v0.23.0

What's Changed

• fix(release): apply manylinux tag before PyPI upload (795396a3a) by AlpinDale
• fix(release): attach all wheels to GitHub releases (#1749) (ba857e9b9) by @AlpinDale
• fix(build): package native artifacts in CUDA wheels (#1748) (d23fda737) by @AlpinDale
• fix(moe): support Kimi K3 DeepGEMM gather shape (#1747) (a4a09745f) by @AlpinDale
• fix(ci): allow execution in ROCm build cache (6586b2889) by AlpinDale
• fix(ci): isolate wheel builder storage (a646aa2c4) by AlpinDale
• fix(ci): pin ROCm backfill build driver (c1279d75a) by AlpinDale
• fix(ci): persist incremental ROCm build state (4ac0eb0bb) by AlpinDale
• docs: correct installer platform list (2be97a6fc) by AlpinDale
• docs: list automatic installer platforms (4f8f6aa52) by AlpinDale
• docs: update stale vLLM references (d41a3ee16) by AlpinDale
• refactor: remove vLLM compatibility names (aae7ceb48) by AlpinDale
• fix(sync): complete repository namespace port (db6c97167) by AlpinDale
• [sync] [ROCm]Migrating Deepseek V3.2 to vllm/models/deepseek_v32/ (#47207) (1a6353405) by AlpinDale
• [sync] [MyPy][1/N] Fix mypy errors in some tests/ directories and enforce follow-imports=silent (#49570) (e1f8de1f6) by AlpinDale
• [sync] docs(security): document Ray cluster trust model and env var propagation (#50397) (6dbfb4973) by AlpinDale
• [sync] [CPU] Bump up CPU kernels to latest version (#50387) (da866b35c) by AlpinDale
• [sync] [New model] Kimi K3 (#50000) (9c06da9f0) by AlpinDale
• [sync] [Test][ROCm] Account for gfx950 FP8 RMSNorm rounding (#49839) (898ae52a7) by AlpinDale
• [sync] [CI] Stabilize speculator memory teardown (#50284) (d01367531) by AlpinDale
• [sync] [CI][ROCm] Stabilize LLM GC teardown check (#50340) (eeb1d2f8c) by AlpinDale
• [sync] [Frontend] Lazily initialize chat media connectors (#49914) (c1c9538a9) by AlpinDale
• [sync] [ROCm] Add AITER FP8 ViT encoder attention (#49937) (a28f50621) by AlpinDale
• [sync] advance vllm-sha to 0028fc8d8 (580346f91) by AlpinDale
• [sync] Add FlashMLA H100 tests to CI, fix them after #32810 (#50322) (aa9483902) by AlpinDale
• fix(sync): remove misplaced derender test duplicate (ff2debcd5) by AlpinDale
• [sync] [ROCm] [CI] Support cached K/V (key/value=None) in Triton prefix-prefill (#48257) (29c4e8aab) by AlpinDale
• [sync] [CI/Build] Limit wheel size check to CUDA 13 (#50357) (9a494a4eb) by AlpinDale
• [sync] [Quantization][Autoround][XPU] Add W4A16(moe) / MXFP4(linear/moe) Support (#47124) (439a390a9) by AlpinDale
• [sync] [PD][Bugfix] Rebase KV lease deadlines onto worker clock (#50326) (7fd2e5aeb) by AlpinDale
• [sync] [BugFix] Fix P/D preemption race condition (#50297) (4e137fb6c) by AlpinDale
• [sync] [Frontend] Add detokenization streaming derender for disaggregated serving (#47301) (45c38624d) by AlpinDale
• [sync] [Bugfix][Kernel] Fix integer overflow in libtorch_stable/activation_kernels.cu (#49660) (6555c10c5) by AlpinDale
• [sync] [Bugfix] Prevent stale multiproc RPC deadlines from becoming unbounded waits (#41357) (53e82a623) by AlpinDale
• [sync] [Frontend] Add diarized_json support for MOSS-Transcribe-Diarize (#48543) (1160c6ba3) by AlpinDale
• [sync] advance vllm-sha to 1cb3fe584 (b1fbc6776) by AlpinDale
• [sync] [CI][ROCm] Fix AMD nightly distributed regressions (#50304) (62a2328bd) by AlpinDale
• [sync] [ROCm][CI] Avoid Ray worker startup env race (#50311) (8bd0dd537) by AlpinDale
• [sync] [Rubin] Enable NVLink all-reduce paths on SM107 (#49647) (57f3ed08f) by AlpinDale
• [sync] feat(grpc): add KV event source discovery (#50033) (d531ff193) by AlpinDale
• [sync] Revert "[Misc][Minimax-M3]add default video_processor (#50092)" (#50313) (cd539bbd9) by AlpinDale
• [sync] [Perf] RMSNorm uncontiguous support, 1.2~3.1x kernel performance improvement (#49750) (2342d461f) by AlpinDale
• [sync] [torch.compile] Compile CustomOp.forward_native for ReLU^2 to avoid raw torch ops inside opaque custom ops (#50244) (2faf40808) by AlpinDale
• [sync] [docs] Add documentation for pynvvideocodec video decoding backend (#49066) (ad80b5774) by AlpinDale
• [sync] fix(step3p5-mtp): honor exclude_modules for the MTP head via prefix (#48883) (62fa777e9) by AlpinDale
• [sync] [KV Connector] Fix NIXL mamba state pairing for multi-slot block tables (#50153) (ad243ca1f) by AlpinDale
• [sync] [ROCm][CI] Fix Kimi K3 KDA on ROCm (#50262) (8fa6ca6e5) by AlpinDale
• [sync] [Spec Decode][Perf] Replicate DSpark Markov head across TP ranks (#49731) (5d0fd48c0) by AlpinDale
• [sync] [BugFix] Fix num_output_placeholders preemption underflow (#48245) (b18d7e7f4) by AlpinDale
• [sync] [ModelRunner V2] Enable sequence pooling for embedding and classification models (#48791) (fead40caa) by AlpinDale
• [sync] [Bugfix][Frontend] Return transcription and translation verbose as float (#49073) (31f0d668f) by AlpinDale
• [sync] [Rust Frontend] Send multimodal tensors in auxiliary frames (#49341) (9bdac0f02) by AlpinDale
• [sync] [CI][Test] Fix pooling truncation test after VLLMError hierarchy change (#50241) (c458c4bee) by AlpinDale
• [sync] [XPU] Route weightless RMSNorm to _C dispatch (#47121) (fd14c4385) by AlpinDale
• [sync] [Kernel][Mamba] Fused-kernel support for align-mode DS-conv state migration with num_accepted_tokens > 1 (#49291) (2de5895a2) by AlpinDale
• [sync] [Bugfix][Multimodal] Include media IO config in MM cache hash (#49975) (da89224cb) by AlpinDale
• [sync] [KV Offload] Move CPUOffloadingSpec onto SharedOffloadRegion (#50094) (8fce75345) by AlpinDale
• fix(installer): support SM103 GPUs (#1743) (63b2c2066) by @AlpinDale
• fix(ci): persist CPU wheel build caches (#1742) (c4c630d7e) by @AlpinDale
• docs: update site branding and installation (334428ea0) by AlpinDale
• fix(ci): restore tracked files between wheel builds (4daf03219) by AlpinDale
• fix(ci): pin Metal build driver during reconciliation (f7b54aab3) by AlpinDale
• fix(ci): reuse stable Metal build environment (6d174669c) by AlpinDale
• fix(ci): refresh Metal CMake cache per wheel (0c2239ce7) by AlpinDale
• fix(ci): reconcile wheels after rewritten history (dd349a2d3) by AlpinDale
• fix(ci): avoid stale Metal CMake cache paths (ec14ea3dc) by AlpinDale
• fix(sync): complete Kimi K3 namespace port (695748f09) by AlpinDale
• fix(config): type video pruning init variable (40b8e7d98) by AlpinDale
• fix(sync): reject patch backup artifacts (ed1f54226) by AlpinDale
• [sync] advance vllm-sha to 9a4e5f953 (4a5f02703) by AlpinDale
• [sync] [Rust Frontend] Add --limit-mm-per-prompt support (#49604) (523db06e4) by AlpinDale
• [sync] [CI] Fix MXFP8 MOE backend selection tests on gfx942 (#50222) (7352def52) by AlpinDale
• [sync] [EC Connector] Add has_pending_push_work (#49582) (c8a03dd11) by AlpinDale
• [sync] [XPU] upgrade to torch 2.13 (#48677) (4bc18d837) by AlpinDale
• [sync] [Misc][Minimax-M3]add default video_processor (#50092) (ffea2fdcd) by AlpinDale
• [sync] [BugFix] eagle draft max position embeddings (#49343) (cc5ccd358) by AlpinDale
• [sync] [Model] Support Qwen3.5 text-only dense and MoE models (#50210) (2c4dd3008) by AlpinDale
• [sync] [Model] Add Kimi K3 support: Python frontend [2/2] (#50093) (ab33c6072) by AlpinDale
• [sync] [Frontend] Reuse prefill token ids on the decode chat path for disaggregated serving (#48145) (587231b7e) by AlpinDale
• [sync] [CPU] Fix FP8 attention scratchpad sizing (#50194) (ddb77ce83) by AlpinDale
• [sync] [XPU][UT][CI] add xpu config to run gpt-oss accuracy in ut and ci (#48703) (b46dbbd16) by AlpinDale
• [sync] [Model] Add Kimi K3 support: model files and kernels [1/N] (#50089) (4fe06dc85) by AlpinDale
• [sync] fused_moe: add VLLM_TRITON_USE_TD tensor-descriptor path (#42436) (abe9ba0eb) by AlpinDale
• [sync] [Test] dynamic_shapes_compilation (#49974) (e4c40830a) by AlpinDale
• [sync] [CPU] Fix s390x builds and update torch version in dockerfile (#50144) (acb4de95a) by AlpinDale
• [sync] Add CachePolicyFactory for pluggable/external eviction policies (#49114) (309170976) by AlpinDale
• [sync] [CI][ROCm] Stabilize Qwen2-VL LoRA test (#50161) (966ff6a06) by AlpinDale
• [sync] [Frontend][Core] Standardize request error handling with VLLMError hierarchy (#49665) (75ee2854e) by AlpinDale
• [sync] [KV Connector] Support NIXL P/D for hybrid MLA+SSM models (#49762) (714fedca7) by AlpinDale
• [sync] Integrate CuTeDSL MoE for ReLU2 NVFP4 (#49580) (cc6d10852) by AlpinDale
• [sync] [CompressedTensors] FP4 Qutlass Integration (#43229) (d7ec31ff1) by AlpinDale
• [sync] [ROCm][CI] Stabilize ngram and suffix correctness test (#50190) (cbc49c305) by AlpinDale
• [sync] [Model] Add Inkling compressed-tensors dynamic FP8 support (#48876) (b45b6a00c) by AlpinDale
• [sync] [BugFix] Stop dummy runs from writing mamba state through stale block-table rows (#49757) (4baa4ea3c) by AlpinDale
• [sync] [Bugfix] Fix /wake_up crash on hybrid models (Mamba/DeltaNet) (#41602) (e1ee37213) by AlpinDale
• [sync] [CI][NIXL] Fix flaky DP+EP test port conflict (#50171) (ed466b35e) by AlpinDale
• [sync] [ROCm][CI] Stabilize ROCm audio streaming test (#50163) (5f29c0afe) by AlpinDale
• [sync] [MXFP8][ROCm] Fix MXFP8 MoE backend selection (#49747) (34ec18a15) by AlpinDale
• [sync] [Test] Make EPD correctness tests configurable for XPU (#50110) (e563cd347) by AlpinDale
• [sync] [Perf] Zero-copy torch.Tensor pickling in shm_broadcast MessageQueue (#48442) (a2b8c0146) by AlpinDale
• [sync] perf: dispatch non-grouped bias-less topk routing methods to fused path (#49618) (1ca7686ed) by AlpinDale
• [sync] [Model] Add Kimi K3 support: Rust frontend [1/2] (#50104) (ba83d2f57) by AlpinDale
• [sync] [Bugfix][Spec Decode] Size DFlash query buffers for cudagraph-padded batches (#50065) (acebb7d21) by AlpinDale
• [sync] [Rust Frontend] Extract shared tracing setup logic into vllm-tracing (#50129) (88cf45f91) by AlpinDale
• [sync] [BugFix] Fix clang spinloop mwaitx include (#45532) (832d96896) by AlpinDale
• [sync] [ROCm][Bugfix] Sanitize AITER paged-MQA logits before sparse top-k for DeepSeek-V4 (#49714) (0a4d537d8) by AlpinDale
• [sync] [Bugfix][Multimodal] Fix video temporal padding estimates (#49030) (710b062bf) by AlpinDale
• feat(ci): backfill platform wheels per commit (#1734) (8366cd407) by @AlpinDale
• feat(packaging): add universal platform installer (#1733) (afa79e0f3) by @AlpinDale
• docs: add Rust frontend deployment guide (#1732) (e788d8540) by @AlpinDale
• docs: overhaul user documentation (#1730) (0b37108b7) by @AlpinDale
• ci: publish tagged release wheels (#1729) (5bd5d3d9c) by @AlpinDale
• fix(loader): log model download status (#1728) (bcc6f2a75) by @AlpinDale
• fix(build): parse decorated wheel index links (e9076dffe) by AlpinDale
• feat(ci): improve nightly wheel index (#1727) (1d0fc1035) by @AlpinDale
• feat(build): download commit-matched precompiled wheels (#1726) (758db6487) by @AlpinDale
• test(kernels): use unified Minimax KV cache layout (b0d00c88b) by AlpinDale
• fix(kv-cache): support DCP with hybrid cache groups (29c336edc) by AlpinDale
• [sync] [compressed-tensors] update find_matched_target order to prioritize fused name matches over class match (#49483) (5994cb617) by AlpinDale
• [sync] [ROCm] Fix and optimize GPT-J-style MRoPE (#49906) (7f6c582bd) by AlpinDale
• [sync] [ROCm] Cache fp32 upcast of static e8m0 weight scale in AITER scaled_mm (#47773) (db0f3969a) by AlpinDale
• [sync] [Rust][Benchmark] Make vllm bench serve Rust delegation opt-in (#50081) (875ee74f9) by AlpinDale
• [sync] [Attention] Skip sparse indexer scoring for dense short prefills (#48407) (9389017cc) by AlpinDale
• [sync] [Bugfix] Don't reuse engine core payload buffer while zmq is sending it (#50053) (2e2bf043a) by AlpinDale
• [sync] [Bugfix][KV Offload][P2P] Scope serve state to fetch rounds (#49877) (1c4a30271) by AlpinDale
• [sync] [Core] Warm up runner-owned Triton kernels before the first request (#49903) (5d9b8d7e9) by AlpinDale
• [sync] [Bugfix][KV Offload] Keep Mamba block span unscaled under DCP (#49964) (2208cfa33) by AlpinDale
• [sync] [Bugfix][Kernel] Fix batch invariance in RMSNorm kernels by pinning block size (#48391) (88134bfb5) by AlpinDale
• [sync] [CI] Wire untethered test files into CI jobs (#49340) (ca72cfcfe) by AlpinDale
• [sync] [Bugfix] Add missing vllm/models/kimi_k3/__init__.py (#50131) (d6dc28353) by AlpinDale
• [sync] [KV Connector] Support NIXL heterogeneous P/D block sizes for hybrid models (#49612) (0924e23d6) by AlpinDale
• [sync] add epilogue hook to flex attention (#45841) (e0b2fdf40) by AlpinDale
• [sync] [PD][NixlPush] Skip extra add_remote_agent step in D->P handshake (#49345) (5e3e7ffde) by AlpinDale
• [sync] [Bugfix] Enhance extra_config handling for layer name suffix matching (#48589) (bb60ac998) by AlpinDale
• feat(elastic-ep): prepare scaling asynchronously (d70c8d15b) by AlpinDale
• [sync] [Build] Fix DeepEP CUDA driver stub linking (#50103) (082ddc1f2) by AlpinDale
• [sync] [Rust Frontend] Align sampling validation with Python (#47494) (269c65348) by AlpinDale
• [sync] [Rust Frontend] Fix finish reason for named tool choices (#49496) (a2b554613) by AlpinDale
• [sync] [CI] Increase Qwen3.5 MTP GSM8K generation length (#49881) (fc7e5d5de) by AlpinDale
• refactor(native): move fork kernels to stable ABI (#1725) (a901b6a3a) by @AlpinDale
• feat(ci): publish wheels for every main commit (5ea849f48) by AlpinDale
• fix(ci): bound retained wheel build layers (2787b2a99) by AlpinDale
• fix(ci): bound nightly wheel upload retries (1634627d5) by AlpinDale
• feat(docs): publish wheel index at /whl (46f0683b4) by AlpinDale
• fix(nightly): encode wheel URLs for Backblaze (88ba736a3) by AlpinDale
• ci(nightly): migrate wheel storage to Backblaze B2 (#1723) (67b7312f3) by @AlpinDale
• docs(tokenizer): restore upstream issue reference (6c8fe08e5) by AlpinDale
• fix(kv-offload): defer unsafe object tier shutdown cleanup (5573e2b8b) by AlpinDale
• fix(kv-offload): preserve object transfer results on retry (47f704cbf) by AlpinDale
• fix(moriio): normalize unknown remote tensor parallel size (f9dd50925) by AlpinDale
• fix(kv-offload): validate filesystem batch lengths (1ce0ffb90) by AlpinDale
• fix(kv-offload): support legacy file mapper configs (afdd582c9) by AlpinDale
• test(quantization): guard unsupported accelerator platforms (bd18237b4) by AlpinDale
• test(kv-offload): normalize object tier fixtures (23ed98de3) by AlpinDale
• test(kv-offload): update filesystem tier config fixture (221a695c3) by AlpinDale
• Fix AITER MXFP4 fake op signature (903e39cf6) by AlpinDale
• Fix CUDA event shim for CPU workers (731a856d8) by AlpinDale
• Fix ABI audit for fork extension (44c52a21f) by AlpinDale
• [sync] [CI] Add PyTorch stable ABI audit check (#48164) (7b41ac1d2) by AlpinDale
• [sync] [Test] Skip ROCm AITER MLA prefill tests on non-ROCm platforms (#49945) (f9ee9b135) by AlpinDale
• [sync] [Rust Frontend] Add ordinary-text tokenizer encoding (#49992) (0eaa6c74b) by AlpinDale
• [sync] [Kimi-K3] Add AttnRes kernels (#50090) (dd792ba3a) by AlpinDale
• [sync] [Bugfix][Spec Decode] Preserve draft buffers across level-2 sleep (#49774) (e01ce3e11) by AlpinDale
• [sync] [Bugfix] Respect cgroup memory limits on all platforms (#49966) (819eff30a) by AlpinDale
• [sync] [Core][Frontend] Add weight version tagging for RL rollouts (#49040) (ae8273152) by AlpinDale
• [sync] [Bugfix] Fix multi-modal support on CPU MRV2 (#50073) (ab6fc7a71) by AlpinDale
• [sync] Remove triton per group quant [ROCm] [Bugfix] (#49621) (42f51b9f5) by AlpinDale
• [sync] [Build] Fix CUDA arch detection producing kernel-less builds on SM121 (#49904) (7ee1f11d2) by AlpinDale
• [sync] [Bugfix] Only pad transformers backend value when it is narrower (#50060) (fc68bf4df) by AlpinDale
• [sync] [Bugfix][KV Offload][OBJ] Preserve job completion during cleanup (#49947) (0d96a931b) by AlpinDale
• [sync] [KV Offload] Make compact secondary identity TP-independent (#49858) (a85e7d5e6) by AlpinDale
• [sync] [Bugfix] Fix DeepseekV4FP8 Quark MXFP4 crash on list-valued weight (#49634) (94f9a5b17) by AlpinDale
• [sync] [AMD] Revert Mxfp4MoeBackend.TRITON_UNFUSED fallback (#46491) (ab4c16643) by AlpinDale
• [sync] [KV-offload][FS] : Batch store/load_block in C (#49152) (648959851) by AlpinDale
• [sync] [Feature] Add VidCom2 video token pruning (#47750) (fdf8bf3f2) by AlpinDale
• [sync] [ROCm][Quark][6/N] Use MXFP4 linear kernel abstraction for aiter backend (#49348) (5515ab5a4) by AlpinDale
• [sync] [XPU] Add online fp8 quantization test (#44513) (a0527d2a1) by AlpinDale
• [sync] [ROCm][DSv3.2] Eliminate per-decode FillFunctor launches in sparse-MLA hot loop (#44527) (cacbe5249) by AlpinDale
• [sync] [ROCm][KVConnector][MoRI-IO] Fix WRITE-mode remote-TP rank collapse (#46332 follow-up) (#47764) (50a2c367f) by AlpinDale
• [sync] [Core][PCP] Select MRV2 when PCP is enabled (#50034) (07cd3b4a5) by AlpinDale
• [sync] [Tests][Spec Decode] Add gemma4 MTP acceptance rates test (#47920) (4befe2dad) by AlpinDale
• [sync] Fix Humming non-gated MoE (#49096) (53341d750) by AlpinDale
• [sync] Fix MQA with tensor parallelism on transformers modeling backend (#49987) (d3906760a) by AlpinDale
• [sync] [ROCm] [BugFix] Fix Quark GLM-5.2 Checkpoint inference: indexer wk per-channel FP8 dequant + missing sparse-MLA metadata fields (#48886) (b95b8740a) by AlpinDale
• [sync] [CI][ROCm] Fix test_ocp_mx_wikitext_correctness reference value (#49690) (d298c4d95) by AlpinDale
• [sync] [CI] Initialize DeepEP FP8 test weights (#49912) (3d0906125) by AlpinDale
• [sync] [ROCm][Quantization][5/N] Refactor quark_moe w8a8-int8 w/ oracle (#46765) (e8743a549) by AlpinDale
• [sync] [Refactor] Remove dead code in multiple files (#49745) (52159d136) by AlpinDale
• [sync] [DSv4 Perf] Adaptive topk width, 1.0% E2E throughput improvement (#50004) (32e1cda88) by AlpinDale
• [sync] [Core] Fail fast when /dev/shm is too small for the shm ring buffer (#48879) (0a130f4b3) by AlpinDale
• ci: retain historical nightly wheels (e377f66f7) by AlpinDale
• ci: validate documentation links (f7a52d70e) by AlpinDale
• test: align Mooncake cache pool fixtures (2783312a4) by AlpinDale
• fix: enable prefix caching for hybrid models (78140fa7e) by AlpinDale
• [sync] [Test] Regression test for hybrid-Mamba eagle cache-peek in Mooncake connector (#43559) (#48361) (68ed3e646) by AlpinDale
• [sync] [Bugfix][CPU] Fall back to torch for unaligned swigluoai on NEON/vec MoE (#49985) (1ab2f81b7) by AlpinDale
• [sync] Fix MLA padding and grouped topk routing in the Transformers modelling backend (#49982) (f2ff44fb3) by AlpinDale
• [sync] [Attention] Integrate FlashAttention 4 SM100 headdim 256 support (#42669) (1e3c55e28) by AlpinDale
• [sync] [MRV2] Always build attn metadata at capture time (#49364) (#49995) (cc8593ac7) by AlpinDale
• [sync] [Bugfix] Changed speech to text chunk timestamp to cumulative approach (#41131) (e48f80bd7) by AlpinDale
• [sync] [Rust Frontend][gRPC] Add server and model discovery (#49491) (340b78cf4) by AlpinDale
• [sync] [Misc][PD] Nixl cleanup get_backend_aware_kv_block_len and virtually_split_kv_in_blocks (#49988) (c8b097f45) by AlpinDale
• [sync] [Bugfix] Fix VLLM_ENFORCE_STRICT_TOOL_CALLING mutation in tests (#49846) (2e0aceee3) by AlpinDale
• [sync] [Perf] Hash videos by source bytes (#49607) (d850b1244) by AlpinDale
• [sync] [MRV2][Performance] Skip no-op FP32 logits materialization (#47711) (f9bb0bdbb) by AlpinDale
• [sync] [Bugfix] Restore truncate_prompt_tokens for Jina rerank/score online (#49963) (ba9231af5) by AlpinDale
• [sync] [Bugfix] Reject pipeline parallelism for DiffusionGemma (#45828) (96dac911b) by AlpinDale
• [sync] [Perf] Tune LL BF16 Router GEMM (#48774) (3129a6815) by AlpinDale
• [sync] [Core] Fix internal LB load-balancing (#49204) (094bc8823) by AlpinDale
• [sync] [Model] Enable EVS for Qwen3.5 (#48912) (90212a8b6) by AlpinDale
• [sync] [Perf] Make merge attention context count a runtime argument (#48739) (1d48268c5) by AlpinDale
• [sync] [Perf] Skip ll_bf16 router GEMM warmup for non-MoE models (#49659) (af28c8de3) by AlpinDale
• [sync] [Bugfix]Reject invalid FlashInfer MNNVL workspaces (#49043) (5d45147c3) by AlpinDale
• [sync] Improve Transformers modelling backend fx tracer (#49957) (b326953dc) by AlpinDale
• [sync] [KV Offloading] Per-request tier filtering with TierFilter/TierMatcher (#48123) (cf8d5d964) by AlpinDale
• [sync] [ROCm] Use backend-default dot precision for ReplaySSM (#49909) (b8b973b12) by AlpinDale
• [sync] [ROCm] [Model] Enable TML inkling (#48841) (d19cd0ab5) by AlpinDale
• [sync] [ROCm][CI] Skip three torchao tests of gfx950 until torchao==0.18 is released (#49732) (76c8f6784) by AlpinDale
• [sync] [Bugfix][ROCm] Use batch DMA for CPU KV cache loads (#49843) (12431ac96) by AlpinDale
• [sync] [communication] [bugfix] fix quickreduce acc error in cudagraph mode (#46913) (3b2b4f5a5) by AlpinDale
• [sync] [Bugfix][CPU] Zero-pad MoE intermediate size for grouped-gemm TP alignment (#49591) (ceb89422f) by AlpinDale
• [sync] [Tokenizer] Use HF config for HF tokenizers (#49907) (5d1b5a5ee) by AlpinDale
• [sync] [XPU][CI] Use platform device in InputBatch V2 test (#49939) (af4ec5cd1) by AlpinDale
• [sync] [Bugfix] Normalize sparse MLA warmup compression ratios (#49392) (f39eaf923) by AlpinDale
• [sync] [Rust Frontend] Keep --max-model-len engine-owned (#49944) (34ceff663) by AlpinDale
• [sync] [CI][ROCm] Reduce kernel test runtime (#49915) (1babe97dc) by AlpinDale
• [sync] [Quantization][INC]Add MXFP8 Linear Support (#47514) (eeabd02c6) by AlpinDale
• [sync] [3/N][Core][KV Connector] Support reliable partial-tail KV offload for sub-block prompts (#49502) (01a65ec07) by AlpinDale
• [sync] [CPU][Spec Decode] Optimize GDN conv path for speculative decoding (#48577) (02685e156) by AlpinDale
• [sync] [CPU][Perf] INT8 Fused MoE Kernel for Arm CPUs (#48637) (270791f42) by AlpinDale
• [sync] [Hardware][Power] Add FAST_EXP for Power (#49571) (aa871e6a0) by AlpinDale
• [sync] [CI] Explicitly tear down speculative decode runners (#49910) (23f2a8534) by AlpinDale
• [sync] [Bugfix][KV Offload][P2P] Fix EngineCore crash reconnecting to a reaped peer (#49823) (7f33df638) by AlpinDale
• [sync] [ROCm] Make vllm_c RMSNorm output contiguous (#49913) (faa8fca40) by AlpinDale
• [sync] [Bugfix] Fix mHC block-M prenorm GEMM cross-row reduction carry-over (#49429) (75753150d) by AlpinDale
• [sync] [Docs] Document NVFP4 GEMM kernel selection and Marlin weight-only fallback (#49376) (25226e967) by AlpinDale
• [sync] [Frontend] expose stream_interval as req sampling param (#49754) (2337c6bde) by AlpinDale
• [sync] [XPU] Enable QK Norm + RoPE fusion pass on XPU (#49394) (a7b3413e1) by AlpinDale
• [sync] [CI][ROCm] Keep global GPU memory cleanup opt-in (#49911) (18188bda2) by AlpinDale
• [sync] [CI][ROCm] Reduce V1 attention test runtime (#49916) (90b60afef) by AlpinDale
• [sync] [Core] Fix gpu<->cpu syncs in MRV2 mamba_hybrid.py (#49736) (94732ce7a) by AlpinDale
• [sync] [Core][KV-transfer] MoRIIO: heterogeneous TP<->DP prefill/decode read routing (#46116) (834a8d3e1) by AlpinDale
• [sync] [BugFix][MRV2] Don't create dummy requests longer than max_model_len (#49751) (d3f090ede) by AlpinDale
• [sync] [Bugfix] Prevent NaN poisoning in xpu_mla_sparse for fully-masked index chunks (#48366) (5b3dfc165) by AlpinDale
• [sync] [ModelRunner V2] Support encoder-only attention (#49331) (96c3d3820) by AlpinDale
• [sync] [Bugfix][CuMem] Make KV-cache wake cleanup tag-safe (#49857) (e36b5916c) by AlpinDale
• [sync] [Core][Distributed] Add process-checkpoint lifecycle hooks for communicators (starting with Flashinfer) (#46877) (43432e1c9) by AlpinDale
• [sync] [Bugfix][KV Offload] Namespace auto cache dtype by effective dtype (#49438) (b028c1f17) by AlpinDale
• [sync] [Bugfix] Fix handling 5D KV cache in kv_postprocess_layout_on_receive (#47791) (c564641e9) by AlpinDale
• [sync] [KV Offload] Fix num_tokens_after_batch for different termination types (#49285) (fd5c9ac92) by AlpinDale
• [sync] [Bugfix] Reject contradictory custom-op directives (#49134) (5192a9d00) by AlpinDale
• [sync] [Bugfix][KV Offload] Bound unaligned SWA loads by physical GPU blocks (#49052) (8bb452a7c) by AlpinDale
• [sync] [KVOffload][P2P] Generic P2P secondary tier: peer lookup and serving via ParentManager (#48021) (b440fbf26) by AlpinDale
• [sync] [Model] Support top_k and top_p sampling for DiffusionGemma (#45429) (c82a9e9b7) by AlpinDale
• [sync] [Perf] Isolate MM preprocessing on its own executor (#49524) (9e715a57f) by AlpinDale
• [sync] [CI] Fix speech correctness check rejecting improved WER (#49853) (c04fad394) by AlpinDale
• [sync] [Perf] DeepSeek-OCR-2 TTFT Optimize (#49531) (bf416a1ec) by AlpinDale
• [sync] [Doc] Add compile cache volume example to the Docker deployment page (#49782) (64f825e62) by AlpinDale
• [sync] [KV Offload] Deduplicate replicated MLA KV in the shared CPU region (#48906) (46cf4a9b1) by AlpinDale
• [sync] [Bugfix][KV Offload] Namespace persistent cache by model runner (#49440) (c462eb154) by AlpinDale
• [sync] [Bugfix] Respect declared attention contract for ColQwen3.5 retrievers (#49372) (1f0a624bb) by AlpinDale
• [sync] [Bugfix][KVConnector] Disable cross-layer KV blocks for per-token-head quant (#49226) (dfd819809) by AlpinDale
• [sync] [CI] Compute speech WER directly with jiwer (#49773) (b11e887bd) by AlpinDale
• [sync] [Kernel] TD operand loads for batched MoE GEMM (moe_mmk) on XPU (#46340) (2d2628c4f) by AlpinDale
• [sync] [Bugfix][KV Offloading] Defer request finalization until final store (#49671) (8128747b7) by AlpinDale
• [sync] [UX] DCP Topology Validation (#49777) (d5f998a41) by AlpinDale
• [sync] [MM][CG] Support ViT CUDA Graph for Gemma-4 (#46837) (bc4d49167) by AlpinDale
• [sync] [Build] Fix for DeepEP manylinux pidfd sycall usage (#49814) (435a924b0) by AlpinDale
• [sync] [CI] Stabilize Pooling Rerank Equivalence Test (#49822) (2bd13da3e) by AlpinDale
• [sync] [Bugfix] Wait for the linear bias before layerwise online processing (#49805) (8a7e730fd) by AlpinDale
• [sync] [Model] Add VaultGemma via Transformers modeling backend (#49803) (6cc855837) by AlpinDale
• [sync] [Perf] Fix moe reduce_scatter perf regression by removing additional comm, 5% E2E throughput gain back. (#48763) (35aa7295e) by AlpinDale
• [sync] [Feature] Add fault tolerance framework (simplified) for DP+EP external LB deployments (#44428) (1217330fc) by AlpinDale
• [sync] [BugFix] Increase the max supported duration for MOSS-TD (#49403) (232ad8626) by AlpinDale
• [sync] [Bugfix][MiniMax-M3] Fix token-major top-k buffer handling in Triton … (#49149) (2aea34103) by AlpinDale
• [sync] [Bugfix][Tool Parser] Fix dropped streaming arguments in Jamba and InternLM2 parsers (#48852) (092fe8329) by AlpinDale
• [sync] [CI] Stop flaky test from downloading model every time (#49800) (2151835dd) by AlpinDale
• [sync] feat[vLLM × v5]: Add audio support for the Transformers backend (#39330) (a0f1a8a09) by AlpinDale
• [sync] Make bare hugging_face imports forbidden (#49726) (6e32b8e2d) by AlpinDale
• [sync] [Core] Keep attention backends eligible for text-only serving of prefix-LM models (#48796) (d7cd23687) by AlpinDale
• [sync] [Model] Remove Ouro (#49786) (85fcd64c7) by AlpinDale
• [sync] [Perf][V1] Skip LRU hash-split in free_blocks when prefix caching is off (#48017) (c9a495ce1) by AlpinDale
• [sync] [Docs] Fix confusing docstring indentation in nemotron_h.py (#49781) (3feb9c053) by AlpinDale
• [sync] [multimodal] Make PyNvVideoCodec decoder concurrency configurable (#49753) (b4827900f) by AlpinDale
• [sync] Stabilize GPU memory teardown between ROCm CI tests (#49242) (69e833183) by AlpinDale
• [sync] [CI] fix compile test | refactor VLLM_DISABLE_COMPILE_CACHE for tests (#49770) (e291bcf98) by AlpinDale
• [sync] [Bugfix][KV Connector][Mooncake] Keep TP-sharded Mamba state out of the KV-head dedup (#49499) (3fae4eee6) by AlpinDale
• [sync] [XPU] add warning for xpu graph limitations (#49419) (e8d4cf434) by AlpinDale
• [sync] [BUGFIX] Fix log capture in KV test (#49655) (39e9e0ec0) by AlpinDale
• [sync] Revert "[Perf][GLM-5.2] Blackwell decode optimizations" (#49768) (478ee214a) by AlpinDale
• [sync] [ROCm][CI] Wait for ROCm VRAM to settle between compiled and eager LL… (#49739) (bf89fd327) by AlpinDale
• [sync] [ROCm][Docker] Drop MORI_GPU_ARCHS so MoRI autodetects the device arch (#49737) (2f6eaaf5a) by AlpinDale
• [sync] [ROCm][CI] Fix XPASS(strict) on mixed audio embeds test (#49733) (11c5b81c4) by AlpinDale
• [sync] [KV Offload][CI] Fall back to buffered I/O without O_DIRECT; fix flaky api-server test (#49734) (184d4a516) by AlpinDale
• [sync] [Model] Remove Plamo2 (#49729) (919da3b86) by AlpinDale
• [sync] [CI] Stabilize memory-sensitive compile and structured output tests (#49749) (c8a3c02b3) by AlpinDale
• [sync] [Bugfix] Support non-uniform page sizes in KVBlockZeroer (#49704) (cd3ec998b) by AlpinDale
• [sync] [AMD][Bugfix][EPLB] Fix elastic EP scaling accuracy on ROCm (#47206) (df4f0d6d0) by AlpinDale
• [sync] [WideEP] Update NCCL to 2.30.7 to enable DeepEPv2 in the vllm/vllm-openai image (#45321) (ca35e1437) by AlpinDale
• [sync] [Bugfix] Register axk1 config to fix A.X-K1 init (#49727) (327c88ae1) by AlpinDale
• [sync] Fix GLM-4.1V video placeholder token ID handling. (#49484) (8d616782d) by AlpinDale
• [sync] [CI] Avoid unnecessary Hugging Face metadata requests (#49508) (9b5266ccb) by AlpinDale
• [sync] [CI] Reuse loaded config for cached tokenizer (#49509) (4b0e7d57b) by AlpinDale
• [sync] Add sm_107 for Rubin (#49387) (57bb8332c) by AlpinDale
• [sync] [Kernel] ReplaySSM: cache SSM inputs for faster Mamba2 standard decode (#48018) (625a5fa4f) by AlpinDale
• [sync] [Bugfix] Accept RFC 2397 parameters in base64 data URLs (#48973) (67738e528) by AlpinDale
• [sync] [UX] Improve data-parallel launch validation (#49124) (030591b71) by AlpinDale
• [sync] [UX] Reject incompatible nested runtime overrides (#49247) (fc1c8b8c2) by AlpinDale
• [sync] [ROCM] Fix AITER Fused AllReduce RMSNorm for Transformers Backend (#49673) (fe665f94a) by AlpinDale
• [sync] [Bugfix][Benchmarks] Restore --skip-tokenizer-init with custom dataset (#49180) (895a0c9a0) by AlpinDale
• [sync] Encoder cache extension hooks (#48218) (9b6f8ddc8) by AlpinDale
• [sync] [Bugfix] Skip linear bias in layerwise reload to avoid corruption (#49586) (776313c93) by AlpinDale
• [sync] [PD][NixlPush][Bugfix] Fix blocking handshake call on writer thread (#49221) (0f3bbbc75) by AlpinDale
• [sync] Remove Quantization test parallelism (#49693) (1e152915b) by AlpinDale
• [sync] [Perf][GLM-5.2] Blackwell decode optimizations (#48597) (a40b3d87d) by AlpinDale
• [sync] Bump Transformers version to 5.14.1 (#49223) (c1e05d447) by AlpinDale
• [sync] [CI] Use explicit devices in IR tests (#49513) (4b7f247d9) by AlpinDale
• [sync] [Bugfix] Fix humming kernel crash when layer.has_bias is None (#48769) (59d88b29e) by AlpinDale
• [sync] [CompressedTensors] DeepSeek4 CT Quantization Support (#41276) (7c3c3ff3b) by AlpinDale
• [sync] [Bugfix] Detect mixed precision in packed KV cache specs (#49623) (baa61ca8a) by AlpinDale
• [sync] [Model] Support llm-compressor Inkling NVFP4 weights (#49258) (18d56625e) by AlpinDale
• [sync] [ROCm][Quantization] Add Quark W4A8 (INT4-FP8) MoE CI coverage (#48050) (a0208e41c) by AlpinDale
• [sync] [ROCm][CI] Prepare AMD mirrors for regating (#49270) (701f8f66d) by AlpinDale
• [sync] [BugFix][LoRA] Skip marlin-backend gpt-oss LoRA tests on XPU (#49385) (b4b8b7317) by AlpinDale
• [sync] [CI][PD] Add hybrid SSM P_TP>D_TP accuracy sweep entry (#49593) (b33db51e6) by AlpinDale
• [sync] [CI] Disable reasoning in Responses smoke test (#49511) (e53f74a46) by AlpinDale
• [sync] [ROCm][CI] Language Models tests tiny-mixtral with aiter fix (#49551) (849c45b3d) by AlpinDale
• [sync] [Bugfix][KV cache] Support sparse-MLA targets with SWA drafts (#48776) (888455ff8) by AlpinDale
• [sync] [Bug] Fix batch invariance rms norm comparison (#49603) (af57381f9) by AlpinDale
• [sync] [Core] Update PyTorch to 2.13.0, torchvision to 0.28.0, triton to 3.7.1 (#48155) (2a2a0dfe9) by AlpinDale
• [sync] [CI][Bugfix] Fix test isolation in block_int8/ptpc_fp8 MoE kernel tests (#49609) (8ab55f0e5) by AlpinDale
• [sync] [CI] Isolate cudagraph tests in child processes (#49510) (8e0e4e9f6) by AlpinDale
• [sync] [DSv4 Perf] Skip topk and router when not needed, 3.4% E2E TTFT improvement for Decode case (#49486) (6b0822f38) by AlpinDale
• [sync] [CI] Use explicit devices in quantization tests (#49512) (53f1eacac) by AlpinDale
• [sync] Revert "[MRV2] Always build attn metadata at capture time" (#49364) (#49451) (f1076032a) by AlpinDale
• [sync] [Bugfix][Structured Output][Spec Decode] Advance grammar across reasoning boundary (#44993) (a3cc9b532) by AlpinDale
• [sync] Fix GPTQ quantized Qwen3.5 MTP weight loading with spec decode (#48816) (3d38c9dad) by AlpinDale
• [sync] [Perf] Defer MM embeds loading off the event loop (#49477) (45c9fe2fa) by AlpinDale
• [sync] [Bugfix][CI/Build] Fix Plamo2 HF runner crash on transformers v5 (_tied_weights_keys list→dict) (#44239) (bb9c1fc1a) by AlpinDale
• [sync] [Bugfix][Spec Decode] Select earliest-completing stop string in check_stop_strings (#49391) (f4135b082) by AlpinDale
• [sync] [Bugfix] handle grammar compilation failures to avoid engine crash (#47312) (d1f85f2df) by AlpinDale
• [sync] [Bugfix][Core] shm_broadcast: bound idle reader waits and release read slots (#45224) (36e199800) by AlpinDale
• [sync] [MRV2][Spec Decode] Avoid rejection sampler OOM by chunking (#48630) (123df4a57) by AlpinDale
• [sync] [Core] Simplify KVBlockZeroer index tensor handling (#48399) (5086576e1) by AlpinDale
• [sync] [MooncakeStore] Re-derive full external hits on stored boundaries (#49481) (0722a9096) by AlpinDale
• [sync] [Perf][KVConnector][Mooncake] Vectorize prepare_value on the KV load path (#48531) (66999c382) by AlpinDale
• [sync] [Misc] Use VLLMValidationError in chat_utils content-part validation (#49217) (c87deabe5) by AlpinDale
• [sync] [Bugfix] Fix DeepSeek-V4 DSpark draft shared-expert padding for TP > 8 (#49415) (0b9024c87) by AlpinDale
• [sync] [Performance][Model] Avoid transient Inkling result allocations (performance, and OOM prevention on smaller memory configurations) (#49487) (ce46f9286) by AlpinDale
• [sync] [Bugfix][Model] Remove SciPy dependency from Inkling scale planning (#49485) (49f90f96e) by AlpinDale
• [sync] [Docs] Re-add Reo.dev analytics beacon (#49474) (2a9e71813) by AlpinDale
• [sync] [Bugfix] Make shared NVFP4 MoE scales writable (#49489) (57fdfa219) by AlpinDale
• [sync] Fused Shared Expert Support for AMD Quark DeepSeek-V4 Model Checkpoints (#48044) (62aa0e6a8) by AlpinDale
• [sync] [BugFix] Handle per-group prefix-hit divergence for hybrid models with KV connector (#48425) (bdc87341b) by AlpinDale
• [sync] [Core][DSV4] Compact MXFP4 indexer KV cache and packed group overlays (#48993) (88d67e980) by AlpinDale
• [sync] [Bugfix] Exclude location-derived path vars from torch.compile cache factors (#47573) (074f6698a) by AlpinDale
• [sync] [Bugfix] Fix DeepGEMM warmup when using FlashInferFp8DeepGEMMDynamicBlockScaledKernel (#49467) (b14dd9be3) by AlpinDale
• [sync] [Bugfix] Retry config read to survive concurrent HF cache refresh (#49001) (2979ce418) by AlpinDale
• [sync] [Bugfix] Restore gather_and_maybe_dequant_cache OOB guard (#49427) (df4a08a7f) by AlpinDale
• [sync] [CI] Fix stale/fragile untethered kernels-root tests (#49423) (3512e5d80) by AlpinDale
• [sync] Bump Flashinfer version to 0.6.15 (#48914) (14b716eca) by AlpinDale
• [sync] [Bugfix][Parser] Fix special tokens (EOS/BOS) leaking into reasoning content (#48748) (8ba1c8260) by AlpinDale
• [sync] [CI] stabilize GDN prefill CuTeDSL test (#49388) (25e71b76b) by AlpinDale
• [sync] [ROCm][CI] skip moe weight padding for eplb (#49350) (98cacd5c4) by AlpinDale
• [sync] [DSv4 Perf] Skip empty c128 kernel launch, around 2x kernel performance improvement. (#48957) (02a755a1d) by AlpinDale
• [sync] [Bugfix][Renderer] Rebuild vision chunk UUIDs in async render path (#49400) (cee86b8d2) by AlpinDale
• [sync] [Build] Bump vllm-flash-attn to C++20-compatible commit for torch-nightly (#49326) (f14458e13) by AlpinDale
• [sync] Upgrade tpu-inference to v0.25.0 (#49431) (be3de5343) by AlpinDale
• [sync] [Rust Frontend] Extract request preparation from the inference path (#49045) (5cb786dd3) by AlpinDale
• [sync] [Rust Frontend][gRPC] Add abort control RPC (#49255) (6b975c986) by AlpinDale
• fix: translate pooling config names (7b27f4a50) by AlpinDale
• [sync] Parallelize pooling request preprocessing (#49153) (9140dd3f5) by AlpinDale
• [sync] [XPU] WA of topk_softplus_sqrt arg mismatch on XPU (#49408) (0812df02b) by AlpinDale
• [sync] [XPU] WA of topk_softmax arg mismatch on XPU (#49395) (0223f74dc) by AlpinDale
• [sync] [PD][Bugfix] Fix NIXL hybrid MLA+mamba heterogeneous TP (#49297) (a798ffd20) by AlpinDale
• [sync] [Renderer] Offload derender CPU work to renderer thread pool (#49396) (20da720ca) by AlpinDale
• [sync] [ROCm] Remove redundant AITER fused_qk_rmsnorm probe (avoids config-time HIP init) (#47992) (e898b8f28) by AlpinDale
• [sync] [ROCm] Upgrade NIXL and UCX (#49251) (bec4c859e) by AlpinDale
• [sync] [MRV2] Always build attn metadata at capture time (#49364) (f764aefa5) by AlpinDale
• [sync] [Bugfix][Attention] Ignore empty MLA context chunks during merge (#49294) (ced037cde) by AlpinDale
• [sync] [MoE Refactor] Migrate MoeWNA16Method quantization method over to using the new MK oracle scheme. (#44120) (6c5704df3) by AlpinDale
• [sync] [Misc] Move PyNvVideoCodec stuff out of gpu worker (#49322) (ae6094835) by AlpinDale
• [sync] [Bugfix][SpecDecode] Scope MTP completeness checks outside bucketed updates (#49178) (6e6d00793) by AlpinDale
• [sync] [CI][Bugfix] Fix ROCm FP8 KV cache dtype in attention backend test (#49380) (649c55a39) by AlpinDale
• [sync] [CI][Bugfix] Fix and wire streaming-input tests (#49356) (a6639aadf) by AlpinDale
• [sync] [Misc] Fix terminal output logo coloring (#49344) (1c3886e2a) by AlpinDale
• [sync] [ROCm][CI] Fix order-dependent failure in test_flash_attn_accepts_handled_fp8_variants (MI355) (#49329) (b0674c6fd) by AlpinDale
• [sync] skip cudagraph/DP padding in topk (#48979) (52e4829d3) by AlpinDale
• [sync] [Bugfix] DFlash fc sized wrong when num_target_layers != num_hidden_layers (#48524) (d73d43795) by AlpinDale
• [sync] [CI] Add gemma-4-E4B-it-assistant to CI gsm8k for GemmaMTP (#49243) (b4137fe2c) by AlpinDale
• [sync] [Bugfix] fix cutalss version upgrade bug, need update MSG new commit (#49016) (04075f622) by AlpinDale
• [sync] [CI/Build][The Rock][BugFix] Use fork method in test_multiproc_executor_multi_node for py 3.14 compat and fix test_multiproc_executor_shutdown_cleanup (#48655) (d2a69fcfc) by AlpinDale
• [sync] [CI] Fix and wire encoder/manager cudagraph unit tests (#49339) (5b17f9a4f) by AlpinDale
• [sync] [CI][Bugfix] Reduce max_model_len in OOT embedding test to fix KV-cache OOM on small GPUs (#49351) (cf213b8e1) by AlpinDale
• [sync] Fix Qwen3-VL M-RoPE on the Transformers modeling backend (grids + compile) (#49292) (3d87ae61e) by AlpinDale
• [sync] [Bugfix][Spec Decode] Restrict embedding-width share guard to EAGLE drafts (#47953) (3b6af1fb0) by AlpinDale
• [sync] [Bugfix] Fix DSA crash under breakable piecewise cudagraphs (#49302) (5eb624d62) by AlpinDale
• [sync] [ROCm][CI] Fix AITER MLA fp8 decode metadata regression test (#48845) (d5626d24f) by AlpinDale
• [sync] Ci/add laguna xs gsm8k (#49241) (a6531ecd4) by AlpinDale
• [sync] [Test] Add DeepSeek MTP parallel-load tests (#41653) (e2fe5049a) by AlpinDale
• [sync] [Bugfix] Prefix-cache metrics double-counted when a KV connector defers requests (#48860) (70e4777b7) by AlpinDale
• [sync] [Bugfix] Handle MLA fallback during FA4 JIT warmup (#49306) (0d86c951c) by AlpinDale
• [sync] [XPU][DeepSeekV4]Add DeepSeek-V4 fuse_index_q SYCL kernel path (#45991) (63ddbb314) by AlpinDale
• ci: harden incremental wheel cache (5c4e46c6a) by AlpinDale
• docs: fix derenderer link (0d6cd7ecd) by AlpinDale
• ci: cache incremental wheel builds (89abe5887) by AlpinDale
• fix: expose CUDA stream shim to Qutlass (e66025eea) by AlpinDale
• fix: support PyTorch 2.13 CUDA stream API (07ce5e872) by AlpinDale
• ci: add manual CUDA wheel publishing (55dcf9dc5) by AlpinDale
• Delete .github/CODEOWNERS (5701f39e5) by @AlpinDale
• [sync] [BugFix] Set graph_pool_id before FULL CUDA graph capture in ModelRunner V2 (#48843) (c0e633ca1) by AlpinDale
• [sync] [Misc] Use VLLMValidationError in chat completion tool and batch validators (#49214) (5df156361) by AlpinDale
• [sync] [Bugfix][CPU] Fix Clang OpenMP build on macOS (#49021) (66584f849) by AlpinDale
• [sync] [bugfix] Fix Cosmos3 Edge checkpoint weights filtering, video loading, prompt expansion (#49190) (84a4ed4b5) by AlpinDale
• [sync] [3/N][KV-Cache Layout Refactor] Standardize Mamba cache; drop get_transfer_cache_regions (#44456) (cba70de1a) by AlpinDale
• [sync] [Rust][Benchmark] Use async HTTP clients (#49295) (23f97c973) by AlpinDale
• [sync] [Misc] Remove old now unsupported max_num_partial_prefills and max_long_partial_prefills (#49244) (34bd514a5) by AlpinDale
• [sync] [Misc][Docs] Fix XPU compute-runtime driver link version mismatch (#49299) (b62a0f526) by AlpinDale
• [sync] Revert "[Sampler] Stop upcasting logits to fp32 in apply_sampling_params" (#48641) (#49033) (b51742ecb) by AlpinDale
• [sync] [Bugfix] Fix Ovis2_5 special tokens for transformers v5 (#47298) (e70f3438c) by AlpinDale
• [sync] Propagate Flash Attention cache configuration to Ray workers (#49177) (7b5973ec0) by AlpinDale
• [sync] [Misc] Add @esmeetu to codeowners for rust/src/bench (#49298) (b4cbfca37) by AlpinDale
• [sync] [Bugfix] Propagate quant_config to LFM2 ShortConv projections (#48917) (6fb9d6e3a) by AlpinDale
• [sync] [Frontend] Support additional sampling parameters for translation API (#45839) (dd4095998) by AlpinDale
• [sync] [Rust][Benchmark] Use tracing for logs (#48937) (b5e4a784f) by AlpinDale
• [sync] [Bugfix][Rust Frontend] Handle zero-column logprobs payloads without panicking (#49113) (d55bcc94d) by AlpinDale
• [sync] Update BGE-M3 token expectations for leading spaces (#49269) (851494ac2) by AlpinDale
• [sync] [CI][NIXL] Isolate concurrent engine internal ports (#49129) (35ccca2b2) by AlpinDale
• [sync] [MRV2] Add encoder cache profiling implementation (#47985) (1ae759cb9) by AlpinDale
• [sync] [Bugfix][KV Offloading] Handle queued request aborts without allocated KV blocks (#49146) (d18ecd878) by AlpinDale
• fix: build qutlass without TORCH_TARGET_VERSION and with correct pybind-disable macro (ffd92b949) by AlpinDale
• fix: incorrect macro name from sync (c9443fc59) by AlpinDale
• fix: grow padding instead of block size when unifying pre-padded kv cache specs (bf01d9926) by AlpinDale
• fix: gate ll_bf16 router gemm warmup to sm90/sm100 and make it non-fatal (edb754721) by AlpinDale
• [sync] [Rust Frontend][gRPC] Add engine-aware health reporting (#48992) (0c04aea83) by AlpinDale
• [sync] [CI] Exercise FA3 FP8 attention on SM90 (#49231) (7e9483941) by AlpinDale
• [sync] Update qutlass cmake for stable abi (#47879) (990249374) by AlpinDale
• [sync] [Bugfix] Fix WSL circular import from pin_memory warning_once (#48444) (412b5f142) by AlpinDale
• [sync] Fixes non-coalesced HBM access in marlin_int4_fp8_preprocess_kernel_awq (#47268) (c321b31fa) by AlpinDale
• [sync] [Bugfix] Count per-group blocks in get_max_concurrency_for_kv_cache_config (#48317) (94ab738a4) by AlpinDale
• [sync] [Cleanup] Remove unused StructuredOutputRequest.status field (#49235) (23a85879f) by AlpinDale
• [sync] [Attention][MLA][DCP] Query replication for MLA decode (DeepSeek-V2/R1 + Kimi-K2.5) (#45964) (d497f5466) by AlpinDale
• [sync] [RL Infra][FlashInfer] Enable router replay output from FlashInfer monolithic MoE kernel (#44214) (896938a8c) by AlpinDale
• [sync] [copy of #45208] CuMem slept-L1 fragmentation accounting (#49208) (05fb2b4b0) by AlpinDale
• [sync] [chore] adjust logo be more friendly to white background terminal (#48938) (13fb27887) by AlpinDale
• [sync] Support loading sample_from_anchor flag from speculators config (#48639) (215a0209a) by AlpinDale
• [sync] [ROCm][CI] fix test_rocm_quick_reduce.py (#49234) (e157cb3fe) by AlpinDale
• [sync] [Bugfix][Spec-Decode] Populate draft seq_lens_cpu_upper_bound for spec-decode attention metadata (#44492) (a96496a44) by AlpinDale
• [sync] [Feat][Perf] Add new warmup infrastructure for JITs (#47451) (ea9cf9378) by AlpinDale
• [sync] [Bugfix][KV Offload] Preserve reachable tails for hybrid SWA groups (#48911) (90ba9236c) by AlpinDale
• [sync] Cosmos3 FP8 ModelOpt/Diffusers remapping (#48952) (773667c08) by AlpinDale
• [sync] [Revert][Bugfix] Restore MiniCPM-V 4.6 ViT QKV weight loader (#49193) (0b15337e9) by AlpinDale
• [sync] [Attention] FlashAttention 4 SM100 FP8 kv cache support (#42569) (99d12a397) by AlpinDale
• [sync] [XPU]add sycl path for Mhc (#47245) (d48b2da6f) by AlpinDale
• [sync] [Rust Frontend] Bump xgrammar-structural-tag and enable local extension (#49161) (40e45dcd3) by AlpinDale
• [sync] [Multimodal] Allow keeping original image mode for ImageIO (#49159) (4d5cc195b) by AlpinDale
• [sync] [KV Offload] Support self-describing KV events with TieringOffloadingSpec (#48679) (316ca8e13) by AlpinDale
• [sync] [Bugfix][Gemma4] Fix ModelOpt mixed-precision MoE config mapping (#48563) (fe50c154c) by AlpinDale
• [sync] [Bugfix] Zero new KV blocks for quantized + sliding-window hybrid caches (#47574) (aff17e39f) by AlpinDale
• [sync] fix(openai): reject non-numeric logprobs with 400 instead of 500 (#49144) (610455aae) by AlpinDale
• [sync] [Bugfix][Rust Frontend] Map missing prompt logprobs for single-token prompts in chat and raw generate (#49111) (3e732c3e2) by AlpinDale
• [sync] [Rust Frontend] Fix macro-based content format detection (#49042) (7f9f35940) by AlpinDale
• [sync] [Rust][Benchmark] Integrate vllm-bench to vllm-rs & vllm CLI (#48930) (ea817fd6e) by AlpinDale
• [sync] [XPU] FP8 o_proj with fp8_bmm and load-time scale transpose (#48334) (3e611d4b8) by AlpinDale
• [sync] [Rust Frontend] Use zero-copy slicing for multimodal tensors (#48781) (54a66e6d1) by AlpinDale
• [sync] [XPU] [MoE] add quant input when prepare for fusedmoe (#47122) (c1c4a399d) by AlpinDale
• [sync] [CPU] fixes heterogeneous NIXL KV transfer into CPU_ATTN decode workers (#47871) (81530e18b) by AlpinDale
• [sync] [XPU][Bugfix] Fix GroupCoordinator device_index (#47295) (90f50f9fa) by AlpinDale
• [sync] [Bugfix][RL] Set vLLM config during weight reload (#45989) (7217f7486) by AlpinDale
• [sync] [Doc] Document blocks_per_chunk in the KV offloading guide (#49100) (e48c90a42) by AlpinDale
• [sync] [Bugfix][KV Offload] Propagate EAGLE mode to SimpleCPU coordinator (#49071) (552e749d8) by AlpinDale
• [sync] [Hardware][CPU] Enable granite-4 model on cpu (#47641) (d13aa3d97) by AlpinDale
• [sync] [Bugfix] Fix logprobs token-string collision from SentencePiece space… (#48674) (359ef740a) by AlpinDale
• [sync] [ROCm][CI] Fix sparse MLA metadata sync fixture (#49128) (243d19d22) by AlpinDale
• [sync] [Bugfix] Fix broken NVVM caused by CuteDSL 4.6.0 (#49108) (3bdd019ac) by AlpinDale
• [sync] [CI] Move compatible 1xL4 jobs to H200 35GB MIG (#43024) (1255bca0f) by AlpinDale
• [sync] [CI/Build][BugFix][The Rock][AMD] Add spawn method in vision examples to avoid reinitialization (#47932) (06fa31b35) by AlpinDale
• [sync] [Bugfix] Reject removed pooling parameters (#48984) (8ccff5686) by AlpinDale
• [sync] [Refactor] Extract StructuredOutputsParams creation logic from Request.to_sampling_params (#49003) (3c85b34b2) by AlpinDale
• [sync] [Core] Add MRV2 virtual-batch PCP for MLA (#46570) (a93099049) by AlpinDale
• [sync] [Multimodal] Automatically fallback to ViT DP when TP is unavailable (#49046) (5f474e538) by AlpinDale
• [sync] [ROCm][CI] Ensure sliding window tests release GPU memory (#49055) (bd673425a) by AlpinDale
• [sync] [Bugfix] Qwen3-VL/Qwen-Omni: honor max_pixels/min_pixels for video prompts (#49015) (ed9dbaeee) by AlpinDale
• [sync] [Front-end] [Messages] Populate num_cache_creation_tokens (#48535) (9fb5b71e9) by AlpinDale
• [sync] [ROCm][DSv3.2][Perf] Cap sparse MLA decode KV-splits with a work-per-split heuristic (#46832) (e09d3669e) by AlpinDale
• [sync] [ci] Move 3 entrypoints tests to h200_35gb queue (#43164) (8282cb042) by AlpinDale
• [sync] [Bugfix] Bump tml-fa4 for cutlass-dsl 4.6 API compatibility (#48988) (f129585ca) by AlpinDale
• [sync] Remove even more unnecessary load_weights methods (#48496) (08589c95c) by AlpinDale
• [sync] [Bugfix] Re-sync parameter tp_rank after process_weights_after_loading (fix replicated / disable_tp weight reload) (#48025) (aefaec15a) by AlpinDale
• [sync] [Perf][Hybrid] Vectorize _copy_mamba_state_block to uint64 for temporal (#48110) (fbf689d2e) by AlpinDale
• [sync] [ROCm][Bugfix] Fix GPT-OSS Quark MXFP4 MoE loading - emulation buffer not block-aligned (#43979) (6d54a6df0) by AlpinDale
• [sync] [LoRA] Optimize TrtLlmLoRAExperts (#48759) (3003a4637) by AlpinDale
• [sync] [Sampler] Stop upcasting logits to fp32 in apply_sampling_params (#48641) (47116bbd7) by AlpinDale
• [sync] [Model] Use standard ModelOpt config for Inkling NVFP4 (#48990) (fbebe738f) by AlpinDale
• [sync] [Frontend]Flatten beam-search beams with itertools.chain instead of sum (#48829) (47627decf) by AlpinDale
• [sync] [Bugfix][V1/V2] Fix prompt_logprobs to respect logprobs_mode (#47680) (6573f2b92) by AlpinDale
• [sync] [Hardware][GPU] Profiler config additional to increase it scope and annotation details (#37524) (842b04ffa) by AlpinDale
• [sync] [Perf] Optimize dsv4 routing using specialized kernel, 2.94% E2E TPOT improvement (#48660) (e10559dc0) by AlpinDale
• [sync] [Loader] Improve InstantTensor loading (#46868) (d389ab9ea) by AlpinDale
• [sync] [Attention] Allow selecting a different attention backend per KV-cache group (#48012) (7d6e8a5de) by AlpinDale
• [sync] [Bugfix][GLM4V] Fix video dummy profiling and memory usage (#48729) (ef9c6a457) by AlpinDale
• [sync] [ROCm][Perf][DSV4] Improve sparse decode reduction occupancy on gfx950 (#48788) (fefbf99a5) by AlpinDale
• [sync] [Bugfix][Tool Parser] Preserve whitespace in parameter values (MiniMax M2, Qwen3, MiniCPM5 XML) (#48846) (a86d91e2d) by AlpinDale
• [sync] [Refactor] Remove deepseek dead code (#48780) (9bad9e7d6) by AlpinDale
• [sync] Fix: Restore data_parallel_size > 1 for use_sequence_parallel_moe (#48849) (c2a69560f) by AlpinDale
• [sync] [Performance] Use CuTe-DSL for FlashInfer MXFP4 quantization (#48417) (4394388dd) by AlpinDale
• [sync] [Bugfix][KV Offloading] Offload last block at request finish and prevent reuse race (#48596) (6e2b3d15e) by AlpinDale
• [sync] [XPU] Bump vllm_xpu_kernels to v0.1.11.1 (#48942) (8bf8f484f) by AlpinDale
• [sync] [Bugfix]Fix transformer backend failed: AttributeError: 'Parameter' object has no attribute 'weight_loader' (#48699) (a041dec36) by AlpinDale
• [sync] [XPU][UT]fix _POSSIBLE_KERNELS error on XPU (#47516) (3dac5c11b) by AlpinDale
• [sync] [XPU] support HND layout (#47975) (ad83ce343) by AlpinDale
• [sync] [Bugfix] Enable FlashAttention MLA prefill for Mistral Small 4 head dims (#48855) (ad9b553c1) by AlpinDale
• [sync] [XPU] allow forcing flash attn for mm_prefix (#48828) (63fc9b72d) by AlpinDale
• feat: support min_p with speculative decoding (#1716) (e38e6a8a8) by @AlpinDale

Full Changelog: v0.22.0...v0.23.0