夜雨聆风今日 174 个合入 PR｜vLLM 54 · SGLang 73 · PyTorch 6 · Python DSL / Compiler 41
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)
vLLM:
• #50089 [Model] Add Kimi K3 support: model files and kernels [1/N] — 性能优化
• #43538 [CI/Perf] Fix malformed serving benchmark config — 问题修复，提升稳定性
• #50244 [torch.compile] Compile CustomOp.forward_native for ReLU^2 to avoid raw torch ops inside opaque custom ops — 性能优化
• #49750 [Perf] RMSNorm uncontiguous support, 1.2~3.1x kernel performance improvement — 性能优化
SGLang:
• #32642 Add a benchmark script for the HPC-Ops bf16xfp32 router GEMM — 路由/分布式调度能力增强
• #32612 Support DCP for Kimi Linear model — 性能优化
• #30614 [Diffusion][Docs] Ascend A2, A3 add basic usage and benchmark results in diffusion cookbook — 扩散模型推理支持/优化
• #32819 feat(kda): optimize PTX prefill kernel — 性能优化
Python DSL / Compiler:
• #2782 [BugFix] Fix warp_reduce truncating int64/uint64 to 32 bits on sm_80+ — 关键路径问题修复
• #3147 [cutedsl] Optimize dense HD64 two-CTA flash attention — 性能优化
• #3195 Benchmark fused QK RoPE against vLLM fused op — 性能优化
• #3200 [examples] Perf-tune the linear-attention kernels further — 性能优化
PyTorch:
• #4641 Remove disabled A100 perf workflow and dead llama-benchmark scripts — 性能优化
2. 硬件支持 (Hardware Support)
SGLang:
• #31409 [AMD] Replace MI325 with MI300 CI Runners — AMD/ROCm 生态适配或优化
• #32760 docker: add Kimi K3 images — 硬件支持
• #32447 [MLX] Fix overlap-loop request bookkeeping and graceful shutdown — 问题修复，提升稳定性
• #32596 sgl-kernel-npu tag update to 2026.7.27 — 硬件支持
vLLM:
• #48677 [XPU] upgrade to torch 2.13 — XPU 平台适配或修复
• #50304 [CI][ROCm] Fix AMD nightly distributed regressions — AMD/ROCm 生态适配或优化
• #47121 [XPU] Route weightless RMSNorm to _C dispatch — XPU 平台适配或修复
• #50190 [ROCm][CI] Stabilize ngram and suffix correctness test — AMD ROCm 平台适配或修复
Python DSL / Compiler:
• #10878 [AMD] Add RDNA4m target — AMD/ROCm 生态适配或优化
• #11078 [AMD][BACKEND] Split multicast groups into smaller subgroups if exceeding HW limit — AMD/ROCm 生态适配或优化
• #3006 [autotuner] H100 matmul seed: a general budget-formula heuristic — 硬件支持
• #11081 [AMD] Fix swapped structured binding in emitFence for buffer atomics — AMD/ROCm 生态适配或优化
3. 新模型/模型架构 (Models & Model Architecture)
SGLang:
• #31563 fix mqa preshuffle layout issue for deepseek v4 — DeepSeek 系列模型支持或优化
• #32228 Skip mamba lock during decoding — 新模型/模型架构
• #32022 fix(qwen3.5): restrict MoE weights to local PP layers — MoE 推理能力优化
• #31747 [AMD] DSv4: bring HIP compress-state pool into the memory_saver KV_CACHE region — KV Cache / 缓存路径优化
vLLM:
• #41602 [Bugfix] Fix /wake_up crash on hybrid models (Mamba/DeltaNet) — 关键路径问题修复
• #50092 [Misc][Minimax-M3]add default video_processor — 视频/多模态场景支持优化
• #49291 [Kernel][Mamba] Fused-kernel support for align-mode DS-conv state migration with num_accepted_tokens > 1 — 新模型/模型架构
• #49757 [BugFix] Stop dummy runs from writing mamba state through stale block-table rows — 关键路径问题修复
4. KV Cache / Connector / Offload
SGLang:
• #31968 [Disagg][NIXL] Fix heterogeneous attn-TP KV transfer for replicated GQA heads (NIXL_ERR_NOT_FOUND) — KV Cache / 缓存路径优化
• #32025 [PD] NIXL connector: shard by destination — KV Cache / Connector / Offload
• #31706 [Elastic EP] fix previously flaky test of test_mooncake_ep_small.py — Mooncake KV Connector 相关优化
• #32743 [diffusion] fix: dual-DiT models crash with (1,)-placeholder weights after compile-time offload — 扩散模型推理支持/优化
vLLM:
• #49975 [Bugfix][Multimodal] Include media IO config in MM cache hash — 关键路径问题修复
• #49114 Add CachePolicyFactory for pluggable/external eviction policies — CI / 测试基础设施调整
• #49762 [KV Connector] Support NIXL P/D for hybrid MLA+SSM models — KV Cache / 缓存路径优化
• #50033 feat(grpc): add KV event source discovery — KV Cache / 缓存路径优化
Python DSL / Compiler:
• #2793 [BugFix] Prevent autotuner cache reuse across different outputs and validation settings — 关键路径问题修复
• #11093 Avoid reference cycle in trition.runtime.jit.compute_cache_key — KV Cache / Connector / Offload
5. 推测解码 (Speculative Decoding)
SGLang:
• #29735 [GDN] Support FlashInfer GDN prefill with extra-buffer radix cache — KV Cache / 缓存路径优化
• #32823 perf(kimi-k3): optimize KDA MTP ns0 memory path — MTP / 投机解码能力增强
• #32396 [EAGLE] Handle NaNs in fused top-k=1 — 推测解码
• #31739 [NPU] adapt dflash v2 on npu — 推测解码
vLLM:
• #49731 [Spec Decode][Perf] Replicate DSpark Markov head across TP ranks — 推测解码链路优化
• #49343 [BugFix] eagle draft max position embeddings — 关键路径问题修复
• #48883 fix(step3p5-mtp): honor exclude_modules for the MTP head via prefix — MTP / 投机解码能力增强
• #50094 [KV Offload] Move CPUOffloadingSpec onto SharedOffloadRegion — KV Cache / 缓存路径优化
6. 量化与精度 (Quantization & Precision)
SGLang:
• #32118 Fix nightly CI: NVFP4 cuda-graph crash, NVILA batching, CuTe paged-KV zero-size, Kimi-VL OOM — KV Cache / 缓存路径优化
• #32699 perf(kimi-k3): fuse MoE-front prep (route + trtllm pack + mxfp8 quant) into one launch — MoE 推理能力优化
• #32013 :bug: [llm][npu][quant] Fix ModelSlim MXFP4 packed weight loading — FP4 量化推理能力增强
• #30768 :sparkles: [llm][npu][quant] Add W8A8 MXFP8 quantization for Qwen3 MoE on Ascend NPU — MoE 推理能力优化
vLLM:
• #43229 [CompressedTensors] FP4 Qutlass Integration — FP4 量化推理能力增强
• #49580 Integrate CuTeDSL MoE for ReLU2 NVFP4 — MoE 推理能力优化
• #50222 [CI] Fix MXFP8 MOE backend selection tests on gfx942 — MoE 推理能力优化
• #50194 [CPU] Fix FP8 attention scratchpad sizing — FP8 量化/算子能力增强
Python DSL / Compiler:
• #3046 Add pretuned NVFP4 decode GEMV benchmark vs vLLM CUTLASS — FP4 量化推理能力增强
PyTorch:
• #4640 Fix ROCm rowwise-fp8 warning crash — FP8 量化/算子能力增强
• #4643 Stop quantize_ from leaking global fp32 matmul precision (complete fix) — 问题修复，提升稳定性
7. 架构重构/工程演进 (Architecture & Refactor)
vLLM:
• #49665 [Frontend][Core] Standardize request error handling with VLLMError hierarchy — 架构重构/工程演进
• #50093 [Model] Add Kimi K3 support: Python frontend [2/2] — 架构重构/工程演进
• #49604 [Rust Frontend] Add --limit-mm-per-prompt support — 架构重构/工程演进
• #48543 [Frontend] Add diarized_json support for MOSS-Transcribe-Diarize — 架构重构/工程演进
SGLang:
• #32104 [EPD][VLM] Fix Kimi-VL 2D encoder grids — 问题修复，提升稳定性
• #32812 refactor(kimi-k3): put the gemm_ar kernels back in-tree and fold the PTX headers into their consumers — 架构/模块重构，提升可维护性
• #31869 [PD+PP] Honor PP consensus for bootstrap and prealloc — 架构重构/工程演进
• #32709 [Refactor] Remove dead allocator backup_state / restore_state — 架构/模块重构，提升可维护性
Python DSL / Compiler:
• #3178 Move Triton launcher into helion/runtime/triton/launcher.py — Triton / Python DSL 内核能力更新
• #2790 fix: refine TCGEN05 architecture guards — 问题修复，提升稳定性
• #2810 [Refactor] Remove unused tilelang.common package — 架构/模块重构，提升可维护性
• #3415 Remove stale SM100 grouped scheduler helper — 调度器路径优化
8. 多模态/Diffusion
vLLM:
• #42436 fused_moe: add VLLM_TRITON_USE_TD tensor-descriptor path — MoE 推理能力优化
• #50241 [CI][Test] Fix pooling truncation test after VLLMError hierarchy change — 问题修复，提升稳定性
• #49066 [docs] Add documentation for pynvvideocodec video decoding backend — 视频/多模态场景支持优化
SGLang:
• #31538 [diffusion] support resident layers for DiT — 扩散模型推理支持/优化
• #32318 Fix FlashInfer MNNVL workspace size check — 问题修复，提升稳定性
• #31361 [diffusion] fix: don't self-kill diffusion worker when PID 1 is the real parent — 扩散模型推理支持/优化
• #32696 feat(diffusion): add regional torch compile — 扩散模型推理支持/优化
9. Bug/稳定性修复 (Bugfix & Stability)
vLLM:
• #41357 [Bugfix] Prevent stale multiproc RPC deadlines from becoming unbounded waits — 关键路径问题修复
• #48245 [BugFix] Fix num_output_placeholders preemption underflow — 关键路径问题修复
Python DSL / Compiler:
• #2800 [BugFix] Reject floating-point predicates in vote intrinsics — 关键路径问题修复
• #2801 [BugFix] Preserve alloc_var initializer dtype — 关键路径问题修复
• #2802 [BugFix] Reject mixed packed x2 operand dtypes — 关键路径问题修复
• #2803 [BugFix] Skip descriptor TMA for device-bound copy bases — 关键路径问题修复
SGLang:
• #32540 fix(reasoning): honor Poolside template thinking defaults — 问题修复，提升稳定性
• #32757 Fix Kimi K3 reasoning leak in the Responses API — 问题修复，提升稳定性
10. CI/测试/基础设施
SGLang:
• #32242 sglang rust server request message — CI/测试/基础设施
• #27089 Disable extra NCCL CUDA event synchronization with symm mem — CI/测试/基础设施
• #28691 Add LFM2.5 embedding model support — CI/测试/基础设施
• #30553 [2/N] elastic-ep: Enable EPLB after scale-up — CI/测试/基础设施
Python DSL / Compiler:
• #11054 Add Windows ARM64 LLVM Build Support — CI/测试/基础设施
• #3140 [cutedsl] Make dtype subtests xdist-safe — CI/测试/基础设施
vLLM:
• #49974 [Test] dynamic_shapes_compilation — CI/测试/基础设施
• #50132 [CI] Add comment-based Buildkite triggers — CI / 测试基础设施调整
• #50211 [CI] Allow PR comment acknowledgements — CI / 测试基础设施调整
• #50197 [CI] Allow comment-triggered builds past pipeline filters — CI / 测试基础设施调整
PyTorch:
• #4644 Split PT2E tests into dedicated CPU and GPU CI workflows — CI / 测试基础设施调整
• #4638 Pin CI to PyTorch 2.13 for release; drop unused torchaudio — CI / 测试基础设施调整
• #189318 Bump pip from 26.0.1 to 26.1.2 in /.ci/docker — CI / 测试基础设施调整
11. 其他值得关注
Python DSL / Compiler:
• #2785 [CUDA] Support arbitrary TMEM layouts — 其他值得关注
• #3143 [cutedsl] Constrain FA4 resource use during autotuning — 其他值得关注
• #3139 [cutedsl] Clean up isolated autotune workers — 其他值得关注
• #3145 [cutedsl] Simplify FA4 role and epilogue lifecycle — 其他值得关注
vLLM:
• #48791 [ModelRunner V2] Enable sequence pooling for embedding and classification models — 其他值得关注
• #49582 [EC Connector] Add has_pending_push_work — 其他值得关注
SGLang:
• #32343 sglang rust server sampling message — 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 5
Python DSL / Compiler TOP 5
四、趋势分析
1. 活跃度对比
• vLLM 合入 54 个 PR。
• SGLang 合入 73 个 PR。
• PyTorch 合入 6 个 PR。
• Python DSL / Compiler 合入 41 个 PR。
2. 技术趋势
多硬件适配： 当日共有 34 个相关 PR，涉及 pytorch/ao、pytorch/helion、sgl-project/sglang、triton-lang/triton、vllm-project/vllm。
• sgl-project/sglang#32642 Add a benchmark script for the HPC-Ops bf16xfp32 router GEMM (https://github.com/sgl-project/sglang/pull/32642)
• sgl-project/sglang#31563 fix mqa preshuffle layout issue for deepseek v4 (https://github.com/sgl-project/sglang/pull/31563)
多模态/扩散： 当日共有 26 个相关 PR，涉及 pytorch/helion、sgl-project/sglang、vllm-project/vllm。
• vllm-project/vllm#42436 fused_moe: add VLLM_TRITON_USE_TD tensor-descriptor path (https://github.com/vllm-project/vllm/pull/42436)
• sgl-project/sglang#32118 Fix nightly CI: NVFP4 cuda-graph crash, NVILA batching, CuTe paged-KV zero-size, Kimi-VL OOM (https://github.com/sgl-project/sglang/pull/32118)
量化技术： 当日共有 20 个相关 PR，涉及 pytorch/ao、pytorch/helion、sgl-project/sglang、vllm-project/vllm。
• sgl-project/sglang#32118 Fix nightly CI: NVFP4 cuda-graph crash, NVILA batching, CuTe paged-KV zero-size, Kimi-VL OOM (https://github.com/sgl-project/sglang/pull/32118)
• vllm-project/vllm#43229 [CompressedTensors] FP4 Qutlass Integration (https://github.com/vllm-project/vllm/pull/43229)
Python DSL/编译： 当日共有 15 个相关 PR，涉及 pytorch/helion、tile-ai/tilelang、vllm-project/vllm。
• vllm-project/vllm#42436 fused_moe: add VLLM_TRITON_USE_TD tensor-descriptor path (https://github.com/vllm-project/vllm/pull/42436)
• vllm-project/vllm#49580 Integrate CuTeDSL MoE for ReLU2 NVFP4 (https://github.com/vllm-project/vllm/pull/49580)
分布式/路由： 当日共有 13 个相关 PR，涉及 sgl-project/sglang、vllm-project/vllm。
• sgl-project/sglang#32642 Add a benchmark script for the HPC-Ops bf16xfp32 router GEMM (https://github.com/sgl-project/sglang/pull/32642)
• sgl-project/sglang#32104 [EPD][VLM] Fix Kimi-VL 2D encoder grids (https://github.com/sgl-project/sglang/pull/32104)
MoE优化： 当日共有 11 个相关 PR，涉及 sgl-project/sglang、triton-lang/triton、vllm-project/vllm。
• vllm-project/vllm#42436 fused_moe: add VLLM_TRITON_USE_TD tensor-descriptor path (https://github.com/vllm-project/vllm/pull/42436)
• sgl-project/sglang#32699 perf(kimi-k3): fuse MoE-front prep (route + trtllm pack + mxfp8 quant) into one launch (https://github.com/sgl-project/sglang/pull/32699)
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。
• 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。
• Python DSL / 编译栈：Triton、Helion、torch.compile 方向活跃，说明高性能内核开发门槛正在下降，适合持续观察可复用 kernel 模式。
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！👋
