共 12 个 commit，涉及 125 个文件，+2375/-2555 行变动。
概要
| 统计项 | 数值 |
|---|---|
| Commit 数 | 12 |
| 变更文件 | 125 |
| 新增行数 | +2375 |
| 删除行数 | -2555 |
Commit 列表
🐛 Bug Fix
- e6d1310 #48984 — [Bugfix] Reject removed pooling parameters (#48984)
- 作者: Taneem Ibrahim | +108/-17 | 9 个文件
Follow-up to #33477 and #33794. The pooling API removed normalize in favor of use_activation, removed the score task in favor of classify, and split the old encode task into token_embed and token_classify. Two validation gaps remained: 1. OpenAI request models allow unknown fields, so /v1/embeddings, /pooling, and /classify silently accepted removed fields. Validating inside each union member prod…
- 7c2acd3 #49015 — [Bugfix] Qwen3-VL/Qwen-Omni: honor max_pixels/min_pixels for video prompts (#49015)
- 作者: SYLAR | +30/-0 | 2 个文件
Fixes #48733. ## Problem For Qwen3-VL and the Qwen-Omni thinkers (Qwen2.5-Omni, Qwen3-Omni), video mm_processor_kwargs={“max_pixels”: …} (and min_pixels) is silently dropped, so the prompt uses the uncapped native resolution and prompt_tokens are inflated. Two things combine: 1. HF video processors read the per-frame pixel budget only from size.{longest,shortest}_edge, not max_pixels. 2. In the …
- c7ce03b #48988 — [Bugfix] Bump tml-fa4 for cutlass-dsl 4.6 API compatibility (#48988)
- 作者: Michael Goin | +1/-1 | 1 个文件
Bump tml-fa4 to include the cutlass-dsl 4.6 CuTe API migration (https://github.com/vllm-project/tml-fa4/pull/3) so Inkling Blackwell attention imports cleanly.
- d96aee0 #48025 — [Bugfix] Re-sync parameter tp_rank after process_weights_after_loading (fix replicated / disable_tp weight reload) (#48025)
- 作者: alexxu-roblox | +23/-7 | 3 个文件
Weight reload can crash for replicated (disable_tp=True) parameters such as the DeepSeek-V2/V3 / GLM MLA fused a-projection (q_a_proj + kv_a_proj_with_mqa when q_lora_rank > 0). BasevLLMParameter.init stamps self.tp_rank with the global rank. It is reconciled to the layer’s tp_rank (0 for disable_tp) by update_param_tp_status(), but that only runs at construction. When a parameter is re-create…
🔧 Refactor
- ac5f38a #49003 — [Refactor] Extract StructuredOutputsParams creation logic from Request.to_sampling_params (#49003)
- 作者: yzong-rh | +95/-98 | 5 个文件
Add extract_structured_outputs() helpers that normalize a request’s StructuredOutputsParams before building sampling params. Facilitate future Parser.adjust_request() implementation of https://github.com/vllm-project/vllm/pull/48116 and https://github.com/vllm-project/vllm/pull/45560 by consolidating the StructuredOutputsParams extraction logic. Deduplicate the structured-output normalization logi…
📦 Other
- b6ff8a2 #46570 — [Core] Add MRV2 virtual-batch PCP for MLA (#46570)
- 作者: Lucas Wilkinson | +1204/-196 | 39 个文件
Adds an MRV2 virtual-batch implementation for prefill context parallelism (PCP), initially targeting MLA and sparse MLA. - Add a stateful PCPManager that partitions the globally scheduled MRV2 InputBatch into rank-local virtual rows while preserving global request state for restore, logits, and postprocessing. - Add MRV2 input-batch plumbing for PCP positions, sequence lengths, logits indices, and…
- 9243e01 #49046 — [Multimodal] Automatically fallback to ViT DP when TP is unavailable (#49046)
- 作者: Isotr0py | +17/-5 | 3 个文件
- Automatically fallback to ViT DP when TP is unavailable ## Test Result —
- a287eb1 #48535 — [Front-end] [Messages] Populate
num_cache_creation_tokens(#48535)- 作者: yzong-rh | +201/-127 | 15 个文件
Follow up to https://github.com/vllm-project/vllm/pull/40912. When –enable-prompt-tokens-details is toggled, compute the number of prompt tokens written to prefix cache and pass it through to OpenAI Chat Completions and eventually to Anthropic Messages API. Ensures num_cache_creation_tokens is populated. Send and examine real requests to Qwen3.6-35B-A3B served via Messages API, both streaming and…
- c233d90 #48496 — Remove even more unnecessary
load_weightsmethods (#48496)- 作者: Harry Mellor | +639/-2076 | 44 个文件
Aside from deleting boilerplate, the new things in this PR are: - Extract and deduplicate weight filters used by many models: - maybe_fuse_shared_experts - the same code was repeated in several models for enabling the AITER fused-shared-experts optimisation - get_spec_layer_idx_from_weight_name/skip_spec_layers - the same code was repeated in several models for skipping the loading of the spec dec…
🧪 CI/Tests
- df362b2 #49055 — [ROCm][CI] Ensure sliding window tests release GPU memory (#49055)
- 作者: Andreas Karatzas | +32/-27 | 1 个文件
Use vllm_runner in the sliding-window correctness test to explicitly shut down the engine and wait for GPU memory to settle between parameterized cases. This prevents the next test from failing its startup memory check while the previous engine is still releasing VRAM. Motivation: - https://buildkite.com/vllm/amd-ci/builds/11005/list?tab=output&jid=019f7474-7e55-4f9c-9769-c5d1663dcd6c#L865
- 29c0ec4 #43164 — [ci] Move 3 entrypoints tests to h200_35gb queue (#43164)
- 作者: Kevin H. Luu | +4/-1 | 2 个文件
- Move Entrypoints Unit Tests, Entrypoints Integration (LLM), and Entrypoints Integration (Responses API) from the implicit gpu_1_queue path to the h200_35gb MIG queue. - Make the basic parser-context smoke request a concise answer so Qwen3 does not exhaust the 6,000-token test context while solving a multiplication prompt on H200. - Refresh the branch against current main while preser…
⚡ Performance
- e942438 #46832 — [ROCm][DSv3.2][Perf] Cap sparse MLA decode KV-splits with a work-per-split heuristic (#46832)
- 作者: frida-andersson | +21/-0 | 1 个文件
[ROCm][DSv3.2][Perf] Cap sparse MLA decode KV-splits with a work-per-split heuristic The ROCm aiter sparse-MLA decode backend calls get_mla_metadata_v1 without max_split_per_batch, so aiter defaults to -1 and splits the decode reduction across every CU (304 on gfx942). The reduce only covers the selected tokens per row (<= index_topk), so for topk=2048 that’s ~7 tokens/split — the reduce kernel …