v0.26.0rc1

v0.26.0rc1 - 2026.09.03

This is the first release candidate of v0.26.0 for vLLM Ascend, aligned with upstream vLLM v0.26.0. This release is a model‑restricted version. Fully validated models include Kimi K3, GLM‑5.2, DeepSeek V4 Flash 0731, DeepSeek V4 Pro 0813. Availability is not guaranteed for other models. For the full test report, see: v0.26.0rc1 Test Conclusion. Please follow the official documentation to get started.

Highlights

• Kimi K3 on Ascend: Added end-to-end Kimi K3 support, including MLA DSpark speculative decoding, no-RoPE MLAPO on Ascend 950, fused QKV projections, fused norm gate and attention residual, and projector rotation. #12950 #13277 #13507 #13989 #13509 #14231
• GLM-5.2 / DeepSeek V4 Flash 0731 / DeepSeek V4 Pro 0813 accuracy and performance: Resolved GLM-5.2 DSpark acceptance regressions and DeepSeek V4 Flash-0731 / Pro-0813 accuracy issues (reasoning-effort alignment, routed SwiGLU limit, and frontend behavior), and reused DSV4 compressor metadata across layers for better performance. #12262 #13531 #13993 #14074 #14397 #14624 #14994
• DeepSeek V4 DSpark: Refactored DSv4 DSpark speculative decoding with aligned SP handling and QuaRot weight support for Qwen3 DSpark. #11431 #12662
• Sparse-attention context parallelism: Added SFA DCP with a replicated indexer, compact KV gather, and C8 support. Enabled P/D disaggregation for DCP with replicate-indexer. #11443 #11870 #11980 #11696
• Deprecation cleanup: Removed layer sharding, FlashComm2, multistream overlap gate, dynamic-batch SLO, weight prefetch, matmul all-reduce fusions, and kv offload in KV Pool to streamline the codebase. #11953 #12117 #11956 #11933 #11949 #12119 #11904

Features

• Added LoRA with unquantized MoE models and AlltoAll + EP + LoRA support. #10977 #12451
• Added Gemma4 E2B and E4B model support with graph execution on A2/A3 and ModelSlim quantization. #11536 #11575 #11791
• Added Step3.5/3.7 Flash support for Ascend 950. #10556
• Added KV sliding window for Eagle3 and DFlash. #10023
• Added ShortRequestFirst scheduling and batch job aware scheduler. #11576 #12240 #12719
• Added NZ static buffers for prefetch offload. #11945
• Added update config and reload weights to NPU Worker. #10126
• Added SFA C8 support: unified packed KV cache layout on A3, and DCP replicated indexer for sparse‑attention paths. #11228 #11870

Hardware and Operator Support

• Expanded Ascend 950 supports: MXFP quant for token dispatch, and CPU binding with topo clusters. #11614 #11717
• Expanded Atlas 300I DUO support: disabled npugraph_ex by default, added Qwen3.5-Dense documentation, and fixed 310P spec decoding accuracy. #10874 #12077 #11918
• Added operator supports: KV quant sparse flash attention, and Triton SwiGLuStep kernel (replaced AscendC‑based fused_gdn_gating). #11626 #11467 #12035

Performance

Unless stated otherwise, these optimizations are selected automatically for the targeted path and need no additional configuration.

• Optimized PCP FA restore and output merge to reduce overhead. #11586
• Vectorized local sequence-length computation in SFA metadata to remove per-request NPU-to-CPU synchronization. #11762
• Avoided H2D synchronization in context-parallel speculative-proposer metadata. #11496
• Optimized DSA-CP local token metadata with a fused Triton kernel and caching. #12193
• Split mixed ChunkedPrefill into separate decode and prefill attention calls. #11948
• Optimized AscendStore key construction and miss-path handling. #12814
• Removed D2H sync in QLIMetadata builder for DSA-CP. #12536
• Bound Mooncake receiver threads to KV cache device. #12126
• Sharded DeepSeek V4 DSpark main projection across TP ranks for better parallelism. #15144
• Reused DeepSeek V4 compressor metadata across layers to avoid redundant computation. #14994

Dependencies

• Upstream vLLM: v0.26.0.
• Python: >= 3.10, < 3.13.
• CANN: 9.1.0.
• PyTorch / torch_npu: 2.10.0 / 2.10.0.post4.
• Triton Ascend: 3.2.2 for A2, A3, and Ascend 950; Triton Ascend is not supported on Atlas 300I DUO.
• Mooncake: 0.3.11.post1 in the release images.

Deprecation and Breaking Changes

• Removed layer sharding. #11953
• Removed FlashComm2 optimization techniques. #12117
• Removed multistream overlap gate. #11956
• Removed dynamic-batch SLO. #11933
• Removed weight prefetch config. #11949
• Removed matmul all-reduce and matmul all-reduce RMSNorm fusions. #12119
• Removed KV offload in KV Pool. #11904
• Removed dp_allreduce_on_npu additional config option. #12496
• Removed PCP from MRV1; use MRV2 for prefill context parallelism. #12592
• Removed custom top-k top-p AscendC implementation in favor of CANN operator. #12232
• Refactored hamming ops and removed sparse action. #12049
• Deprecated ASCEND_BUFFER_POOL environment variable; use ASCEND_ENABLE_USE_FABRIC_MEM=1 or HCCL_INTRA_ROCE_ENABLE=1. #13856

Ready to Deprecate

The following features are deprecated or subject to change:

• Added deprecation warnings for W4A8 linear, W4A8 MoE per-group, and W8A8 PDMix MoE quantization. These quantization paths will be removed in a future release. #13850
• The mega_moe_max_tokens and enable_fused_mc2 configurations in additional-config will be moved into a new dedicated dict for centralized maintenance in the next release.
• DeepSeek-V3, DeepSeek-V3.1, and DeepSeek-R1 model support will be removed in v0.28.0.
• The reduce sampling feature is experimental and will be deprecated in future releases. We will follow the upstream vLLM community for the batch-sharded sampling feature.

Documentation

• Added Kimi-K3 deployment guide on Ascend. #14125
• Refreshed DeepSeek-V4, GLM-5.2, Kimi-K2.x, and Gemma4 model deployment guides. #11525 #11534 #11055
• Added KV Pool and Memcache SSD cache configuration guides. #12600 #12670
• Added RLHF LoRA HTTP API example and e2e RLHF async pause/resume example. #14672 #12300

Stability and Bug Fixes

• Fixed GDN state and graph-dispatch accuracy regressions across P/D, PCP, MTP, and DCP. #11195 #11893 #12027
• Fixed Qwen3.5 PCP + chunked prefill accuracy and small-chunk problems. #11508 #11807
• Fixed DeepSeek V4 MXFP routing precision and W4A16MXFP communication accuracy. #11663 #12585
• Fixed KV-transfer ordering, Mooncake grouping, and AscendStore hash/lease handling. #11887 #12252 #12012
• Fixed DCP/DP service hangs and restricted recompute scheduler to decode nodes. #12036 #11265 #11537
• Fixed 310P spec decoding accuracy, SDMA errors, and memory calculation. #11918 #11678 #11959
• Fixed DSpark multi-DP bugs and SP handling. #12612 #12777
• Fixed GLM-5.2 DSpark W8A8 eager acceptance. #12262
• Fixed fused_infer_attention contiguous errors in MLA-CP and GQA paths. #11986 #12025
• Fixed DSA-CP invalid local sequence lengths for graph padding. #12840
• Fixed invalid GVA handling in layerwise KV Pool. #12705
• Fixed AllGather MoE finalize order with DP and PCP. #11921
• Fixed batch invariance not taking effect. #11818
• Fixed Qwen3.x KV cache binding for multiple layers. #11470
• Fixed spec decode rejection sampling length-1 bug in Triton. #11458
• Fixed CPU binding device mapping. #15518
• Backported structured output fixes for vLLM 0.26. #15433
• Invalidated dummy slots before attention metadata build. #15409
• Skipped generate_mtp_attention_mask_for_decode with sparse attention to avoid hostbound. #15130
• Selected DeepSeek V4 effort mapping by checkpoint config. #14952

Known Issues

• Kimi K3 DSpark speculative decoding has a low draft acceptance rate for long input sequences, reducing the expected performance benefit. #15237
• DeepSeek-V4-Flash may show a performance regression for 128K-input/1K-output workloads when pooling is enabled, especially for single-prefix requests. #15649
• Qwen3-32B W8A8/QuaRot may show a small performance regression in PD-disaggregated or pooling-enabled scenarios on v0.26.0rc. #15295
• DeepSeek-V3.1 may hang or fail during PD-disaggregated serving and high-concurrency PD-mixed inference. These scenarios are outside the supported export-model scope. #14911
• Qwen3-235B-A22B W8A8 and Qwen3-235B W4A8 may show lower performance than the baseline in some A3 PD-mixed aclgraph scenarios. #15296
• GLM5.1-W4A4C8-mxfp4 may show a performance regression on Ascend 950 due to CPU-binding and load-balancing effects. #15268
• Qwen3.6-27B DFlash with mixed sliding-window and full attention requires the V2 model runner. #15650
• DeepSeek-V3.2-W8A8 may show a partial performance regression in A3 single-node mixed deployment. #15651
• DeepSeek-V4 sleep mode with sleep_mode_extra_cleanup may fail to recover after wakeup. Sleep mode is mainly intended for training scenarios. #15648
• KV pooling may cause a small single-prefix performance regression in some long-context workloads, including DeepSeek-V4-Flash 128K-1K and Qwen3-32B A2 PD-mixed scenarios. Investigation is ongoing. #15649 #15295 #12234
• MiniMax-M2.7 W8A8 shows a performance regression in the A3 single-node scenario. The 128K-input/1K-output double-node result also fell below the baseline threshold, but more samples are needed to determine whether it is a persistent regression or normal performance variation. #15658