7b650) | Both delegate PCP compatibility validation to the PCP manager. | | #53183 | `4aab2b0ebed20343efe543c633f71b3c1336d5b8`<br>[vllm/config/vllm.py](https://github.com/vllm-project/vllm/pull/53183/files#diff-bee6813076031d3ca1edc903c1b02b81e4676519afc562ce3fefe37f20c7b650) | Both expose the V1 unsupported-feature helper used by the existing Ascend MRV1 feature filter. | #### Additional inherited contracts | vllm-ascend change | Why it is required | Upstream cause and direct link | Lane | Verification | |---|---|---|---|---| | `_310p/worker/v2/model_state.py`: remove tuple-return and RecoverSSM v0.28 fallbacks | v0.29.0 and fixed main both return `dict[MambaSpec, list[int]]` from `get_mamba_groups` and both initialize `recoverssm`; keeping the old fallback would preserve an unsupported third contract | [v0.29.0 mamba groups](https://github.com/vllm-project/vllm/blob/98dff2a81d747d1dba01a47f939f48c3526d4206/vllm/v1/worker/mamba_utils.py#L691-L704), [fixed-main mamba groups](https://github.com/vllm-project/vllm/blob/84030bbe3d74d99bad477a3d2e37a973ccd8865c/vllm/v1/worker/mamba_utils.py#L691-L704), [v0.29.0 RecoverSSM](https://github.com/vllm-project/vllm/blob/98dff2a81d747d1dba01a47f939f48c3526d4206/vllm/v1/worker/gpu/model_states/mamba_hybrid.py#L80-L103), [fixed-main RecoverSSM](https://github.com/vllm-project/vllm/blob/84030bbe3d74d99bad477a3d2e37a973ccd8865c/vllm/v1/worker/gpu/model_states/mamba_hybrid.py#L80-L103) | both, common implementation | Existing constructor UT now asserts the parent-created RecoverSSM value is retained; Ruff and compileall pass | | `worker/v2/model_runner.py`: always forward `kv_cache_allocation_context` | v0.29.0 and fixed main both accept this keyword; the new-base gate was specifically for v0.28.0 | [v0.29.0 signature](https://github.com/vllm-project/vllm/blob/98dff2a81d747d1dba01a47f939f48c3526d4206/vllm/v1/worker/gpu/model_runner.py#L533-L538), [fixed-main signature](https://github.com/vllm-project/vllm/blob/84030bbe3d74d99bad477a3d2e37a973ccd8865c/vllm/v1/worker/gpu/model_runner.py#L561-L566), [vllm-ascend vllm-project#16791](https://github.com/vllm-project/vllm-ascend/pull/16791/files) | both, common implementation | Existing UT continues to assert the exact context object reaches the parent; Ruff and compileall pass | | `_310p/worker/v2/model_runner.py`: select the release KV-zeroing contract with `vllm_version_is("0.29.0")` | v0.29.0 exposes `KVCacheGroupSpec.is_eagle_group` but lacks `SpeculativeConfig.use_eagle_block_drop`; fixed main added the method | vLLM [#53388 diff](https://github.com/vllm-project/vllm/pull/53388/files), commit [`481839ad9e5ebf87aecb54fa5c9d986bd5ea4b81`](vllm-project/vllm@481839a); [v0.29.0 group field](https://github.com/vllm-project/vllm/blob/98dff2a81d747d1dba01a47f939f48c3526d4206/vllm/v1/kv_cache_interface.py#L1200), [fixed-main helper](https://github.com/vllm-project/vllm/blob/84030bbe3d74d99bad477a3d2e37a973ccd8865c/vllm/config/speculative.py#L1874-L1876) | release differs from main | Existing two-path KV-zeroing UT retained and renamed for v0.29.0; Ruff and compileall pass | | existing DFlash kernel UT: remove v0.28-only kwarg omission | the current Ascend kernel accepts the CP arguments and the only excluded lane was v0.28.0, which this PR replaces | [vllm-ascend vllm-project#15098](https://github.com/vllm-project/vllm-ascend/pull/15098/files) | both, common invocation | Existing NPU test remains enabled with all assertions; Current-head CI pending | | Ascend change | Why / upstream cause | Lane | Verification | |---|---|---|---| | `patch/platform/patch_parallel_config.py`, registration, and global patch documentation | Allow Ascend PCP+DP by removing the generic GPU restriction, following vLLM [#54523](https://github.com/vllm-project/vllm/pull/54523/files), commit `7c2f1ff4958eaf0818405e9192c71608fe4a16b1`. Preserve dynamic `parallel_config.current_platform` lookup and rebuild ParallelConfig → SpeculativeConfig → VllmConfig. | v0.29.0 only | Existing configuration/EPLB UTs and historical PCP+DP NPU execution passed; current-head CI pending. | | `ops/triton/v2/sample/categorical_sample.py`, existing Gumbel UT | Both supported contracts require `is_drafting`, from [#54282](https://github.com/vllm-project/vllm/pull/54282/files), `fe755c88995ad468882517b6c4bdd60138d46a3a`; preserve the inherited optimized kernel and use a common wrapper. | Both | Existing positive drafting assertion retained; current-head CI pending. | | `patch/platform/patch_kv_cache_coordinator.py`, existing prefix-cache UT | Both support `cache_hit_alignment_tokens`, introduced by [#53598](https://github.com/vllm-project/vllm/pull/53598/files), `2ba984a5d06db414f3b2474fe9338faf6cd80a1c`; retire the v0.28-only write-mask branch. | Both | Existing assertions retained; current-head CI pending. | #### Rebase and retired-fallback evidence | Change | Exact source evidence | Decision | |---|---|---| | `tests/e2e/conftest.py::PROMPT_CONFIGS[HunyuanOCR]` | vLLM [#53272](https://github.com/vllm-project/vllm/pull/53272/files#diff-0852f1e9753819abe4f85380abdc0660fd4a998f44bd7bb73d68f462bd776d95), `d53b1c2efc0ca7161c3844f51ca9acbcdb7129d5` removes native Hunyuan V1/VL before v0.29.0; both exact supported sources lack it. The old conditional came from vllm-ascend [vllm-project#14898](https://github.com/vllm-project/vllm-ascend/pull/14898/files#diff-824bc3e5c28226a3c0f4581750fa93a7da7001aff9e0ff4aee4131e3edd7b144), `e5118d151314ae18e56c0b63aa8dd00d294adc22`. | Remove the v0.28.0 selector and use the existing exclusion for both supported lanes. This does not migrate the protected SFA PCP skip or add a v0.29.0 accuracy skip. | | `vllm_ascend/worker/model_runner_v1.py` NaN helper imports and `tests/ut/worker/test_model_runner_v1_nan_detection.py` | vLLM [#50323](https://github.com/vllm-project/vllm/pull/50323/files#diff-80ee7e2a62f9dcfbb8a312dc4e3948557e97ef187290daebbcae1e28596bda29), `12292d94b25869be2af6b6d4f8eea6c2445e935f`; both exact sources export `nans_to_dict`, `gpu_sync_allowed`, and `raise_if_nan_logits`. The v0.28.0 fallback originated in vllm-ascend [vllm-project#14898](https://github.com/vllm-project/vllm-ascend/pull/14898/files#diff-c49594855b615477bbc34f06d2d423a7dd84c021a7925cd1f61fdb79cb814c08), `e5118d151314ae18e56c0b63aa8dd00d294adc22`. | Delete import-failure/`None` fallbacks and the existing UT's obsolete availability skip; assertions remain unchanged and execute on both lanes. | | `vllm_ascend/worker/v2/aclgraph_utils.py` | vllm-ascend [vllm-project#16409](https://github.com/vllm-project/vllm-ascend/pull/16409/files#diff-dc1c8a80dadf23627467bf04d6bc59fe13c6d075b7d2564d0acc07dc97d921f3), `799801feef347469d5e9b39374b9210e0d4f7431`. | Preserve the upstream revert while resolving the real rebase conflict; do not reintroduce the reverted graph-update behavior. | | `vllm_ascend/_310p/worker/v2/model_runner.py` | vllm-ascend [vllm-project#16043](https://github.com/vllm-project/vllm-ascend/pull/16043/files#diff-2f27611a7cfbbdb780b9e88157f05d14d6dfe65406c8d66a5ba1f6949124f5d2), `d4d2957e5208c2f464d4625c05920bd29ea233cb`. | Preserve the newly merged 310P MTP copy tracking and remove only the v0.28.0 descriptor/allocation branch. | </details> </details> ### Does this PR introduce _any_ user-facing change? Yes. The supported vLLM release and default container builds move to v0.29.0, while the fixed main remains supported. The vllm-ascend package version does not change. Existing 0.29-only PCP+DP compatibility is restored. ### How was this patch tested? - Final E2E result for head `a3b76b073a201851454e873aa89fc2254992fc06`: [run 35504646654](https://github.com/vllm-project/vllm-ascend/actions/runs/35504646654) succeeded (39 successful jobs, 6 skipped). Raw logs from all 32 selected NPU jobs confirm the requested Ascend head, integration base `2de71b594319bde52c8bded69eba154f50be8e75`, and the actual vLLM pins/installations: fixed main `84030bbe3d74d99bad477a3d2e37a973ccd8865c` / `0.1.dev1+g84030bbe3.empty`, release `98dff2a81d747d1dba01a47f939f48c3526d4206` / `0.29.0+empty`. Each lane totals **563 passed, 35 skipped, 1 xfailed** across its selected pytest invocations. Skips/xfails are not passes. The resulting rebased integration commit is not printed and is not inferred. Actual release CPU and failed/cancelled image variants remain gaps. - Current-head CI update (2026-09-20): [main CPU UT](https://github.com/vllm-project/vllm-ascend/actions/runs/35504646654/job/106062951955) passed **5157 tests**, with **67 skipped**; actual installed vLLM was `0.1.dev1+g84030bbe3.empty`. Ascend checkout was the current PR head; the log does not print the full resulting integration head/base. Pre-commit and mypy passed. NPU/E2E results are recorded above; actual release CPU remains unverified. - Image build is partially blocked by infrastructure: A5 amd64 [Ubuntu](https://github.com/vllm-project/vllm-ascend/actions/runs/35504646536/job/106062362005) and [openEuler](https://github.com/vllm-project/vllm-ascend/actions/runs/35504646536/job/106062362068) failed before reading the Dockerfile because BuildKit could not create a snapshot temporary directory (`no space left on device`). No release compatibility code change is justified by this failure; cancelled variants remain unverified. The successful [310P openEuler arm64 build](https://github.com/vllm-project/vllm-ascend/actions/runs/35504646536/job/106062361987) explicitly checked out release `98dff2a81d747d1dba01a47f939f48c3526d4206` and installed `0.29.0+empty`; this is build evidence, not runtime UT coverage. - Current head: `a3b76b073a201851454e873aa89fc2254992fc06`, rebased onto the baseline above. Previous-head CPU [job 106060931458](https://github.com/vllm-project/vllm-ascend/actions/runs/35503843936/job/106060931458) stopped during Ascend integration rebase after vllm-project#16803 merged, before any UT ran. Its vLLM checkout was the fixed main and installation reported `0.1.dev1+g84030bbe3.empty`. The two import conflicts are resolved; fresh-head CI was retriggered by the push. - Remote official v0.29.0 tag resolved to the exact SHA above; both source pins inspected. - All 95 changed Python files pass Ruff lint, Ruff format and AST parsing; `git diff --check` passes. Dockerfile tag/marker consistency checked across all eight variants. - Full pre-commit invocation: Ruff, codespell, typos, clang-format, markdownlint, actionlint, package-init, forbidden-import and boolean-context checks passed. Bash-dependent hooks cannot run on this Windows host; Python launcher hooks exit 9009. Full CI lint remains pending. - Actual main/release CPU/NPU and image builds are not claimed locally: the required Linux/NPU/container environment is unavailable. New PR E2E and image-build CI are requested. Existing CPU workflow runs fixed main only, so actual release CPU remains a validation gap. - vllm-project#16393's historical green CI is not a substitute for this new head. No new test functions, workflow changes, golden/threshold changes or additional skips are introduced beyond restoring vllm-project#16393. Inherited 0.28-only tests/skips are not counted as passes. - vLLM main: vllm-project/vllm@84030bb Signed-off-by: shenzhao <shenzhao9@huawei.com> Co-authored-by: shenzhao <shenzhao9@huawei.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What this PR does / why we need it?
The AllGather MoE path gathers hidden states and router logits in DP-then-PCP order, while hash routing previously gathered
input_idsonly across DP. With PCP enabled, this leaves fewer token IDs than router-logit rows and causes hash routing to fail.This PR:
Does this PR introduce any user-facing change?
No API or configuration change. Hash-based MoE routing now works with the existing PCP configuration.
How was this patch tested?
A5 PCP4 (
DP1 × TP1 × PCP4): service started and ran normally. Full GPQA Diamond (enable_thinking=false) achieved 76.26% accuracy (151/198), with 198/198 requests successful and a 56.68% DSpark acceptance rate.vLLM main: vllm-project/vllm@84030bb