夜雨聆风今日 78 个合入 PR｜vLLM 32 · SGLang 37 · PyTorch 1 · Python DSL / Compiler 8
一、PR 准确统计
统计口径：仅统计 Merge 的 PR，按 GitHub PR URL 去重；单仓 GitHub Search API 异常会记录到 raw_prs.json.errors。
二、重点关注 PR 分类
1. 性能优化 (Performance)
SGLang:
• #29771 [MoE] Consolidate ungrouped + grouped gate/topk onto one Triton router (#26771) — faster than AOT on B200/H100/H200, at parity with flashinfer — MoE 推理能力优化
• #29822 [AMD] Accept ROCm tensors in JIT kernel TensorMatcher + register 4 kernel tests — AMD/ROCm 生态适配或优化
• #27704 [Diffusion] Add profiling support and fix VBench dataset handling in bench_offline_throughput — 扩散模型推理支持/优化
• #30016 [diffusion] feat: performance_mode=speed enables torch.compile by default — 扩散模型推理支持/优化
Python DSL / Compiler:
• #2509 [JIT][Cache] Cache PyTorch extensions and perf wheels — 性能优化
vLLM:
• #47523 [Rust Frontend] Speed up chat roundtrip tests — 性能优化
• #47517 [Doc] Fix VLM2Vec benchmark chat template path — 问题修复，提升稳定性
2. 硬件支持 (Hardware Support)
vLLM:
• #46656 New stable abi cleanup — 硬件支持
• #47566 [Bugfix][Multimodal] Normalize direct PIL image inputs — 关键路径问题修复
• #47035 [ROCm] Fix encoder-decoder cross-attention KV layout aliasing — KV Cache / 缓存路径优化
• #45246 [CI] Enable sccache for Rust build under CUDA/ROCm — AMD ROCm 平台适配或修复
SGLang:
• #29703 [Anthropic] Fix missing cache_read_input_tokens in streaming responses — 问题修复，提升稳定性
• #29986 [AMD]: hot-patch transformers dynamic_module_utils symlink bug — AMD/ROCm 生态适配或优化
• #29908 [Apple Silicon] Add labeler config — 硬件支持
• #29807 Add XPU CI job monitor workflow — XPU 平台适配或修复
Python DSL / Compiler:
• #10793 [PROTON] Skip AMD tests in periodic flushing due to current issues — AMD/ROCm 生态适配或优化
3. 新模型/模型架构 (Models & Model Architecture)
vLLM:
• #47217 [Bugfix][Gemma4] Keep image bidirectional attention within the sliding window — 关键路径问题修复
• #47155 [GLM4V] Avoid GLM4V processor init during startup metadata reads — 新模型/模型架构
• #47539 [CI Bugfix] Lazily import Qwen warmup dependencies — Qwen 系列模型支持或修复
• #47480 [ROCm][CI] Adding qwen3 dp4 eplb — AMD ROCm 平台适配或修复
SGLang:
• #27914 [Intel GPU] DeepSeek V4 6/N: use sgl-kernel implemetation of flash_mla_with_kvcache on XPU — KV Cache / 缓存路径优化
• #29365 [CP] Consolidate decode-context-parallel (DCP) helpers into layers/dcp/ — 新模型/模型架构
• #29988 [dsv4] Trigger MHC prenorm prewarm at weight-load time with rank sync — 新模型/模型架构
• #29619 [DeepSeek-V4] Add an opt-in non-paged indexer for long-context prefill — DeepSeek 系列模型支持或优化
Python DSL / Compiler:
• #2513 [BugFix] Fix DeepSeek V3.2 topk threshold on exact-boundary inputs — DeepSeek 系列模型支持或优化
4. KV Cache / Connector / Offload
SGLang:
• #29631 [diffusion][cache-dit] add cache-dit support for Ideogram 4 — 扩散模型推理支持/优化
5. 推测解码 (Speculative Decoding)
vLLM:
• #46853 Add Laguna XS.2.1 DFlash drafter support — 推测解码
SGLang:
• #29995 [Spec] Remove the ServerArgs clone + global save/restore hack from DFlashWorkerV2 — 推测解码链路优化
6. 量化与精度 (Quantization & Precision)
SGLang:
• #30021 [CI] Add GLM52 NVFP4 MTP B200 tests — MTP / 投机解码能力增强
• #29937 [NPU] [DOC] add missing DEEP_NORMAL_MODE_USE_INT8_QUANT for w8a8+deepep scenarios — 量化与精度
• #29918 [AMD] Gate broken CK block-FP8 GEMM shapes to aiter-triton-GEMM to fix ROCm 7.0 Qwen3.5 accuracy — FP8 量化/算子能力增强
vLLM:
• #47220 [AMD][EPLB] Enable EPLB for Quark OCP MXFP4 MoE — MoE 推理能力优化
• #46456 [CI] intel CI: add quantization and awq case for xpu — XPU 平台适配或修复
7. 架构重构/工程演进 (Architecture & Refactor)
SGLang:
• #29017 [model-gateway] PD router: cancel paired decode when prefill fails — 路由/分布式调度能力增强
• #29915 [router] Log every engine /abort_request with a router_reason label + Prom counter — 路由/分布式调度能力增强
vLLM:
• #44682 [Bugfix][Rust Frontend] Tolerate out-of-vocab prompt ids in detokenizer — 关键路径问题修复
• #47384 [Bugfix][Frontend] Fix batch chat endpoint corrupting logprobs when return_token_ids is set — 关键路径问题修复
• #46684 [Rust Frontend] add repetition_detection support to sampling params — 架构重构/工程演进
• #47530 [Rust Frontend] Bump llm-multimodal version — 架构重构/工程演进
8. Bug/稳定性修复 (Bugfix & Stability)
Python DSL / Compiler:
• #2469 [BugFix] Fix LayoutInference divide-by-zero on non-power-of-two broadcast — 关键路径问题修复
• #10789 [Consan] Fix flaky multi-CTA CLC test — 问题修复，提升稳定性
SGLang:
• #29798 fix: avoid DSA indexer CPU seq lens fallback — 问题修复，提升稳定性
• #29874 [LoRA] DSA indexer targets + MoE-LoRA cuda-graph and RL adapter-reload fixes — MoE 推理能力优化
• #30018 [Fix] Turn off dsa indexer fusion by default — 问题修复，提升稳定性
• #30025 fix: reorder DSA indexer dual-stream ops to avoid CUDA graph stream explosion — CUDA Graph 执行路径优化
vLLM:
• #47483 [ModelRunner V2][BugFix] Free all model refs on shutdown — 关键路径问题修复
• #47465 [CI] Pin modelscope version to fix test breakage — 问题修复，提升稳定性
• #47485 [BugFix] Derive FlashInfer Q dtype from resolved per-group builder state — 关键路径问题修复
9. CI/测试/基础设施
SGLang:
• #29447 [CI] Add per-stage NVIDIA model inventory tool — CI / 测试基础设施调整
• #29932 add mimo-v2-flash model tutorial — CI/测试/基础设施
PyTorch:
• #188925 [release/2.13] Strip +PTX from CUDA arch list on release/RC builds (#188914) — CI/测试/基础设施
Python DSL / Compiler:
• #2512 Bump transformers from 5.0.0rc3 to 5.3.0 in /examples/bitnet-1.58b — CI/测试/基础设施
vLLM:
• #47467 [CPU][Build] Enable oneDNN ITT task collection by default for CPU primitive-level profiling — CI/测试/基础设施
10. 文档/Cookbook
SGLang:
• #29990 docs: sync LMSYS SGLang blog cards — 文档/Cookbook
11. 其他值得关注
Python DSL / Compiler:
• #2987 [pretuned kernels] cute scaled_mm on b200 — 其他值得关注
• #2514 [Transform][Layout] Avoid thread-indexed replicated fragment readback — 其他值得关注
vLLM:
• #47102 Add Triton Backend for Unlimited-OCR R-SWA — Triton / Python DSL 内核能力更新
• #47452 Move Roberta remaining nn.Embedding to VocabParallelEmbedding — 其他值得关注
三、最值得关注的 TOP 5 PR（每个项目）
vLLM TOP 5
SGLang TOP 5
PyTorch TOP 1
Python DSL / Compiler TOP 5
四、趋势分析
1. 活跃度对比
• vLLM 合入 32 个 PR。
• SGLang 合入 37 个 PR。
• PyTorch 合入 1 个 PR。
• Python DSL / Compiler 合入 8 个 PR。
2. 技术趋势
多硬件适配： 当日共有 31 个相关 PR，涉及 sgl-project/sglang、tile-ai/tilelang、triton-lang/triton、vllm-project/vllm。
• sgl-project/sglang#30021 [CI] Add GLM52 NVFP4 MTP B200 tests (https://github.com/sgl-project/sglang/pull/30021)
• sgl-project/sglang#29937 [NPU] [DOC] add missing DEEP_NORMAL_MODE_USE_INT8_QUANT for w8a8+deepep scenarios (https://github.com/sgl-project/sglang/pull/29937)
多模态/扩散： 当日共有 9 个相关 PR，涉及 sgl-project/sglang、vllm-project/vllm。
• sgl-project/sglang#29631 [diffusion][cache-dit] add cache-dit support for Ideogram 4 (https://github.com/sgl-project/sglang/pull/29631)
• vllm-project/vllm#47566 [Bugfix][Multimodal] Normalize direct PIL image inputs (https://github.com/vllm-project/vllm/pull/47566)
分布式/路由： 当日共有 6 个相关 PR，涉及 sgl-project/sglang、vllm-project/vllm。
• sgl-project/sglang#29771 [MoE] Consolidate ungrouped + grouped gate/topk onto one Triton router (#26771) — faster than AOT on B200/H100/H200, at parity with flashinfer (https://github.com/sgl-project/sglang/pull/29771)
• sgl-project/sglang#29017 [model-gateway] PD router: cancel paired decode when prefill fails (https://github.com/sgl-project/sglang/pull/29017)
量化技术： 当日共有 5 个相关 PR，涉及 sgl-project/sglang、vllm-project/vllm。
• sgl-project/sglang#30021 [CI] Add GLM52 NVFP4 MTP B200 tests (https://github.com/sgl-project/sglang/pull/30021)
• sgl-project/sglang#29937 [NPU] [DOC] add missing DEEP_NORMAL_MODE_USE_INT8_QUANT for w8a8+deepep scenarios (https://github.com/sgl-project/sglang/pull/29937)
Serving运行时： 当日共有 5 个相关 PR，涉及 sgl-project/sglang。
• sgl-project/sglang#29771 [MoE] Consolidate ungrouped + grouped gate/topk onto one Triton router (#26771) — faster than AOT on B200/H100/H200, at parity with flashinfer (https://github.com/sgl-project/sglang/pull/29771)
• sgl-project/sglang#29995 [Spec] Remove the ServerArgs clone + global save/restore hack from DFlashWorkerV2 (https://github.com/sgl-project/sglang/pull/29995)
MoE优化： 当日共有 4 个相关 PR，涉及 sgl-project/sglang、vllm-project/vllm。
• sgl-project/sglang#29771 [MoE] Consolidate ungrouped + grouped gate/topk onto one Triton router (#26771) — faster than AOT on B200/H100/H200, at parity with flashinfer (https://github.com/sgl-project/sglang/pull/29771)
• vllm-project/vllm#47220 [AMD][EPLB] Enable EPLB for Quark OCP MXFP4 MoE (https://github.com/vllm-project/vllm/pull/47220)
3. 对推理栈建设的启示
• 多硬件适配加速：vLLM / SGLang / Triton 在 AMD、CPU、CUDA Graph 等方向持续更新，值得持续跟踪异构硬件适配策略。
• KV Cache 与 Serving 路径优化：KV Offload、调度器、路由和缓存指标相关 PR 密集，值得映射到长上下文与服务化性能优化。
• 低精度与 MoE 能力：FP8/FP4/INT4/MoE 后端持续演进，可作为低成本推理能力建设参考。
关注「AI 每日参」，每天同步主流 AI 开源社区关键进展。
本文为自动生成的开源 PR 日报，重点关注 vLLM、SGLang、PyTorch、Triton、Helion 等社区。
恭喜你完成今日份的 AI 进化！里程碑已达成：🚩
别忘了顺手解锁 "点赞+在看+转发" 隐藏成就。
记得点亮 星标，防止由于算法调皮导致咱们"走散"。
撤了，明天同一时间见！👋
