[PERF] Decouple projections from GDN custom op #27512

vadiklyutiy · 2025-10-25T20:46:47Z

Purpose

This PR is refactoring of GDN.

The main goal is to allow wider using of torch.compile.

Separated forward pass of GDN attention into three distinct pieces: Input Projection, Core Attention, Output Projection. Before projections was in the GDN custom op and were not covered by torch.compile.
Added RMSNormGated class that implements torch native gated rmsnorm and use it for GDN. torch.compile creates a good code for RMSNormGated even better than custom triton kernel used before.

Functional Test Result

lm_eval
Before

|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|↑  |0.8491|±  |0.0099|
|     |       |strict-match    |     5|exact_match|↑  |0.8059|±  |0.0109|

After

|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value |   |Stderr|
|-----|------:|----------------|-----:|-----------|---|-----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|↑  |0.8544|±  |0.0097|
|     |       |strict-match    |     5|exact_match|↑  |0.8127|±  |0.0107|

Perf Test Result

Server

VLLM_ATTENTION_BACKEND=FLASH_ATTN VLLM_USE_FLASHINFER_MOE_FP16=1 vllm serve Qwen/Qwen3-Next-80B-A3B-Instruct -tp 4 --enable-expert-parallel --no-enable-prefix-caching --async-scheduling --max_cudagraph_capture_size=2048

Prefill

vllm bench serve --backend vllm --model Qwen/Qwen3-Next-80B-A3B-Instruct --endpoint /v1/completions --dataset-name random --random-input 8192 --random-output 1 --max-concurrency 512 --num-prompt 512 --ignore-eos

Before: Total Token throughput (tok/s): 104098.78
After: Total Token throughput (tok/s): 105270.70
Speedup: 1.1%

Decode1

vllm bench serve --backend vllm --model Qwen/Qwen3-Next-80B-A3B-Instruct --endpoint /v1/completions --dataset-name random --random-input 32 --random-output 1024 --max-concurrency 512 --num-prompt 512 --ignore-eos

Before: Output token throughput (tok/s): 19212.17
After: Output token throughput (tok/s): 22384.37
Speedup: 16.5%

Decode2

vllm bench serve --backend vllm --model Qwen/Qwen3-Next-80B-A3B-Instruct --endpoint /v1/completions --dataset-name random --random-input 32 --random-output 1024 --max-concurrency 1024 --num-prompt 1024 --ignore-eos

Before: Output token throughput (tok/s): 28821.37
After: Output token throughput (tok/s): 30298.90
Speed up: 5.1%

Decode3
Server

VLLM_ATTENTION_BACKEND=FLASH_ATTN VLLM_USE_FLASHINFER_MOE_FP16=1 vllm serve Qwen/Qwen3-Next-80B-A3B-Instruct -tp 4 --enable-expert-parallel --no-enable-prefix-caching --async-scheduling

(without increasing --max_cudagraph_capture_size)

vllm bench serve --backend vllm --model Qwen/Qwen3-Next-80B-A3B-Instruct --endpoint /v1/completions --dataset-name random --random-input 32 --random-output 1024 --max-concurrency 1024 --num-prompt 1024 --ignore-eos

Before: Output token throughput (tok/s): 16586.93
After: Output token throughput (tok/s): 18953.92
Speed up: 14.3%

gemini-code-assist

Code Review

This pull request refactors the Gated Delta Net (GDN) attention mechanism to improve torch.compile compatibility and performance. By decoupling the input/output projections from the core custom operator and introducing a native PyTorch RMSNormGated layer, the changes yield significant decode throughput improvements. The refactoring is well-executed and the code is clear. I have one high-severity suggestion regarding a local import in a performance-critical path, which should be moved to the top level of the module to adhere to best practices and avoid potential overhead.

vllm/model_executor/layers/layernorm.py

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

vllm/model_executor/models/qwen3_next.py

ZJY0516 · 2025-10-27T09:07:50Z

CC @heheda12345

vadiklyutiy · 2025-10-28T23:09:58Z

@ALL
Could you please take a look at this PR

mergify · 2025-10-30T19:11:58Z

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @vadiklyutiy.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

ProExpertProg

LGTM but I'll let someone more familiar with Qwen3 approve

vllm/model_executor/layers/layernorm.py

vadiklyutiy · 2025-10-31T00:46:07Z

LGTM but I'll let someone more familiar with Qwen3 approve

is there someone familiar with Qwen3-next except @sighingnow ?

mgoin · 2025-10-31T01:58:43Z

cc @tlrmchlsmth

heheda12345

LGTM!

heheda12345 · 2025-10-31T06:01:08Z

@codex review

chatgpt-codex-connector · 2025-10-31T06:06:53Z

Codex Review: Didn't find any major issues. Delightful!

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

youkaichao

looks good! cc @zhiyuan1i we should be able to do similar optimization for kimi linear.

ZJY0516 · 2025-10-31T10:50:51Z

looks good! cc @zhiyuan1i we should be able to do similar optimization for kimi linear.

I have made a pr #27871 for kimi.

vadiklyutiy · 2025-11-01T16:38:37Z

Regarding CI fails. It looks like they are not related to this PR and a lot of latest commits to fail also has similar fails.
Is it ok to merge this PR?

heheda12345 · 2025-11-01T22:46:25Z

@codex review

chatgpt-codex-connector · 2025-11-01T22:51:10Z

Codex Review: Didn't find any major issues. Hooray!

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

vadiklyutiy · 2025-11-03T20:26:53Z

CI results are weird. I can't repeat them locally.
Try to disable some parts to localize

heheda12345 · 2025-11-03T22:51:52Z

Can you fix the pre-commit and rebase your PR from main? There are many CI fixes on main branch these days.

Signed-off-by: Vadim Gimpelson <[email protected]>

vadiklyutiy · 2025-11-04T10:37:13Z

Can you fix the pre-commit and rebase your PR from main? There are many CI fixes on main branch these days.

Rebased. Still the same and still can't reproduce locally :(

…27512)" This reverts commit 5fd8f02.

…7512) Signed-off-by: Vadim Gimpelson <[email protected]>

Signed-off-by: Vadim Gimpelson <[email protected]>

* add fault_report_addr in FaultToleranceConfig * add handle fault&get_fault_info api Signed-off-by: w00689259 <[email protected]> * remove fault_report_address in CoreEngineActorManager __init__ Signed-off-by: a798347923 <[email protected]> * ruff format Signed-off-by: a798347923 <[email protected]> * add handle fault&get_fault_info api Signed-off-by: w00689259 <[email protected]> * fix one bug. Signed-off-by: fangyuchu <[email protected]> * add fault_report_port in FaultToleranceConfig Signed-off-by: a798347923 <[email protected]> * add zmq_addr concatenate with fault_report_addr and fault_report_port Signed-off-by: a798347923 <[email protected]> * fault reporter bug fix Signed-off-by: w00689259 <[email protected]> * fault reporter bug fix Signed-off-by: w00689259 <[email protected]> * fault reporter bug fix Signed-off-by: w00689259 <[email protected]> * fault reporter bug fix Signed-off-by: w00689259 <[email protected]> * fault reporter bug fix Signed-off-by: w00689259 <[email protected]> * fault reporter bug fix Signed-off-by: w00689259 <[email protected]> * fix some bug * fault reporter bug fix Signed-off-by: w00689259 <[email protected]> * fault reporter bug fix Signed-off-by: w00689259 <[email protected]> * remove fault_report_addr in FaultToleranceConfig Signed-off-by: a798347923 <[email protected]> * refactor: relocate method serialization functions to serial_util.py Signed-off-by: fangyuchu <[email protected]> * fix actor bug * fix actor bug * add engine_core_cmd_addr in FaultToleranceConfig Signed-off-by: a798347923 <[email protected]> * add and use _stop_worker_execution in EngineCoreGuard Signed-off-by: a798347923 <[email protected]> * add and use run in WorkerGuard Signed-off-by: a798347923 <[email protected]> * fix actor bug * fix bug * fix sentinel * fix bug vllm/v1/engine/core.py:847: error: Missing positional argument "tp_size" in call to "EngineCoreGuard" Signed-off-by: a798347923 <[email protected]> * fix bug error: Missing positional arguments "length", "byteorder" in call to "to_bytes" of "int" Signed-off-by: a798347923 <[email protected]> * fix bug in fault tolerance mode Signed-off-by: w00689259 <[email protected]> * fix bug in fault tolerance mode Signed-off-by: w00689259 <[email protected]> * change fault_report_port to internal_fault_report_port add external_fault_notify_port Signed-off-by: a798347923 <[email protected]> * change fault_report_port to internal_fault_report_port add external_fault_notify_port Signed-off-by: a798347923 <[email protected]> * add _recv_cmd func use deserialize_method_call and run_method in run func Signed-off-by: a798347923 <[email protected]> * Update core.py fix bug error: Need type annotation for "kwargs" (hint: "kwargs: dict[<type>, <type>] = ...") Signed-off-by: a798347923 <[email protected]> * add self.ctx.term() in shutdown() Signed-off-by: a798347923 <[email protected]> * changed import deserialize_method_call,serialize_method_call Signed-off-by: a798347923 <[email protected]> * changed init worker_guard in init_device Signed-off-by: a798347923 <[email protected]> * Update core.py add import serialize_method_call Signed-off-by: a798347923 <[email protected]> * Update gpu_worker.py changed init WorkerGuard in init_device Signed-off-by: a798347923 <[email protected]> * Update gpu_worker.py FIX BUG self.worker_guard: WorkerGuard|None = None Signed-off-by: a798347923 <[email protected]> * Update gpu_worker.py fix bug error: Argument 1 to "deserialize_method_call" has incompatible type "str | None"; expected "str" [arg-type] Signed-off-by: a798347923 <[email protected]> * Update gpu_worker.py ruff format Signed-off-by: a798347923 <[email protected]> * Update core.py ruff-format Signed-off-by: a798347923 <[email protected]> * actively send exception information Signed-off-by: w00689259 <[email protected]> * actively send exception information Signed-off-by: w00689259 <[email protected]> * actively send exception information Signed-off-by: w00689259 <[email protected]> * change engine_core_cmd_addr(str) to engine_core_cmd_addrs(list[str]) in EngineZmqAddresses Signed-off-by: a798347923 <[email protected]> * change engine_core_cmd_addr(str) to engine_core_cmd_addrs(list[str]) in EngineZmqAddresses Signed-off-by: a798347923 <[email protected]> * Update utils.py delete engine_core_cmd_addr in EngineZmqAddresses Signed-off-by: a798347923 <[email protected]> * Remove redundant configuration: fault-pub-port Signed-off-by: fangyuchu <[email protected]> * Send pause instructions after receiving fault info in ClientGuard Signed-off-by: fangyuchu <[email protected]> * change engine_core_guard_identities from dict[int, bytes] to list[bytes] Signed-off-by: a798347923 <[email protected]> * fix bug "only the worker guard of engine core 0 can receive messages sent from engine core guard Signed-off-by: a798347923 <[email protected]> * change local_rank to rank_in_group in WorkerGuard Signed-off-by: a798347923 <[email protected]> * changed del self.client_cmd_registry[int(unhealthy_engine.engine_id)] Signed-off-by: a798347923 <[email protected]> * add gloo communication timeout * fix some bug * add stateless_process_group gloo_comm_timeout * reconstruct fault receiver&fault handler Signed-off-by: w00689259 <[email protected]> * fix some bug * reconstruct fault receiver&fault handler Signed-off-by: w00689259 <[email protected]> * reconstruct fault receiver&fault handler Signed-off-by: w00689259 <[email protected]> * fix return format Signed-off-by: w00689259 <[email protected]> * fix return format Signed-off-by: w00689259 <[email protected]> * fix return format Signed-off-by: w00689259 <[email protected]> * add abort request * fix some bug * fix some bug * fix some bug * add dt for client guard Signed-off-by: w00689259 <[email protected]> * add dt for client guard Signed-off-by: w00689259 <[email protected]> * add dt for client guard Signed-off-by: w00689259 <[email protected]> * Implementation of two types of pause: a soft one by using flag signals and a hard one by aborting nccl communicators. Signed-off-by: fangyuchu <[email protected]> * Refine certain log forms and fix a minor bug in pause function. Signed-off-by: fangyuchu <[email protected]> * Refactor and abstract the recv_msg logic in CG,ECG,WG. Signed-off-by: fangyuchu <[email protected]> * [Frontend] Align finish_reason when tool is called with OpenAI (vllm-project#25054) Signed-off-by: Sungyoon Jeong <[email protected]> Co-authored-by: Chauncey <[email protected]> * [Hybrid] Pass kernel block size to builders (vllm-project#27753) Signed-off-by: Thomas Parnell <[email protected]> * [Bugfix] Padded Eagle Specdec with Chunked Prefill (vllm-project#26263) Signed-off-by: Rémi Delacourt <[email protected]> Signed-off-by: Rémi Delacourt <[email protected]> Signed-off-by: remi <[email protected]> Co-authored-by: Benjamin Chislett <[email protected]> * [XPU]Refine Dockerfile.xpu, avoid oneccl dependency issue (vllm-project#27964) Signed-off-by: Kunshang Ji <[email protected]> * Add and check method uuid when sending commands and receiving results. Signed-off-by: fangyuchu <[email protected]> * Add ORCA endpoint load metrics support (vllm-project#24905) Signed-off-by: Misha Efimov <[email protected]> * [CI/Build] Remove the flaky gpt-oss lora test (vllm-project#27966) Signed-off-by: Jee Jee Li <[email protected]> * Abstract the logic of sending instructions and waiting responses from FaultHandler Signed-off-by: fangyuchu <[email protected]> * [Model] Add PaddleOCR-VL Model Support (vllm-project#27758) Signed-off-by: zhangyue <[email protected]> Signed-off-by: Roger Wang <[email protected]> Signed-off-by: Isotr0py <[email protected]> Signed-off-by: zhangyue66 <[email protected]> Co-authored-by: Roger Wang <[email protected]> Co-authored-by: Isotr0py <[email protected]> * Add options in EngineCoreGuard to recv execution results from WorkerGuard Signed-off-by: fangyuchu <[email protected]> * Early exit for MoE LoRA kernels (vllm-project#27131) Signed-off-by: gnovack <[email protected]> Co-authored-by: Jee Jee Li <[email protected]> * [Bugfix] Skip gs:// model paths for speculator detection (vllm-project#27846) Signed-off-by: Peter Schuurman <[email protected]> * [BUG] Make 'binary' default option for saving torch compile artifacts when using standalone_compile (vllm-project#27616) Signed-off-by: ahao-anyscale <[email protected]> * [CI/Testing] Add basic single node dual batch overlap test (vllm-project#27235) Signed-off-by: Lucas Wilkinson <[email protected]> * [Spec Decode] Integrate Suffix Decoding from Arctic Inference (vllm-project#25784) Co-authored-by: Aurick Qiao <[email protected]> * [Feature][Benchmarks] Support `inf` burstiness (vllm-project#26941) Signed-off-by: Sophie du Couédic <[email protected]> * [Bugfix][Qwen][Multimodal] Move Qwen2_5_vl sdpa to custom op and reenable compile (vllm-project#27764) Signed-off-by: Lucas Kabela <[email protected]> * [Bugfix] change FlashMLA reorder_batch_threshold (vllm-project#27777) Signed-off-by: Matthew Bonanni <[email protected]> * [Docs] add runai_streamer_sharded to LoadConfig (vllm-project#27937) Signed-off-by: Andy Xie <[email protected]> * Add TP parameter to attention tests (vllm-project#27683) Signed-off-by: Matthew Bonanni <[email protected]> * [Bugfix][plugin] fla crash on plugin (vllm-project#27322) * [Bugfix] Fix MoE Routing Simulation (vllm-project#28002) Signed-off-by: Tyler Michael Smith <[email protected]> * Remove the tpu docker image nightly build. (vllm-project#27997) Signed-off-by: Qiliang Cui <[email protected]> * [Bugfix][ROCm] Fix ViT rotary embeddings for torch.compile compatibility on ROCm (vllm-project#27748) Signed-off-by: vllmellm <[email protected]> * [LoRA] Lora shrink swizzle (vllm-project#27694) Signed-off-by: li2haipeng <[email protected]> Signed-off-by: Haipeng Li <[email protected]> Co-authored-by: Jee Jee Li <[email protected]> * [Refactor] Lazy import tool_parser (vllm-project#27974) Signed-off-by: chaunceyjiang <[email protected]> * [NIXL][XPU] Pin NIXL version to 0.7.0 (vllm-project#27849) Signed-off-by: zhenwei-intel <[email protected]> * [Metrics] Enable sleep state metric outside of dev mode (vllm-project#27867) Signed-off-by: Mark McLoughlin <[email protected]> * [Bug] Batch invariant: Fix flash attn MLA `RuntimeError: scheduler_metadata must have shape (metadata_size)` (vllm-project#27884) * [CPU]Improve dynamic 4bit moe performance (vllm-project#27240) Signed-off-by: Zhang Xiangze <[email protected]> * [CI/Build] Update LM Eval Version in AMD CI (vllm-project#27944) Signed-off-by: zhewenli <[email protected]> * [KV Connector] Make KVCacheConfig an explicit constructor argument (vllm-project#27887) Signed-off-by: Mark McLoughlin <[email protected]> * [Model] fix ernie45 reasoning_parser (vllm-project#27973) Signed-off-by: wangyafeng <[email protected]> * [CI/Build] Fix OpenAI API correctness on AMD CI (vllm-project#28022) Signed-off-by: zhewenli <[email protected]> * [BugFix][Performance] Restore flashinfer autotuning for all scenarios (vllm-project#27904) * Support worker reinitialization after hard pause; add task queue in FaultHandler to ensure sequential task execution Signed-off-by: fangyuchu <[email protected]> * resolve conflicts Signed-off-by: w00689259 <[email protected]> * resolve conflicts Signed-off-by: w00689259 <[email protected]> * resolve conflicts Signed-off-by: w00689259 <[email protected]> * Load tuned fused_moe_lora shrink and expand kernel configs separately (vllm-project#27435) Signed-off-by: Yu Gong <[email protected]> Co-authored-by: Jee Jee Li <[email protected]> * resolve conflicts Signed-off-by: w00689259 <[email protected]> * resolve conflicts Signed-off-by: w00689259 <[email protected]> * resolve conflicts Signed-off-by: w00689259 <[email protected]> * Support using Int4PreshuffledTensor after loading (vllm-project#26066) Signed-off-by: Jerry Zhang <[email protected]> * [Core] Enable StatLogger in LLMEngine (vllm-project#28020) Signed-off-by: Zhuohan Li <[email protected]> * [Model][Bugfix] fix pipeline parallelism support for NemotronH (vllm-project#27968) Signed-off-by: Tomer Asida <[email protected]> * [Model] add optimal triton fused moe configs for NemotronH MoE (vllm-project#27967) Signed-off-by: Tomer Asida <[email protected]> * [Kernels] Isolate modular kernel code from FusedMoEMethodBase subclasses. (vllm-project#27123) * [BugFix] Fix incorrect preallocated sampled_token_ids tensor size (vllm-project#28025) Signed-off-by: Nick Hill <[email protected]> * [Perf] SM100 - add swap AB optimization to CUTLASS FP8 GEMM (vllm-project#27284) Signed-off-by: Faqin Zhong <[email protected]> Co-authored-by: Faqin Zhong <[email protected]> Co-authored-by: Michael Goin <[email protected]> * [PERF] Decouple projections from GDN custom op (vllm-project#27512) Signed-off-by: Vadim Gimpelson <[email protected]> * [model] Add support for openPangu_Ultra_MoE (vllm-project#27521) Signed-off-by: yuantao <[email protected]> Signed-off-by: yt0428 <[email protected]> Co-authored-by: Jee Jee Li <[email protected]> * [PerfFix] Avoid separate thread for MP executor shm spin (vllm-project#28012) Signed-off-by: Nick Hill <[email protected]> * [AsyncScheduling] Don't schedule past request max_tokens (vllm-project#27922) Signed-off-by: Nick Hill <[email protected]> * Remove deprecated `--rope-scaling` and `--rope-theta` (vllm-project#28006) Signed-off-by: Harry Mellor <[email protected]> * [ROCm][Perf] New design on ROCm AITER MHA backend Implementation (vllm-project#25763) Signed-off-by: ganyi <[email protected]> * Added disable rule to track files under benchmarks/lib (vllm-project#28048) Signed-off-by: Nadav Kluger <[email protected]> * [Multimodal] Make MediaConnector extensible. (vllm-project#27759) Signed-off-by: Chenheli Hua <[email protected]> * [ROCm] gemm_a16w16 upstreaming (vllm-project#26969) Signed-off-by: Aleksandr Malyshev <[email protected]> Co-authored-by: Aleksandr Malyshev <[email protected]> * Revert "[PERF] Decouple projections from GDN custom op" (vllm-project#28080) Signed-off-by: Vadim Gimpelson <[email protected]> * add engine core ut Signed-off-by: w00689259 <[email protected]> * add engine core ut Signed-off-by: w00689259 <[email protected]> * [Qwen3-Next] MOE configs for A100-SXM4-80GB TP4 TP8 (vllm-project#27740) * [XPU] Add gpt-oss model support for Intel GPU (vllm-project#27786) Signed-off-by: Kunshang Ji <[email protected]> * [CI/Build] Enable some fixed tests in AMD CI (vllm-project#28078) Signed-off-by: zhewenli <[email protected]> * [V0 deprecation] Remove VLLM_USE_V1 usage in most modules (vllm-project#27955) Signed-off-by: wangxiyuan <[email protected]> * [Bugfix] Fix encoder-only model support for transformers backend (vllm-project#28021) Signed-off-by: Isotr0py <[email protected]> Signed-off-by: Harry Mellor <[email protected]> Co-authored-by: Harry Mellor <[email protected]> * [BugFix] Fix DCP Assert (AssertionError: DCP not support reorder_batch_threshold > 1 now.) (vllm-project#28100) Signed-off-by: Lucas Wilkinson <[email protected]> * [Model, Core] Support Granite Speech & LoRA for STT (vllm-project#24455) * [Refactor] Lazy-loaded reasoning_parser (vllm-project#28092) Signed-off-by: chaunceyjiang <[email protected]> * [Refactor] to simplify and extract the shared logic between chat completion and responses (vllm-project#27961) Signed-off-by: chaunceyjiang <[email protected]> * [bugfix] fix wrong `dcp_local_seq_lens` calc (vllm-project#27518) Signed-off-by: Qiu <[email protected]> * [Hybrid allocator + kv connector] revert connector test changes related to hybrid allocator (vllm-project#28011) Signed-off-by: KuntaiDu <[email protected]> * [Misc] fix import error for DeepSeekR1ReasoningParser (vllm-project#28114) Signed-off-by: chaunceyjiang <[email protected]> * Fix excessive logging noise by reducing the log level of the MinimaxM2ToolParser import success message (vllm-project#27635) Signed-off-by: minatoaquaMK2 <[email protected]> * Bugfix: Cutlass FP8 FusedMoE bad scaling factors (vllm-project#27255) Signed-off-by: Amir Klein <[email protected]> Co-authored-by: Tyler Michael Smith <[email protected]> Co-authored-by: Michael Goin <[email protected]> * [Graph Partition][Cache] Use inductor partition ops config (vllm-project#27702) Signed-off-by: Boyuan Feng <[email protected]> * [XPU] Enable custom routing functions in IPEX for Llama4 (vllm-project#28004) Signed-off-by: frost-intel <[email protected]> * add kimi reasoning parser (vllm-project#28128) Signed-off-by: wangzhengtao <[email protected]> Co-authored-by: wangzhengtao <[email protected]> * [DCP] check return_lse for all layers in dcp (vllm-project#27929) Signed-off-by: Chen Zhang <[email protected]> * [BugFix] Support EP/DP + EPLB with MTP (vllm-project#25311) Signed-off-by: ilmarkov <[email protected]> Signed-off-by: Sage Moore <[email protected]> Co-authored-by: Sage Moore <[email protected]> Co-authored-by: Tyler Michael Smith <[email protected]> Co-authored-by: Lucas Wilkinson <[email protected]> * Enabling cooperative multi-gpu tests on multi-gpu nodes (vllm-project#27986) Signed-off-by: Alexei V. Ivanov <[email protected]> * [ROCm][MLA] Support block-size > 1 for AITER MLA backend (vllm-project#27224) Signed-off-by: ganyi <[email protected]> Co-authored-by: wuhuikx <[email protected]> * [Bugfix] Validate custom logits processor xargs for online serving (vllm-project#27560) Signed-off-by: Isotr0py <[email protected]> * [misc] add vLLM Beijing Meetup (vllm-project#28127) Signed-off-by: Jiaju Zhang <[email protected]> * [Kernel] Fuse computation of g and beta for Gated Delta Net (vllm-project#28095) Signed-off-by: zjy0516 <[email protected]> * [Core] add support for reasoning parser plugins (vllm-project#28075) Signed-off-by: walter beller-morales <[email protected]> * [Bugfix] vLLM should check Inductor config for compile cache enablement status (vllm-project#27637) Signed-off-by: Yanan Cao <[email protected]> * [FlashInfer] Avoid FlashInfer block_size 16 + head_size 256 on blackwell (vllm-project#27994) Signed-off-by: Chen Zhang <[email protected]> * [CI]: Add LMCacheConnector Unit Tests (vllm-project#27852) Signed-off-by: Samuel Shen <[email protected]> Co-authored-by: Samuel Shen <[email protected]> Co-authored-by: Yihua Cheng <[email protected]> * [Feature] Extend batch invariant torch.compile to B200 (vllm-project#27856) Signed-off-by: PaulZhang12 <[email protected]> * [Bugfix] Fix Qwen3-Reranker-8B load (vllm-project#28117) Signed-off-by: wang.yuqi <[email protected]> * [Docs] Clean up README_TUNING.md (vllm-project#28088) Signed-off-by: windsonsea <[email protected]> * [Hardware][IBM Z] Optimize s390x Dockerfile (vllm-project#28023) Signed-off-by: Rehan Khan <[email protected]> * [Chore] Remove Nemotron-Nano-VL config copy (vllm-project#28126) Signed-off-by: Isotr0py <[email protected]> * [Docs] Add guide to debugging vLLM-torch.compile integration (vllm-project#28094) Signed-off-by: Richard Zou <[email protected]> * [Feature]: Add corrupted request metric to V1 metrics system. (vllm-project#27306) Signed-off-by: atalhens <[email protected]> * [CI/Build] Update checking logic in cutlass_group_gemm_supported (vllm-project#27948) Signed-off-by: zhewenli <[email protected]> * [CI/Build] Fix `test_defaults_with_usage_context` in AMD CI (vllm-project#27926) Signed-off-by: zhewenli <[email protected]> * [Core][Hybrid allocator + connector 2/n] Unify `remove_skipped_blocks` by `get_last_useful_token` (vllm-project#25431) Signed-off-by: KuntaiDu <[email protected]> * [Debugging] Add annotation for easier trace analysis (vllm-project#22496) * [PERF] Decouple projections from GDN custom op. Attempt 2 (vllm-project#28083) Signed-off-by: Vadim Gimpelson <[email protected]> * [Bug] Fix cpu disable shared_experts `VLLM_DISABLE_SHARED_EXPERTS_STREAM` (vllm-project#28157) Signed-off-by: yewentao256 <[email protected]> * [Bug] Fix env string `"0"` same to `True` (vllm-project#28159) Signed-off-by: yewentao256 <[email protected]> * Ensure WorkerGuard command execution returns result; fix missing set_device when TP>1 Signed-off-by: fangyuchu <[email protected]> * [Feature] Enable TP + EP `shared_experts` overlap with router, 3.7% E2E performance improvement (vllm-project#28164) Signed-off-by: yewentao256 <[email protected]> * [CI Failure] `nm-testing/Qwen2-0.5B-Instruct-FP8-SkipQKV` was removed from HF. Skip it in tests (vllm-project#28170) Signed-off-by: Vadim Gimpelson <[email protected]> * [Misc] Remove the duplicate code (vllm-project#28111) Signed-off-by: chaunceyjiang <[email protected]> * rename& format logger Signed-off-by: w00689259 <[email protected]> * rename& format logger Signed-off-by: w00689259 <[email protected]> * feat(nccl): enable non-blocking NCCL communicators to support ncclCommAbort Signed-off-by: fangyuchu <[email protected]> --------- Signed-off-by: w00689259 <[email protected]> Signed-off-by: a798347923 <[email protected]> Signed-off-by: fangyuchu <[email protected]> Signed-off-by: zWaNg3 <[email protected]> Signed-off-by: a798347923 <[email protected]> Signed-off-by: Sungyoon Jeong <[email protected]> Signed-off-by: Thomas Parnell <[email protected]> Signed-off-by: Rémi Delacourt <[email protected]> Signed-off-by: Rémi Delacourt <[email protected]> Signed-off-by: remi <[email protected]> Signed-off-by: Kunshang Ji <[email protected]> Signed-off-by: Misha Efimov <[email protected]> Signed-off-by: Jee Jee Li <[email protected]> Signed-off-by: zhangyue <[email protected]> Signed-off-by: Roger Wang <[email protected]> Signed-off-by: Isotr0py <[email protected]> Signed-off-by: zhangyue66 <[email protected]> Signed-off-by: gnovack <[email protected]> Signed-off-by: Peter Schuurman <[email protected]> Signed-off-by: ahao-anyscale <[email protected]> Signed-off-by: Lucas Wilkinson <[email protected]> Signed-off-by: Sophie du Couédic <[email protected]> Signed-off-by: Lucas Kabela <[email protected]> Signed-off-by: Matthew Bonanni <[email protected]> Signed-off-by: Andy Xie <[email protected]> Signed-off-by: Tyler Michael Smith <[email protected]> Signed-off-by: Qiliang Cui <[email protected]> Signed-off-by: vllmellm <[email protected]> Signed-off-by: li2haipeng <[email protected]> Signed-off-by: Haipeng Li <[email protected]> Signed-off-by: chaunceyjiang <[email protected]> Signed-off-by: zhenwei-intel <[email protected]> Signed-off-by: Mark McLoughlin <[email protected]> Signed-off-by: Zhang Xiangze <[email protected]> Signed-off-by: zhewenli <[email protected]> Signed-off-by: wangyafeng <[email protected]> Signed-off-by: Yu Gong <[email protected]> Signed-off-by: Jerry Zhang <[email protected]> Signed-off-by: Zhuohan Li <[email protected]> Signed-off-by: Tomer Asida <[email protected]> Signed-off-by: Nick Hill <[email protected]> Signed-off-by: Faqin Zhong <[email protected]> Signed-off-by: Vadim Gimpelson <[email protected]> Signed-off-by: yuantao <[email protected]> Signed-off-by: yt0428 <[email protected]> Signed-off-by: Harry Mellor <[email protected]> Signed-off-by: ganyi <[email protected]> Signed-off-by: Nadav Kluger <[email protected]> Signed-off-by: Chenheli Hua <[email protected]> Signed-off-by: Aleksandr Malyshev <[email protected]> Signed-off-by: wangxiyuan <[email protected]> Signed-off-by: Qiu <[email protected]> Signed-off-by: KuntaiDu <[email protected]> Signed-off-by: minatoaquaMK2 <[email protected]> Signed-off-by: Amir Klein <[email protected]> Signed-off-by: Boyuan Feng <[email protected]> Signed-off-by: frost-intel <[email protected]> Signed-off-by: wangzhengtao <[email protected]> Signed-off-by: Chen Zhang <[email protected]> Signed-off-by: ilmarkov <[email protected]> Signed-off-by: Sage Moore <[email protected]> Signed-off-by: Alexei V. Ivanov <[email protected]> Signed-off-by: Jiaju Zhang <[email protected]> Signed-off-by: zjy0516 <[email protected]> Signed-off-by: walter beller-morales <[email protected]> Signed-off-by: Yanan Cao <[email protected]> Signed-off-by: Samuel Shen <[email protected]> Signed-off-by: PaulZhang12 <[email protected]> Signed-off-by: wang.yuqi <[email protected]> Signed-off-by: windsonsea <[email protected]> Signed-off-by: Rehan Khan <[email protected]> Signed-off-by: Richard Zou <[email protected]> Signed-off-by: atalhens <[email protected]> Signed-off-by: yewentao256 <[email protected]> Co-authored-by: fangyuchu <[email protected]> Co-authored-by: a798347923 <[email protected]> Co-authored-by: w00689259 <[email protected]> Co-authored-by: fangyuchu <[email protected]> Co-authored-by: TianZhuo <[email protected]> Co-authored-by: a798347923 <[email protected]> Co-authored-by: Sungyoon Jeong <[email protected]> Co-authored-by: Chauncey <[email protected]> Co-authored-by: Thomas Parnell <[email protected]> Co-authored-by: Rémi Delacourt <[email protected]> Co-authored-by: Benjamin Chislett <[email protected]> Co-authored-by: Kunshang Ji <[email protected]> Co-authored-by: Misha Efimov <[email protected]> Co-authored-by: Jee Jee Li <[email protected]> Co-authored-by: zhang-prog <[email protected]> Co-authored-by: Roger Wang <[email protected]> Co-authored-by: Isotr0py <[email protected]> Co-authored-by: gnovack <[email protected]> Co-authored-by: pwschuurman <[email protected]> Co-authored-by: ahao-anyscale <[email protected]> Co-authored-by: Lucas Wilkinson <[email protected]> Co-authored-by: Aurick Qiao <[email protected]> Co-authored-by: Aurick Qiao <[email protected]> Co-authored-by: Sophie du Couédic <[email protected]> Co-authored-by: Lucas Kabela <[email protected]> Co-authored-by: Matthew Bonanni <[email protected]> Co-authored-by: Ning Xie <[email protected]> Co-authored-by: Hank_ <[email protected]> Co-authored-by: Tyler Michael Smith <[email protected]> Co-authored-by: QiliangCui <[email protected]> Co-authored-by: vllmellm <[email protected]> Co-authored-by: li2haipeng <[email protected]> Co-authored-by: liuzhenwei <[email protected]> Co-authored-by: Mark McLoughlin <[email protected]> Co-authored-by: Wentao Ye <[email protected]> Co-authored-by: xiangze-arm <[email protected]> Co-authored-by: Zhewen Li <[email protected]> Co-authored-by: CSWYF3634076 <[email protected]> Co-authored-by: Varun Sundar Rabindranath <[email protected]> Co-authored-by: yugong333 <[email protected]> Co-authored-by: Jerry Zhang <[email protected]> Co-authored-by: Zhuohan Li <[email protected]> Co-authored-by: tomeras91 <[email protected]> Co-authored-by: bnellnm <[email protected]> Co-authored-by: Nick Hill <[email protected]> Co-authored-by: lyrisz <[email protected]> Co-authored-by: Faqin Zhong <[email protected]> Co-authored-by: Michael Goin <[email protected]> Co-authored-by: Vadim Gimpelson <[email protected]> Co-authored-by: yt0428 <[email protected]> Co-authored-by: Harry Mellor <[email protected]> Co-authored-by: Pleaplusone <[email protected]> Co-authored-by: nadavkluger <[email protected]> Co-authored-by: Chenheli Hua <[email protected]> Co-authored-by: Aleksandr Malyshev <[email protected]> Co-authored-by: Aleksandr Malyshev <[email protected]> Co-authored-by: tou <[email protected]> Co-authored-by: wangxiyuan <[email protected]> Co-authored-by: Alex Brooks <[email protected]> Co-authored-by: Qiu <[email protected]> Co-authored-by: Kuntai Du <[email protected]> Co-authored-by: Eric Yue <[email protected]> Co-authored-by: amirkl94 <[email protected]> Co-authored-by: Boyuan Feng <[email protected]> Co-authored-by: Frost Mitchell <[email protected]> Co-authored-by: bigmoyan <[email protected]> Co-authored-by: wangzhengtao <[email protected]> Co-authored-by: Chen Zhang <[email protected]> Co-authored-by: Ilya Markov <[email protected]> Co-authored-by: Sage Moore <[email protected]> Co-authored-by: Alexei-V-Ivanov-AMD <[email protected]> Co-authored-by: wuhuikx <[email protected]> Co-authored-by: Jiaju Zhang <[email protected]> Co-authored-by: Jiangyun Zhu <[email protected]> Co-authored-by: Walter Beller-Morales <[email protected]> Co-authored-by: gmagogsfm <[email protected]> Co-authored-by: Samuel Shen <[email protected]> Co-authored-by: Samuel Shen <[email protected]> Co-authored-by: Yihua Cheng <[email protected]> Co-authored-by: Paul Zhang <[email protected]> Co-authored-by: wang.yuqi <[email protected]> Co-authored-by: Michael Yao <[email protected]> Co-authored-by: R3hankhan <[email protected]> Co-authored-by: Richard Zou <[email protected]> Co-authored-by: Snehlata <[email protected]> Co-authored-by: Dayeol Lee <[email protected]>

Signed-off-by: Vadim Gimpelson <[email protected]>

vadiklyutiy requested review from ProExpertProg, WoosukKwon, hmellor, houseroad, mgoin, robertgshaw2-redhat, sighingnow, simon-mo, tlrmchlsmth, yewentao256 and youkaichao as code owners October 25, 2025 20:46

mergify bot added the qwen Related to Qwen models label Oct 25, 2025

gemini-code-assist bot reviewed Oct 25, 2025

View reviewed changes

vllm/model_executor/layers/layernorm.py Show resolved Hide resolved

chatgpt-codex-connector bot reviewed Oct 25, 2025

View reviewed changes

vllm/model_executor/models/qwen3_next.py Show resolved Hide resolved

mergify bot added the needs-rebase label Oct 30, 2025

ProExpertProg reviewed Oct 30, 2025

View reviewed changes

vllm/model_executor/layers/layernorm.py Outdated Show resolved Hide resolved

vllm/model_executor/layers/layernorm.py Outdated Show resolved Hide resolved

mergify bot removed the needs-rebase label Oct 31, 2025

heheda12345 approved these changes Oct 31, 2025

View reviewed changes

heheda12345 enabled auto-merge (squash) October 31, 2025 06:07

github-actions bot added the ready ONLY add when PR is ready to merge/full CI is needed label Oct 31, 2025

youkaichao approved these changes Oct 31, 2025

View reviewed changes

youkaichao disabled auto-merge October 31, 2025 13:42

vadiklyutiy force-pushed the vadim/refac-gdn branch 2 times, most recently from 39cd84b to 090a44b Compare November 1, 2025 00:05

heheda12345 enabled auto-merge (squash) November 1, 2025 22:53

sighingnow approved these changes Nov 3, 2025

View reviewed changes

auto-merge was automatically disabled November 3, 2025 09:07
Head branch was pushed to by a user without write access

vadiklyutiy force-pushed the vadim/refac-gdn branch 2 times, most recently from 70d2048 to c3197c6 Compare November 3, 2025 12:21

vadiklyutiy added 4 commits November 4, 2025 11:21

GDN refacoring

8f45f6d

Signed-off-by: Vadim Gimpelson <[email protected]>

fix

c955727

Signed-off-by: Vadim Gimpelson <[email protected]>

Make RMSNormGated a CustomOp

d710a90

Signed-off-by: Vadim Gimpelson <[email protected]>

fix pre-commit

f276f69

Signed-off-by: Vadim Gimpelson <[email protected]>

vadiklyutiy force-pushed the vadim/refac-gdn branch from 7f164be to f276f69 Compare November 4, 2025 07:22

simon-mo merged commit 5fd8f02 into vllm-project:main Nov 4, 2025
54 of 58 checks passed

vadiklyutiy added a commit to CentML/vllm that referenced this pull request Nov 4, 2025

Revert "[PERF] Decouple projections from GDN custom op (vllm-project#…

3d3da84

…27512)" This reverts commit 5fd8f02.

vadiklyutiy mentioned this pull request Nov 4, 2025

Revert "[PERF] Decouple projections from GDN custom op" #28080

Merged

vadiklyutiy added a commit to CentML/vllm that referenced this pull request Nov 4, 2025

Revert [PERF] Decouple projections from GDN custom op (vllm-project#2…

fb128a1

…7512) Signed-off-by: Vadim Gimpelson <[email protected]>

vadiklyutiy deleted the vadim/refac-gdn branch November 5, 2025 00:41

vadiklyutiy mentioned this pull request Nov 5, 2025

[PERF] Decouple projections from GDN custom op. Attempt 2 #28083

Merged

juliendenize pushed a commit to juliendenize/vllm that referenced this pull request Nov 6, 2025

[PERF] Decouple projections from GDN custom op (vllm-project#27512)

b3bc1f8

Signed-off-by: Vadim Gimpelson <[email protected]>

ZhengHongming888 pushed a commit to ZhengHongming888/vllm that referenced this pull request Nov 8, 2025

[PERF] Decouple projections from GDN custom op (vllm-project#27512)

5ff3e3e

Signed-off-by: Vadim Gimpelson <[email protected]>

Uh oh!

[PERF] Decouple projections from GDN custom op #27512

[PERF] Decouple projections from GDN custom op #27512

Conversation

vadiklyutiy commented Oct 25, 2025 • edited by github-actions bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Purpose

Functional Test Result

Perf Test Result

Uh oh!

gemini-code-assist bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

Uh oh!

chatgpt-codex-connector bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

Uh oh!

ZJY0516 commented Oct 27, 2025

Uh oh!

vadiklyutiy commented Oct 28, 2025

Uh oh!

mergify bot commented Oct 30, 2025

Uh oh!

ProExpertProg left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

vadiklyutiy commented Oct 31, 2025

Uh oh!

mgoin commented Oct 31, 2025

Uh oh!

heheda12345 left a comment

Choose a reason for hiding this comment

Uh oh!

heheda12345 commented Oct 31, 2025

Uh oh!

chatgpt-codex-connector bot commented Oct 31, 2025

Uh oh!

youkaichao left a comment

Choose a reason for hiding this comment

Uh oh!

ZJY0516 commented Oct 31, 2025

Uh oh!

vadiklyutiy commented Nov 1, 2025

Uh oh!

heheda12345 commented Nov 1, 2025

Uh oh!

chatgpt-codex-connector bot commented Nov 1, 2025

Uh oh!

vadiklyutiy commented Nov 3, 2025

Uh oh!

heheda12345 commented Nov 3, 2025

Uh oh!

vadiklyutiy commented Nov 4, 2025

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

8 participants

vadiklyutiy commented Oct 25, 2025 •

edited by github-actions bot

Loading