Update vLLM to v0.27.1

[swjeong9]

vLLM v0.26.0 → v0.27.1 업데이트 보고서

upstream(vllm-project/vllm)의 v0.27.1 태그를 main(v0.26.0 기반)에 머지한 브랜치입니다.

개요

항목	내용
기준 버전	v0.26.0 → v0.27.1 (v0.27.0 + 패치 1건)
릴리스 일자	v0.27.0: 2026-08-10, v0.27.1: 2026-08-11
커밋 수	577 커밋 (기여자 242명)
변경 규모	전체 1,650개 파일 (+138,056 / −25,921), vllm/ 코어만 776개 파일 (+65,208 / −13,952)
충돌 해소	1건 (vllm/v1/attention/ops/triton_merge_attn_states.py, 아래 상세)

⚠️ 환경 관련 Breaking Change

• PyTorch 2.13.0 / torchvision 0.28.0 / Triton 3.7.1로 업그레이드 ([Core] Update PyTorch to 2.13.0, torchvision to 0.28.0, triton to 3.7.1 vllm-project/vllm#48155). XPU([XPU] upgrade to torch 2.13 vllm-project/vllm#48677), CPU(Update torch version to 2.13.0+cpu vllm-project/vllm#50412)도 torch 2.13으로 통일. 클라우드 인스턴스의 venv를 재구성해야 하므로, 인스턴스 프로비저닝/셋업 스크립트의 의존성 설치 부분 점검이 필요합니다.
• 기타 의존성: Transformers 5.14.1(Bump Transformers version to 5.14.1 vllm-project/vllm#49223), FlashInfer 0.6.16.post3(Bump Flashinfer version to 0.6.16.post3 vllm-project/vllm#50892), NCCL 2.30.7(공식 이미지에서 DeepEPv2 활성화, [WideEP] Update NCCL to 2.30.7 to enable DeepEPv2 in the vllm/vllm-openai image vllm-project/vllm#45321), AITER 0.1.19([ROCm]: bump AITER to 0.1.19 vllm-project/vllm#49361).
• 제거된 기능: max_num_partial_prefills / max_long_partial_prefills 인자 제거([Misc] Remove old now unsupported max_num_partial_prefills and max_long_partial_prefills vllm-project/vllm#49244), Plamo2([Model] Remove Plamo2 vllm-project/vllm#49729)·Ouro([Model] Remove Ouro vllm-project/vllm#49786) 모델 제거.

본 프로젝트(P/D Disaggregation·MoE/EP 연구) 관련 핵심 변경

P/D Disaggregation

• NIXL P/D가 하이브리드(MLA+SSM) 모델 지원 ([KV Connector] Support NIXL P/D for hybrid MLA+SSM models  vllm-project/vllm#49762) 및 하이브리드 모델의 이종(heterogeneous) P/D 블록 사이즈 지원 ([KV Connector] Support NIXL heterogeneous P/D block sizes for hybrid models vllm-project/vllm#49612).
• MoRIIO 이종 TP↔DP prefill/decode read 라우팅 ([Core][KV-transfer] MoRIIO: heterogeneous TP<->DP prefill/decode read routing vllm-project/vllm#46116).
• Decode 측 lookup 비활성화 옵션((feat): optionally disable lookup on PD decode vllm-project/vllm#50498), decode chat 경로에서 prefill 토큰 ID 재사용([Frontend] Reuse prefill token ids on the decode chat path for disaggregated serving vllm-project/vllm#48145), NixlPush의 D→P 핸드셰이크 단계 축소([PD][NixlPush] Skip extra add_remote_agent step in D->P handshake vllm-project/vllm#49345).
• 버그픽스: P/D preemption race([BugFix] Fix P/D preemption race condition vllm-project/vllm#50297), KV lease deadline을 worker clock 기준으로 재계산([PD][Bugfix] Rebase KV lease deadlines onto worker clock vllm-project/vllm#50326), NIXL 하이브리드 MLA+mamba 이종 TP([PD][Bugfix] Fix NIXL hybrid MLA+mamba heterogeneous TP vllm-project/vllm#49297), internal LB 로드밸런싱([Core] Fix internal LB load-balancing vllm-project/vllm#49204).
• Mooncake: KV load 경로의 prepare_value 벡터화([Perf][KVConnector][Mooncake] Vectorize prepare_value on the KV load path vllm-project/vllm#48531).

대규모 서빙 / EP·DP

• DP+EP external LB 배포용 fault tolerance 프레임워크 신규 도입 ([Feature] Add fault tolerance framework (simplified) for DP+EP external LB deployments vllm-project/vllm#44428).
• Elastic EP: 스케일링 async 준비 단계([Elastic EP] Async preparation vllm-project/vllm#47288), non-contiguous weight 전송 픽스([Elastic EP] Fix non-contiguous weight transfers vllm-project/vllm#50641).

MoE 성능/구조

• FusedMoE → FusedMoEFactory 리팩토링 ([MoE Refactor] Rename FusedMoE to FusedMoEFactory vllm-project/vllm#44941) — MoE 레이어를 직접 계측하는 프로파일링 코드는 시그니처 변경 여부 확인 필요.
• MoE reduce_scatter 회귀 수정으로 E2E 처리량 5% 회복([Perf] Fix moe reduce_scatter perf regression by removing additional comm, 5% E2E throughput gain back. vllm-project/vllm#48763), bias 없는 non-grouped topk 라우팅의 fused path 디스패치(perf: dispatch non-grouped bias-less topk routing methods to fused path vllm-project/vllm#49618), LL BF16 router GEMM 튜닝([Perf] Tune LL BF16 Router GEMM vllm-project/vllm#48774), VLLM_TRITON_USE_TD로 Triton tensor-descriptor fused MoE 경로 제공(fused_moe: add VLLM_TRITON_USE_TD tensor-descriptor path vllm-project/vllm#42436), topk에서 cudagraph/DP 패딩 스킵(skip cudagraph/DP padding in topk vllm-project/vllm#48979).

KV Offloading

• 범용 P2P secondary tier(peer lookup/serving, [KVOffload][P2P] Generic P2P secondary tier: peer lookup and serving via ParentManager vllm-project/vllm#48021), per-request tier 필터링(TierFilter/TierMatcher, [KV Offloading] Per-request tier filtering with TierFilter/TierMatcher vllm-project/vllm#48123), TieringOffloadingSpec 기반 self-describing KV 이벤트([KV Offload] Support self-describing KV events with TieringOffloadingSpec vllm-project/vllm#48679), CachePolicyFactory로 eviction 정책 플러그인화(Add CachePolicyFactory for pluggable/external eviction policies vllm-project/vllm#49114).
• MLA KV의 shared CPU region 중복 제거([KV Offload] Deduplicate replicated MLA KV in the shared CPU region vllm-project/vllm#48906), CPUOffloadingSpec single-copy MLA 레이아웃([KV Offload] Enable single-copy MLA layout for CPUOffloadingSpec vllm-project/vllm#50301), 병렬화 구성에 독립적인 per-layer canonical KV 페이지 매핑([KV Connector] Add per-layer canonical KV page mappings for parallelism-agnostic offload vllm-project/vllm#48408).

일반 주요 변경 (Highlights)

모델 지원

• Kimi K3 풀스택 지원: 모델/커널([Model] Add Kimi K3 support: model files and kernels [1/N] vllm-project/vllm#50089, [New model] Kimi K3 vllm-project/vllm#50000), Python([Model] Add Kimi K3 support: Python frontend [2/2] vllm-project/vllm#50093)·Rust([Model] Add Kimi K3 support: Rust frontend [1/2] vllm-project/vllm#50104) 프론트엔드, AttnRes 커널([Kimi-K3] Add AttnRes kernels vllm-project/vllm#50090), DeepGEMM([Kimi K3 Bug] Fix deepgemm support for kimi k3 vllm-project/vllm#50458), compressed-tensors 양자화 체크포인트([Compressed-Tensors] Support Kimi-K3 quantized models vllm-project/vllm#50500), shared expert 샤딩 옵션([Kimi-K3] Add option to shard the shared expert instead of replicating vllm-project/vllm#50656).
• 신규 모델: Qwen3.5 dense·MoE([Model] Support Qwen3.5 text-only dense and MoE models vllm-project/vllm#50210), K-EXAONE-2.0-750B-A37B([Model] Add K-EXAONE-2.0-750B-A37B vllm-project/vllm#50524), VaultGemma([Model] Add VaultGemma via Transformers modeling backend vllm-project/vllm#49803), jina-embeddings-v5-text-nano([Model] Support jina-embeddings-v5-text-nano (EuroBERT encoder backbone) vllm-project/vllm#50688).

엔진 코어

• JIT warmup 인프라 신설([1/N][Feat][Perf] Add new warmup infrastructure for JITs vllm-project/vllm#47451) + runner 소유 Triton 커널의 사전 warmup([Core] Warm up runner-owned Triton kernels before the first request vllm-project/vllm#49903)으로 첫 요청 컴파일 지연 제거.
• FlashAttention 4 (SM100): FP8 KV cache([Attention] FlashAttention 4 SM100 FP8 kv cache support vllm-project/vllm#42569), headdim-256([Attention] Integrate FlashAttention 4 SM100 headdim 256 support vllm-project/vllm#42669).
• merge-attention context count의 런타임 인자화([Perf] Make merge attention context count a runtime argument vllm-project/vllm#48739) — 이번 머지의 충돌 지점.
• Model Runner V2 확장: encoder-only attention([ModelRunner V2] Support encoder-only attention vllm-project/vllm#49331), embedding/classification용 sequence pooling([ModelRunner V2] Enable sequence pooling for embedding and classification models vllm-project/vllm#48791), CPU 멀티모달([Bugfix] Fix multi-modal support on CPU MRV2 vllm-project/vllm#50073), multi-layer MTP speculator([Model Runner V2][Spec Decode] Add multi-layer MTP speculator vllm-project/vllm#48892), PCP의 MRV2 선택([Core][PCP] Select MRV2 when PCP is enabled vllm-project/vllm#50034).
• DeepSeek-V4 성능: sequence parallelism([DSV4] Implement Sequence Parallelism vllm-project/vllm#46789), 빈 c128 런치 스킵으로 커널 ~2x([DSv4 Perf] Skip empty c128 kernel launch, around 2x kernel performance improvement. vllm-project/vllm#48957), TTFT 3.4%([DSv4 Perf] Skip topk and router when not needed, 3.4% E2E TTFT improvement for Decode case vllm-project/vllm#49486)·3.9%([DSv4 Perf] Optimize workspace reuse for eager break vllm-project/vllm#49236) 개선, PP 버퍼 448 MiB 절감([DSv4 Perf] Fix redundant memory allocation and copy for dsv4 pp buffer, 448 MiB GPU memory saved vllm-project/vllm#50312) 등.

하드웨어

• NVIDIA Rubin(sm_107) 타깃 추가(Add sm_107 for Rubin vllm-project/vllm#49387)와 SM107 NVLink all-reduce([Rubin] Enable NVLink all-reduce paths on SM107 vllm-project/vllm#49647). ROCm gfx1250 활성화(Enable gfx1250 ROCm architecture vllm-project/vllm#46516), XPU QK Norm+RoPE fusion([XPU] Enable QK Norm + RoPE fusion pass on XPU vllm-project/vllm#49394), Arm CPU INT8 fused MoE 커널([CPU][Perf] INT8 Fused MoE Kernel for Arm CPUs vllm-project/vllm#48637).

API / 프론트엔드

• Rust 프론트엔드에 gRPC 컨트롤 플레인 추가: 헬스 리포팅([Rust Frontend][gRPC] Add engine-aware health reporting vllm-project/vllm#48992), abort([Rust Frontend][gRPC] Add abort control RPC vllm-project/vllm#49255), 서버/모델 디스커버리([Rust Frontend][gRPC] Add server and model discovery vllm-project/vllm#49491), KV 이벤트 소스 디스커버리([Rust Frontend][gRPC] Add KV event source discovery vllm-project/vllm#50033). vllm-bench의 vllm CLI 통합([Rust][Benchmark] Integrate vllm-bench to vllm-rs & vllm CLI vllm-project/vllm#48930).
• VLLMError 계층으로 요청 에러 처리 표준화([Frontend][Core] Standardize request error handling with VLLMError hierarchy vllm-project/vllm#49665), Cohere chat v2 API([Frontend] Cohere chat v2 api support vllm-project/vllm#47189), Anthropic Messages API cache_salt([Frontend] Add cache_salt support to Anthropic Messages API vllm-project/vllm#49498).

양자화

• FP4 Qutlass(compressed-tensors) 통합([CompressedTensors] FP4 Qutlass Integration vllm-project/vllm#43229), CuTeDSL MoE ReLU2 NVFP4(Integrate CuTeDSL MoE for ReLU2 NVFP4 vllm-project/vllm#49580), TurboQuant KV quant 모드([Bugfix][TurboQuant] Add KV quant mode for turboquant  vllm-project/vllm#50533), DeepSeek-V4([CompressedTensors] DeepSeek4 CT Quantization Support vllm-project/vllm#41276)·Kimi-K3([Compressed-Tensors] Support Kimi-K3 quantized models vllm-project/vllm#50500) compressed-tensors 체크포인트.

v0.27.1 패치 내용

• v0.27.0 위의 패치 릴리스 1건: quantized DSpark Markov head 지원 (Support quantized DSpark Markov heads vllm-project/vllm#50424).

머지 충돌 해소 상세

vllm/v1/attention/ops/triton_merge_attn_states.py 1건:

• 원인: main에 cherry-pick되어 있던 MLA 빈 컨텍스트 청크 버그픽스([Bugfix][Attention] Ignore empty MLA context chunks during merge vllm-project/vllm#49294)와 upstream의 merge-attention stride 리팩토링([Perf] Make merge attention context count a runtime argument vllm-project/vllm#48739, output_lse를 head/token stride 인자로 접근)이 같은 tl.store 라인에서 충돌.
• 해소: 해당 버그픽스는 v0.27.1에 이미 포함되어 있으므로(upstream 060b5f61d) upstream(stride 파라미터) 쪽을 채택. 머지 결과 파일 전체가 stride 방식 시그니처를 사용함을 확인.

---

🤖 Generated with Claude Code