共 15 个 commit，涉及 51 个文件，+1662/-1008 行变动。

概要

统计项 数值
Commit 数 15
变更文件 51
新增行数 +1662
删除行数 -1008

Commit 列表

📦 Other

  • 96add73 #50641 — [Elastic EP] Fix non-contiguous weight transfers (#50641)
    • 作者: Itay Alroy | +11/-7 | 1 个文件

    Elastic EP weight transfer passed each weight tensor directly to PyNccl send API, but PyNccl transfers each tensor as a flat contiguous memory and does not account for tensor strides. For non-contiguous tensor views, such as some DeepSeek-V3 weights, this caused new workers to receive incorrect weights during scale-up. Instead, pack non-contiguous weights into a temporary contiguous buffer before …

  • e2fa285 #49934 — [1/N] Unify multiple-path encoder cuda graph support (#49934)
    • 作者: Isotr0py | +192/-387 | 7 个文件

    PLEASE FILL IN THE PR DESCRIPTION HERE ENSURING ALL CHECKLIST ITEMS (AT THE BOTTOM) HAVE BEEN CONSIDERED. - Separate from #49432 to avoid massive design changes. - Currently, for multi-path encoder like deepseek-ocr, we assumed that there’s two path (global/local) for encoder cuda graph. However, it’s not a good design for encoder cg manager to be aware of the special case global/local path. - Thi…

  • 38a466e #46789 — [DSV4] Implement Sequence Parallelism (#46789)
    • 作者: Woosuk Kwon | +138/-29 | 7 个文件

    Implement DeepSeek V4 sequence parallelism using the collective path introduced for Kimi K3. - Move the Kimi K3 SP helpers into a shared model-ops module and use the TP device communicator’s optimized custom_all_gather and custom_reduce_scatter, with the existing generic collectives as fallback. - Keep token padding, sharding, padding-mask propagation, and output trimming in the DeepSeek V4 model …

  • dc818c1 #45560 — [GPT-OSS] Strict tool call and constrained decoding for Harmony (#45560)
    • 作者: yzong-rh | +632/-511 | 10 个文件

    Following #45003 and #45396, attempt to implement constrained decoding for tool calls for Harmony models. Related: https://github.com/vllm-project/vllm/pull/35904. Also “embed content constraints in structural tags”. ### Problem vLLM’s constrained decoding for the Harmony delays enforcing response_format constraints (json_object, json_schema, structural_tag) until <|channel|>final…<|message|> is…

  • 127a7ce #50608 — Add @hongxiayang as code owner for AMD-specific model files and ROCm docs (#50608)
    • 作者: Hongxia Yang | +7/-3 | 1 个文件

    Shepherding and maintaining several components of ROCm-specific code base. ## Test Result —

  • 03c782e #50673 — [model registry] some simple typos (#50673)
    • 作者: Ning Xie | +26/-26 | 2 个文件
    1. the first return value for inspect_model_cls is model info instead of model class. Make it more understandable. 2. typo fix. NA ## Test Result NA —
  • 39f55ff #49608 — [Core] Offload raw-prompt preprocessing to renderer thread pool in AsyncLLM (#49608)
    • 作者: Almog Tavor | +37/-12 | 2 个文件

    Fixes #49317. AsyncLLM.add_request() calls InputProcessor.process_inputs() synchronously on the asyncio event loop thread. For raw prompts this includes tokenization and HF multimodal preprocessing (image resize/patchify into pixel_values), which holds the event loop for the whole duration. Under concurrent multimodal requests everything serializes before reaching the EngineCore scheduler, so max_…

  • 63e78ce #49611 — [Benchmark] Add probe requests to vllm bench serve (#49611)
    • 作者: Guan-Ming Chiu | +106/-0 | 3 个文件
    • Addresses part of #16353: no benchmark for how concurrent MM requests affect each other - Per review, implemented in vllm bench serve instead of a separate mm-processor mode - –probe-request-rate sends single-token text-only probes alongside the main workload (bypassing –max-concurrency) and reports their latency separately - Server: vllm serve Qwen/Qwen2.5-VL-3B-Instruct –max-model-len 8192 …
  • 652ba59 #50574 — [Model Runner V2] Enable encoder token embedding (#50574)
    • 作者: Taneem Ibrahim | +106/-7 | 4 个文件

    This PR enables the token_embed pooling task on Model Runner V2 for encoder-only models, and wires the existing LateInteractionRunner into the V2 pooling path so ColBERT-style scoring works under MRV2. This is the third task in the MRV2 pooling enablement: | Task | Status | | — | — | | embed, classify | #48791 (merged) | | token_classify | #50293 (merged) | | token_embed | this PR | ## Test Pl…

  • 3986b96 #50498 — (feat): optionally disable lookup on PD decode (#50498)
    • 作者: Jeff (Junze) Ma | +67/-3 | 5 个文件

    The original implementation is done by @ivanium . This PR introduces the enable_lookup flag for MooncakeStoreConnector, to be added in kv_connector_extra_config with kv_consumer role. This flag is for a PD setup where decode instance doesn’t read from or write to the Mooncake, and only contributes its segment for extending prefill worker KV cache capacity. e.g. - This avoid the potential overhead …

  • 77469c9 #50476 — [ROCm][MLA] Mask the AITER MLA small-head verify flatten causally (#50476)
    • 作者: Matt Qin | +284/-10 | 2 个文件

    [ROCm][MLA] Mask the AITER MLA small-head verify flatten causally On ROCM_AITER_MLA, a multi-token speculative verify block with fewer than 16 query heads per rank is flattened into single-token Gluon decodes with no causality check (vllm/v1/attention/backends/mla/rocm_aiter_mla.py:1006-1009), and every resulting row is handed its request’s entire paged-KV range, per_req_len[row_req] (rocm_a…

🐛 Bug Fix

  • 0601850 #50704 — [Bugfix][Models] Accept Qwen3_5MoeTextConfig in Qwen3_5MoeProcessingInfo for transformers 5.x compatibility (#50704)
    • 作者: Rafli Surya Wijaya | +5/-1 | 1 个文件

    Fix #50428: vLLM fails to load Qwen3.5-MoE text-only models (e.g. Qwen3.6-35B-A3B) when transformers 5.x is installed. In transformers 5.x, the top-level config class for Qwen3.5-MoE text-only models was renamed from Qwen3_5MoeConfig (the multimodal VL wrapper) to Qwen3_5MoeTextConfig. Qwen3_5MoeProcessingInfo.get_hf_config() hardcoded the wrapper type, so loading any Qwen3.5-MoE text-only model w…

  • c67fe49 #50701 — [Bugfix][Doc] Fix references to FusedMoE in doc (#50701)
    • 作者: bnellnm | +11/-11 | 2 个文件

    The rename of FusedMoE -> FusedMoEFactory did not update related uses in the docs. This PR replaces most instances of FusedMoE with “fused MoE” since it is more of a generic description of the layer. One specific link to the code is replaced with MoERunner where the naive all2all dispatcher is controlled. cc @mgoin @DarkLight1337 CI ## Test Result

  • ab06486 #46805 — [Bugfix][Kernel] Fix dangling temporary in AWQ gemm torch::stable::sum dim arg (#46805)
    • 作者: xiaolinchen | +5/-1 | 1 个文件

    awq_gemm (in csrc/libtorch_stable/quantization/awq/gemm_kernels.cu) ends with: The scalar 0 is wrapped into a single-element HeaderOnlyArrayRef whose constructor only stores a pointer to the temporary. When this sum call is not inlined, that temporary’s stack slot is overwritten (by the adjacent CUDA launch-config setup) before the boxed dispatcher reads it → the serialized dim becomes gar…

🦀 Rust Frontend

  • 81a42d3 #49498 — [Frontend] Add cache_salt support to Anthropic Messages API (#49498)
    • 作者: Cao Qian | +35/-0 | 3 个文件

    Closes #46688. The OpenAI-compatible Chat Completions / Completions APIs accept a cache_salt for explicit prefix-cache isolation in multi-user deployments. The Anthropic Messages API (/v1/messages) did not expose it, so clients built on the Anthropic schema had to switch API formats just to get deterministic prefix-cache isolation. ## Change - Add cache_salt: str | None to AnthropicMessagesRequest…