共 16 个 commit，涉及 60 个文件，+1197/-382 行变动。

概要

统计项 数值
Commit 数 16
变更文件 60
新增行数 +1197
删除行数 -382

Commit 列表

⚡ Performance

  • d272418 #46026 — [Perf] Optimize Qwen3-VL multi-video prompt processing (#46026)
    • 作者: L丶 | +124/-26 | 2 个文件

    Replace text-level prompt expansion (decode + string replace + re-tokenize) with token-level replacement in _call_hf_processor. Co-authored-by: DeepSeek Optimize _call_hf_processor in Qwen3-VL for multi-video inputs. Problem: Each video’s expanded token IDs were decoded to text and inserted into the prompt via prompt.replace(), causing the prompt to grow with every video. The final super()._ca…

📦 Other

  • 7ff7f5c #46125 — Revert “Fix Stale Encoder Cache After Weight Update” (#46125)
    • 作者: Sumanth R Hegde | +0/-12 | 2 个文件

    Reverts vllm-project/vllm#45093 The decision to reset prefix cache should be left to the user. We should not be silently resetting prefix cache after every weight update I’m not sure about encoder cache though (i.e is it typical to reset encoder cache for multimodal async RL?)

  • 0fbf42a #46046 — [ROCm] Fix VRAM not freed in test_phi3v (#46046)
    • 作者: djramic | +2/-1 | 2 个文件

    test_phi3v.py::test_models_image fails on ROCm CI (MI300) with: ValueError: Memory of devices devices=[0] not free after dur_s=240.04 (threshold=‘0.100’) The HF reference forward in _run_test runs without disabling autograd. This keeps activations alive, so HfRunner.exit can’t free the VRAM (~22 GiB stays resident) and the wait_for_rocm_memory_to_settle() guard times out after 240s wit…

  • ca7e1f2 #45975 — Move CI failure diagnosis docs into ci-fails-buildkite skill (#45975)
    • 作者: Vadim Gimpelson | +40/-15 | 4 个文件

    Add Buildkite CI-failure diagnosis skill - Moves the Buildkite CI-failure diagnosis instructions out of AGENTS.md into a dedicated .claude/skills/ci-fails-buildkite skill, and updates supporting config so the skill is tracked. - Also refreshes CODEOWNERS for the GDN linear-attention path. Was refactored but CODEOWNERS wasn’t updated. ## Skill trigger examples The skill loads automatically from n…

  • dec860f #46176 — [ROCm] Use vLLM’s fp8 quant max in AITER hipBLASLt accuracy test (#46176)
    • 作者: djramic | +2/-1 | 1 个文件

    Fixes test_hipb_mm_kernel_forward_accuracy on MI300. The test quantized its fp8 reference with torch.finfo(e4m3fnuz).max (240.0), while the kernel uses vLLM’s get_fp8_min_max(), which on fnuz intentionally returns 224.0. That ~7% scale mismatch gave max_diff ≈ 10 and tripped the strict assert_close, even though the kernel was correct. The fix sources the test’s fp8 max from get_fp8_min_max(), so t…

  • 0a49fb2 #46181 — Fix dead link in docs (#46181)
    • 作者: Harry Mellor | +1/-1 | 1 个文件

    Bamba has been deleted. NemotronH is derived from Bamba

  • 0119213 #44577 — [DSv4] Pack KV caches into contiguous per-block allocations for DeepSeek V4 (#44577)
    • 作者: Tyler Michael Smith | +344/-22 | 7 个文件

    For DeepSeek V4, pack all layer data contiguously per block so that KV connectors can send/receive one region per block. Full-attention MLA + SWA/compressor caches share one contiguous allocation per block. Each layer gets an as_strided view with storage_offset into the packed backing tensor. The packed backing tensor and block_stride are passed through the cross-layer KV cache registration API so…

  • b9a7cd4 #45415 — [12/n] final _C library kernel migration (#45415)
    • 作者: Chris Leonard | +239/-212 | 17 个文件

    This PR continues the libtorch stable ABI migration (see https://github.com/vllm-project/vllm/issues/26946) for vLLM and is the final _C library kernels to move to the _C_stable_libtorch library. The PR moves csrc/quantization/activation_kernels.cu to csrc/libtorch_stable/quantization/activation_kernels.cu, along with the weak_ref_tensor (defined in ops.h), silu_and_mul_quant, and persistent_maske…

  • ec67d7a #40367 — [xpu] bump up vllm-xpu-kernels v0.1.10 and upgrade 2618 umd (#40367)
    • 作者: Kunshang Ji | +10/-9 | 4 个文件

    https://github.com/intel/compute-runtime/releases/tag/26.14.37833.4 ## Test Result —

🧪 CI/Tests

  • dced290 #46024 — [Hardware][AMD][CI] Fix e2e core test group (#46024)
    • 作者: Matt | +18/-14 | 3 个文件

    This PR fixes the e2e core test group. Cascade attention is only supported on the FLASH_ATTN and FLASHINFER backends, neither of which is currently supported on AMD. pytest -v -s v1/e2e/general –ignore v1/e2e/general/test_async_scheduling.py ## Test Result The test group passes. It is run as part of AMD CI. cc @AndreasKaratzas —

  • e6cd891 #46109 — [ROCm][CI] Skip Qwen3.5-35B-A3B-MXFP4-AITER-TP2 for non gfx950 (#46109)
    • 作者: Charlie Fu | +18/-0 | 2 个文件
  • 4a083cc #46180 — [ROCm][CI] Pin test_rocm_compressed_tensors_w8a8 to TRITON_ATTN (#46180)
    • 作者: Micah Williamson | +5/-2 | 2 个文件

    AMD: Kernels Quantization Test 2 (mi325_1) is failing on main ever since https://github.com/vllm-project/vllm/pull/44446 was merged. There is a seg fault in the ROCM_ATTN backend, so we’re pinning to TRITON_ATTN to unblock CI while I investigate the fix (see https://github.com/vllm-project/vllm/issues/46179).

  • 4a8abf3 #46173 — [Test] Migrate test_openai_schema.py to schemathesis 4.x (#46173)
    • 作者: Ben Browning | +61/-48 | 4 个文件

    Follow-up to #45675 which upgraded schemathesis to 4.x for a security fix. That PR updated the pinned versions but not the test code, which uses APIs that changed across the major version boundary. In addition to migrating to the new APIs, this also had to bump the schemathesis lower bound to >= 4.0.0 in requirement files because there are breaking changes between schemathesis 3.x and 4.x. And, sc…

🐛 Bug Fix

  • 93bad11 #45255 — [Bugfix] Fix gridDim.y overflow for large row counts (#45255)
    • 作者: JasonLi314 | +71/-12 | 2 个文件

    Fixes #45099. Prior code mixed up cuda gridDim.x and gridDim.y, mapping the row dimension (mn) to gridDim.y, which CUDA caps at 65535. This fix: * Assigns row dimension mn to gridDim.x instead of gridDim.y, as grid x limit is capped at 2^31 - 1. * Fixed the launch guard, which previously checked both grid dims against INT32_MAX. Now grid x is checked at INT32_MAX while grid y is checked against 65…

  • 859e4d4 #46159 — [Bugfix][Parser] Fix U+FFFD leak at reasoning-to-content transition in engine parsers (#46159)
    • 作者: Ben Browning | +191/-7 | 5 个文件

    Flush the reasoning parser’s engine lexer at the reasoning→content transition instead of decoding isolated token IDs, which produces U+FFFD when byte-fallback tokens span delta boundaries. This fixes a source of unicode replacement characters getting streamed back to clients first identified in https://github.com/vllm-project/vllm/pull/45919#issuecomment-4732432638 ### Unit Tests ### Manual Test w…

  • 69bdd34 #46038 — [Bugfix] Fall back to Pydantic loc for param in validation errors (#46038)
    • 作者: Muhammad Fawaz | +71/-0 | 2 个文件

    Previously, param was only populated for errors carrying a custom VLLMValidationError in their Pydantic ctx, leaving it None for ordinary missing-field/wrong-type validation errors even though the field name was available from error[’loc’]. Fixes the param field in validation error responses being silently None for ordinary Pydantic validation failures (missing required fields, wrong types), even …