vllm-ascend

There is no code change in this release. It just contains a quick fix for Dockerfile to make the official Docker image build as expected. If you're using v0.30.0rc1 official docker image, please upgrade to v0.30.0rc2 instead. Otherwise, Please just ignore this release.

Highlights

• Upgrade CANN to 9.2.0_20260930220649 in Dockerfile to make sure the official Docker image build as expected.

Full Changelog: https://github.com/vllm-project/vllm-ascend/compare/v0.30.0rc1...v0.30.0rc2

v0.30.0rc1 - 2026.10.12

This is vLLM Ascend v0.30.0rc1, a model-focused release built on the v0.29 release line and aligned with upstream vLLM v0.30.0 via the main2main pipeline. This release focuses on four flagship models: DeepSeek V4.1, GLM-5.3, GLM-5.3-Flash, and Qwen3.8-27B, and hardens the shared execution stacks they rely on (Model Runner V2, PCP/DCP parallelism, KV Cache Pool, and speculative decoding). Model or hardware combinations outside the scope declared below are not part of the release claims. Please follow the official documentation to get started.

Highlights

• DeepSeek V4.1: Added DeepSeek V4.1 support on Model Runner V2, enabled torch.compile, and sped up startup via in-place rotary embedding, reused filtered RoPE lookups, and removed redundant DSV4.1 indexer/SP copies; added the TurboQuant 4-bit latent KV cache for DSA sparse models (DeepSeek V4 Flash; A2/A3 only), including packed KV cache in the V1 runner; fixed ModelSlim-quantized DSV4 weight layout and added AscendStore preparation for V4.1, and packaged attention operators (QLI/QSLI/MQSMLA) with DSL. #17927 #16899 #17726 #17676 #18026 #18081 #16548 #16978 #17067 #17165 #17483 #16973
• GLM-5.3: Enabled sparse SFA C8 (INT8 packed KV cache) for GLM-5.3-Flash and indexer MXFP4 for GLM-5.2, added the TurboQuant 4-bit latent KV cache for GLM-5.2/5.3 SFA (A2/A3 only), fixed GLM MRV2 issues and GLM-5.2 DSpark sequence lengths/per-request FIA bounds, fixed GLM-5.3-Flash small slot page inflation, and backported always-on reasoning for GLM-5.3. #17115 #17949 #18060 #17613 #17571 #17545 #17876
• GLM-5.3-Flash: Added KV Pool support for non-layerwise pooling and Mamba in Mooncake layerwise, added layerwise memcache DRAM/SSD pooling, optimized Triton KPool and gated norm, supported GQA decode context parallelism and SFA-only DSA CP, and provided P/D and KV cache pool deployment guides. #16945 #17287 #17698 #17542 #16478 #17990 #17987 #17773 #17946
• Qwen3.8-27B: Supported non-contiguous KV cache in hybrid linear-attention models, respected per-layer KV cache dtype and kept KV cache contiguous for paged attention, restored C8 V scales on hybrid cache page reassignment, restored Mamba block tables before deferred PP postprocessing, added the FLA NPU backend for GDN prefill, and f

v0.27.1rc1 - 2026.09.24

This is the first release candidate of v0.27.1 for vLLM Ascend, aligned with upstream vLLM v0.27.1. This model-focused release adds MiniMax-M3 support on Atlas 800 A3 and 950DT Products, and new DeepSeek-V4-Flash support on 950DT Products. Other models and hardware combinations, such as Atlas 800 A2, were not sufficiently validated for this release. Models and hardware combinations outside this declared scope are not part of the v0.27.1rc1 release claim. Please follow the official documentation to get started.

Highlights

• MiniMax-M3 on Atlas 800 A3 and 950DT Products: Added BF16, W8A8, and ModelOpt-compatible MXFP8 serving paths, with optimized MSA index-score, top-k, FlashDecoding, tensor-parallel communication, and the 950DT Products FP8 sparse-attention path. #14412 #14539 #13142 #13987 #15897 #15926
• DeepSeek-V4-Flash on 950DT Products: Added the validated DSpark serving path, including DSpark integration with DSA-CP and SparseFlashMla-based BF16 KV sparse attention. Eligible compressor, indexer, sparse-attention, DSpark, and local DSA-CP metadata is built asynchronously on Model Runner V1. #13249 #13506 #13716 #15776
• Speculative-decoding and serving stability: Hardened sequence-length boundaries, empty-context handling, draft-state publication, rejection sampling, request token validation, and distributed MemCache initialization. #16513 #16720 #17276 #16215 #16161 #17308

Features

• Added a ModelOpt-compatible MXFP8 path for MiniMax-M3. With --kv-cache-dtype fp8, MSA and index caches use native FP8 E4M3 while dense GQA layers use BF16 unless explicitly configured otherwise. #14412 #15897 #15926
• Added DeepSeek-V4-Flash DSpark on 950DT Products, including DSA-CP integration and SparseFlashMla-based BF16 KV sparse attention. #13249 #13506 #13716
• Added DeepSeek-V4-Flash layerwise KV-cache reuse support for physical layers with multiple main cache specifications. #17081
• Integrated vLLM StartPlan so repeat startups can reuse measured KV-cache budgets and graph-capture results. #15654
• Integrated the GMM-SiTU operator for supported MoE paths. #15871

Hardware and Operator Support

v0.23.0.post1 - 2026.09.21

This is the first post release of vLLM Ascend v0.23.0. It includes the fixes, dependency updates, CI changes, and documentation updates merged into the v0.23.0 release branch after the v0.23.0 tag. Please follow the official documentation to get started.

Bug Fixes

• Fixed stale KV-cache writes during DP-aligned dummy runs by invalidating per-KV-group slot mappings before attention metadata construction. #15362
• Fixed MTP overlay prefix-cache precision on Atlas 300I DUO and kept W8A8 MXFP8 transformed buffers stable across RL weight reloads in ACL Graph mode. #14336 #13905

Other Changes

• Pinned the KV Pool dependencies to memfabric_hybrid==1.2.0 and memcache_hybrid==1.2.0. #14352
• Consolidated installation guidance, GLM-5/5.2 and Kimi-K3 deployment instructions, PD and 310P notes, release metadata, navigation titles, English comments, and Chinese translations. #15449 #14242 #14338 #14634 #14698 #14713 #14906 #15998 #16106 #14382 #14387 #14436 #14579 #14684 #15534 #16202
• Added release-branch nightly and weekly model configurations and installed concurrent-log-handler in release images. #14559 #14645 #14739 #16191