[Feature][KV Pool] Support and optimize Mooncake hybrid layerwise transfer by Eric-dot · Pull Request #16747 · vllm-project/vllm-ascend

Eric-dot · 2026-09-17T03:46:00Z

What this PR does / why we need it?

This PR adds Mooncake layerwise KV pool support for hybrid-attention cache layouts and optimizes the layerwise data path for long shared-prefix workloads such as DeepSeek-V4 Flash.

Build group-aware Mooncake range sessions for multiple KV cache groups, including different block sizes and sparse reachability masks.
Derive object identity from the canonical layout digest, group ID, and per-group block size without duplicating cache-family labels in the key.
Normalize scheduler/worker layout fingerprints across UniformTypeKVCacheSpecs packaging while retaining dtype and cache-geometry isolation.
Skip objects that are already readable in the pool before opening PUT sessions, reducing redundant writes for repeated prefixes and concurrent DP ranks.
Keep future-layer GET prefetch asynchronous and use a backend-internal PUT backlog of eight tasks instead of draining both queues after every attention layer.
Preserve transfer safety by draining both queues during full teardown and waiting for the final layer SAVE before source buffers can be reused.
Preserve supported retention arguments across cache-manager signatures and add a no-split range-metadata fast path.
Document the hybrid layout, internal timing policy, limitations, and validation workflow.

The asynchronous policy is backend-internal and exposes no runtime tuning variables. It keeps future-layer GET prefetch running and allows up to eight unfinished SEND tasks; errors and teardown still drain both queues completely.

Does this PR introduce any user-facing change?

Yes. Mooncake layerwise KV pool transfer can be used with hybrid-attention cache groups. The transfer scheduling policy is internal and introduces no new environment variables.

How was this patch tested?

Rebased onto current vllm-ascend/main (7a0aaebca).
ruff check passed for all changed Python files and related unit tests.
python -m compileall passed for the changed Python modules.
test_kv_transfer.py: 33 passed, 26 subtests passed in the local mock environment.
Targeted layout/retention regression tests: 4 passed.
Isolated lifecycle probes verified final-layer SAVE waiting and full receive-queue draining.
Markdown lint passed for the feature guide.

Real NPU and distributed Mooncake performance validation is still required and is not claimed by this PR.

RFC

模型	卡数	数据类型	是否开 layerwise	是否池化	输入	输出	External prefix cache	Prefix cache	P并行配置	D并行配置	并发数	总请求数	TTFT/s
DeepSeek-V4-Flash	8	W4A8	否	否	128K	1	90%	√	DP4	DP4	32	128	21.9
DeepSeek-V4-Flash	8	W4A8	是	是	128K	1	90%	/	DP4	DP4	32	128	22.1

vLLM main: vllm-project/vllm@84030bb

gemini-code-assist · 2026-09-17T03:47:21Z

Summary of Changes

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

This pull request introduces multi-KV-group support for Mooncake layerwise KV pool transfers, specifically targeting hybrid-attention cache layouts. It optimizes the data path for long shared-prefix workloads by enabling asynchronous prefetching and configurable PUT backlogs, while also ensuring transfer safety and providing new tuning parameters for performance fine-tuning.

Highlights

Multi-KV-group support: Added Mooncake layerwise KV pool support for hybrid-attention cache layouts, enabling better handling of complex attention patterns.
Performance optimizations: Implemented asynchronous prefetching and a configurable PUT backlog to improve data path efficiency for long shared-prefix workloads.
Transfer safety: Enhanced transfer safety by ensuring proper queue draining during teardown and adding configurable fence policies to manage transfer tails.
Configuration: Introduced new tuning variables VLLM_ASCEND_KVPOOL_FENCE_DRAIN_RECV and VLLM_ASCEND_KVPOOL_FENCE_SEND_BACKLOG for performance fine-tuning.

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 Assist

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

Invoking Gemini

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

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

Customization

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

Limitations & Feedback

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

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

github-actions · 2026-09-17T03:47:51Z

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

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

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

gemini-code-assist

Code Review

Suggested PR Title:\n\nmarkdown\n[Attention][Feature] Add Mooncake Layerwise Hybrid Attention for Multi-KV-Group Support\n\n\nSuggested PR Summary:\n\nmarkdown\n### What this PR does / why we need it?\nThis pull request adds multi-KV-group support to the Mooncake layerwise range/session implementation, targeting attention layouts such as DeepSeek-V4 where sliding-window KV, compressed KV, indexer caches, and compressor state belong to different KV groups. It integrates compute windows into Ascend DSA, FA, and SFA attention paths, introduces group-aware objects, and adds a configurable asynchronous pipeline.\n\nFeedback on the implementation highlights several critical issues:\n- An `AttributeError` in `kv_transfer.py` due to direct access of `_put_started_keys_lock` and `_put_started_keys` on the transfer thread instead of the worker.\n- Missing cleanup of `_group_load_indices` on exception in `_handle_range_layer_task`.\n- Potential `AttributeError` when accessing `_current_mooncake_last_chunk_req_ids` in `pool_worker.py`.\n- Signature generation issues in `group_cache_spec_signature` when handling `__slots__` or rank-dependent attributes during JSON serialization.\n\n### Does this PR introduce _any_ user-facing change?\nYes, it introduces new configuration options via environment variables (`VLLM_ASCEND_KVPOOL_FENCE_DRAIN_RECV` and `VLLM_ASCEND_KVPOOL_FENCE_SEND_BACKLOG`) and supports multi-KV-group attention layouts.\n\n### How was this patch tested?\nThe patch includes new unit tests in `test_mooncake_hybrid.py` and an opt-in end-to-end NPU smoke/benchmark script in `mooncake_layerwise_hybrid.py`.\n

Pz1116

LGTM

Eric-dot · 2026-09-18T09:51:40Z

/rerun
[Bot]: rerun completed.

Rerun (failed jobs only):

E2E

github-actions · 2026-09-19T13:20:20Z

This pull request has conflicts, please resolve those before we can evaluate the pull request.

github-actions · 2026-09-20T05:34:51Z

This pull request has conflicts, please resolve those before we can evaluate the pull request.

Add group-aware block-key range sessions for hybrid KV cache layouts, including canonical layout keys, backend-specific readability checks, independent group lifecycle tracking, and sparse reachability handling. Overlap ranged transfers with attention compute, keep bounded asynchronous prefetch and save backlogs, and preserve final-layer and teardown safety. Add CPU mock coverage and an English feature guide. Signed-off-by: fangrongcan <17343701736@163.com>

…#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` `_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` `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` `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` `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` `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` `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` `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` `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` `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` `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` `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` `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` `_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` `_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` `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` `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` `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` `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` `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` `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` `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` `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` `__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` `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` `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` `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` `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` `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` `_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` `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` `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` `_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` `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` `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` `__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` `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` `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` `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` `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` `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` `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` `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` `_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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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>

…lm-project#16798 This reverts commit 34bb51f. Restore FIA and speculative-decoding graph replay to the pre-UpdatableGraph path. Keep later unrelated changes already present on the branch, including attention_transfer_window from vllm-project#16747 and the vLLM 0.29.0 PCP import from vllm-project#17004. Signed-off-by: shiqiangA <shiqiang15@huawei.com> Co-authored-by: Cursor <cursoragent@cursor.com>

…lm-project#16798 This reverts commit 34bb51f. Restore FIA and speculative-decoding graph replay to the pre-UpdatableGraph path. Keep later unrelated changes already present after vllm-project#17004, including attention_transfer_window from vllm-project#16747 and the vLLM 0.29.0 PCP import. Signed-off-by: shiqiangA <shiqiang15@huawei.com> Co-authored-by: Cursor <cursoragent@cursor.com>

…nsfer (vllm-project#16747) ### What this PR does / why we need it? This PR adds Mooncake layerwise KV pool support for hybrid-attention cache layouts and optimizes the layerwise data path for long shared-prefix workloads such as DeepSeek-V4 Flash. - Build group-aware Mooncake range sessions for multiple KV cache groups, including different block sizes and sparse reachability masks. - Derive object identity from the canonical layout digest, group ID, and per-group block size without duplicating cache-family labels in the key. - Normalize scheduler/worker layout fingerprints across `UniformTypeKVCacheSpecs` packaging while retaining dtype and cache-geometry isolation. - Skip objects that are already readable in the pool before opening PUT sessions, reducing redundant writes for repeated prefixes and concurrent DP ranks. - Keep future-layer GET prefetch asynchronous and use a backend-internal PUT backlog of eight tasks instead of draining both queues after every attention layer. - Preserve transfer safety by draining both queues during full teardown and waiting for the final layer SAVE before source buffers can be reused. - Preserve supported retention arguments across cache-manager signatures and add a no-split range-metadata fast path. - Document the hybrid layout, internal timing policy, limitations, and validation workflow. The asynchronous policy is backend-internal and exposes no runtime tuning variables. It keeps future-layer GET prefetch running and allows up to eight unfinished SEND tasks; errors and teardown still drain both queues completely. ### Does this PR introduce _any_ user-facing change? Yes. Mooncake layerwise KV pool transfer can be used with hybrid-attention cache groups. The transfer scheduling policy is internal and introduces no new environment variables. ### How was this patch tested? - Rebased onto current `vllm-ascend/main` (`7a0aaebca`). - `ruff check` passed for all changed Python files and related unit tests. - `python -m compileall` passed for the changed Python modules. - `test_kv_transfer.py`: 33 passed, 26 subtests passed in the local mock environment. - Targeted layout/retention regression tests: 4 passed. - Isolated lifecycle probes verified final-layer SAVE waiting and full receive-queue draining. - Markdown lint passed for the feature guide. Real NPU and distributed Mooncake performance validation is still required and is not claimed by this PR. - [RFC][rfc] [rfc]: vllm-project#16759 [vllm-main]: vllm-project/vllm@84030bb | 模型 | 卡数 | 数据类型 | 是否开 layerwise | 是否池化 | 输入 | 输出 | External prefix cache | Prefix cache | P并行配置 | D并行配置 | 并发数 | 总请求数 | TTFT/s | |---|---:|---|---|---|---:|---:|---:|:---:|---|---|---:|---:|---:| | DeepSeek-V4-Flash | 8 | W4A8 | 否 | 否 | 128K | 1 | 90% | √ | DP4 | DP4 | 32 | 128 | 21.9 | | DeepSeek-V4-Flash | 8 | W4A8 | 是 | 是 | 128K | 1 | 90% | / | DP4 | DP4 | 32 | 128 | 22.1 | - vLLM main: vllm-project/vllm@84030bb Signed-off-by: fangrongcan <17343701736@163.com>

…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` `_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` `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` `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` `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` `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` `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` `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` `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` `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` `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` `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` `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` `_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` `_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` `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` `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` `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` `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` `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` `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` `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` `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` `__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` `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` `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` `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` `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` `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` `_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` `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` `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` `_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` `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` `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` `__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` `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` `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` `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` `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` `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` `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` `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` `_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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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>

…nsfer (vllm-project#16747) ### What this PR does / why we need it? This PR adds Mooncake layerwise KV pool support for hybrid-attention cache layouts and optimizes the layerwise data path for long shared-prefix workloads such as DeepSeek-V4 Flash. - Build group-aware Mooncake range sessions for multiple KV cache groups, including different block sizes and sparse reachability masks. - Derive object identity from the canonical layout digest, group ID, and per-group block size without duplicating cache-family labels in the key. - Normalize scheduler/worker layout fingerprints across `UniformTypeKVCacheSpecs` packaging while retaining dtype and cache-geometry isolation. - Skip objects that are already readable in the pool before opening PUT sessions, reducing redundant writes for repeated prefixes and concurrent DP ranks. - Keep future-layer GET prefetch asynchronous and use a backend-internal PUT backlog of eight tasks instead of draining both queues after every attention layer. - Preserve transfer safety by draining both queues during full teardown and waiting for the final layer SAVE before source buffers can be reused. - Preserve supported retention arguments across cache-manager signatures and add a no-split range-metadata fast path. - Document the hybrid layout, internal timing policy, limitations, and validation workflow. The asynchronous policy is backend-internal and exposes no runtime tuning variables. It keeps future-layer GET prefetch running and allows up to eight unfinished SEND tasks; errors and teardown still drain both queues completely. ### Does this PR introduce _any_ user-facing change? Yes. Mooncake layerwise KV pool transfer can be used with hybrid-attention cache groups. The transfer scheduling policy is internal and introduces no new environment variables. ### How was this patch tested? - Rebased onto current `vllm-ascend/main` (`7a0aaebca`). - `ruff check` passed for all changed Python files and related unit tests. - `python -m compileall` passed for the changed Python modules. - `test_kv_transfer.py`: 33 passed, 26 subtests passed in the local mock environment. - Targeted layout/retention regression tests: 4 passed. - Isolated lifecycle probes verified final-layer SAVE waiting and full receive-queue draining. - Markdown lint passed for the feature guide. Real NPU and distributed Mooncake performance validation is still required and is not claimed by this PR. - [RFC][rfc] [rfc]: vllm-project#16759 [vllm-main]: vllm-project/vllm@84030bb | 模型 | 卡数 | 数据类型 | 是否开 layerwise | 是否池化 | 输入 | 输出 | External prefix cache | Prefix cache | P并行配置 | D并行配置 | 并发数 | 总请求数 | TTFT/s | |---|---:|---|---|---|---:|---:|---:|:---:|---|---|---:|---:|---:| | DeepSeek-V4-Flash | 8 | W4A8 | 否 | 否 | 128K | 1 | 90% | √ | DP4 | DP4 | 32 | 128 | 21.9 | | DeepSeek-V4-Flash | 8 | W4A8 | 是 | 是 | 128K | 1 | 90% | / | DP4 | DP4 | 32 | 128 | 22.1 | - vLLM main: vllm-project/vllm@84030bb Signed-off-by: fangrongcan <17343701736@163.com>

…vllm-project#17004) Reintroduce the vLLM v0.29.0 release upgrade from vllm-project#16393, reverted by 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 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: 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> 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 (`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` `_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` `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` `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` `get_310p_rope_state` | gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/attention/attention_v1.py` `module imports/dispatch` | 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` `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` `module imports/dispatch` | gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/attention/mla_v1.py` `module imports/dispatch` | gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/core/dyntra_lb_scheduler.py` `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` `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` `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` `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` `_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` `_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` `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` `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` `get_kv_cache_spec` | gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/models/glm5next/cache_config.py` `make_tensor` | gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/models/glm5next/kv_cache.py` `get_kv_cache_spec` | gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/models/kimi_k3_dspark.py` `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` `get_kv_cache_spec` | gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/patch/platform/patch_balance_schedule.py` `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` `__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` `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` `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` `bind_kv_cache` | 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` `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` `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` `_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` `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` `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` `_get_ascend_mamba_state_copy_funcs`, `_allocate_kv_cache_tensors`, `_prepare_inputs`, `_dummy_run`, `_reshape_kv_cache_tensors`, `get_kv_cache_spec`, NaN helpers | 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` `capture` | #53869 makes `pcp_manager` common to both lanes. The rebase preserves vllm-ascend v0.28.0 capture branch. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/worker/v2/attn_utils.py` `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` `__init__`, `init_block_table_layout_tensors`, `compute_slot_mappings` | #51718, gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/worker/v2/model_runner.py` `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` `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` `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` `init_speculator` | gated as detailed below. | Static checked; prior-head dual-version results below; rebased CI pending | | `vllm_ascend/worker/v2/spec_decode/autoregressive/speculator.py` `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` `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` `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` `_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 | | Upstream change | Full commit SHA / direct diff | Actual supported contracts and branch decision | |---|---|---| | #51718 | `8bdc70ec7b379279ec0152343239c2d50aced687` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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. | | 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 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 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. | | 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> 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. - 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>

…nsfer (vllm-project#16747) ### What this PR does / why we need it? This PR adds Mooncake layerwise KV pool support for hybrid-attention cache layouts and optimizes the layerwise data path for long shared-prefix workloads such as DeepSeek-V4 Flash. - Build group-aware Mooncake range sessions for multiple KV cache groups, including different block sizes and sparse reachability masks. - Derive object identity from the canonical layout digest, group ID, and per-group block size without duplicating cache-family labels in the key. - Normalize scheduler/worker layout fingerprints across `UniformTypeKVCacheSpecs` packaging while retaining dtype and cache-geometry isolation. - Skip objects that are already readable in the pool before opening PUT sessions, reducing redundant writes for repeated prefixes and concurrent DP ranks. - Keep future-layer GET prefetch asynchronous and use a backend-internal PUT backlog of eight tasks instead of draining both queues after every attention layer. - Preserve transfer safety by draining both queues during full teardown and waiting for the final layer SAVE before source buffers can be reused. - Preserve supported retention arguments across cache-manager signatures and add a no-split range-metadata fast path. - Document the hybrid layout, internal timing policy, limitations, and validation workflow. The asynchronous policy is backend-internal and exposes no runtime tuning variables. It keeps future-layer GET prefetch running and allows up to eight unfinished SEND tasks; errors and teardown still drain both queues completely. ### Does this PR introduce _any_ user-facing change? Yes. Mooncake layerwise KV pool transfer can be used with hybrid-attention cache groups. The transfer scheduling policy is internal and introduces no new environment variables. ### How was this patch tested? - Rebased onto current `vllm-ascend/main` (`7a0aaebca`). - `ruff check` passed for all changed Python files and related unit tests. - `python -m compileall` passed for the changed Python modules. - `test_kv_transfer.py`: 33 passed, 26 subtests passed in the local mock environment. - Targeted layout/retention regression tests: 4 passed. - Isolated lifecycle probes verified final-layer SAVE waiting and full receive-queue draining. - Markdown lint passed for the feature guide. Real NPU and distributed Mooncake performance validation is still required and is not claimed by this PR. - [RFC][rfc] [rfc]: vllm-project#16759 [vllm-main]: vllm-project/vllm@84030bb | 模型 | 卡数 | 数据类型 | 是否开 layerwise | 是否池化 | 输入 | 输出 | External prefix cache | Prefix cache | P并行配置 | D并行配置 | 并发数 | 总请求数 | TTFT/s | |---|---:|---|---|---|---:|---:|---:|:---:|---|---|---:|---:|---:| | DeepSeek-V4-Flash | 8 | W4A8 | 否 | 否 | 128K | 1 | 90% | √ | DP4 | DP4 | 32 | 128 | 21.9 | | DeepSeek-V4-Flash | 8 | W4A8 | 是 | 是 | 128K | 1 | 90% | / | DP4 | DP4 | 32 | 128 | 22.1 | - vLLM main: vllm-project/vllm@84030bb Signed-off-by: fangrongcan <17343701736@163.com>

…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` `_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` `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` `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` `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` `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` `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` `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` `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` `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` `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` `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` `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` `_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` `_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` `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` `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` `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` `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` `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` `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` `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` `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` `__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` `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` `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` `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` `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` `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` `_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` `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` `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` `_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` `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` `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` `__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` `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` `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` `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` `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` `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` `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` `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` `_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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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` [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>

Eric-dot requested review from LCAIZJ, Yikun and weijinqian0 as code owners September 17, 2026 03:46

github-actions Bot added documentation Improvements or additions to documentation module:tests module:core labels Sep 17, 2026

gemini-code-assist Bot reviewed Sep 17, 2026

View reviewed changes

Eric-dot mentioned this pull request Sep 17, 2026

[RFC]: Mooncake Layerwise KV Pool Support for Hybrid-Attention Models #16759

Open

Eric-dot changed the title ~~[Performance][KV Pool] Optimize Mooncake hybrid layerwise transfer~~ [Feature][KV Pool] Support and optimize Mooncake hybrid layerwise transfer Sep 17, 2026

weijinqian0 reviewed Sep 17, 2026

View reviewed changes

Comment thread vllm_ascend/distributed/kv_transfer/kv_pool/ascend_store/pool_scheduler.py Outdated

github-actions Bot removed the module:core label Sep 18, 2026

Pz1116 added the ready-precise run selected e2e test for pr label Sep 18, 2026

Pz1116 approved these changes Sep 18, 2026

View reviewed changes

weijinqian0 previously approved these changes Sep 18, 2026

View reviewed changes

Eric-dot dismissed weijinqian0’s stale review via c0d35e1 September 18, 2026 08:10

nwpu-zxr approved these changes Sep 18, 2026

View reviewed changes

github-actions Bot added the merge-conflicts label Sep 19, 2026

github-actions Bot removed the merge-conflicts label Sep 19, 2026

Eric-dot force-pushed the mooncake-layerwise-hybrid-attention branch from 7f13bec to 1d7e0a7 Compare September 19, 2026 15:11

weijinqian0 previously approved these changes Sep 20, 2026

View reviewed changes

weijinqian0 reviewed Sep 20, 2026

View reviewed changes

Comment thread vllm_ascend/distributed/kv_transfer/kv_pool/ascend_store/pool_scheduler.py Outdated

Eric-dot dismissed weijinqian0’s stale review via b2eafc6 September 20, 2026 02:35

weijinqian0 previously approved these changes Sep 20, 2026

View reviewed changes

weijinqian0 enabled auto-merge (squash) September 20, 2026 03:05

auto-merge was automatically disabled September 20, 2026 03:54
Head branch was pushed to by a user without write access

Eric-dot dismissed weijinqian0’s stale review via d033570 September 20, 2026 03:54

Eric-dot force-pushed the mooncake-layerwise-hybrid-attention branch from d033570 to e93688c Compare September 20, 2026 04:09

github-actions Bot added the merge-conflicts label Sep 20, 2026

Eric-dot force-pushed the mooncake-layerwise-hybrid-attention branch from e93688c to 9ba5fba Compare September 20, 2026 06:09

github-actions Bot removed the merge-conflicts label Sep 20, 2026

weijinqian0 previously approved these changes Sep 20, 2026

View reviewed changes

weijinqian0 enabled auto-merge (squash) September 20, 2026 06:31

auto-merge was automatically disabled September 20, 2026 07:30
Head branch was pushed to by a user without write access

Eric-dot dismissed weijinqian0’s stale review via e87611f September 20, 2026 07:30

Eric-dot force-pushed the mooncake-layerwise-hybrid-attention branch from 9ba5fba to e87611f Compare September 20, 2026 07:30

weijinqian0 enabled auto-merge (squash) September 20, 2026 09:21

weijinqian0 approved these changes Sep 20, 2026

View reviewed changes

weijinqian0 merged commit ba7bc4d into vllm-project:main Sep 20, 2026
36 checks passed

zhao-stack mentioned this pull request Sep 20, 2026

[Misc][Main2Main] Upgrade vLLM to v0.29.0 with matching Docker images #17004

Merged

bowgneo mentioned this pull request Sep 20, 2026

[BugFix][KV Pool] Fix stale Mamba group reference #17023

Merged

shiqiangA mentioned this pull request Sep 21, 2026

[Bugfix] Revert host-side ACL graph parameter updates from #1… #17087

Closed

shiqiangA mentioned this pull request Sep 21, 2026

[test][Revert] Revert host-side ACL graph parameter updates from #1… #17111

Closed

Pz1116 mentioned this pull request Sep 22, 2026

[Feature][KV Pool] Support safe Mooncake layerwise PUTs for recurrent state #17194

Draft

Navigation Menu

[Feature][KV Pool] Support and optimize Mooncake hybrid layerwise transfer - #16747

[Feature][KV Pool] Support and optimize Mooncake hybrid layerwise transfer#16747
weijinqian0 merged 1 commit into
vllm-project:mainfrom
Eric-dot:mooncake-layerwise-hybrid-attention

Eric-dot commented Sep 17, 2026 •

edited by github-actions Bot

Loading

Uh oh!

gemini-code-assist Bot commented Sep 17, 2026

Uh oh!

github-actions Bot commented Sep 17, 2026

Uh oh!

gemini-code-assist Bot left a comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Pz1116 left a comment

Uh oh!

Eric-dot commented Sep 18, 2026 •

edited by vllm-ascend-ci

Loading

Uh oh!

github-actions Bot commented Sep 19, 2026

Uh oh!

Uh oh!

github-actions Bot commented Sep 20, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Conversation

Eric-dot commented Sep 17, 2026 • edited by github-actions Bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What this PR does / why we need it?

Does this PR introduce any user-facing change?

How was this patch tested?

Uh oh!

gemini-code-assist Bot commented Sep 17, 2026

Summary of Changes

Highlights

Footnotes

Uh oh!

github-actions Bot commented Sep 17, 2026

Uh oh!

gemini-code-assist Bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Pz1116 left a comment

Choose a reason for hiding this comment

Uh oh!

Eric-dot commented Sep 18, 2026 • edited by vllm-ascend-ci Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

github-actions Bot commented Sep 19, 2026

Uh oh!

Uh oh!

github-actions Bot commented Sep 20, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Eric-dot commented Sep 17, 2026 •

edited by github-actions Bot

Loading

Eric-dot commented Sep 18, 2026 •

edited by vllm-ascend-ci

Loading