diff --git a/.github/workflows/validate-skills.yml b/.github/workflows/validate-skills.yml index 20ed399..66e364b 100644 --- a/.github/workflows/validate-skills.yml +++ b/.github/workflows/validate-skills.yml @@ -32,7 +32,7 @@ jobs: if [[ "$filename" =~ [A-Z] ]] || [[ "$filename" =~ [[:space:]] ]] || [[ "$filename" =~ _ ]]; then invalid_files+=("$file") fi - done < <(find skills -name "*.md" -not -name "README.md") + done < <(find skills -name "*.md" -not -name "README.md" -not -name "SKILL.md") if [ ${#invalid_files[@]} -gt 0 ]; then echo "❌ The following files don't follow naming convention (lowercase with hyphens):" diff --git a/skills/upstream/vllm-ascend-releasing-note/SKILL.md b/skills/upstream/vllm-ascend-releasing-note/SKILL.md new file mode 100644 index 0000000..3c95f29 --- /dev/null +++ b/skills/upstream/vllm-ascend-releasing-note/SKILL.md @@ -0,0 +1,63 @@ +--- +name: vLLM Ascend Releasing Note Writer +description: You are a releasing note writer for vLLM Ascend project (vllm-project/vllm-ascend). You are responsible for writing release notes for vLLM Ascend. +--- + +# vLLM Ascend Releasing Note Writer Skill + +## Overview +You should use the `ref-past-release-notes-highlight.md` as style and cateogry reference. Always read these first. + +## When to use this skill +When a new version of vLLM Ascend is released, you should use this skill to write the release notes. + +## How to use it +0. all ouput files should be saved under `vllm-ascend-releasing-note/output/$version` folder + +1. Use the `fetch_commits-optimize.py` script to fetch the commits between the previous and current version. + +``` +uv run python fetch_commits-optimize.py --base-tag $LAST_TAG --head-tag $NEW_TAG --output 0-current-raw-commits.md +``` +`0-current-raw-commits.md` is your raw data input. + +2. Use the `commit-analysis-draft.csv` tool to analyze the commits and put them into the correct section. +`1-commit-analysis-draft.csv` is your workspace for commit by commit analysis for which commit goes into which section, whether can be ignored, and why. You can create auxilariy files in `tmp` folder. + * You should check each commit. They are put into rows in the CSV file. + * The CSV should have headers `title`, `pr number`, `user facing impact/summary`, `catgory`, `decision`, `reason`. Please brainstorm other fields as you see fit. + +3. Draft the highlights note, and save it to `2-highlights-note-draft.md`. +4. Edit the draft highlights note in `2-highlights-note-draft.md`, and save it to `3-highlights-note-edit.md`. You should double and triple check with the raw commits + analysis. You can leave any uncertainty and doubts in the file, and we will discuss them together. +5. Use the format `This is the $NUMBER release candidate of $VERSION for vLLM Ascend. Please follow the [official doc](https://docs.vllm.ai/projects/ascend/en/latest) to get started.`. + +## Writing style +1. To keep simple, you should only save one level of headings, starting with ###, which may include the following categories follow below order: +### Highlights +### Features +### Hardware and Operator Support +### Performance +### Dependencies +### Deprecation & Breaking Changes +### Documentation +### Others + +2. Additional Inclusion Criteria +* User experience improvements (CLI enhancements, better error messages, configuration flexibility) +* Core feature (PD Disaggregation, KVCaceh, Graph mode, CP/SP, quantization) +* Breaking changes and deprecations (always include with clear impact description) +* Significant infrastructure changes (elastic scaling, distributed serving, hardware support) +* Major dependency updates (CANN/torch_npu/triton-ascend/MoonCake/Ray/transformers versions, critical library updates) +* Binary/deployment improvements (size reductions, Docker enhancements) +* Default behavior changes (default models, configuration changes that affect all users) +* Hardware compatibility expansions (310P, A2, A3, A5 support) +In the end we don't want to miss any important changes. But also don't want to spam the notes with unnecessary details. + +3. Section Organization Guidelines +* **Model Support first**: Most immediately visible to users, should lead the highlights +* **Group by user impact**: Hardware/performance should focus on what users experience, not internal optimizations +* **Provide usage context**: Include relevant flags, configuration options, and practical usage information +* **Technical detail level**: Explain what features enable rather than just listing technical changes + +4. Writing Tips +* Look up the PR if you are not sure about the details. The PR number at the end (#12345) can be looked up via vllm-project/vllm#12345. To get the description, you just need to call https://api.github.com/repos/vllm-project/vllm/pulls/12345 and look at the body field. +* When writing the highlights, don't be too verbose. Focus exclusively on what users should know. \ No newline at end of file diff --git a/skills/upstream/vllm-ascend-releasing-note/output/v0.13.0/0-current-raw-commits.md b/skills/upstream/vllm-ascend-releasing-note/output/v0.13.0/0-current-raw-commits.md new file mode 100644 index 0000000..b0ef825 --- /dev/null +++ b/skills/upstream/vllm-ascend-releasing-note/output/v0.13.0/0-current-raw-commits.md @@ -0,0 +1,994 @@ +[CI]Fix test nightly workflow. (#3603) +Reapply "[MoE] [Refactor] Remove manual memory cleanup (#3365)" (#3483) (#3365) +fix : support chunked_prefill with deepseek_mtp (#2711) +[Misc] clean up useless function (#3348) +[Feat] Dynamic Batch Feature (#3490) +[Feat] add native kvcache offload (#3433) +[CI] Multi-Node CI scalable (#3611) +clean up uesless ut test (#3622) +unify logic between aclgraph and torchair (#3560) +[Fix] Fixes attribute error in MLA implementation (#3618) +[BugFix][main] Fix quantization related mtp bug with patch (#3620) +[Feat] Prefetching Attention QKV Linear Weight With `AddRmsNormQuant` Custom Op (#3517) +[Doc] Upgrade docker run command (#3645) +[main][refactor] refactor SequenceRowParallelOp forward (#3616) +[Doc] Update the modelslim website from gitee to gitcode. (#3615) +[BugFix] fix deepseek torchair precision (#3624) +perf : optimize memory for deepseek mtp (#2713) +[Misc] Add a model loader that utilizes HCCL for weight loading (#2888) +[TEST]Add initial multi modal cases for nightly test and deepseek-r1 tests (#3631) +[Structured Output] Replace `apply_grammar_bitmask()` method with that in vllm to avoid maintenance (#2524) +[Bugfix] fix delay free prefill req & D node support prefix cache (#3607) +[Doc] Update the Pangu Pro MoE tutorials. (#3651) +[Test] add a new Qwen3-32b-int8 test case with feature_stack3 (#3676) +[main][bugfix] Add 'layer_type' param to get_pergroup_param() for compatibility (#3682) +[BugFix]fix deepseek torchair recompile (#3678) +support cp&dcp (#3260) +[MoE][Multistream] Avoid performing communication in extra stream. (#3582) +[Benchmark] Upgrade benchmark args for new vllm version (#3218) +[UT] Fix test_sample_recovered_tokens_pytorch_autoregressive (#3434) +remove useless code (#3685) +[TEST]Add initial prefix cache case for nightly test (#3709) +[1/N][Refactor] Refactor code to adapt with vllm main (#3612) +[BugFix] Check all expert maps when using muilty instance. (#3576) +[TEST]Add initial multi modal cases of Qwen2.5-VL-32B-Instruct for nightly test (#3707) +[Bugfix] The server fails to locate the request, leading to the server hanging. (#3703) +[Main][Perf] Add fused matmul/reduce-scatter kernel for performance optimization. (#3693) +[Refactor] Refactor Ascend attention implementation forward (#3714) +[Feat] Add mrope fusion op (#3708) +[BugFix][P/D] Modify the recalculation logic to prevent waiting requests from filling up the D node KVCache (#3641) +[CI][Doc] Optimize multi-node CI (#3565) +[UT][fix] Add missing get_ascend_config mock to NPUWorker initialization tests (#3729) +[Misc] Limit ray version (#3660) +[CI] Skip ops test for e2e (#3665) +Update version doc (#3599) +[BugFix][Core] Fix a bug running multi-modal with ascend_scheduler (#3675) +[Bugfix] Fix zero attention output in qwen3-next (#3572) +[Refactor] optimize _prepare_inputs method in eagle_proposer (#3296) +[BugFix] Comment out newly added vlm e2e. (#3736) +[Test] Add e2e test and accuracy test for Qwen3-Next-80B-A3B-Instruct (#3450) +[Doc] Update supported models (#3481) +[Refactor] [MoE] Rename moe-related classes & files (#3646) +[Test] add test for prefix cache feature of deepseek (#3733) +[bugfixfix] correct _register function place for mooncacke (#3747) +Upgrade to new vllm commit (#3719) +[main] remove dbo code (#3712) +add qwq testcase (#3757) +[BugFix] Fix Qwen3-next break (#3428) +[Installation] limit opencv-python-headless version to resolve numpy version conflict (#3713) +[feat]dcp pcp support aclgraph (#3731) +[CI] Add custom op to nightly (#3765) +[CI] Enable 2 jobs for nightly test (#3781) +Bump actions/download-artifact from 5 to 6 (#3787) +Bump actions/upload-artifact from 4 to 5 (#3786) +[Doc][Example][Bugfix] Elements in local_device_ids should be casted … (#3782) +[bugfix][main]fix proxy decode bug (#3750) +[MM][Doc] Update online serving tutorials for `Qwen2-Audio` (#3606) +support prefill cache mode use fia op (#3696) +[Doc] Update FAQ (#3792) +【Bugfix】bugfix for weight load of kimi-k2 (#3798) +[TEST]Add 2P1D multi node cases for nightly test (#3764) +[CI] Add multi-node test case for a2 (#3805) +Upgrade to 0.11.1 newest vllm commit (#3762) +[CI] Fix nightly CI (#3821) +[Main][Bugfix]Avoid using the fusion operator in the MOE model (#3834) +[TEST]Add aisbench log and A2 cases (#3841) +[long_seq_optim] BSND to TND and FA_UPDATE replacement (#3778) +[P/D] force with_prefill true after allreduce in kv producer (#3768) +fix qwen3next full graph break. (#3812) +[Doc] Update doc (#3836) +[HybridKV][Bugfix] Fix Hybrid kvcache sharing bug in same attention type (#3760) +[Bugfix] [MoE] fix error in deepseek when using allgather (#3824) +[Perf] Delete redundant operations in model_runner and forward_context (#3677) +[CI]pin vllm commit id (#3861) +[CI] Optimize nightly CI (#3858) +[BugFix] deepseek torchair adapt for torch_npu version (#3862) +[CI]Fix eplb nightly tests. (#3863) +fix mooncake layerwise connector (#3849) +bugfix for mtp fullgraph (#3845) +[BugFix] Fix mlapo accuracy problem related with weight processing. (#3850) +[CI]Fix oom of deepseek-eplb nigtly test. (#3884) +Add FAQ for docker pull error on Kylin OS (#3870) +[UT] fix skip ut test for test_utils (#3803) +[Doc] Remove modeling doc (#3789) +[Build] Force torch version (#3791) +[FEAT] Refactor spec decode to support efficient padded speculation (#3528) +[Model][3/N] Refactor sfa into mla and remove deepseek_v3_2.py (#3769) +[feature] Prompt Embeddings Support for v1 Engine (#3026) +[TEST]Add MALPO for aclgraph in nightly test (#3894) +[BugFix]Fix group list type of mc2. (#3864) +[bugfix] layerwise D first plan (#3866) +[CI] Optimize nightly CI (#3898) +[bugfix]cancel tokenize for layerwise_proxy (#3914) +add new e2e tests case for aclgraph memory (#3879) +mfix bug when max_seqs=14 in mtp=2 scenario and raise error when cudagraph_capture_sizes can't be an integer multiple of uniform_decode_query_lentp (#3910) +[Bugfix] Fix MTP support for lmhead_tensor_parallel_size (#3915) +[Test] Add new test model for aclgraph single_request (#3888) +[main][bugfix] fix valueError in static_forward_context when prefix is empty (#3924) +[E2E][MM] Add e2e tests for InternVL model (#3796) +[feature] support pcp + mtp (with pd disaggregate) (#3822) +[Doc] Update doc for release notese (#3853) +Update torch-npu version to 2.7.1 (#3896) +[CI][Nightly] Correct the commit hash available for mooncake (#3943) +[Perf] Move attention update stream out of loop to optimize performance (#3848) +[Feat][UT] Support Deepseekv32 FULL_DECODE_ONLY mode and add unit test of sfa_v1 (#3763) +correct bug to fix the value of max_num_tokens (#3933) +[CI][Nightly] Fix mooncake build (#3958) +[Test] Add new e2e test use deepseek-v2-lite in ge graph mode (#3937) +revert TND modify when dcp pcp (#3948) +Quality enhancement: Immediately interrupt execution when memory OOM (#3932) +[Test] Add accuracy test for qwen3-8b-w8a8 (#3799) +[BugFix] Fix deepseek v3.2 mtp bug. (#3900) +[Test]Add accuracy test for multiple models (#3823) +[PD Disaggregation]Set adxl engine as default backend and update README (#3761) +[TEST]Add full graph for multimodal nightly tests (#3968) +[Perf] move quant before allgather in Allgather EP (#3420) +[ModelRunner][Refactor] Refactor kv cache tensor initialization logic (#3106) +[Test] Add accuracy test for qwen3-30b-a3b-w8a8 (#3807) +[Doc] Refactor the DeepSeek-V3.2-Exp tutorial. (#3871) +support qwen3-next full_decode_only mode. (#3949) +[docs] add aclgraph developer guide (#3683) +[Doc] Update version policy (#3999) +[Doc] add mtp doc (#3770) +[docs] Add kv pool developer guide (#3752) +[Doc]Add developer guide of eplb. (#3759) +[main][doc][kv_pool]Add adxl timeout parameter in kv pool user guide (#4012) +[Test] Refactor accuracy test to nightly test (#3814) +[P/D]Make kv-transfer env variable take effect & Fix load-balance proxy (#3981) +[Feat](Mooncake) Supports multiple input suffixes for global_segment_size (#3690) +[feat]decode convert bsnd to tnd and fix bug when pcp and dcp (#3980) +[TEST]Update nightly acc test standard (#4032) +[CI] Quick fix mooncake for nightly-ci (#4028) +[Bugfix] Add constraints for sequence parallelism (#4014) +[BugFix][main] Adapted to torch_npu.npu_fused_infer_attention_score (#4025) +[main][bugfix] Fix a rare bug triggered by _npu_paged_attention in FULL_DECODE_ONLY mode (#3986) +[long_seq] fix A2 accuracy problem (#4030) +[Feat] update op for mla (#4000) +[UT] Add new ut case for aclgraph in auto enable (#4031) +[Doc] Add model feature matrix table. (#4040) +[Feat] Adapted mtp function to Qwen3-next (#3918) +[BugFix]Fix group list type of mc2. (#4047) +[CI]Fix eplb ci. (#4052) +[Bugfix] fix sleepmode level2 e2e test (#4019) +[P/D][BugFix]Fix proxy format processing errors & Layerwise connector performance optimization (#4043) +[BugFix] Improve the performance of prefixcache features (#4022) +[Bugfix]fix pcp dcp attn aclgraph (#4066) +[Info][main] Corrected the errors in the information (#4055) +[TEST]Add qwen3-235b-w8a8 and qwen3-30b-w8a8 nightly test (#3973) +[Doc] Remove extra MLAPO installation step for DeepSeek-V3.2. (#4024) +[docs] [P/D] add feature guide for disaggregated-prefill (#3950) +[Feat] flashcomm_v2 optim solution (#3232) +[Feature][Build] Upgrade the minimum version to 3.10 (#3926) +[Fix] fix Qwen2-Audio-7B-Instruct accuracy test (#4017) +[Typo] LLama has been changed to Llama (#4089) +[Core] Restore scheduling logic under default configuration (#3967) +[Doc] add qwen3 w4a4 tutorial (#4076) +[BugFix] Fixes Qwen3-Next enable nz accuracy problem (#4058) +[Doc] Add release note for v0.11.0rc1 (#3931) +[main][Bugfix] Fix ngram precision issue and open e2e ngram test (#4090) +[feature] chunkprefill support pcp&dcp (#3801) +[Doc] Recover installation doc to use pip install (#4109) +[Test] Add nightly test for DeepSeek-V3.2-Exp (#3908) +[Fix] Refactor and fix dist test to e2e full test (#3808) +[Fixbug] Fix ut test (#4116) +Remove VLLM_USE_V1 (#4086) +[CI] Integrate mooncake to vllm-ascend base image (#4062) +[TEST]Update nightly cases and add mtpx (#4111) +oproj TP support acl graph (#4073) +[Test][Accuracy] Add accuracy evaluation config for InternVL3_5-8B (#3964) +[Misc][Doc] Add service profiling feature with user guide (#3756) +[Perf] Remove D2H operations to imporve performance (#4063) +[Doc] Fix DeepSeek-3.2-Exp doc, remove v0.11.0rc0 outdated infos. (#4095) +fix fullgraph in ds. (#4016) +[feature] support pcp + mtp (in pd co-locate scenario) (#4098) +[main][bugfix] Change seq_lens in dummy attn_metadata to max_query_len (#4097) +[CI] Fix nightly-ci (#4159) +Upgrade to 0.11.1 newest vllm commit (#3982) +[Perf] fix async copy for async scheduling (#4113) +[Perf] [MoE] optimize all2allv (#3738) +[Bugfix] fix mtp profile run error where main model and mtp model use different quantization (#4102) +[BugFix] adapted e2e tests for Qwen3-next-mtp (#4160) +[BugFix] Fix kv_no_split not contiguous (#3594) +[Info][main] Correct the mistake in information documents (#4157) +[Test]Add ut test qwen3_moe and sfa (#4121) +[CI] Remove unsupported python 3.9 format check (#4172) +[CI] Add daily images build for nightly ci (#3989) +[long_seq_Feat] support chunk prefill (#4158) +[CI] Add multi-nodes EPLB configs of DeepSeek-R1-W8A8 & Qwen3-235B-W8A8 (#4144) +[Bugfix] fix cannot import name get_mp_context (#4174) +[Feat] Adds a utility for printing from within ACL graphs (#4162) +[Platform] Add import_kernels interface (#3694) +[Misc] Add benchmark results into `.gitignore` (#4200) +[Test] Add deepseek v3.2 exp nightly test (#4191) +[CI] Fix no space left in build wheel CI. (#4215) +support FULL graph mode for GQA (#3970) +[TEST]Update prefixcache perf threshold for qwen3-32b-int8 (#4220) +make vllm-ascend work well in developer mode (#4179) +[Bugfix]Fix moe error when sp chunked the hidden_states (#4212) +[main][misc]change default capture size for Qwen3-MoE when using full dp (#4199) +[feature] Mooncake_connector support pcp/dcp (#4183) +[P/D] pd proxy support ipv6 (#4161) +[bugfix] fix proxy hen host ip using domain name (#4243) +[Fix] Sorts aclgraph batch sizes in ascending order (#4230) +[refactor]support gatingtopk operator generalization (#2958) +[bugfix] pcp + mtp acl graph bugfix (#4221) +[CI] Fix kubernetes failed to resolve ip by dns name (#4240) +[Bugfix] fix hang in async scheduling (#4233) +remove get_metadata_cls (#4087) +[doc]fix readme for kv pool user guide (#4271) +[Docs] Improve the AISBench multi-modal testing docs (#4255) +[misc] clean up get_metadata_cls (#4276) +[long seq feat]GQA support long-prefill-token-threshold and fixbug (#4209) +[Bugfix] fix nightly multi-node EPLB tests' "DYNAMIC_EPLB=true" environment not working (#4223) +avoid mrope fusion op when running qwen2.5-vl on a+x machine (#4270) +[Test] Add tests for the multi-node DeepSeek-V2-Lite network in GE Graph (#4039) +[CI] Add mla ut (#4280) +[Feat] Support MTP to running in full graph mode (#3892) +[Test] quick fix mla ut (#4318) +[Test] Add ACL graph capture/replay DP test (#4259) +[Feat][BugFix]Support the Qwen3-Next-80B-A3B-Instruct quantization model&Fix the NZ issue (#4245) +eplb redundant expert bugfix (#4291) +[Readme] EPLB Support Scenarios (#4314) +[MM][Bugfix] Add error log for VL models when enabling FLASHCOMM (#4272) +[Feat][Doc] Add a load_balance_dp_proxy in examples and external dp doc. (#4265) +[Test] Add ut test for torchair (#4287) +[bugfix] bugfix for PD disaggregate (#4319) +[EPLB] Eplb Verify Fix (#4333) +[Doc] add release note for v0.11.0rc2 (#4348) +[BugFix] Fix some issues caused by the ascending order of cudagraph_capture_sizes (#4338) +[Bugfix][KV Pool]fix get_ip import in mooncake_store (#4355) +[Doc]Add single node PD disaggregation instructions (#4337) +[CI] Fix nightly CI for A2 series (#3825) +[Doc] Upgrade multi-node doc (#4365) +Change the first letter to uppercase (#4375) +[Fix] Remove unnecessary NPU synchronization in MTP proposer (#4325) +[TEST]Update deepseek mtpx acc cases standard (#4321) +Drop 0.11.0 support (#4377) +[Fix] fix aclgraph e2e test. (#4131) +[Refactor] remove moe type of multicast. (#4224) +[Bugfix][MoE] enable force_load_balance in aclgraph (#4366) +[feature] vllm-ascend support msprobe (eager mode dump) (#4241) +Bump actions/checkout from 4 to 6 (#4380) +[Bugfix]Fix the hang issue of multimodal model when running with DP>1 (#4392) +Document error correction (#4422) +[Bugfix] fix patch typo (#4351) +[bugfix]Return the Transformer version from 4.57.2 to 4.57.1 (#4423) +[Bugfix] use module-level import for patched function in Qwen3Next (#4354) +[MM][Bugfix] Minor fix for VL model verification (#4384) +[misc] Remove useless patch_logits (#4252) +[TEST] Delete Comment (#4427) +mkdir triton package and move triton files (#4420) +upgrade to vllm 0.11.2 (#4400) +[CI] clean up ci (#4452) +[refact] unified soc_version code (#4359) +Change comment location (#4432) +[UT] Fix ut test (#4472) +chip type judgement code optimization (#4485) +[CI][Nightly] Support local debugging for multi-node CI test cases (#4489) +[BugFix] Adapted Qwen3-Next eager mode to v0.11.2 (#4477) +[bugfix] fix ray start failed: local_world_size cannot little than visible device count error (#4457) +[feature] Add Custom Op grouped_matmul_swiglu_quant (#4431) +[TEST] Add eagle proposer ut (#4447) +[main]Upgrade cann to 8.3rc2 (#4350) +[Quantization] Support compressed tensors w8a8 static and w8a8 dynamic weight (#4036) +[MM][Model][Perf] Remove Qwen2.5-VL modeling files and add patch for VisionAttention (#4349) +[P/D] Add readme for PD separation (#4182) +[Doc]Delete equals sign (#4537) +[Kernel] add custom op GmmSwigluQuantWeightNzTensorList (#3804) +[Feature][main]reconstruction kvpool connector to ascend connector (#4438) +【OPS】qwen3-next support triton chunk_gated_delta_rule ops (#4070) +update triton package url (#4552) +[Bugfix] Fix model run _npu_flash_attention hang issue (#4410) +[Doc] Add single NPU tutorial for Qwen2.5-Omni-7B (#4446) +Update triton package name (#4563) +[bugfix] dep ineffective (#4417) +[P/D] [bugfix] add get_kv_connector_handshake_metadata func for 0.11.2 (#4567) +drop ascend scheduler (#4498) +improve soc version (#4522) +[MM][Model] Remove Qwen2-VL modeling files (#4534) +Move mla to ops module (#4575) +[Bugfix] fix dp parallel + tp > 1 offline inference port conflict (#4539) +remove qwen3-next model file (#4573) +[feature]Pooling Features and PCP Adaptation (#4143) +Revert "drop ascend scheduler" (#4580) +[CI] Skip test_ngram_correctness as the oom issue block CI (#4578) +[bugfix] Repair the problem of moe model accuracy caused by version upgrade. (#4562) +[Bugfix] Fix kvpool precision synchronization (#4574) +[feature] Support W8A8 PD-Mix Quantization (#4235) +[EPLB][Ops] Integerate grouped_matmul_swiglu_quant_weight_nz_tensor_list operator into dynamic EPLB (#4216) +[OPS] add bmm_transpose ops (#3990) +[BugFix] Fix Qwen2.5_Omni vision customized op attr err (#4568) +[Bugfix] Resolve MTP > 1 issue when lm head tp > 1 (#4254) +Bump actions/setup-python from 6.0.0 to 6.1.0 (#4591) +[Bugfix] Fix bug with establishing the flashcomm2 and pp communication domains. (#4458) +[Kernel] add triton kernels for sampling (#4550) +add _cann_ops_custom gitignore (#4605) +[Feature] Integrate Suffix Spec Decoding (#4045) +upgrade torch npu version (#4433) +[Bugfix] PCP adaptation for VLLM v0.11.2 modifications (#4604) +[Bug_fix] fix torchair o_proj forward parameter (#4166) +[CI] drop ascend scheduler test (#4582) +[Feat] shared expert dp for deepseek_mtp (#3811) +fix qwenvl pd smoke test error (#4597) +[Test] Add accuracy nightly test for new models (#4262) +[Doc] Fix DeepSeek-V3.2-Exp doc, add docker command. (#4479) +[Test] Add GLM-4.5 nightly test (#4225) +[Bugfix] Remove ModelSlim-"M4 Quantization". (#4589) +[main][bugfix] bugfix for qwen3 moe quantization (#4599) +[MM][Model] Remove Qwen3-VL modeling files (#4577) +[CI]enable chunked prefill by default (#4569) +[Refactor] Remove redundant attention operator branches. (#4531) +[Bugfix] Fix Qwen2.5-Omni-7B accuarcy test (#4556) +[Bugfix]Fix eplb enable when using mtp float weights. (#4571) +Bump actions/checkout from 4.3.1 to 6.0.0 (#4592) +Revert "[Bugfix] Fix Qwen2.5-Omni-7B accuarcy test (#4556)" (#4556) +[CI] Drop ascend scheduler from test (#4613) +add hyperlink (#4588) +[Doc]clean up ascend scheduler config from doc (#4612) +[Doc] Add tutorial for Qwen3-Coder-30B-A3B (#4391) +[Ops][Triton] Add a triton kernel supporting partial rope. (#4413) +clean up model module (#4611) +[Doc] Refactor the DeepSeek-V3.1 tutorial. (#4399) +[performance] Enhance performance after enabling min_p (#4529) +【doc fix】doc fix: deepseekv3.1 (#4645) +[Bugfix] fix custom op GmmSwigluQuantWeightNzTensorList (#4593) +upgrade vLLM to main (#4608) +[kernel] add AscendC op: lightning_indexer and sparse_flash_attention (#4625) +[Doc] add release note for v0.11.0rc3 (#4646) +fix typo (#4657) +[Model] Add qwen3Next support in Main (#4596) +[Feat] MTP support DeepSeekV3.2 (#4465) +[Fix] Fix FIA `query` and `query_start_loc` shape mismatch error (#4518) +[CI] Fix ut ci: no space on the device (#4662) +[Misc] Add cann custom ops to `.gitignore` (#4670) +fix custom ops env set error (#4675) +[Core] Encoder separation for Encode-Prefill-Decode Disaggregation (#4176) +upgrade vLLM to 0.12.0 tag (#4647) +Remove cancel for main to main check (#4685) +Adopt inductor fusion and define quantization fusion pass (#4168) +Remove ascend schuduler ut (#4684) +【fix】ops gatingtopk fix nightly ci error (#4340) +[MM][Patch] Remove patch for cos/sin cache (#4672) +[Nightly] Optimize nightly CI (#4509) +[Misc] Upgrade vllm vllm commit to 2025_12_04 (#4690) +add `dispatch_gmm_combine` kernel (#3532) +[Bugfix] Quick hot fix for nightly CI (#4727) +[Doc] Update vLLM version in doc (#4691) +Drop ascend scheduler (#4623) +[long_seq] remove long_seq env (#4660) +Update comment doc (#4731) +[BugFix][Triton] Fix ub overflow bug of sample_recover_tokens_kernel (#4673) +[Bugifx] fix quant_apply_mlp w1_scale type error & fix getting num_local_expert (#4632) +[P/D][main] Clean connector history information (#4650) +[CI] Fix unit test fault `no space left` (#4728) +【main】[Doc]add 2P1D instruction for single node (#4716) +[Refactor] 1/N Refactor attention_v1 & extract attention_cp (#4628) +[Bugfix]fix bmm_transpose ops for cann version (#4653) +rm vanilla attn (#4558) +mlapo add qdown output (#4707) +[Bugfix] fix mtp and eagle aclgraph bug (#4710) +support async mtp (#4511) +[BugFix] Fix eagle3 accuracy problem when enforce_eager=True (#4521) +[Kernel] add custom op DispatchGmmCombineDecode (#4139) +[Feat]enable sfa cp for dsv3.2 (#4702) +Support DeepSeekV3.2 with MLAPO operator (#4753) +[P/D] check kv extra config and del hccl backend (#4547) +[BugFix] Refactor ACL graph size adjustment for speculative decoding (#4640) +[Feat] Add Euler xlite graph wrapper support (#4526) +fix synchronize error of exceeds_max_model_len d2h copy (#4708) +[CI] Fix ngram & suffix test oom (#4755) +Deepseek Mtp model uses the lm_head and embedding from the main model (#2790) +remove useless patch (#4699) +[Op] DeepSeekV3.2 support bmm_transpose operator (#4631) +[EPLB] Add log Info for moe_load Imbalance Ratio (#4482) +[Fix] skip xlite e2e test (#4786) +[Bugfix] Fix Dcp dimension mismatch when enable Mlapo (#4687) +[Kernel] add custom moe ops for prefill (#4194) +Bump actions/checkout from 6.0.0 to 6.0.1 (#4772) +[Doc] Add Qwen3-235B tutorial (#4358) +[DP] Fix dp padding logic in dummyrun (#4705) +[MOE]move weight transpose to wakeup for RL secnarios (#4626) +Fix incorrect MLAPO weight release in PD mixex scenarios. (#4774) +Revert "[Kernel] add custom moe ops for prefill" (#4806) +[Bugfix] Add the check for a null VllmConfig (#4749) +[CI] Skip `test_suffix_correctness` (#4820) +[Docs]fix the configuration conflicts in documentation (#4823) +[CI] Optimize CI time (#4821) +[KVPOOl]Support pp (#4761) +[Feat] Multi-stream for eplb heat collection and aggregation (#4214) +[kernel] Adapt DispatchGmmCombineDecode operator to parameters of small operators (#4790) +[CI] Increase HCCL_BUFFSIZE for A3 (#4838) +[Bugfix]fix bmm_transpose ops in dsv32 (#4791) +[UT]add pcp aclgraph ut (#4804) +[Usability]local_buffer_size support for units: GB, MB, KB, B (#4829) +[Refactor] 2/N Unify all mask generation methods and cache mask (#4779) +[CI] Setup github proxy for self_hosted runners (#4841) +[Fix] Add extra warmup run count for MC2 on specific SoC version (#4843) +[Bugfix] Disable the dispatch_ffn_combine kernel in MTP path (#4751) +[P/D][main]Offline the llmdatadist connector related parts of the code and files. (#4780) +Add gsm8k accuracy test for multi-note Qwen3-235B-A22B (#4802) +[bugfix] fix quant method validation bug (#4831) +[Kernel] add custom op MatmulAllreduceAddRmsnorm (#4606) +Drop torchair (#4814) +[Nightly] Optimize nightly online test logger info (#4798) +[Test] Temporarily skips Qwen3-30B-A3B-W8A8 data parallel test case (#4857) +add e2e test for mtp async_scheduling (#4826) +[Model] Support pooling models (#3122) +[CI]Cleanup accurary test (#4861) +[CI] Use offline mode for modelscope (#4875) +[Feat] Support native Kimi-K2-Thinking native W4A16 quantized experts weights (#4516) +mooncake connector support pipeline parallel & fix pp with flashcomm1 (#4054) +add multi_npu_qwen3_dense tutorials (#4543) +[CI] fix lint (#4888) +[Kernel] Add moe normal ops (#4810) +[Bugfix] Fix out-of-bounds access to token_id due to uninitialized logprobs (#4248) +[FEAT] Support DeepSeek-V3.2 with `FULL_DECODE_ONLY` mode (#4706) +Fixed the performance degradation issue in post-processing in speculative decoding scenarios. (#4849) +[Bugfix] Support for mlapo in deepseekv3.1 w4a8 (#4828) +[Feature] Support npuhraph_ex backend (#4700) +[perf][dsv3.2][async_scheduling] improve dsv3.2 performance by eliminating HD synchronization (#4805) +[BugFix][main] Adapted Qwen3-Next-MTP to chunked prefill (#4770) +Update patch doc (#4869) +Remove COMPILE_CUSTOM_KERNELS env (#4864) +Remove VLLM_ASCEND_ENABLE_TOPK_TOPP_OPTIMIZATION (#4860) +Remove useless env (#4858) +add DeepSeek-R1 tutorial. (#4666) +[Doc]Add tutorial document for qwen-VL-Dense (#3516) +[Doc] Add local running multi-node nightly test case guide (#4884) +[E2E] Remove unused PD-disaggreate scripts in E2E test. (#4837) +[E2E] Refactor the e2e testcases. (#4789) +[E2E] Optimize nightly testcase. (#4886) +[feat] mlapo add bf16 no_quant support (#4852) +cleanup useless torchair logic (#4856) +[Feat] Flashcomm2 use o_shared linear (#4188) +[OPS] support triton causal_conv1d_fn ops (#4119) +【Bugfix】bugfix_for_bmm_transpose (#4899) +[Bugfix] Fix the bug in sfa-cp under multi-DP scenarios. (#4850) +[feat] apply flashcomm1 on bailing (#4868) +[Bugfix] support mtp kv transfer and pp partition by hand in kv transfer (#4892) +[doc] Add Qwen2.5 tutorials (#4636) +[Fix] Delete redundant variable (#4903) +[Fusion] normalize fusion naming and enable e2e test (#4693) +[Bugfix] Prevent engine hang during KVCacheSendingThread startup (#4754) +[CI] speed up ut (#4901) +Remove mindie_turbo (#4896) +[Doc] Update structured output doc with upstream link (#4015) +Refactor CI workflow (#4912) +[CI] Cancel whl build when submitting a new commit (#4925) +[CI]cleanup e2e test (#4800) +[Doc] Update tutorial index (#4920) +[bugfix][refactor] fix recompute_scheduler break with vllm 0.12.0 & support async scheduling & refactor recompute_scheduler.py (#4895) +[Performance] Pre-issued exponential distribution operator. (#4908) +[CI] refect e2e test (#4799) +[MoE][TorchAir] Remove FusedMoEState (#4927) +[main][Bugfix] Remove the ZMQ communication setup on the D node (#4926) +[Feat] Add custom Embedding tensor model parallel (#2616) +[Bugfix] bugfix for moe_mlp (#4822) +BugFix: Resolve PolicyFlashlb warm up function attribute error (#4741) +[CI] fix light test (#4954) +【doc】Add model feature matrix (#4950) +[Doc] Upgrade outdated doc (#4957) +update qwen2.5vl readme (#4938) +vllm-ascend support Ascend950 with Qwen dense model. (#4228) +[usability]Modify the default value of the protocol to ascend (#4959) +[Nightly] Remove gen_ranktable logic (#4941) +[Feature] model_runner refactor (#4764) +[doc][main] Correct mistakes in doc (#4945) +[CI] Add mtp_proposer ut (#4397) +[Bugfix] Pass vllm_config to kv_connector_no_forward in NPUModelRunner (#4970) +[Bugfix] fix eagle proposer (#4971) +[bugfix] asyncscheduler bug fix (#4968) +[doc][main] Correct more doc mistakes (#4958) +Revert "[Bugfix] support mtp kv transfer and pp partition by hand in kv transfer (#4892)" (#4892) +[perf] replace all_reduce for kv_consumer and support different num_tokens among all ranks (#4983) +[CI] Pull latest vllm-ascend src before tests (#4988) +add release note for 0.12.0 (#4995) +[Fix] Fixes issues in MTP with async scheduling and ACL graph (#4963) +[Perf]enable prefill flashcommon3 (#4065) +[CI] CI refactor (#4928) +add ut for model runner (#4991) +[Misc] Update pooling example (#5002) +[CI][Bugfix] Fix scheduleroutput has no attr get error in prompt logprobs (#4998) +Add Qwen3-Next tutorials (#4607) +[Refactor]3/N Refactor mla_v1.py & extract mla_cp (#4933) +[main][BugFix] Fixed an accuracy bug of Qwen3-next-MTP when batched inferring (#4932) +Bump actions/upload-artifact from 5 to 6 (#5014) +[bugfix] Fix dummy-run and multi-node issues in MoE routing and MTP (#4947) +[Doc ] Supplement kvpool user guide (#5013) +[Test]update accuracy test of models (#4911) +[Bugfix] Fix the bug in initializing the shared_weight communication domain in sfa-cp, and fix the mtp weight load in pp>1 situation (#4913) +[Bugfix] Add support for PP intermediate value types in graph mode (#4902) +[Bugfix] qwen3-vl-235b-w8a8 load weight ERROR when start service (#4292) +update release note for suffix decoding (#5009) +[Graph][Fusion] Add AddRMSNorm(with bias) and Quant Fusion Pattern (#5011) +[UT]add pcp dcp ut (#4949) +[Bugfix] fix the incorrect use of python's sum on tensors. (#4655) +[Misc] Upgrade vllm hash to 12_14 (#5000) +[CI] Delete deepseek3.2-exp nightly test (#5028) +[E2E] Collect test run time. (#5018) +[doc]Modify quantization tutorials (#5026) +[KVPool]Fix PP get bug (#5007) +[Attention] Temporarily add back pa for small batch sizes. (#4765) +[Cleanup] Remove unused attn_metadata parameter from Proposer classes (#4862) +[bugfix] [main] Fix KV cache query inconsistency across different TP ranks in the KV Pool (#5030) +[Bugfix] Fix precision issues in moe_mlp (vllm-ascend main) (#5025) +[Bugfix] Fix the attn_metadata is None (#5038) +[Misc] Upgrade vllm commit hash to 1215 (#5029) +[Fix]Revert temporary skip on mtp1/mtp2 correctness tests (aclgraph fix) (#5039) +[Core][Worker] Add UCMConnector for KV Cache Offloading (#4411) +[Bugfix] dynamic eplb does't use fused_alltoall (#4919) +Bump actions/checkout from 4 to 6 (#5015) +[Feat] Refactor rejection sampler (#4975) +[bugfix] Fix mooncake kvpool accuracy issue (#4976) +[Refactor] Remove the process patches of Qwen2.5-VL and Qwen2.5-Omni (#5035) +[Doc] Upgrade some outdated doc (#5062) +[ModelRunner] apply_grammer uses vllm function (#4974) +[Bugfix] fix fastapi version (#5047) +[BugFix]Fix FIA input err in DSv3.1 (#5059) +[Doc] Add user guide of speculative decoding (#5074) +Add release note for v0.11.0 (#4918) +[bugfix] matmul_allreduce_add_rmsnorm aclnn interface (#5082) +【Feature】refactor npu_modelrunner for profile_run (#4993) +Add a Mooncake installation tutorial for kv pool and update Mooncake installation tutorial (#5069) +[Bugfix] EPLB nightly deepseek (#5095) +[Nightly] Upgrade single node test to latest main (#5101) +[Nightly][BugFix] Install triton for nightly e2e op test. (#5096) +[Feat] Support async_scheduler and disable_padded_drafter_batch in eagle (#4893) +[bugfix] fix mtp accept rate (#5093) +Upgrade vllm commit hash to 1216 (#5053) +[Fusion] [Graph] Add qknorm rope fusion operator (#4711) +[UT]add the UT of pcp and dcp in the attention_cp file (#5054) +[Bugfix] Fix DeepSeek FIA error in async_scheduling with mtp (#5046) +[feat]pd disaggregated support cross-machine (#5008) +[CI] Fix UT (#5106) +[main] rename device type (#5099) +[main][doc] Instructions for using permissions added to docker (#5092) +[Pangu][MoE] Remove PanguProMoEV1 related code (#5088) +[model] Support PanguUltraMoE (#4615) +[UT] add pcp&dcp UT for mla_cp (#4953) +[BugFix] Fix mooncake bug in PCP scenario (#5055) +[Bugfix][MoE] Remove All2All in w4a8_dynamic (#4977) +Fix a data conversion bug introduced by commit 3b7eb51 in main#4655 (#5115) +[Refactor] 4/N Distinguish the branches based on the applicable scenarios of PA and FIA Ops. (#5081) +[Bugfix]delele profile_run in model_runner (#5122) +[Fix] Synchronize the host query_start_loc with device values to prevent shape mismatches (#5134) +fix profile run for vl model (#5136) +enable npugraph_ex (#5120) +[Doc] add qwen3 reranker (#5086) +[UT] Add model_runner pcp related UTs (#4951) +qwen3_next add triton ops : fused_qkvzba_split_reshape (#4788) +[test] add w4a8 accuracy case (#5110) +[feat] proxy support elastic scaling (#5063) +[Fix] Fix DeepSeek V3.2 "no attr" error (#5147) +[UT]Ut for function cumsum_group_list in moe_mlp (ref #5025) (#5036) +[UT] Add mooncake ut test (#5080) +fixed fused alltoall execute all reduce (#5109) +Qwen3-Next：Update the gpu-memory-utilization parameter to 0.7 (#5129) +[Bugfix] fix pipeline parallelism bug introduced by async-scheduling refactor work (#4973) +implement model runner v2 basic framework (#5051) +fix: use batch_matmul_transpose operator in MLA _v_up_proj for better performance (#5142) +Nominate new maintainers @zzzzwwjj @realliujiaxu @LCAIZJ (#5152) +feat: implement high-performance Triton kernels for rejection sampling (#4830) +[Feat] Support MLP_TP feature, exclude MOE layer (#4999) +[Graph][Fusion]Add new pattern for AddRmsnormQuant with SP. (#5077) +[Fix] Refines decode mode padding condition for uniform queries (#5164) +fix vl pd smoke error (#5103) +[BugFix]Fix incorrect get_current_vllm_config (#5121) +[Nightly] Avoid max_model_len being smaller than the decoder prompt to prevent single-node-accuray-tests from failing (#5174) +[Doc] Refact benchmark doc (#5173) +[Bugfix] Fix in_profile_run in mtp_proposer dummy_run (#5165) +[Doc][P/D] Fix MooncakeConnector's name (#5172) +[BugFix] Fix top_p,top_k issue with EAGLE and add top_p,top_k in EAGLE e2e (#5131) +[bugfix] Use FUSED_MC2 MoE comm path for the op `dispatch_ffn_combine` (#5156) +[2/N][Pangu][MoE] Remove Pangu Related Code (#5130) +[Bugfix] install trition for test_custom_op (#5112) +support basic long_seq feature st (#5140) +【Doc】Deepseekv3.1/R1 doc enhancement (#4827) +[BugFix]Fix precision issue for LoRA feature (#4141) +[refactor] refactor weight trans nz and transpose (#4878) +[Image] Refactor image build (#5175) +[Doc] Add a perf tune section (#5127) +Add Qwen3-VL-235B-A22B-Instruct tutorials (#5167) +[Refactor] remove some metadata variables in attention_v1. (#5160) +[CI] Improve CI (#5078) +[Feature] Add token mask for DispatchGmmCombineDecode operator (#5171) +[pref] qwen3_next add triton ops : fused_sigmoid_gating_delta_rule_update (#4818) +[Doc]Add the user_guide doc file regarding fine-grained TP. (#5084) +restore matmul_allreduce_add_rmsnrom aclnn interface (#5119) +[CI] Fix image merge bug (#5197) +[CI] Use offline mode for nightly test (#5187) +Drop 0.12.0 support (#5146) +[CI] unblock CI on suffix spec decoding (#4813) +[e2e] add pcp e2e (#5141) +[CI] fix lint (#5216) +[lint]clean code (#5218) +[Fix] Delete pooling redundant code (#4940) +[Performance] Add async exponential while model executing (#4501) +[BugFix]Fix wrong _cos, _sin instantiation (#5154) +[Feature]Use DispatchGmmCombineDecode operator to replace MC2(Optional) (#5040) +[Perf] vectorize PCP/DCP loops in attention_cp.py (#4944) +[Perf] vectorize PCP/DCP loops in mla_v1.py (#5003) +[Misc] Cleanup useless print and logger (#5220) +[Doc] Fix DeepSeek-V3.2 tutorial. (#5190) +[bugfix][ACLGraph][MTP] deletes `cudagraph_batch_sizes` in `MtpProposer` (#5183) +[task] Add fused gdn gating triton kernel (#4304) +[CustomOp] Register AscendMMEncoderAttention CustomOp and remove related patch (#4750) +[misc][FlashComm1][ACLGraph] Incompatibility between Flashcomm1 and FULL_DECODE_ONLY. (#5200) +Bump actions/upload-artifact from 4 to 6 (#5233) +Bump actions/checkout from 4 to 6 (#5234) +[Doc] Update readme (#5226) +[1/N][Eagle3] Aligns auxiliary hidden state usage for eagle3 models (#5162) +[Triton]support swiglu_quant triton in w4a8 (#5161) +[feature] support pcp + mtp in full graph (#4572) +[Feat]Xlite Qwen3-vl Support (#5228) +[bugfix] fix w8a8dynamic fused_moe trans nz (#5199) +[Bugfix] Implement multimodal_cpu_fields in model runner (#5196) +[TEST]Update mm param --mm-processor-cache-gb (#5242) +[Bugfix] Use hf_text_config instead of hf_config to support multimodal PD-Disaggregated (#5205) +[Refactor] move the metadata from attention_v1 to util(ready for extract common_cp) & realize Ascendmetadata inherit from the parent class. (#5203) +[refactor] Remove unnecessary attributes from set_ascend_forward_context (#5204) +[Doc] Update the weight download URL. (#5238) +[Main] [Patch] support balance scheduling patch (#5212) +[Doc] Add new contributors and relative scripts. (#5070) +[CustomOp] Register AscendApplyRotaryEmb CustomOp and remove related patch (#4667) +[Doc] fix docs set rope_theta value is 10e6 in qwen3-235b model (#5258) +[ModelRunner] Add hunyuan-vl basic support (#5151) +[KV-Sharing] Support KV-Sharing feature in CLA models (#4138) +[EPLB][CI] Add dynamic EPLB CI for qwen3-moe (#5179) +[CI] Add Triton Ascend in CI (#4921) +[CI]refactor: standardize test case naming convention (#5243) +[test]Corrected the Qwen3-Omni-30B-A3B-Instruct accuracy test configuration in nightly tests. (#5195) +[main][Refactor] Remove `with_prefill` parameter from `set_ascend_forward_context` (#5094) +[Doc] Added deploying on k8s with kthena (#4674) +[CI] Mock spawn for vlm tests (#5279) +[CI] refect e2e ci test (#5246) +[Refactor][MoE] Reuse vLLM's all_reduce logic (#5189) +[Bugfix] quick fix balance scheduling patch (#5281) +update to vllm 12-19 (#5223) +fix transformer version to 4.57.3 (#5250) +[Refactor]5/N Extract common code of mla_v1.py & extract mla_cp (#5097) +[CI] Add skipped testcases. (#5254) +[E2E] Optimize e2e test. (#5091) +[bugfix] remove the EP buffer allocation introduced by fused-op dispatch_ffn_c… (#5284) +[Doc] Add pa_shape_list description to qwen dense tutorial (#5225) +Update vllm pin to 12.24 (#5307) +[CI] Skip some failed ops tests (#5309) +[perf][bugfix] improve performance of rejection sampler and eliminate HD synchronize in TopKTopPSampler (#4154) +[quantization] Add w8a16 quantization support (#4541) +Cleanup uesless env (#5270) +Revert [KV-Sharing] Support KV-Sharing feature in CLA models (#4138) (#4138) +[Kernel] add l2norm triton kernel (#4595) +Add MagicMTP(block verify) and Triton optimization (#4443) +[CI] add xlite e2e test (#5305) +[E2E Refactor] Enable skipped e2e case (#5287) +[BugFix] Fix num_pcp_pads Assignment Issues (#5273) +[bugfix] fix Error 'ValueError: Duplicate layer name' (#5280) +Remove VLLM_ASCEND_ENABLE_DENSE_OPTIMIZE (#5272) +fix e2e rejection-sampler error (#5341) +[Bugfix] fix pcp 128K break (#5266) +[Bugfix] fix xlite decode-only e2e test (#5354) +[doc] update using command (#5373) +[Bugfix] Fix Qwen P/D Disaggregation accuracy issue (#5340) +[CI] Skip failed test cases to recover CI (#5368) +[FIX] Update _causal_conv1d_update_kernel for Efficient Conv State Handling on NPU (#5322) +[BugFix][Fusion] Patch compile backend to make fusion available (#5308) +move contiguous in fused_sigmoid_gating_delta_rule_update to model_runner_v1 (#5274) +[Nightly] Initial logging for nightly multi-node testing (#5362) +Update vllm pin to 12.25 (#5342) +cleanup ascend config (#5296) +[E2E] Optimize the E2E test time. (#5294) +Revert "Add MagicMTP(block verify) and Triton optimization (#4443)" (#4443) +[doc] add developer guide for PCP&DCP (#5372) +[Bugfix] Fix unsuitable moe_comm_type under ep=1 scenario (#5388) +[doc] Add context parallel user guide (#5358) +[Doc] update R1/V3.1 doc (#5383) +[Feature] Enhance all-reduce skipping logic for MoE models in NPUModelRunner (#5329) +[TEST]Add sending request with and without chat (#5286) +rollback causal_conv1d_fn to torch ops & update qwen3Next doc (#5391) +[bugfix] Fix MHA model runtime error in aclgraph mode (#5397) +[Feature] Remove the transpose step after attention and switch to transpose_batchmatmul (#5390) +Update vllm pin to 12.26 (#5378) +[CI] Add qwen-235b-a22b a2 multi-node test (#5393) +[Build] Add installation script of fused_infer_attention_score kernel with flash decoding (#5402) +[Test] Add acceptance test for eagle/eagle3 (#5366) +[TEST]Add vllm bench (#5306) +MLA prefill preformance optimization (#5275) +Revert "MLA prefill preformance optimization (#5275)" (#5275) +[Doc]add long sequence tutorials (#5364) +[bugfix][main]KV Pool for KV Transfer in PD Disaggregation Scenarios (#5398) +[BugFix] Fix npu-cpu offloading interface change bug. (#5290) +[Doc] modify pcp tutorials (#5411) +[bugfix] solve dp scenario Host-Device sync (#5298) +[Doc] add long_sequence feature user guide (#5343) +[Doc] delete environment variable HCCL_OP_EXPANSION_MODE in DeepSeekV3.1/R1 (#5419) +[feat] enable hierarchical mc2 ops on A2 by default (#5300) +[doc] Update Qwen3-235B doc for reproducing latest performance (#5323) +[Bugfix] fix greedy temperature detection (#5417) +Revert "[feat] enable hierarchical mc2 ops on A2 by default (#5300)" (#5300) +[Doc] Modify DeepSeek-R1/V3.1 documentation (#5426) +[DOC]Fix model weight download links (#5436) +[Doc] Update DeepSeek V3.1/R1 2P1D doc (#5387) +[Misc] fast fail for exiting if tools/install_flash_infer_attention_score_ops_a2.sh (#5422) +[Doc]modify pcp tutorial doc (#5440) +[bugfix] fix typo of _skip_all_reduce_across_dp_group (#5435) +Fix nightly (#5413) +[Bugfix] Correctly handle the output shape in multimodal attention (#5443) +[ReleaseNote] Add release note for v0.13.0rc1 (#5334) +update vllm pin to 12.27 (#5412) +[Refactor] cache cos/sin in mla & remove parameter model in builder. (#5277) +[Refactor]6/N Extract common code of class AscendMLAImpl (#5314) +[EPLB][refactor] Modification of the initialization logic for expert_map and log2phy（depend on pr5285） (#5311) +[Feature] Support to use fullgraph with eagle (#5118) +Optimize some rejectsampler functions to make npu op launch non-blocking (#4587) +[feature] fia support sliding windows (#5239) +[Feature] support eager mode in model runner v2 (#5210) +[Refactor][Triton] Move reject sample triton kernels into ops/triton (#5324) +[Refactor][EAGLE] 1/N delete __init__ in mtp_proposer (#5176) +[OP] add custom op aclnnMoeInitRoutingCustom (#5251) +[Kernel]update csrc cmakelist for open-source cann (#5458) +Update corresponding vllm commit ID to 12 29 (#5475) +[refactor] refactor model runner capture model (#5230) +moe_gating_top_k (#5271) +[CI]update triton ascend version (#5392) +[Doc] Fix issue link for 0.12.0 (#5500) +Revert "moe_gating_top_k" (#5512) +Docs: Remove deprecated --task parameter for embedding models (#5257) +[1/N] Refactor nightly test structure (#5479) +[3/N][Nightly] Move ops tests to nightly (#5538) +[Doc] Add new contributors. (#5537) +[2/N] Upgrade nightly doc (#5534) +[smoke][bugfix] moe_init_routing_v2 active_expert_range use int type (#5521) +[main][test] Refactor the mtp and eagle test case (#5326) +[Feature] Refactor PCP &DCP related code (#5214) +[Main2Main] Upgrade vllm commit to 1230 (#5495) +[Bugfix] Fix mm_merge (#5249) +[feature] mooncake support pcp/dcp in common conditions (#5224) +[Feature] Support kv nz feature for DeepSeek decode node in disagg-prefill scenario (#3072) +[Refactor] Formatting output types related to FuseMoE (#5481) +[P/D] Improve the performance of Layerwise Connector (#5303) +[Bugfix] fix the precision issues that may raise from the inter-layer reuse of the workspace in certain scenarios (#5522) +[Model] Add LongCat-Flash (#3833) +[Graph][Fusion] Add AddRMSNorm(with bias) (#5491) +[P/D] Bugfix zmq send/receive failed (#5503) +[Nightly] Trigger image build for nightly (#5547) +Bump actions/upload-artifact from 4 to 6 (#5466) +Bump actions/download-artifact from 4 to 7 (#5465) +[CI] Add multi-nodes longseq configs of DeepSeek-R1-W8A8 & Qwen3-235B-W8A8 (#5381) +Cleanup pass config override (#5283) +[Doc] Fix spelling mistake of environment variable name ASCEND_RT_VISIBLE_DEVICES in Doc (#5570) +[Feat][main] Supported to use full-graph with Qwen3-Next-MTP (#5477) +[Feat] enable hierarchical mc2 ops on A2 by default (#5545) +[CI] Move longseq Nightly CI (#5577) +[Perf][PCP][DCP] add multi-stream for GQA to enable computation-communication overlap (#5382) +[Recover] [Bugfix] support mtp kv transfer and pp partition by hand in kv transfer (#4892) (revert in #4981) (#4892) +[Doc] Fix typo in ASCEND_RT_VISIBLE_DEVICES (#5581) +[bugfix](pcp) expand max_num_tokens for pcp pad (#5478) +[BugFix]Disable dispatch_gmm_combine_decode operator when mtp drafter model uses non-w8a8 while main model uses w8a8, or drafter model is eagle series (#5293) +[KVPOOL]decode save kvcache (#5168) +[refactor](UT,PCP,DCP) refactor pcp&dcp patches in UTs (#5505) +[Doc]modify the quantization user guide and add a quantization adaptation developer guide (#5554) +[bugfix]update bishengir source envs (#5582) +[Bugfix] Fix chunk prefill bug for long_sequence feature (#5444) +[Bugfix] Fix weight transpose in RL scenarios (#5567) +[Doc] update supported models (#5379) +[CI] skip xlite-decode-only e2e test (#5407) +[Doc] eval-type not support service but server (#2920) +MLA prefill preformance optimization (#5456) +[Refactor][EAGLE] 2/N: load model and generate token (#5437) +[Bugfix] fix pcp + eplb error (#5561) +[Doc] add new doc for mooncake: PD-Colocated cross-node multi-instance validation of Mooncake's KV Cache reuse and performance. (#5415) +[BugFix][kernel] fix matmul_allreduce_add_rmsnorm_kernel (#5335) +feat: implement high-performance Triton kernels for rejection sampling: optimization for rejection_random_sample_kernel (#5259) +[Feat][Spec] Optimize token index calculation in spec decode with Triton kernel (#5356) +[Refactor]7/N Extract common code to common_cp (#5490) +[BugFix][Fusion] Fix graph fusion failure problem (#5253) +Add the requirement of arctic-inference which speculative decoding with suffix_decode (#5045) +[Doc] Add NNAL installation guide and requirements (#5235) +Docs: Add A3 Docker image guidance for Atlas A3 machines (#5256) +[CI] Download models from ms (#5405) +[UT]add triton ops ut : test_fused_qkvzba_split_reshape_cat (#5474) +[bugfix] fix test_camem failed with triton-ascend (#5492) +[Bugfix] record cos and sin cache in AscendRotaryEmbedding (#5516) +[P/D]Remove mooncake kvpool unused parameter `local_hostname` (#5574) +[CI] update triton-ascend version (#5584) +[docs] Correct image about prefill phase of PCP (#5598) +[perf] Fix MLAPO weight disposal for KV-consumer MLA in PD-mix deploy... (#5192) +[TRITON][TEST]Add nightly test for triton split_qkv_rmsnorm_rope (#5267) +Revert "[Feat] enable hierarchical mc2 ops on A2 by default (#5545)" (#5545) +[BugFix] Fix Smoke Testing Bug for DSR1 longseq (#5613) +[CI] mv ops to correct path (#5615) +[Main2Main] Upgrade vllm commit to 0105 (#5595) +[UT][PCP&DCP] UT for block_table.py (#5032) +[CI]update bisheng version (#5621) +[Main2Main] Upgrade vllm commit to 0106 (#5617) +[CI] Specify the version of xlite (#5612) +[MM][Bugfix] Update `hf_config` to `hf_text_config` (#5319) +[Refactor][EAGLE] 3/N delete redundant methods in mtp_proposer (#5420) +Bugfix: Align expert map shapes with redundant experts in EPLB adjustment (#5285) +[Bugfix] Remove swa parameter of fia (#5602) +[Nightly][Test] Add Qwen3-Next-80B-A3B-Instruct-W8A8 nightly test (#5616) +[Misc] Remove useless weight loader patch (#5619) +[P/D] Performance enhancement of Layerwise connector in TP asymmetric scenarios (#5540) +Revert "[BugFix][Fusion] Fix graph fusion failure problem (#5253)" (#5253) +[Bugfix] fix dcp_only bug and add e2e accuracy test for dcp only and pcp only (#5565) +[Graph][Fusion] Add AddRMSNormSPPattern and AddRMSNormSPPatternWithBias (#5569) +[Feature] implement basic framework for batch invariant (#5517) +[Refactor] Cleanup platform (#5566) +[bugfix (pcp)] fix chunked prefill accurancy issue (#5647) +[CI] Add DeepSeek-V3.2-W8A8 nightly ci test (#5371) +[Feature]EPLB:Adapt DispatchGmmCombineDecode operator to eplb tensor list and expert token numbers (#5552) +[Bugfix] Revert pr4214 multi-stream collect expert hotpot (#5529) +[Bugfix]Add register_kv_cache in ucm_connector (#5657) +[misc]Add Kimi-K2 series to CI model list (#5656) +[CI] cleanup single/multi-card test (#5623) +[CI] Bump lm-eval version to v0.4.9.2 (#5655) +[CI] Add workflow to cancel running workflows on PR close (#5646) +[Bugfix] fix resource are insufficient when pcp and piecewise (#5377) +[Bugfix] Fix the graph capture failure issue in the eagle3+full scenario. (#5553) +[CI] move image and wheel job to schedule way (#5685) +[Refactor] Fix AttentionMaskBuilder singleton and remove redundant pcp_prefill_mask (#4870) +[Refactor] Import global var form vllm instead of overwirte it (#5469) +[Tests] Add qwen3-8b nightly test (#5597) +[BugFix][Fusion] Fix graph fusion failure problem (#5676) +[1/N][CI] Refactor accuracy test (#5400) +[Kernel] Add moe_gating_top_k operator support for Ascend NPU (#5579) +[BugFix][P/D] Fix pre-create link parameter error (#5694) +[refactor] Refactor the interface for shard weight and remove the flashcomm2 o_shared interface. (#5181) +[bugfix] adapt to new implemented get_kv_cache_spec in cpuoffload connector (#4311) +[Feature] add the magicmtp speculative decoding acceleration algorithm (#5542) +Optimize the print info format when deprecated code is used in vllm-ascend (#5696) +[CI] fix image build tag (#5703) +[EPLB][CI] EPLB add aclgraph and redundant expert ci (#5625) +[CI] Drop outdated cases (#5709) +[CI] Fix image build workflow_dispatch error (#5717) +[Feat][Bugfix][main] Adapted SP to eagle3 (#5562) +[bugfix] Support dsv3.2 enable both mtp and full_decode_only (#5679) +[Doc] Add Qwen3-Omni-30B-A3B-Thinking Tutorials (#3991) +[Fix] Fixes speculative decode indexing and unpad condition for attention metadata (#5626) +[CI] Add triton ascend in nightly CI (#5716) +[feature]dcp&pcp support mlapo (#5672) +[CI] Remove workflow_dispatch way for image build (#5742) +[Nightly] Move ops to the correct path (#5642) +[OP] Enable custom op aclnnMoeInitRoutingCustom (#5332) +[CI] Add qwen3 next ci (#5395) +[Doc] add PaddleOCR-VL tutorials guide (#5556) +[BugFix][DS 3.2] Fix ds indexer accuracy problem caused by rope. (#4641) +[Doc][fix] Fix the title of the document for the layer_sharding feature (#5759) +[CI] lint and ut use self_hosted runner (#5652) +[BugFix] NetLoader: No backend type associated with device type npu (#5700) +[CI] Accuracy issue of qwen3-next-w8a8 nightly test fix. (#5746) +[BugFix] Xlite: Bypass the padding of the graph mode in non-MTP cases to obtain the correct decode num. (#5711) +[CustomOp] support TensorList for dispatchFFNCombine (#5665) +[CI] Avoid lint and ut for PR push (#5762) +[BufFix]Fix the error when using Ascend custom operators with rank=128 (#5394) +[Refactor] Replace the implementations of o_proj, q_b_proj, and kv_b_proj with custom_op for sharded CP (#5698) +[Bugfix] Fix matmul allreduce precision issue by using original weight (#4939) +[Feature] GLM4.6 support mtp with fullgraph (#5460) +[CI]Add Disaggregated PD Nightly Test for Qwen3-235B and Qwen3-VL-235B (#5502) +support mxfp8 quantization (qwen dense) (#5723) +[Doc] Add GLM4.5 GLM4.6 doc (#5740) +[bugfix] Fixing KV Pool Memory Retention and Performance Degradation Issues (#5751) +[P/D][bugfix]Fix the PCP port mapping error issue (#5706) +[Feat] flashcomm2+oshard Generalized (#4723) +adapt to minimax_m2 (#5624) +[P/D] layerwise connector supports DeepSeek-V3.2 sparse attention && Distribute transfer tasks to redundant kv_head cards (#5722) +[main][bugfix] Fix fullgraph padding bug in mtp eagle refactor (#5692) +[Perf] Supports compute-communication overlap in the forward of sfa_v1 in the Sharded-CP feature. (#5701) +[Feature] Support for cross-attention and whisper model (#5592) +[0.13.0][doc] correct doc url (#5791) +[0.13.0][CI] disable main CI (#5792) +[0.13.0][Cherry Pick] cherry pick from 5638 Update pd readme (#5811) +[0.13.0][cherry-pick][bugfix](cp) align max_context_chunk to cp_virtual_block_size (#5782) +[0.13.0][Bugfix] bugfix for the order of dummy run pad and sync (#5778) +[0.13.0][Patch] AscendLoRAModelManager.__init__ (#5800) +[cherry-pick][BugFix] Support setting tp=1 for the Eagle draft model to take effect (#5804) +[v0.13.0][Bugfix] Support ALL D-Nodes in fullgraph when running MTP in PD (#5786) +[0.13.0][cherry-pick]enable ep32 for dispatch_ffn_combine (#5788) +[P/D] [CherryPick] 5846 fix layerwise connector for decoder tp size > num kv he… (#5857) +[0.13.0][cherry-pick][bugfix]Synchronize memcache adaptation on A2 (#5842) +[0.13.0][cherry-pick][Bugfix] Fixed an accuracy problem of sp with eagle3 (#5814) +[0.13.0][cherry-pick][P/D] bugfix for p node force free requset (#5431) (#5431) +[0.13.0][cherry-pick][bugfix](cp) replace None with zeros/inf tensor to avoid TypeError (#5844) +[v0.13.0][bugfix] patch set cudagraph size (#5860) +[v0.13.0][Bugfix] Fix acc bug when enbale dispatch_gmm_combine_decode and eplb[RFC: issue 5476] (#5836) +[0.13.0][Bugfix] Fix memory inconsistency in cross-process shared memory (#5779) +[v0.13.0][cherry-pick][BugFix] Fix DispatchGmmCombineDecode acc bug when big batch (#5873) +[v0.13.0][bugfix]Fix graph sync (#5809) +[0.13.0][cherry-pick][bugfix]support dsv3.2 enable both mtp and full_decode_only (#5849) (#5849) +Revert "[BugFix] Support setting tp=1 for the Eagle draft model to take effect(#5519) (#5519) +enbale qwen3-vl model fc1 feature (#5848) +Revert "[v0.13.0][bugfix]Fix graph sync (#5809)" (#5809) +[Performance]use triton mrope for Qwen3-VL (#5827) +[P/D]The issue of solving the force-free secondary release request, which causes the node to crash. (#5970) +[0.13.0][bugfix] fix mooncake kv cache transfer when one P has multi nodes (#5961) +[0.13.0][Feature] Support fine-grained shared expert overlap (#5962) +[0.13.0][Bugfix] fix bug of pcp+mtp+async scheduler (#5995) +[0.13.0][Bugfix] Add `synced_cudagraph_mode` to limit mixed graph modes in dp ranks (#6011) +【0.13.0】【bugfix】Resolved memory deallocation failure in the pooling layer under re-computation workloads. (#6056) +[0.13.0][cherry-pick][bugfix] fix bug of triton mrope (#6009) +[0.13.0][Bugfix] fix pcp aclgraph qwen FIA bug (#6038) +[Bugfix]Fixed precision issues caused by pooled request pooling (#6057) +[0.13.0][Bugfix] Fixed an problem related to embeddings sharing (#5972) +[v0.13.0][Bugfix] Fix XliteModelRunner init failed when aclgraph is enabled (#5887) +[0.13.0][Bugfix] Fix setting of `speculative_config.enforce_eager` for dsv32 (#5958) +Revert "[0.13.0][cherry-pick][bugfix] fix bug of triton mrope" (#6075) +[0.13.0][cherry-pick][bugfix] fix the complex and potentially problematic generate_kv_idx. (#5955) +[0.13.0][CI]fix for CI lint (#6093) +[EPLB][Bugfix] Dispatch Allgather use log2phy if enable eplb (#5933) (#5933) +[EPLB][Bugfix][v0.13.0] Incorporate the warm up of the EPLB into the profile run. (#6099) +[0.13.0][Doc] Supplement PD separation parameters of DeepSeek V3.1 (#6054) +[v0.13.0][CI] Upgrade to CANN 8.5.0 (#6101) +[0.13.0][Bugfix] Fix Triton operator usage for multimodal models based on `the mrope_interleaved` parameter (#6074) +[v0.13.0][BugFix][Cherry Pick] Fix input parameter bug of dispatch_gmm_combine_decode (#5931) +[Feature][Cherry Pick]Enable DispatchGmmCombineDecode when eagle is moe with w8a8, or not moe (#6081) +[v0.13.0][Bugfix] Fix the input constraints checks for the mlapo and bmm_transpose operators (#5764) (#5764) +[EPLB] Config Rename wrapper (#6111) +[v0.13.0][cherry-pick][BugFix]converting pa get_workspace back to capturing (#6108) +[cherry-pick][BugFix] Support setting tp=1 for the Eagle draft model to take effect (#6095) +[kv_cache] support multi_block_pool (#6106) +[CI] Skip some persistently stuck ut cases (#6133) +[0.13.0][CI]Add triton ascend version to 3.2.0 (#6105) +[0.13.0][cherry-pick][CP&SP] Integrate FIA operator in mla_cp._forward_decode (#6046) +[0.13.0][cherry-pick][BugFix] fix 3vl dense model load quant weight (#6103) +[0.13.0][cherry-pick] Reset incompatible config (#6118) +[0.13.0][Bugfix] Remove `use_aclgraph` in mtp_proposer and use `use_cuda_graph` (#6102) +[0.13.0][BugFix][cherry-pick]hccl bufferSize check for dispatch_ffn_combine (#6131) +Mix placement (#6086) +[v0.13.0][Feature] Support DSA-CP for Hybrid scenario (#5702) (#5702) +[0.13.0][P/D][PCP]bugfix pcp force free twice caused logger error (#6132) +[EPLB][Bugfix] Do not refresh parameters when eplb_config is not passed (#6160) +[0.13.0][Doc] update supported features (#6150) +[0.13.0][cherry-pick][CP&SP] Remove CP Redundant Variables after FIA operator enables for CANN 8.5 (#6039) +[v0.13.0][cherry-pick][Bugfix] Fix seq_lens reset issue causing performance degradation (#6166) +[0.13.0][Bugix] fix kv pcp+pooling+pd separation bug (#6152) +[Doc] Document translation (#6066) +[v0.13.0][bugfix] fix capture shape in sp_eagle_fullgraph (#6159) +[0.13.0][Feat] Merge the multi eagle graphs to one graph (#6178) +[0.13.0][BugFix] Avoided a bug of `torch_npu.npu_mm_reduce_scatter_base` when sp size >= 16 (#6167) +[0.13.0][BugFix]bug fix for dispatch_ffn_combine (#6157) +[0.13.0]Add has_connector_metadata (#6154) +[v0.13.0]skip eagle dp allreduce (#6162) +[0.13.0][KVCache] Support different page sizes (#6171) +[Bugfix] Fix the issue of the acceptance rate decline for Qwen3-30B-A3B-EAGLE3 (#6139) +[0.13.0][cherry-pick] addrmsnorm op support bias (#6140) +[Doc] Refresh doc for 0.13.0 (#6184) +[Doc] Add release note for 0.13.0rc2 (#6208) +[0.13.0] [BugFix] buildwheel dependency install (#6211) +[ci] Fix docker image build (#6215) +[0.13.0] [BugFix] Fix build wheel (#6220) +[Bugfix][v0.13.0] Fix a bug when cherry-pick from main (#6209) +[0.13.0][Bugfix] Avoided a bug of drafter when `dp` and `sp` are enabled (#6224) +[Inductor][v0.13.0]Adapt AddRmsNormQuant pass to new addrmsnormBias operator (#6210) +[v0.13.0][cherry-pick][BugFix] Fix moe_load accumulation error in ACL graph mode (#6258) +[CI] migrate single card runner to hk (#6260) +[Refactor] use the count of kv_cache_group to create multi_block_table (#6116) +[Misc][v0.13.0]Removes unnecessary graph size re-initialization (#6281) +[0.13.0][KVCache] Prioritize using a hybrid manager to manage different types of kvcache (#6289) +[CI] Update pta to 2.8.0.post2 (#6287) +[0.13.0][cherry-pick][BugFix][CI]Fix DeepSeek-R1-W8A8-longseq nightly CI (#6337) +[CI] Add per pr image build for nightly test (#6353) +[CI] Cherry pick nightly test from main (#6365) +[cherry-pick][BugFix] Disable enable_shared_expert_dp by default if tensor_parallel_size=1 (#6363) +[0.13.0][Bugfix] Fix hash conflict due to reset incompatible configuations (#6330) +[0.13.0][Bugfix] Fix FIA operator validation error in Eagle scenario with CANN 8.5 (#6284) +[CI]Limit transformers version (#6373) +[Doc] Reranker guide remove deprecate task option (#6380) +[0.13.0][cherry-pick][bugfix](CP,MLA) fix wrong slot_mapping of decode for mixed p/d batch (#6346) +[0.13.0][Profiler] Fix profiler bug (#6383) +[0.13.0][cherry-pick][bugfix](pcp,gqa) set kv_inverse_idx_for_chunk and cp_kv_recover_idx_for_chunk to None when dcp only (#6318) +[0.13.0][Bugfix]Fix of Pooling Code (#6146) +[0.13.0][cherry-pick][Bugfix][CI] Specify tensorflow version in accuracy test to avoid segmentation fault (#6292) (#6292) +[0.13.0][bugfix]Raise exception for Omni models with FLASHCOMM enabled (#6392) +[CI][Nightly] Correct the nightly image build ref (#6396) +[P/D][0.13.0]Add ssl cert for metaserver proxy (#6400) +[0.13.0]fix patch cudagraph size (#6397) +fix: resolve sync bug in DispathFFNCombine when expert num per card is 32 (#6422) +work around: reset the None reference count to prevent it from droppi… (#6441) +[v0.13.0][Eagle3]Extend PR #5786 to eagle3 (#6443) +[0.13.0][cherry-pick]pick from 6310 to fix rope op (#6444) +[CI] change ds32 cudagraph_sizes (#6399) +[bugfix][0.13.0]fix bug in dispatch_ffn_combine kernel (#6464) +[v0.13.0][Lora][BugFix] Fix crash on base model requests with LoRA enabled (#6457) +Bugfix: Pre-compile EPLB algorithm successfully in subprocess under graph mode (#6472) +[0.13.0][Bugfix] fix npu memory is not released in cp (#6479) +[0.13.0][Bugfix] Fix problematic dummy_run & improper input_batch_size in eagle (#6518) diff --git a/skills/upstream/vllm-ascend-releasing-note/output/v0.13.0/1-commit-analysis-draft.csv b/skills/upstream/vllm-ascend-releasing-note/output/v0.13.0/1-commit-analysis-draft.csv new file mode 100644 index 0000000..e5b0917 --- /dev/null +++ b/skills/upstream/vllm-ascend-releasing-note/output/v0.13.0/1-commit-analysis-draft.csv @@ -0,0 +1,995 @@ +title,pr_number,user_facing_impact,category,decision,reason +[CI]Fix test nightly workflow. (#3603),3603,[CI]Fix test nightly workflow.,Ignore,Ignore,Internal/CI/test change +"Reapply ""[MoE] [Refactor] Remove manual memory cleanup (#3365)"" (#3483) (#3365)",3365,"Reapply ""[MoE] [Refactor] Remove manual memory cleanup",Ignore,Ignore,Internal/CI/test change +fix : support chunked_prefill with deepseek_mtp (#2711),2711,fix : support chunked_prefill with deepseek_mtp,Highlights,Include,Major feature +[Misc] clean up useless function (#3348),3348,[Misc] clean up useless function,Ignore,Ignore,Internal/CI/test change +[Feat] Dynamic Batch Feature (#3490),3490,[Feat] Dynamic Batch Feature,Features,Include,New feature +[Feat] add native kvcache offload (#3433),3433,[Feat] add native kvcache offload,Features,Include,New feature +[CI] Multi-Node CI scalable (#3611),3611,[CI] Multi-Node CI scalable,Ignore,Ignore,Internal/CI/test change +clean up uesless ut test (#3622),3622,clean up uesless ut test,Others,Include,Miscellaneous +unify logic between aclgraph and torchair (#3560),3560,unify logic between aclgraph and torchair,Highlights,Include,Major feature +[Fix] Fixes attribute error in MLA implementation (#3618),3618,[Fix] Fixes attribute error in MLA implementation,Features,Include,New feature +[BugFix][main] Fix quantization related mtp bug with patch (#3620),3620,[BugFix][main] Fix quantization related mtp bug with patch,Highlights,Include,Major feature +[Feat] Prefetching Attention QKV Linear Weight With `AddRmsNormQuant` Custom Op (#3517),3517,[Feat] Prefetching Attention QKV Linear Weight With `AddRmsNormQuant` Custom Op,Hardware and Operator Support,Include,Operator/hardware support +[Doc] Upgrade docker run command (#3645),3645,[Doc] Upgrade docker run command,Documentation,Include,Documentation update +[main][refactor] refactor SequenceRowParallelOp forward (#3616),3616,[main][refactor] refactor SequenceRowParallelOp forward,Ignore,Ignore,Internal/CI/test change +[Doc] Update the modelslim website from gitee to gitcode. (#3615),3615,[Doc] Update the modelslim website from gitee to gitcode.,Documentation,Include,Documentation update +[BugFix] fix deepseek torchair precision (#3624),3624,[BugFix] fix deepseek torchair precision,Others,Include,Bug fix +perf : optimize memory for deepseek mtp (#2713),2713,perf : optimize memory for deepseek mtp,Highlights,Include,Major feature +[Misc] Add a model loader that utilizes HCCL for weight loading (#2888),2888,[Misc] Add a model loader that utilizes HCCL for weight loading,Others,Include,Miscellaneous +[TEST]Add initial multi modal cases for nightly test and deepseek-r1 tests (#3631),3631,[TEST]Add initial multi modal cases for nightly test and deepseek-r1 tests,Highlights,Include,Model support +[Structured Output] Replace `apply_grammar_bitmask()` method with that in vllm to avoid maintenance (#2524),2524,[Structured Output] Replace `apply_grammar_bitmask()` method with that in vllm to avoid maintenance,Others,Include,Miscellaneous +[Bugfix] fix delay free prefill req & D node support prefix cache (#3607),3607,[Bugfix] fix delay free prefill req & D node support prefix cache,Features,Include,New feature +[Doc] Update the Pangu Pro MoE tutorials. (#3651),3651,[Doc] Update the Pangu Pro MoE tutorials.,Documentation,Include,Model tutorial +[Test] add a new Qwen3-32b-int8 test case with feature_stack3 (#3676),3676,[Test] add a new Qwen3-32b-int8 test case with feature_stack3,Others,Include,Miscellaneous +[main][bugfix] Add 'layer_type' param to get_pergroup_param() for compatibility (#3682),3682,[main][bugfix] Add 'layer_type' param to get_pergroup_param() for compatibility,Others,Include,Bug fix +[BugFix]fix deepseek torchair recompile (#3678),3678,[BugFix]fix deepseek torchair recompile,Others,Include,Bug fix +support cp&dcp (#3260),3260,support cp&dcp,Highlights,Include,Major feature +[MoE][Multistream] Avoid performing communication in extra stream. (#3582),3582,[MoE][Multistream] Avoid performing communication in extra stream.,Others,Include,Miscellaneous +[Benchmark] Upgrade benchmark args for new vllm version (#3218),3218,[Benchmark] Upgrade benchmark args for new vllm version,Others,Include,Miscellaneous +[UT] Fix test_sample_recovered_tokens_pytorch_autoregressive (#3434),3434,[UT] Fix test_sample_recovered_tokens_pytorch_autoregressive,Ignore,Ignore,Internal/CI/test change +remove useless code (#3685),3685,remove useless code,Others,Include,Miscellaneous +[TEST]Add initial prefix cache case for nightly test (#3709),3709,[TEST]Add initial prefix cache case for nightly test,Ignore,Ignore,Internal/CI/test change +[1/N][Refactor] Refactor code to adapt with vllm main (#3612),3612,[1/N][Refactor] Refactor code to adapt with vllm main,Ignore,Ignore,Internal/CI/test change +[BugFix] Check all expert maps when using muilty instance. (#3576),3576,[BugFix] Check all expert maps when using muilty instance.,Others,Include,Bug fix +[TEST]Add initial multi modal cases of Qwen2.5-VL-32B-Instruct for nightly test (#3707),3707,[TEST]Add initial multi modal cases of Qwen2.5-VL-32B-Instruct for nightly test,Others,Include,Miscellaneous +"[Bugfix] The server fails to locate the request, leading to the server hanging. (#3703)",3703,"[Bugfix] The server fails to locate the request, leading to the server hanging.",Others,Include,Bug fix +[Main][Perf] Add fused matmul/reduce-scatter kernel for performance optimization. (#3693),3693,[Main][Perf] Add fused matmul/reduce-scatter kernel for performance optimization.,Performance,Include,Operator fusion +[Refactor] Refactor Ascend attention implementation forward (#3714),3714,[Refactor] Refactor Ascend attention implementation forward,Ignore,Ignore,Internal/CI/test change +[Feat] Add mrope fusion op (#3708),3708,[Feat] Add mrope fusion op,Performance,Include,Operator fusion +[BugFix][P/D] Modify the recalculation logic to prevent waiting requests from filling up the D node KVCache (#3641),3641,[BugFix][P/D] Modify the recalculation logic to prevent waiting requests from filling up the D node KVCache,Others,Include,Bug fix +[CI][Doc] Optimize multi-node CI (#3565),3565,[CI][Doc] Optimize multi-node CI,Ignore,Ignore,Internal/CI/test change +[UT][fix] Add missing get_ascend_config mock to NPUWorker initialization tests (#3729),3729,[UT][fix] Add missing get_ascend_config mock to NPUWorker initialization tests,Ignore,Ignore,Internal/CI/test change +[Misc] Limit ray version (#3660),3660,[Misc] Limit ray version,Others,Include,Miscellaneous +[CI] Skip ops test for e2e (#3665),3665,[CI] Skip ops test for e2e,Ignore,Ignore,Internal/CI/test change +Update version doc (#3599),3599,Update version doc,Others,Include,Miscellaneous +[BugFix][Core] Fix a bug running multi-modal with ascend_scheduler (#3675),3675,[BugFix][Core] Fix a bug running multi-modal with ascend_scheduler,Others,Include,Bug fix +[Bugfix] Fix zero attention output in qwen3-next (#3572),3572,[Bugfix] Fix zero attention output in qwen3-next,Highlights,Include,Model support +[Refactor] optimize _prepare_inputs method in eagle_proposer (#3296),3296,[Refactor] optimize _prepare_inputs method in eagle_proposer,Ignore,Ignore,Internal/CI/test change +[BugFix] Comment out newly added vlm e2e. (#3736),3736,[BugFix] Comment out newly added vlm e2e.,Others,Include,Bug fix +[Test] Add e2e test and accuracy test for Qwen3-Next-80B-A3B-Instruct (#3450),3450,[Test] Add e2e test and accuracy test for Qwen3-Next-80B-A3B-Instruct,Highlights,Include,Model support +[Doc] Update supported models (#3481),3481,[Doc] Update supported models,Documentation,Include,Documentation update +[Refactor] [MoE] Rename moe-related classes & files (#3646),3646,[Refactor] [MoE] Rename moe-related classes & files,Ignore,Ignore,Internal/CI/test change +[Test] add test for prefix cache feature of deepseek (#3733),3733,[Test] add test for prefix cache feature of deepseek,Others,Include,Miscellaneous +[bugfixfix] correct _register function place for mooncacke (#3747),3747,[bugfixfix] correct _register function place for mooncacke,Others,Include,Miscellaneous +Upgrade to new vllm commit (#3719),3719,Upgrade to new vllm commit,Others,Include,Miscellaneous +[main] remove dbo code (#3712),3712,[main] remove dbo code,Others,Include,Miscellaneous +add qwq testcase (#3757),3757,add qwq testcase,Others,Include,Miscellaneous +[BugFix] Fix Qwen3-next break (#3428),3428,[BugFix] Fix Qwen3-next break,Highlights,Include,Model support +[Installation] limit opencv-python-headless version to resolve numpy version conflict (#3713),3713,[Installation] limit opencv-python-headless version to resolve numpy version conflict,Others,Include,Miscellaneous +[feat]dcp pcp support aclgraph (#3731),3731,[feat]dcp pcp support aclgraph,Highlights,Include,Major feature +[CI] Add custom op to nightly (#3765),3765,[CI] Add custom op to nightly,Hardware and Operator Support,Include,Operator/hardware support +[CI] Enable 2 jobs for nightly test (#3781),3781,[CI] Enable 2 jobs for nightly test,Features,Include,New feature +Bump actions/download-artifact from 5 to 6 (#3787),3787,Bump actions/download-artifact from 5 to 6,Ignore,Ignore,Internal/CI/test change +Bump actions/upload-artifact from 4 to 5 (#3786),3786,Bump actions/upload-artifact from 4 to 5,Ignore,Ignore,Internal/CI/test change +[Doc][Example][Bugfix] Elements in local_device_ids should be casted … (#3782),3782,[Doc][Example][Bugfix] Elements in local_device_ids should be casted …,Documentation,Include,Documentation update +[bugfix][main]fix proxy decode bug (#3750),3750,[bugfix][main]fix proxy decode bug,Others,Include,Bug fix +[MM][Doc] Update online serving tutorials for `Qwen2-Audio` (#3606),3606,[MM][Doc] Update online serving tutorials for `Qwen2-Audio`,Documentation,Include,Model tutorial +support prefill cache mode use fia op (#3696),3696,support prefill cache mode use fia op,Hardware and Operator Support,Include,Operator/hardware support +[Doc] Update FAQ (#3792),3792,[Doc] Update FAQ,Documentation,Include,Documentation update +【Bugfix】bugfix for weight load of kimi-k2 (#3798),3798,【Bugfix】bugfix for weight load of kimi-k2,Highlights,Include,Model support +[TEST]Add 2P1D multi node cases for nightly test (#3764),3764,[TEST]Add 2P1D multi node cases for nightly test,Others,Include,Miscellaneous +[CI] Add multi-node test case for a2 (#3805),3805,[CI] Add multi-node test case for a2,Ignore,Ignore,Internal/CI/test change +Upgrade to 0.11.1 newest vllm commit (#3762),3762,Upgrade to 0.11.1 newest vllm commit,Others,Include,Miscellaneous +[CI] Fix nightly CI (#3821),3821,[CI] Fix nightly CI,Others,Include,Miscellaneous +[Main][Bugfix]Avoid using the fusion operator in the MOE model (#3834),3834,[Main][Bugfix]Avoid using the fusion operator in the MOE model,Performance,Include,Operator fusion +[TEST]Add aisbench log and A2 cases (#3841),3841,[TEST]Add aisbench log and A2 cases,Hardware and Operator Support,Include,Operator/hardware support +[long_seq_optim] BSND to TND and FA_UPDATE replacement (#3778),3778,[long_seq_optim] BSND to TND and FA_UPDATE replacement,Performance,Include,Performance optimization +[P/D] force with_prefill true after allreduce in kv producer (#3768),3768,[P/D] force with_prefill true after allreduce in kv producer,Others,Include,Miscellaneous +fix qwen3next full graph break. (#3812),3812,fix qwen3next full graph break.,Highlights,Include,Major feature +[Doc] Update doc (#3836),3836,[Doc] Update doc,Documentation,Include,Documentation update +[HybridKV][Bugfix] Fix Hybrid kvcache sharing bug in same attention type (#3760),3760,[HybridKV][Bugfix] Fix Hybrid kvcache sharing bug in same attention type,Others,Include,Bug fix +[Bugfix] [MoE] fix error in deepseek when using allgather (#3824),3824,[Bugfix] [MoE] fix error in deepseek when using allgather,Others,Include,Bug fix +[Perf] Delete redundant operations in model_runner and forward_context (#3677),3677,[Perf] Delete redundant operations in model_runner and forward_context,Performance,Include,Performance optimization +[CI]pin vllm commit id (#3861),3861,[CI]pin vllm commit id,Ignore,Ignore,Internal/CI/test change +[CI] Optimize nightly CI (#3858),3858,[CI] Optimize nightly CI,Performance,Include,Performance optimization +[BugFix] deepseek torchair adapt for torch_npu version (#3862),3862,[BugFix] deepseek torchair adapt for torch_npu version,Dependencies,Include,Dependency update +[CI]Fix eplb nightly tests. (#3863),3863,[CI]Fix eplb nightly tests.,Ignore,Ignore,Internal/CI/test change +fix mooncake layerwise connector (#3849),3849,fix mooncake layerwise connector,Highlights,Include,Major feature +bugfix for mtp fullgraph (#3845),3845,bugfix for mtp fullgraph,Highlights,Include,Major feature +[BugFix] Fix mlapo accuracy problem related with weight processing. (#3850),3850,[BugFix] Fix mlapo accuracy problem related with weight processing.,Hardware and Operator Support,Include,Operator/hardware support +[CI]Fix oom of deepseek-eplb nigtly test. (#3884),3884,[CI]Fix oom of deepseek-eplb nigtly test.,Ignore,Ignore,Internal/CI/test change +Add FAQ for docker pull error on Kylin OS (#3870),3870,Add FAQ for docker pull error on Kylin OS,Others,Include,Miscellaneous +[UT] fix skip ut test for test_utils (#3803),3803,[UT] fix skip ut test for test_utils,Ignore,Ignore,Internal/CI/test change +[Doc] Remove modeling doc (#3789),3789,[Doc] Remove modeling doc,Documentation,Include,Documentation update +[Build] Force torch version (#3791),3791,[Build] Force torch version,Others,Include,Miscellaneous +[FEAT] Refactor spec decode to support efficient padded speculation (#3528),3528,[FEAT] Refactor spec decode to support efficient padded speculation,Features,Include,New feature +[Model][3/N] Refactor sfa into mla and remove deepseek_v3_2.py (#3769),3769,[Model][3/N] Refactor sfa into mla and remove deepseek_v3_2.py,Hardware and Operator Support,Include,Operator/hardware support +[feature] Prompt Embeddings Support for v1 Engine (#3026),3026,[feature] Prompt Embeddings Support for v1 Engine,Features,Include,New feature +[TEST]Add MALPO for aclgraph in nightly test (#3894),3894,[TEST]Add MALPO for aclgraph in nightly test,Highlights,Include,Major feature +[BugFix]Fix group list type of mc2. (#3864),3864,[BugFix]Fix group list type of mc2.,Others,Include,Bug fix +[bugfix] layerwise D first plan (#3866),3866,[bugfix] layerwise D first plan,Others,Include,Bug fix +[CI] Optimize nightly CI (#3898),3898,[CI] Optimize nightly CI,Performance,Include,Performance optimization +[bugfix]cancel tokenize for layerwise_proxy (#3914),3914,[bugfix]cancel tokenize for layerwise_proxy,Others,Include,Bug fix +add new e2e tests case for aclgraph memory (#3879),3879,add new e2e tests case for aclgraph memory,Highlights,Include,Major feature +mfix bug when max_seqs=14 in mtp=2 scenario and raise error when cudagraph_capture_sizes can't be an integer multiple of uniform_decode_query_lentp (#3910),3910,mfix bug when max_seqs=14 in mtp=2 scenario and raise error when cudagraph_capture_sizes can't be an integer multiple of uniform_decode_query_lentp,Highlights,Include,Major feature +[Bugfix] Fix MTP support for lmhead_tensor_parallel_size (#3915),3915,[Bugfix] Fix MTP support for lmhead_tensor_parallel_size,Highlights,Include,Major feature +[Test] Add new test model for aclgraph single_request (#3888),3888,[Test] Add new test model for aclgraph single_request,Highlights,Include,Major feature +[main][bugfix] fix valueError in static_forward_context when prefix is empty (#3924),3924,[main][bugfix] fix valueError in static_forward_context when prefix is empty,Others,Include,Bug fix +[E2E][MM] Add e2e tests for InternVL model (#3796),3796,[E2E][MM] Add e2e tests for InternVL model,Highlights,Include,Model support +[feature] support pcp + mtp (with pd disaggregate) (#3822),3822,[feature] support pcp + mtp (with pd disaggregate),Highlights,Include,Major feature +[Doc] Update doc for release notese (#3853),3853,[Doc] Update doc for release notese,Documentation,Include,Documentation update +Update torch-npu version to 2.7.1 (#3896),3896,Update torch-npu version to 2.7.1,Dependencies,Include,Dependency update +[CI][Nightly] Correct the commit hash available for mooncake (#3943),3943,[CI][Nightly] Correct the commit hash available for mooncake,Highlights,Include,Major feature +[Perf] Move attention update stream out of loop to optimize performance (#3848),3848,[Perf] Move attention update stream out of loop to optimize performance,Performance,Include,Performance optimization +[Feat][UT] Support Deepseekv32 FULL_DECODE_ONLY mode and add unit test of sfa_v1 (#3763),3763,[Feat][UT] Support Deepseekv32 FULL_DECODE_ONLY mode and add unit test of sfa_v1,Ignore,Ignore,Internal/CI/test change +correct bug to fix the value of max_num_tokens (#3933),3933,correct bug to fix the value of max_num_tokens,Others,Include,Miscellaneous +[CI][Nightly] Fix mooncake build (#3958),3958,[CI][Nightly] Fix mooncake build,Highlights,Include,Major feature +[Test] Add new e2e test use deepseek-v2-lite in ge graph mode (#3937),3937,[Test] Add new e2e test use deepseek-v2-lite in ge graph mode,Others,Include,Miscellaneous +revert TND modify when dcp pcp (#3948),3948,revert TND modify when dcp pcp,Ignore,Ignore,Internal/CI/test change +Quality enhancement: Immediately interrupt execution when memory OOM (#3932),3932,Quality enhancement: Immediately interrupt execution when memory OOM,Others,Include,Miscellaneous +[Test] Add accuracy test for qwen3-8b-w8a8 (#3799),3799,[Test] Add accuracy test for qwen3-8b-w8a8,Others,Include,Miscellaneous +[BugFix] Fix deepseek v3.2 mtp bug. (#3900),3900,[BugFix] Fix deepseek v3.2 mtp bug.,Highlights,Include,Major feature +[Test]Add accuracy test for multiple models (#3823),3823,[Test]Add accuracy test for multiple models,Others,Include,Miscellaneous +[PD Disaggregation]Set adxl engine as default backend and update README (#3761),3761,[PD Disaggregation]Set adxl engine as default backend and update README,Highlights,Include,Major feature +[TEST]Add full graph for multimodal nightly tests (#3968),3968,[TEST]Add full graph for multimodal nightly tests,Highlights,Include,Major feature +[Perf] move quant before allgather in Allgather EP (#3420),3420,[Perf] move quant before allgather in Allgather EP,Performance,Include,Performance optimization +[ModelRunner][Refactor] Refactor kv cache tensor initialization logic (#3106),3106,[ModelRunner][Refactor] Refactor kv cache tensor initialization logic,Ignore,Ignore,Internal/CI/test change +[Test] Add accuracy test for qwen3-30b-a3b-w8a8 (#3807),3807,[Test] Add accuracy test for qwen3-30b-a3b-w8a8,Hardware and Operator Support,Include,Operator/hardware support +[Doc] Refactor the DeepSeek-V3.2-Exp tutorial. (#3871),3871,[Doc] Refactor the DeepSeek-V3.2-Exp tutorial.,Documentation,Include,Model tutorial +support qwen3-next full_decode_only mode. (#3949),3949,support qwen3-next full_decode_only mode.,Highlights,Include,Major feature +[docs] add aclgraph developer guide (#3683),3683,[docs] add aclgraph developer guide,Highlights,Include,Major feature +[Doc] Update version policy (#3999),3999,[Doc] Update version policy,Documentation,Include,Documentation update +[Doc] add mtp doc (#3770),3770,[Doc] add mtp doc,Highlights,Include,Major feature +[docs] Add kv pool developer guide (#3752),3752,[docs] Add kv pool developer guide,Highlights,Include,Major feature +[Doc]Add developer guide of eplb. (#3759),3759,[Doc]Add developer guide of eplb.,Highlights,Include,Major feature +[main][doc][kv_pool]Add adxl timeout parameter in kv pool user guide (#4012),4012,[main][doc][kv_pool]Add adxl timeout parameter in kv pool user guide,Highlights,Include,Major feature +[Test] Refactor accuracy test to nightly test (#3814),3814,[Test] Refactor accuracy test to nightly test,Others,Include,Miscellaneous +[P/D]Make kv-transfer env variable take effect & Fix load-balance proxy (#3981),3981,[P/D]Make kv-transfer env variable take effect & Fix load-balance proxy,Others,Include,Miscellaneous +[Feat](Mooncake) Supports multiple input suffixes for global_segment_size (#3690),3690,[Feat](Mooncake) Supports multiple input suffixes for global_segment_size,Highlights,Include,Major feature +[feat]decode convert bsnd to tnd and fix bug when pcp and dcp (#3980),3980,[feat]decode convert bsnd to tnd and fix bug when pcp and dcp,Highlights,Include,Major feature +[TEST]Update nightly acc test standard (#4032),4032,[TEST]Update nightly acc test standard,Others,Include,Miscellaneous +[CI] Quick fix mooncake for nightly-ci (#4028),4028,[CI] Quick fix mooncake for nightly-ci,Highlights,Include,Major feature +[Bugfix] Add constraints for sequence parallelism (#4014),4014,[Bugfix] Add constraints for sequence parallelism,Highlights,Include,Major feature +[BugFix][main] Adapted to torch_npu.npu_fused_infer_attention_score (#4025),4025,[BugFix][main] Adapted to torch_npu.npu_fused_infer_attention_score,Performance,Include,Operator fusion +[main][bugfix] Fix a rare bug triggered by _npu_paged_attention in FULL_DECODE_ONLY mode (#3986),3986,[main][bugfix] Fix a rare bug triggered by _npu_paged_attention in FULL_DECODE_ONLY mode,Highlights,Include,Major feature +[long_seq] fix A2 accuracy problem (#4030),4030,[long_seq] fix A2 accuracy problem,Hardware and Operator Support,Include,Operator/hardware support +[Feat] update op for mla (#4000),4000,[Feat] update op for mla,Features,Include,New feature +[UT] Add new ut case for aclgraph in auto enable (#4031),4031,[UT] Add new ut case for aclgraph in auto enable,Ignore,Ignore,Internal/CI/test change +[Doc] Add model feature matrix table. (#4040),4040,[Doc] Add model feature matrix table.,Documentation,Include,Documentation update +[Feat] Adapted mtp function to Qwen3-next (#3918),3918,[Feat] Adapted mtp function to Qwen3-next,Highlights,Include,Major feature +[BugFix]Fix group list type of mc2. (#4047),4047,[BugFix]Fix group list type of mc2.,Others,Include,Bug fix +[CI]Fix eplb ci. (#4052),4052,[CI]Fix eplb ci.,Ignore,Ignore,Internal/CI/test change +[Bugfix] fix sleepmode level2 e2e test (#4019),4019,[Bugfix] fix sleepmode level2 e2e test,Ignore,Ignore,Internal/CI/test change +[P/D][BugFix]Fix proxy format processing errors & Layerwise connector performance optimization (#4043),4043,[P/D][BugFix]Fix proxy format processing errors & Layerwise connector performance optimization,Performance,Include,Performance optimization +[BugFix] Improve the performance of prefixcache features (#4022),4022,[BugFix] Improve the performance of prefixcache features,Performance,Include,Performance optimization +[Bugfix]fix pcp dcp attn aclgraph (#4066),4066,[Bugfix]fix pcp dcp attn aclgraph,Highlights,Include,Major feature +[Info][main] Corrected the errors in the information (#4055),4055,[Info][main] Corrected the errors in the information,Others,Include,Miscellaneous +[TEST]Add qwen3-235b-w8a8 and qwen3-30b-w8a8 nightly test (#3973),3973,[TEST]Add qwen3-235b-w8a8 and qwen3-30b-w8a8 nightly test,Others,Include,Miscellaneous +[Doc] Remove extra MLAPO installation step for DeepSeek-V3.2. (#4024),4024,[Doc] Remove extra MLAPO installation step for DeepSeek-V3.2.,Documentation,Include,Model tutorial +[docs] [P/D] add feature guide for disaggregated-prefill (#3950),3950,[docs] [P/D] add feature guide for disaggregated-prefill,Highlights,Include,Major feature +[Feat] flashcomm_v2 optim solution (#3232),3232,[Feat] flashcomm_v2 optim solution,Performance,Include,Performance optimization +[Feature][Build] Upgrade the minimum version to 3.10 (#3926),3926,[Feature][Build] Upgrade the minimum version to 3.10,Features,Include,New feature +[Fix] fix Qwen2-Audio-7B-Instruct accuracy test (#4017),4017,[Fix] fix Qwen2-Audio-7B-Instruct accuracy test,Ignore,Ignore,Internal/CI/test change +[Typo] LLama has been changed to Llama (#4089),4089,[Typo] LLama has been changed to Llama,Others,Include,Miscellaneous +[Core] Restore scheduling logic under default configuration (#3967),3967,[Core] Restore scheduling logic under default configuration,Others,Include,Miscellaneous +[Doc] add qwen3 w4a4 tutorial (#4076),4076,[Doc] add qwen3 w4a4 tutorial,Documentation,Include,Documentation update +[BugFix] Fixes Qwen3-Next enable nz accuracy problem (#4058),4058,[BugFix] Fixes Qwen3-Next enable nz accuracy problem,Highlights,Include,Model support +[Doc] Add release note for v0.11.0rc1 (#3931),3931,[Doc] Add release note for v0.11.0rc1,Documentation,Include,Documentation update +[main][Bugfix] Fix ngram precision issue and open e2e ngram test (#4090),4090,[main][Bugfix] Fix ngram precision issue and open e2e ngram test,Ignore,Ignore,Internal/CI/test change +[feature] chunkprefill support pcp&dcp (#3801),3801,[feature] chunkprefill support pcp&dcp,Highlights,Include,Major feature +[Doc] Recover installation doc to use pip install (#4109),4109,[Doc] Recover installation doc to use pip install,Documentation,Include,Documentation update +[Test] Add nightly test for DeepSeek-V3.2-Exp (#3908),3908,[Test] Add nightly test for DeepSeek-V3.2-Exp,Highlights,Include,Model support +[Fix] Refactor and fix dist test to e2e full test (#3808),3808,[Fix] Refactor and fix dist test to e2e full test,Ignore,Ignore,Internal/CI/test change +[Fixbug] Fix ut test (#4116),4116,[Fixbug] Fix ut test,Ignore,Ignore,Internal/CI/test change +Remove VLLM_USE_V1 (#4086),4086,Remove VLLM_USE_V1,Others,Include,Miscellaneous +[CI] Integrate mooncake to vllm-ascend base image (#4062),4062,[CI] Integrate mooncake to vllm-ascend base image,Ignore,Ignore,Internal/CI/test change +[TEST]Update nightly cases and add mtpx (#4111),4111,[TEST]Update nightly cases and add mtpx,Highlights,Include,Major feature +oproj TP support acl graph (#4073),4073,oproj TP support acl graph,Features,Include,New feature +[Test][Accuracy] Add accuracy evaluation config for InternVL3_5-8B (#3964),3964,[Test][Accuracy] Add accuracy evaluation config for InternVL3_5-8B,Highlights,Include,Model support +[Misc][Doc] Add service profiling feature with user guide (#3756),3756,[Misc][Doc] Add service profiling feature with user guide,Documentation,Include,Documentation update +[Perf] Remove D2H operations to imporve performance (#4063),4063,[Perf] Remove D2H operations to imporve performance,Performance,Include,Performance optimization +"[Doc] Fix DeepSeek-3.2-Exp doc, remove v0.11.0rc0 outdated infos. (#4095)",4095,"[Doc] Fix DeepSeek-3.2-Exp doc, remove v0.11.0rc0 outdated infos.",Documentation,Include,Documentation update +fix fullgraph in ds. (#4016),4016,fix fullgraph in ds.,Others,Include,Miscellaneous +[feature] support pcp + mtp (in pd co-locate scenario) (#4098),4098,[feature] support pcp + mtp (in pd co-locate scenario),Highlights,Include,Major feature +[main][bugfix] Change seq_lens in dummy attn_metadata to max_query_len (#4097),4097,[main][bugfix] Change seq_lens in dummy attn_metadata to max_query_len,Others,Include,Bug fix +[CI] Fix nightly-ci (#4159),4159,[CI] Fix nightly-ci,Others,Include,Miscellaneous +Upgrade to 0.11.1 newest vllm commit (#3982),3982,Upgrade to 0.11.1 newest vllm commit,Others,Include,Miscellaneous +[Perf] fix async copy for async scheduling (#4113),4113,[Perf] fix async copy for async scheduling,Performance,Include,Performance optimization +[Perf] [MoE] optimize all2allv (#3738),3738,[Perf] [MoE] optimize all2allv,Performance,Include,Performance optimization +[Bugfix] fix mtp profile run error where main model and mtp model use different quantization (#4102),4102,[Bugfix] fix mtp profile run error where main model and mtp model use different quantization,Highlights,Include,Major feature +[BugFix] adapted e2e tests for Qwen3-next-mtp (#4160),4160,[BugFix] adapted e2e tests for Qwen3-next-mtp,Ignore,Ignore,Internal/CI/test change +[BugFix] Fix kv_no_split not contiguous (#3594),3594,[BugFix] Fix kv_no_split not contiguous,Others,Include,Bug fix +[Info][main] Correct the mistake in information documents (#4157),4157,[Info][main] Correct the mistake in information documents,Others,Include,Miscellaneous +[Test]Add ut test qwen3_moe and sfa (#4121),4121,[Test]Add ut test qwen3_moe and sfa,Ignore,Ignore,Internal/CI/test change +[CI] Remove unsupported python 3.9 format check (#4172),4172,[CI] Remove unsupported python 3.9 format check,Ignore,Ignore,Internal/CI/test change +[CI] Add daily images build for nightly ci (#3989),3989,[CI] Add daily images build for nightly ci,Others,Include,Miscellaneous +[long_seq_Feat] support chunk prefill (#4158),4158,[long_seq_Feat] support chunk prefill,Features,Include,New feature +[CI] Add multi-nodes EPLB configs of DeepSeek-R1-W8A8 & Qwen3-235B-W8A8 (#4144),4144,[CI] Add multi-nodes EPLB configs of DeepSeek-R1-W8A8 & Qwen3-235B-W8A8,Ignore,Ignore,Internal/CI/test change +[Bugfix] fix cannot import name get_mp_context (#4174),4174,[Bugfix] fix cannot import name get_mp_context,Others,Include,Bug fix +[Feat] Adds a utility for printing from within ACL graphs (#4162),4162,[Feat] Adds a utility for printing from within ACL graphs,Features,Include,New feature +[Platform] Add import_kernels interface (#3694),3694,[Platform] Add import_kernels interface,Hardware and Operator Support,Include,Operator/hardware support +[Misc] Add benchmark results into `.gitignore` (#4200),4200,[Misc] Add benchmark results into `.gitignore`,Others,Include,Miscellaneous +[Test] Add deepseek v3.2 exp nightly test (#4191),4191,[Test] Add deepseek v3.2 exp nightly test,Highlights,Include,Model support +[CI] Fix no space left in build wheel CI. (#4215),4215,[CI] Fix no space left in build wheel CI.,Ignore,Ignore,Internal/CI/test change +support FULL graph mode for GQA (#3970),3970,support FULL graph mode for GQA,Highlights,Include,Major feature +[TEST]Update prefixcache perf threshold for qwen3-32b-int8 (#4220),4220,[TEST]Update prefixcache perf threshold for qwen3-32b-int8,Others,Include,Miscellaneous +make vllm-ascend work well in developer mode (#4179),4179,make vllm-ascend work well in developer mode,Others,Include,Miscellaneous +[Bugfix]Fix moe error when sp chunked the hidden_states (#4212),4212,[Bugfix]Fix moe error when sp chunked the hidden_states,Others,Include,Bug fix +[main][misc]change default capture size for Qwen3-MoE when using full dp (#4199),4199,[main][misc]change default capture size for Qwen3-MoE when using full dp,Others,Include,Miscellaneous +[feature] Mooncake_connector support pcp/dcp (#4183),4183,[feature] Mooncake_connector support pcp/dcp,Highlights,Include,Major feature +[P/D] pd proxy support ipv6 (#4161),4161,[P/D] pd proxy support ipv6,Features,Include,New feature +[bugfix] fix proxy hen host ip using domain name (#4243),4243,[bugfix] fix proxy hen host ip using domain name,Others,Include,Bug fix +[Fix] Sorts aclgraph batch sizes in ascending order (#4230),4230,[Fix] Sorts aclgraph batch sizes in ascending order,Highlights,Include,Major feature +[refactor]support gatingtopk operator generalization (#2958),2958,[refactor]support gatingtopk operator generalization,Ignore,Ignore,Internal/CI/test change +[bugfix] pcp + mtp acl graph bugfix (#4221),4221,[bugfix] pcp + mtp acl graph bugfix,Highlights,Include,Major feature +[CI] Fix kubernetes failed to resolve ip by dns name (#4240),4240,[CI] Fix kubernetes failed to resolve ip by dns name,Ignore,Ignore,Internal/CI/test change +[Bugfix] fix hang in async scheduling (#4233),4233,[Bugfix] fix hang in async scheduling,Performance,Include,Performance optimization +remove get_metadata_cls (#4087),4087,remove get_metadata_cls,Others,Include,Miscellaneous +[doc]fix readme for kv pool user guide (#4271),4271,[doc]fix readme for kv pool user guide,Highlights,Include,Major feature +[Docs] Improve the AISBench multi-modal testing docs (#4255),4255,[Docs] Improve the AISBench multi-modal testing docs,Others,Include,Miscellaneous +[misc] clean up get_metadata_cls (#4276),4276,[misc] clean up get_metadata_cls,Ignore,Ignore,Internal/CI/test change +[long seq feat]GQA support long-prefill-token-threshold and fixbug (#4209),4209,[long seq feat]GQA support long-prefill-token-threshold and fixbug,Features,Include,New feature +"[Bugfix] fix nightly multi-node EPLB tests' ""DYNAMIC_EPLB=true"" environment not working (#4223)",4223,"[Bugfix] fix nightly multi-node EPLB tests' ""DYNAMIC_EPLB=true"" environment not working",Ignore,Ignore,Internal/CI/test change +avoid mrope fusion op when running qwen2.5-vl on a+x machine (#4270),4270,avoid mrope fusion op when running qwen2.5-vl on a+x machine,Performance,Include,Operator fusion +[Test] Add tests for the multi-node DeepSeek-V2-Lite network in GE Graph (#4039),4039,[Test] Add tests for the multi-node DeepSeek-V2-Lite network in GE Graph,Others,Include,Miscellaneous +[CI] Add mla ut (#4280),4280,[CI] Add mla ut,Ignore,Ignore,Internal/CI/test change +[Feat] Support MTP to running in full graph mode (#3892),3892,[Feat] Support MTP to running in full graph mode,Highlights,Include,Major feature +[Test] quick fix mla ut (#4318),4318,[Test] quick fix mla ut,Ignore,Ignore,Internal/CI/test change +[Test] Add ACL graph capture/replay DP test (#4259),4259,[Test] Add ACL graph capture/replay DP test,Others,Include,Miscellaneous +[Feat][BugFix]Support the Qwen3-Next-80B-A3B-Instruct quantization model&Fix the NZ issue (#4245),4245,[Feat][BugFix]Support the Qwen3-Next-80B-A3B-Instruct quantization model&Fix the NZ issue,Highlights,Include,Model support +eplb redundant expert bugfix (#4291),4291,eplb redundant expert bugfix,Highlights,Include,Major feature +[Readme] EPLB Support Scenarios (#4314),4314,[Readme] EPLB Support Scenarios,Highlights,Include,Major feature +[MM][Bugfix] Add error log for VL models when enabling FLASHCOMM (#4272),4272,[MM][Bugfix] Add error log for VL models when enabling FLASHCOMM,Performance,Include,Performance optimization +[Feat][Doc] Add a load_balance_dp_proxy in examples and external dp doc. (#4265),4265,[Feat][Doc] Add a load_balance_dp_proxy in examples and external dp doc.,Documentation,Include,Documentation update +[Test] Add ut test for torchair (#4287),4287,[Test] Add ut test for torchair,Ignore,Ignore,Internal/CI/test change +[bugfix] bugfix for PD disaggregate (#4319),4319,[bugfix] bugfix for PD disaggregate,Others,Include,Bug fix +[EPLB] Eplb Verify Fix (#4333),4333,[EPLB] Eplb Verify Fix,Highlights,Include,Major feature +[Doc] add release note for v0.11.0rc2 (#4348),4348,[Doc] add release note for v0.11.0rc2,Documentation,Include,Documentation update +[BugFix] Fix some issues caused by the ascending order of cudagraph_capture_sizes (#4338),4338,[BugFix] Fix some issues caused by the ascending order of cudagraph_capture_sizes,Others,Include,Bug fix +[Bugfix][KV Pool]fix get_ip import in mooncake_store (#4355),4355,[Bugfix][KV Pool]fix get_ip import in mooncake_store,Highlights,Include,Major feature +[Doc]Add single node PD disaggregation instructions (#4337),4337,[Doc]Add single node PD disaggregation instructions,Highlights,Include,Major feature +[CI] Fix nightly CI for A2 series (#3825),3825,[CI] Fix nightly CI for A2 series,Hardware and Operator Support,Include,Operator/hardware support +[Doc] Upgrade multi-node doc (#4365),4365,[Doc] Upgrade multi-node doc,Documentation,Include,Documentation update +Change the first letter to uppercase (#4375),4375,Change the first letter to uppercase,Others,Include,Miscellaneous +[Fix] Remove unnecessary NPU synchronization in MTP proposer (#4325),4325,[Fix] Remove unnecessary NPU synchronization in MTP proposer,Highlights,Include,Major feature +[TEST]Update deepseek mtpx acc cases standard (#4321),4321,[TEST]Update deepseek mtpx acc cases standard,Highlights,Include,Major feature +Drop 0.11.0 support (#4377),4377,Drop 0.11.0 support,Features,Include,New feature +[Fix] fix aclgraph e2e test. (#4131),4131,[Fix] fix aclgraph e2e test.,Ignore,Ignore,Internal/CI/test change +[Refactor] remove moe type of multicast. (#4224),4224,[Refactor] remove moe type of multicast.,Ignore,Ignore,Internal/CI/test change +[Bugfix][MoE] enable force_load_balance in aclgraph (#4366),4366,[Bugfix][MoE] enable force_load_balance in aclgraph,Highlights,Include,Major feature +[feature] vllm-ascend support msprobe (eager mode dump) (#4241),4241,[feature] vllm-ascend support msprobe (eager mode dump),Features,Include,New feature +Bump actions/checkout from 4 to 6 (#4380),4380,Bump actions/checkout from 4 to 6,Ignore,Ignore,Internal/CI/test change +[Bugfix]Fix the hang issue of multimodal model when running with DP>1 (#4392),4392,[Bugfix]Fix the hang issue of multimodal model when running with DP>1,Others,Include,Bug fix +Document error correction (#4422),4422,Document error correction,Others,Include,Miscellaneous +[Bugfix] fix patch typo (#4351),4351,[Bugfix] fix patch typo,Others,Include,Bug fix +[bugfix]Return the Transformer version from 4.57.2 to 4.57.1 (#4423),4423,[bugfix]Return the Transformer version from 4.57.2 to 4.57.1,Others,Include,Bug fix +[Bugfix] use module-level import for patched function in Qwen3Next (#4354),4354,[Bugfix] use module-level import for patched function in Qwen3Next,Others,Include,Bug fix +[MM][Bugfix] Minor fix for VL model verification (#4384),4384,[MM][Bugfix] Minor fix for VL model verification,Others,Include,Bug fix +[misc] Remove useless patch_logits (#4252),4252,[misc] Remove useless patch_logits,Others,Include,Miscellaneous +[TEST] Delete Comment (#4427),4427,[TEST] Delete Comment,Others,Include,Miscellaneous +mkdir triton package and move triton files (#4420),4420,mkdir triton package and move triton files,Hardware and Operator Support,Include,Operator/hardware support +upgrade to vllm 0.11.2 (#4400),4400,upgrade to vllm 0.11.2,Others,Include,Miscellaneous +[CI] clean up ci (#4452),4452,[CI] clean up ci,Ignore,Ignore,Internal/CI/test change +[refact] unified soc_version code (#4359),4359,[refact] unified soc_version code,Others,Include,Miscellaneous +Change comment location (#4432),4432,Change comment location,Others,Include,Miscellaneous +[UT] Fix ut test (#4472),4472,[UT] Fix ut test,Ignore,Ignore,Internal/CI/test change +chip type judgement code optimization (#4485),4485,chip type judgement code optimization,Performance,Include,Performance optimization +[CI][Nightly] Support local debugging for multi-node CI test cases (#4489),4489,[CI][Nightly] Support local debugging for multi-node CI test cases,Features,Include,New feature +[BugFix] Adapted Qwen3-Next eager mode to v0.11.2 (#4477),4477,[BugFix] Adapted Qwen3-Next eager mode to v0.11.2,Highlights,Include,Model support +[bugfix] fix ray start failed: local_world_size cannot little than visible device count error (#4457),4457,[bugfix] fix ray start failed: local_world_size cannot little than visible device count error,Others,Include,Bug fix +[feature] Add Custom Op grouped_matmul_swiglu_quant (#4431),4431,[feature] Add Custom Op grouped_matmul_swiglu_quant,Hardware and Operator Support,Include,Operator/hardware support +[TEST] Add eagle proposer ut (#4447),4447,[TEST] Add eagle proposer ut,Ignore,Ignore,Internal/CI/test change +[main]Upgrade cann to 8.3rc2 (#4350),4350,[main]Upgrade cann to 8.3rc2,Others,Include,Miscellaneous +[Quantization] Support compressed tensors w8a8 static and w8a8 dynamic weight (#4036),4036,[Quantization] Support compressed tensors w8a8 static and w8a8 dynamic weight,Features,Include,New feature +[MM][Model][Perf] Remove Qwen2.5-VL modeling files and add patch for VisionAttention (#4349),4349,[MM][Model][Perf] Remove Qwen2.5-VL modeling files and add patch for VisionAttention,Performance,Include,Performance optimization +[P/D] Add readme for PD separation (#4182),4182,[P/D] Add readme for PD separation,Documentation,Include,Documentation update +[Doc]Delete equals sign (#4537),4537,[Doc]Delete equals sign,Documentation,Include,Documentation update +[Kernel] add custom op GmmSwigluQuantWeightNzTensorList (#3804),3804,[Kernel] add custom op GmmSwigluQuantWeightNzTensorList,Hardware and Operator Support,Include,Operator/hardware support +[Feature][main]reconstruction kvpool connector to ascend connector (#4438),4438,[Feature][main]reconstruction kvpool connector to ascend connector,Features,Include,New feature +【OPS】qwen3-next support triton chunk_gated_delta_rule ops (#4070),4070,【OPS】qwen3-next support triton chunk_gated_delta_rule ops,Highlights,Include,Model support +update triton package url (#4552),4552,update triton package url,Hardware and Operator Support,Include,Operator/hardware support +[Bugfix] Fix model run _npu_flash_attention hang issue (#4410),4410,[Bugfix] Fix model run _npu_flash_attention hang issue,Hardware and Operator Support,Include,Operator/hardware support +[Doc] Add single NPU tutorial for Qwen2.5-Omni-7B (#4446),4446,[Doc] Add single NPU tutorial for Qwen2.5-Omni-7B,Documentation,Include,Model tutorial +Update triton package name (#4563),4563,Update triton package name,Hardware and Operator Support,Include,Operator/hardware support +[bugfix] dep ineffective (#4417),4417,[bugfix] dep ineffective,Others,Include,Bug fix +[P/D] [bugfix] add get_kv_connector_handshake_metadata func for 0.11.2 (#4567),4567,[P/D] [bugfix] add get_kv_connector_handshake_metadata func for 0.11.2,Others,Include,Bug fix +drop ascend scheduler (#4498),4498,drop ascend scheduler,Others,Include,Miscellaneous +improve soc version (#4522),4522,improve soc version,Others,Include,Miscellaneous +[MM][Model] Remove Qwen2-VL modeling files (#4534),4534,[MM][Model] Remove Qwen2-VL modeling files,Others,Include,Miscellaneous +Move mla to ops module (#4575),4575,Move mla to ops module,Others,Include,Miscellaneous +[Bugfix] fix dp parallel + tp > 1 offline inference port conflict (#4539),4539,[Bugfix] fix dp parallel + tp > 1 offline inference port conflict,Others,Include,Bug fix +remove qwen3-next model file (#4573),4573,remove qwen3-next model file,Highlights,Include,Model support +[feature]Pooling Features and PCP Adaptation (#4143),4143,[feature]Pooling Features and PCP Adaptation,Highlights,Include,Major feature +"Revert ""drop ascend scheduler"" (#4580)",4580,"Revert ""drop ascend scheduler""",Ignore,Ignore,Internal/CI/test change +[CI] Skip test_ngram_correctness as the oom issue block CI (#4578),4578,[CI] Skip test_ngram_correctness as the oom issue block CI,Ignore,Ignore,Internal/CI/test change +[bugfix] Repair the problem of moe model accuracy caused by version upgrade. (#4562),4562,[bugfix] Repair the problem of moe model accuracy caused by version upgrade.,Others,Include,Bug fix +[Bugfix] Fix kvpool precision synchronization (#4574),4574,[Bugfix] Fix kvpool precision synchronization,Others,Include,Bug fix +[feature] Support W8A8 PD-Mix Quantization (#4235),4235,[feature] Support W8A8 PD-Mix Quantization,Features,Include,New feature +[EPLB][Ops] Integerate grouped_matmul_swiglu_quant_weight_nz_tensor_list operator into dynamic EPLB (#4216),4216,[EPLB][Ops] Integerate grouped_matmul_swiglu_quant_weight_nz_tensor_list operator into dynamic EPLB,Highlights,Include,Major feature +[OPS] add bmm_transpose ops (#3990),3990,[OPS] add bmm_transpose ops,Others,Include,Miscellaneous +[BugFix] Fix Qwen2.5_Omni vision customized op attr err (#4568),4568,[BugFix] Fix Qwen2.5_Omni vision customized op attr err,Others,Include,Bug fix +[Bugfix] Resolve MTP > 1 issue when lm head tp > 1 (#4254),4254,[Bugfix] Resolve MTP > 1 issue when lm head tp > 1,Highlights,Include,Major feature +Bump actions/setup-python from 6.0.0 to 6.1.0 (#4591),4591,Bump actions/setup-python from 6.0.0 to 6.1.0,Ignore,Ignore,Internal/CI/test change +[Bugfix] Fix bug with establishing the flashcomm2 and pp communication domains. (#4458),4458,[Bugfix] Fix bug with establishing the flashcomm2 and pp communication domains.,Performance,Include,Performance optimization +[Kernel] add triton kernels for sampling (#4550),4550,[Kernel] add triton kernels for sampling,Hardware and Operator Support,Include,Operator/hardware support +add _cann_ops_custom gitignore (#4605),4605,add _cann_ops_custom gitignore,Others,Include,Miscellaneous +[Feature] Integrate Suffix Spec Decoding (#4045),4045,[Feature] Integrate Suffix Spec Decoding,Highlights,Include,Major feature +upgrade torch npu version (#4433),4433,upgrade torch npu version,Others,Include,Miscellaneous +[Bugfix] PCP adaptation for VLLM v0.11.2 modifications (#4604),4604,[Bugfix] PCP adaptation for VLLM v0.11.2 modifications,Highlights,Include,Major feature +[Bug_fix] fix torchair o_proj forward parameter (#4166),4166,[Bug_fix] fix torchair o_proj forward parameter,Others,Include,Miscellaneous +[CI] drop ascend scheduler test (#4582),4582,[CI] drop ascend scheduler test,Ignore,Ignore,Internal/CI/test change +[Feat] shared expert dp for deepseek_mtp (#3811),3811,[Feat] shared expert dp for deepseek_mtp,Highlights,Include,Major feature +fix qwenvl pd smoke test error (#4597),4597,fix qwenvl pd smoke test error,Ignore,Ignore,Internal/CI/test change +[Test] Add accuracy nightly test for new models (#4262),4262,[Test] Add accuracy nightly test for new models,Others,Include,Miscellaneous +"[Doc] Fix DeepSeek-V3.2-Exp doc, add docker command. (#4479)",4479,"[Doc] Fix DeepSeek-V3.2-Exp doc, add docker command.",Documentation,Include,Model tutorial +[Test] Add GLM-4.5 nightly test (#4225),4225,[Test] Add GLM-4.5 nightly test,Highlights,Include,Model support +"[Bugfix] Remove ModelSlim-""M4 Quantization"". (#4589)",4589,"[Bugfix] Remove ModelSlim-""M4 Quantization"".",Others,Include,Bug fix +[main][bugfix] bugfix for qwen3 moe quantization (#4599),4599,[main][bugfix] bugfix for qwen3 moe quantization,Others,Include,Bug fix +[MM][Model] Remove Qwen3-VL modeling files (#4577),4577,[MM][Model] Remove Qwen3-VL modeling files,Highlights,Include,Model support +[CI]enable chunked prefill by default (#4569),4569,[CI]enable chunked prefill by default,Ignore,Ignore,Internal/CI/test change +[Refactor] Remove redundant attention operator branches. (#4531),4531,[Refactor] Remove redundant attention operator branches.,Ignore,Ignore,Internal/CI/test change +[Bugfix] Fix Qwen2.5-Omni-7B accuarcy test (#4556),4556,[Bugfix] Fix Qwen2.5-Omni-7B accuarcy test,Ignore,Ignore,Internal/CI/test change +[Bugfix]Fix eplb enable when using mtp float weights. (#4571),4571,[Bugfix]Fix eplb enable when using mtp float weights.,Highlights,Include,Major feature +Bump actions/checkout from 4.3.1 to 6.0.0 (#4592),4592,Bump actions/checkout from 4.3.1 to 6.0.0,Ignore,Ignore,Internal/CI/test change +"Revert ""[Bugfix] Fix Qwen2.5-Omni-7B accuarcy test (#4556)"" (#4556)",4556,"Revert ""[Bugfix] Fix Qwen2.5-Omni-7B accuarcy test",Ignore,Ignore,Internal/CI/test change +[CI] Drop ascend scheduler from test (#4613),4613,[CI] Drop ascend scheduler from test,Ignore,Ignore,Internal/CI/test change +add hyperlink (#4588),4588,add hyperlink,Others,Include,Miscellaneous +[Doc]clean up ascend scheduler config from doc (#4612),4612,[Doc]clean up ascend scheduler config from doc,Documentation,Include,Documentation update +[Doc] Add tutorial for Qwen3-Coder-30B-A3B (#4391),4391,[Doc] Add tutorial for Qwen3-Coder-30B-A3B,Hardware and Operator Support,Include,Operator/hardware support +[Ops][Triton] Add a triton kernel supporting partial rope. (#4413),4413,[Ops][Triton] Add a triton kernel supporting partial rope.,Hardware and Operator Support,Include,Operator/hardware support +clean up model module (#4611),4611,clean up model module,Others,Include,Miscellaneous +[Doc] Refactor the DeepSeek-V3.1 tutorial. (#4399),4399,[Doc] Refactor the DeepSeek-V3.1 tutorial.,Documentation,Include,Model tutorial +[performance] Enhance performance after enabling min_p (#4529),4529,[performance] Enhance performance after enabling min_p,Performance,Include,Performance optimization +【doc fix】doc fix: deepseekv3.1 (#4645),4645,【doc fix】doc fix: deepseekv3.1,Others,Include,Miscellaneous +[Bugfix] fix custom op GmmSwigluQuantWeightNzTensorList (#4593),4593,[Bugfix] fix custom op GmmSwigluQuantWeightNzTensorList,Hardware and Operator Support,Include,Operator/hardware support +upgrade vLLM to main (#4608),4608,upgrade vLLM to main,Others,Include,Miscellaneous +[kernel] add AscendC op: lightning_indexer and sparse_flash_attention (#4625),4625,[kernel] add AscendC op: lightning_indexer and sparse_flash_attention,Hardware and Operator Support,Include,Operator/hardware support +[Doc] add release note for v0.11.0rc3 (#4646),4646,[Doc] add release note for v0.11.0rc3,Documentation,Include,Documentation update +fix typo (#4657),4657,fix typo,Others,Include,Miscellaneous +[Model] Add qwen3Next support in Main (#4596),4596,[Model] Add qwen3Next support in Main,Features,Include,New feature +[Feat] MTP support DeepSeekV3.2 (#4465),4465,[Feat] MTP support DeepSeekV3.2,Highlights,Include,Major feature +[Fix] Fix FIA `query` and `query_start_loc` shape mismatch error (#4518),4518,[Fix] Fix FIA `query` and `query_start_loc` shape mismatch error,Hardware and Operator Support,Include,Operator/hardware support +[CI] Fix ut ci: no space on the device (#4662),4662,[CI] Fix ut ci: no space on the device,Ignore,Ignore,Internal/CI/test change +[Misc] Add cann custom ops to `.gitignore` (#4670),4670,[Misc] Add cann custom ops to `.gitignore`,Hardware and Operator Support,Include,Operator/hardware support +fix custom ops env set error (#4675),4675,fix custom ops env set error,Hardware and Operator Support,Include,Operator/hardware support +[Core] Encoder separation for Encode-Prefill-Decode Disaggregation (#4176),4176,[Core] Encoder separation for Encode-Prefill-Decode Disaggregation,Highlights,Include,Major feature +upgrade vLLM to 0.12.0 tag (#4647),4647,upgrade vLLM to 0.12.0 tag,Others,Include,Miscellaneous +Remove cancel for main to main check (#4685),4685,Remove cancel for main to main check,Others,Include,Miscellaneous +Adopt inductor fusion and define quantization fusion pass (#4168),4168,Adopt inductor fusion and define quantization fusion pass,Performance,Include,Operator fusion +Remove ascend schuduler ut (#4684),4684,Remove ascend schuduler ut,Others,Include,Miscellaneous +【fix】ops gatingtopk fix nightly ci error (#4340),4340,【fix】ops gatingtopk fix nightly ci error,Others,Include,Miscellaneous +[MM][Patch] Remove patch for cos/sin cache (#4672),4672,[MM][Patch] Remove patch for cos/sin cache,Others,Include,Miscellaneous +[Nightly] Optimize nightly CI (#4509),4509,[Nightly] Optimize nightly CI,Performance,Include,Performance optimization +[Misc] Upgrade vllm vllm commit to 2025_12_04 (#4690),4690,[Misc] Upgrade vllm vllm commit to 2025_12_04,Others,Include,Miscellaneous +add `dispatch_gmm_combine` kernel (#3532),3532,add `dispatch_gmm_combine` kernel,Hardware and Operator Support,Include,Operator/hardware support +[Bugfix] Quick hot fix for nightly CI (#4727),4727,[Bugfix] Quick hot fix for nightly CI,Others,Include,Bug fix +[Doc] Update vLLM version in doc (#4691),4691,[Doc] Update vLLM version in doc,Documentation,Include,Documentation update +Drop ascend scheduler (#4623),4623,Drop ascend scheduler,Others,Include,Miscellaneous +[long_seq] remove long_seq env (#4660),4660,[long_seq] remove long_seq env,Others,Include,Miscellaneous +Update comment doc (#4731),4731,Update comment doc,Others,Include,Miscellaneous +[BugFix][Triton] Fix ub overflow bug of sample_recover_tokens_kernel (#4673),4673,[BugFix][Triton] Fix ub overflow bug of sample_recover_tokens_kernel,Hardware and Operator Support,Include,Operator/hardware support +[Bugifx] fix quant_apply_mlp w1_scale type error & fix getting num_local_expert (#4632),4632,[Bugifx] fix quant_apply_mlp w1_scale type error & fix getting num_local_expert,Others,Include,Miscellaneous +[P/D][main] Clean connector history information (#4650),4650,[P/D][main] Clean connector history information,Others,Include,Miscellaneous +[CI] Fix unit test fault `no space left` (#4728),4728,[CI] Fix unit test fault `no space left`,Ignore,Ignore,Internal/CI/test change +【main】[Doc]add 2P1D instruction for single node (#4716),4716,【main】[Doc]add 2P1D instruction for single node,Documentation,Include,Documentation update +[Refactor] 1/N Refactor attention_v1 & extract attention_cp (#4628),4628,[Refactor] 1/N Refactor attention_v1 & extract attention_cp,Ignore,Ignore,Internal/CI/test change +[Bugfix]fix bmm_transpose ops for cann version (#4653),4653,[Bugfix]fix bmm_transpose ops for cann version,Others,Include,Bug fix +rm vanilla attn (#4558),4558,rm vanilla attn,Others,Include,Miscellaneous +mlapo add qdown output (#4707),4707,mlapo add qdown output,Hardware and Operator Support,Include,Operator/hardware support +[Bugfix] fix mtp and eagle aclgraph bug (#4710),4710,[Bugfix] fix mtp and eagle aclgraph bug,Highlights,Include,Major feature +support async mtp (#4511),4511,support async mtp,Highlights,Include,Major feature +[BugFix] Fix eagle3 accuracy problem when enforce_eager=True (#4521),4521,[BugFix] Fix eagle3 accuracy problem when enforce_eager=True,Highlights,Include,Major feature +[Kernel] add custom op DispatchGmmCombineDecode (#4139),4139,[Kernel] add custom op DispatchGmmCombineDecode,Hardware and Operator Support,Include,Operator/hardware support +[Feat]enable sfa cp for dsv3.2 (#4702),4702,[Feat]enable sfa cp for dsv3.2,Hardware and Operator Support,Include,Operator/hardware support +Support DeepSeekV3.2 with MLAPO operator (#4753),4753,Support DeepSeekV3.2 with MLAPO operator,Hardware and Operator Support,Include,Operator/hardware support +[P/D] check kv extra config and del hccl backend (#4547),4547,[P/D] check kv extra config and del hccl backend,Others,Include,Miscellaneous +[BugFix] Refactor ACL graph size adjustment for speculative decoding (#4640),4640,[BugFix] Refactor ACL graph size adjustment for speculative decoding,Highlights,Include,Major feature +[Feat] Add Euler xlite graph wrapper support (#4526),4526,[Feat] Add Euler xlite graph wrapper support,Highlights,Include,Major feature +fix synchronize error of exceeds_max_model_len d2h copy (#4708),4708,fix synchronize error of exceeds_max_model_len d2h copy,Others,Include,Miscellaneous +[CI] Fix ngram & suffix test oom (#4755),4755,[CI] Fix ngram & suffix test oom,Ignore,Ignore,Internal/CI/test change +Deepseek Mtp model uses the lm_head and embedding from the main model (#2790),2790,Deepseek Mtp model uses the lm_head and embedding from the main model,Highlights,Include,Major feature +remove useless patch (#4699),4699,remove useless patch,Others,Include,Miscellaneous +[Op] DeepSeekV3.2 support bmm_transpose operator (#4631),4631,[Op] DeepSeekV3.2 support bmm_transpose operator,Hardware and Operator Support,Include,Operator/hardware support +[EPLB] Add log Info for moe_load Imbalance Ratio (#4482),4482,[EPLB] Add log Info for moe_load Imbalance Ratio,Highlights,Include,Major feature +[Fix] skip xlite e2e test (#4786),4786,[Fix] skip xlite e2e test,Ignore,Ignore,Internal/CI/test change +[Bugfix] Fix Dcp dimension mismatch when enable Mlapo (#4687),4687,[Bugfix] Fix Dcp dimension mismatch when enable Mlapo,Highlights,Include,Major feature +[Kernel] add custom moe ops for prefill (#4194),4194,[Kernel] add custom moe ops for prefill,Hardware and Operator Support,Include,Operator/hardware support +Bump actions/checkout from 6.0.0 to 6.0.1 (#4772),4772,Bump actions/checkout from 6.0.0 to 6.0.1,Ignore,Ignore,Internal/CI/test change +[Doc] Add Qwen3-235B tutorial (#4358),4358,[Doc] Add Qwen3-235B tutorial,Documentation,Include,Documentation update +[DP] Fix dp padding logic in dummyrun (#4705),4705,[DP] Fix dp padding logic in dummyrun,Others,Include,Miscellaneous +[MOE]move weight transpose to wakeup for RL secnarios (#4626),4626,[MOE]move weight transpose to wakeup for RL secnarios,Others,Include,Miscellaneous +Fix incorrect MLAPO weight release in PD mixex scenarios. (#4774),4774,Fix incorrect MLAPO weight release in PD mixex scenarios.,Hardware and Operator Support,Include,Operator/hardware support +"Revert ""[Kernel] add custom moe ops for prefill"" (#4806)",4806,"Revert ""[Kernel] add custom moe ops for prefill""",Ignore,Ignore,Internal/CI/test change +[Bugfix] Add the check for a null VllmConfig (#4749),4749,[Bugfix] Add the check for a null VllmConfig,Others,Include,Bug fix +[CI] Skip `test_suffix_correctness` (#4820),4820,[CI] Skip `test_suffix_correctness`,Ignore,Ignore,Internal/CI/test change +[Docs]fix the configuration conflicts in documentation (#4823),4823,[Docs]fix the configuration conflicts in documentation,Documentation,Include,Documentation update +[CI] Optimize CI time (#4821),4821,[CI] Optimize CI time,Ignore,Ignore,Internal/CI/test change +[KVPOOl]Support pp (#4761),4761,[KVPOOl]Support pp,Features,Include,New feature +[Feat] Multi-stream for eplb heat collection and aggregation (#4214),4214,[Feat] Multi-stream for eplb heat collection and aggregation,Highlights,Include,Major feature +[kernel] Adapt DispatchGmmCombineDecode operator to parameters of small operators (#4790),4790,[kernel] Adapt DispatchGmmCombineDecode operator to parameters of small operators,Hardware and Operator Support,Include,Operator/hardware support +[CI] Increase HCCL_BUFFSIZE for A3 (#4838),4838,[CI] Increase HCCL_BUFFSIZE for A3,Ignore,Ignore,Internal/CI/test change +[Bugfix]fix bmm_transpose ops in dsv32 (#4791),4791,[Bugfix]fix bmm_transpose ops in dsv32,Others,Include,Bug fix +[UT]add pcp aclgraph ut (#4804),4804,[UT]add pcp aclgraph ut,Ignore,Ignore,Internal/CI/test change +"[Usability]local_buffer_size support for units: GB, MB, KB, B (#4829)",4829,"[Usability]local_buffer_size support for units: GB, MB, KB, B",Features,Include,New feature +[Refactor] 2/N Unify all mask generation methods and cache mask (#4779),4779,[Refactor] 2/N Unify all mask generation methods and cache mask,Ignore,Ignore,Internal/CI/test change +[CI] Setup github proxy for self_hosted runners (#4841),4841,[CI] Setup github proxy for self_hosted runners,Ignore,Ignore,Internal/CI/test change +[Fix] Add extra warmup run count for MC2 on specific SoC version (#4843),4843,[Fix] Add extra warmup run count for MC2 on specific SoC version,Others,Include,Bug fix +[Bugfix] Disable the dispatch_ffn_combine kernel in MTP path (#4751),4751,[Bugfix] Disable the dispatch_ffn_combine kernel in MTP path,Highlights,Include,Major feature +[P/D][main]Offline the llmdatadist connector related parts of the code and files. (#4780),4780,[P/D][main]Offline the llmdatadist connector related parts of the code and files.,Others,Include,Miscellaneous +Add gsm8k accuracy test for multi-note Qwen3-235B-A22B (#4802),4802,Add gsm8k accuracy test for multi-note Qwen3-235B-A22B,Hardware and Operator Support,Include,Operator/hardware support +[bugfix] fix quant method validation bug (#4831),4831,[bugfix] fix quant method validation bug,Others,Include,Bug fix +[Kernel] add custom op MatmulAllreduceAddRmsnorm (#4606),4606,[Kernel] add custom op MatmulAllreduceAddRmsnorm,Hardware and Operator Support,Include,Operator/hardware support +Drop torchair (#4814),4814,Drop torchair,Others,Include,Miscellaneous +[Nightly] Optimize nightly online test logger info (#4798),4798,[Nightly] Optimize nightly online test logger info,Performance,Include,Performance optimization +[Test] Temporarily skips Qwen3-30B-A3B-W8A8 data parallel test case (#4857),4857,[Test] Temporarily skips Qwen3-30B-A3B-W8A8 data parallel test case,Hardware and Operator Support,Include,Operator/hardware support +add e2e test for mtp async_scheduling (#4826),4826,add e2e test for mtp async_scheduling,Highlights,Include,Major feature +[Model] Support pooling models (#3122),3122,[Model] Support pooling models,Highlights,Include,Major feature +[CI]Cleanup accurary test (#4861),4861,[CI]Cleanup accurary test,Ignore,Ignore,Internal/CI/test change +[CI] Use offline mode for modelscope (#4875),4875,[CI] Use offline mode for modelscope,Ignore,Ignore,Internal/CI/test change +[Feat] Support native Kimi-K2-Thinking native W4A16 quantized experts weights (#4516),4516,[Feat] Support native Kimi-K2-Thinking native W4A16 quantized experts weights,Highlights,Include,Model support +mooncake connector support pipeline parallel & fix pp with flashcomm1 (#4054),4054,mooncake connector support pipeline parallel & fix pp with flashcomm1,Highlights,Include,Major feature +add multi_npu_qwen3_dense tutorials (#4543),4543,add multi_npu_qwen3_dense tutorials,Documentation,Include,Documentation update +[CI] fix lint (#4888),4888,[CI] fix lint,Ignore,Ignore,Internal/CI/test change +[Kernel] Add moe normal ops (#4810),4810,[Kernel] Add moe normal ops,Hardware and Operator Support,Include,Operator/hardware support +[Bugfix] Fix out-of-bounds access to token_id due to uninitialized logprobs (#4248),4248,[Bugfix] Fix out-of-bounds access to token_id due to uninitialized logprobs,Others,Include,Bug fix +[FEAT] Support DeepSeek-V3.2 with `FULL_DECODE_ONLY` mode (#4706),4706,[FEAT] Support DeepSeek-V3.2 with `FULL_DECODE_ONLY` mode,Highlights,Include,Major feature +Fixed the performance degradation issue in post-processing in speculative decoding scenarios. (#4849),4849,Fixed the performance degradation issue in post-processing in speculative decoding scenarios.,Highlights,Include,Major feature +[Bugfix] Support for mlapo in deepseekv3.1 w4a8 (#4828),4828,[Bugfix] Support for mlapo in deepseekv3.1 w4a8,Hardware and Operator Support,Include,Operator/hardware support +[Feature] Support npuhraph_ex backend (#4700),4700,[Feature] Support npuhraph_ex backend,Features,Include,New feature +[perf][dsv3.2][async_scheduling] improve dsv3.2 performance by eliminating HD synchronization (#4805),4805,[perf][dsv3.2][async_scheduling] improve dsv3.2 performance by eliminating HD synchronization,Performance,Include,Performance optimization +[BugFix][main] Adapted Qwen3-Next-MTP to chunked prefill (#4770),4770,[BugFix][main] Adapted Qwen3-Next-MTP to chunked prefill,Highlights,Include,Major feature +Update patch doc (#4869),4869,Update patch doc,Others,Include,Miscellaneous +Remove COMPILE_CUSTOM_KERNELS env (#4864),4864,Remove COMPILE_CUSTOM_KERNELS env,Hardware and Operator Support,Include,Operator/hardware support +Remove VLLM_ASCEND_ENABLE_TOPK_TOPP_OPTIMIZATION (#4860),4860,Remove VLLM_ASCEND_ENABLE_TOPK_TOPP_OPTIMIZATION,Performance,Include,Performance optimization +Remove useless env (#4858),4858,Remove useless env,Others,Include,Miscellaneous +add DeepSeek-R1 tutorial. (#4666),4666,add DeepSeek-R1 tutorial.,Documentation,Include,Model tutorial +[Doc]Add tutorial document for qwen-VL-Dense (#3516),3516,[Doc]Add tutorial document for qwen-VL-Dense,Documentation,Include,Documentation update +[Doc] Add local running multi-node nightly test case guide (#4884),4884,[Doc] Add local running multi-node nightly test case guide,Documentation,Include,Documentation update +[E2E] Remove unused PD-disaggreate scripts in E2E test. (#4837),4837,[E2E] Remove unused PD-disaggreate scripts in E2E test.,Others,Include,Miscellaneous +[E2E] Refactor the e2e testcases. (#4789),4789,[E2E] Refactor the e2e testcases.,Others,Include,Miscellaneous +[E2E] Optimize nightly testcase. (#4886),4886,[E2E] Optimize nightly testcase.,Performance,Include,Performance optimization +[feat] mlapo add bf16 no_quant support (#4852),4852,[feat] mlapo add bf16 no_quant support,Hardware and Operator Support,Include,Operator/hardware support +cleanup useless torchair logic (#4856),4856,cleanup useless torchair logic,Ignore,Ignore,Internal/CI/test change +[Feat] Flashcomm2 use o_shared linear (#4188),4188,[Feat] Flashcomm2 use o_shared linear,Performance,Include,Performance optimization +[OPS] support triton causal_conv1d_fn ops (#4119),4119,[OPS] support triton causal_conv1d_fn ops,Hardware and Operator Support,Include,Operator/hardware support +【Bugfix】bugfix_for_bmm_transpose (#4899),4899,【Bugfix】bugfix_for_bmm_transpose,Others,Include,Miscellaneous +[Bugfix] Fix the bug in sfa-cp under multi-DP scenarios. (#4850),4850,[Bugfix] Fix the bug in sfa-cp under multi-DP scenarios.,Hardware and Operator Support,Include,Operator/hardware support +[feat] apply flashcomm1 on bailing (#4868),4868,[feat] apply flashcomm1 on bailing,Performance,Include,Performance optimization +[Bugfix] support mtp kv transfer and pp partition by hand in kv transfer (#4892),4892,[Bugfix] support mtp kv transfer and pp partition by hand in kv transfer,Highlights,Include,Major feature +[doc] Add Qwen2.5 tutorials (#4636),4636,[doc] Add Qwen2.5 tutorials,Documentation,Include,Documentation update +[Fix] Delete redundant variable (#4903),4903,[Fix] Delete redundant variable,Others,Include,Bug fix +[Fusion] normalize fusion naming and enable e2e test (#4693),4693,[Fusion] normalize fusion naming and enable e2e test,Performance,Include,Operator fusion +[Bugfix] Prevent engine hang during KVCacheSendingThread startup (#4754),4754,[Bugfix] Prevent engine hang during KVCacheSendingThread startup,Others,Include,Bug fix +[CI] speed up ut (#4901),4901,[CI] speed up ut,Ignore,Ignore,Internal/CI/test change +Remove mindie_turbo (#4896),4896,Remove mindie_turbo,Others,Include,Miscellaneous +[Doc] Update structured output doc with upstream link (#4015),4015,[Doc] Update structured output doc with upstream link,Documentation,Include,Documentation update +Refactor CI workflow (#4912),4912,Refactor CI workflow,Others,Include,Miscellaneous +[CI] Cancel whl build when submitting a new commit (#4925),4925,[CI] Cancel whl build when submitting a new commit,Ignore,Ignore,Internal/CI/test change +[CI]cleanup e2e test (#4800),4800,[CI]cleanup e2e test,Ignore,Ignore,Internal/CI/test change +[Doc] Update tutorial index (#4920),4920,[Doc] Update tutorial index,Documentation,Include,Documentation update +[bugfix][refactor] fix recompute_scheduler break with vllm 0.12.0 & support async scheduling & refactor recompute_scheduler.py (#4895),4895,[bugfix][refactor] fix recompute_scheduler break with vllm 0.12.0 & support async scheduling & refactor recompute_scheduler.py,Ignore,Ignore,Internal/CI/test change +[Performance] Pre-issued exponential distribution operator. (#4908),4908,[Performance] Pre-issued exponential distribution operator.,Hardware and Operator Support,Include,Operator/hardware support +[CI] refect e2e test (#4799),4799,[CI] refect e2e test,Ignore,Ignore,Internal/CI/test change +[MoE][TorchAir] Remove FusedMoEState (#4927),4927,[MoE][TorchAir] Remove FusedMoEState,Performance,Include,Operator fusion +[main][Bugfix] Remove the ZMQ communication setup on the D node (#4926),4926,[main][Bugfix] Remove the ZMQ communication setup on the D node,Others,Include,Bug fix +[Feat] Add custom Embedding tensor model parallel (#2616),2616,[Feat] Add custom Embedding tensor model parallel,Features,Include,New feature +[Bugfix] bugfix for moe_mlp (#4822),4822,[Bugfix] bugfix for moe_mlp,Others,Include,Bug fix +BugFix: Resolve PolicyFlashlb warm up function attribute error (#4741),4741,BugFix: Resolve PolicyFlashlb warm up function attribute error,Others,Include,Miscellaneous +[CI] fix light test (#4954),4954,[CI] fix light test,Ignore,Ignore,Internal/CI/test change +【doc】Add model feature matrix (#4950),4950,【doc】Add model feature matrix,Hardware and Operator Support,Include,Operator/hardware support +[Doc] Upgrade outdated doc (#4957),4957,[Doc] Upgrade outdated doc,Documentation,Include,Documentation update +update qwen2.5vl readme (#4938),4938,update qwen2.5vl readme,Documentation,Include,Documentation update +vllm-ascend support Ascend950 with Qwen dense model. (#4228),4228,vllm-ascend support Ascend950 with Qwen dense model.,Hardware and Operator Support,Include,Operator/hardware support +[usability]Modify the default value of the protocol to ascend (#4959),4959,[usability]Modify the default value of the protocol to ascend,Others,Include,Miscellaneous +[Nightly] Remove gen_ranktable logic (#4941),4941,[Nightly] Remove gen_ranktable logic,Others,Include,Miscellaneous +[Feature] model_runner refactor (#4764),4764,[Feature] model_runner refactor,Features,Include,New feature +[doc][main] Correct mistakes in doc (#4945),4945,[doc][main] Correct mistakes in doc,Documentation,Include,Documentation update +[CI] Add mtp_proposer ut (#4397),4397,[CI] Add mtp_proposer ut,Ignore,Ignore,Internal/CI/test change +[Bugfix] Pass vllm_config to kv_connector_no_forward in NPUModelRunner (#4970),4970,[Bugfix] Pass vllm_config to kv_connector_no_forward in NPUModelRunner,Others,Include,Bug fix +[Bugfix] fix eagle proposer (#4971),4971,[Bugfix] fix eagle proposer,Highlights,Include,Major feature +[bugfix] asyncscheduler bug fix (#4968),4968,[bugfix] asyncscheduler bug fix,Performance,Include,Performance optimization +[doc][main] Correct more doc mistakes (#4958),4958,[doc][main] Correct more doc mistakes,Documentation,Include,Documentation update +"Revert ""[Bugfix] support mtp kv transfer and pp partition by hand in kv transfer (#4892)"" (#4892)",4892,"Revert ""[Bugfix] support mtp kv transfer and pp partition by hand in kv transfer",Ignore,Ignore,Internal/CI/test change +[perf] replace all_reduce for kv_consumer and support different num_tokens among all ranks (#4983),4983,[perf] replace all_reduce for kv_consumer and support different num_tokens among all ranks,Performance,Include,Performance optimization +[CI] Pull latest vllm-ascend src before tests (#4988),4988,[CI] Pull latest vllm-ascend src before tests,Ignore,Ignore,Internal/CI/test change +add release note for 0.12.0 (#4995),4995,add release note for 0.12.0,Others,Include,Miscellaneous +[Fix] Fixes issues in MTP with async scheduling and ACL graph (#4963),4963,[Fix] Fixes issues in MTP with async scheduling and ACL graph,Highlights,Include,Major feature +[Perf]enable prefill flashcommon3 (#4065),4065,[Perf]enable prefill flashcommon3,Performance,Include,Performance optimization +[CI] CI refactor (#4928),4928,[CI] CI refactor,Ignore,Ignore,Internal/CI/test change +add ut for model runner (#4991),4991,add ut for model runner,Others,Include,Miscellaneous +[Misc] Update pooling example (#5002),5002,[Misc] Update pooling example,Highlights,Include,Major feature +[CI][Bugfix] Fix scheduleroutput has no attr get error in prompt logprobs (#4998),4998,[CI][Bugfix] Fix scheduleroutput has no attr get error in prompt logprobs,Ignore,Ignore,Internal/CI/test change +Add Qwen3-Next tutorials (#4607),4607,Add Qwen3-Next tutorials,Documentation,Include,Model tutorial +[Refactor]3/N Refactor mla_v1.py & extract mla_cp (#4933),4933,[Refactor]3/N Refactor mla_v1.py & extract mla_cp,Ignore,Ignore,Internal/CI/test change +[main][BugFix] Fixed an accuracy bug of Qwen3-next-MTP when batched inferring (#4932),4932,[main][BugFix] Fixed an accuracy bug of Qwen3-next-MTP when batched inferring,Highlights,Include,Major feature +Bump actions/upload-artifact from 5 to 6 (#5014),5014,Bump actions/upload-artifact from 5 to 6,Ignore,Ignore,Internal/CI/test change +[bugfix] Fix dummy-run and multi-node issues in MoE routing and MTP (#4947),4947,[bugfix] Fix dummy-run and multi-node issues in MoE routing and MTP,Highlights,Include,Major feature +[Doc ] Supplement kvpool user guide (#5013),5013,[Doc ] Supplement kvpool user guide,Documentation,Include,Documentation update +[Test]update accuracy test of models (#4911),4911,[Test]update accuracy test of models,Others,Include,Miscellaneous +"[Bugfix] Fix the bug in initializing the shared_weight communication domain in sfa-cp, and fix the mtp weight load in pp>1 situation (#4913)",4913,"[Bugfix] Fix the bug in initializing the shared_weight communication domain in sfa-cp, and fix the mtp weight load in pp>1 situation",Highlights,Include,Major feature +[Bugfix] Add support for PP intermediate value types in graph mode (#4902),4902,[Bugfix] Add support for PP intermediate value types in graph mode,Features,Include,New feature +[Bugfix] qwen3-vl-235b-w8a8 load weight ERROR when start service (#4292),4292,[Bugfix] qwen3-vl-235b-w8a8 load weight ERROR when start service,Highlights,Include,Model support +update release note for suffix decoding (#5009),5009,update release note for suffix decoding,Highlights,Include,Major feature +[Graph][Fusion] Add AddRMSNorm(with bias) and Quant Fusion Pattern (#5011),5011,[Graph][Fusion] Add AddRMSNorm(with bias) and Quant Fusion Pattern,Performance,Include,Operator fusion +[UT]add pcp dcp ut (#4949),4949,[UT]add pcp dcp ut,Ignore,Ignore,Internal/CI/test change +[Bugfix] fix the incorrect use of python's sum on tensors. (#4655),4655,[Bugfix] fix the incorrect use of python's sum on tensors.,Others,Include,Bug fix +[Misc] Upgrade vllm hash to 12_14 (#5000),5000,[Misc] Upgrade vllm hash to 12_14,Others,Include,Miscellaneous +[CI] Delete deepseek3.2-exp nightly test (#5028),5028,[CI] Delete deepseek3.2-exp nightly test,Others,Include,Miscellaneous +[E2E] Collect test run time. (#5018),5018,[E2E] Collect test run time.,Others,Include,Miscellaneous +[doc]Modify quantization tutorials (#5026),5026,[doc]Modify quantization tutorials,Documentation,Include,Documentation update +[KVPool]Fix PP get bug (#5007),5007,[KVPool]Fix PP get bug,Others,Include,Miscellaneous +[Attention] Temporarily add back pa for small batch sizes. (#4765),4765,[Attention] Temporarily add back pa for small batch sizes.,Others,Include,Miscellaneous +[Cleanup] Remove unused attn_metadata parameter from Proposer classes (#4862),4862,[Cleanup] Remove unused attn_metadata parameter from Proposer classes,Ignore,Ignore,Internal/CI/test change +[bugfix] [main] Fix KV cache query inconsistency across different TP ranks in the KV Pool (#5030),5030,[bugfix] [main] Fix KV cache query inconsistency across different TP ranks in the KV Pool,Highlights,Include,Major feature +[Bugfix] Fix precision issues in moe_mlp (vllm-ascend main) (#5025),5025,[Bugfix] Fix precision issues in moe_mlp (vllm-ascend main),Others,Include,Bug fix +[Bugfix] Fix the attn_metadata is None (#5038),5038,[Bugfix] Fix the attn_metadata is None,Others,Include,Bug fix +[Misc] Upgrade vllm commit hash to 1215 (#5029),5029,[Misc] Upgrade vllm commit hash to 1215,Others,Include,Miscellaneous +[Fix]Revert temporary skip on mtp1/mtp2 correctness tests (aclgraph fix) (#5039),5039,[Fix]Revert temporary skip on mtp1/mtp2 correctness tests (aclgraph fix),Ignore,Ignore,Internal/CI/test change +[Core][Worker] Add UCMConnector for KV Cache Offloading (#4411),4411,[Core][Worker] Add UCMConnector for KV Cache Offloading,Highlights,Include,Major feature +[Bugfix] dynamic eplb does't use fused_alltoall (#4919),4919,[Bugfix] dynamic eplb does't use fused_alltoall,Highlights,Include,Major feature +Bump actions/checkout from 4 to 6 (#5015),5015,Bump actions/checkout from 4 to 6,Ignore,Ignore,Internal/CI/test change +[Feat] Refactor rejection sampler (#4975),4975,[Feat] Refactor rejection sampler,Highlights,Include,Major feature +[bugfix] Fix mooncake kvpool accuracy issue (#4976),4976,[bugfix] Fix mooncake kvpool accuracy issue,Highlights,Include,Major feature +[Refactor] Remove the process patches of Qwen2.5-VL and Qwen2.5-Omni (#5035),5035,[Refactor] Remove the process patches of Qwen2.5-VL and Qwen2.5-Omni,Ignore,Ignore,Internal/CI/test change +[Doc] Upgrade some outdated doc (#5062),5062,[Doc] Upgrade some outdated doc,Documentation,Include,Documentation update +[ModelRunner] apply_grammer uses vllm function (#4974),4974,[ModelRunner] apply_grammer uses vllm function,Others,Include,Miscellaneous +[Bugfix] fix fastapi version (#5047),5047,[Bugfix] fix fastapi version,Others,Include,Bug fix +[BugFix]Fix FIA input err in DSv3.1 (#5059),5059,[BugFix]Fix FIA input err in DSv3.1,Hardware and Operator Support,Include,Operator/hardware support +[Doc] Add user guide of speculative decoding (#5074),5074,[Doc] Add user guide of speculative decoding,Highlights,Include,Major feature +Add release note for v0.11.0 (#4918),4918,Add release note for v0.11.0,Others,Include,Miscellaneous +[bugfix] matmul_allreduce_add_rmsnorm aclnn interface (#5082),5082,[bugfix] matmul_allreduce_add_rmsnorm aclnn interface,Others,Include,Bug fix +【Feature】refactor npu_modelrunner for profile_run (#4993),4993,【Feature】refactor npu_modelrunner for profile_run,Others,Include,Miscellaneous +Add a Mooncake installation tutorial for kv pool and update Mooncake installation tutorial (#5069),5069,Add a Mooncake installation tutorial for kv pool and update Mooncake installation tutorial,Highlights,Include,Major feature +[Bugfix] EPLB nightly deepseek (#5095),5095,[Bugfix] EPLB nightly deepseek,Highlights,Include,Major feature +[Nightly] Upgrade single node test to latest main (#5101),5101,[Nightly] Upgrade single node test to latest main,Others,Include,Miscellaneous +[Nightly][BugFix] Install triton for nightly e2e op test. (#5096),5096,[Nightly][BugFix] Install triton for nightly e2e op test.,Ignore,Ignore,Internal/CI/test change +[Feat] Support async_scheduler and disable_padded_drafter_batch in eagle (#4893),4893,[Feat] Support async_scheduler and disable_padded_drafter_batch in eagle,Highlights,Include,Major feature +[bugfix] fix mtp accept rate (#5093),5093,[bugfix] fix mtp accept rate,Highlights,Include,Major feature +Upgrade vllm commit hash to 1216 (#5053),5053,Upgrade vllm commit hash to 1216,Others,Include,Miscellaneous +[Fusion] [Graph] Add qknorm rope fusion operator (#4711),4711,[Fusion] [Graph] Add qknorm rope fusion operator,Performance,Include,Operator fusion +[UT]add the UT of pcp and dcp in the attention_cp file (#5054),5054,[UT]add the UT of pcp and dcp in the attention_cp file,Ignore,Ignore,Internal/CI/test change +[Bugfix] Fix DeepSeek FIA error in async_scheduling with mtp (#5046),5046,[Bugfix] Fix DeepSeek FIA error in async_scheduling with mtp,Highlights,Include,Major feature +[feat]pd disaggregated support cross-machine (#5008),5008,[feat]pd disaggregated support cross-machine,Features,Include,New feature +[CI] Fix UT (#5106),5106,[CI] Fix UT,Ignore,Ignore,Internal/CI/test change +[main] rename device type (#5099),5099,[main] rename device type,Others,Include,Miscellaneous +[main][doc] Instructions for using permissions added to docker (#5092),5092,[main][doc] Instructions for using permissions added to docker,Documentation,Include,Documentation update +[Pangu][MoE] Remove PanguProMoEV1 related code (#5088),5088,[Pangu][MoE] Remove PanguProMoEV1 related code,Highlights,Include,Model support +[model] Support PanguUltraMoE (#4615),4615,[model] Support PanguUltraMoE,Highlights,Include,Model support +[UT] add pcp&dcp UT for mla_cp (#4953),4953,[UT] add pcp&dcp UT for mla_cp,Ignore,Ignore,Internal/CI/test change +[BugFix] Fix mooncake bug in PCP scenario (#5055),5055,[BugFix] Fix mooncake bug in PCP scenario,Highlights,Include,Major feature +[Bugfix][MoE] Remove All2All in w4a8_dynamic (#4977),4977,[Bugfix][MoE] Remove All2All in w4a8_dynamic,Others,Include,Bug fix +Fix a data conversion bug introduced by commit 3b7eb51 in main#4655 (#5115),4655,Fix a data conversion bug introduced by commit 3b7eb51 in main#4655,Others,Include,Miscellaneous +[Refactor] 4/N Distinguish the branches based on the applicable scenarios of PA and FIA Ops. (#5081),5081,[Refactor] 4/N Distinguish the branches based on the applicable scenarios of PA and FIA Ops.,Ignore,Ignore,Internal/CI/test change +[Bugfix]delele profile_run in model_runner (#5122),5122,[Bugfix]delele profile_run in model_runner,Others,Include,Bug fix +[Fix] Synchronize the host query_start_loc with device values to prevent shape mismatches (#5134),5134,[Fix] Synchronize the host query_start_loc with device values to prevent shape mismatches,Others,Include,Bug fix +fix profile run for vl model (#5136),5136,fix profile run for vl model,Others,Include,Miscellaneous +enable npugraph_ex (#5120),5120,enable npugraph_ex,Highlights,Include,Major feature +[Doc] add qwen3 reranker (#5086),5086,[Doc] add qwen3 reranker,Documentation,Include,Documentation update +[UT] Add model_runner pcp related UTs (#4951),4951,[UT] Add model_runner pcp related UTs,Ignore,Ignore,Internal/CI/test change +qwen3_next add triton ops : fused_qkvzba_split_reshape (#4788),4788,qwen3_next add triton ops : fused_qkvzba_split_reshape,Performance,Include,Operator fusion +[test] add w4a8 accuracy case (#5110),5110,[test] add w4a8 accuracy case,Others,Include,Miscellaneous +[feat] proxy support elastic scaling (#5063),5063,[feat] proxy support elastic scaling,Features,Include,New feature +"[Fix] Fix DeepSeek V3.2 ""no attr"" error (#5147)",5147,"[Fix] Fix DeepSeek V3.2 ""no attr"" error",Highlights,Include,Model support +[UT]Ut for function cumsum_group_list in moe_mlp (ref #5025) (#5036),5025,[UT]Ut for function cumsum_group_list in moe_mlp (ref #5025),Ignore,Ignore,Internal/CI/test change +[UT] Add mooncake ut test (#5080),5080,[UT] Add mooncake ut test,Ignore,Ignore,Internal/CI/test change +fixed fused alltoall execute all reduce (#5109),5109,fixed fused alltoall execute all reduce,Performance,Include,Operator fusion +Qwen3-Next：Update the gpu-memory-utilization parameter to 0.7 (#5129),5129,Qwen3-Next：Update the gpu-memory-utilization parameter to 0.7,Highlights,Include,Model support +[Bugfix] fix pipeline parallelism bug introduced by async-scheduling refactor work (#4973),4973,[Bugfix] fix pipeline parallelism bug introduced by async-scheduling refactor work,Performance,Include,Performance optimization +implement model runner v2 basic framework (#5051),5051,implement model runner v2 basic framework,Features,Include,New feature +fix: use batch_matmul_transpose operator in MLA _v_up_proj for better performance (#5142),5142,fix: use batch_matmul_transpose operator in MLA _v_up_proj for better performance,Hardware and Operator Support,Include,Operator/hardware support +Nominate new maintainers @zzzzwwjj @realliujiaxu @LCAIZJ (#5152),5152,Nominate new maintainers @zzzzwwjj @realliujiaxu @LCAIZJ,Others,Include,Miscellaneous +feat: implement high-performance Triton kernels for rejection sampling (#4830),4830,feat: implement high-performance Triton kernels for rejection sampling,Hardware and Operator Support,Include,Operator/hardware support +"[Feat] Support MLP_TP feature, exclude MOE layer (#4999)",4999,"[Feat] Support MLP_TP feature, exclude MOE layer",Features,Include,New feature +[Graph][Fusion]Add new pattern for AddRmsnormQuant with SP. (#5077),5077,[Graph][Fusion]Add new pattern for AddRmsnormQuant with SP.,Performance,Include,Operator fusion +[Fix] Refines decode mode padding condition for uniform queries (#5164),5164,[Fix] Refines decode mode padding condition for uniform queries,Others,Include,Bug fix +fix vl pd smoke error (#5103),5103,fix vl pd smoke error,Others,Include,Miscellaneous +[BugFix]Fix incorrect get_current_vllm_config (#5121),5121,[BugFix]Fix incorrect get_current_vllm_config,Others,Include,Bug fix +[Nightly] Avoid max_model_len being smaller than the decoder prompt to prevent single-node-accuray-tests from failing (#5174),5174,[Nightly] Avoid max_model_len being smaller than the decoder prompt to prevent single-node-accuray-tests from failing,Others,Include,Miscellaneous +[Doc] Refact benchmark doc (#5173),5173,[Doc] Refact benchmark doc,Documentation,Include,Documentation update +[Bugfix] Fix in_profile_run in mtp_proposer dummy_run (#5165),5165,[Bugfix] Fix in_profile_run in mtp_proposer dummy_run,Highlights,Include,Major feature +[Doc][P/D] Fix MooncakeConnector's name (#5172),5172,[Doc][P/D] Fix MooncakeConnector's name,Highlights,Include,Major feature +"[BugFix] Fix top_p,top_k issue with EAGLE and add top_p,top_k in EAGLE e2e (#5131)",5131,"[BugFix] Fix top_p,top_k issue with EAGLE and add top_p,top_k in EAGLE e2e",Highlights,Include,Major feature +[bugfix] Use FUSED_MC2 MoE comm path for the op `dispatch_ffn_combine` (#5156),5156,[bugfix] Use FUSED_MC2 MoE comm path for the op `dispatch_ffn_combine`,Performance,Include,Operator fusion +[2/N][Pangu][MoE] Remove Pangu Related Code (#5130),5130,[2/N][Pangu][MoE] Remove Pangu Related Code,Highlights,Include,Model support +[Bugfix] install trition for test_custom_op (#5112),5112,[Bugfix] install trition for test_custom_op,Ignore,Ignore,Internal/CI/test change +support basic long_seq feature st (#5140),5140,support basic long_seq feature st,Features,Include,New feature +【Doc】Deepseekv3.1/R1 doc enhancement (#4827),4827,【Doc】Deepseekv3.1/R1 doc enhancement,Others,Include,Miscellaneous +[BugFix]Fix precision issue for LoRA feature (#4141),4141,[BugFix]Fix precision issue for LoRA feature,Others,Include,Bug fix +[refactor] refactor weight trans nz and transpose (#4878),4878,[refactor] refactor weight trans nz and transpose,Ignore,Ignore,Internal/CI/test change +[Image] Refactor image build (#5175),5175,[Image] Refactor image build,Others,Include,Miscellaneous +[Doc] Add a perf tune section (#5127),5127,[Doc] Add a perf tune section,Documentation,Include,Documentation update +Add Qwen3-VL-235B-A22B-Instruct tutorials (#5167),5167,Add Qwen3-VL-235B-A22B-Instruct tutorials,Documentation,Include,Model tutorial +[Refactor] remove some metadata variables in attention_v1. (#5160),5160,[Refactor] remove some metadata variables in attention_v1.,Ignore,Ignore,Internal/CI/test change +[CI] Improve CI (#5078),5078,[CI] Improve CI,Ignore,Ignore,Internal/CI/test change +[Feature] Add token mask for DispatchGmmCombineDecode operator (#5171),5171,[Feature] Add token mask for DispatchGmmCombineDecode operator,Hardware and Operator Support,Include,Operator/hardware support +[pref] qwen3_next add triton ops : fused_sigmoid_gating_delta_rule_update (#4818),4818,[pref] qwen3_next add triton ops : fused_sigmoid_gating_delta_rule_update,Performance,Include,Operator fusion +[Doc]Add the user_guide doc file regarding fine-grained TP. (#5084),5084,[Doc]Add the user_guide doc file regarding fine-grained TP.,Documentation,Include,Documentation update +restore matmul_allreduce_add_rmsnrom aclnn interface (#5119),5119,restore matmul_allreduce_add_rmsnrom aclnn interface,Others,Include,Miscellaneous +[CI] Fix image merge bug (#5197),5197,[CI] Fix image merge bug,Ignore,Ignore,Internal/CI/test change +[CI] Use offline mode for nightly test (#5187),5187,[CI] Use offline mode for nightly test,Others,Include,Miscellaneous +Drop 0.12.0 support (#5146),5146,Drop 0.12.0 support,Features,Include,New feature +[CI] unblock CI on suffix spec decoding (#4813),4813,[CI] unblock CI on suffix spec decoding,Ignore,Ignore,Internal/CI/test change +[e2e] add pcp e2e (#5141),5141,[e2e] add pcp e2e,Highlights,Include,Major feature +[CI] fix lint (#5216),5216,[CI] fix lint,Ignore,Ignore,Internal/CI/test change +[lint]clean code (#5218),5218,[lint]clean code,Ignore,Ignore,Internal/CI/test change +[Fix] Delete pooling redundant code (#4940),4940,[Fix] Delete pooling redundant code,Highlights,Include,Major feature +[Performance] Add async exponential while model executing (#4501),4501,[Performance] Add async exponential while model executing,Performance,Include,Performance optimization +"[BugFix]Fix wrong _cos, _sin instantiation (#5154)",5154,"[BugFix]Fix wrong _cos, _sin instantiation",Others,Include,Bug fix +[Feature]Use DispatchGmmCombineDecode operator to replace MC2(Optional) (#5040),5040,[Feature]Use DispatchGmmCombineDecode operator to replace MC2(Optional),Hardware and Operator Support,Include,Operator/hardware support +[Perf] vectorize PCP/DCP loops in attention_cp.py (#4944),4944,[Perf] vectorize PCP/DCP loops in attention_cp.py,Highlights,Include,Major feature +[Perf] vectorize PCP/DCP loops in mla_v1.py (#5003),5003,[Perf] vectorize PCP/DCP loops in mla_v1.py,Highlights,Include,Major feature +[Misc] Cleanup useless print and logger (#5220),5220,[Misc] Cleanup useless print and logger,Ignore,Ignore,Internal/CI/test change +[Doc] Fix DeepSeek-V3.2 tutorial. (#5190),5190,[Doc] Fix DeepSeek-V3.2 tutorial.,Documentation,Include,Model tutorial +[bugfix][ACLGraph][MTP] deletes `cudagraph_batch_sizes` in `MtpProposer` (#5183),5183,[bugfix][ACLGraph][MTP] deletes `cudagraph_batch_sizes` in `MtpProposer`,Highlights,Include,Major feature +[task] Add fused gdn gating triton kernel (#4304),4304,[task] Add fused gdn gating triton kernel,Performance,Include,Operator fusion +[CustomOp] Register AscendMMEncoderAttention CustomOp and remove related patch (#4750),4750,[CustomOp] Register AscendMMEncoderAttention CustomOp and remove related patch,Others,Include,Miscellaneous +[misc][FlashComm1][ACLGraph] Incompatibility between Flashcomm1 and FULL_DECODE_ONLY. (#5200),5200,[misc][FlashComm1][ACLGraph] Incompatibility between Flashcomm1 and FULL_DECODE_ONLY.,Highlights,Include,Major feature +Bump actions/upload-artifact from 4 to 6 (#5233),5233,Bump actions/upload-artifact from 4 to 6,Ignore,Ignore,Internal/CI/test change +Bump actions/checkout from 4 to 6 (#5234),5234,Bump actions/checkout from 4 to 6,Ignore,Ignore,Internal/CI/test change +[Doc] Update readme (#5226),5226,[Doc] Update readme,Documentation,Include,Documentation update +[1/N][Eagle3] Aligns auxiliary hidden state usage for eagle3 models (#5162),5162,[1/N][Eagle3] Aligns auxiliary hidden state usage for eagle3 models,Highlights,Include,Major feature +[Triton]support swiglu_quant triton in w4a8 (#5161),5161,[Triton]support swiglu_quant triton in w4a8,Hardware and Operator Support,Include,Operator/hardware support +[feature] support pcp + mtp in full graph (#4572),4572,[feature] support pcp + mtp in full graph,Highlights,Include,Major feature +[Feat]Xlite Qwen3-vl Support (#5228),5228,[Feat]Xlite Qwen3-vl Support,Highlights,Include,Major feature +[bugfix] fix w8a8dynamic fused_moe trans nz (#5199),5199,[bugfix] fix w8a8dynamic fused_moe trans nz,Performance,Include,Operator fusion +[Bugfix] Implement multimodal_cpu_fields in model runner (#5196),5196,[Bugfix] Implement multimodal_cpu_fields in model runner,Features,Include,New feature +[TEST]Update mm param --mm-processor-cache-gb (#5242),5242,[TEST]Update mm param --mm-processor-cache-gb,Others,Include,Miscellaneous +[Bugfix] Use hf_text_config instead of hf_config to support multimodal PD-Disaggregated (#5205),5205,[Bugfix] Use hf_text_config instead of hf_config to support multimodal PD-Disaggregated,Features,Include,New feature +[Refactor] move the metadata from attention_v1 to util(ready for extract common_cp) & realize Ascendmetadata inherit from the parent class. (#5203),5203,[Refactor] move the metadata from attention_v1 to util(ready for extract common_cp) & realize Ascendmetadata inherit from the parent class.,Ignore,Ignore,Internal/CI/test change +[refactor] Remove unnecessary attributes from set_ascend_forward_context (#5204),5204,[refactor] Remove unnecessary attributes from set_ascend_forward_context,Ignore,Ignore,Internal/CI/test change +[Doc] Update the weight download URL. (#5238),5238,[Doc] Update the weight download URL.,Documentation,Include,Documentation update +[Main] [Patch] support balance scheduling patch (#5212),5212,[Main] [Patch] support balance scheduling patch,Features,Include,New feature +[Doc] Add new contributors and relative scripts. (#5070),5070,[Doc] Add new contributors and relative scripts.,Documentation,Include,Documentation update +[CustomOp] Register AscendApplyRotaryEmb CustomOp and remove related patch (#4667),4667,[CustomOp] Register AscendApplyRotaryEmb CustomOp and remove related patch,Others,Include,Miscellaneous +[Doc] fix docs set rope_theta value is 10e6 in qwen3-235b model (#5258),5258,[Doc] fix docs set rope_theta value is 10e6 in qwen3-235b model,Documentation,Include,Documentation update +[ModelRunner] Add hunyuan-vl basic support (#5151),5151,[ModelRunner] Add hunyuan-vl basic support,Highlights,Include,Model support +[KV-Sharing] Support KV-Sharing feature in CLA models (#4138),4138,[KV-Sharing] Support KV-Sharing feature in CLA models,Highlights,Include,Major feature +[EPLB][CI] Add dynamic EPLB CI for qwen3-moe (#5179),5179,[EPLB][CI] Add dynamic EPLB CI for qwen3-moe,Ignore,Ignore,Internal/CI/test change +[CI] Add Triton Ascend in CI (#4921),4921,[CI] Add Triton Ascend in CI,Ignore,Ignore,Internal/CI/test change +[CI]refactor: standardize test case naming convention (#5243),5243,[CI]refactor: standardize test case naming convention,Ignore,Ignore,Internal/CI/test change +[test]Corrected the Qwen3-Omni-30B-A3B-Instruct accuracy test configuration in nightly tests. (#5195),5195,[test]Corrected the Qwen3-Omni-30B-A3B-Instruct accuracy test configuration in nightly tests.,Hardware and Operator Support,Include,Operator/hardware support +[main][Refactor] Remove `with_prefill` parameter from `set_ascend_forward_context` (#5094),5094,[main][Refactor] Remove `with_prefill` parameter from `set_ascend_forward_context`,Ignore,Ignore,Internal/CI/test change +[Doc] Added deploying on k8s with kthena (#4674),4674,[Doc] Added deploying on k8s with kthena,Documentation,Include,Documentation update +[CI] Mock spawn for vlm tests (#5279),5279,[CI] Mock spawn for vlm tests,Ignore,Ignore,Internal/CI/test change +[CI] refect e2e ci test (#5246),5246,[CI] refect e2e ci test,Ignore,Ignore,Internal/CI/test change +[Refactor][MoE] Reuse vLLM's all_reduce logic (#5189),5189,[Refactor][MoE] Reuse vLLM's all_reduce logic,Ignore,Ignore,Internal/CI/test change +[Bugfix] quick fix balance scheduling patch (#5281),5281,[Bugfix] quick fix balance scheduling patch,Others,Include,Bug fix +update to vllm 12-19 (#5223),5223,update to vllm 12-19,Others,Include,Miscellaneous +fix transformer version to 4.57.3 (#5250),5250,fix transformer version to 4.57.3,Others,Include,Miscellaneous +[Refactor]5/N Extract common code of mla_v1.py & extract mla_cp (#5097),5097,[Refactor]5/N Extract common code of mla_v1.py & extract mla_cp,Ignore,Ignore,Internal/CI/test change +[CI] Add skipped testcases. (#5254),5254,[CI] Add skipped testcases.,Ignore,Ignore,Internal/CI/test change +[E2E] Optimize e2e test. (#5091),5091,[E2E] Optimize e2e test.,Performance,Include,Performance optimization +[bugfix] remove the EP buffer allocation introduced by fused-op dispatch_ffn_c… (#5284),5284,[bugfix] remove the EP buffer allocation introduced by fused-op dispatch_ffn_c…,Performance,Include,Operator fusion +[Doc] Add pa_shape_list description to qwen dense tutorial (#5225),5225,[Doc] Add pa_shape_list description to qwen dense tutorial,Documentation,Include,Documentation update +Update vllm pin to 12.24 (#5307),5307,Update vllm pin to 12.24,Others,Include,Miscellaneous +[CI] Skip some failed ops tests (#5309),5309,[CI] Skip some failed ops tests,Ignore,Ignore,Internal/CI/test change +[perf][bugfix] improve performance of rejection sampler and eliminate HD synchronize in TopKTopPSampler (#4154),4154,[perf][bugfix] improve performance of rejection sampler and eliminate HD synchronize in TopKTopPSampler,Highlights,Include,Major feature +[quantization] Add w8a16 quantization support (#4541),4541,[quantization] Add w8a16 quantization support,Features,Include,New feature +Cleanup uesless env (#5270),5270,Cleanup uesless env,Ignore,Ignore,Internal/CI/test change +Revert [KV-Sharing] Support KV-Sharing feature in CLA models (#4138) (#4138),4138,Revert [KV-Sharing] Support KV-Sharing feature in CLA models,Ignore,Ignore,Internal/CI/test change +[Kernel] add l2norm triton kernel (#4595),4595,[Kernel] add l2norm triton kernel,Hardware and Operator Support,Include,Operator/hardware support +Add MagicMTP(block verify) and Triton optimization (#4443),4443,Add MagicMTP(block verify) and Triton optimization,Highlights,Include,Major feature +[CI] add xlite e2e test (#5305),5305,[CI] add xlite e2e test,Ignore,Ignore,Internal/CI/test change +[E2E Refactor] Enable skipped e2e case (#5287),5287,[E2E Refactor] Enable skipped e2e case,Features,Include,New feature +[BugFix] Fix num_pcp_pads Assignment Issues (#5273),5273,[BugFix] Fix num_pcp_pads Assignment Issues,Highlights,Include,Major feature +[bugfix] fix Error 'ValueError: Duplicate layer name' (#5280),5280,[bugfix] fix Error 'ValueError: Duplicate layer name',Others,Include,Bug fix +Remove VLLM_ASCEND_ENABLE_DENSE_OPTIMIZE (#5272),5272,Remove VLLM_ASCEND_ENABLE_DENSE_OPTIMIZE,Performance,Include,Performance optimization +fix e2e rejection-sampler error (#5341),5341,fix e2e rejection-sampler error,Others,Include,Miscellaneous +[Bugfix] fix pcp 128K break (#5266),5266,[Bugfix] fix pcp 128K break,Highlights,Include,Major feature +[Bugfix] fix xlite decode-only e2e test (#5354),5354,[Bugfix] fix xlite decode-only e2e test,Ignore,Ignore,Internal/CI/test change +[doc] update using command (#5373),5373,[doc] update using command,Documentation,Include,Documentation update +[Bugfix] Fix Qwen P/D Disaggregation accuracy issue (#5340),5340,[Bugfix] Fix Qwen P/D Disaggregation accuracy issue,Others,Include,Bug fix +[CI] Skip failed test cases to recover CI (#5368),5368,[CI] Skip failed test cases to recover CI,Ignore,Ignore,Internal/CI/test change +[FIX] Update _causal_conv1d_update_kernel for Efficient Conv State Handling on NPU (#5322),5322,[FIX] Update _causal_conv1d_update_kernel for Efficient Conv State Handling on NPU,Hardware and Operator Support,Include,Operator/hardware support +[BugFix][Fusion] Patch compile backend to make fusion available (#5308),5308,[BugFix][Fusion] Patch compile backend to make fusion available,Performance,Include,Operator fusion +move contiguous in fused_sigmoid_gating_delta_rule_update to model_runner_v1 (#5274),5274,move contiguous in fused_sigmoid_gating_delta_rule_update to model_runner_v1,Performance,Include,Operator fusion +[Nightly] Initial logging for nightly multi-node testing (#5362),5362,[Nightly] Initial logging for nightly multi-node testing,Others,Include,Miscellaneous +Update vllm pin to 12.25 (#5342),5342,Update vllm pin to 12.25,Others,Include,Miscellaneous +cleanup ascend config (#5296),5296,cleanup ascend config,Ignore,Ignore,Internal/CI/test change +[E2E] Optimize the E2E test time. (#5294),5294,[E2E] Optimize the E2E test time.,Performance,Include,Performance optimization +"Revert ""Add MagicMTP(block verify) and Triton optimization (#4443)"" (#4443)",4443,"Revert ""Add MagicMTP(block verify) and Triton optimization",Ignore,Ignore,Internal/CI/test change +[doc] add developer guide for PCP&DCP (#5372),5372,[doc] add developer guide for PCP&DCP,Highlights,Include,Major feature +[Bugfix] Fix unsuitable moe_comm_type under ep=1 scenario (#5388),5388,[Bugfix] Fix unsuitable moe_comm_type under ep=1 scenario,Others,Include,Bug fix +[doc] Add context parallel user guide (#5358),5358,[doc] Add context parallel user guide,Highlights,Include,Major feature +[Doc] update R1/V3.1 doc (#5383),5383,[Doc] update R1/V3.1 doc,Documentation,Include,Documentation update +[Feature] Enhance all-reduce skipping logic for MoE models in NPUModelRunner (#5329),5329,[Feature] Enhance all-reduce skipping logic for MoE models in NPUModelRunner,Features,Include,New feature +[TEST]Add sending request with and without chat (#5286),5286,[TEST]Add sending request with and without chat,Ignore,Ignore,Internal/CI/test change +rollback causal_conv1d_fn to torch ops & update qwen3Next doc (#5391),5391,rollback causal_conv1d_fn to torch ops & update qwen3Next doc,Others,Include,Miscellaneous +[bugfix] Fix MHA model runtime error in aclgraph mode (#5397),5397,[bugfix] Fix MHA model runtime error in aclgraph mode,Highlights,Include,Major feature +[Feature] Remove the transpose step after attention and switch to transpose_batchmatmul (#5390),5390,[Feature] Remove the transpose step after attention and switch to transpose_batchmatmul,Features,Include,New feature +Update vllm pin to 12.26 (#5378),5378,Update vllm pin to 12.26,Others,Include,Miscellaneous +[CI] Add qwen-235b-a22b a2 multi-node test (#5393),5393,[CI] Add qwen-235b-a22b a2 multi-node test,Ignore,Ignore,Internal/CI/test change +[Build] Add installation script of fused_infer_attention_score kernel with flash decoding (#5402),5402,[Build] Add installation script of fused_infer_attention_score kernel with flash decoding,Performance,Include,Operator fusion +[Test] Add acceptance test for eagle/eagle3 (#5366),5366,[Test] Add acceptance test for eagle/eagle3,Highlights,Include,Major feature +[TEST]Add vllm bench (#5306),5306,[TEST]Add vllm bench,Others,Include,Miscellaneous +MLA prefill preformance optimization (#5275),5275,MLA prefill preformance optimization,Performance,Include,Performance optimization +"Revert ""MLA prefill preformance optimization (#5275)"" (#5275)",5275,"Revert ""MLA prefill preformance optimization",Ignore,Ignore,Internal/CI/test change +[Doc]add long sequence tutorials (#5364),5364,[Doc]add long sequence tutorials,Documentation,Include,Documentation update +[bugfix][main]KV Pool for KV Transfer in PD Disaggregation Scenarios (#5398),5398,[bugfix][main]KV Pool for KV Transfer in PD Disaggregation Scenarios,Highlights,Include,Major feature +[BugFix] Fix npu-cpu offloading interface change bug. (#5290),5290,[BugFix] Fix npu-cpu offloading interface change bug.,Others,Include,Bug fix +[Doc] modify pcp tutorials (#5411),5411,[Doc] modify pcp tutorials,Highlights,Include,Major feature +[bugfix] solve dp scenario Host-Device sync (#5298),5298,[bugfix] solve dp scenario Host-Device sync,Others,Include,Bug fix +[Doc] add long_sequence feature user guide (#5343),5343,[Doc] add long_sequence feature user guide,Documentation,Include,Documentation update +[Doc] delete environment variable HCCL_OP_EXPANSION_MODE in DeepSeekV3.1/R1 (#5419),5419,[Doc] delete environment variable HCCL_OP_EXPANSION_MODE in DeepSeekV3.1/R1,Documentation,Include,Documentation update +[feat] enable hierarchical mc2 ops on A2 by default (#5300),5300,[feat] enable hierarchical mc2 ops on A2 by default,Hardware and Operator Support,Include,Operator/hardware support +[doc] Update Qwen3-235B doc for reproducing latest performance (#5323),5323,[doc] Update Qwen3-235B doc for reproducing latest performance,Performance,Include,Performance optimization +[Bugfix] fix greedy temperature detection (#5417),5417,[Bugfix] fix greedy temperature detection,Others,Include,Bug fix +"Revert ""[feat] enable hierarchical mc2 ops on A2 by default (#5300)"" (#5300)",5300,"Revert ""[feat] enable hierarchical mc2 ops on A2 by default",Ignore,Ignore,Internal/CI/test change +[Doc] Modify DeepSeek-R1/V3.1 documentation (#5426),5426,[Doc] Modify DeepSeek-R1/V3.1 documentation,Documentation,Include,Model tutorial +[DOC]Fix model weight download links (#5436),5436,[DOC]Fix model weight download links,Documentation,Include,Documentation update +[Doc] Update DeepSeek V3.1/R1 2P1D doc (#5387),5387,[Doc] Update DeepSeek V3.1/R1 2P1D doc,Documentation,Include,Documentation update +[Misc] fast fail for exiting if tools/install_flash_infer_attention_score_ops_a2.sh (#5422),5422,[Misc] fast fail for exiting if tools/install_flash_infer_attention_score_ops_a2.sh,Hardware and Operator Support,Include,Operator/hardware support +[Doc]modify pcp tutorial doc (#5440),5440,[Doc]modify pcp tutorial doc,Highlights,Include,Major feature +[bugfix] fix typo of _skip_all_reduce_across_dp_group (#5435),5435,[bugfix] fix typo of _skip_all_reduce_across_dp_group,Others,Include,Bug fix +Fix nightly (#5413),5413,Fix nightly,Others,Include,Miscellaneous +[Bugfix] Correctly handle the output shape in multimodal attention (#5443),5443,[Bugfix] Correctly handle the output shape in multimodal attention,Others,Include,Bug fix +[ReleaseNote] Add release note for v0.13.0rc1 (#5334),5334,[ReleaseNote] Add release note for v0.13.0rc1,Others,Include,Miscellaneous +update vllm pin to 12.27 (#5412),5412,update vllm pin to 12.27,Others,Include,Miscellaneous +[Refactor] cache cos/sin in mla & remove parameter model in builder. (#5277),5277,[Refactor] cache cos/sin in mla & remove parameter model in builder.,Ignore,Ignore,Internal/CI/test change +[Refactor]6/N Extract common code of class AscendMLAImpl (#5314),5314,[Refactor]6/N Extract common code of class AscendMLAImpl,Ignore,Ignore,Internal/CI/test change +[EPLB][refactor] Modification of the initialization logic for expert_map and log2phy（depend on pr5285） (#5311),5311,[EPLB][refactor] Modification of the initialization logic for expert_map and log2phy（depend on pr5285）,Ignore,Ignore,Internal/CI/test change +[Feature] Support to use fullgraph with eagle (#5118),5118,[Feature] Support to use fullgraph with eagle,Highlights,Include,Major feature +Optimize some rejectsampler functions to make npu op launch non-blocking (#4587),4587,Optimize some rejectsampler functions to make npu op launch non-blocking,Performance,Include,Performance optimization +[feature] fia support sliding windows (#5239),5239,[feature] fia support sliding windows,Hardware and Operator Support,Include,Operator/hardware support +[Feature] support eager mode in model runner v2 (#5210),5210,[Feature] support eager mode in model runner v2,Features,Include,New feature +[Refactor][Triton] Move reject sample triton kernels into ops/triton (#5324),5324,[Refactor][Triton] Move reject sample triton kernels into ops/triton,Ignore,Ignore,Internal/CI/test change +[Refactor][EAGLE] 1/N delete __init__ in mtp_proposer (#5176),5176,[Refactor][EAGLE] 1/N delete __init__ in mtp_proposer,Ignore,Ignore,Internal/CI/test change +[OP] add custom op aclnnMoeInitRoutingCustom (#5251),5251,[OP] add custom op aclnnMoeInitRoutingCustom,Hardware and Operator Support,Include,Operator/hardware support +[Kernel]update csrc cmakelist for open-source cann (#5458),5458,[Kernel]update csrc cmakelist for open-source cann,Hardware and Operator Support,Include,Operator/hardware support +Update corresponding vllm commit ID to 12 29 (#5475),5475,Update corresponding vllm commit ID to 12 29,Others,Include,Miscellaneous +[refactor] refactor model runner capture model (#5230),5230,[refactor] refactor model runner capture model,Ignore,Ignore,Internal/CI/test change +moe_gating_top_k (#5271),5271,moe_gating_top_k,Others,Include,Miscellaneous +[CI]update triton ascend version (#5392),5392,[CI]update triton ascend version,Ignore,Ignore,Internal/CI/test change +[Doc] Fix issue link for 0.12.0 (#5500),5500,[Doc] Fix issue link for 0.12.0,Documentation,Include,Documentation update +"Revert ""moe_gating_top_k"" (#5512)",5512,"Revert ""moe_gating_top_k""",Ignore,Ignore,Internal/CI/test change +Docs: Remove deprecated --task parameter for embedding models (#5257),5257,Docs: Remove deprecated --task parameter for embedding models,Deprecation & Breaking Changes,Include,Breaking change +[1/N] Refactor nightly test structure (#5479),5479,[1/N] Refactor nightly test structure,Others,Include,Miscellaneous +[3/N][Nightly] Move ops tests to nightly (#5538),5538,[3/N][Nightly] Move ops tests to nightly,Others,Include,Miscellaneous +[Doc] Add new contributors. (#5537),5537,[Doc] Add new contributors.,Documentation,Include,Documentation update +[2/N] Upgrade nightly doc (#5534),5534,[2/N] Upgrade nightly doc,Others,Include,Miscellaneous +[smoke][bugfix] moe_init_routing_v2 active_expert_range use int type (#5521),5521,[smoke][bugfix] moe_init_routing_v2 active_expert_range use int type,Others,Include,Bug fix +[main][test] Refactor the mtp and eagle test case (#5326),5326,[main][test] Refactor the mtp and eagle test case,Highlights,Include,Major feature +[Feature] Refactor PCP &DCP related code (#5214),5214,[Feature] Refactor PCP &DCP related code,Highlights,Include,Major feature +[Main2Main] Upgrade vllm commit to 1230 (#5495),5495,[Main2Main] Upgrade vllm commit to 1230,Others,Include,Miscellaneous +[Bugfix] Fix mm_merge (#5249),5249,[Bugfix] Fix mm_merge,Others,Include,Bug fix +[feature] mooncake support pcp/dcp in common conditions (#5224),5224,[feature] mooncake support pcp/dcp in common conditions,Highlights,Include,Major feature +[Feature] Support kv nz feature for DeepSeek decode node in disagg-prefill scenario (#3072),3072,[Feature] Support kv nz feature for DeepSeek decode node in disagg-prefill scenario,Highlights,Include,Major feature +[Refactor] Formatting output types related to FuseMoE (#5481),5481,[Refactor] Formatting output types related to FuseMoE,Ignore,Ignore,Internal/CI/test change +[P/D] Improve the performance of Layerwise Connector (#5303),5303,[P/D] Improve the performance of Layerwise Connector,Performance,Include,Performance optimization +[Bugfix] fix the precision issues that may raise from the inter-layer reuse of the workspace in certain scenarios (#5522),5522,[Bugfix] fix the precision issues that may raise from the inter-layer reuse of the workspace in certain scenarios,Others,Include,Bug fix +[Model] Add LongCat-Flash (#3833),3833,[Model] Add LongCat-Flash,Highlights,Include,Model support +[Graph][Fusion] Add AddRMSNorm(with bias) (#5491),5491,[Graph][Fusion] Add AddRMSNorm(with bias),Performance,Include,Operator fusion +[P/D] Bugfix zmq send/receive failed (#5503),5503,[P/D] Bugfix zmq send/receive failed,Others,Include,Miscellaneous +[Nightly] Trigger image build for nightly (#5547),5547,[Nightly] Trigger image build for nightly,Others,Include,Miscellaneous +Bump actions/upload-artifact from 4 to 6 (#5466),5466,Bump actions/upload-artifact from 4 to 6,Ignore,Ignore,Internal/CI/test change +Bump actions/download-artifact from 4 to 7 (#5465),5465,Bump actions/download-artifact from 4 to 7,Ignore,Ignore,Internal/CI/test change +[CI] Add multi-nodes longseq configs of DeepSeek-R1-W8A8 & Qwen3-235B-W8A8 (#5381),5381,[CI] Add multi-nodes longseq configs of DeepSeek-R1-W8A8 & Qwen3-235B-W8A8,Ignore,Ignore,Internal/CI/test change +Cleanup pass config override (#5283),5283,Cleanup pass config override,Ignore,Ignore,Internal/CI/test change +[Doc] Fix spelling mistake of environment variable name ASCEND_RT_VISIBLE_DEVICES in Doc (#5570),5570,[Doc] Fix spelling mistake of environment variable name ASCEND_RT_VISIBLE_DEVICES in Doc,Documentation,Include,Documentation update +[Feat][main] Supported to use full-graph with Qwen3-Next-MTP (#5477),5477,[Feat][main] Supported to use full-graph with Qwen3-Next-MTP,Highlights,Include,Major feature +[Feat] enable hierarchical mc2 ops on A2 by default (#5545),5545,[Feat] enable hierarchical mc2 ops on A2 by default,Hardware and Operator Support,Include,Operator/hardware support +[CI] Move longseq Nightly CI (#5577),5577,[CI] Move longseq Nightly CI,Others,Include,Miscellaneous +[Perf][PCP][DCP] add multi-stream for GQA to enable computation-communication overlap (#5382),5382,[Perf][PCP][DCP] add multi-stream for GQA to enable computation-communication overlap,Highlights,Include,Major feature +[Recover] [Bugfix] support mtp kv transfer and pp partition by hand in kv transfer (#4892) (revert in #4981) (#4892),4892,[Recover] [Bugfix] support mtp kv transfer and pp partition by hand in kv transfer,Ignore,Ignore,Internal/CI/test change +[Doc] Fix typo in ASCEND_RT_VISIBLE_DEVICES (#5581),5581,[Doc] Fix typo in ASCEND_RT_VISIBLE_DEVICES,Documentation,Include,Documentation update +[bugfix](pcp) expand max_num_tokens for pcp pad (#5478),5478,[bugfix](pcp) expand max_num_tokens for pcp pad,Highlights,Include,Major feature +"[BugFix]Disable dispatch_gmm_combine_decode operator when mtp drafter model uses non-w8a8 while main model uses w8a8, or drafter model is eagle series (#5293)",5293,"[BugFix]Disable dispatch_gmm_combine_decode operator when mtp drafter model uses non-w8a8 while main model uses w8a8, or drafter model is eagle series",Highlights,Include,Major feature +[KVPOOL]decode save kvcache (#5168),5168,[KVPOOL]decode save kvcache,Others,Include,Miscellaneous +"[refactor](UT,PCP,DCP) refactor pcp&dcp patches in UTs (#5505)",5505,"[refactor](UT,PCP,DCP) refactor pcp&dcp patches in UTs",Ignore,Ignore,Internal/CI/test change +[Doc]modify the quantization user guide and add a quantization adaptation developer guide (#5554),5554,[Doc]modify the quantization user guide and add a quantization adaptation developer guide,Documentation,Include,Documentation update +[bugfix]update bishengir source envs (#5582),5582,[bugfix]update bishengir source envs,Others,Include,Bug fix +[Bugfix] Fix chunk prefill bug for long_sequence feature (#5444),5444,[Bugfix] Fix chunk prefill bug for long_sequence feature,Others,Include,Bug fix +[Bugfix] Fix weight transpose in RL scenarios (#5567),5567,[Bugfix] Fix weight transpose in RL scenarios,Others,Include,Bug fix +[Doc] update supported models (#5379),5379,[Doc] update supported models,Documentation,Include,Documentation update +[CI] skip xlite-decode-only e2e test (#5407),5407,[CI] skip xlite-decode-only e2e test,Ignore,Ignore,Internal/CI/test change +[Doc] eval-type not support service but server (#2920),2920,[Doc] eval-type not support service but server,Documentation,Include,Documentation update +MLA prefill preformance optimization (#5456),5456,MLA prefill preformance optimization,Performance,Include,Performance optimization +[Refactor][EAGLE] 2/N: load model and generate token (#5437),5437,[Refactor][EAGLE] 2/N: load model and generate token,Ignore,Ignore,Internal/CI/test change +[Bugfix] fix pcp + eplb error (#5561),5561,[Bugfix] fix pcp + eplb error,Highlights,Include,Major feature +[Doc] add new doc for mooncake: PD-Colocated cross-node multi-instance validation of Mooncake's KV Cache reuse and performance. (#5415),5415,[Doc] add new doc for mooncake: PD-Colocated cross-node multi-instance validation of Mooncake's KV Cache reuse and performance.,Highlights,Include,Major feature +[BugFix][kernel] fix matmul_allreduce_add_rmsnorm_kernel (#5335),5335,[BugFix][kernel] fix matmul_allreduce_add_rmsnorm_kernel,Hardware and Operator Support,Include,Operator/hardware support +feat: implement high-performance Triton kernels for rejection sampling: optimization for rejection_random_sample_kernel (#5259),5259,feat: implement high-performance Triton kernels for rejection sampling: optimization for rejection_random_sample_kernel,Hardware and Operator Support,Include,Operator/hardware support +[Feat][Spec] Optimize token index calculation in spec decode with Triton kernel (#5356),5356,[Feat][Spec] Optimize token index calculation in spec decode with Triton kernel,Hardware and Operator Support,Include,Operator/hardware support +[Refactor]7/N Extract common code to common_cp (#5490),5490,[Refactor]7/N Extract common code to common_cp,Ignore,Ignore,Internal/CI/test change +[BugFix][Fusion] Fix graph fusion failure problem (#5253),5253,[BugFix][Fusion] Fix graph fusion failure problem,Performance,Include,Operator fusion +Add the requirement of arctic-inference which speculative decoding with suffix_decode (#5045),5045,Add the requirement of arctic-inference which speculative decoding with suffix_decode,Highlights,Include,Major feature +[Doc] Add NNAL installation guide and requirements (#5235),5235,[Doc] Add NNAL installation guide and requirements,Documentation,Include,Documentation update +Docs: Add A3 Docker image guidance for Atlas A3 machines (#5256),5256,Docs: Add A3 Docker image guidance for Atlas A3 machines,Hardware and Operator Support,Include,Operator/hardware support +[CI] Download models from ms (#5405),5405,[CI] Download models from ms,Ignore,Ignore,Internal/CI/test change +[UT]add triton ops ut : test_fused_qkvzba_split_reshape_cat (#5474),5474,[UT]add triton ops ut : test_fused_qkvzba_split_reshape_cat,Ignore,Ignore,Internal/CI/test change +[bugfix] fix test_camem failed with triton-ascend (#5492),5492,[bugfix] fix test_camem failed with triton-ascend,Ignore,Ignore,Internal/CI/test change +[Bugfix] record cos and sin cache in AscendRotaryEmbedding (#5516),5516,[Bugfix] record cos and sin cache in AscendRotaryEmbedding,Others,Include,Bug fix +[P/D]Remove mooncake kvpool unused parameter `local_hostname` (#5574),5574,[P/D]Remove mooncake kvpool unused parameter `local_hostname`,Highlights,Include,Major feature +[CI] update triton-ascend version (#5584),5584,[CI] update triton-ascend version,Ignore,Ignore,Internal/CI/test change +[docs] Correct image about prefill phase of PCP (#5598),5598,[docs] Correct image about prefill phase of PCP,Highlights,Include,Major feature +[perf] Fix MLAPO weight disposal for KV-consumer MLA in PD-mix deploy... (#5192),5192,[perf] Fix MLAPO weight disposal for KV-consumer MLA in PD-mix deploy...,Hardware and Operator Support,Include,Operator/hardware support +[TRITON][TEST]Add nightly test for triton split_qkv_rmsnorm_rope (#5267),5267,[TRITON][TEST]Add nightly test for triton split_qkv_rmsnorm_rope,Hardware and Operator Support,Include,Operator/hardware support +"Revert ""[Feat] enable hierarchical mc2 ops on A2 by default (#5545)"" (#5545)",5545,"Revert ""[Feat] enable hierarchical mc2 ops on A2 by default",Ignore,Ignore,Internal/CI/test change +[BugFix] Fix Smoke Testing Bug for DSR1 longseq (#5613),5613,[BugFix] Fix Smoke Testing Bug for DSR1 longseq,Ignore,Ignore,Internal/CI/test change +[CI] mv ops to correct path (#5615),5615,[CI] mv ops to correct path,Ignore,Ignore,Internal/CI/test change +[Main2Main] Upgrade vllm commit to 0105 (#5595),5595,[Main2Main] Upgrade vllm commit to 0105,Others,Include,Miscellaneous +[UT][PCP&DCP] UT for block_table.py (#5032),5032,[UT][PCP&DCP] UT for block_table.py,Ignore,Ignore,Internal/CI/test change +[CI]update bisheng version (#5621),5621,[CI]update bisheng version,Ignore,Ignore,Internal/CI/test change +[Main2Main] Upgrade vllm commit to 0106 (#5617),5617,[Main2Main] Upgrade vllm commit to 0106,Others,Include,Miscellaneous +[CI] Specify the version of xlite (#5612),5612,[CI] Specify the version of xlite,Ignore,Ignore,Internal/CI/test change +[MM][Bugfix] Update `hf_config` to `hf_text_config` (#5319),5319,[MM][Bugfix] Update `hf_config` to `hf_text_config`,Others,Include,Bug fix +[Refactor][EAGLE] 3/N delete redundant methods in mtp_proposer (#5420),5420,[Refactor][EAGLE] 3/N delete redundant methods in mtp_proposer,Ignore,Ignore,Internal/CI/test change +Bugfix: Align expert map shapes with redundant experts in EPLB adjustment (#5285),5285,Bugfix: Align expert map shapes with redundant experts in EPLB adjustment,Highlights,Include,Major feature +[Bugfix] Remove swa parameter of fia (#5602),5602,[Bugfix] Remove swa parameter of fia,Hardware and Operator Support,Include,Operator/hardware support +[Nightly][Test] Add Qwen3-Next-80B-A3B-Instruct-W8A8 nightly test (#5616),5616,[Nightly][Test] Add Qwen3-Next-80B-A3B-Instruct-W8A8 nightly test,Highlights,Include,Model support +[Misc] Remove useless weight loader patch (#5619),5619,[Misc] Remove useless weight loader patch,Others,Include,Miscellaneous +[P/D] Performance enhancement of Layerwise connector in TP asymmetric scenarios (#5540),5540,[P/D] Performance enhancement of Layerwise connector in TP asymmetric scenarios,Performance,Include,Performance optimization +"Revert ""[BugFix][Fusion] Fix graph fusion failure problem (#5253)"" (#5253)",5253,"Revert ""[BugFix][Fusion] Fix graph fusion failure problem",Ignore,Ignore,Internal/CI/test change +[Bugfix] fix dcp_only bug and add e2e accuracy test for dcp only and pcp only (#5565),5565,[Bugfix] fix dcp_only bug and add e2e accuracy test for dcp only and pcp only,Ignore,Ignore,Internal/CI/test change +[Graph][Fusion] Add AddRMSNormSPPattern and AddRMSNormSPPatternWithBias (#5569),5569,[Graph][Fusion] Add AddRMSNormSPPattern and AddRMSNormSPPatternWithBias,Performance,Include,Operator fusion +[Feature] implement basic framework for batch invariant (#5517),5517,[Feature] implement basic framework for batch invariant,Features,Include,New feature +[Refactor] Cleanup platform (#5566),5566,[Refactor] Cleanup platform,Ignore,Ignore,Internal/CI/test change +[bugfix (pcp)] fix chunked prefill accurancy issue (#5647),5647,[bugfix (pcp)] fix chunked prefill accurancy issue,Highlights,Include,Major feature +[CI] Add DeepSeek-V3.2-W8A8 nightly ci test (#5371),5371,[CI] Add DeepSeek-V3.2-W8A8 nightly ci test,Highlights,Include,Model support +[Feature]EPLB:Adapt DispatchGmmCombineDecode operator to eplb tensor list and expert token numbers (#5552),5552,[Feature]EPLB:Adapt DispatchGmmCombineDecode operator to eplb tensor list and expert token numbers,Highlights,Include,Major feature +[Bugfix] Revert pr4214 multi-stream collect expert hotpot (#5529),5529,[Bugfix] Revert pr4214 multi-stream collect expert hotpot,Ignore,Ignore,Internal/CI/test change +[Bugfix]Add register_kv_cache in ucm_connector (#5657),5657,[Bugfix]Add register_kv_cache in ucm_connector,Highlights,Include,Major feature +[misc]Add Kimi-K2 series to CI model list (#5656),5656,[misc]Add Kimi-K2 series to CI model list,Highlights,Include,Model support +[CI] cleanup single/multi-card test (#5623),5623,[CI] cleanup single/multi-card test,Ignore,Ignore,Internal/CI/test change +[CI] Bump lm-eval version to v0.4.9.2 (#5655),5655,[CI] Bump lm-eval version to v0.4.9.2,Ignore,Ignore,Internal/CI/test change +[CI] Add workflow to cancel running workflows on PR close (#5646),5646,[CI] Add workflow to cancel running workflows on PR close,Ignore,Ignore,Internal/CI/test change +[Bugfix] fix resource are insufficient when pcp and piecewise (#5377),5377,[Bugfix] fix resource are insufficient when pcp and piecewise,Highlights,Include,Major feature +[Bugfix] Fix the graph capture failure issue in the eagle3+full scenario. (#5553),5553,[Bugfix] Fix the graph capture failure issue in the eagle3+full scenario.,Highlights,Include,Major feature +[CI] move image and wheel job to schedule way (#5685),5685,[CI] move image and wheel job to schedule way,Ignore,Ignore,Internal/CI/test change +[Refactor] Fix AttentionMaskBuilder singleton and remove redundant pcp_prefill_mask (#4870),4870,[Refactor] Fix AttentionMaskBuilder singleton and remove redundant pcp_prefill_mask,Ignore,Ignore,Internal/CI/test change +[Refactor] Import global var form vllm instead of overwirte it (#5469),5469,[Refactor] Import global var form vllm instead of overwirte it,Ignore,Ignore,Internal/CI/test change +[Tests] Add qwen3-8b nightly test (#5597),5597,[Tests] Add qwen3-8b nightly test,Others,Include,Miscellaneous +[BugFix][Fusion] Fix graph fusion failure problem (#5676),5676,[BugFix][Fusion] Fix graph fusion failure problem,Performance,Include,Operator fusion +[1/N][CI] Refactor accuracy test (#5400),5400,[1/N][CI] Refactor accuracy test,Ignore,Ignore,Internal/CI/test change +[Kernel] Add moe_gating_top_k operator support for Ascend NPU (#5579),5579,[Kernel] Add moe_gating_top_k operator support for Ascend NPU,Hardware and Operator Support,Include,Operator/hardware support +[BugFix][P/D] Fix pre-create link parameter error (#5694),5694,[BugFix][P/D] Fix pre-create link parameter error,Others,Include,Bug fix +[refactor] Refactor the interface for shard weight and remove the flashcomm2 o_shared interface. (#5181),5181,[refactor] Refactor the interface for shard weight and remove the flashcomm2 o_shared interface.,Ignore,Ignore,Internal/CI/test change +[bugfix] adapt to new implemented get_kv_cache_spec in cpuoffload connector (#4311),4311,[bugfix] adapt to new implemented get_kv_cache_spec in cpuoffload connector,Features,Include,New feature +[Feature] add the magicmtp speculative decoding acceleration algorithm (#5542),5542,[Feature] add the magicmtp speculative decoding acceleration algorithm,Highlights,Include,Major feature +Optimize the print info format when deprecated code is used in vllm-ascend (#5696),5696,Optimize the print info format when deprecated code is used in vllm-ascend,Performance,Include,Performance optimization +[CI] fix image build tag (#5703),5703,[CI] fix image build tag,Ignore,Ignore,Internal/CI/test change +[EPLB][CI] EPLB add aclgraph and redundant expert ci (#5625),5625,[EPLB][CI] EPLB add aclgraph and redundant expert ci,Ignore,Ignore,Internal/CI/test change +[CI] Drop outdated cases (#5709),5709,[CI] Drop outdated cases,Ignore,Ignore,Internal/CI/test change +[CI] Fix image build workflow_dispatch error (#5717),5717,[CI] Fix image build workflow_dispatch error,Ignore,Ignore,Internal/CI/test change +[Feat][Bugfix][main] Adapted SP to eagle3 (#5562),5562,[Feat][Bugfix][main] Adapted SP to eagle3,Highlights,Include,Major feature +[bugfix] Support dsv3.2 enable both mtp and full_decode_only (#5679),5679,[bugfix] Support dsv3.2 enable both mtp and full_decode_only,Highlights,Include,Major feature +[Doc] Add Qwen3-Omni-30B-A3B-Thinking Tutorials (#3991),3991,[Doc] Add Qwen3-Omni-30B-A3B-Thinking Tutorials,Hardware and Operator Support,Include,Operator/hardware support +[Fix] Fixes speculative decode indexing and unpad condition for attention metadata (#5626),5626,[Fix] Fixes speculative decode indexing and unpad condition for attention metadata,Highlights,Include,Major feature +[CI] Add triton ascend in nightly CI (#5716),5716,[CI] Add triton ascend in nightly CI,Hardware and Operator Support,Include,Operator/hardware support +[feature]dcp&pcp support mlapo (#5672),5672,[feature]dcp&pcp support mlapo,Highlights,Include,Major feature +[CI] Remove workflow_dispatch way for image build (#5742),5742,[CI] Remove workflow_dispatch way for image build,Ignore,Ignore,Internal/CI/test change +[Nightly] Move ops to the correct path (#5642),5642,[Nightly] Move ops to the correct path,Others,Include,Miscellaneous +[OP] Enable custom op aclnnMoeInitRoutingCustom (#5332),5332,[OP] Enable custom op aclnnMoeInitRoutingCustom,Hardware and Operator Support,Include,Operator/hardware support +[CI] Add qwen3 next ci (#5395),5395,[CI] Add qwen3 next ci,Ignore,Ignore,Internal/CI/test change +[Doc] add PaddleOCR-VL tutorials guide (#5556),5556,[Doc] add PaddleOCR-VL tutorials guide,Documentation,Include,Model tutorial +[BugFix][DS 3.2] Fix ds indexer accuracy problem caused by rope. (#4641),4641,[BugFix][DS 3.2] Fix ds indexer accuracy problem caused by rope.,Others,Include,Bug fix +[Doc][fix] Fix the title of the document for the layer_sharding feature (#5759),5759,[Doc][fix] Fix the title of the document for the layer_sharding feature,Documentation,Include,Documentation update +[CI] lint and ut use self_hosted runner (#5652),5652,[CI] lint and ut use self_hosted runner,Ignore,Ignore,Internal/CI/test change +[BugFix] NetLoader: No backend type associated with device type npu (#5700),5700,[BugFix] NetLoader: No backend type associated with device type npu,Others,Include,Bug fix +[CI] Accuracy issue of qwen3-next-w8a8 nightly test fix. (#5746),5746,[CI] Accuracy issue of qwen3-next-w8a8 nightly test fix.,Highlights,Include,Model support +[BugFix] Xlite: Bypass the padding of the graph mode in non-MTP cases to obtain the correct decode num. (#5711),5711,[BugFix] Xlite: Bypass the padding of the graph mode in non-MTP cases to obtain the correct decode num.,Highlights,Include,Major feature +[CustomOp] support TensorList for dispatchFFNCombine (#5665),5665,[CustomOp] support TensorList for dispatchFFNCombine,Features,Include,New feature +[CI] Avoid lint and ut for PR push (#5762),5762,[CI] Avoid lint and ut for PR push,Ignore,Ignore,Internal/CI/test change +[BufFix]Fix the error when using Ascend custom operators with rank=128 (#5394),5394,[BufFix]Fix the error when using Ascend custom operators with rank=128,Hardware and Operator Support,Include,Operator/hardware support +"[Refactor] Replace the implementations of o_proj, q_b_proj, and kv_b_proj with custom_op for sharded CP (#5698)",5698,"[Refactor] Replace the implementations of o_proj, q_b_proj, and kv_b_proj with custom_op for sharded CP",Ignore,Ignore,Internal/CI/test change +[Bugfix] Fix matmul allreduce precision issue by using original weight (#4939),4939,[Bugfix] Fix matmul allreduce precision issue by using original weight,Others,Include,Bug fix +[Feature] GLM4.6 support mtp with fullgraph (#5460),5460,[Feature] GLM4.6 support mtp with fullgraph,Highlights,Include,Major feature +[CI]Add Disaggregated PD Nightly Test for Qwen3-235B and Qwen3-VL-235B (#5502),5502,[CI]Add Disaggregated PD Nightly Test for Qwen3-235B and Qwen3-VL-235B,Highlights,Include,Model support +support mxfp8 quantization (qwen dense) (#5723),5723,support mxfp8 quantization (qwen dense),Highlights,Include,Major feature +[Doc] Add GLM4.5 GLM4.6 doc (#5740),5740,[Doc] Add GLM4.5 GLM4.6 doc,Documentation,Include,Model tutorial +[bugfix] Fixing KV Pool Memory Retention and Performance Degradation Issues (#5751),5751,[bugfix] Fixing KV Pool Memory Retention and Performance Degradation Issues,Highlights,Include,Major feature +[P/D][bugfix]Fix the PCP port mapping error issue (#5706),5706,[P/D][bugfix]Fix the PCP port mapping error issue,Highlights,Include,Major feature +[Feat] flashcomm2+oshard Generalized (#4723),4723,[Feat] flashcomm2+oshard Generalized,Performance,Include,Performance optimization +adapt to minimax_m2 (#5624),5624,adapt to minimax_m2,Highlights,Include,Model support +[P/D] layerwise connector supports DeepSeek-V3.2 sparse attention && Distribute transfer tasks to redundant kv_head cards (#5722),5722,[P/D] layerwise connector supports DeepSeek-V3.2 sparse attention && Distribute transfer tasks to redundant kv_head cards,Highlights,Include,Model support +[main][bugfix] Fix fullgraph padding bug in mtp eagle refactor (#5692),5692,[main][bugfix] Fix fullgraph padding bug in mtp eagle refactor,Highlights,Include,Major feature +[Perf] Supports compute-communication overlap in the forward of sfa_v1 in the Sharded-CP feature. (#5701),5701,[Perf] Supports compute-communication overlap in the forward of sfa_v1 in the Sharded-CP feature.,Hardware and Operator Support,Include,Operator/hardware support +[Feature] Support for cross-attention and whisper model (#5592),5592,[Feature] Support for cross-attention and whisper model,Highlights,Include,Major feature +[0.13.0][doc] correct doc url (#5791),5791,[0.13.0][doc] correct doc url,Documentation,Include,Documentation update +[0.13.0][CI] disable main CI (#5792),5792,[0.13.0][CI] disable main CI,Ignore,Ignore,Internal/CI/test change +[0.13.0][Cherry Pick] cherry pick from 5638 Update pd readme (#5811),5811,[0.13.0][Cherry Pick] cherry pick from 5638 Update pd readme,Documentation,Include,Documentation update +[0.13.0][cherry-pick][bugfix](cp) align max_context_chunk to cp_virtual_block_size (#5782),5782,[0.13.0][cherry-pick][bugfix](cp) align max_context_chunk to cp_virtual_block_size,Others,Include,Bug fix +[0.13.0][Bugfix] bugfix for the order of dummy run pad and sync (#5778),5778,[0.13.0][Bugfix] bugfix for the order of dummy run pad and sync,Others,Include,Bug fix +[0.13.0][Patch] AscendLoRAModelManager.__init__ (#5800),5800,[0.13.0][Patch] AscendLoRAModelManager.__init__,Others,Include,Miscellaneous +[cherry-pick][BugFix] Support setting tp=1 for the Eagle draft model to take effect (#5804),5804,[cherry-pick][BugFix] Support setting tp=1 for the Eagle draft model to take effect,Highlights,Include,Major feature +[v0.13.0][Bugfix] Support ALL D-Nodes in fullgraph when running MTP in PD (#5786),5786,[v0.13.0][Bugfix] Support ALL D-Nodes in fullgraph when running MTP in PD,Highlights,Include,Major feature +[0.13.0][cherry-pick]enable ep32 for dispatch_ffn_combine (#5788),5788,[0.13.0][cherry-pick]enable ep32 for dispatch_ffn_combine,Features,Include,New feature +[P/D] [CherryPick] 5846 fix layerwise connector for decoder tp size > num kv he… (#5857),5857,[P/D] [CherryPick] 5846 fix layerwise connector for decoder tp size > num kv he…,Others,Include,Miscellaneous +[0.13.0][cherry-pick][bugfix]Synchronize memcache adaptation on A2 (#5842),5842,[0.13.0][cherry-pick][bugfix]Synchronize memcache adaptation on A2,Hardware and Operator Support,Include,Operator/hardware support +[0.13.0][cherry-pick][Bugfix] Fixed an accuracy problem of sp with eagle3 (#5814),5814,[0.13.0][cherry-pick][Bugfix] Fixed an accuracy problem of sp with eagle3,Highlights,Include,Major feature +[0.13.0][cherry-pick][P/D] bugfix for p node force free requset (#5431) (#5431),5431,[0.13.0][cherry-pick][P/D] bugfix for p node force free requset,Others,Include,Miscellaneous +[0.13.0][cherry-pick][bugfix](cp) replace None with zeros/inf tensor to avoid TypeError (#5844),5844,[0.13.0][cherry-pick][bugfix](cp) replace None with zeros/inf tensor to avoid TypeError,Others,Include,Bug fix +[v0.13.0][bugfix] patch set cudagraph size (#5860),5860,[v0.13.0][bugfix] patch set cudagraph size,Others,Include,Bug fix +[v0.13.0][Bugfix] Fix acc bug when enbale dispatch_gmm_combine_decode and eplb[RFC: issue 5476] (#5836),5836,[v0.13.0][Bugfix] Fix acc bug when enbale dispatch_gmm_combine_decode and eplb[RFC: issue 5476],Highlights,Include,Major feature +[0.13.0][Bugfix] Fix memory inconsistency in cross-process shared memory (#5779),5779,[0.13.0][Bugfix] Fix memory inconsistency in cross-process shared memory,Others,Include,Bug fix +[v0.13.0][cherry-pick][BugFix] Fix DispatchGmmCombineDecode acc bug when big batch (#5873),5873,[v0.13.0][cherry-pick][BugFix] Fix DispatchGmmCombineDecode acc bug when big batch,Others,Include,Bug fix +[v0.13.0][bugfix]Fix graph sync (#5809),5809,[v0.13.0][bugfix]Fix graph sync,Others,Include,Bug fix +[0.13.0][cherry-pick][bugfix]support dsv3.2 enable both mtp and full_decode_only (#5849) (#5849),5849,[0.13.0][cherry-pick][bugfix]support dsv3.2 enable both mtp and full_decode_only,Highlights,Include,Major feature +"Revert ""[BugFix] Support setting tp=1 for the Eagle draft model to take effect(#5519) (#5519)",5519,"Revert ""[BugFix] Support setting tp=1 for the Eagle draft model to take effect",Ignore,Ignore,Internal/CI/test change +enbale qwen3-vl model fc1 feature (#5848),5848,enbale qwen3-vl model fc1 feature,Highlights,Include,Model support +"Revert ""[v0.13.0][bugfix]Fix graph sync (#5809)"" (#5809)",5809,"Revert ""[v0.13.0][bugfix]Fix graph sync",Ignore,Ignore,Internal/CI/test change +[Performance]use triton mrope for Qwen3-VL (#5827),5827,[Performance]use triton mrope for Qwen3-VL,Highlights,Include,Model support +"[P/D]The issue of solving the force-free secondary release request, which causes the node to crash. (#5970)",5970,"[P/D]The issue of solving the force-free secondary release request, which causes the node to crash.",Others,Include,Miscellaneous +[0.13.0][bugfix] fix mooncake kv cache transfer when one P has multi nodes (#5961),5961,[0.13.0][bugfix] fix mooncake kv cache transfer when one P has multi nodes,Highlights,Include,Major feature +[0.13.0][Feature] Support fine-grained shared expert overlap (#5962),5962,[0.13.0][Feature] Support fine-grained shared expert overlap,Performance,Include,Performance optimization +[0.13.0][Bugfix] fix bug of pcp+mtp+async scheduler (#5995),5995,[0.13.0][Bugfix] fix bug of pcp+mtp+async scheduler,Highlights,Include,Major feature +[0.13.0][Bugfix] Add `synced_cudagraph_mode` to limit mixed graph modes in dp ranks (#6011),6011,[0.13.0][Bugfix] Add `synced_cudagraph_mode` to limit mixed graph modes in dp ranks,Others,Include,Bug fix +【0.13.0】【bugfix】Resolved memory deallocation failure in the pooling layer under re-computation workloads. (#6056),6056,【0.13.0】【bugfix】Resolved memory deallocation failure in the pooling layer under re-computation workloads.,Highlights,Include,Major feature +[0.13.0][cherry-pick][bugfix] fix bug of triton mrope (#6009),6009,[0.13.0][cherry-pick][bugfix] fix bug of triton mrope,Hardware and Operator Support,Include,Operator/hardware support +[0.13.0][Bugfix] fix pcp aclgraph qwen FIA bug (#6038),6038,[0.13.0][Bugfix] fix pcp aclgraph qwen FIA bug,Highlights,Include,Major feature +[Bugfix]Fixed precision issues caused by pooled request pooling (#6057),6057,[Bugfix]Fixed precision issues caused by pooled request pooling,Highlights,Include,Major feature +[0.13.0][Bugfix] Fixed an problem related to embeddings sharing (#5972),5972,[0.13.0][Bugfix] Fixed an problem related to embeddings sharing,Others,Include,Bug fix +[v0.13.0][Bugfix] Fix XliteModelRunner init failed when aclgraph is enabled (#5887),5887,[v0.13.0][Bugfix] Fix XliteModelRunner init failed when aclgraph is enabled,Highlights,Include,Major feature +[0.13.0][Bugfix] Fix setting of `speculative_config.enforce_eager` for dsv32 (#5958),5958,[0.13.0][Bugfix] Fix setting of `speculative_config.enforce_eager` for dsv32,Highlights,Include,Major feature +"Revert ""[0.13.0][cherry-pick][bugfix] fix bug of triton mrope"" (#6075)",6075,"Revert ""[0.13.0][cherry-pick][bugfix] fix bug of triton mrope""",Ignore,Ignore,Internal/CI/test change +[0.13.0][cherry-pick][bugfix] fix the complex and potentially problematic generate_kv_idx. (#5955),5955,[0.13.0][cherry-pick][bugfix] fix the complex and potentially problematic generate_kv_idx.,Others,Include,Bug fix +[0.13.0][CI]fix for CI lint (#6093),6093,[0.13.0][CI]fix for CI lint,Ignore,Ignore,Internal/CI/test change +[EPLB][Bugfix] Dispatch Allgather use log2phy if enable eplb (#5933) (#5933),5933,[EPLB][Bugfix] Dispatch Allgather use log2phy if enable eplb,Highlights,Include,Major feature +[EPLB][Bugfix][v0.13.0] Incorporate the warm up of the EPLB into the profile run. (#6099),6099,[EPLB][Bugfix][v0.13.0] Incorporate the warm up of the EPLB into the profile run.,Highlights,Include,Major feature +[0.13.0][Doc] Supplement PD separation parameters of DeepSeek V3.1 (#6054),6054,[0.13.0][Doc] Supplement PD separation parameters of DeepSeek V3.1,Documentation,Include,Documentation update +[v0.13.0][CI] Upgrade to CANN 8.5.0 (#6101),6101,[v0.13.0][CI] Upgrade to CANN 8.5.0,Ignore,Ignore,Internal/CI/test change +[0.13.0][Bugfix] Fix Triton operator usage for multimodal models based on `the mrope_interleaved` parameter (#6074),6074,[0.13.0][Bugfix] Fix Triton operator usage for multimodal models based on `the mrope_interleaved` parameter,Hardware and Operator Support,Include,Operator/hardware support +[v0.13.0][BugFix][Cherry Pick] Fix input parameter bug of dispatch_gmm_combine_decode (#5931),5931,[v0.13.0][BugFix][Cherry Pick] Fix input parameter bug of dispatch_gmm_combine_decode,Others,Include,Bug fix +"[Feature][Cherry Pick]Enable DispatchGmmCombineDecode when eagle is moe with w8a8, or not moe (#6081)",6081,"[Feature][Cherry Pick]Enable DispatchGmmCombineDecode when eagle is moe with w8a8, or not moe",Highlights,Include,Major feature +[v0.13.0][Bugfix] Fix the input constraints checks for the mlapo and bmm_transpose operators (#5764) (#5764),5764,[v0.13.0][Bugfix] Fix the input constraints checks for the mlapo and bmm_transpose operators,Hardware and Operator Support,Include,Operator/hardware support +[EPLB] Config Rename wrapper (#6111),6111,[EPLB] Config Rename wrapper,Highlights,Include,Major feature +[v0.13.0][cherry-pick][BugFix]converting pa get_workspace back to capturing (#6108),6108,[v0.13.0][cherry-pick][BugFix]converting pa get_workspace back to capturing,Others,Include,Bug fix +[cherry-pick][BugFix] Support setting tp=1 for the Eagle draft model to take effect (#6095),6095,[cherry-pick][BugFix] Support setting tp=1 for the Eagle draft model to take effect,Highlights,Include,Major feature +[kv_cache] support multi_block_pool (#6106),6106,[kv_cache] support multi_block_pool,Features,Include,New feature +[CI] Skip some persistently stuck ut cases (#6133),6133,[CI] Skip some persistently stuck ut cases,Ignore,Ignore,Internal/CI/test change +[0.13.0][CI]Add triton ascend version to 3.2.0 (#6105),6105,[0.13.0][CI]Add triton ascend version to 3.2.0,Ignore,Ignore,Internal/CI/test change +[0.13.0][cherry-pick][CP&SP] Integrate FIA operator in mla_cp._forward_decode (#6046),6046,[0.13.0][cherry-pick][CP&SP] Integrate FIA operator in mla_cp._forward_decode,Hardware and Operator Support,Include,Operator/hardware support +[0.13.0][cherry-pick][BugFix] fix 3vl dense model load quant weight (#6103),6103,[0.13.0][cherry-pick][BugFix] fix 3vl dense model load quant weight,Others,Include,Bug fix +[0.13.0][cherry-pick] Reset incompatible config (#6118),6118,[0.13.0][cherry-pick] Reset incompatible config,Others,Include,Miscellaneous +[0.13.0][Bugfix] Remove `use_aclgraph` in mtp_proposer and use `use_cuda_graph` (#6102),6102,[0.13.0][Bugfix] Remove `use_aclgraph` in mtp_proposer and use `use_cuda_graph`,Highlights,Include,Major feature +[0.13.0][BugFix][cherry-pick]hccl bufferSize check for dispatch_ffn_combine (#6131),6131,[0.13.0][BugFix][cherry-pick]hccl bufferSize check for dispatch_ffn_combine,Others,Include,Bug fix +Mix placement (#6086),6086,Mix placement,Others,Include,Miscellaneous +[v0.13.0][Feature] Support DSA-CP for Hybrid scenario (#5702) (#5702),5702,[v0.13.0][Feature] Support DSA-CP for Hybrid scenario,Features,Include,New feature +[0.13.0][P/D][PCP]bugfix pcp force free twice caused logger error (#6132),6132,[0.13.0][P/D][PCP]bugfix pcp force free twice caused logger error,Highlights,Include,Major feature +[EPLB][Bugfix] Do not refresh parameters when eplb_config is not passed (#6160),6160,[EPLB][Bugfix] Do not refresh parameters when eplb_config is not passed,Highlights,Include,Major feature +[0.13.0][Doc] update supported features (#6150),6150,[0.13.0][Doc] update supported features,Documentation,Include,Documentation update +[0.13.0][cherry-pick][CP&SP] Remove CP Redundant Variables after FIA operator enables for CANN 8.5 (#6039),6039,[0.13.0][cherry-pick][CP&SP] Remove CP Redundant Variables after FIA operator enables for CANN 8.5,Hardware and Operator Support,Include,Operator/hardware support +[v0.13.0][cherry-pick][Bugfix] Fix seq_lens reset issue causing performance degradation (#6166),6166,[v0.13.0][cherry-pick][Bugfix] Fix seq_lens reset issue causing performance degradation,Performance,Include,Performance optimization +[0.13.0][Bugix] fix kv pcp+pooling+pd separation bug (#6152),6152,[0.13.0][Bugix] fix kv pcp+pooling+pd separation bug,Highlights,Include,Major feature +[Doc] Document translation (#6066),6066,[Doc] Document translation,Documentation,Include,Documentation update +[v0.13.0][bugfix] fix capture shape in sp_eagle_fullgraph (#6159),6159,[v0.13.0][bugfix] fix capture shape in sp_eagle_fullgraph,Highlights,Include,Major feature +[0.13.0][Feat] Merge the multi eagle graphs to one graph (#6178),6178,[0.13.0][Feat] Merge the multi eagle graphs to one graph,Highlights,Include,Major feature +[0.13.0][BugFix] Avoided a bug of `torch_npu.npu_mm_reduce_scatter_base` when sp size >= 16 (#6167),6167,[0.13.0][BugFix] Avoided a bug of `torch_npu.npu_mm_reduce_scatter_base` when sp size >= 16,Dependencies,Include,Dependency update +[0.13.0][BugFix]bug fix for dispatch_ffn_combine (#6157),6157,[0.13.0][BugFix]bug fix for dispatch_ffn_combine,Others,Include,Bug fix +[0.13.0]Add has_connector_metadata (#6154),6154,[0.13.0]Add has_connector_metadata,Others,Include,Miscellaneous +[v0.13.0]skip eagle dp allreduce (#6162),6162,[v0.13.0]skip eagle dp allreduce,Highlights,Include,Major feature +[0.13.0][KVCache] Support different page sizes (#6171),6171,[0.13.0][KVCache] Support different page sizes,Features,Include,New feature +[Bugfix] Fix the issue of the acceptance rate decline for Qwen3-30B-A3B-EAGLE3 (#6139),6139,[Bugfix] Fix the issue of the acceptance rate decline for Qwen3-30B-A3B-EAGLE3,Highlights,Include,Major feature +[0.13.0][cherry-pick] addrmsnorm op support bias (#6140),6140,[0.13.0][cherry-pick] addrmsnorm op support bias,Features,Include,New feature +[Doc] Refresh doc for 0.13.0 (#6184),6184,[Doc] Refresh doc for 0.13.0,Documentation,Include,Documentation update +[Doc] Add release note for 0.13.0rc2 (#6208),6208,[Doc] Add release note for 0.13.0rc2,Documentation,Include,Documentation update +[0.13.0] [BugFix] buildwheel dependency install (#6211),6211,[0.13.0] [BugFix] buildwheel dependency install,Dependencies,Include,Dependency update +[ci] Fix docker image build (#6215),6215,[ci] Fix docker image build,Ignore,Ignore,Internal/CI/test change +[0.13.0] [BugFix] Fix build wheel (#6220),6220,[0.13.0] [BugFix] Fix build wheel,Others,Include,Bug fix +[Bugfix][v0.13.0] Fix a bug when cherry-pick from main (#6209),6209,[Bugfix][v0.13.0] Fix a bug when cherry-pick from main,Others,Include,Bug fix +[0.13.0][Bugfix] Avoided a bug of drafter when `dp` and `sp` are enabled (#6224),6224,[0.13.0][Bugfix] Avoided a bug of drafter when `dp` and `sp` are enabled,Features,Include,New feature +[Inductor][v0.13.0]Adapt AddRmsNormQuant pass to new addrmsnormBias operator (#6210),6210,[Inductor][v0.13.0]Adapt AddRmsNormQuant pass to new addrmsnormBias operator,Hardware and Operator Support,Include,Operator/hardware support +[v0.13.0][cherry-pick][BugFix] Fix moe_load accumulation error in ACL graph mode (#6258),6258,[v0.13.0][cherry-pick][BugFix] Fix moe_load accumulation error in ACL graph mode,Others,Include,Bug fix +[CI] migrate single card runner to hk (#6260),6260,[CI] migrate single card runner to hk,Ignore,Ignore,Internal/CI/test change +[Refactor] use the count of kv_cache_group to create multi_block_table (#6116),6116,[Refactor] use the count of kv_cache_group to create multi_block_table,Ignore,Ignore,Internal/CI/test change +[Misc][v0.13.0]Removes unnecessary graph size re-initialization (#6281),6281,[Misc][v0.13.0]Removes unnecessary graph size re-initialization,Others,Include,Miscellaneous +[0.13.0][KVCache] Prioritize using a hybrid manager to manage different types of kvcache (#6289),6289,[0.13.0][KVCache] Prioritize using a hybrid manager to manage different types of kvcache,Others,Include,Miscellaneous +[CI] Update pta to 2.8.0.post2 (#6287),6287,[CI] Update pta to 2.8.0.post2,Ignore,Ignore,Internal/CI/test change +[0.13.0][cherry-pick][BugFix][CI]Fix DeepSeek-R1-W8A8-longseq nightly CI (#6337),6337,[0.13.0][cherry-pick][BugFix][CI]Fix DeepSeek-R1-W8A8-longseq nightly CI,Highlights,Include,Model support +[CI] Add per pr image build for nightly test (#6353),6353,[CI] Add per pr image build for nightly test,Others,Include,Miscellaneous +[CI] Cherry pick nightly test from main (#6365),6365,[CI] Cherry pick nightly test from main,Others,Include,Miscellaneous +[cherry-pick][BugFix] Disable enable_shared_expert_dp by default if tensor_parallel_size=1 (#6363),6363,[cherry-pick][BugFix] Disable enable_shared_expert_dp by default if tensor_parallel_size=1,Features,Include,New feature +[0.13.0][Bugfix] Fix hash conflict due to reset incompatible configuations (#6330),6330,[0.13.0][Bugfix] Fix hash conflict due to reset incompatible configuations,Others,Include,Bug fix +[0.13.0][Bugfix] Fix FIA operator validation error in Eagle scenario with CANN 8.5 (#6284),6284,[0.13.0][Bugfix] Fix FIA operator validation error in Eagle scenario with CANN 8.5,Highlights,Include,Major feature +[CI]Limit transformers version (#6373),6373,[CI]Limit transformers version,Ignore,Ignore,Internal/CI/test change +[Doc] Reranker guide remove deprecate task option (#6380),6380,[Doc] Reranker guide remove deprecate task option,Deprecation & Breaking Changes,Include,Breaking change +"[0.13.0][cherry-pick][bugfix](CP,MLA) fix wrong slot_mapping of decode for mixed p/d batch (#6346)",6346,"[0.13.0][cherry-pick][bugfix](CP,MLA) fix wrong slot_mapping of decode for mixed p/d batch",Others,Include,Bug fix +[0.13.0][Profiler] Fix profiler bug (#6383),6383,[0.13.0][Profiler] Fix profiler bug,Others,Include,Miscellaneous +"[0.13.0][cherry-pick][bugfix](pcp,gqa) set kv_inverse_idx_for_chunk and cp_kv_recover_idx_for_chunk to None when dcp only (#6318)",6318,"[0.13.0][cherry-pick][bugfix](pcp,gqa) set kv_inverse_idx_for_chunk and cp_kv_recover_idx_for_chunk to None when dcp only",Highlights,Include,Major feature +[0.13.0][Bugfix]Fix of Pooling Code (#6146),6146,[0.13.0][Bugfix]Fix of Pooling Code,Highlights,Include,Major feature +[0.13.0][cherry-pick][Bugfix][CI] Specify tensorflow version in accuracy test to avoid segmentation fault (#6292) (#6292),6292,[0.13.0][cherry-pick][Bugfix][CI] Specify tensorflow version in accuracy test to avoid segmentation fault,Ignore,Ignore,Internal/CI/test change +[0.13.0][bugfix]Raise exception for Omni models with FLASHCOMM enabled (#6392),6392,[0.13.0][bugfix]Raise exception for Omni models with FLASHCOMM enabled,Performance,Include,Performance optimization +[CI][Nightly] Correct the nightly image build ref (#6396),6396,[CI][Nightly] Correct the nightly image build ref,Others,Include,Miscellaneous +[P/D][0.13.0]Add ssl cert for metaserver proxy (#6400),6400,[P/D][0.13.0]Add ssl cert for metaserver proxy,Others,Include,Miscellaneous +[0.13.0]fix patch cudagraph size (#6397),6397,[0.13.0]fix patch cudagraph size,Others,Include,Miscellaneous +fix: resolve sync bug in DispathFFNCombine when expert num per card is 32 (#6422),6422,fix: resolve sync bug in DispathFFNCombine when expert num per card is 32,Others,Include,Miscellaneous +work around: reset the None reference count to prevent it from droppi… (#6441),6441,work around: reset the None reference count to prevent it from droppi…,Others,Include,Miscellaneous +[v0.13.0][Eagle3]Extend PR #5786 to eagle3 (#6443),5786,[v0.13.0][Eagle3]Extend PR #5786 to eagle3,Highlights,Include,Major feature +[0.13.0][cherry-pick]pick from 6310 to fix rope op (#6444),6444,[0.13.0][cherry-pick]pick from 6310 to fix rope op,Others,Include,Miscellaneous +[CI] change ds32 cudagraph_sizes (#6399),6399,[CI] change ds32 cudagraph_sizes,Ignore,Ignore,Internal/CI/test change +[bugfix][0.13.0]fix bug in dispatch_ffn_combine kernel (#6464),6464,[bugfix][0.13.0]fix bug in dispatch_ffn_combine kernel,Hardware and Operator Support,Include,Operator/hardware support +[v0.13.0][Lora][BugFix] Fix crash on base model requests with LoRA enabled (#6457),6457,[v0.13.0][Lora][BugFix] Fix crash on base model requests with LoRA enabled,Features,Include,New feature +Bugfix: Pre-compile EPLB algorithm successfully in subprocess under graph mode (#6472),6472,Bugfix: Pre-compile EPLB algorithm successfully in subprocess under graph mode,Highlights,Include,Major feature +[0.13.0][Bugfix] fix npu memory is not released in cp (#6479),6479,[0.13.0][Bugfix] fix npu memory is not released in cp,Others,Include,Bug fix +[0.13.0][Bugfix] Fix problematic dummy_run & improper input_batch_size in eagle (#6518),6518,[0.13.0][Bugfix] Fix problematic dummy_run & improper input_batch_size in eagle,Highlights,Include,Major feature diff --git a/skills/upstream/vllm-ascend-releasing-note/output/v0.13.0/2-highlights-note-draft.md b/skills/upstream/vllm-ascend-releasing-note/output/v0.13.0/2-highlights-note-draft.md new file mode 100644 index 0000000..a586932 --- /dev/null +++ b/skills/upstream/vllm-ascend-releasing-note/output/v0.13.0/2-highlights-note-draft.md @@ -0,0 +1,69 @@ +This is the 2nd release candidate of v0.13.0 for vLLM Ascend. Please follow the [official doc](https://docs.vllm.ai/projects/ascend/en/latest) to get started. + +### Highlights + +**Model Support** +* **Qwen3-Next**: Added support for Qwen3-Next-80B-A3B-Instruct with full graph mode, MTP, quantization, and NZ optimization (#3450, #3572, #3428, #3918, #4058, #4245, #4070, #4477) +* **DeepSeek-R1 & DeepSeek-V3.2-Exp**: Added support for DeepSeek-R1 multimodal capabilities and DeepSeek-V3.2-Exp with MTP support (#3631, #3900, #3908, #4191) +* **InternVL**: Added support for InternVL models with e2e tests and accuracy evaluation (#3796, #3964) +* **Kimi-K2**: Fixed weight loading for Kimi-K2 model (#3798) + +**Core Features** +* **Context Parallel & Sequence Parallel (CP/SP)**: Added support for context parallel (PCP) and data context parallel (DCP) with ACLGraph, MTP, and chunked prefill (#3260, #3731, #3801, #4066, #4098, #4183) +* **Full Graph Mode (ACLGraph)**: Enhanced full graph mode with GQA support, memory optimizations, and unified logic between ACLGraph and Torchair (#3560, #3970, #3812, #3879, #3888, #3894) +* **Multi-Token Prediction (MTP)**: Improved MTP support with chunked prefill, quantization, full graph mode, and PCP/DCP integration (#2711, #2713, #3620, #3845, #3910, #3915, #4102, #4111) +* **PD Disaggregation**: Set ADXL engine as default backend for disaggregated prefill with improved documentation and feature guides (#3761, #3950, #4012) +* **KV Pool & Mooncake**: Enhanced KV pool with developer guide and Mooncake connector support for PCP/DCP with multiple input suffixes (#3690, #3752, #3849, #4183) +* **Full Decode Only Mode**: Added support for Qwen3-Next in full_decode_only mode with bug fixes for rare edge cases (#3949, #3986) + +**Developer Documentation** +* Added comprehensive developer guides for ACLGraph, MTP, KV Pool, EPLB, and PD disaggregation features (#3683, #3770, #3752, #3759, #3950) + +### Features + +* **Sampling & Decoding**: Added support for min_p sampling parameter with performance enhancements (#4529) +* **Quantization**: Enhanced quantization support with inductor fusion and dynamic quantization fusion pass (#4168) +* **Prefix Caching**: Improved performance of prefix cache features (#4022) + +### Hardware and Operator Support + +* **Operator Fusion**: Added fused matmul/reduce-scatter kernel, mrope fusion op, and npu_fused_infer_attention_score support (#3693, #3708, #4025) +* **Custom Operators**: Added Triton chunk_gated_delta_rule ops for Qwen3-Next (#4070) +* **MLA/SFA**: Refactored SFA into MLA architecture (#3769) +* **Chip Optimization**: Optimized chip type judgement code and added machine-specific fusion op handling (#4485, #4270) + +### Performance + +* **Async Scheduling**: Improved async scheduling with async copy fixes and eliminated HD synchronization for DeepSeek-V3.2 (#4113, #4233, #4805) +* **FlashComm**: Enhanced FlashComm v2 optimization with o_shared linear and communication domain fixes (#3232, #4188, #4458) +* **Memory Optimization**: Optimized memory usage for DeepSeek MTP and removed redundant D2H operations (#2713, #4063) +* **MoE Optimization**: Optimized all2allv for MoE models (#3738) +* **Attention Optimization**: Moved attention update stream out of loop and converted BSND to TND format for long sequence optimization (#3848, #3778) +* **Quantization Performance**: Moved quantization before allgather in Allgather EP (#3420) +* **Model Runner**: Deleted redundant operations in model_runner and forward_context (#3677) +* **Layerwise Connector**: Performance optimization for layerwise connector (#4043) +* **Vision Models**: Removed Qwen2.5-VL modeling files and added patch for VisionAttention performance (#4349) +* **Sampling**: Removed VLLM_ASCEND_ENABLE_TOPK_TOPP_OPTIMIZATION flag (#4860) + +### Dependencies + +* **torch-npu**: Updated to version 2.7.1 with bug fixes for npu_mm_reduce_scatter_base when sp size >= 16 (#3896, #4433, #6167) +* **vLLM**: Upgraded to vLLM 0.11.2 (#4400) +* **Build Dependencies**: Fixed buildwheel dependency installation (#6211) + +### Deprecation & Breaking Changes + +* **Qwen3-Next Model Files**: Removed Qwen3-Next model files (now using upstream implementation) (#4573) +* **Qwen2.5-VL Modeling**: Removed Qwen2.5-VL modeling files in favor of patch-based approach (#4349) + +### Documentation + +* **Model Tutorials**: Added and refactored tutorials for DeepSeek-V3.2-Exp with simplified MLAPO installation steps (#3871, #4024) +* **Feature Guides**: Added comprehensive guides for disaggregated-prefill, ACLGraph, MTP, KV Pool, and EPLB (#3950, #3683, #3770, #3752, #3759) +* **KV Pool**: Added ADXL timeout parameter documentation in KV pool user guide (#4012) + +### Others + +* **Error Handling**: Added error log for VL models when enabling FLASHCOMM (#4272) +* **Bug Fixes**: Fixed various bugs including MOE fusion operator usage, prefix cache performance, and async scheduling hangs (#3834, #4022, #4233) +* **Testing**: Enhanced nightly CI optimization and added comprehensive test coverage for new features (#3858, #3898, #4509, #4798, #4886) diff --git a/skills/upstream/vllm-ascend-releasing-note/output/v0.13.0/3-highlights-note-edit.md b/skills/upstream/vllm-ascend-releasing-note/output/v0.13.0/3-highlights-note-edit.md new file mode 100644 index 0000000..c29ef3e --- /dev/null +++ b/skills/upstream/vllm-ascend-releasing-note/output/v0.13.0/3-highlights-note-edit.md @@ -0,0 +1,108 @@ +This is the final release of v0.13.0 for vLLM Ascend. Please follow the [official doc](https://docs.vllm.ai/projects/ascend/en/v0.13.0/) to get started. + +### Highlights + +**Model Support** +- **Qwen3-Next**: Full support for Qwen3-Next series including 80B-A3B-Instruct with full graph mode, MTP, quantization (W8A8), NZ optimization, and chunked prefill. Fixed multiple accuracy and stability issues. [#3450](https://github.com/vllm-project/vllm-ascend/pull/3450) [#3572](https://github.com/vllm-project/vllm-ascend/pull/3572) [#3428](https://github.com/vllm-project/vllm-ascend/pull/3428) [#3918](https://github.com/vllm-project/vllm-ascend/pull/3918) [#4058](https://github.com/vllm-project/vllm-ascend/pull/4058) [#4245](https://github.com/vllm-project/vllm-ascend/pull/4245) [#4070](https://github.com/vllm-project/vllm-ascend/pull/4070) [#4477](https://github.com/vllm-project/vllm-ascend/pull/4477) [#4770](https://github.com/vllm-project/vllm-ascend/pull/4770) +- **DeepSeek-R1 & DeepSeek-V3.2**: Added support for DeepSeek-R1 multimodal capabilities and improved DeepSeek-V3.2 with MTP support, performance optimizations, and async scheduling enhancements. [#3631](https://github.com/vllm-project/vllm-ascend/pull/3631) [#3900](https://github.com/vllm-project/vllm-ascend/pull/3900) [#3908](https://github.com/vllm-project/vllm-ascend/pull/3908) [#4191](https://github.com/vllm-project/vllm-ascend/pull/4191) [#4805](https://github.com/vllm-project/vllm-ascend/pull/4805) +- **InternVL**: Added support for InternVL models with comprehensive e2e tests and accuracy evaluation. [#3796](https://github.com/vllm-project/vllm-ascend/pull/3796) [#3964](https://github.com/vllm-project/vllm-ascend/pull/3964) +- **LongCat-Flash**: Added support for LongCat-Flash model. [#3833](https://github.com/vllm-project/vllm-ascend/pull/3833) +- **minimax_m2**: Added support for minimax_m2 model. [#5624](https://github.com/vllm-project/vllm-ascend/pull/5624) +- **Whisper & Cross-Attention**: Added support for cross-attention and Whisper models. [#5592](https://github.com/vllm-project/vllm-ascend/pull/5592) +- **Pooling Models**: Added support for pooling models with PCP adaptation and fixed multiple pooling-related bugs. [#3122](https://github.com/vllm-project/vllm-ascend/pull/3122) [#4143](https://github.com/vllm-project/vllm-ascend/pull/4143) [#6056](https://github.com/vllm-project/vllm-ascend/pull/6056) [#6057](https://github.com/vllm-project/vllm-ascend/pull/6057) [#6146](https://github.com/vllm-project/vllm-ascend/pull/6146) +- **PanguUltraMoE**: Added support for PanguUltraMoE model. [#4615](https://github.com/vllm-project/vllm-ascend/pull/4615) + +**Core Features** +- **Context Parallel (PCP/DCP)**: [Experimental] Added comprehensive support for Prefill Context Parallel (PCP) and Decode Context Parallel (DCP) with ACLGraph, MTP, chunked prefill, MLAPO, and Mooncake connector integration. This is an experimental feature - feedback welcome. [#3260](https://github.com/vllm-project/vllm-ascend/pull/3260) [#3731](https://github.com/vllm-project/vllm-ascend/pull/3731) [#3801](https://github.com/vllm-project/vllm-ascend/pull/3801) [#3980](https://github.com/vllm-project/vllm-ascend/pull/3980) [#4066](https://github.com/vllm-project/vllm-ascend/pull/4066) [#4098](https://github.com/vllm-project/vllm-ascend/pull/4098) [#4183](https://github.com/vllm-project/vllm-ascend/pull/4183) [#5672](https://github.com/vllm-project/vllm-ascend/pull/5672) +- **Full Graph Mode (ACLGraph)**: Enhanced full graph mode with GQA support, memory optimizations, unified logic between ACLGraph and Torchair, and improved stability. [#3560](https://github.com/vllm-project/vllm-ascend/pull/3560) [#3970](https://github.com/vllm-project/vllm-ascend/pull/3970) [#3812](https://github.com/vllm-project/vllm-ascend/pull/3812) [#3879](https://github.com/vllm-project/vllm-ascend/pull/3879) [#3888](https://github.com/vllm-project/vllm-ascend/pull/3888) [#3894](https://github.com/vllm-project/vllm-ascend/pull/3894) [#5118](https://github.com/vllm-project/vllm-ascend/pull/5118) +- **Multi-Token Prediction (MTP)**: Significantly improved MTP support with chunked prefill for DeepSeek, quantization support, full graph mode, PCP/DCP integration, and async scheduling. MTP now works in most cases and is recommended for use. [#2711](https://github.com/vllm-project/vllm-ascend/pull/2711) [#2713](https://github.com/vllm-project/vllm-ascend/pull/2713) [#3620](https://github.com/vllm-project/vllm-ascend/pull/3620) [#3845](https://github.com/vllm-project/vllm-ascend/pull/3845) [#3910](https://github.com/vllm-project/vllm-ascend/pull/3910) [#3915](https://github.com/vllm-project/vllm-ascend/pull/3915) [#4102](https://github.com/vllm-project/vllm-ascend/pull/4102) [#4111](https://github.com/vllm-project/vllm-ascend/pull/4111) [#4770](https://github.com/vllm-project/vllm-ascend/pull/4770) [#5477](https://github.com/vllm-project/vllm-ascend/pull/5477) +- **Eagle Speculative Decoding**: Eagle spec decode now works with full graph mode and is more stable. [#5118](https://github.com/vllm-project/vllm-ascend/pull/5118) [#4893](https://github.com/vllm-project/vllm-ascend/pull/4893) [#5804](https://github.com/vllm-project/vllm-ascend/pull/5804) +- **PD Disaggregation**: Set ADXL engine as default backend for disaggregated prefill with improved performance and stability. Added support for KV NZ feature for DeepSeek decode node. [#3761](https://github.com/vllm-project/vllm-ascend/pull/3761) [#3950](https://github.com/vllm-project/vllm-ascend/pull/3950) [#5008](https://github.com/vllm-project/vllm-ascend/pull/5008) [#3072](https://github.com/vllm-project/vllm-ascend/pull/3072) +- **KV Pool & Mooncake**: Enhanced KV pool with Mooncake connector support for PCP/DCP, multiple input suffixes, and improved performance of Layerwise Connector. [#3690](https://github.com/vllm-project/vllm-ascend/pull/3690) [#3752](https://github.com/vllm-project/vllm-ascend/pull/3752) [#3849](https://github.com/vllm-project/vllm-ascend/pull/3849) [#4183](https://github.com/vllm-project/vllm-ascend/pull/4183) [#5303](https://github.com/vllm-project/vllm-ascend/pull/5303) +- **EPLB (Elastic Prefill Load Balancing)**: EPLB is now more stable with many bug fixes. Mix placement now works. [#6086](https://github.com/vllm-project/vllm-ascend/pull/6086) +- **Full Decode Only Mode**: Added support for Qwen3-Next and DeepSeekv32 in full_decode_only mode with bug fixes. [#3949](https://github.com/vllm-project/vllm-ascend/pull/3949) [#3986](https://github.com/vllm-project/vllm-ascend/pull/3986) [#3763](https://github.com/vllm-project/vllm-ascend/pull/3763) +- **Model Runner V2**: Added basic support for Model Runner V2, the next generation of vLLM. It will be used by default in future releases. [#5210](https://github.com/vllm-project/vllm-ascend/pull/5210) + +### Features + +- **W8A16 Quantization**: Added new W8A16 quantization method support. [#4541](https://github.com/vllm-project/vllm-ascend/pull/4541) +- **UCM Connector**: Added UCMConnector for KV Cache Offloading. [#4411](https://github.com/vllm-project/vllm-ascend/pull/4411) +- **Batch Invariant**: Implemented basic framework for batch invariant feature. [#5517](https://github.com/vllm-project/vllm-ascend/pull/5517) +- **Sampling**: Enhanced sampling with async_scheduler and disable_padded_drafter_batch support in Eagle. [#4893](https://github.com/vllm-project/vllm-ascend/pull/4893) + +### Hardware and Operator Support + +- **Custom Operators**: Added multiple custom operators including: + - Fused matmul/reduce-scatter kernel [#3693](https://github.com/vllm-project/vllm-ascend/pull/3693) + - mrope fusion op [#3708](https://github.com/vllm-project/vllm-ascend/pull/3708) + - Triton chunk_gated_delta_rule ops for Qwen3-Next [#4070](https://github.com/vllm-project/vllm-ascend/pull/4070) + - l2norm triton kernel [#4595](https://github.com/vllm-project/vllm-ascend/pull/4595) + - RejectSampler, MoeInitRoutingCustom, DispatchFFNCombine custom ops +- **Operator Fusion**: Added AddRmsnormQuant fusion pattern with SP support and inductor fusion for quantization. [#5077](https://github.com/vllm-project/vllm-ascend/pull/5077) [#4168](https://github.com/vllm-project/vllm-ascend/pull/4168) +- **MLA/SFA**: Refactored SFA into MLA architecture for better maintainability. [#3769](https://github.com/vllm-project/vllm-ascend/pull/3769) +- **FIA Operator**: Adapted to npu_fused_infer_attention_score with flash decoding function. To optimize performance in small batch size scenarios, this attention operator is now available. Please refer to item 22 in [FAQs](https://docs.vllm.ai/projects/ascend/en/latest/faqs.html) to enable it. [#4025](https://github.com/vllm-project/vllm-ascend/pull/4025) +- **CANN 8.5 Support**: Removed CP redundant variables after FIA operator enables for CANN 8.5. [#6039](https://github.com/vllm-project/vllm-ascend/pull/6039) + +### Performance + +Many custom ops and triton kernels were added in this release to speed up model performance: + +- **DeepSeek Performance**: Improved performance for DeepSeek V3.2 by eliminating HD synchronization in async scheduling and optimizing memory usage for MTP. [#4805](https://github.com/vllm-project/vllm-ascend/pull/4805) [#2713](https://github.com/vllm-project/vllm-ascend/pull/2713) +- **Qwen3-Next Performance**: Improved performance with Triton ops and optimizations. [#5664](https://github.com/vllm-project/vllm-ascend/pull/5664) [#5984](https://github.com/vllm-project/vllm-ascend/pull/5984) [#5765](https://github.com/vllm-project/vllm-ascend/pull/5765) +- **FlashComm**: Enhanced FlashComm v2 optimization with o_shared linear and communication domain fixes. [#3232](https://github.com/vllm-project/vllm-ascend/pull/3232) [#4188](https://github.com/vllm-project/vllm-ascend/pull/4188) [#4458](https://github.com/vllm-project/vllm-ascend/pull/4458) [#5848](https://github.com/vllm-project/vllm-ascend/pull/5848) +- **MoE Optimization**: Optimized all2allv for MoE models and enhanced all-reduce skipping logic. [#3738](https://github.com/vllm-project/vllm-ascend/pull/3738) [#5329](https://github.com/vllm-project/vllm-ascend/pull/5329) +- **Attention Optimization**: Moved attention update stream out of loop, converted BSND to TND format for long sequence optimization, and removed transpose step after attention switching to transpose_batchmatmul. [#3848](https://github.com/vllm-project/vllm-ascend/pull/3848) [#3778](https://github.com/vllm-project/vllm-ascend/pull/3778) [#5390](https://github.com/vllm-project/vllm-ascend/pull/5390) +- **Quantization Performance**: Moved quantization before allgather in Allgather EP. [#3420](https://github.com/vllm-project/vllm-ascend/pull/3420) +- **Layerwise Connector**: Improved performance of Layerwise Connector. [#5303](https://github.com/vllm-project/vllm-ascend/pull/5303) +- **Prefix Cache**: Improved performance of prefix cache features. [#4022](https://github.com/vllm-project/vllm-ascend/pull/4022) +- **Async Scheduling**: Fixed async copy and eliminated hangs in async scheduling. [#4113](https://github.com/vllm-project/vllm-ascend/pull/4113) [#4233](https://github.com/vllm-project/vllm-ascend/pull/4233) +- **Memory Operations**: Removed redundant D2H operations and deleted redundant operations in model_runner. [#4063](https://github.com/vllm-project/vllm-ascend/pull/4063) [#3677](https://github.com/vllm-project/vllm-ascend/pull/3677) +- **Rope Embedding**: Optimized rope embedding with triton kernel for huge performance gain. [#5918](https://github.com/vllm-project/vllm-ascend/pull/5918) +- **Sampling**: Added support for advanced apply_top_k_top_p without top_k constraint. [#6098](https://github.com/vllm-project/vllm-ascend/pull/6098) +- **Multimodal**: Parallelized Q/K/V padding in AscendMMEncoderAttention for better performance. [#6204](https://github.com/vllm-project/vllm-ascend/pull/6204) + +### Dependencies + +- **CANN**: Upgraded to 8.5.0 [#6112](https://github.com/vllm-project/vllm-ascend/pull/6112) +- **torch-npu**: Upgraded to 2.8.0.post1. Please note that the post version will not be installed by default. Please install it by hand from [pypi mirror](https://mirrors.huaweicloud.com/ascend/repos/pypi/torch-npu/). [#3896](https://github.com/vllm-project/vllm-ascend/pull/3896) [#4433](https://github.com/vllm-project/vllm-ascend/pull/4433) +- **triton-ascend**: Upgraded to 3.2.0 [#6105](https://github.com/vllm-project/vllm-ascend/pull/6105) +- **vLLM**: Upgraded to 0.13.0 and dropped 0.12.0 support. [#5146](https://github.com/vllm-project/vllm-ascend/pull/5146) +- **Transformers**: Upgraded to >= 4.57.3 [#5250](https://github.com/vllm-project/vllm-ascend/pull/5250) + +### Deprecation & Breaking Changes + +- **CPUOffloadingConnector** is deprecated. We'll remove it in the next release. It'll be replaced by CPUOffload feature from vLLM in the future. +- **EPLB config options** have been moved to `eplb_config` in [additional config](https://docs.vllm.ai/projects/ascend/en/latest/user_guide/configuration/additional_config.html). The old ones will be removed in the next release. +- **ProfileExecuteDuration** [feature](https://docs.vllm.ai/projects/ascend/en/latest/developer_guide/performance_and_debug/profile_execute_duration.html) is deprecated. It's replaced by `ObservabilityConfig` from vLLM. +- **Ascend Scheduler** has been dropped. [#4623](https://github.com/vllm-project/vllm-ascend/pull/4623) +- **Torchair** has been dropped. [#4814](https://github.com/vllm-project/vllm-ascend/pull/4814) +- **VLLM_ASCEND_ENABLE_DENSE_OPTIMIZE** is removed and `VLLM_ASCEND_ENABLE_PREFETCH_MLP` is recommended to replace as they were always enabled together. [#5272](https://github.com/vllm-project/vllm-ascend/pull/5272) +- **VLLM_ENABLE_FUSED_EXPERTS_ALLGATHER_EP** is dropped now. [#5270](https://github.com/vllm-project/vllm-ascend/pull/5270) +- **VLLM_ASCEND_ENABLE_NZ** is disabled for float weight case, since we noticed that the performance is not good in some float cases. Feel free to set it to 2 if you make sure it works for your case. [#4878](https://github.com/vllm-project/vllm-ascend/pull/4878) +- **chunked_prefill_for_mla** in `additional_config` is dropped now. [#5296](https://github.com/vllm-project/vllm-ascend/pull/5296) +- **dump_config** in `additional_config` is renamed to `dump_config_path` and the type is changed from `dict` to `string`. [#5296](https://github.com/vllm-project/vllm-ascend/pull/5296) +- **--task parameter** for embedding models is deprecated. [#5257](https://github.com/vllm-project/vllm-ascend/pull/5257) +- **The value of VLLM_ASCEND_ENABLE_MLAPO** env will be set to True by default in the next release. It'll be enabled in decode node by default. Please note that this feature will cost more memory. If you are memory sensitive, please set it to False. + +### Documentation + +- Added comprehensive developer guides for ACLGraph, MTP, KV Pool, EPLB, and PD disaggregation features +- Added tutorials for multiple models including DeepSeek-V3.2-Exp, Qwen3-Next, and various multimodal models +- Updated FAQ and configuration documentation + +### Others + +- **OOM Fix**: OOM error on VL models is fixed now. We're keeping observing it. If you hit OOM problem again, please submit an issue. [#5136](https://github.com/vllm-project/vllm-ascend/pull/5136) +- **Qwen3-Next-MTP Accuracy**: Fixed an accuracy bug of Qwen3-Next-MTP when batched inferring. [#4932](https://github.com/vllm-project/vllm-ascend/pull/4932) +- **ZMQ Bug Fix**: Fixed zmq send/receive failed bug. [#5503](https://github.com/vllm-project/vllm-ascend/pull/5503) +- **Weight Transpose**: Fixed weight transpose in RL scenarios. [#5567](https://github.com/vllm-project/vllm-ascend/pull/5567) +- **Eagle3 SP**: Adapted SP to eagle3. [#5562](https://github.com/vllm-project/vllm-ascend/pull/5562) +- **GLM4.6 MTP**: GLM4.6 now supports MTP with fullgraph. [#5460](https://github.com/vllm-project/vllm-ascend/pull/5460) +- **Flashcomm2 Oshard**: Flashcomm2 now works with oshard generalized feature. [#4723](https://github.com/vllm-project/vllm-ascend/pull/4723) +- **Fine-grained Shared Expert Overlap**: Support fine-grained shared expert overlap. [#5962](https://github.com/vllm-project/vllm-ascend/pull/5962) + +### Known Issues + +- Qwen3-Next doesn't support long sequence scenario, and we should limit `gpu-memory-utilization` according to the doc to run Qwen3-Next. We'll improve it in the next release. +- The functional break on Qwen3-Next when the input/output is around 3.5k/1.5k is fixed, but it introduces a regression on performance. We'll fix it in the next release. [#5357](https://github.com/vllm-project/vllm-ascend/issues/5357) +- There is a precision issue with curl on ultra-short sequences in DeepSeek-V3.2. We'll fix it in the next release. [#5370](https://github.com/vllm-project/vllm-ascend/issues/5370) diff --git a/skills/upstream/vllm-ascend-releasing-note/output/v0.13.0/tmp/analyze_commits.py b/skills/upstream/vllm-ascend-releasing-note/output/v0.13.0/tmp/analyze_commits.py new file mode 100644 index 0000000..549f0de --- /dev/null +++ b/skills/upstream/vllm-ascend-releasing-note/output/v0.13.0/tmp/analyze_commits.py @@ -0,0 +1,170 @@ +#!/usr/bin/env python3 +"""Script to analyze commits and categorize them for release notes.""" + +import re +import csv + +def extract_pr_number(title): + """Extract PR number from commit title.""" + match = re.search(r'#(\d+)', title) + return match.group(1) if match else "" + +def categorize_commit(title): + """Categorize commit based on title patterns.""" + title_lower = title.lower() + + # Ignore patterns + ignore_patterns = [ + r'\[ci\](?!.*nightly)', # CI changes (except nightly) + r'\[test\].*ut', # Unit test only + r'\[ut\]', # Unit test + r'bump actions/', # GitHub actions bumps + r'revert', # Reverts + r'\[misc\].*clean', # Cleanup + r'\[refactor\](?!.*user)', # Pure refactoring + r'fix.*test', # Test fixes + r'cleanup', # Cleanup + r'\[lint\]', # Lint + ] + + for pattern in ignore_patterns: + if re.search(pattern, title_lower): + return "Ignore", "Internal/CI/test change" + + # Highlights patterns + if any(keyword in title_lower for keyword in [ + 'pd disaggregation', 'disaggregated-prefill', 'encoder separation', + 'kv cache offload', 'ucm', 'mooncake', 'kv pool', + 'full_decode_only', 'full graph', 'aclgraph', + 'mtp', 'eagle', 'speculative', 'suffix', 'rejection sampler', + 'pcp', 'dcp', 'context parallel', 'sequence parallel', + 'quantization.*w8a8', 'quantization.*w4a8', 'mxfp8', + 'eplb', 'dynamic eplb', + 'xlite', 'npugraph_ex', + 'pooling', 'cross-attention', 'whisper', + 'kv-sharing', 'kv nz', + ]): + return "Highlights", "Major feature" + + # Model support + if any(keyword in title_lower for keyword in [ + 'qwen3-next', 'qwen3-vl', 'qwen2.5-omni', 'qwen2-audio', + 'deepseek-v3', 'deepseek-r1', 'deepseek v3.2', + 'glm-4', 'glm4', 'internvl', 'hunyuan', + 'pangu', 'longcat', 'minimax', 'paddleocr', + 'kimi-k2', + ]): + if '[doc]' in title_lower or 'tutorial' in title_lower: + return "Documentation", "Model tutorial" + return "Highlights", "Model support" + + # Hardware and Operator Support + if any(keyword in title_lower for keyword in [ + 'a2', 'a3', 'a5', '310p', '950', 'ascend950', + 'custom op', 'operator', 'triton', 'kernel', + 'flash_attention', 'fused', 'fusion', + 'mlapo', 'sfa', 'fia', + ]): + if 'fusion' in title_lower or 'fused' in title_lower: + return "Performance", "Operator fusion" + return "Hardware and Operator Support", "Operator/hardware support" + + # Performance + if any(keyword in title_lower for keyword in [ + '[perf]', 'performance', 'optim', 'improve.*performance', + 'async', 'overlap', 'multi-stream', 'flashcomm', + ]): + return "Performance", "Performance optimization" + + # Dependencies + if any(keyword in title_lower for keyword in [ + 'upgrade.*cann', 'upgrade.*torch', 'upgrade.*vllm', + 'torch-npu', 'torch_npu', 'triton-ascend', + 'mooncake.*version', 'ray.*version', + 'transformers.*version', 'upgrade.*version', + 'dependency', 'dependencies', + ]): + return "Dependencies", "Dependency update" + + # Breaking changes + if any(keyword in title_lower for keyword in [ + 'breaking', 'deprecat', 'drop.*support', + 'remove.*scheduler', 'drop.*scheduler', + ]): + return "Deprecation & Breaking Changes", "Breaking change" + + # Documentation + if any(keyword in title_lower for keyword in [ + '[doc]', 'documentation', 'tutorial', 'readme', + 'user guide', 'developer guide', + ]): + return "Documentation", "Documentation update" + + # Features + if any(keyword in title_lower for keyword in [ + '[feat]', '[feature]', 'support', 'add.*feature', + 'enable', 'implement', + ]): + return "Features", "New feature" + + # Bugfixes + if any(keyword in title_lower for keyword in [ + '[bugfix]', '[fix]', 'fix.*bug', 'fix.*error', + 'fix.*issue', 'fix.*accuracy', + ]): + return "Others", "Bug fix" + + return "Others", "Miscellaneous" + +def analyze_commits(input_file, output_file): + """Analyze commits and write to CSV.""" + with open(input_file, 'r') as f: + lines = f.readlines() + + commits = [] + for line in lines: + line = line.strip() + if not line: + continue + + # Use the line as-is (no "- " prefix in this format) + title = line + + pr_number = extract_pr_number(title) + category, reason = categorize_commit(title) + decision = "Ignore" if category == "Ignore" else "Include" + + # Generate user-facing impact + impact = title.split('(#')[0].strip() if '(#' in title else title + + commits.append({ + 'title': title, + 'pr_number': pr_number, + 'user_facing_impact': impact, + 'category': category, + 'decision': decision, + 'reason': reason + }) + + # Write to CSV + with open(output_file, 'w', newline='', encoding='utf-8') as f: + fieldnames = ['title', 'pr_number', 'user_facing_impact', 'category', 'decision', 'reason'] + writer = csv.DictWriter(f, fieldnames=fieldnames) + writer.writeheader() + writer.writerows(commits) + + # Print summary + category_counts = {} + for commit in commits: + cat = commit['category'] + category_counts[cat] = category_counts.get(cat, 0) + 1 + + print(f"Total commits: {len(commits)}") + print("\nCategory breakdown:") + for cat, count in sorted(category_counts.items(), key=lambda x: -x[1]): + print(f" {cat}: {count}") + +if __name__ == '__main__': + input_file = '0-current-raw-commits.md' + output_file = '1-commit-analysis-draft.csv' + analyze_commits(input_file, output_file) diff --git a/skills/upstream/vllm-ascend-releasing-note/references/ref-past-release-notes-highlight.md b/skills/upstream/vllm-ascend-releasing-note/references/ref-past-release-notes-highlight.md new file mode 100644 index 0000000..3c27595 --- /dev/null +++ b/skills/upstream/vllm-ascend-releasing-note/references/ref-past-release-notes-highlight.md @@ -0,0 +1,198 @@ +## v0.14.0rc1 - 2026.01.26 + +This is the first release candidate of v0.14.0 for vLLM Ascend. Please follow the [official doc](https://docs.vllm.ai/projects/ascend/en/latest) to get started. This release includes all the changes in v0.13.0rc2. So We just list the differences from v0.13.0rc2. If you are upgrading from v0.13.0rc1, please read both v0.14.0rc1 and v0.13.0rc2 release notes. + +### Highlights + +- 310P support is back now. In this release, only basic dense and vl models are supported with eager mode. We'll keep improving and maintaining the support for 310P. [#5776](https://github.com/vllm-project/vllm-ascend/pull/5776) +- Support compressed tensors moe w8a8-int8 quantization. [#5718](https://github.com/vllm-project/vllm-ascend/pull/5718) +- Support Medusa speculative decoding. [#5668](https://github.com/vllm-project/vllm-ascend/pull/5668) +- Support Eagle3 speculative decoding for Qwen3vl. [#4848](https://github.com/vllm-project/vllm-ascend/pull/4848) + +### Features + +- Xlite Backend supports Qwen3 MoE now. [#5951](https://github.com/vllm-project/vllm-ascend/pull/5951) +- Support DSA-CP for PD-mix deployment case. [#5702](https://github.com/vllm-project/vllm-ascend/pull/5702) +- Add support of new W4A4_LAOS_DYNAMIC quantization method. [#5143](https://github.com/vllm-project/vllm-ascend/pull/5143) + +### Performance + +- The performance of Qwen3-next has been improved. [#5664](https://github.com/vllm-project/vllm-ascend/pull/5664) [#5984](https://github.com/vllm-project/vllm-ascend/pull/5984) [#5765](https://github.com/vllm-project/vllm-ascend/pull/5765) +- The CPU bind logic and performance has been improved. [#5555](https://github.com/vllm-project/vllm-ascend/pull/5555) +- Merge Q/K split to simplify AscendApplyRotaryEmb for better performance. [#5799](https://github.com/vllm-project/vllm-ascend/pull/5799) +- Add Matmul Allreduce Rmsnorm fusion Pass. It's disabled by default. Set `fuse_allreduce_rms=True` in `--additional_config` to enable it. [#5034](https://github.com/vllm-project/vllm-ascend/pull/5034) +- Optimize rope embedding with triton kernel for huge performance gain. [#5918](https://github.com/vllm-project/vllm-ascend/pull/5918) +- support advanced apply_top_k_top_p without top_k constraint. [#6098](https://github.com/vllm-project/vllm-ascend/pull/6098) +- Parallelize Q/K/V padding in AscendMMEncoderAttention for better performance. [#6204](https://github.com/vllm-project/vllm-ascend/pull/6204) + +### Others + +- model runner v2 support triton of penalty. [#5854](https://github.com/vllm-project/vllm-ascend/pull/5854) +- model runner v2 support eagle spec decoding. [#5840](https://github.com/vllm-project/vllm-ascend/pull/5840) +- Fix multi-modal inference OOM issues by setting `expandable_segments:True` by default. [#5855](https://github.com/vllm-project/vllm-ascend/pull/5855) +- `VLLM_ASCEND_ENABLE_MLAPO` is set to `True` by default. It's enabled automatically on decode node in PD deployment case. Please note that this feature will cost more memory. If you are memory sensitive, please set it to False. [#5952](https://github.com/vllm-project/vllm-ascend/pull/5952) +- SSL config can be set to kv_extra_config for PD deployment with mooncake layerwise connector. [#5875](https://github.com/vllm-project/vllm-ascend/pull/5875) +- support `--max_model_len=auto`. [#6193](https://github.com/vllm-project/vllm-ascend/pull/6193) + +### Dependencies + +- torch-npu is upgraded to 2.9.0 [#6112](https://github.com/vllm-project/vllm-ascend/pull/6112) + +### Deprecation & Breaking Changes + +- EPLB config options is moved to `eplb_config` in [additional config](https://docs.vllm.ai/projects/ascend/en/latest/user_guide/configuration/additional_config.html). The old ones are removed in this release. +- The profiler envs, such as `VLLM_TORCH_PROFILER_DIR` and `VLLM_TORCH_PROFILER_WITH_PROFILE_MEMORY` do not work with vLLM Ascend now. Please use vLLM `--profiler-config` parameters instead. [#5928](https://github.com/vllm-project/vllm-ascend/pull/5928) + +### Known Issues + +- If you hit the pickle error from `EngineCore` process sometimes, please cherry-pick the [PR](https://github.com/vllm-project/vllm/pull/32022) into your local vLLM code. This known issue will be fixed in vLLM in the next release. + +## v0.13.0rc2 - 2026.01.24 + +This is the second release candidate of v0.13.0 for vLLM Ascend. In this rc release, we fixed lots of bugs and improved the performance of many models. Please follow the [official doc](https://docs.vllm.ai/projects/ascend/en/v0.13.0/) to get started. Any feedback is welcome to help us to improve the final version of v0.13.0. + +### Highlights + +We mainly focus on quality and performance improvement in this release. The spec decode, graph mode, context parallel and EPLB have been improved significantly. A lot of bugs have been fixed and the performance has been improved for DeepSeek3.1/3.2, Qwen3 Dense/MOE models. + +### Features + +- implement basic framework for batch invariant [#5517](https://github.com/vllm-project/vllm-ascend/pull/5517) +- Eagle spec decode feature now works with full graph mode. [#5118](https://github.com/vllm-project/vllm-ascend/pull/5118) +- Context Parallel(PCP&DCP) feature is more stable now. And it works for most case. Please try it out. +- MTP and eagle spec decode feature now works in most cases. And it's suggested to use them in most cases. +- EPLB feature more stable now. Many bugs have been fixed. Mix placement works now [#6086](https://github.com/vllm-project/vllm-ascend/pull/6086) +- Support kv nz feature for DeepSeek decode node in disagg-prefill scenario [#3072](https://github.com/vllm-project/vllm-ascend/pull/3072) + +### Model Support + +- LongCat-Flash is supported now.[#3833](https://github.com/vllm-project/vllm-ascend/pull/3833) +- minimax_m2 is supported now. [#5624](https://github.com/vllm-project/vllm-ascend/pull/5624) +- Support for cross-attention and whisper models [#5592](https://github.com/vllm-project/vllm-ascend/pull/5592) + +### Performance + +- Many custom ops and triton kernels are added in this release to speed up the performance of models. Such as `RejectSampler`, `MoeInitRoutingCustom`, `DispatchFFNCombine` and so on. +- Improved the performance of Layerwise Connector [#5303](https://github.com/vllm-project/vllm-ascend/pull/5303) + +### Others + +- Basic support Model Runner v2. Model Runner V2 is the next generation of vLLM. It will be used by default in the future release. [#5210](https://github.com/vllm-project/vllm-ascend/pull/5210) +- Fixed a bug that the zmq send/receive may failed [#5503](https://github.com/vllm-project/vllm-ascend/pull/5503) +- Supported to use full-graph with Qwen3-Next-MTP [#5477](https://github.com/vllm-project/vllm-ascend/pull/5477) +- Fix weight transpose in RL scenarios [#5567](https://github.com/vllm-project/vllm-ascend/pull/5567) +- Adapted SP to eagle3 [#5562](https://github.com/vllm-project/vllm-ascend/pull/5562) +- Context Parallel(PCP&DCP) support mlapo [#5672](https://github.com/vllm-project/vllm-ascend/pull/5672) +- GLM4.6 support mtp with fullgraph [#5460](https://github.com/vllm-project/vllm-ascend/pull/5460) +- Flashcomm2 now works with oshard generalized feature [#4723](https://github.com/vllm-project/vllm-ascend/pull/4723) +- Support setting tp=1 for the Eagle draft model [#5804](https://github.com/vllm-project/vllm-ascend/pull/5804) +- Flashcomm1 feature now works with qwen3-vl [#5848](https://github.com/vllm-project/vllm-ascend/pull/5848) +- Support fine-grained shared expert overlap [#5962](https://github.com/vllm-project/vllm-ascend/pull/5962) + +### Dependencies + +- CANN is upgraded to 8.5.0 +- torch-npu is upgraded to 2.8.0.post1. Please note that the post version will not be installed by default. Please install it by hand from [pypi mirror](https://mirrors.huaweicloud.com/ascend/repos/pypi/torch-npu/). +- triton-ascend is upgraded to 3.2.0 + +### Deprecation & Breaking Changes + +- `CPUOffloadingConnector` is deprecated. We'll remove it in the next release. It'll be replaced by CPUOffload feature from vLLM in the future. +- eplb config options is moved to `eplb_config` in [additional config](https://docs.vllm.ai/projects/ascend/en/latest/user_guide/configuration/additional_config.html). The old ones will be removed in the next release. +- `ProfileExecuteDuration` [feature](https://docs.vllm.ai/projects/ascend/en/latest/developer_guide/performance_and_debug/profile_execute_duration.html) is deprecated. It's replaced by `ObservabilityConfig` from vLLM. +- The value of `VLLM_ASCEND_ENABLE_MLAPO` env will be set to True by default in the next release. It'll be enabled in decode node by default. Please note that this feature will cost more memory. If you are memory sensitive, please set it to False. + +## v0.13.0rc1 - 2025.12.27 + +This is the first release candidate of v0.13.0 for vLLM Ascend. We landed lots of bug fix, performance improvement and feature support in this release. Any feedback is welcome to help us to improve vLLM Ascend. Please follow the [official doc](https://docs.vllm.ai/projects/ascend/en/latest) to get started. + +### Highlights + +- Improved the performance of DeepSeek V3.2, please refer to [tutorials](https://docs.vllm.ai/projects/ascend/en/latest/tutorials/DeepSeek-V3.2.html) +- Qwen3-Next MTP with chunked prefill is supported now [#4770](https://github.com/vllm-project/vllm-ascend/pull/4770), please refer to [tutorials](https://docs.vllm.ai/projects/ascend/en/latest/tutorials/Qwen3-Next.html) +- [Experimental] Prefill Context Parallel and Decode Context Parallel are supported, but notice that it is an experimental feature now, welcome any feedback. please refer to [context parallel feature guide](https://docs.vllm.ai/projects/ascend/en/latest/user_guide/feature_guide/context_parallel.html) + +### Features + +- Support openPangu Ultra MoE [4615](https://github.com/vllm-project/vllm-ascend/pull/4615) +- A new quantization method W8A16 is supported now. [#4541](https://github.com/vllm-project/vllm-ascend/pull/4541) +- Cross-machine Disaggregated Prefill is supported now. [#5008](https://github.com/vllm-project/vllm-ascend/pull/5008) +- Add UCMConnector for KV Cache Offloading. [#4411](https://github.com/vllm-project/vllm-ascend/pull/4411) +- Support async_scheduler and disable_padded_drafter_batch in eagle. [#4893](https://github.com/vllm-project/vllm-ascend/pull/4893) +- Support pcp + mtp in full graph mode. [#4572](https://github.com/vllm-project/vllm-ascend/pull/4572) +- Enhance all-reduce skipping logic for MoE models in NPUModelRunner [#5329](https://github.com/vllm-project/vllm-ascend/pull/5329) + +### Performance + +Some general performance improvement: + +- Add l2norm triton kernel [#4595](https://github.com/vllm-project/vllm-ascend/pull/4595) +- Add new pattern for AddRmsnormQuant with SP, which could only take effect in graph mode. [#5077](https://github.com/vllm-project/vllm-ascend/pull/5077) +- Add async exponential while model executing. [#4501](https://github.com/vllm-project/vllm-ascend/pull/4501) +- Remove the transpose step after attention and switch to transpose_batchmatmul [#5390](https://github.com/vllm-project/vllm-ascend/pull/5390) +- To optimize the performance in small batch size scenario, an attention operator with flash decoding function is offered, please refer to item 22 in [FAQs](https://docs.vllm.ai/projects/ascend/en/latest/faqs.html) to enable it. + +### Other + +- OOM error on VL models is fixed now. We're keeping observing it, if you hit OOM problem again, please submit an issue. [#5136](https://github.com/vllm-project/vllm-ascend/pull/5136) +- Fixed an accuracy bug of Qwen3-Next-MTP when batched inferring. [#4932](https://github.com/vllm-project/vllm-ascend/pull/4932) +- Fix npu-cpu offloading interface change bug. [#5290](https://github.com/vllm-project/vllm-ascend/pull/5290) +- Fix MHA model runtime error in aclgraph mode [#5397](https://github.com/vllm-project/vllm-ascend/pull/5397) +- Fix unsuitable moe_comm_type under ep=1 scenario [#5388](https://github.com/vllm-project/vllm-ascend/pull/5388) + +### Deprecation & Breaking Changes + +- `VLLM_ASCEND_ENABLE_DENSE_OPTIMIZE` is removed and `VLLM_ASCEND_ENABLE_PREFETCH_MLP` is recommend to replace as they always be enabled together. [#5272](https://github.com/vllm-project/vllm-ascend/pull/5272) +- `VLLM_ENABLE_FUSED_EXPERTS_ALLGATHER_EP` is dropped now. [#5270](https://github.com/vllm-project/vllm-ascend/pull/5270) +- `VLLM_ASCEND_ENABLE_NZ` is disabled for float weight case, since we notice that the performance is not good in some float case. Feel free to set it to 2 if you make sure it works for your case. [#4878](https://github.com/vllm-project/vllm-ascend/pull/4878) +- `chunked_prefill_for_mla` in `additional_config` is dropped now. [#5296](https://github.com/vllm-project/vllm-ascend/pull/5296) +- `dump_config` in `additional_config` is renamed to `dump_config_path` and the type is change from `dict` to `string`. [#5296](https://github.com/vllm-project/vllm-ascend/pull/5296) + +### Dependencies + +- vLLM version has been upgraded to 0.13.0 and drop 0.12.0 support. [#5146](https://github.com/vllm-project/vllm-ascend/pull/5146) +- Transformer version has been upgraded >= 4.57.3 [#5250](https://github.com/vllm-project/vllm-ascend/pull/5250) + +### Known Issues + +- Qwen3-Next doesn't support long sequence scenario, and we should limit `gpu-memory-utilization` according to the doc to run Qwen3-Next. We'll improve it in the next release +- The functional break on Qwen3-Next when the input/output is around 3.5k/1.5k is fixed, but it introduces a regression on performance. We'll fix it in next release. [#5357](https://github.com/vllm-project/vllm-ascend/issues/5357) +- There is a precision issue with curl on ultra-short sequences in DeepSeek-V3.2. We'll fix it in next release. [#5370](https://github.com/vllm-project/vllm-ascend/issues/5370) + +## v0.11.0 - 2025.12.16 + +We're excited to announce the release of v0.11.0 for vLLM Ascend. This is the official release for v0.11.0. Please follow the [official doc](https://docs.vllm.ai/projects/ascend/en/v0.11.0) to get started. We'll consider to release post version in the future if needed. This release note will only contain the important change and note from v0.11.0rc3. + +### Highlights + +- Improved the performance for deepseek 3/3.1. [#3995](https://github.com/vllm-project/vllm-ascend/pull/3995) +- Fixed the accuracy bug for qwen3-vl. [#4811](https://github.com/vllm-project/vllm-ascend/pull/4811) +- Improved the performance of sample. [#4153](https://github.com/vllm-project/vllm-ascend/pull/4153) +- Eagle3 is back now. [#4721](https://github.com/vllm-project/vllm-ascend/pull/4721) + +### Other + +- Improved the performance for kimi-k2. [#4555](https://github.com/vllm-project/vllm-ascend/pull/4555) +- Fixed a quantization bug for deepseek3.2-exp. [#4797](https://github.com/vllm-project/vllm-ascend/pull/4797) +- Fixed qwen3-vl-moe bug under high concurrency. [#4658](https://github.com/vllm-project/vllm-ascend/pull/4658) +- Fixed an accuracy bug for Prefill Decode disaggregation case. [#4437](https://github.com/vllm-project/vllm-ascend/pull/4437) +- Fixed some bugs for EPLB [#4576](https://github.com/vllm-project/vllm-ascend/pull/4576) [#4777](https://github.com/vllm-project/vllm-ascend/pull/4777) +- Fixed the version incompatibility issue for openEuler docker image. [#4745](https://github.com/vllm-project/vllm-ascend/pull/4745) + +### Deprecation announcement + +- LLMdatadist connector has been deprecated, it'll be removed in v0.12.0rc1 +- Torchair graph has been deprecated, it'll be removed in v0.12.0rc1 +- Ascend scheduler has been deprecated, it'll be removed in v0.12.0rc1 + +### Upgrade notice + +- torch-npu is upgraded to 2.7.1.post1. Please note that the package is pushed to [pypi mirror](https://mirrors.huaweicloud.com/ascend/repos/pypi/torch-npu/). So it's hard to add it to auto dependence. Please install it by yourself. +- CANN is upgraded to 8.3.rc2. + +### Known Issues + +- Qwen3-Next doesn't support expert parallel and MTP features in this release. And it'll be oom if the input is too long. We'll improve it in the next release +- Deepseek 3.2 only work with torchair graph mode in this release. We'll make it work with aclgraph mode in the next release. +- Qwen2-audio doesn't work by default. Temporary solution is to set `--gpu-memory-utilization` to a suitable value, such as 0.8. +- CPU bind feature doesn't work if more than one vLLM instance is running on the same node. diff --git a/skills/upstream/vllm-ascend-releasing-note/scripts/fetch_commits-optimize.py b/skills/upstream/vllm-ascend-releasing-note/scripts/fetch_commits-optimize.py new file mode 100644 index 0000000..3c2a610 --- /dev/null +++ b/skills/upstream/vllm-ascend-releasing-note/scripts/fetch_commits-optimize.py @@ -0,0 +1,1587 @@ +#!/usr/bin/env python3 +""" +Fetch all commits between two tags from a GitHub repository. +Usage: python fetch_commits.py [--token YOUR_GITHUB_TOKEN] +""" + +import argparse +import os +import re + +import dotenv +import requests + +# Load .env.local first (higher priority), then .env as fallback +dotenv.load_dotenv(".env.local") +dotenv.load_dotenv() # .env as fallback + + +def get_github_token(): + """Get GitHub token from environment or argument.""" + return os.environ.get("GITHUB_TOKEN") or os.environ.get("GH_TOKEN") + + +def resolve_tag_to_sha(base_url: str, tag: str, headers: dict) -> str: + """Resolve a tag name to its commit SHA.""" + print(f"Resolving tag {tag}...") + + tag_resp = requests.get(f"{base_url}/git/refs/tags/{tag}", headers=headers) + if tag_resp.status_code != 200: + raise Exception(f"Failed to get tag {tag}: {tag_resp.text}") + + tag_data = tag_resp.json() + sha = tag_data["object"]["sha"] + + # If it's an annotated tag, we need to get the commit it points to + if tag_data["object"]["type"] == "tag": + tag_obj_resp = requests.get(f"{base_url}/git/tags/{sha}", headers=headers) + if tag_obj_resp.status_code == 200: + sha = tag_obj_resp.json()["object"]["sha"] + + return sha + + +def resolve_commit_sha(base_url: str, commit_ref: str, headers: dict) -> str: + """Resolve a commit reference (SHA or short SHA) to full SHA.""" + print(f"Resolving commit {commit_ref}...") + + commit_resp = requests.get(f"{base_url}/commits/{commit_ref}", headers=headers) + if commit_resp.status_code != 200: + raise Exception(f"Failed to get commit {commit_ref}: {commit_resp.text}") + + return commit_resp.json()["sha"] + + +def get_default_branch_head(base_url: str, headers: dict) -> tuple[str, str]: + """ + Get the HEAD commit of the default branch. + + Returns: + Tuple of (branch_name, head_sha) + """ + print("Getting default branch HEAD...") + + # Get repository info to find default branch + repo_resp = requests.get(base_url, headers=headers) + if repo_resp.status_code != 200: + raise Exception(f"Failed to get repository info: {repo_resp.text}") + + default_branch = repo_resp.json()["default_branch"] + print(f" Default branch: {default_branch}") + + # Get the HEAD commit of the default branch + branch_resp = requests.get(f"{base_url}/branches/{default_branch}", headers=headers) + if branch_resp.status_code != 200: + raise Exception(f"Failed to get branch {default_branch}: {branch_resp.text}") + + head_sha = branch_resp.json()["commit"]["sha"] + print(f" HEAD: {head_sha[:8]}") + + return (default_branch, head_sha) + + +def get_all_tags(base_url: str, headers: dict) -> list[dict]: + """Get all tags from the repository with their commit SHAs and dates.""" + print("Fetching all tags...") + + all_tags = [] + page = 1 + per_page = 100 + + while True: + resp = requests.get( + f"{base_url}/tags", + headers=headers, + params={"per_page": per_page, "page": page}, + ) + + if resp.status_code != 200: + raise Exception(f"Failed to get tags: {resp.text}") + + tags = resp.json() + if not tags: + break + + all_tags.extend(tags) + page += 1 + + if len(tags) < per_page: + break + + print(f" Found {len(all_tags)} tags") + return all_tags + + +def get_commit_date(base_url: str, sha: str, headers: dict) -> str: + """Get the commit date for a given SHA.""" + commit_resp = requests.get(f"{base_url}/commits/{sha}", headers=headers) + if commit_resp.status_code != 200: + return None + return commit_resp.json()["commit"]["committer"]["date"] + + +def find_previous_tag( + base_url: str, head_sha: str, headers: dict, tag_pattern: str | None = None +) -> tuple[str, str] | None: + """ + Find the most recent tag that is an ancestor of the given commit. + + Uses git history to find tags that are reachable from the commit. + + Args: + base_url: GitHub API base URL + head_sha: The commit SHA to search from + headers: Request headers + tag_pattern: Optional regex pattern to filter tags (e.g., r'^v\\d+\\.\\d+\\.\\d+$') + + Returns: + Tuple of (tag_name, tag_sha) or None if no tag found + """ + print(f"Finding previous tag before commit {head_sha[:8]}...") + + # Get the date of the head commit + head_date = get_commit_date(base_url, head_sha, headers) + if not head_date: + print(" Warning: Could not get head commit date") + return None + + print(f" Head commit date: {head_date}") + + # Get all tags + all_tags = get_all_tags(base_url, headers) + + # Filter tags by pattern if provided + if tag_pattern: + import re + + pattern = re.compile(tag_pattern) + all_tags = [t for t in all_tags if pattern.match(t["name"])] + print(f" After pattern filter: {len(all_tags)} tags") + + # For each tag, check if it's an ancestor of head_sha and get its date + tag_candidates = [] + + for tag in all_tags: + tag_name = tag["name"] + tag_commit_sha = tag["commit"]["sha"] + + # Skip if this is the same commit as head + if tag_commit_sha == head_sha: + continue + + # Check if this tag's commit is an ancestor of head + compare_resp = requests.get( + f"{base_url}/compare/{tag_commit_sha}...{head_sha}", headers=headers + ) + + if compare_resp.status_code != 200: + continue + + compare_data = compare_resp.json() + + # If tag is behind head (status = "behind" or "ahead"), it's an ancestor + # We want tags where the comparison shows head is ahead + if compare_data.get("status") in ["ahead", "diverged"]: + # Get the tag's commit date + tag_date = get_commit_date(base_url, tag_commit_sha, headers) + if tag_date and tag_date < head_date: + tag_candidates.append( + { + "name": tag_name, + "sha": tag_commit_sha, + "date": tag_date, + "ahead_by": compare_data.get("ahead_by", 0), + } + ) + print( + f" Found candidate: {tag_name} ({compare_data.get('ahead_by', 0)} commits behind)" + ) + + if not tag_candidates: + print(" No previous tag found") + return None + + # Sort by date (most recent first) or by ahead_by (smallest first) + # Using ahead_by gives us the closest tag + tag_candidates.sort(key=lambda x: x["ahead_by"]) + + best_tag = tag_candidates[0] + print(f" Selected: {best_tag['name']} ({best_tag['ahead_by']} commits behind)") + + return (best_tag["name"], best_tag["sha"]) + + +def fetch_commits_between_tags( + owner: str, repo: str, base_tag: str, head_tag: str, token: str | None = None +) -> list[dict]: + """ + Fetch all commits between two tags by walking the commit graph. + + This method traverses from head_tag back to base_tag, collecting all commits. + It properly handles the commit history and doesn't rely on date filtering. + + Args: + owner: Repository owner (e.g., 'vllm-project') + repo: Repository name (e.g., 'vllm') + base_tag: Base tag (older, e.g., 'v0.11.2') + head_tag: Head tag (newer, e.g., 'v0.12.0') + token: Optional GitHub token for higher rate limits + + Returns: + List of commit dictionaries + """ + headers = { + "Accept": "application/vnd.github.v3+json", + } + if token: + headers["Authorization"] = f"token {token}" + + base_url = f"https://api.github.com/repos/{owner}/{repo}" + + # Resolve tags to commit SHAs + base_sha = resolve_tag_to_sha(base_url, base_tag, headers) + head_sha = resolve_tag_to_sha(base_url, head_tag, headers) + + print(f"\nBase SHA ({base_tag}): {base_sha}") + print(f"Head SHA ({head_tag}): {head_sha}") + + # First, use Compare API to get total commit count (for progress info) + print(f"\nComparing {base_tag}...{head_tag}...") + compare_resp = requests.get( + f"{base_url}/compare/{base_sha}...{head_sha}", headers=headers + ) + if compare_resp.status_code == 200: + compare_data = compare_resp.json() + total_commits = compare_data.get("total_commits", "unknown") + print(f"Total commits to fetch: {total_commits}") + + # Walk the commit history from head to base + # We use the commits API starting from head_sha and stop when we reach base_sha + all_commits = [] + seen_shas = set() + seen_shas.add(base_sha) # Don't include the base commit itself + + # BFS traversal of commit graph + to_visit = [head_sha] + page_count = 0 + + print(f"\nFetching commits from {head_tag} back to {base_tag}...") + + while to_visit: + current_sha = to_visit.pop(0) + + if current_sha in seen_shas: + continue + + seen_shas.add(current_sha) + + # Fetch commit details + commit_resp = requests.get(f"{base_url}/commits/{current_sha}", headers=headers) + + if commit_resp.status_code != 200: + print(f" Warning: Failed to fetch commit {current_sha[:8]}") + continue + + commit = commit_resp.json() + all_commits.append(commit) + + # Add parent commits to visit queue + for parent in commit.get("parents", []): + parent_sha = parent["sha"] + if parent_sha not in seen_shas: + to_visit.append(parent_sha) + + # Progress logging + if len(all_commits) % 50 == 0: + page_count += 1 + print(f" Fetched {len(all_commits)} commits...") + + print(f" Completed: {len(all_commits)} commits fetched") + + return all_commits + + +def fetch_commits_by_date_range( + owner: str, + repo: str, + since: str, + until: str, + token: str | None = None, + branch: str | None = None, +) -> list[dict]: + """ + Fetch all commits within a date range. + + Args: + owner: Repository owner (e.g., 'vllm-project') + repo: Repository name (e.g., 'vllm') + since: Start date (ISO 8601 format, e.g., '2025-01-01' or '2025-01-01T00:00:00Z') + until: End date (ISO 8601 format, e.g., '2025-01-31' or '2025-01-31T23:59:59Z') + token: Optional GitHub token for higher rate limits + branch: Optional branch name (defaults to repository's default branch) + + Returns: + List of commit dictionaries + """ + headers = { + "Accept": "application/vnd.github.v3+json", + } + if token: + headers["Authorization"] = f"token {token}" + + base_url = f"https://api.github.com/repos/{owner}/{repo}" + per_page = 100 + + # Normalize date format - add time if not present + if len(since) == 10: # YYYY-MM-DD format + since = f"{since}T00:00:00Z" + if len(until) == 10: # YYYY-MM-DD format + until = f"{until}T23:59:59Z" + + print(f"\nFetching commits from {since} to {until}...") + if branch: + print(f" Branch: {branch}") + + all_commits = [] + page = 1 + + while True: + params = {"since": since, "until": until, "per_page": per_page, "page": page} + if branch: + params["sha"] = branch + + response = requests.get(f"{base_url}/commits", headers=headers, params=params) + + if response.status_code != 200: + raise Exception(f"Failed to fetch commits: {response.text}") + + commits = response.json() + if not commits: + break + + all_commits.extend(commits) + print( + f" Page {page}: fetched {len(commits)} commits (total: {len(all_commits)})" + ) + + if len(commits) < per_page: + break + + page += 1 + + print(f" Completed: {len(all_commits)} commits fetched") + return all_commits + + +def get_merge_base( + base_url: str, base_sha: str, head_sha: str, headers: dict +) -> str | None: + """ + Get the merge base (common ancestor) of two commits. + + Args: + base_url: GitHub API base URL for the repo + base_sha: First commit SHA + head_sha: Second commit SHA + headers: Request headers + + Returns: + Merge base commit SHA, or None if not found + """ + # GitHub Compare API returns merge_base_commit + compare_resp = requests.get( + f"{base_url}/compare/{base_sha}...{head_sha}", + headers=headers, + ) + + if compare_resp.status_code != 200: + return None + + compare_data = compare_resp.json() + merge_base = compare_data.get("merge_base_commit", {}).get("sha") + return merge_base + + +def fetch_commits_by_walking_history( + base_url: str, + base_sha: str, + head_sha: str, + base_tag: str, + head_tag: str, + headers: dict, + stop_sha: str | None = None, +) -> list[dict]: + """ + Fetch commits by walking the commit history from head to a stop point. + + This method correctly handles release branches with cherry-picks. + It walks the head's commit history until it reaches the stop commit. + + Args: + base_url: GitHub API base URL for the repo + base_sha: Base commit SHA (for display purposes) + head_sha: Head commit SHA (newer) + base_tag: Display name for base reference + head_tag: Display name for head reference + headers: Request headers + stop_sha: SHA to stop at (if None, uses base_sha) + + Returns: + List of commit dictionaries (excluding stop commit) + """ + per_page = 100 + all_commits = [] + page = 1 + target_sha = stop_sha or base_sha + + print(f"\nWalking commit history from {head_tag} back to {base_tag}...") + print(f" Stop SHA: {target_sha[:8]}") + + while True: + response = requests.get( + f"{base_url}/commits", + headers=headers, + params={"sha": head_sha, "per_page": per_page, "page": page}, + ) + + if response.status_code != 200: + print(f" Warning: API error on page {page}, stopping") + break + + commits = response.json() + if not commits: + print(f" No more commits found on page {page}") + break + + found_stop = False + for commit in commits: + if commit["sha"] == target_sha: + # Reached stop commit, stop (don't include it) + found_stop = True + break + all_commits.append(commit) + + print( + f" Page {page}: fetched {len(commits)} commits (total: {len(all_commits)})" + ) + + if found_stop: + print(f" Reached stop commit ({target_sha[:8]})") + break + + if len(commits) < per_page: + print(" Warning: Reached end of history without finding stop commit") + break + + page += 1 + + return all_commits + + +def fetch_commits_between_tags_fast( + owner: str, + repo: str, + base_tag: str, + head_tag: str, + token: str | None = None, + head_is_commit: bool = False, + base_is_commit: bool = False, +) -> list[dict]: + """ + Fetch all commits between two tags using GitHub Compare API with pagination. + + This properly fetches only the commits between the two tags. + Automatically handles diverged branches (e.g., release branches with cherry-picks) + by falling back to walking the commit history. + + Args: + owner: Repository owner (e.g., 'vllm-project') + repo: Repository name (e.g., 'vllm') + base_tag: Base tag (older, e.g., 'v0.11.2') or commit SHA if base_is_commit=True + head_tag: Head tag (newer, e.g., 'v0.12.0') or commit SHA if head_is_commit=True + token: Optional GitHub token for higher rate limits + head_is_commit: If True, treat head_tag as a commit SHA instead of a tag + base_is_commit: If True, treat base_tag as a commit SHA instead of a tag + + Returns: + List of commit dictionaries + """ + headers = { + "Accept": "application/vnd.github.v3+json", + } + if token: + headers["Authorization"] = f"token {token}" + + base_url = f"https://api.github.com/repos/{owner}/{repo}" + per_page = 100 + + # Resolve to commit SHAs + if base_is_commit: + base_sha = resolve_commit_sha(base_url, base_tag, headers) + else: + base_sha = resolve_tag_to_sha(base_url, base_tag, headers) + + if head_is_commit: + head_sha = resolve_commit_sha(base_url, head_tag, headers) + else: + head_sha = resolve_tag_to_sha(base_url, head_tag, headers) + + print(f"\nBase SHA ({base_tag}): {base_sha}") + print(f"Head SHA ({head_tag}): {head_sha}") + + # Use Compare API to check relationship and get commits + print(f"\nComparing {base_tag}...{head_tag}...") + compare_resp = requests.get( + f"{base_url}/compare/{base_sha}...{head_sha}", + headers=headers, + params={"per_page": per_page}, + ) + + if compare_resp.status_code != 200: + raise Exception(f"Failed to compare: {compare_resp.text}") + + compare_data = compare_resp.json() + status = compare_data.get("status", "unknown") + total_commits = compare_data.get("total_commits", 0) + + print(f" Comparison status: {status}") + print(f" Total commits: {total_commits}") + + # Get merge_base for potential fallback + merge_base = compare_data.get("merge_base_commit", {}).get("sha") + + # If branches have diverged (e.g., release branch with cherry-picks), + # we need to filter by PR numbers to avoid duplicates + is_diverged = status == "diverged" + + if is_diverged: + print("\n Branches have diverged (likely a release branch scenario)") + print(" Will filter by PR numbers to handle cherry-picks...") + if merge_base: + print(f" Merge base: {merge_base[:8]}") + + # Use Compare API results + all_commits = compare_data.get("commits", []) + print(f" Initial fetch: {len(all_commits)} commits") + + if len(all_commits) >= total_commits: + print(" All commits fetched in initial response") + return all_commits + + # Need to paginate - try Compare API pagination first + page = 1 + while len(all_commits) < total_commits: + page += 1 + print(f" Fetching page {page}...") + + compare_resp = requests.get( + f"{base_url}/compare/{base_sha}...{head_sha}", + headers=headers, + params={"per_page": per_page, "page": page}, + ) + + if compare_resp.status_code != 200: + # Compare API doesn't support pagination well for large diffs + print(" Compare API pagination not supported, using commit walk...") + break + + page_data = compare_resp.json() + page_commits = page_data.get("commits", []) + + if not page_commits: + break + + all_commits.extend(page_commits) + print( + f" Page {page}: got {len(page_commits)} commits (total: {len(all_commits)})" + ) + + # If we still don't have all commits, walk the history + if len(all_commits) < total_commits: + print( + f"\n Need to fetch remaining {total_commits - len(all_commits)} commits via history walk..." + ) + + # For diverged branches, use merge_base as stop point + # For non-diverged, use base_sha + stop_sha = merge_base if (status == "diverged" and merge_base) else base_sha + + # Get commits we already have + seen_shas = {c["sha"] for c in all_commits} + + # Walk from head, collecting commits not already seen, until we reach stop point + walk_commits = [] + walk_page = 1 + found_stop = False + + while len(all_commits) + len(walk_commits) < total_commits and not found_stop: + response = requests.get( + f"{base_url}/commits", + headers=headers, + params={"sha": head_sha, "per_page": per_page, "page": walk_page}, + ) + + if response.status_code != 200: + print(f" Warning: API error on page {walk_page}") + break + + commits = response.json() + if not commits: + break + + for commit in commits: + sha = commit["sha"] + if sha == stop_sha: + found_stop = True + break + if sha not in seen_shas: + seen_shas.add(sha) + walk_commits.append(commit) + + print( + f" Walk page {walk_page}: found {len(walk_commits)} additional commits" + ) + walk_page += 1 + + # Combine: Compare API commits first (they're in order), then walk commits + # Actually, we should return all unique commits + all_commits.extend(walk_commits) + print(f" Total after walk: {len(all_commits)} commits") + + # For diverged branches, filter out commits whose PRs are already in base release + # This handles cherry-picks that exist in both releases + if is_diverged: + print(f"\n Filtering out PRs already in {base_tag}...") + + # Get base release commits to extract PR numbers + print(f" Fetching {base_tag} commits...") + base_commits = [] + base_page = 1 + while True: + response = requests.get( + f"{base_url}/commits", + headers=headers, + params={"sha": base_sha, "per_page": per_page, "page": base_page}, + ) + if response.status_code != 200: + break + commits = response.json() + if not commits: + break + + for commit in commits: + if merge_base and commit["sha"] == merge_base: + break + base_commits.append(commit) + else: + base_page += 1 + continue + break + + print(f" Found {len(base_commits)} commits in {base_tag}") + + # Extract PR numbers from base commits + base_pr_numbers = set() + for commit in base_commits: + message = commit.get("commit", {}).get("message", "") + pr_num = extract_pr_number(message) + if pr_num: + base_pr_numbers.add(pr_num) + print(f" Found {len(base_pr_numbers)} unique PRs in {base_tag}") + + # Filter out commits whose PR is already in base + filtered_commits = [] + skipped_count = 0 + for commit in all_commits: + message = commit.get("commit", {}).get("message", "") + pr_num = extract_pr_number(message) + if pr_num and pr_num in base_pr_numbers: + skipped_count += 1 + continue + filtered_commits.append(commit) + + print(f" Skipped {skipped_count} commits (PRs already in {base_tag})") + print(f" Final count: {len(filtered_commits)} new commits in {head_tag}") + return filtered_commits + + return all_commits + + +def extract_contributors(commits: list[dict]) -> dict: + """ + Extract unique contributors from commits. + + Returns a dict with: + - contributors: set of (login, name) tuples + - by_login: dict mapping login -> contributor info + - by_email: dict mapping email -> contributor info (for commits without GitHub user) + """ + contributors_by_login = {} + contributors_by_email = {} + + for commit in commits: + # Try to get GitHub user info first (author field) + author = commit.get("author") + if author and author.get("login"): + login = author["login"] + if login not in contributors_by_login: + contributors_by_login[login] = { + "login": login, + "name": commit.get("commit", {}).get("author", {}).get("name", ""), + "email": commit.get("commit", {}) + .get("author", {}) + .get("email", ""), + "avatar_url": author.get("avatar_url", ""), + "html_url": author.get("html_url", ""), + "commits": 0, + } + contributors_by_login[login]["commits"] += 1 + else: + # Fallback to git author info + git_author = commit.get("commit", {}).get("author", {}) + email = git_author.get("email", "") + name = git_author.get("name", "") + + if email and email not in contributors_by_email: + contributors_by_email[email] = { + "login": None, + "name": name, + "email": email, + "avatar_url": "", + "html_url": "", + "commits": 0, + } + if email: + contributors_by_email[email]["commits"] += 1 + + return { + "by_login": contributors_by_login, + "by_email": contributors_by_email, + "total": len(contributors_by_login) + len(contributors_by_email), + } + + +def get_tag_date(base_url: str, tag: str, headers: dict) -> str: + """Get the date of a tag's commit.""" + # First resolve the tag to a commit SHA + tag_resp = requests.get(f"{base_url}/git/refs/tags/{tag}", headers=headers) + if tag_resp.status_code != 200: + return None + + tag_data = tag_resp.json() + sha = tag_data["object"]["sha"] + + # If it's an annotated tag, get the underlying commit + if tag_data["object"]["type"] == "tag": + tag_obj_resp = requests.get(f"{base_url}/git/tags/{sha}", headers=headers) + if tag_obj_resp.status_code == 200: + sha = tag_obj_resp.json()["object"]["sha"] + + # Get the commit date + commit_resp = requests.get(f"{base_url}/commits/{sha}", headers=headers) + if commit_resp.status_code == 200: + return commit_resp.json()["commit"]["committer"]["date"] + + return None + + +def check_contributor_is_new( + owner: str, repo: str, login: str, before_date: str, headers: dict +) -> bool: + """ + Check if a contributor has any commits before a given date. + + Returns True if this is their first contribution (no commits before the date). + """ + base_url = f"https://api.github.com/repos/{owner}/{repo}" + + # Search for commits by this author before the base tag date + response = requests.get( + f"{base_url}/commits", + headers=headers, + params={"author": login, "until": before_date, "per_page": 1}, + ) + + if response.status_code == 200: + commits = response.json() + # If no commits found before the date, they're a new contributor + return len(commits) == 0 + + return False + + +def find_first_contribution(commits: list[dict], login: str) -> dict | None: + """ + Find the first (earliest) contribution by a user in the commit list. + + Returns the commit dict or None. + """ + user_commits = [] + for commit in commits: + author = commit.get("author") + if author and author.get("login") == login: + user_commits.append(commit) + + # Commits are usually newest first, so reverse to get oldest first + if user_commits: + return user_commits[-1] # Last one is the oldest/first contribution + return None + + +def calculate_new_contributors_via_generate_notes( + owner: str, + repo: str, + base_tag: str, + head_tag: str, + token: str | None = None, +) -> list[dict]: + """ + Calculate new contributors using GitHub's generate-notes API. + + This is more accurate than checking commit history because GitHub + tracks contributor status internally. + + Args: + owner: Repository owner + repo: Repository name + base_tag: The base tag (older version) + head_tag: The head tag (newer version) + token: GitHub token + + Returns: + List of new contributor info dicts with login and first_pr fields + """ + import re + import subprocess + + print("\nGetting new contributors via GitHub generate-notes API...") + + # Use gh CLI to call the generate-notes API + cmd = [ + "gh", "api", f"repos/{owner}/{repo}/releases/generate-notes", + "-f", f"tag_name={head_tag}", + "-f", f"target_commitish={head_tag}", + "-f", f"previous_tag_name={base_tag}", + "--jq", ".body" + ] + + try: + result = subprocess.run(cmd, capture_output=True, text=True, timeout=60) + if result.returncode != 0: + print(f" Warning: gh CLI failed: {result.stderr}") + return [] + + body = result.stdout + + # Parse new contributors from the generated notes + # Format: "* @username made their first contribution in https://github.com/owner/repo/pull/12345" + pattern = r'\* @(\S+) made their first contribution in https://github\.com/[^/]+/[^/]+/pull/(\d+)' + matches = re.findall(pattern, body) + + new_contributors = [] + for login, pr_number in matches: + new_contributors.append({ + "login": login, + "first_pr": pr_number, + }) + + print(f" Found {len(new_contributors)} new contributors") + return new_contributors + + except subprocess.TimeoutExpired: + print(" Warning: gh CLI timed out") + return [] + except FileNotFoundError: + print(" Warning: gh CLI not found, falling back to legacy method") + return [] + + +def calculate_new_contributors( + commits: list[dict], + current_contributors: dict, + owner: str, + repo: str, + base_tag: str, + head_tag: str = "", + token: str | None = None, +) -> list[dict]: + """ + Calculate which contributors are new (first-time) in this release. + + First tries GitHub's generate-notes API (more accurate), then falls back + to checking commit history if that fails. + + Args: + commits: List of commits in the current release + current_contributors: Output from extract_contributors() + owner: Repository owner + repo: Repository name + base_tag: The base tag (older version) + head_tag: The head tag (newer version) + token: GitHub token + + Returns: + List of new contributor info dicts with first_pr field + """ + # Try the accurate method first (via generate-notes API) + if head_tag: + new_contributors = calculate_new_contributors_via_generate_notes( + owner=owner, + repo=repo, + base_tag=base_tag, + head_tag=head_tag, + token=token, + ) + if new_contributors: + return new_contributors + + # Fall back to legacy method (checking commit history) + print("\nFalling back to legacy new contributor detection...") + + headers = { + "Accept": "application/vnd.github.v3+json", + } + if token: + headers["Authorization"] = f"token {token}" + + base_url = f"https://api.github.com/repos/{owner}/{repo}" + + # Get the date of the base tag + print("Getting base tag date...") + base_date = get_tag_date(base_url, base_tag, headers) + if not base_date: + print(f" Warning: Could not get date for tag {base_tag}") + return [] + + print(f" Base tag date: {base_date}") + + new_contributors = [] + logins = list(current_contributors["by_login"].keys()) + total = len(logins) + + print(f"\nChecking {total} contributors for first-time status...") + + for i, login in enumerate(logins): + if (i + 1) % 20 == 0: + print(f" Checked {i + 1}/{total} contributors...") + + is_new = check_contributor_is_new(owner, repo, login, base_date, headers) + + if is_new: + info = current_contributors["by_login"][login].copy() + + # Find their first PR in this release + first_commit = find_first_contribution(commits, login) + if first_commit: + message = first_commit.get("commit", {}).get("message", "") + pr_number = extract_pr_number(message) + info["first_pr"] = pr_number + info["first_commit_sha"] = first_commit.get("sha", "")[:8] + + new_contributors.append(info) + + print(f" Found {len(new_contributors)} new contributors (legacy method)") + + return new_contributors + + +def generate_contributor_stats( + commits: list[dict], + owner: str, + repo: str, + base_tag: str, + head_tag: str, + token: str | None = None, + check_new: bool = True, +) -> dict: + """ + Generate contributor statistics for the release. + + Returns a dict with all statistics data. + """ + print("\n" + "=" * 60) + print("CONTRIBUTOR STATISTICS") + print("=" * 60) + + # Extract contributors from current commits + contributors = extract_contributors(commits) + + print(f"\nTotal commits: {len(commits)}") + print(f"Total contributors: {contributors['total']}") + print(f" - With GitHub account: {len(contributors['by_login'])}") + print(f" - Without GitHub account (by email): {len(contributors['by_email'])}") + + new_count = 0 + new_contributors_list = [] + + if check_new: + # Calculate new contributors (tries GitHub generate-notes API first, then falls back to commit history) + new_contributors_list = calculate_new_contributors( + commits=commits, + current_contributors=contributors, + owner=owner, + repo=repo, + base_tag=base_tag, + head_tag=head_tag, + token=token, + ) + new_count = len(new_contributors_list) + + print(f"\nNew contributors (first-time): {new_count}") + + if new_contributors_list: + print("\nNew contributors list:") + for c in sorted(new_contributors_list, key=lambda x: x["login"].lower()): + pr_info = f" in #{c['first_pr']}" if c.get("first_pr") else "" + print(f" - @{c['login']} made their first contribution{pr_info}") + + # Print summary line for release notes + print("\n" + "-" * 60) + print("RELEASE NOTES SUMMARY LINE:") + print("-" * 60) + if check_new: + summary_line = f"This release features {len(commits)} commits from {contributors['total']} contributors ({new_count} new)!" + else: + summary_line = f"This release features {len(commits)} commits from {contributors['total']} contributors!" + print(summary_line) + print("-" * 60) + + # Get all contributors sorted by commit count + all_contributors_list = list(contributors["by_login"].values()) + list( + contributors["by_email"].values() + ) + sorted_contributors = sorted( + all_contributors_list, key=lambda x: x["commits"], reverse=True + ) + + # Print top contributors + print("\nTop contributors by commit count:") + for i, c in enumerate(sorted_contributors[:20], 1): + if c.get("login"): + print(f" {i:2}. @{c['login']:20} - {c['commits']:3} commits") + else: + print( + f" {i:2}. {c['name']:20} - {c['commits']:3} commits (no GitHub account)" + ) + + return { + "total_commits": len(commits), + "total_contributors": contributors["total"], + "new_contributors": new_count if check_new else None, + "new_contributors_list": new_contributors_list, + "contributors": contributors, + "sorted_contributors": sorted_contributors, + "summary_line": summary_line, + "base_tag": base_tag, + "head_tag": head_tag, + "owner": owner, + "repo": repo, + } + + +def save_contributor_stats(stats: dict, output_file: str, owner: str, repo: str): + """ + Save contributor statistics to a markdown file. + + Args: + stats: Statistics dict from generate_contributor_stats() + output_file: Output file path + owner: Repository owner + repo: Repository name + """ + lines = [] + + # Header + lines.append(f"# Contributor Statistics: {stats['base_tag']} → {stats['head_tag']}") + lines.append("") + + # Summary for release notes + lines.append("## Release Notes Summary") + lines.append("") + lines.append(f"> {stats['summary_line']}") + lines.append("") + + # Overview stats + lines.append("## Overview") + lines.append("") + lines.append(f"- **Total Commits**: {stats['total_commits']}") + lines.append(f"- **Total Contributors**: {stats['total_contributors']}") + if stats["new_contributors"] is not None: + lines.append(f"- **New Contributors**: {stats['new_contributors']}") + lines.append("") + + # Top contributors table + lines.append("## Top Contributors") + lines.append("") + lines.append("| Rank | Contributor | Commits |") + lines.append("|------|-------------|---------|") + + for i, c in enumerate(stats["sorted_contributors"][:30], 1): + if c.get("login"): + contributor_link = f"[@{c['login']}](https://github.com/{c['login']})" + else: + contributor_link = c["name"] + lines.append(f"| {i} | {contributor_link} | {c['commits']} |") + + lines.append("") + + # New contributors section + if stats["new_contributors_list"]: + lines.append("## New Contributors 🎉") + lines.append("") + + sorted_new = sorted( + stats["new_contributors_list"], key=lambda x: x["login"].lower() + ) + for c in sorted_new: + pr_num = c.get("first_pr") + if pr_num: + pr_link = f"https://github.com/{owner}/{repo}/pull/{pr_num}" + lines.append( + f"* @{c['login']} made their first contribution in {pr_link}" + ) + else: + lines.append(f"* @{c['login']} made their first contribution") + lines.append("") + + # All contributors section (collapsed) + lines.append("## All Contributors") + lines.append("") + lines.append("
") + lines.append("Click to expand full list") + lines.append("") + lines.append("| Contributor | Commits |") + lines.append("|-------------|---------|") + + for c in stats["sorted_contributors"]: + if c.get("login"): + contributor_link = f"[@{c['login']}](https://github.com/{c['login']})" + else: + contributor_link = c["name"] + lines.append(f"| {contributor_link} | {c['commits']} |") + + lines.append("") + lines.append("
") + lines.append("") + + # Write to file + with open(output_file, "w", encoding="utf-8") as f: + f.write("\n".join(lines)) + + print(f"\nSaved contributor statistics to {output_file}") + + +def extract_pr_number(message: str) -> str | None: + """Extract PR number from commit message.""" + # Common patterns: (#12345), (https://github.com/.../pull/12345) + patterns = [ + r"\(#(\d+)\)", # (#12345) + r"pull/(\d+)", # https://github.com/.../pull/12345 + r"#(\d+)$", # #12345 at end + ] + + for pattern in patterns: + match = re.search(pattern, message) + if match: + return match.group(1) + return None + + +def format_commit_message( + commit: dict, + owner: str, + repo: str, + include_sha: bool = False, + include_date: bool = False, +) -> str: + """ + Format a commit message for the output file. + + Format: [Category] Description in https://github.com/owner/repo/pull/XXXX + or: [Category] Description (#XXXX) + + If include_sha is True, prepends the full SHA: `sha` Message (#XXXX) + If include_date is True, prepends the date: [YYYY-MM-DD] Message (#XXXX) + """ + message = commit["commit"]["message"] + sha = commit.get("sha", "") + + # Get commit date (use committer date for when it was merged) + commit_date = "" + if include_date: + date_str = commit.get("commit", {}).get("committer", {}).get("date", "") + if date_str: + # Parse ISO format and extract date part (YYYY-MM-DD) + commit_date = date_str[:10] + + # Get the first line of the commit message + first_line = message.split("\n")[0].strip() + + # Extract PR number if present + pr_number = extract_pr_number(first_line) + + # Clean up the message - remove existing PR references for reformatting + clean_message = first_line + clean_message = re.sub(r"\s*\(#\d+\)\s*$", "", clean_message) + clean_message = re.sub( + r"\s*https://github\.com/[^/]+/[^/]+/pull/\d+\s*", "", clean_message + ) + clean_message = re.sub(r"\s+in\s*$", "", clean_message) + clean_message = clean_message.strip() + + # Format output + if pr_number: + # Check if message already contains the full URL pattern + if f"https://github.com/{owner}/{repo}/pull/" in first_line: + formatted = first_line + else: + formatted = f"{clean_message} (#{pr_number})" + else: + formatted = clean_message + + # Prepend metadata if requested + prefix_parts = [] + if include_date and commit_date: + prefix_parts.append(f"[{commit_date}]") + if include_sha and sha: + prefix_parts.append(f"`{sha}`") + + if prefix_parts: + formatted = f"{' '.join(prefix_parts)} {formatted}" + + return formatted + + +def save_commits_to_file( + commits: list[dict], + output_file: str, + owner: str, + repo: str, + sort_mode: str = "chronological", + include_sha: bool = False, + include_date: bool = False, +): + """ + Save formatted commits to a markdown file. + + Args: + commits: List of commit dictionaries + output_file: Output file path + owner: Repository owner + repo: Repository name + sort_mode: "chronological" (newest first, like GitHub), + "alphabetical" (by commit message), + "reverse" (oldest first) + include_sha: If True, include full commit SHA in output + include_date: If True, include commit date in output + """ + print(f"\nFormatting and saving {len(commits)} commits to {output_file}...") + + formatted_lines = [] + for commit in commits: + formatted = format_commit_message( + commit, owner, repo, include_sha=include_sha, include_date=include_date + ) + formatted_lines.append(formatted) + + # Sort based on mode + if sort_mode == "alphabetical": + formatted_lines.sort(key=lambda x: x.lower()) + print(" Sorted alphabetically by commit message") + elif sort_mode == "reverse": + formatted_lines.reverse() + print(" Sorted chronologically (oldest first)") + else: + # chronological - keep original order (newest first, as returned by API) + print(" Keeping chronological order (newest first)") + + with open(output_file, "w", encoding="utf-8") as f: + for line in formatted_lines: + f.write(line + "\n") + + print(f"Saved {len(formatted_lines)} commits to {output_file}") + + +def main(): + parser = argparse.ArgumentParser( + description="Fetch commits between two GitHub tags or between a tag and a commit" + ) + parser.add_argument( + "--owner", + default="vllm-project", + help="Repository owner (default: vllm-project)", + ) + parser.add_argument( + "--repo", default="vllm", help="Repository name (default: vllm)" + ) + parser.add_argument( + "--base-tag", + help="Base tag (older, e.g., v0.11.2). If not provided with --head-commit, will auto-detect previous tag.", + ) + parser.add_argument( + "--head-tag", + help="Head tag (newer, e.g., v0.12.0). Use this OR --head-commit. If neither specified, uses HEAD of default branch.", + ) + parser.add_argument( + "--head-commit", + help="Head commit SHA (can be short or full). If not specified and no --head-tag, uses HEAD of default branch.", + ) + parser.add_argument( + "--tag-pattern", + default=r"^v\d+\.\d+\.\d+$", + help="Regex pattern to filter tags when auto-detecting previous tag (default: ^v\\d+\\.\\d+\\.\\d+$)", + ) + parser.add_argument( + "--output", + default="0-current-raw-commits.md", + help="Output file (default: 0-current-raw-commits.md)", + ) + parser.add_argument("--token", help="GitHub token (or set GITHUB_TOKEN env var)") + parser.add_argument( + "--slow", + action="store_true", + help="Use slower but more thorough commit-by-commit fetching", + ) + parser.add_argument( + "--sort", + choices=["chronological", "alphabetical", "reverse"], + default="chronological", + help="Sort mode: chronological (newest first, like GitHub), alphabetical (by message), reverse (oldest first)", + ) + parser.add_argument( + "--stats", action="store_true", help="Generate and save contributor statistics" + ) + parser.add_argument( + "--stats-output", + default="0-contributor-stats.md", + help="Output file for contributor statistics (default: 0-contributor-stats.md)", + ) + parser.add_argument( + "--no-new-check", + action="store_true", + help="Skip checking for new contributors (faster, avoids extra API calls)", + ) + parser.add_argument( + "--include-sha", + action="store_true", + help="Include full commit SHA in output (format: `sha` message)", + ) + parser.add_argument( + "--include-date", + action="store_true", + help="Include commit date in output (format: [YYYY-MM-DD] message)", + ) + parser.add_argument( + "--since", + help="Fetch commits since this date (ISO 8601: YYYY-MM-DD or YYYY-MM-DDTHH:MM:SSZ). Use with --until for date range mode.", + ) + parser.add_argument( + "--until", + help="Fetch commits until this date (ISO 8601: YYYY-MM-DD or YYYY-MM-DDTHH:MM:SSZ). Use with --since for date range mode.", + ) + parser.add_argument( + "--branch", + help="Branch to fetch commits from (only used with --since/--until date range mode)", + ) + + args = parser.parse_args() + + # Validate arguments + if args.head_tag and args.head_commit: + parser.error("Cannot specify both --head-tag and --head-commit") + + # Check for date range mode + date_range_mode = args.since is not None or args.until is not None + if date_range_mode: + if not args.since or not args.until: + parser.error( + "Both --since and --until must be specified for date range mode" + ) + if args.head_tag or args.head_commit or args.base_tag: + parser.error( + "Cannot use --since/--until with --head-tag, --head-commit, or --base-tag" + ) + + token = args.token or get_github_token() + + if not token: + print("Warning: No GitHub token provided. Rate limits will be stricter.") + print("Set GITHUB_TOKEN environment variable or use --token argument.") + print() + + headers = { + "Accept": "application/vnd.github.v3+json", + } + if token: + headers["Authorization"] = f"token {token}" + + base_url = f"https://api.github.com/repos/{args.owner}/{args.repo}" + + try: + # Date range mode + if date_range_mode: + print(f"\n{'=' * 60}") + print(f"Fetching commits by date range: {args.since} → {args.until}") + if args.branch: + print(f"Branch: {args.branch}") + print(f"{'=' * 60}") + + commits = fetch_commits_by_date_range( + owner=args.owner, + repo=args.repo, + since=args.since, + until=args.until, + token=token, + branch=args.branch, + ) + + print(f"\nTotal commits found: {len(commits)}") + + save_commits_to_file( + commits=commits, + output_file=args.output, + owner=args.owner, + repo=args.repo, + sort_mode=args.sort, + include_sha=args.include_sha, + include_date=args.include_date, + ) + + # Stats not fully supported in date range mode (no base_tag for new contributor check) + if args.stats: + print( + "\nNote: Contributor statistics in date range mode won't check for new contributors." + ) + stats = generate_contributor_stats( + commits=commits, + owner=args.owner, + repo=args.repo, + base_tag=args.since, + head_tag=args.until, + token=token, + check_new=False, # Can't check new contributors without a base tag + ) + save_contributor_stats( + stats=stats, + output_file=args.stats_output, + owner=args.owner, + repo=args.repo, + ) + + return + + # Tag/commit mode (existing logic) + # Determine head reference + head_is_commit = False + head_ref = None + head_display_name = None + + if args.head_tag: + head_ref = args.head_tag + head_is_commit = False + head_display_name = args.head_tag + elif args.head_commit: + head_ref = args.head_commit + head_is_commit = True + head_display_name = ( + args.head_commit[:8] if len(args.head_commit) > 8 else args.head_commit + ) + else: + # Auto-detect HEAD of default branch + branch_name, head_sha = get_default_branch_head(base_url, headers) + head_ref = head_sha + head_is_commit = True + head_display_name = f"{branch_name} ({head_sha[:8]})" + + base_tag = args.base_tag + base_is_commit = False + + # Auto-detect previous tag if needed + if not base_tag and head_is_commit: + print("Auto-detecting previous tag...") + head_sha = resolve_commit_sha(base_url, head_ref, headers) + + result = find_previous_tag( + base_url=base_url, + head_sha=head_sha, + headers=headers, + tag_pattern=args.tag_pattern, + ) + + if result is None: + raise Exception( + "Could not find a previous tag. Please specify --base-tag manually." + ) + + base_tag, _ = result + print(f"\nUsing auto-detected base tag: {base_tag}") + elif not base_tag: + parser.error("Must specify --base-tag when using --head-tag") + + print(f"\n{'=' * 60}") + print(f"Fetching commits: {base_tag} → {head_display_name}") + print(f"{'=' * 60}") + + if args.slow: + # Note: slow mode doesn't support commit SHA yet, only tags + if head_is_commit: + print( + "Warning: --slow mode with --head-commit not fully supported, using fast mode" + ) + commits = fetch_commits_between_tags_fast( + owner=args.owner, + repo=args.repo, + base_tag=base_tag, + head_tag=head_ref, + token=token, + head_is_commit=head_is_commit, + base_is_commit=base_is_commit, + ) + else: + commits = fetch_commits_between_tags_fast( + owner=args.owner, + repo=args.repo, + base_tag=base_tag, + head_tag=head_ref, + token=token, + head_is_commit=head_is_commit, + base_is_commit=base_is_commit, + ) + + print(f"\nTotal commits found: {len(commits)}") + + save_commits_to_file( + commits=commits, + output_file=args.output, + owner=args.owner, + repo=args.repo, + sort_mode=args.sort, + include_sha=args.include_sha, + include_date=args.include_date, + ) + + # Generate and save contributor statistics if requested + if args.stats: + stats = generate_contributor_stats( + commits=commits, + owner=args.owner, + repo=args.repo, + base_tag=base_tag, + head_tag=head_display_name, + token=token, + check_new=not args.no_new_check, + ) + save_contributor_stats( + stats=stats, + output_file=args.stats_output, + owner=args.owner, + repo=args.repo, + ) + + except Exception as e: + print(f"Error: {e}") + raise + + +if __name__ == "__main__": + main()