Repository navigation
Add ninja to dependency - #21
Merged
Merged
Conversation
hongxiayang
pushed a commit
to hongxiayang/vllm
that referenced
this pull request
Feb 13, 2024
slyalin
pushed a commit
to slyalin/vllm
that referenced
this pull request
Apr 3, 2024
…ock_size [CPU] Support for larger block_size
tdg5
pushed a commit
to tdg5/vllm
that referenced
this pull request
Apr 25, 2024
Fix more logging lint errors
z103cb
referenced
this pull request
in z103cb/opendatahub_vllm
May 7, 2024
Signed-off-by: Nick Hill <nickhill@us.ibm.com> Co-authored-by: Daniel Clark <daniel.clark@ibm.com>
Closed
wuhuikx
pushed a commit
to wuhuikx/vllm
that referenced
this pull request
Mar 27, 2025
make package version control by setuptools_scm to keep the same with vllm Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
1 task done
1 task done
1 task done
1 task done
zyongye
pushed a commit
to zyongye/vllm
that referenced
this pull request
Aug 5, 2025
zyongye
pushed a commit
to zyongye/vllm
that referenced
this pull request
Aug 6, 2025
heheda12345
pushed a commit
to heheda12345/vllm
that referenced
this pull request
Sep 29, 2025
Co-authored-by: Lucia Fang <fanglu@meta.com>
1 task done
inkcherry
pushed a commit
to inkcherry/vllm
that referenced
this pull request
Nov 6, 2025
Co-authored-by: root <root@smc300x-ccs-aus-gpue77e.prov.aus.ccs.cpe.ice.amd.com>
4 tasks
yma11
pushed a commit
to yma11/vllm
that referenced
this pull request
Nov 12, 2025
* remove duplicated code Signed-off-by: Kunshang Ji <kunshang.ji@intel.com> * remove more Signed-off-by: Kunshang Ji <kunshang.ji@intel.com> --------- Signed-off-by: Kunshang Ji <kunshang.ji@intel.com>
dik654
pushed a commit
to dik654/vllm-for-study
that referenced
this pull request
Nov 18, 2025
New Industry Use Cases (vllm-project#21-30): - vllm-project#21 Game Development: AI game testing + balance tuning - vllm-project#22 Construction: Vision AI safety inspection - vllm-project#23 Agriculture/Smart Farm: Crop monitoring + pest detection - vllm-project#24 Government/Public: Document automation + citizen services - vllm-project#25 Energy/Utilities: Grid monitoring + anomaly detection - vllm-project#26 Environment/Sustainability: Carbon tracking + ESG reporting - vllm-project#27 Fashion/Apparel: Trend analysis + inventory optimization - vllm-project#28 Sports/Fitness: Performance analytics + tactical analysis - vllm-project#29 Automotive/Mobility: Autonomous driving simulation - vllm-project#30 Space/Aerospace: Satellite image analysis Advanced Architecture Patterns: 1. Event-Driven Pattern: Webhook → Event Bus → Agent triggers 2. Streaming Pattern: Large dataset processing with chunking 3. Batch Processing Pattern: Celery-based parallel processing 4. Circuit Breaker Pattern: Fault tolerance + auto recovery 5. CQRS + Event Sourcing: Command/Query separation 6. Saga Pattern: Distributed transaction management Guide now covers: - 30+ industry-specific MCP implementations - 6 production-ready architecture patterns - Real-world scalability solutions - Enterprise integration strategies - Total: 8,672 lines (from 7,249)
1 task done
1 task done
1 task done
1 task done
yuezhu1
added a commit
to yuezhu1/vllm
that referenced
this pull request
Mar 30, 2026
- Add _lora_slots field on LoRAModelManager, decoupled from lora_config so dynamic scaling does not mutate the original config object - Add _evict_adapters_to_fit() hook on base class (raises on overflow); LRUCacheLoRAModelManager overrides it with LRU eviction and rebuilds _active_adapters cache with new capacity (cachetools maxsize read-only) - Implement resize_lora_slots() on base class: validates, evicts, calls reallocate_lora_weights() on all modules, empty_cache() once, resizes lora_index_to_id, updates _lora_slots - Step 7 (re-load surviving adapters) intentionally omitted — weights are preserved via GPU-to-GPU copy in reallocate_lora_weights(); comment notes what to do if a remote weight store is introduced in future - Add tests/lora/test_lora_model_manager_resize.py: 6 CPU-only unit tests covering validation, no-op, grow, LRU shrink, base-class overflow raise, and empty_cache() called exactly once Closes vllm-project#11 Closes vllm-project#21 AI assistance was used; all changed lines reviewed by the submitter. Co-authored-by: Claude Signed-off-by: Yue Zhu <Yue.Zhu@ibm.com>
amd-xavierwang
pushed a commit
to amd-xavierwang/vllm
that referenced
this pull request
Apr 8, 2026
…WEEP (vllm-project#21) ## Motivation Reduce built time by disabling sweep functions by default ## Summary - Wraps sweep functions in skinny_gemms_int4.cu and skinny_gemms_int8.cu with ifdef VLLM_SKINNY_GEMM_SWEEP, matching the existing pattern in skinny_gemms.cu - Also guards the corresponding forward declarations in ops.h and torch binding registrations in torch_bindings.cpp ## Build time improvement - Before: ~1236 seconds (~20.6 minutes) - After: ~213 seconds (~3.6 minutes) - Reduction: ~83% ## Test plan - Clean build succeeds without errors
JaimeCernuda
pushed a commit
to JaimeCernuda/vllm-kvlobotomy
that referenced
this pull request
Apr 17, 2026
Added OLMo-2-style post-norm forward alongside the existing pre-norm (LLaMA-style) path. Dispatched via _is_post_norm flag computed once from the first layer's attribute signature. Changes: - Pre-attention layernorm: pre-norm reads from input_layernorm; post-norm sends the raw residual stream into qkv_proj (no pre-attn norm). - qk_norm: if the attention module exposes _apply_qk_norm (OLMo-2), call it between qkv_proj and rotary_emb. - Post-attention layernorm: pre-norm uses the LLaMA-style residual+norm- in-one-call; post-norm applies post_attention_layernorm to the attn output, then adds residual. - MLP: pre-norm path is unchanged; post-norm adds post_feedforward_layernorm after MLP and adds residual explicitly. Issue vllm-project#21.
jxmorris12
added a commit
to jxmorris12/vllm
that referenced
this pull request
Apr 21, 2026
* [Bugfix] fix DeepSeek R1 with CUTLASS MLA Broken on B200 (#33637)
Signed-off-by: chaunceyjiang <chaunceyjiang@gmail.com>
* [release] Minor fixes to release annotation (#33849)
Signed-off-by: Kevin H. Luu <khluu000@gmail.com>
* [CI][Bugfix]: return McpCall for built-in MCP tools in non-streaming mode (#32762)
Signed-off-by: Andreas Karatzas <akaratza@amd.com>
* Revert "[Attention][FA3] Update FA3 to include new swizzle optimization" (#33841)
* [Minor] Include `StreamingInput` in inputs package (#33856)
Signed-off-by: Nick Hill <nickhill123@gmail.com>
* [docs] fix unintentional misspellings (#33863)
Signed-off-by: rinbaro <ilgomishra@gmail.com>
* [CI][AMD][BugFix] Ensure VLLM_ROCM_USE_AITER is set so test_rocm_aiter_topk.py can run correctly (#33840)
Signed-off-by: Randall Smith <Randall.Smith@amd.com>
* [2/N] move responses/serving _make_response_output_items logic to parser (#33281)
Signed-off-by: Andrew Xia <axia@fb.com>
Signed-off-by: Andrew Xia <axia@meta.com>
Co-authored-by: Andrew Xia <axia@fb.com>
* [CI/Build] Parallelize CPU CI tests (#33778)
Signed-off-by: jiang1.li <jiang1.li@intel.com>
* [Bugfix] Fix ScoreMultiModalParam multi-document scoring returning single result (#33837)
Signed-off-by: Andreas Karatzas <akaratza@amd.com>
Signed-off-by: wang.yuqi <yuqi.wang@daocloud.io>
Co-authored-by: wang.yuqi <yuqi.wang@daocloud.io>
* [CPU][BugFix] Allow w8a8 oneDNN quantized matmul to support 3D inputs (#33727)
Signed-off-by: Fadi Arafeh <fadi.arafeh@arm.com>
* [CI/Build] Fix CPU CI test case title (#33870)
Signed-off-by: jiang1.li <jiang1.li@intel.com>
* [Perf] Optimize the performance of structured output + reasoning (#33557)
Signed-off-by: chaunceyjiang <chaunceyjiang@gmail.com>
* [KV Connector][Metrics] Do not count local prefix cache hits in connector queries (#30522)
Signed-off-by: Mark McLoughlin <markmc@redhat.com>
* [Bugfix] Kimi-K2 grouped_topk usage for Flashinfer monolithic kernels. (#33858)
Signed-off-by: Pavani Majety <pmajety@nvidia.com>
* [Refactor] Move `task` outside of `PoolingParams.verify` (#33796)
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
Signed-off-by: wang.yuqi <yuqi.wang@daocloud.io>
Co-authored-by: wang.yuqi <yuqi.wang@daocloud.io>
* [ROCm][Bugfix][CI] Fix hybrid models and their tests (Mamba/Jamba/Bamba) (#32710)
Signed-off-by: Andreas Karatzas <akaratza@amd.com>
Signed-off-by: Matthew Wong <Matthew.Wong2@amd.com>
Co-authored-by: Matthew Wong <Matthew.Wong2@amd.com>
* Enable Cross layers KV cache layout at NIXL Connector V2 (#33339)
Signed-off-by: Liran Schour <lirans@il.ibm.com>
Signed-off-by: liranschour <liranschour@users.noreply.github.com>
Co-authored-by: Or Ozeri <or@ozery.com>
Co-authored-by: Nicolò Lucchesi <nicolo.lucchesi@gmail.com>
Co-authored-by: Nicolò Lucchesi <nlucches@redhat.com>
* [perf] Integrate flashinfer concat_mla_k (#31171)
* [Bugfix] Fix Kimi-K2.5 NVFP4 checkpoints weight loading (#33876)
Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
* [Refactor] Clean up input preprocessing (#33687)
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
* [Bugfix] Fix corner case of sparse embedding (#33886)
Signed-off-by: wang.yuqi <yuqi.wang@daocloud.io>
* [Docs] Add bart-plugin to docs (#33905)
Signed-off-by: NickLucche <nlucches@redhat.com>
* [Bugfix] Fix step3p5 parser when using mtp (#33690)
Signed-off-by: mariohong <mariohong128@gmail.com>
* [Feat][RL][1/2] Native Weight Syncing API: NCCL (#31943)
Signed-off-by: ahao-anyscale <ahao@anyscale.com>
Signed-off-by: Aaron Hao <ahao@anyscale.com>
Co-authored-by: SumanthRH <sumanthrh99@gmail.com>
* [BugFix] Fix LoRA Fp8 (#33879)
Signed-off-by: Daniel Serebrenik <daserebrenik@nvidia.com>
* [Spec Decode] Unified Parallel Drafting (#32887)
Signed-off-by: Benjamin Chislett <bchislett@nvidia.com>
* [Misc] Add debug logs (#33931)
Signed-off-by: NickLucche <nlucches@redhat.com>
* [Bugfix] Fix swapped engine_ids in NIXL Llama 4 local attention path (#33795)
Signed-off-by: Yoray Zack <yorayz@nvidia.com>
* [Moe Refactor] Make Inplace Flag for FusedMoEModularKernel part of the constructor (#33375)
Signed-off-by: Bill Nell <bnell@redhat.com>
Co-authored-by: Robert Shaw <114415538+robertgshaw2-redhat@users.noreply.github.com>
* [Models] Consolidate Deepseek-OCR2 processor (#33909)
Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
* [Bugfix] Suppress non-TTY color output on the process name part of the log (#29714)
Signed-off-by: Tsukasa OI <floss_llm@irq.a4lg.com>
* Fix tokenizer test for renamed attr on Transformers v5 (#33902)
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
* [Misc] Rename `translations` to `speech_to_text` for OAI serving component (#33904)
Signed-off-by: NickLucche <nlucches@redhat.com>
* [Bugfix] Fix DSV3.2 NVFP4 (#33932)
Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>
* [Bugfix] Make MM batching more robust (#33817)
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
* [Minor] Sort safetensors files to ensure deterministic loading order (#33491)
Signed-off-by: Lihao Ran <imlihao.ran@gmail.com>
Signed-off-by: mgoin <mgoin64@gmail.com>
Co-authored-by: mgoin <mgoin64@gmail.com>
* Adds padding and perf improvements to wvSplitK_fp8 (#33527)
Signed-off-by: Hashem Hashemi <hashem.hashemi@amd.com>
* [Bugfix] Fix DeepSeek v3.2 tokenizer outputting None issue (#33832)
Signed-off-by: wzhao18 <wzhao18.sz@gmail.com>
* [Feature] OTEL tracing during loading (#31162)
* [Perf] Disable clean_logits in deepgemm fp8_mqa_logits kernel (#33568)
* [Docs] Add reo analytics (#33957)
Signed-off-by: simon-mo <simon.mo@hey.com>
* fix(ROCm): Make flash_attn import optional in MLA attention (#33511)
Signed-off-by: rabi <ramishra@redhat.com>
* feat(frontend): early-fail tokenization guard for user requests (#31366)
Signed-off-by: limingliang <limingliang@stepfun.com>
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
Co-authored-by: limingliang <limingliang@stepfun.com>
Co-authored-by: DarkLight1337 <tlleungac@connect.ust.hk>
* [Misc] Update code for encoder-decoder models (#33900)
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
* [CPU] Add BF16 Kernel type for s390x (#33788)
Signed-off-by: Rehan Khan <Rehan.Khan7@ibm.com>
* [XPU][4/N] add mxfp4 moe model support (#33679)
Signed-off-by: Kunshang Ji <kunshang.ji@intel.com>
* [XPU]Replace pip in docker.xpu with uv pip (#31112)
Signed-off-by: sihao.li <sihao.li@intel.com>
* Onboard voyage-4-nano (#33720)
Signed-off-by: Chengcheng Pei <chengchengpei@outlook.com>
Signed-off-by: chengchengpei <5881383+chengchengpei@users.noreply.github.com>
Co-authored-by: chengchengpei <5881383+chengchengpei@users.noreply.github.com>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
* [cpu][performance] CPU Paged Attention NEON BFMMLA BF16 Implementation (#32263)
Signed-off-by: Gassan <gassan.salama@arm.com>
* Fix `main` pre-commit (#33975)
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
* support view_from_cpu_tensor on XPU (#33868)
Signed-off-by: Xinyu Chen <xinyu1.chen@intel.com>
* Consolidate and fix forbidden import `pre-commit` checks (#33982)
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
* [PaddleOCR-VL] Add BC for transformers 5.0 config (#33976)
Signed-off-by: zhangyue66 <zhangyue66@baidu.com>
* Bump HF Hub client to get bug fix (#33984)
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
* [CPU][BugFix] Fix loading of w8a8int models with bias (#33582)
Signed-off-by: Fadi Arafeh <fadi.arafeh@arm.com>
* [torch.compile] Reorganize vllm/compilation and tests/compile (0/N for vLLM IR) (#33731)
Signed-off-by: Luka Govedič <lgovedic@redhat.com>
Signed-off-by: ProExpertProg <luka.govedic@gmail.com>
Signed-off-by: Luka Govedič <ProExpertProg@users.noreply.github.com>
* [Bugfix][Model] Support LoRA on Qwen3 Output Embedding (#29816)
Signed-off-by: kurt <kurt@thinkingmachines.ai>
* [Docs] Improve documentation (#33799)
Co-authored-by: Soren Dreano <soren@numind.ai>
Co-authored-by: Wentao Ye <44945378+yewentao256@users.noreply.github.com>
* Update `WeightTransferConfig` to be more standard like the others (#33989)
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
* [Bugfix] Fix models and tests for transformers v5 (#33977)
Signed-off-by: raushan <raushan@huggingface.co>
Signed-off-by: Raushan Turganbay <raushan.turganbay@alumni.nu.edu.kz>
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Co-authored-by: Cyrus Leung <cyrus.tl.leung@gmail.com>
Co-authored-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
* [FIX] guidance: use max(vocab_size, len(tokenizer)) for n_vocab (#33509)
Signed-off-by: Frederic Odermatt <frederic.odermatt@44ai.ch>
* [ROCm][AITER] Fix AITER import regression for explicit backend selection (#33749)
Signed-off-by: Andreas Karatzas <akaratza@amd.com>
* [Docs] Add sections on process architecture and minimum CPU resources (#33940)
It seems users can be confused about vLLM's performance when running
with very small amounts of CPU cores available. We are missing a clear
overview of what vLLM's process architecture is, so I added this along with
some diagrams in arch_overview.md, and included a section on CPU resource
recommendations in optimization.md
Signed-off-by: mgoin <mgoin64@gmail.com>
* [Model] Support MiniCPM-o 4.5 (#33431)
Signed-off-by: caitianchi <caitianchi@modelbest.cn>
Signed-off-by: tc-mb <caitianchi@modelbest.cn>
Co-authored-by: mslv <mslv@baai.ac.cn>
* [Refactor] Consolidate sequence normalization and enc-dec parsing (#33928)
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
* [XPU][5/N] add wna16 xpu kernel (#33973)
Signed-off-by: Zhu, Zufang <zufang.zhu@intel.com>
* [Docs] Update link to Benchmark CLI documentation (#33254)
Signed-off-by: Eldar Kurtić <8884008+eldarkurtic@users.noreply.github.com>
* [Bugfix] Fix the issue where tool calling does not work when using fast detokenization with dsv32 (#33964)
Signed-off-by: chaunceyjiang <chaunceyjiang@gmail.com>
* [Log] Optimize duplicate startup log (#33944)
Signed-off-by: yewentao256 <zhyanwentao@126.com>
* [KV Connector] Add missing method overrides to MultiConnector (#33292)
Signed-off-by: Seiji Eicher <seiji@anyscale.com>
* [DOC] [ROCm] Update docker deployment doc (#33971)
Signed-off-by: vllmellm <vllm.ellm@embeddedllm.com>
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Co-authored-by: TJian <tunjian.tan@embeddedllm.com>
Co-authored-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
* [Model Runner V2] support apply penalty for spec decode (#33251)
Signed-off-by: zhuhaoran <zhuhaoran.zhr@alibaba-inc.com>
* [Refactor] Remove align block size logic in `moe_permute` (#33449)
Signed-off-by: yewentao256 <zhyanwentao@126.com>
* [Rocm][Bugfix] Fix dtype not same for gemm_a4w4 op (#33734)
Signed-off-by: charlifu <charlifu@amd.com>
* [Bugfix] Fix no attribute error of SharedFusedMoE (DeepSeek-V3.1 as test model) (#33993)
Signed-off-by: xuebwang-amd <xuebwang@amd.com>
* [Fix] Fix `logprobs=0` handling for `/inference/v1/generate` endpoint (#34010)
Signed-off-by: SumanthRH <sumanthrh99@gmail.com>
* Fix RoutingMethodType logic (#33919)
Signed-off-by: Dimitrios Bariamis <12195802+dbari@users.noreply.github.com>
Signed-off-by: mgoin <mgoin64@gmail.com>
Co-authored-by: Dimitrios Bariamis <12195802+dbari@users.noreply.github.com>
Co-authored-by: mgoin <mgoin64@gmail.com>
* [bugfix] [ROCm] Fix premature CUDA initialization in platform detection (#33941)
Signed-off-by: Kourosh Hakhamaneshi <kourosh@anyscale.com>
* [Feat][RL] Pause and Resume with keep requests for single engine (#32351)
Signed-off-by: ahao-anyscale <ahao@anyscale.com>
Signed-off-by: Aaron Hao <ahao@anyscale.com>
Co-authored-by: Robert Shaw <114415538+robertgshaw2-redhat@users.noreply.github.com>
* [Bugfix] Fix QK Norm+RoPE fusion pattern matching on B200+FP8 (#33967)
Signed-off-by: Ikenna <ikennachifo@gmail.com>
Co-authored-by: Luka Govedič <ProExpertProg@users.noreply.github.com>
* [Bugfix] Fix Whisper tokenization (#34011)
Signed-off-by: NickLucche <nlucches@redhat.com>
* [CI][AMD]Bugfix] Check that model_config is not None in enable_norm_pad_fusion (#34007)
Signed-off-by: Randall Smith <Randall.Smith@amd.com>
* [Bugfix] Fix _fused_moe_lora_expand signature mismatch (#33821)
Signed-off-by: Xin Yang <xyangx@amazon.com>
* [Misc] Add backward-compatible import aliases for renamed translations module (#34015)
Signed-off-by: Kourosh Hakhamaneshi <kourosh@anyscale.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
* [ModelRunner V2] Revert token rank comparison difference for now (#34017)
Signed-off-by: Nick Hill <nickhill123@gmail.com>
* fix description in plugin_system.md (#33999)
* [Revert] Add util `handle_deprecated` back (#33998)
Signed-off-by: yewentao256 <zhyanwentao@126.com>
* [Kernel] Add enable_sm120_or_later for SM121 (DGX Spark) CUTLASS support (#33517)
Signed-off-by: code4me2 <velvetmoon222999@gmail.com>
* [Misc] Make `PlaceholderRange.get_num_embeds` a method (#34035)
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
* [ROCm][CI] Pinning lm-eval version to resolve multi-modal small eval bug (#34038)
Signed-off-by: Andreas Karatzas <akaratza@amd.com>
* Fix spelling errors (#33978)
* [Misc] Simplify `get_max_tokens` (#34036)
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
* [CI][Build] Pin grpcio-tools==1.78.0 (#34048)
Signed-off-by: wang.yuqi <noooop@126.com>
Co-authored-by: Cyrus Leung <cyrus.tl.leung@gmail.com>
* [Renderer] Define `render_cmpl` and `render_chat` (#34039)
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
* [Kernel] Add KernelConfig flag to enable/disable FlashInfer autotune (#34006)
Signed-off-by: Mohammad Miadh Angkad <176301910+mmangkad@users.noreply.github.com>
Signed-off-by: Luka Govedič <ProExpertProg@users.noreply.github.com>
Co-authored-by: Luka Govedič <ProExpertProg@users.noreply.github.com>
* [torch.compile] Stop compiling identical artifacts (#34003)
Signed-off-by: Richard Zou <zou3519@gmail.com>
* Enable Eagle3 speculative decoding for Mistral3ForConditionalGeneration to support eagle3 (#33939)
Signed-off-by: Akintunde Oladipo <akintunde.oladipo@servicenow.com>
Signed-off-by: TundeAtSN <akintunde.oladipo@servicenow.com>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
* [Frontend]Add support for transcriptions and translations to run_batch (#33934)
Signed-off-by: Pooya Davoodi <pooya.davoodi@parasail.io>
Signed-off-by: Cyrus Leung <cyrus.tl.leung@gmail.com>
Co-authored-by: Cyrus Leung <cyrus.tl.leung@gmail.com>
* [Model] Enable Step3p5ForCausalLM testing (#33755)
Signed-off-by: Jee Jee Li <pandaleefree@gmail.com>
* [PluggableLayer][3/N] Apply PluggableLayer to mamba layers. (#33660)
Signed-off-by: whx-sjtu <2952154980@qq.com>
* move checks out of `unified_kv_cache_update` custom op (#33943)
Signed-off-by: Rohan138 <rohanpotdar138@gmail.com>
* Update DeepGEMM version pin in Dockerfile to match #32479 (#33935)
Signed-off-by: Zifei Tong <zifeitong@gmail.com>
Signed-off-by: Wentao Ye <44945378+yewentao256@users.noreply.github.com>
Co-authored-by: Wentao Ye <44945378+yewentao256@users.noreply.github.com>
* Make directory exist ok for ray spinning up multiple replicas on a single instance (#33604)
Signed-off-by: Jiang Wu <jwu@cclgroup.com>
* Perf tuning and expansion of cases covered for wvSplitKrc (#33493)
Signed-off-by: Hashem Hashemi <hashem.hashemi@amd.com>
* [Doc] Fix run_batch docs (#34056)
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
* [CI/Build] Skip GCS test (#34057)
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
* [ROCm][Bugfix] fix act_quant_fusion module import error (#34069)
Signed-off-by: Andreas Karatzas <akaratza@amd.com>
* [Perf] Simplify DeepseekV32 tokenizer, ensure fast detokenization used (#33855)
Signed-off-by: Nick Hill <nickhill123@gmail.com>
* [ROCm] [CI] Reduce Resource of two test groups (#34059)
Signed-off-by: tjtanaa <tunjian.tan@embeddedllm.com>
* Add embedding input functionality for disabled modalities [remake] (#32493)
Signed-off-by: Reagan Lee <“reaganjlee@gmail.com”>
Signed-off-by: Reagan Lee <reaganjlee@gmail.com>
Signed-off-by: Reagan Lee <96998476+reaganjlee@users.noreply.github.com>
Co-authored-by: Reagan Lee <“reaganjlee@gmail.com”>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
* [Revert] Fix performance regression for GLM-4.7-GPTQ decode and MTP acceptance rate (#33771)
Signed-off-by: aabbccddwasd <aabbccddwasd@qq.com>
* [BugFix] Change support no act and mul for marlin (#34088)
Signed-off-by: Tomer Natan <tbarnatan@computelab-frontend-8.nvidia.com>
Co-authored-by: Tomer Natan <tbarnatan@computelab-frontend-8.nvidia.com>
* [torch.compile] Add an option to force-enable the MOE cold start optimization (#33735)
Signed-off-by: Richard Zou <zou3519@gmail.com>
* glm 4.6 fused tuned inference config for B200 (#32958)
* Add support for ModelOpt MXFP8 dense models (#33786)
Signed-off-by: Daniel Serebrenik <daserebrenik@nvidia.com>
* [Release 2.10] Update to Torch 2.10 - final release (#30525)
* [bug-fix] supported_tasks is breaking backward compatibility at init_app_state (#34027)
Signed-off-by: Kourosh Hakhamaneshi <kourosh@anyscale.com>
Signed-off-by: kourosh hakhamaneshi <31483498+kouroshHakha@users.noreply.github.com>
Co-authored-by: Cyrus Leung <cyrus.tl.leung@gmail.com>
* [Tiny] Rename encoder budget file to more specific name (#34103)
Signed-off-by: Reagan Lee <“reaganjlee@gmail.com”>
Co-authored-by: Reagan Lee <“reaganjlee@gmail.com”>
* [Frontend][last/5] Make pooling entrypoints request schema consensus. (#31127)
Signed-off-by: wang.yuqi <yuqi.wang@daocloud.io>
* [BugFix] Fix `fastsafetensors` TP all procs using all GPUs (#34070)
Signed-off-by: Nick Hill <nickhill123@gmail.com>
Co-authored-by: Cyrus Leung <tlleungac@connect.ust.hk>
* fix(cpu): fix mla_decode compilation on x86 without AVX512 (#34052)
Signed-off-by: ihb2032 <hebome@foxmail.com>
Co-authored-by: root <root@LAPTOP-FKNHV411.localdomain>
* [Model] GLM adaptation (#34124)
* [CI] Remove empty image_size_factors for fuyu, glm4_1v, glm_ocr (#34107)
Signed-off-by: Andreas Karatzas <akaratza@amd.com>
* [ASR] Fix audio benchmark and add RTFx metric (#32300)
Signed-off-by: Ekagra Ranjan <3116519+ekagra-ranjan@users.noreply.github.com>
Co-authored-by: Nicolò Lucchesi <nicolo.lucchesi@gmail.com>
* [Fix] [CPU Backend] : Prepack weights for w8a8 oneDNN matmul (#33901)
Signed-off-by: nikhil-arm <nikhil.gupta2@arm.com>
* [XPU][6/N] add xpu scaled_mm kernel (#34117)
Signed-off-by: Zhu, Zufang <zufang.zhu@intel.com>
* [MODEL] Adding Support for Qwen3.5 Models (#34110)
Signed-off-by: JJJYmmm <1650675829@qq.com>
Signed-off-by: JJJYmmm <92386084+JJJYmmm@users.noreply.github.com>
Signed-off-by: Roger Wang <hey@rogerw.io>
Co-authored-by: wulipc <wulipc@users.noreply.github.com>
Co-authored-by: ywang96 <ywang96@users.noreply.github.com>
Co-authored-by: Isotr0py <Isotr0py@users.noreply.github.com>
Co-authored-by: Isotr0py <2037008807@qq.com>
Co-authored-by: Roger Wang <hey@rogerw.io>
* [Misc] Fix up attention benchmarks (#33810)
Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>
Co-authored-by: Matthew Bonanni <mbonanni@redhat.com>
* [UX] Add `--language-model-only` for hybrid models (#34120)
Signed-off-by: Roger Wang <hey@rogerw.io>
* [CI][torch.compile] Fix incorrect filtering for E2E fusion tests on B200 (#34031)
Signed-off-by: Luka Govedič <lgovedic@redhat.com>
* Add NUMA Core binding in nixl_connector for CPU xPyD (#32365)
Signed-off-by: Hongming Zheng <hongming.zheng@intel.com>
Signed-off-by: ZhengHongming888 <hongming.zheng@intel.com>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
* [Kernel] FlashInfer: switch allreduce fusion to unified API (#33985)
Signed-off-by: Mohammad Miadh Angkad <176301910+mmangkad@users.noreply.github.com>
* [Bugfix] Fix shared expert input for latent MoE in EP+DP (Nemotron-H) (#34087)
Signed-off-by: Tomer Natan <tbarnatan@nvidia.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
* [Kernel] use flashinfer for gdn prefill (#32846)
Signed-off-by: zjy0516 <riverclouds.zhu@qq.com>
* [Bugfix] Avoid duplicate k-proj weight emission in helper (#34142)
Signed-off-by: Artus KG <artuskg@gmail.com>
* [Bugfix] Voxtral prompt/audio placeholder alignment (#34140)
Signed-off-by: Artus KG <artuskg@gmail.com>
* [ROCm] update triton branch to support gpt-oss models for gfx11xx devices (#34032)
Signed-off-by: Hongxia Yang <hongxia.yang@amd.com>
* [torch.compile][Fusion] Fix attention fusion pass removing kv_udpate op. (#33945)
Signed-off-by: charlifu <charlifu@amd.com>
* [ModelRunner V2][BugFix] Fix `max_query_len` calculation (#34167)
Signed-off-by: Nick Hill <nickhill123@gmail.com>
* [Doc] Add DCP support to attention backend doc (#33936)
* [Bugfix][ROCm][GPT-OSS] Use old triton_kernels implementation on ROCm if the new API is not available (#34153)
Signed-off-by: Gregory Shtrasberg <Gregory.Shtrasberg@amd.com>
* [structured output] validate unsupported json features first (#33233)
Signed-off-by: Andy Xie <andy.xning@gmail.com>
Co-authored-by: Chauncey <chaunceyjiang@gmail.com>
Co-authored-by: Russell Bryant <rbryant@redhat.com>
* [LMCache] Token Base IPC API (#34175)
Signed-off-by: Oasis-Git <ayw.sirius19@gmail.com>
* [Bugfix] Adopt `ChunkGatedDeltaRule` for Qwen3.5 (#34198)
Signed-off-by: Roger Wang <hey@rogerw.io>
* [ROCm][Bugfix] Resolve Dynamo tracing crash from amdsmi calls in on_gfx* arch detection (#34108)
Signed-off-by: Andreas Karatzas <akaratza@amd.com>
* [Bugfix][Core] Fix CPU memory leak from Request reference cycle in prefix caching (#34183)
Signed-off-by: Roger Wang <hey@rogerw.io>
* [Doc] Update usage of `--limit-mm-per-prompt` (#34148)
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
* [CI/Build] Relax `test_mcp_tool_call` (#34204)
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
* [Bugfix] Fix DP Attention Padding in Dummy Run (#34187)
Signed-off-by: Benjamin Chislett <bchislett@nvidia.com>
Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
Co-authored-by: Benjamin Chislett <bchislett@nvidia.com>
* [Bugfix] Add `--trust-remote-code` to dataset bench args (#34208)
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
* [responsesAPI] fix simpleContext streaming output_messages (#34188)
Signed-off-by: Andrew Xia <axia@meta.com>
Signed-off-by: Andrew Xia <axia@fb.com>
Co-authored-by: Andrew Xia <axia@fb.com>
* [Bugfix] Sort hf_weights_files in fastsafetensors_weights_iterator to match #33491 (#34190)
Signed-off-by: Balaxxe <136368465+jaim12005@users.noreply.github.com>
* [Frontend][CI] Consolidate instrumentator entrypoints (#34123)
Signed-off-by: wang.yuqi <yuqi.wang@daocloud.io>
* [BugFix] Avoid prefix cache hit in the same schedule step for mamba layers (#29387)
Signed-off-by: Chen Zhang <zhangch99@outlook.com>
* [Perf] Optimize detokenizer python logic (#32975)
Signed-off-by: yewentao256 <zhyanwentao@126.com>
Signed-off-by: Wentao Ye <44945378+yewentao256@users.noreply.github.com>
Co-authored-by: Nick Hill <nhill@redhat.com>
* Revert #34208 (#34216)
* [Bugfix] Fix memory inconsistency in cross-process shared memory (#32022)
Signed-off-by: Zetong Li <slippersss@126.com>
* [Bugfix] Fix `--trust-remote-code` conflict (#34218)
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
* [Docs] Fix format error in KV load failure recovery doc (#34137)
Signed-off-by: Jaebok Lee <jaebok9541@naver.com>
* [Bugfix] Fix FI kernel`chunk_gated_delta_rule` output shape for Qwen3.5 (#34219)
Signed-off-by: Roger Wang <hey@rogerw.io>
* Add flagos in MiniCPM-o (#34126)
Signed-off-by: tc-mb <caitianchi@modelbest.cn>
Signed-off-by: Vincent-Xiao <vincent.xiao.me@gmail.com>
Co-authored-by: Vincent-Xiao <vincent.xiao.me@gmail.com>
* [Misc] allow specify is_mm_prefix_lm in hf_config (#34215)
* Stop testing for slow tokenizers as they will not exist soon (#34235)
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
* [V1][BugFix] Fix EAGLE3 encoder cache miss with disable_chunked_mm_input (#34220)
Signed-off-by: KrxGu <krishom70@gmail.com>
* Bump `mamba-ssm` version in CI for Transformers v5 compatibility (#34233)
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
* add --insecure arg to the vllm bench to skip TLS (#34026)
Signed-off-by: Fan Yang <yan9fan@meta.com>
Co-authored-by: Fan Yang <yan9fan@meta.com>
* Support benchmarking of Geospatial models (#33922)
Signed-off-by: Michele Gazzetti <michele.gazzetti1@ibm.com>
* [ROCm][Quantization] GPT_OSS in amd-quark format model loading and emulations (#29008)
Signed-off-by: xuebwang-amd <xuebwang@amd.com>
Signed-off-by: Robert Shaw <robshaw@redhat.com>
Co-authored-by: Robert Shaw <robshaw@redhat.com>
Co-authored-by: Robert Shaw <114415538+robertgshaw2-redhat@users.noreply.github.com>
* [compile] Enable AOT compile with 2.10 in trunk. (#34155)
Signed-off-by: Zhengxu Chen <zhxchen17@meta.com>
* [Perf][Kernel] Add faster topKperRow decode kernel for DeepSeek-V3.2 sparse attention (#33680)
Signed-off-by: LopezCastroRoberto <rocastro@redhat.com>
Signed-off-by: Roberto L. Castro <38211239+LopezCastroRoberto@users.noreply.github.com>
Co-authored-by: Claude Sonnet 4.5 <noreply@anthropic.com>
* [Core][BugFix] Fix PP KV cache sharding memory validation (#33698)
Signed-off-by: junuxyz <216036880+junuxyz@users.noreply.github.com>
* [BUGFIX] Fix accuracy bugs in Qwen3-Next MTP (#34077)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
* [Docs] Speed up build environment set-up (#34240)
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
* [Model Runner V2] Use pinned memory for write_contents (#34222)
Signed-off-by: Woosuk Kwon <woosuk@inferact.ai>
* Minor cleanup for Voxtral (#34247)
Signed-off-by: Andy Lo <andy@mistral.ai>
* [UX nit] Fix non-default api_server_count message (#34152)
Signed-off-by: mgoin <mgoin64@gmail.com>
* [Misc] Introduce ec_both role EC (encoder cache) connector (#34182)
Signed-off-by: Qi Wang <qiwa@nvidia.com>
* Convert online APIs to use Renderer (#34084)
Signed-off-by: Reagan Lee <“reaganjlee@gmail.com”>
Co-authored-by: Reagan Lee <“reaganjlee@gmail.com”>
* [Bugfix] Fix weights offloading for sleep mode (#32947)
Signed-off-by: Jarno Seppänen <jseppanen@nvidia.com>
Co-authored-by: Tyler Michael Smith <tyler@neuralmagic.com>
* [Benchmarks] Fix attention benchmark smoke test (#34269)
Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>
* [Bugfix] Fix mamba cache dtype for Qwen3.5 (#34200)
Signed-off-by: Roger Wang <hey@rogerw.io>
* [SM100] Resubmit FMHA FP8 prefill for MLA (#31195)
Signed-off-by: Pavani Majety <pmajety@nvidia.com>
* [Feature] Warn about unrecognized environment variables (#33581)
Signed-off-by: Gregory Shtrasberg <Gregory.Shtrasberg@amd.com>
* [Perf] Move eplb rebalance algo to async thread (#30888)
Signed-off-by: ilmarkov <markovilya197@gmail.com>
Signed-off-by: Tyler Michael Smith <tlrmchlsmth@gmail.com>
Co-authored-by: Tyler Michael Smith <tyler@neuralmagic.com>
Co-authored-by: Tyler Michael Smith <tlrmchlsmth@gmail.com>
* [BugFix] Fix async EPLB hang with DeepEP LL all2all backend (#32860)
Signed-off-by: ilmarkov <markovilya197@gmail.com>
* [Misc][Spec Decode] support different load config for draft model (#34022)
Signed-off-by: zzhengkai <zzhengkai@devgpu049.ldc1.facebook.com>
Co-authored-by: zzhengkai <zzhengkai@devgpu049.ldc1.facebook.com>
* [torch.compile] Disable recursive pre_grad_passes (#34092)
Signed-off-by: Richard Zou <zou3519@gmail.com>
* [Misc] Add pre-commit hook to catch boolean ops in with-statements (#34271)
Signed-off-by: Tyler Michael Smith <tlrmchlsmth@gmail.com>
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
* [CI] Add pip caching to cleanup_pr_body workflow (#32979)
Signed-off-by: 7. Sun <jhao.sun@gmail.com>
* [MoE Refactor] Introduce MoERunner abstraction and move execution logic from FusedMoE to DefaultMoERunner (#32344)
Signed-off-by: Bill Nell <bnell@redhat.com>
* [ROCm][CI] Fix test_sequence_parallel.py location in AMD CI pipeline (#34280)
Signed-off-by: Micah Williamson <micah.williamson@amd.com>
* [Misc] Add run one batch script that supports profiling (#32968)
Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
* [Bugfix] Fix Worker.load_model context-manager composition for sleep mode (#34021)
Signed-off-by: tianshu.yu <tianshuyu.formal@gmail.com>
* [Redo] Add `--trust-remote-code` to dataset bench args (#34251)
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
* [torch.compile] Stop doing unnecessary FakeTensorProp in PiecewiseCompileInterpreter (#34093)
Signed-off-by: Richard Zou <zou3519@gmail.com>
* [WideEP] Fix nvfp4 DeepEP High Throughput All2All backend (#33738)
Signed-off-by: Tyler Michael Smith <tlrmchlsmth@gmail.com>
Co-authored-by: Robert Shaw <114415538+robertgshaw2-redhat@users.noreply.github.com>
* [Misc] Clean up validation logic in input processor (#34144)
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
* [Bugfix][DeepSeek-V3.2] fix fp8 kvcache type cast (#33884)
Signed-off-by: Kebe <mail@kebe7jun.com>
* [Kernel] Apply 256bit LDG/STG To Activation Kernels (#33022)
Signed-off-by: Dzerzhinsky <256908701+AstroVoyager7@users.noreply.github.com>
Signed-off-by: Дзержи́нский <256908701+AstroVoyager7@users.noreply.github.com>
Co-authored-by: Wentao Ye <44945378+yewentao256@users.noreply.github.com>
* [XPU][7/N] enable xpu fp8 moe (#34202)
Signed-off-by: Zhu, Zufang <zufang.zhu@intel.com>
* [Plugin] Simplify IO Processor Plugin interface (#34236)
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
* [Bugfix] Fix benchmark_moe.py inplace assertion with torch >= 2.9 (#34149)
Signed-off-by: Matthias Gehre <matthias.gehre@amd.com>
* Threshold fix wvSplitk for occasional CI fails (#34013)
Signed-off-by: Hashem Hashemi <hashem.hashemi@amd.com>
* [ModelBash][DSR1 NVFp4] Removed Bf16 Bias Cast (#34298)
Signed-off-by: Robert Shaw <robshaw@redhat.com>
Co-authored-by: Robert Shaw <robshaw@redhat.com>
* [Bugfix] Fix fused MoE IMA (sans chunking) by using int64 for strides (#34279)
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
* [Bugfix] Fix weight naming in Qwen3.5 (#34313)
Signed-off-by: Roger Wang <hey@rogerw.io>
* [CPU] Enable FP16 (Half dtype) support for s390x (#34116)
Signed-off-by: Rehan Khan <Rehan.Khan7@ibm.com>
* [model] support FunASR model (#33247)
Signed-off-by: zixiao <shunli.dsl@alibaba-inc.com>
Co-authored-by: zixiao <shunli.dsl@alibaba-inc.com>
* [XPU][9/N] clean up existing ipex code/doc (#34111)
Signed-off-by: Kunshang Ji <kunshang.ji@intel.com>
* [Chore] Move `BaseRenderer` to `base.py` (#34308)
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
* [torch.compile] Enable AR+rms fusion by default available for `-O2` (#34299)
Signed-off-by: Luka Govedič <lgovedic@redhat.com>
* [Misc] Bump `fastsafetensors` version for latest fixes (#34273)
Signed-off-by: Nick Hill <nickhill123@gmail.com>
* [Doc] Update Marlin support matrix for Turing (#34319)
Signed-off-by: Tianqi Ren <tianqi.r@outlook.com>
* [Frontend] Exploit tokenizers "new stream" in FastIncrementalDetokenizer (#34217)
Signed-off-by: Nick Hill <nickhill123@gmail.com>
* Patch protobuf for CVE-2026-0994 (#34253)
Signed-off-by: Seiji Eicher <seiji@anyscale.com>
Co-authored-by: Kevin H. Luu <khluu000@gmail.com>
* [Docs] Reduce time spent generating API docs (#34255)
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
* [Bugfix][CPU] Fix llama4 inference on CPU (#34321)
Signed-off-by: jiang1.li <jiang1.li@intel.com>
* Make Qwen3VL compatible with Transformers v5 (#34262)
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Signed-off-by: Roger Wang <hey@rogerw.io>
Co-authored-by: Roger Wang <hey@rogerw.io>
* Make JAIS compatible with Transformers v5 (#34264)
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
* [NVIDIA][test] Tests for flashinfer TRTLLM BF16 MoE (#33715)
Signed-off-by: Linda-Stadter <57756729+Linda-Stadter@users.noreply.github.com>
Co-authored-by: Pavani Majety <pmajety@nvidia.com>
* Responses harmony system message structured (#34268)
Signed-off-by: Adam Binford <adamq43@gmail.com>
* Reapply [Attention][FA3] Update FA3 to include new swizzle optimization (#34043)
Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
* Don't try and run GLM-ASR with remote code (#34352)
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
* [Bugfix]: Fix ROCm fusion attn test; use AttentionBackend utils to create kv cache (#33948)
Signed-off-by: Rohan138 <rohanpotdar138@gmail.com>
* [ROCm] [aiter] Split KV cache update for AiterFlashAttention (#33681)
Signed-off-by: kliuae <kuanfu.liu@embeddedllm.com>
* [Docs] Fix typo ("defult") and double spacing (#34348)
Signed-off-by: SorenDreano <71752785+SorenDreano@users.noreply.github.com>
Co-authored-by: Soren Dreano <soren@numind.ai>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
* [CI][BugFix] Fix silent failure in shellcheck hook and baseline exist… (#32458)
Signed-off-by: junuxyz <216036880+junuxyz@users.noreply.github.com>
* [Model Runner V2] Init cuda graph pool when necessary (#33217)
Signed-off-by: Xinyu Chen <xinyu1.chen@intel.com>
* [Multimodal] Expose `mm_processor_kwargs` for `DummyInputsBuilder` (#34330)
Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
* [Bugfix] fix default is_neox_style is True for deepseek (#34353)
Signed-off-by: dongxinyu03 <dongxinyu03@baidu.com>
* [Bugfix] Enable attn quantization of Llama-4 by correctly permuting scales for rope (int8, fp8) (#34243)
Signed-off-by: Your Name <you@example.com>
Co-authored-by: Your Name <you@example.com>
* [ROCm] [CI] fix test_unrecognized_env (#34350)
Signed-off-by: tjtanaa <tunjian.tan@embeddedllm.com>
* [GPT-OSS] Remove unnecessary contiguous (#34337)
Signed-off-by: elvischenv <219235043+elvischenv@users.noreply.github.com>
* Add cartridge (prefix) benchmark configs to CI workflows
The prefix_latency and prefix_throughput configs existed but weren't
being run by any workflow. Each benchmark workflow now runs both the
base and cartridge configs using the shared server support.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* update flashinfer
* update wheel
* update cuda and flashinfer
* downgrade
* update tests
* experimental: implement pipelining
* add pipeline test
* configure PR to actually run
* bugfix
* loosen TPOT threshold for catridge latency
* improve pipelining
* simplify pipelining impl
---------
Signed-off-by: chaunceyjiang <chaunceyjiang@gmail.com>
Signed-off-by: Kevin H. Luu <khluu000@gmail.com>
Signed-off-by: Andreas Karatzas <akaratza@amd.com>
Signed-off-by: Nick Hill <nickhill123@gmail.com>
Signed-off-by: rinbaro <ilgomishra@gmail.com>
Signed-off-by: Randall Smith <Randall.Smith@amd.com>
Signed-off-by: Andrew Xia <axia@fb.com>
Signed-off-by: Andrew Xia <axia@meta.com>
Signed-off-by: jiang1.li <jiang1.li@intel.com>
Signed-off-by: wang.yuqi <yuqi.wang@daocloud.io>
Signed-off-by: Fadi Arafeh <fadi.arafeh@arm.com>
Signed-off-by: Mark McLoughlin <markmc@redhat.com>
Signed-off-by: Pavani Majety <pmajety@nvidia.com>
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
Signed-off-by: Matthew Wong <Matthew.Wong2@amd.com>
Signed-off-by: Liran Schour <lirans@il.ibm.com>
Signed-off-by: liranschour <liranschour@users.noreply.github.com>
Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
Signed-off-by: NickLucche <nlucches@redhat.com>
Signed-off-by: mariohong <mariohong128@gmail.com>
Signed-off-by: ahao-anyscale <ahao@anyscale.com>
Signed-off-by: Aaron Hao <ahao@anyscale.com>
Signed-off-by: Daniel Serebrenik <daserebrenik@nvidia.com>
Signed-off-by: Benjamin Chislett <bchislett@nvidia.com>
Signed-off-by: Yoray Zack <yorayz@nvidia.com>
Signed-off-by: Bill Nell <bnell@redhat.com>
Signed-off-by: Tsukasa OI <floss_llm@irq.a4lg.com>
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>
Signed-off-by: Lihao Ran <imlihao.ran@gmail.com>
Signed-off-by: mgoin <mgoin64@gmail.com>
Signed-off-by: Hashem Hashemi <hashem.hashemi@amd.com>
Signed-off-by: wzhao18 <wzhao18.sz@gmail.com>
Signed-off-by: simon-mo <simon.mo@hey.com>
Signed-off-by: rabi <ramishra@redhat.com>
Signed-off-by: limingliang <limingliang@stepfun.com>
Signed-off-by: Rehan Khan <Rehan.Khan7@ibm.com>
Signed-off-by: Kunshang Ji <kunshang.ji@intel.com>
Signed-off-by: sihao.li <sihao.li@intel.com>
Signed-off-by: Chengcheng Pei <chengchengpei@outlook.com>
Signed-off-by: chengchengpei <5881383+chengchengpei@users.noreply.github.com>
Signed-off-by: Gassan <gassan.salama@arm.com>
Signed-off-by: Xinyu Chen <xinyu1.chen@intel.com>
Signed-off-by: zhangyue66 <zhangyue66@baidu.com>
Signed-off-by: Luka Govedič <lgovedic@redhat.com>
Signed-off-by: ProExpertProg <luka.govedic@gmail.com>
Signed-off-by: Luka Govedič <ProExpertProg@users.noreply.github.com>
Signed-off-by: kurt <kurt@thinkingmachines.ai>
Signed-off-by: raushan <raushan@huggingface.co>
Signed-off-by: Raushan Turganbay <raushan.turganbay@alumni.nu.edu.kz>
Signed-off-by: Frederic Odermatt <frederic.odermatt@44ai.ch>
Signed-off-by: caitianchi <caitianchi@modelbest.cn>
Signed-off-by: tc-mb <caitianchi@modelbest.cn>
Signed-off-by: Zhu, Zufang <zufang.zhu@intel.com>
Signed-off-by: Eldar Kurtić <8884008+eldarkurtic@users.noreply.github.com>
Signed-off-by: yewentao256 <zhyanwentao@126.com>
Signed-off-by: Seiji Eicher <seiji@anyscale.com>
Signed-off-by: vllmellm <vllm.ellm@embeddedllm.com>
Signed-off-by: zhuhaoran <zhuhaoran.zhr@alibaba-inc.com>
Signed-off-by: charlifu <charlifu@amd.com>
Signed-off-by: xuebwang-amd <xuebwang@amd.com>
Signed-off-by: SumanthRH <sumanthrh99@gmail.com>
Signed-off-by: Dimitrios Bariamis <12195802+dbari@users.noreply.github.com>
Signed-off-by: Kourosh Hakhamaneshi <kourosh@anyscale.com>
Signed-off-by: Ikenna <ikennachifo@gmail.com>
Signed-off-by: Xin Yang <xyangx@amazon.com>
Signed-off-by: code4me2 <velvetmoon222999@gmail.com>
Signed-off-by: wang.yuqi <noooop@126.com>
Signed-off-by: Mohammad Miadh Angkad <176301910+mmangkad@users.noreply.github.com>
Signed-off-by: Richard Zou <zou3519@gmail.com>
Signed-off-by: Akintunde Oladipo <akintunde.oladipo@servicenow.com>
Signed-off-by: TundeAtSN <akintunde.oladipo@servicenow.com>
Signed-off-by: Pooya Davoodi <pooya.davoodi@parasail.io>
Signed-off-by: Cyrus Leung <cyrus.tl.leung@gmail.com>
Signed-off-by: Jee Jee Li <pandaleefree@gmail.com>
Signed-off-by: whx-sjtu <2952154980@qq.com>
Signed-off-by: Rohan138 <rohanpotdar138@gmail.com>
Signed-off-by: Zifei Tong <zifeitong@gmail.com>
Signed-off-by: Wentao Ye <44945378+yewentao256@users.noreply.github.com>
Signed-off-by: Jiang Wu <jwu@cclgroup.com>
Signed-off-by: tjtanaa <tunjian.tan@embeddedllm.com>
Signed-off-by: Reagan Lee <“reaganjlee@gmail.com”>
Signed-off-by: Reagan Lee <reaganjlee@gmail.com>
Signed-off-by: Reagan Lee <96998476+reaganjlee@users.noreply.github.com>
Signed-off-by: aabbccddwasd <aabbccddwasd@qq.com>
Signed-off-by: Tomer Natan <tbarnatan@computelab-frontend-8.nvidia.com>
Signed-off-by: kourosh hakhamaneshi <31483498+kouroshHakha@users.noreply.github.com>
Signed-off-by: ihb2032 <hebome@foxmail.com>
Signed-off-by: Ekagra Ranjan <3116519+ekagra-ranjan@users.noreply.github.com>
Signed-off-by: nikhil-arm <nikhil.gupta2@arm.com>
Signed-off-by: JJJYmmm <1650675829@qq.com>
Signed-off-by: JJJYmmm <92386084+JJJYmmm@users.noreply.github.com>
Signed-off-by: Roger Wang <hey@rogerw.io>
Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
Signed-off-by: Hongming Zheng <hongming.zheng@intel.com>
Signed-off-by: ZhengHongming888 <hongming.zheng@intel.com>
Signed-off-by: Tomer Natan <tbarnatan@nvidia.com>
Signed-off-by: zjy0516 <riverclouds.zhu@qq.com>
Signed-off-by: Artus KG <artuskg@gmail.com>
Signed-off-by: Hongxia Yang <hongxia.yang@amd.com>
Signed-off-by: Gregory Shtrasberg <Gregory.Shtrasberg@amd.com>
Signed-off-by: Andy Xie <andy.xning@gmail.com>
Signed-off-by: Oasis-Git <ayw.sirius19@gmail.com>
Signed-off-by: Balaxxe <136368465+jaim12005@users.noreply.github.com>
Signed-off-by: Chen Zhang <zhangch99@outlook.com>
Signed-off-by: Zetong Li <slippersss@126.com>
Signed-off-by: Jaebok Lee <jaebok9541@naver.com>
Signed-off-by: Vincent-Xiao <vincent.xiao.me@gmail.com>
Signed-off-by: KrxGu <krishom70@gmail.com>
Signed-off-by: Fan Yang <yan9fan@meta.com>
Signed-off-by: Michele Gazzetti <michele.gazzetti1@ibm.com>
Signed-off-by: Robert Shaw <robshaw@redhat.com>
Signed-off-by: Zhengxu Chen <zhxchen17@meta.com>
Signed-off-by: LopezCastroRoberto <rocastro@redhat.com>
Signed-off-by: Roberto L. Castro <38211239+LopezCastroRoberto@users.noreply.github.com>
Signed-off-by: junuxyz <216036880+junuxyz@users.noreply.github.com>
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
Signed-off-by: Woosuk Kwon <woosuk@inferact.ai>
Signed-off-by: Andy Lo <andy@mistral.ai>
Signed-off-by: Qi Wang <qiwa@nvidia.com>
Signed-off-by: Jarno Seppänen <jseppanen@nvidia.com>
Signed-off-by: ilmarkov <markovilya197@gmail.com>
Signed-off-by: Tyler Michael Smith <tlrmchlsmth@gmail.com>
Signed-off-by: zzhengkai <zzhengkai@devgpu049.ldc1.facebook.com>
Signed-off-by: 7. Sun <jhao.sun@gmail.com>
Signed-off-by: Micah Williamson <micah.williamson@amd.com>
Signed-off-by: tianshu.yu <tianshuyu.formal@gmail.com>
Signed-off-by: Kebe <mail@kebe7jun.com>
Signed-off-by: Dzerzhinsky <256908701+AstroVoyager7@users.noreply.github.com>
Signed-off-by: Дзержи́нский <256908701+AstroVoyager7@users.noreply.github.com>
Signed-off-by: Matthias Gehre <matthias.gehre@amd.com>
Signed-off-by: zixiao <shunli.dsl@alibaba-inc.com>
Signed-off-by: Tianqi Ren <tianqi.r@outlook.com>
Signed-off-by: Linda-Stadter <57756729+Linda-Stadter@users.noreply.github.com>
Signed-off-by: Adam Binford <adamq43@gmail.com>
Signed-off-by: kliuae <kuanfu.liu@embeddedllm.com>
Signed-off-by: SorenDreano <71752785+SorenDreano@users.noreply.github.com>
Signed-off-by: dongxinyu03 <dongxinyu03@baidu.com>
Signed-off-by: Your Name <you@example.com>
Signed-off-by: elvischenv <219235043+elvischenv@users.noreply.github.com>
Co-authored-by: Chauncey <chaunceyjiang@gmail.com>
Co-authored-by: Kevin H. Luu <khluu000@gmail.com>
Co-authored-by: Andreas Karatzas <akaratza@amd.com>
Co-authored-by: Luka Govedič <ProExpertProg@users.noreply.github.com>
Co-authored-by: Nick Hill <nhill@redhat.com>
Co-authored-by: rinbaro <ilgomishra@gmail.com>
Co-authored-by: rasmith <Randall.Smith@amd.com>
Co-authored-by: Andrew Xia <axia@meta.com>
Co-authored-by: Andrew Xia <axia@fb.com>
Co-authored-by: Li, Jiang <jiang1.li@intel.com>
Co-authored-by: wang.yuqi <yuqi.wang@daocloud.io>
Co-authored-by: Fadi Arafeh <115173828+fadara01@users.noreply.github.com>
Co-authored-by: Mark McLoughlin <markmc@redhat.com>
Co-authored-by: Pavani Majety <pmajety@nvidia.com>
Co-authored-by: Cyrus Leung <tlleungac@connect.ust.hk>
Co-authored-by: Matthew Wong <Matthew.Wong2@amd.com>
Co-authored-by: liranschour <liranschour@users.noreply.github.com>
Co-authored-by: Or Ozeri <or@ozery.com>
Co-authored-by: Nicolò Lucchesi <nicolo.lucchesi@gmail.com>
Co-authored-by: Nicolò Lucchesi <nlucches@redhat.com>
Co-authored-by: jiahanc <173873397+jiahanc@users.noreply.github.com>
Co-authored-by: Isotr0py <mozf@mail2.sysu.edu.cn>
Co-authored-by: Mario Hong <86880754+mariohong128@users.noreply.github.com>
Co-authored-by: Aaron Hao <ahao@anyscale.com>
Co-authored-by: SumanthRH <sumanthrh99@gmail.com>
Co-authored-by: danisereb <daserebrenik@nvidia.com>
Co-authored-by: Benjamin Chislett <bchislett@nvidia.com>
Co-authored-by: zackyoray <yorayz@nvidia.com>
Co-authored-by: bnellnm <49004751+bnellnm@users.noreply.github.com>
Co-authored-by: Robert Shaw <114415538+robertgshaw2-redhat@users.noreply.github.com>
Co-authored-by: Tsukasa OI <floss_llm@irq.a4lg.com>
Co-authored-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Co-authored-by: Matthew Bonanni <mbonanni@redhat.com>
Co-authored-by: Lumosis <30372757+Lumosis@users.noreply.github.com>
Co-authored-by: mgoin <mgoin64@gmail.com>
Co-authored-by: Hashem Hashemi <159079214+amd-hhashemi@users.noreply.github.com>
Co-authored-by: Wei Zhao <51183510+wzhao18@users.noreply.github.com>
Co-authored-by: emricksini-h <emrick.birivoutin@hcompany.ai>
Co-authored-by: Xin Yang <105740670+xyang16@users.noreply.github.com>
Co-authored-by: Simon Mo <simon.mo@hey.com>
Co-authored-by: Rabi Mishra <ramishra@redhat.com>
Co-authored-by: Mingliang Li <limingliang0527@gmail.com>
Co-authored-by: limingliang <limingliang@stepfun.com>
Co-authored-by: R3hankhan <Rehan.Khan7@ibm.com>
Co-authored-by: Kunshang Ji <kunshang.ji@intel.com>
Co-authored-by: sihao_li <165983188+1643661061leo@users.noreply.github.com>
Co-authored-by: chengchengpei <5881383+chengchengpei@users.noreply.github.com>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
Co-authored-by: Gassan Salama <gassan.salama@arm.com>
Co-authored-by: Xinyu Chen <xinyu1.chen@intel.com>
Co-authored-by: zhang-prog <69562787+zhang-prog@users.noreply.github.com>
Co-authored-by: Kurt Shuster <shuster.kurt@gmail.com>
Co-authored-by: SorenDreano <71752785+SorenDreano@users.noreply.github.com>
Co-authored-by: Soren Dreano <soren@numind.ai>
Co-authored-by: Wentao Ye <44945378+yewentao256@users.noreply.github.com>
Co-authored-by: Raushan Turganbay <raushan@huggingface.co>
Co-authored-by: Cyrus Leung <cyrus.tl.leung@gmail.com>
Co-authored-by: FredericOdermatt <50372080+FredericOdermatt@users.noreply.github.com>
Co-authored-by: tc-mb <157115220+tc-mb@users.noreply.github.com>
Co-authored-by: mslv <mslv@baai.ac.cn>
Co-authored-by: zofia <110436990+zufangzhu@users.noreply.github.com>
Co-authored-by: Eldar Kurtić <8884008+eldarkurtic@users.noreply.github.com>
Co-authored-by: Seiji Eicher <58963096+eicherseiji@users.noreply.github.com>
Co-authored-by: vllmellm <vllm.ellm@embeddedllm.com>
Co-authored-by: TJian <tunjian.tan@embeddedllm.com>
Co-authored-by: zhrrr <43847754+izhuhaoran@users.noreply.github.com>
Co-authored-by: Charlie Fu <charlifu@amd.com>
Co-authored-by: xuebwang-amd <xuebwang@amd.com>
Co-authored-by: Sumanth R Hegde <39546518+SumanthRH@users.noreply.github.com>
Co-authored-by: Dimitrios Bariamis <dbari@users.noreply.github.com>
Co-authored-by: Dimitrios Bariamis <12195802+dbari@users.noreply.github.com>
Co-authored-by: kourosh hakhamaneshi <31483498+kouroshHakha@users.noreply.github.com>
Co-authored-by: Ikenna <ikennachifo@gmail.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: 果冻虾仁 <guodong@apache.org>
Co-authored-by: Vel <110626982+Code4me2@users.noreply.github.com>
Co-authored-by: lukec <118525388+sleepcoo@users.noreply.github.com>
Co-authored-by: Mohammad Miadh Angkad <176301910+mmangkad@users.noreply.github.com>
Co-authored-by: Richard Zou <zou3519@users.noreply.github.com>
Co-authored-by: TundeAtSN <akintunde.oladipo@servicenow.com>
Co-authored-by: Pooya Davoodi <pooya.davoodi@parasail.io>
Co-authored-by: Jee Jee Li <pandaleefree@gmail.com>
Co-authored-by: whx <56632993+whx-sjtu@users.noreply.github.com>
Co-authored-by: Rohan Potdar <66227218+Rohan138@users.noreply.github.com>
Co-authored-by: zifeitong <zifeitong@gmail.com>
Co-authored-by: Jiang Wu <jwu@cclgroup.com>
Co-authored-by: Reagan Lee <96998476+reaganjlee@users.noreply.github.com>
Co-authored-by: Reagan Lee <“reaganjlee@gmail.com”>
Co-authored-by: aabbccddwasd <140953076+aabbccddwasd@users.noreply.github.com>
Co-authored-by: TomerBN-Nvidia <tbarnatan@nvidia.com>
Co-authored-by: Tomer Natan <tbarnatan@computelab-frontend-8.nvidia.com>
Co-authored-by: navmarri14 <nmarri@roblox.com>
Co-authored-by: Andrey Talman <atalman@fb.com>
Co-authored-by: ihb2032 <40718643+ihb2032@users.noreply.github.com>
Co-authored-by: root <root@LAPTOP-FKNHV411.localdomain>
Co-authored-by: Ekagra Ranjan <3116519+ekagra-ranjan@users.noreply.github.com>
Co-authored-by: Nikhil Gupta <nikhil.gupta2@arm.com>
Co-authored-by: JJJYmmm <92386084+JJJYmmm@users.noreply.github.com>
Co-authored-by: wulipc <wulipc@users.noreply.github.com>
Co-authored-by: ywang96 <ywang96@users.noreply.github.com>
Co-authored-by: Isotr0py <Isotr0py@users.noreply.github.com>
Co-authored-by: Isotr0py <2037008807@qq.com>
Co-authored-by: Roger Wang <hey@rogerw.io>
Co-authored-by: Lucas Wilkinson <LucasWilkinson@users.noreply.github.com>
Co-authored-by: ZhengHongming888 <hongming.zheng@intel.com>
Co-authored-by: Jiangyun Zhu <riverclouds.zhu@qq.com>
Co-authored-by: Artus Krohn-Grimberghe <artuskg@users.noreply.github.com>
Co-authored-by: Hongxia Yang <62075498+hongxiayang@users.noreply.github.com>
Co-authored-by: Gregory Shtrasberg <156009573+gshtras@users.noreply.github.com>
Co-authored-by: Ning Xie <andy.xning@gmail.com>
Co-authored-by: Russell Bryant <rbryant@redhat.com>
Co-authored-by: Yuwei An <ayw.sirius19@gmail.com>
Co-authored-by: Balaxxe <136368465+jaim12005@users.noreply.github.com>
Co-authored-by: Chen Zhang <zhangch99@outlook.com>
Co-authored-by: Zetong Li <48438720+slippersss@users.noreply.github.com>
Co-authored-by: zzaebok <44357534+zzaebok@users.noreply.github.com>
Co-authored-by: Vincent-Xiao <vincent.xiao.me@gmail.com>
Co-authored-by: Phúc H. Lê Khắc <lkhphuc@pm.me>
Co-authored-by: Krish Gupta <krishom70@gmail.com>
Co-authored-by: Fan Yang <fanyang.real@gmail.com>
Co-authored-by: Fan Yang <yan9fan@meta.com>
Co-authored-by: mgazz <michele.gazzetti1@ibm.com>
Co-authored-by: Robert Shaw <robshaw@redhat.com>
Co-authored-by: Zhengxu Chen <zhxchen17@meta.com>
Co-authored-by: Roberto L. Castro <38211239+LopezCastroRoberto@users.noreply.github.com>
Co-authored-by: Claude Sonnet 4.5 <noreply@anthropic.com>
Co-authored-by: junuxyz <216036880+junuxyz@users.noreply.github.com>
Co-authored-by: Vadim Gimpelson <156319763+vadiklyutiy@users.noreply.github.com>
Co-authored-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>
Co-authored-by: Andy Lo <andy@mistral.ai>
Co-authored-by: Qi Wang <wqstu1@gmail.com>
Co-authored-by: J Seppänen <83203+jseppanen@users.noreply.github.com>
Co-authored-by: Tyler Michael Smith <tyler@neuralmagic.com>
Co-authored-by: Ilya Markov <markovilya197@gmail.com>
Co-authored-by: Tyler Michael Smith <tlrmchlsmth@gmail.com>
Co-authored-by: Zhengkai Zhang <33679250+ZhengkaiZ@users.noreply.github.com>
Co-authored-by: zzhengkai <zzhengkai@devgpu049.ldc1.facebook.com>
Co-authored-by: 7. Sun <jhao.sun@gmail.com>
Co-authored-by: Micah Williamson <micah.williamson@amd.com>
Co-authored-by: tianshu-Michael-yu <101950379+tianshu-Michael-yu@users.noreply.github.com>
Co-authored-by: Kebe <mail@kebe7jun.com>
Co-authored-by: Дзержи́нский <256908701+AstroVoyager7@users.noreply.github.com>
Co-authored-by: Matthias Gehre <matthias.gehre@amd.com>
Co-authored-by: AllenDou <allen.dou@hotmail.com>
Co-authored-by: zixiao <shunli.dsl@alibaba-inc.com>
Co-authored-by: Tianqi Ren <tianqi.r@outlook.com>
Co-authored-by: Linda <57756729+Linda-Stadter@users.noreply.github.com>
Co-authored-by: Adam Binford <adamq43@gmail.com>
Co-authored-by: kliuae <17350011+kliuae@users.noreply.github.com>
Co-authored-by: Xinyu Dong <dongxinyu03@baidu.com>
Co-authored-by: Your Name <you@example.com>
Co-authored-by: elvischenv <219235043+elvischenv@users.noreply.github.com>
heliubj18
added a commit
to heliubj18/vllm
that referenced
this pull request
Aug 8, 2026
Both passed in vllm-project#21 - kernels-moe-test at 223m09s once the timeout counted parallelism, kernels-quantization-test at 110m25s once the NVFP4 shard was staged. Leaving them in the allowlist put five and a half hours of GPU queue ahead of v1-others-cpu, which is the step vllm-project#22 was meant to verify: its 30 missing-repo failures are fixed now that Qwen3-0.6B, Qwen1.5-7B and Qwen2.5-VL-3B-Instruct are staged. The comment above the list already said an entry that has passed is queue spent re-proving it. It now names these two, so the next reader does not have to re-derive why they are gone. Co-authored-by: Claude Opus 5
heliubj18
added a commit
to heliubj18/vllm
that referenced
this pull request
Aug 8, 2026
Every step the allowlist was ever narrowed to is now accounted for, so the debugging aid has served its purpose: kernels-attention-test passed 119m25s in vllm-project#17 kernels-quantization-test passed 110m25s in vllm-project#21 kernels-moe-test passed 223m09s in vllm-project#21 sequence-parallel passed 41m32s in vllm-project#15 v1-others-cpu 402 cases / 1 failure in vllm-project#23 engine-1-gpu denylisted, upstream fork bug model-runner-v2-distributed denylisted, same bug model-runner-v2-distributed-2-gpus joins the denylist. It is the same flashinfer-poisons-fork bug as engine-1-gpu, reached through test_basic_correctness.py:23 instead: importing tests/utils runs the chain, then tests/utils.py:1745 forks, and the MultiProcExecutor worker in the child dies at gpu_worker.py:369. The log carries the fork DeprecationWarning at that exact line. Keeping it was tempting because only the first of its three commands is affected - the two DP commands do not import tests/utils and both pass by hand (test_async_llm_dp 14 selected of 18, test_eagle_dp 1 passed in 145s). But they have never run in CI: the first command always crashes first and a Buildkite step is atomic. So keeping the step buys no coverage, only a red build 43 seconds in, which under fail_fast would cancel every GPU step behind it. Recovering those two needs command-level filtering, which the generator does not have. v1-others-cpu is left in: its 30 missing-repo failures are fixed now that Qwen3-0.6B, Qwen1.5-7B and Qwen2.5-VL-3B-Instruct are staged as config-only (12M each - they are constructed through ModelConfig(...), not loaded). Its one remaining failure wants world size 2 while holding one card; upstream runs that step on a CPU image where ParallelConfig does not check GPU count. That is an image-type difference, not a card allocation bug, and it is not worth a per-key exception. Co-authored-by: Claude Opus 5
aoshen02
referenced
this pull request
in aoshen02/vllm
Aug 14, 2026
refactor: derive trace replay capacity from request state
grYe99
pushed a commit
to grYe99/vllm
that referenced
this pull request
Sep 17, 2026
Sync upstream into dc-main (422 commits behind)
This was referenced Oct 15, 2026
2 of 6 tasks
KexyBiscuit
pushed a commit
to tai-core/vllm-tracking
that referenced
this pull request
Nov 4, 2026
… work (vllm-project#21) * feat(glm5next): shard prefill PCP cache pages Allocate uniquely owned MLA and compressed-index pages independently of DCP while retaining KDA head sharding. Compress complete pools once, finish partial pools on tail owners, and assemble native pages through bounded collectives or MUSA IPC. Preserve group-aware COW and account for page workspace capacity at initialization. Group engine logic under vllm.v1.pcp and provide logical DP8/KDA TP8 handoff descriptors with source-page lifetime tracking. Validation: 149 PCP CPU tests passed, 5 skipped on this branch. Eight-rank IPC, pool, prefix, batch, full-logit and KDA-state comparisons passed during development. The existing mainline FlashKDA interface is inherited from the base. Requires the MUSA IPC companion commit 15e0f5f3252bd14fe7c5f32223b6ab6f73878897. Co-authored-by: Codex <noreply@openai.com> * perf(glm5next): reuse PCP producer rows and balance pool-aligned work Reuse encoded MLA and index rows, fuse current/history composition with unique owner storage, and share immutable routes within one forward. Balance complete kpool spans independently of persistent page ownership to remove excess padding. Account for native IPC workspace without charging collective gather buffers. Use the existing mainline KDA implementation inherited from the base; this commit does not modify KDA backend selection or calling conventions. Validation during development: 155 PCP tests passed (5 skipped). Eight-rank cache/COW/boundary checks passed. Full logits, index top-k and generated outputs matched an independent complete-cache reference, including prefix reuse. Original 128-request 8K-input benchmark: 19,831.93 token/s. The engine-side capacity estimator still uses the full logical pool and is addressed separately. Requires MUSA companion commit bdcf771f2dae05ac60d05505c1cec06e2769f744. Co-authored-by: Codex <noreply@openai.com> * fix(pcp): plan physical cache capacity and clear reused block tables Plan logical capacity from sharded attention pages, independent KDA pools and bounded page workspaces within the measured per-worker budget. Use the same accounting for admission, length estimates, allocation and reporting. Clear inactive columns when gathering sharded PCP block tables so a shorter request cannot materialize stale pages left by a previous long request. Validation: PCP and physical-planner regressions; prefill comparisons against independent 4e4fa1a075 sources covering 8K, chunked 32K, prefix reuse, 1M, short-after-long, fixed unequal batches and pool/page boundaries. Full 154880-wide logits, tokens, API logprobs, index top-k and effective KDA states match exactly with identical scheduling on all eight ranks. A 16K batch limit retains 1M context; the original 128-request benchmark reaches 21.18K/23.33K total tokens/s before/after a 1M request. Co-authored-by: Codex <noreply@openai.com> * fix(pcp): collect and remap only valid block-table prefixes Honor the block-table contract at the PCP reader: derive live page counts from effective sequence lengths and compressed chunk offsets, collect only those row prefixes, and remap into compact tables without reading stale tails. Keep root token lengths distinct from compressed pool lengths and include compression units in forward routing cache keys. Restore the original generic block-table gather interface and remove its PCP-specific tail clearing. Validation: PCP suite (166 passed, 6 skipped), targeted empty-page and routing regressions, and seven model scenarios including 1M, prefix hits, fixed unequal batches and pool/page boundaries. Full logits, generated tokens, API logprobs, index top-k and effective KDA states match 5bfdbc9be2 exactly on eight ranks. Original prefill benchmark after 1M: 23398.12 total tokens/s, 128 successes and no failures. Test services cleaned. Co-authored-by: Codex <noreply@openai.com> * refactor(pcp): resolve cache placement from layer specifications Declare page placement, query alignment and shared ownership groups in PCPCacheSpec at the model/cache-layer boundary. Keep GLM-specific support checks in the model implementation. Resolve the declaration in core and carry PCPPageLayout in KVCacheConfig. Worker allocation and PCP token partitioning consume this resolved layout rather than independently identifying the model. Budget bytes no longer select the cache layout. Replace layer-name-based owner pairing with explicit ownership-group validation. Keep capacity orchestration in core and make physical repacking operate on existing tensor descriptors. Remove the PCP allocation module's reverse imports from core. Preserve independent recurrent-state pools and support declared attention-only placement without assuming KDA exists. Validation: PCP, allocation, six packing-order, spec merge/serialization, and generic capacity regressions passed. Seven real model scenarios, including 1M, prefix reuse, unequal batches and pool/page boundaries, have identical full logits, output token IDs/logprobs, top-k and KDA states. Fresh-service benchmark: 21,356.13 total tokens/s, 128 successes, no failures. Original scripts unchanged; test services cleaned. Co-authored-by: Codex <noreply@openai.com> --------- Co-authored-by: Codex <noreply@openai.com>
KexyBiscuit
pushed a commit
to tai-core/vllm-tracking
that referenced
this pull request
Nov 4, 2026
… work (vllm-project#21) * feat(glm5next): shard prefill PCP cache pages Allocate uniquely owned MLA and compressed-index pages independently of DCP while retaining KDA head sharding. Compress complete pools once, finish partial pools on tail owners, and assemble native pages through bounded collectives or MUSA IPC. Preserve group-aware COW and account for page workspace capacity at initialization. Group engine logic under vllm.v1.pcp and provide logical DP8/KDA TP8 handoff descriptors with source-page lifetime tracking. Validation: 149 PCP CPU tests passed, 5 skipped on this branch. Eight-rank IPC, pool, prefix, batch, full-logit and KDA-state comparisons passed during development. The existing mainline FlashKDA interface is inherited from the base. Requires the MUSA IPC companion commit 15e0f5f3252bd14fe7c5f32223b6ab6f73878897. Co-authored-by: Codex <noreply@openai.com> * perf(glm5next): reuse PCP producer rows and balance pool-aligned work Reuse encoded MLA and index rows, fuse current/history composition with unique owner storage, and share immutable routes within one forward. Balance complete kpool spans independently of persistent page ownership to remove excess padding. Account for native IPC workspace without charging collective gather buffers. Use the existing mainline KDA implementation inherited from the base; this commit does not modify KDA backend selection or calling conventions. Validation during development: 155 PCP tests passed (5 skipped). Eight-rank cache/COW/boundary checks passed. Full logits, index top-k and generated outputs matched an independent complete-cache reference, including prefix reuse. Original 128-request 8K-input benchmark: 19,831.93 token/s. The engine-side capacity estimator still uses the full logical pool and is addressed separately. Requires MUSA companion commit bdcf771f2dae05ac60d05505c1cec06e2769f744. Co-authored-by: Codex <noreply@openai.com> * fix(pcp): plan physical cache capacity and clear reused block tables Plan logical capacity from sharded attention pages, independent KDA pools and bounded page workspaces within the measured per-worker budget. Use the same accounting for admission, length estimates, allocation and reporting. Clear inactive columns when gathering sharded PCP block tables so a shorter request cannot materialize stale pages left by a previous long request. Validation: PCP and physical-planner regressions; prefill comparisons against independent 4e4fa1a075 sources covering 8K, chunked 32K, prefix reuse, 1M, short-after-long, fixed unequal batches and pool/page boundaries. Full 154880-wide logits, tokens, API logprobs, index top-k and effective KDA states match exactly with identical scheduling on all eight ranks. A 16K batch limit retains 1M context; the original 128-request benchmark reaches 21.18K/23.33K total tokens/s before/after a 1M request. Co-authored-by: Codex <noreply@openai.com> * fix(pcp): collect and remap only valid block-table prefixes Honor the block-table contract at the PCP reader: derive live page counts from effective sequence lengths and compressed chunk offsets, collect only those row prefixes, and remap into compact tables without reading stale tails. Keep root token lengths distinct from compressed pool lengths and include compression units in forward routing cache keys. Restore the original generic block-table gather interface and remove its PCP-specific tail clearing. Validation: PCP suite (166 passed, 6 skipped), targeted empty-page and routing regressions, and seven model scenarios including 1M, prefix hits, fixed unequal batches and pool/page boundaries. Full logits, generated tokens, API logprobs, index top-k and effective KDA states match 5bfdbc9be2 exactly on eight ranks. Original prefill benchmark after 1M: 23398.12 total tokens/s, 128 successes and no failures. Test services cleaned. Co-authored-by: Codex <noreply@openai.com> * refactor(pcp): resolve cache placement from layer specifications Declare page placement, query alignment and shared ownership groups in PCPCacheSpec at the model/cache-layer boundary. Keep GLM-specific support checks in the model implementation. Resolve the declaration in core and carry PCPPageLayout in KVCacheConfig. Worker allocation and PCP token partitioning consume this resolved layout rather than independently identifying the model. Budget bytes no longer select the cache layout. Replace layer-name-based owner pairing with explicit ownership-group validation. Keep capacity orchestration in core and make physical repacking operate on existing tensor descriptors. Remove the PCP allocation module's reverse imports from core. Preserve independent recurrent-state pools and support declared attention-only placement without assuming KDA exists. Validation: PCP, allocation, six packing-order, spec merge/serialization, and generic capacity regressions passed. Seven real model scenarios, including 1M, prefix reuse, unequal batches and pool/page boundaries, have identical full logits, output token IDs/logprobs, top-k and KDA states. Fresh-service benchmark: 21,356.13 total tokens/s, 128 successes, no failures. Original scripts unchanged; test services cleaned. Co-authored-by: Codex <noreply@openai.com> --------- Co-authored-by: Codex <noreply@openai.com>
KexyBiscuit
pushed a commit
to tai-core/vllm-tracking
that referenced
this pull request
Nov 4, 2026
… work (vllm-project#21) * feat(glm5next): shard prefill PCP cache pages Allocate uniquely owned MLA and compressed-index pages independently of DCP while retaining KDA head sharding. Compress complete pools once, finish partial pools on tail owners, and assemble native pages through bounded collectives or MUSA IPC. Preserve group-aware COW and account for page workspace capacity at initialization. Group engine logic under vllm.v1.pcp and provide logical DP8/KDA TP8 handoff descriptors with source-page lifetime tracking. Validation: 149 PCP CPU tests passed, 5 skipped on this branch. Eight-rank IPC, pool, prefix, batch, full-logit and KDA-state comparisons passed during development. The existing mainline FlashKDA interface is inherited from the base. Requires the MUSA IPC companion commit 15e0f5f3252bd14fe7c5f32223b6ab6f73878897. Co-authored-by: Codex <noreply@openai.com> * perf(glm5next): reuse PCP producer rows and balance pool-aligned work Reuse encoded MLA and index rows, fuse current/history composition with unique owner storage, and share immutable routes within one forward. Balance complete kpool spans independently of persistent page ownership to remove excess padding. Account for native IPC workspace without charging collective gather buffers. Use the existing mainline KDA implementation inherited from the base; this commit does not modify KDA backend selection or calling conventions. Validation during development: 155 PCP tests passed (5 skipped). Eight-rank cache/COW/boundary checks passed. Full logits, index top-k and generated outputs matched an independent complete-cache reference, including prefix reuse. Original 128-request 8K-input benchmark: 19,831.93 token/s. The engine-side capacity estimator still uses the full logical pool and is addressed separately. Requires MUSA companion commit bdcf771f2dae05ac60d05505c1cec06e2769f744. Co-authored-by: Codex <noreply@openai.com> * fix(pcp): plan physical cache capacity and clear reused block tables Plan logical capacity from sharded attention pages, independent KDA pools and bounded page workspaces within the measured per-worker budget. Use the same accounting for admission, length estimates, allocation and reporting. Clear inactive columns when gathering sharded PCP block tables so a shorter request cannot materialize stale pages left by a previous long request. Validation: PCP and physical-planner regressions; prefill comparisons against independent 4e4fa1a075 sources covering 8K, chunked 32K, prefix reuse, 1M, short-after-long, fixed unequal batches and pool/page boundaries. Full 154880-wide logits, tokens, API logprobs, index top-k and effective KDA states match exactly with identical scheduling on all eight ranks. A 16K batch limit retains 1M context; the original 128-request benchmark reaches 21.18K/23.33K total tokens/s before/after a 1M request. Co-authored-by: Codex <noreply@openai.com> * fix(pcp): collect and remap only valid block-table prefixes Honor the block-table contract at the PCP reader: derive live page counts from effective sequence lengths and compressed chunk offsets, collect only those row prefixes, and remap into compact tables without reading stale tails. Keep root token lengths distinct from compressed pool lengths and include compression units in forward routing cache keys. Restore the original generic block-table gather interface and remove its PCP-specific tail clearing. Validation: PCP suite (166 passed, 6 skipped), targeted empty-page and routing regressions, and seven model scenarios including 1M, prefix hits, fixed unequal batches and pool/page boundaries. Full logits, generated tokens, API logprobs, index top-k and effective KDA states match 5bfdbc9be2 exactly on eight ranks. Original prefill benchmark after 1M: 23398.12 total tokens/s, 128 successes and no failures. Test services cleaned. Co-authored-by: Codex <noreply@openai.com> * refactor(pcp): resolve cache placement from layer specifications Declare page placement, query alignment and shared ownership groups in PCPCacheSpec at the model/cache-layer boundary. Keep GLM-specific support checks in the model implementation. Resolve the declaration in core and carry PCPPageLayout in KVCacheConfig. Worker allocation and PCP token partitioning consume this resolved layout rather than independently identifying the model. Budget bytes no longer select the cache layout. Replace layer-name-based owner pairing with explicit ownership-group validation. Keep capacity orchestration in core and make physical repacking operate on existing tensor descriptors. Remove the PCP allocation module's reverse imports from core. Preserve independent recurrent-state pools and support declared attention-only placement without assuming KDA exists. Validation: PCP, allocation, six packing-order, spec merge/serialization, and generic capacity regressions passed. Seven real model scenarios, including 1M, prefix reuse, unequal batches and pool/page boundaries, have identical full logits, output token IDs/logprobs, top-k and KDA states. Fresh-service benchmark: 21,356.13 total tokens/s, 128 successes, no failures. Original scripts unchanged; test services cleaned. Co-authored-by: Codex <noreply@openai.com> --------- Co-authored-by: Codex <noreply@openai.com>
KexyBiscuit
pushed a commit
to tai-core/vllm-tracking
that referenced
this pull request
Nov 5, 2026
… work (vllm-project#21) * feat(glm5next): shard prefill PCP cache pages Allocate uniquely owned MLA and compressed-index pages independently of DCP while retaining KDA head sharding. Compress complete pools once, finish partial pools on tail owners, and assemble native pages through bounded collectives or MUSA IPC. Preserve group-aware COW and account for page workspace capacity at initialization. Group engine logic under vllm.v1.pcp and provide logical DP8/KDA TP8 handoff descriptors with source-page lifetime tracking. Validation: 149 PCP CPU tests passed, 5 skipped on this branch. Eight-rank IPC, pool, prefix, batch, full-logit and KDA-state comparisons passed during development. The existing mainline FlashKDA interface is inherited from the base. Requires the MUSA IPC companion commit 15e0f5f3252bd14fe7c5f32223b6ab6f73878897. Co-authored-by: Codex <noreply@openai.com> * perf(glm5next): reuse PCP producer rows and balance pool-aligned work Reuse encoded MLA and index rows, fuse current/history composition with unique owner storage, and share immutable routes within one forward. Balance complete kpool spans independently of persistent page ownership to remove excess padding. Account for native IPC workspace without charging collective gather buffers. Use the existing mainline KDA implementation inherited from the base; this commit does not modify KDA backend selection or calling conventions. Validation during development: 155 PCP tests passed (5 skipped). Eight-rank cache/COW/boundary checks passed. Full logits, index top-k and generated outputs matched an independent complete-cache reference, including prefix reuse. Original 128-request 8K-input benchmark: 19,831.93 token/s. The engine-side capacity estimator still uses the full logical pool and is addressed separately. Requires MUSA companion commit bdcf771f2dae05ac60d05505c1cec06e2769f744. Co-authored-by: Codex <noreply@openai.com> * fix(pcp): plan physical cache capacity and clear reused block tables Plan logical capacity from sharded attention pages, independent KDA pools and bounded page workspaces within the measured per-worker budget. Use the same accounting for admission, length estimates, allocation and reporting. Clear inactive columns when gathering sharded PCP block tables so a shorter request cannot materialize stale pages left by a previous long request. Validation: PCP and physical-planner regressions; prefill comparisons against independent 4e4fa1a075 sources covering 8K, chunked 32K, prefix reuse, 1M, short-after-long, fixed unequal batches and pool/page boundaries. Full 154880-wide logits, tokens, API logprobs, index top-k and effective KDA states match exactly with identical scheduling on all eight ranks. A 16K batch limit retains 1M context; the original 128-request benchmark reaches 21.18K/23.33K total tokens/s before/after a 1M request. Co-authored-by: Codex <noreply@openai.com> * fix(pcp): collect and remap only valid block-table prefixes Honor the block-table contract at the PCP reader: derive live page counts from effective sequence lengths and compressed chunk offsets, collect only those row prefixes, and remap into compact tables without reading stale tails. Keep root token lengths distinct from compressed pool lengths and include compression units in forward routing cache keys. Restore the original generic block-table gather interface and remove its PCP-specific tail clearing. Validation: PCP suite (166 passed, 6 skipped), targeted empty-page and routing regressions, and seven model scenarios including 1M, prefix hits, fixed unequal batches and pool/page boundaries. Full logits, generated tokens, API logprobs, index top-k and effective KDA states match 5bfdbc9be2 exactly on eight ranks. Original prefill benchmark after 1M: 23398.12 total tokens/s, 128 successes and no failures. Test services cleaned. Co-authored-by: Codex <noreply@openai.com> * refactor(pcp): resolve cache placement from layer specifications Declare page placement, query alignment and shared ownership groups in PCPCacheSpec at the model/cache-layer boundary. Keep GLM-specific support checks in the model implementation. Resolve the declaration in core and carry PCPPageLayout in KVCacheConfig. Worker allocation and PCP token partitioning consume this resolved layout rather than independently identifying the model. Budget bytes no longer select the cache layout. Replace layer-name-based owner pairing with explicit ownership-group validation. Keep capacity orchestration in core and make physical repacking operate on existing tensor descriptors. Remove the PCP allocation module's reverse imports from core. Preserve independent recurrent-state pools and support declared attention-only placement without assuming KDA exists. Validation: PCP, allocation, six packing-order, spec merge/serialization, and generic capacity regressions passed. Seven real model scenarios, including 1M, prefix reuse, unequal batches and pool/page boundaries, have identical full logits, output token IDs/logprobs, top-k and KDA states. Fresh-service benchmark: 21,356.13 total tokens/s, 128 successes, no failures. Original scripts unchanged; test services cleaned. Co-authored-by: Codex <noreply@openai.com> --------- Co-authored-by: Codex <noreply@openai.com>
KexyBiscuit
pushed a commit
to tai-core/vllm-tracking
that referenced
this pull request
Nov 5, 2026
… work (vllm-project#21) * feat(glm5next): shard prefill PCP cache pages Allocate uniquely owned MLA and compressed-index pages independently of DCP while retaining KDA head sharding. Compress complete pools once, finish partial pools on tail owners, and assemble native pages through bounded collectives or MUSA IPC. Preserve group-aware COW and account for page workspace capacity at initialization. Group engine logic under vllm.v1.pcp and provide logical DP8/KDA TP8 handoff descriptors with source-page lifetime tracking. Validation: 149 PCP CPU tests passed, 5 skipped on this branch. Eight-rank IPC, pool, prefix, batch, full-logit and KDA-state comparisons passed during development. The existing mainline FlashKDA interface is inherited from the base. Requires the MUSA IPC companion commit 15e0f5f3252bd14fe7c5f32223b6ab6f73878897. Co-authored-by: Codex <noreply@openai.com> * perf(glm5next): reuse PCP producer rows and balance pool-aligned work Reuse encoded MLA and index rows, fuse current/history composition with unique owner storage, and share immutable routes within one forward. Balance complete kpool spans independently of persistent page ownership to remove excess padding. Account for native IPC workspace without charging collective gather buffers. Use the existing mainline KDA implementation inherited from the base; this commit does not modify KDA backend selection or calling conventions. Validation during development: 155 PCP tests passed (5 skipped). Eight-rank cache/COW/boundary checks passed. Full logits, index top-k and generated outputs matched an independent complete-cache reference, including prefix reuse. Original 128-request 8K-input benchmark: 19,831.93 token/s. The engine-side capacity estimator still uses the full logical pool and is addressed separately. Requires MUSA companion commit bdcf771f2dae05ac60d05505c1cec06e2769f744. Co-authored-by: Codex <noreply@openai.com> * fix(pcp): plan physical cache capacity and clear reused block tables Plan logical capacity from sharded attention pages, independent KDA pools and bounded page workspaces within the measured per-worker budget. Use the same accounting for admission, length estimates, allocation and reporting. Clear inactive columns when gathering sharded PCP block tables so a shorter request cannot materialize stale pages left by a previous long request. Validation: PCP and physical-planner regressions; prefill comparisons against independent 4e4fa1a075 sources covering 8K, chunked 32K, prefix reuse, 1M, short-after-long, fixed unequal batches and pool/page boundaries. Full 154880-wide logits, tokens, API logprobs, index top-k and effective KDA states match exactly with identical scheduling on all eight ranks. A 16K batch limit retains 1M context; the original 128-request benchmark reaches 21.18K/23.33K total tokens/s before/after a 1M request. Co-authored-by: Codex <noreply@openai.com> * fix(pcp): collect and remap only valid block-table prefixes Honor the block-table contract at the PCP reader: derive live page counts from effective sequence lengths and compressed chunk offsets, collect only those row prefixes, and remap into compact tables without reading stale tails. Keep root token lengths distinct from compressed pool lengths and include compression units in forward routing cache keys. Restore the original generic block-table gather interface and remove its PCP-specific tail clearing. Validation: PCP suite (166 passed, 6 skipped), targeted empty-page and routing regressions, and seven model scenarios including 1M, prefix hits, fixed unequal batches and pool/page boundaries. Full logits, generated tokens, API logprobs, index top-k and effective KDA states match 5bfdbc9be2 exactly on eight ranks. Original prefill benchmark after 1M: 23398.12 total tokens/s, 128 successes and no failures. Test services cleaned. Co-authored-by: Codex <noreply@openai.com> * refactor(pcp): resolve cache placement from layer specifications Declare page placement, query alignment and shared ownership groups in PCPCacheSpec at the model/cache-layer boundary. Keep GLM-specific support checks in the model implementation. Resolve the declaration in core and carry PCPPageLayout in KVCacheConfig. Worker allocation and PCP token partitioning consume this resolved layout rather than independently identifying the model. Budget bytes no longer select the cache layout. Replace layer-name-based owner pairing with explicit ownership-group validation. Keep capacity orchestration in core and make physical repacking operate on existing tensor descriptors. Remove the PCP allocation module's reverse imports from core. Preserve independent recurrent-state pools and support declared attention-only placement without assuming KDA exists. Validation: PCP, allocation, six packing-order, spec merge/serialization, and generic capacity regressions passed. Seven real model scenarios, including 1M, prefix reuse, unequal batches and pool/page boundaries, have identical full logits, output token IDs/logprobs, top-k and KDA states. Fresh-service benchmark: 21,356.13 total tokens/s, 128 successes, no failures. Original scripts unchanged; test services cleaned. Co-authored-by: Codex <noreply@openai.com> --------- Co-authored-by: Codex <noreply@openai.com>
KexyBiscuit
pushed a commit
to tai-core/vllm-tracking
that referenced
this pull request
Nov 5, 2026
… work (vllm-project#21) * feat(glm5next): shard prefill PCP cache pages Allocate uniquely owned MLA and compressed-index pages independently of DCP while retaining KDA head sharding. Compress complete pools once, finish partial pools on tail owners, and assemble native pages through bounded collectives or MUSA IPC. Preserve group-aware COW and account for page workspace capacity at initialization. Group engine logic under vllm.v1.pcp and provide logical DP8/KDA TP8 handoff descriptors with source-page lifetime tracking. Validation: 149 PCP CPU tests passed, 5 skipped on this branch. Eight-rank IPC, pool, prefix, batch, full-logit and KDA-state comparisons passed during development. The existing mainline FlashKDA interface is inherited from the base. Requires the MUSA IPC companion commit 15e0f5f3252bd14fe7c5f32223b6ab6f73878897. Co-authored-by: Codex <noreply@openai.com> * perf(glm5next): reuse PCP producer rows and balance pool-aligned work Reuse encoded MLA and index rows, fuse current/history composition with unique owner storage, and share immutable routes within one forward. Balance complete kpool spans independently of persistent page ownership to remove excess padding. Account for native IPC workspace without charging collective gather buffers. Use the existing mainline KDA implementation inherited from the base; this commit does not modify KDA backend selection or calling conventions. Validation during development: 155 PCP tests passed (5 skipped). Eight-rank cache/COW/boundary checks passed. Full logits, index top-k and generated outputs matched an independent complete-cache reference, including prefix reuse. Original 128-request 8K-input benchmark: 19,831.93 token/s. The engine-side capacity estimator still uses the full logical pool and is addressed separately. Requires MUSA companion commit bdcf771f2dae05ac60d05505c1cec06e2769f744. Co-authored-by: Codex <noreply@openai.com> * fix(pcp): plan physical cache capacity and clear reused block tables Plan logical capacity from sharded attention pages, independent KDA pools and bounded page workspaces within the measured per-worker budget. Use the same accounting for admission, length estimates, allocation and reporting. Clear inactive columns when gathering sharded PCP block tables so a shorter request cannot materialize stale pages left by a previous long request. Validation: PCP and physical-planner regressions; prefill comparisons against independent 4e4fa1a075 sources covering 8K, chunked 32K, prefix reuse, 1M, short-after-long, fixed unequal batches and pool/page boundaries. Full 154880-wide logits, tokens, API logprobs, index top-k and effective KDA states match exactly with identical scheduling on all eight ranks. A 16K batch limit retains 1M context; the original 128-request benchmark reaches 21.18K/23.33K total tokens/s before/after a 1M request. Co-authored-by: Codex <noreply@openai.com> * fix(pcp): collect and remap only valid block-table prefixes Honor the block-table contract at the PCP reader: derive live page counts from effective sequence lengths and compressed chunk offsets, collect only those row prefixes, and remap into compact tables without reading stale tails. Keep root token lengths distinct from compressed pool lengths and include compression units in forward routing cache keys. Restore the original generic block-table gather interface and remove its PCP-specific tail clearing. Validation: PCP suite (166 passed, 6 skipped), targeted empty-page and routing regressions, and seven model scenarios including 1M, prefix hits, fixed unequal batches and pool/page boundaries. Full logits, generated tokens, API logprobs, index top-k and effective KDA states match 5bfdbc9be2 exactly on eight ranks. Original prefill benchmark after 1M: 23398.12 total tokens/s, 128 successes and no failures. Test services cleaned. Co-authored-by: Codex <noreply@openai.com> * refactor(pcp): resolve cache placement from layer specifications Declare page placement, query alignment and shared ownership groups in PCPCacheSpec at the model/cache-layer boundary. Keep GLM-specific support checks in the model implementation. Resolve the declaration in core and carry PCPPageLayout in KVCacheConfig. Worker allocation and PCP token partitioning consume this resolved layout rather than independently identifying the model. Budget bytes no longer select the cache layout. Replace layer-name-based owner pairing with explicit ownership-group validation. Keep capacity orchestration in core and make physical repacking operate on existing tensor descriptors. Remove the PCP allocation module's reverse imports from core. Preserve independent recurrent-state pools and support declared attention-only placement without assuming KDA exists. Validation: PCP, allocation, six packing-order, spec merge/serialization, and generic capacity regressions passed. Seven real model scenarios, including 1M, prefix reuse, unequal batches and pool/page boundaries, have identical full logits, output token IDs/logprobs, top-k and KDA states. Fresh-service benchmark: 21,356.13 total tokens/s, 128 successes, no failures. Original scripts unchanged; test services cleaned. Co-authored-by: Codex <noreply@openai.com> --------- Co-authored-by: Codex <noreply@openai.com>
KexyBiscuit
pushed a commit
to tai-core/vllm-tracking
that referenced
this pull request
Nov 5, 2026
… work (vllm-project#21) * feat(glm5next): shard prefill PCP cache pages Allocate uniquely owned MLA and compressed-index pages independently of DCP while retaining KDA head sharding. Compress complete pools once, finish partial pools on tail owners, and assemble native pages through bounded collectives or MUSA IPC. Preserve group-aware COW and account for page workspace capacity at initialization. Group engine logic under vllm.v1.pcp and provide logical DP8/KDA TP8 handoff descriptors with source-page lifetime tracking. Validation: 149 PCP CPU tests passed, 5 skipped on this branch. Eight-rank IPC, pool, prefix, batch, full-logit and KDA-state comparisons passed during development. The existing mainline FlashKDA interface is inherited from the base. Requires the MUSA IPC companion commit 15e0f5f3252bd14fe7c5f32223b6ab6f73878897. Co-authored-by: Codex <noreply@openai.com> * perf(glm5next): reuse PCP producer rows and balance pool-aligned work Reuse encoded MLA and index rows, fuse current/history composition with unique owner storage, and share immutable routes within one forward. Balance complete kpool spans independently of persistent page ownership to remove excess padding. Account for native IPC workspace without charging collective gather buffers. Use the existing mainline KDA implementation inherited from the base; this commit does not modify KDA backend selection or calling conventions. Validation during development: 155 PCP tests passed (5 skipped). Eight-rank cache/COW/boundary checks passed. Full logits, index top-k and generated outputs matched an independent complete-cache reference, including prefix reuse. Original 128-request 8K-input benchmark: 19,831.93 token/s. The engine-side capacity estimator still uses the full logical pool and is addressed separately. Requires MUSA companion commit bdcf771f2dae05ac60d05505c1cec06e2769f744. Co-authored-by: Codex <noreply@openai.com> * fix(pcp): plan physical cache capacity and clear reused block tables Plan logical capacity from sharded attention pages, independent KDA pools and bounded page workspaces within the measured per-worker budget. Use the same accounting for admission, length estimates, allocation and reporting. Clear inactive columns when gathering sharded PCP block tables so a shorter request cannot materialize stale pages left by a previous long request. Validation: PCP and physical-planner regressions; prefill comparisons against independent 4e4fa1a075 sources covering 8K, chunked 32K, prefix reuse, 1M, short-after-long, fixed unequal batches and pool/page boundaries. Full 154880-wide logits, tokens, API logprobs, index top-k and effective KDA states match exactly with identical scheduling on all eight ranks. A 16K batch limit retains 1M context; the original 128-request benchmark reaches 21.18K/23.33K total tokens/s before/after a 1M request. Co-authored-by: Codex <noreply@openai.com> * fix(pcp): collect and remap only valid block-table prefixes Honor the block-table contract at the PCP reader: derive live page counts from effective sequence lengths and compressed chunk offsets, collect only those row prefixes, and remap into compact tables without reading stale tails. Keep root token lengths distinct from compressed pool lengths and include compression units in forward routing cache keys. Restore the original generic block-table gather interface and remove its PCP-specific tail clearing. Validation: PCP suite (166 passed, 6 skipped), targeted empty-page and routing regressions, and seven model scenarios including 1M, prefix hits, fixed unequal batches and pool/page boundaries. Full logits, generated tokens, API logprobs, index top-k and effective KDA states match 5bfdbc9be2 exactly on eight ranks. Original prefill benchmark after 1M: 23398.12 total tokens/s, 128 successes and no failures. Test services cleaned. Co-authored-by: Codex <noreply@openai.com> * refactor(pcp): resolve cache placement from layer specifications Declare page placement, query alignment and shared ownership groups in PCPCacheSpec at the model/cache-layer boundary. Keep GLM-specific support checks in the model implementation. Resolve the declaration in core and carry PCPPageLayout in KVCacheConfig. Worker allocation and PCP token partitioning consume this resolved layout rather than independently identifying the model. Budget bytes no longer select the cache layout. Replace layer-name-based owner pairing with explicit ownership-group validation. Keep capacity orchestration in core and make physical repacking operate on existing tensor descriptors. Remove the PCP allocation module's reverse imports from core. Preserve independent recurrent-state pools and support declared attention-only placement without assuming KDA exists. Validation: PCP, allocation, six packing-order, spec merge/serialization, and generic capacity regressions passed. Seven real model scenarios, including 1M, prefix reuse, unequal batches and pool/page boundaries, have identical full logits, output token IDs/logprobs, top-k and KDA states. Fresh-service benchmark: 21,356.13 total tokens/s, 128 successes, no failures. Original scripts unchanged; test services cleaned. Co-authored-by: Codex <noreply@openai.com> --------- Co-authored-by: Codex <noreply@openai.com>
KexyBiscuit
pushed a commit
to tai-core/vllm-tracking
that referenced
this pull request
Nov 5, 2026
… work (vllm-project#21) * feat(glm5next): shard prefill PCP cache pages Allocate uniquely owned MLA and compressed-index pages independently of DCP while retaining KDA head sharding. Compress complete pools once, finish partial pools on tail owners, and assemble native pages through bounded collectives or MUSA IPC. Preserve group-aware COW and account for page workspace capacity at initialization. Group engine logic under vllm.v1.pcp and provide logical DP8/KDA TP8 handoff descriptors with source-page lifetime tracking. Validation: 149 PCP CPU tests passed, 5 skipped on this branch. Eight-rank IPC, pool, prefix, batch, full-logit and KDA-state comparisons passed during development. The existing mainline FlashKDA interface is inherited from the base. Requires the MUSA IPC companion commit 15e0f5f3252bd14fe7c5f32223b6ab6f73878897. Co-authored-by: Codex <noreply@openai.com> * perf(glm5next): reuse PCP producer rows and balance pool-aligned work Reuse encoded MLA and index rows, fuse current/history composition with unique owner storage, and share immutable routes within one forward. Balance complete kpool spans independently of persistent page ownership to remove excess padding. Account for native IPC workspace without charging collective gather buffers. Use the existing mainline KDA implementation inherited from the base; this commit does not modify KDA backend selection or calling conventions. Validation during development: 155 PCP tests passed (5 skipped). Eight-rank cache/COW/boundary checks passed. Full logits, index top-k and generated outputs matched an independent complete-cache reference, including prefix reuse. Original 128-request 8K-input benchmark: 19,831.93 token/s. The engine-side capacity estimator still uses the full logical pool and is addressed separately. Requires MUSA companion commit bdcf771f2dae05ac60d05505c1cec06e2769f744. Co-authored-by: Codex <noreply@openai.com> * fix(pcp): plan physical cache capacity and clear reused block tables Plan logical capacity from sharded attention pages, independent KDA pools and bounded page workspaces within the measured per-worker budget. Use the same accounting for admission, length estimates, allocation and reporting. Clear inactive columns when gathering sharded PCP block tables so a shorter request cannot materialize stale pages left by a previous long request. Validation: PCP and physical-planner regressions; prefill comparisons against independent 4e4fa1a075 sources covering 8K, chunked 32K, prefix reuse, 1M, short-after-long, fixed unequal batches and pool/page boundaries. Full 154880-wide logits, tokens, API logprobs, index top-k and effective KDA states match exactly with identical scheduling on all eight ranks. A 16K batch limit retains 1M context; the original 128-request benchmark reaches 21.18K/23.33K total tokens/s before/after a 1M request. Co-authored-by: Codex <noreply@openai.com> * fix(pcp): collect and remap only valid block-table prefixes Honor the block-table contract at the PCP reader: derive live page counts from effective sequence lengths and compressed chunk offsets, collect only those row prefixes, and remap into compact tables without reading stale tails. Keep root token lengths distinct from compressed pool lengths and include compression units in forward routing cache keys. Restore the original generic block-table gather interface and remove its PCP-specific tail clearing. Validation: PCP suite (166 passed, 6 skipped), targeted empty-page and routing regressions, and seven model scenarios including 1M, prefix hits, fixed unequal batches and pool/page boundaries. Full logits, generated tokens, API logprobs, index top-k and effective KDA states match 5bfdbc9be2 exactly on eight ranks. Original prefill benchmark after 1M: 23398.12 total tokens/s, 128 successes and no failures. Test services cleaned. Co-authored-by: Codex <noreply@openai.com> * refactor(pcp): resolve cache placement from layer specifications Declare page placement, query alignment and shared ownership groups in PCPCacheSpec at the model/cache-layer boundary. Keep GLM-specific support checks in the model implementation. Resolve the declaration in core and carry PCPPageLayout in KVCacheConfig. Worker allocation and PCP token partitioning consume this resolved layout rather than independently identifying the model. Budget bytes no longer select the cache layout. Replace layer-name-based owner pairing with explicit ownership-group validation. Keep capacity orchestration in core and make physical repacking operate on existing tensor descriptors. Remove the PCP allocation module's reverse imports from core. Preserve independent recurrent-state pools and support declared attention-only placement without assuming KDA exists. Validation: PCP, allocation, six packing-order, spec merge/serialization, and generic capacity regressions passed. Seven real model scenarios, including 1M, prefix reuse, unequal batches and pool/page boundaries, have identical full logits, output token IDs/logprobs, top-k and KDA states. Fresh-service benchmark: 21,356.13 total tokens/s, 128 successes, no failures. Original scripts unchanged; test services cleaned. Co-authored-by: Codex <noreply@openai.com> --------- Co-authored-by: Codex <noreply@openai.com>
KexyBiscuit
pushed a commit
to tai-core/vllm-tracking
that referenced
this pull request
Nov 5, 2026
… work (vllm-project#21) * feat(glm5next): shard prefill PCP cache pages Allocate uniquely owned MLA and compressed-index pages independently of DCP while retaining KDA head sharding. Compress complete pools once, finish partial pools on tail owners, and assemble native pages through bounded collectives or MUSA IPC. Preserve group-aware COW and account for page workspace capacity at initialization. Group engine logic under vllm.v1.pcp and provide logical DP8/KDA TP8 handoff descriptors with source-page lifetime tracking. Validation: 149 PCP CPU tests passed, 5 skipped on this branch. Eight-rank IPC, pool, prefix, batch, full-logit and KDA-state comparisons passed during development. The existing mainline FlashKDA interface is inherited from the base. Requires the MUSA IPC companion commit 15e0f5f3252bd14fe7c5f32223b6ab6f73878897. Co-authored-by: Codex <noreply@openai.com> * perf(glm5next): reuse PCP producer rows and balance pool-aligned work Reuse encoded MLA and index rows, fuse current/history composition with unique owner storage, and share immutable routes within one forward. Balance complete kpool spans independently of persistent page ownership to remove excess padding. Account for native IPC workspace without charging collective gather buffers. Use the existing mainline KDA implementation inherited from the base; this commit does not modify KDA backend selection or calling conventions. Validation during development: 155 PCP tests passed (5 skipped). Eight-rank cache/COW/boundary checks passed. Full logits, index top-k and generated outputs matched an independent complete-cache reference, including prefix reuse. Original 128-request 8K-input benchmark: 19,831.93 token/s. The engine-side capacity estimator still uses the full logical pool and is addressed separately. Requires MUSA companion commit bdcf771f2dae05ac60d05505c1cec06e2769f744. Co-authored-by: Codex <noreply@openai.com> * fix(pcp): plan physical cache capacity and clear reused block tables Plan logical capacity from sharded attention pages, independent KDA pools and bounded page workspaces within the measured per-worker budget. Use the same accounting for admission, length estimates, allocation and reporting. Clear inactive columns when gathering sharded PCP block tables so a shorter request cannot materialize stale pages left by a previous long request. Validation: PCP and physical-planner regressions; prefill comparisons against independent 4e4fa1a075 sources covering 8K, chunked 32K, prefix reuse, 1M, short-after-long, fixed unequal batches and pool/page boundaries. Full 154880-wide logits, tokens, API logprobs, index top-k and effective KDA states match exactly with identical scheduling on all eight ranks. A 16K batch limit retains 1M context; the original 128-request benchmark reaches 21.18K/23.33K total tokens/s before/after a 1M request. Co-authored-by: Codex <noreply@openai.com> * fix(pcp): collect and remap only valid block-table prefixes Honor the block-table contract at the PCP reader: derive live page counts from effective sequence lengths and compressed chunk offsets, collect only those row prefixes, and remap into compact tables without reading stale tails. Keep root token lengths distinct from compressed pool lengths and include compression units in forward routing cache keys. Restore the original generic block-table gather interface and remove its PCP-specific tail clearing. Validation: PCP suite (166 passed, 6 skipped), targeted empty-page and routing regressions, and seven model scenarios including 1M, prefix hits, fixed unequal batches and pool/page boundaries. Full logits, generated tokens, API logprobs, index top-k and effective KDA states match 5bfdbc9be2 exactly on eight ranks. Original prefill benchmark after 1M: 23398.12 total tokens/s, 128 successes and no failures. Test services cleaned. Co-authored-by: Codex <noreply@openai.com> * refactor(pcp): resolve cache placement from layer specifications Declare page placement, query alignment and shared ownership groups in PCPCacheSpec at the model/cache-layer boundary. Keep GLM-specific support checks in the model implementation. Resolve the declaration in core and carry PCPPageLayout in KVCacheConfig. Worker allocation and PCP token partitioning consume this resolved layout rather than independently identifying the model. Budget bytes no longer select the cache layout. Replace layer-name-based owner pairing with explicit ownership-group validation. Keep capacity orchestration in core and make physical repacking operate on existing tensor descriptors. Remove the PCP allocation module's reverse imports from core. Preserve independent recurrent-state pools and support declared attention-only placement without assuming KDA exists. Validation: PCP, allocation, six packing-order, spec merge/serialization, and generic capacity regressions passed. Seven real model scenarios, including 1M, prefix reuse, unequal batches and pool/page boundaries, have identical full logits, output token IDs/logprobs, top-k and KDA states. Fresh-service benchmark: 21,356.13 total tokens/s, 128 successes, no failures. Original scripts unchanged; test services cleaned. Co-authored-by: Codex <noreply@openai.com> --------- Co-authored-by: Codex <noreply@openai.com>
KexyBiscuit
pushed a commit
to tai-core/vllm-tracking
that referenced
this pull request
Nov 5, 2026
… work (vllm-project#21) * feat(glm5next): shard prefill PCP cache pages Allocate uniquely owned MLA and compressed-index pages independently of DCP while retaining KDA head sharding. Compress complete pools once, finish partial pools on tail owners, and assemble native pages through bounded collectives or MUSA IPC. Preserve group-aware COW and account for page workspace capacity at initialization. Group engine logic under vllm.v1.pcp and provide logical DP8/KDA TP8 handoff descriptors with source-page lifetime tracking. Validation: 149 PCP CPU tests passed, 5 skipped on this branch. Eight-rank IPC, pool, prefix, batch, full-logit and KDA-state comparisons passed during development. The existing mainline FlashKDA interface is inherited from the base. Requires the MUSA IPC companion commit 15e0f5f3252bd14fe7c5f32223b6ab6f73878897. Co-authored-by: Codex <noreply@openai.com> * perf(glm5next): reuse PCP producer rows and balance pool-aligned work Reuse encoded MLA and index rows, fuse current/history composition with unique owner storage, and share immutable routes within one forward. Balance complete kpool spans independently of persistent page ownership to remove excess padding. Account for native IPC workspace without charging collective gather buffers. Use the existing mainline KDA implementation inherited from the base; this commit does not modify KDA backend selection or calling conventions. Validation during development: 155 PCP tests passed (5 skipped). Eight-rank cache/COW/boundary checks passed. Full logits, index top-k and generated outputs matched an independent complete-cache reference, including prefix reuse. Original 128-request 8K-input benchmark: 19,831.93 token/s. The engine-side capacity estimator still uses the full logical pool and is addressed separately. Requires MUSA companion commit bdcf771f2dae05ac60d05505c1cec06e2769f744. Co-authored-by: Codex <noreply@openai.com> * fix(pcp): plan physical cache capacity and clear reused block tables Plan logical capacity from sharded attention pages, independent KDA pools and bounded page workspaces within the measured per-worker budget. Use the same accounting for admission, length estimates, allocation and reporting. Clear inactive columns when gathering sharded PCP block tables so a shorter request cannot materialize stale pages left by a previous long request. Validation: PCP and physical-planner regressions; prefill comparisons against independent 4e4fa1a075 sources covering 8K, chunked 32K, prefix reuse, 1M, short-after-long, fixed unequal batches and pool/page boundaries. Full 154880-wide logits, tokens, API logprobs, index top-k and effective KDA states match exactly with identical scheduling on all eight ranks. A 16K batch limit retains 1M context; the original 128-request benchmark reaches 21.18K/23.33K total tokens/s before/after a 1M request. Co-authored-by: Codex <noreply@openai.com> * fix(pcp): collect and remap only valid block-table prefixes Honor the block-table contract at the PCP reader: derive live page counts from effective sequence lengths and compressed chunk offsets, collect only those row prefixes, and remap into compact tables without reading stale tails. Keep root token lengths distinct from compressed pool lengths and include compression units in forward routing cache keys. Restore the original generic block-table gather interface and remove its PCP-specific tail clearing. Validation: PCP suite (166 passed, 6 skipped), targeted empty-page and routing regressions, and seven model scenarios including 1M, prefix hits, fixed unequal batches and pool/page boundaries. Full logits, generated tokens, API logprobs, index top-k and effective KDA states match 5bfdbc9be2 exactly on eight ranks. Original prefill benchmark after 1M: 23398.12 total tokens/s, 128 successes and no failures. Test services cleaned. Co-authored-by: Codex <noreply@openai.com> * refactor(pcp): resolve cache placement from layer specifications Declare page placement, query alignment and shared ownership groups in PCPCacheSpec at the model/cache-layer boundary. Keep GLM-specific support checks in the model implementation. Resolve the declaration in core and carry PCPPageLayout in KVCacheConfig. Worker allocation and PCP token partitioning consume this resolved layout rather than independently identifying the model. Budget bytes no longer select the cache layout. Replace layer-name-based owner pairing with explicit ownership-group validation. Keep capacity orchestration in core and make physical repacking operate on existing tensor descriptors. Remove the PCP allocation module's reverse imports from core. Preserve independent recurrent-state pools and support declared attention-only placement without assuming KDA exists. Validation: PCP, allocation, six packing-order, spec merge/serialization, and generic capacity regressions passed. Seven real model scenarios, including 1M, prefix reuse, unequal batches and pool/page boundaries, have identical full logits, output token IDs/logprobs, top-k and KDA states. Fresh-service benchmark: 21,356.13 total tokens/s, 128 successes, no failures. Original scripts unchanged; test services cleaned. Co-authored-by: Codex <noreply@openai.com> --------- Co-authored-by: Codex <noreply@openai.com>
KexyBiscuit
pushed a commit
to tai-core/vllm-tracking
that referenced
this pull request
Nov 5, 2026
… work (vllm-project#21) * feat(glm5next): shard prefill PCP cache pages Allocate uniquely owned MLA and compressed-index pages independently of DCP while retaining KDA head sharding. Compress complete pools once, finish partial pools on tail owners, and assemble native pages through bounded collectives or MUSA IPC. Preserve group-aware COW and account for page workspace capacity at initialization. Group engine logic under vllm.v1.pcp and provide logical DP8/KDA TP8 handoff descriptors with source-page lifetime tracking. Validation: 149 PCP CPU tests passed, 5 skipped on this branch. Eight-rank IPC, pool, prefix, batch, full-logit and KDA-state comparisons passed during development. The existing mainline FlashKDA interface is inherited from the base. Requires the MUSA IPC companion commit 15e0f5f3252bd14fe7c5f32223b6ab6f73878897. Co-authored-by: Codex <noreply@openai.com> * perf(glm5next): reuse PCP producer rows and balance pool-aligned work Reuse encoded MLA and index rows, fuse current/history composition with unique owner storage, and share immutable routes within one forward. Balance complete kpool spans independently of persistent page ownership to remove excess padding. Account for native IPC workspace without charging collective gather buffers. Use the existing mainline KDA implementation inherited from the base; this commit does not modify KDA backend selection or calling conventions. Validation during development: 155 PCP tests passed (5 skipped). Eight-rank cache/COW/boundary checks passed. Full logits, index top-k and generated outputs matched an independent complete-cache reference, including prefix reuse. Original 128-request 8K-input benchmark: 19,831.93 token/s. The engine-side capacity estimator still uses the full logical pool and is addressed separately. Requires MUSA companion commit bdcf771f2dae05ac60d05505c1cec06e2769f744. Co-authored-by: Codex <noreply@openai.com> * fix(pcp): plan physical cache capacity and clear reused block tables Plan logical capacity from sharded attention pages, independent KDA pools and bounded page workspaces within the measured per-worker budget. Use the same accounting for admission, length estimates, allocation and reporting. Clear inactive columns when gathering sharded PCP block tables so a shorter request cannot materialize stale pages left by a previous long request. Validation: PCP and physical-planner regressions; prefill comparisons against independent 4e4fa1a075 sources covering 8K, chunked 32K, prefix reuse, 1M, short-after-long, fixed unequal batches and pool/page boundaries. Full 154880-wide logits, tokens, API logprobs, index top-k and effective KDA states match exactly with identical scheduling on all eight ranks. A 16K batch limit retains 1M context; the original 128-request benchmark reaches 21.18K/23.33K total tokens/s before/after a 1M request. Co-authored-by: Codex <noreply@openai.com> * fix(pcp): collect and remap only valid block-table prefixes Honor the block-table contract at the PCP reader: derive live page counts from effective sequence lengths and compressed chunk offsets, collect only those row prefixes, and remap into compact tables without reading stale tails. Keep root token lengths distinct from compressed pool lengths and include compression units in forward routing cache keys. Restore the original generic block-table gather interface and remove its PCP-specific tail clearing. Validation: PCP suite (166 passed, 6 skipped), targeted empty-page and routing regressions, and seven model scenarios including 1M, prefix hits, fixed unequal batches and pool/page boundaries. Full logits, generated tokens, API logprobs, index top-k and effective KDA states match 5bfdbc9be2 exactly on eight ranks. Original prefill benchmark after 1M: 23398.12 total tokens/s, 128 successes and no failures. Test services cleaned. Co-authored-by: Codex <noreply@openai.com> * refactor(pcp): resolve cache placement from layer specifications Declare page placement, query alignment and shared ownership groups in PCPCacheSpec at the model/cache-layer boundary. Keep GLM-specific support checks in the model implementation. Resolve the declaration in core and carry PCPPageLayout in KVCacheConfig. Worker allocation and PCP token partitioning consume this resolved layout rather than independently identifying the model. Budget bytes no longer select the cache layout. Replace layer-name-based owner pairing with explicit ownership-group validation. Keep capacity orchestration in core and make physical repacking operate on existing tensor descriptors. Remove the PCP allocation module's reverse imports from core. Preserve independent recurrent-state pools and support declared attention-only placement without assuming KDA exists. Validation: PCP, allocation, six packing-order, spec merge/serialization, and generic capacity regressions passed. Seven real model scenarios, including 1M, prefix reuse, unequal batches and pool/page boundaries, have identical full logits, output token IDs/logprobs, top-k and KDA states. Fresh-service benchmark: 21,356.13 total tokens/s, 128 successes, no failures. Original scripts unchanged; test services cleaned. Co-authored-by: Codex <noreply@openai.com> --------- Co-authored-by: Codex <noreply@openai.com>
KexyBiscuit
pushed a commit
to tai-core/vllm-tracking
that referenced
this pull request
Nov 5, 2026
… work (vllm-project#21) * feat(glm5next): shard prefill PCP cache pages Allocate uniquely owned MLA and compressed-index pages independently of DCP while retaining KDA head sharding. Compress complete pools once, finish partial pools on tail owners, and assemble native pages through bounded collectives or MUSA IPC. Preserve group-aware COW and account for page workspace capacity at initialization. Group engine logic under vllm.v1.pcp and provide logical DP8/KDA TP8 handoff descriptors with source-page lifetime tracking. Validation: 149 PCP CPU tests passed, 5 skipped on this branch. Eight-rank IPC, pool, prefix, batch, full-logit and KDA-state comparisons passed during development. The existing mainline FlashKDA interface is inherited from the base. Requires the MUSA IPC companion commit 15e0f5f3252bd14fe7c5f32223b6ab6f73878897. Co-authored-by: Codex <noreply@openai.com> * perf(glm5next): reuse PCP producer rows and balance pool-aligned work Reuse encoded MLA and index rows, fuse current/history composition with unique owner storage, and share immutable routes within one forward. Balance complete kpool spans independently of persistent page ownership to remove excess padding. Account for native IPC workspace without charging collective gather buffers. Use the existing mainline KDA implementation inherited from the base; this commit does not modify KDA backend selection or calling conventions. Validation during development: 155 PCP tests passed (5 skipped). Eight-rank cache/COW/boundary checks passed. Full logits, index top-k and generated outputs matched an independent complete-cache reference, including prefix reuse. Original 128-request 8K-input benchmark: 19,831.93 token/s. The engine-side capacity estimator still uses the full logical pool and is addressed separately. Requires MUSA companion commit bdcf771f2dae05ac60d05505c1cec06e2769f744. Co-authored-by: Codex <noreply@openai.com> * fix(pcp): plan physical cache capacity and clear reused block tables Plan logical capacity from sharded attention pages, independent KDA pools and bounded page workspaces within the measured per-worker budget. Use the same accounting for admission, length estimates, allocation and reporting. Clear inactive columns when gathering sharded PCP block tables so a shorter request cannot materialize stale pages left by a previous long request. Validation: PCP and physical-planner regressions; prefill comparisons against independent 4e4fa1a075 sources covering 8K, chunked 32K, prefix reuse, 1M, short-after-long, fixed unequal batches and pool/page boundaries. Full 154880-wide logits, tokens, API logprobs, index top-k and effective KDA states match exactly with identical scheduling on all eight ranks. A 16K batch limit retains 1M context; the original 128-request benchmark reaches 21.18K/23.33K total tokens/s before/after a 1M request. Co-authored-by: Codex <noreply@openai.com> * fix(pcp): collect and remap only valid block-table prefixes Honor the block-table contract at the PCP reader: derive live page counts from effective sequence lengths and compressed chunk offsets, collect only those row prefixes, and remap into compact tables without reading stale tails. Keep root token lengths distinct from compressed pool lengths and include compression units in forward routing cache keys. Restore the original generic block-table gather interface and remove its PCP-specific tail clearing. Validation: PCP suite (166 passed, 6 skipped), targeted empty-page and routing regressions, and seven model scenarios including 1M, prefix hits, fixed unequal batches and pool/page boundaries. Full logits, generated tokens, API logprobs, index top-k and effective KDA states match 5bfdbc9be2 exactly on eight ranks. Original prefill benchmark after 1M: 23398.12 total tokens/s, 128 successes and no failures. Test services cleaned. Co-authored-by: Codex <noreply@openai.com> * refactor(pcp): resolve cache placement from layer specifications Declare page placement, query alignment and shared ownership groups in PCPCacheSpec at the model/cache-layer boundary. Keep GLM-specific support checks in the model implementation. Resolve the declaration in core and carry PCPPageLayout in KVCacheConfig. Worker allocation and PCP token partitioning consume this resolved layout rather than independently identifying the model. Budget bytes no longer select the cache layout. Replace layer-name-based owner pairing with explicit ownership-group validation. Keep capacity orchestration in core and make physical repacking operate on existing tensor descriptors. Remove the PCP allocation module's reverse imports from core. Preserve independent recurrent-state pools and support declared attention-only placement without assuming KDA exists. Validation: PCP, allocation, six packing-order, spec merge/serialization, and generic capacity regressions passed. Seven real model scenarios, including 1M, prefix reuse, unequal batches and pool/page boundaries, have identical full logits, output token IDs/logprobs, top-k and KDA states. Fresh-service benchmark: 21,356.13 total tokens/s, 128 successes, no failures. Original scripts unchanged; test services cleaned. Co-authored-by: Codex <noreply@openai.com> --------- Co-authored-by: Codex <noreply@openai.com>
cxywhite
added a commit
to cxywhite/vllm
that referenced
this pull request
Nov 7, 2026
…context runtime environ mutation racing torch getenv)
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
The compilation time of flash-attn can be drastically reduced if ninja is installed. Related issue: Dao-AILab/flash-attention#150