feat: Merge kernels from vLLM and FlashInfer#63
Merged
drunkcoding merged 4 commits intofeature/qwenfrom Feb 16, 2026
Merged
Conversation
drunkcoding
added a commit
that referenced
this pull request
Feb 17, 2026
* add openai api support * add test scripts, update readme, update api * Fix: Undefined Symbol Compilation Error (#37) * reformat code vllm style * add threadsafe queues * fix compilation error --------- Co-authored-by: xly <leyang.xue@ed.ac.uk> * Refactor code for better performance (#38) * reformat code vllm style * add threadsafe queues * fix compilation error * split files and remove queuing * performance improvement * remove error dependency * add try lock return check * fix header dependency * fix hard coded number --------- Co-authored-by: xly <leyang.xue@ed.ac.uk> * CI: add pre commit format ci (#40) * ci: add pre commit format ci * fix: add requirements for linting * fix: format code before merge * fix: update local clang format version * Chore: rename organization name & optimize CI (#41) * reformat code vllm style * add threadsafe queues * fix compilation error * split files and remove queuing * performance improvement * remove error dependency * add try lock return check * fix header dependency * fix hard coded number * update CI using cuda docker image * repo consistency * pr template fix * format doc * delete gpu option, add --no-install-recommends * add cuda matrix and remove cuda full package install * remove publish container * change team name to efficient moe --------- Co-authored-by: xly <leyang.xue@ed.ac.uk> * CI: fix not a git repository in CI (#43) * CI: fix missing sudo in apt install (#44) * CI: fix missing sudo (#45) * CI: revert os matrix in CI (#46) * CI: add missing apt update after installing deb file (#47) * Doc: Update README example to DeepSeek and Suppress Warning (#49) * reformat code vllm style * add threadsafe queues * fix compilation error * split files and remove queuing * performance improvement * remove error dependency * add try lock return check * fix header dependency * fix hard coded number * update CI using cuda docker image * repo consistency * pr template fix * format doc * delete gpu option, add --no-install-recommends * add cuda matrix and remove cuda full package install * remove publish container * change team name to efficient moe * update readme example to deepseek and supress warning * format * revert CI changes to main version --------- Co-authored-by: xly <leyang.xue@ed.ac.uk> * CI: do not build test if document update (#52) * reformat code vllm style * add threadsafe queues * fix compilation error * split files and remove queuing * performance improvement * remove error dependency * add try lock return check * fix header dependency * fix hard coded number * update CI using cuda docker image * repo consistency * pr template fix * format doc * delete gpu option, add --no-install-recommends * add cuda matrix and remove cuda full package install * remove publish container * change team name to efficient moe * update readme example to deepseek and supress warning * format * revert CI changes to main version * update readme conda env and ignore doc update in build and release * fix wildcard --------- Co-authored-by: xly <leyang.xue@ed.ac.uk> * format and change to deepseek in example * fix format * remove unused files * fix api server token id device * feat: Introduce Local Server for OpenAI-Compatible APIs (#4) * update table format * improve table clarity * init code commit * add openai api support * add test scripts, update readme, update api * format and change to deepseek in example * fix format * remove unused files * fix api server token id device --------- Co-authored-by: xly <leyang.xue@ed.ac.uk> Co-authored-by: Yao <fuyao3860@gmail.com> * fix gen broken * update readme links * cancel concurrent job * set dense node to device * sparse node set cpu * remove OS def * use update to date clang-format * fix setuptools version * fix setuptools version for python 3.8 * keep single cuda version in publish * feat: set parameter to device before serving (#56) * update table format * improve table clarity * init code commit * add openai api support * add test scripts, update readme, update api * format and change to deepseek in example * fix format * remove unused files * fix api server token id device * fix gen broken * update readme links * cancel concurrent job * set dense node to device * sparse node set cpu * remove OS def * use update to date clang-format * fix setuptools version * fix setuptools version for python 3.8 * keep single cuda version in publish --------- Co-authored-by: xly <leyang.xue@ed.ac.uk> Co-authored-by: Yao <fuyao3860@gmail.com> * add max length in gen openai * fix cache race condition * all param init at host * add qwen3 * ubuntu lts and build * pre-commit ubuntu version * router weights update overlap * rename deepseek_v2 and reduce torch kernel launch * fix import * fix build and fix bug * fix citation linebreak * fix typo * fix dtype size * remove comments * fix example * pr update init * remove comment and unify deepseek preroute * feat: Merge kernels from vLLM and FlashInfer (#63) * new allocator * add kernel compilation * stable topk --------- Co-authored-by: xly <leyang.xue@ed.ac.uk> --------- Co-authored-by: Yao <fuyao3860@gmail.com> Co-authored-by: xly <leyang.xue@ed.ac.uk> Co-authored-by: Zhan Lu <51200935+lausannel@users.noreply.github.com> Co-authored-by: Yao Fu <yao.fu.aisys@gmail.com>
drunkcoding
added a commit
that referenced
this pull request
Mar 19, 2026
* update table format * improve table clarity * init code commit * doc: add flashattention installation guide and change toc * feat: remove libaio dependency * remove spdlog dependency * misc: remove unused code and dependencies * misc: remove commented-out code and unused imports * fix: cuda oom due to safe tensors open * remove gcc-12 requirement * gptq disable exllama * fix: key error in offload set * add forward and call (#7) * add forward and call * fix a bug * feat: support grok-1 model * update API note and install * Feature/expert parallel (#9) * add back expert parallel by id hash * add grok ep * fix mistral typo * accom cuda copy bug * sync after compute * fix:sync to make sure that input is ready --------- Co-authored-by: xly <leyang.xue@ed.ac.uk> Co-authored-by: luzhan <513964121@qq.com> * fix tokenizer in example * Xly/deepseek (#34) * add override QuantLinear (#29) Co-authored-by: xly <leyang.xue@ed.ac.uk> * use torch streampool * format * working deepspeed backend * fix: revert apply_rotary_pos_emb in deepseek * fix busy waiting * fix deepseek flashattn * add deepseek v3 * format and fix multigpu deepseek bug * with device caching allocator * add on-demand lock cache --------- Co-authored-by: xly <leyang.xue@ed.ac.uk> Co-authored-by: lausannel <513964121@qq.com> * Upstream (#72) * Fix: Undefined Symbol Compilation Error (#37) * reformat code vllm style * add threadsafe queues * fix compilation error --------- Co-authored-by: xly <leyang.xue@ed.ac.uk> * Refactor code for better performance (#38) * reformat code vllm style * add threadsafe queues * fix compilation error * split files and remove queuing * performance improvement * remove error dependency * add try lock return check * fix header dependency * fix hard coded number --------- Co-authored-by: xly <leyang.xue@ed.ac.uk> * CI: add pre commit format ci (#40) * ci: add pre commit format ci * fix: add requirements for linting * fix: format code before merge * fix: update local clang format version * Chore: rename organization name & optimize CI (#41) * reformat code vllm style * add threadsafe queues * fix compilation error * split files and remove queuing * performance improvement * remove error dependency * add try lock return check * fix header dependency * fix hard coded number * update CI using cuda docker image * repo consistency * pr template fix * format doc * delete gpu option, add --no-install-recommends * add cuda matrix and remove cuda full package install * remove publish container * change team name to efficient moe --------- Co-authored-by: xly <leyang.xue@ed.ac.uk> * CI: fix not a git repository in CI (#43) * CI: fix missing sudo in apt install (#44) * CI: fix missing sudo (#45) * CI: revert os matrix in CI (#46) * CI: add missing apt update after installing deb file (#47) * Doc: Update README example to DeepSeek and Suppress Warning (#49) * reformat code vllm style * add threadsafe queues * fix compilation error * split files and remove queuing * performance improvement * remove error dependency * add try lock return check * fix header dependency * fix hard coded number * update CI using cuda docker image * repo consistency * pr template fix * format doc * delete gpu option, add --no-install-recommends * add cuda matrix and remove cuda full package install * remove publish container * change team name to efficient moe * update readme example to deepseek and supress warning * format * revert CI changes to main version --------- Co-authored-by: xly <leyang.xue@ed.ac.uk> * CI: do not build test if document update (#52) * reformat code vllm style * add threadsafe queues * fix compilation error * split files and remove queuing * performance improvement * remove error dependency * add try lock return check * fix header dependency * fix hard coded number * update CI using cuda docker image * repo consistency * pr template fix * format doc * delete gpu option, add --no-install-recommends * add cuda matrix and remove cuda full package install * remove publish container * change team name to efficient moe * update readme example to deepseek and supress warning * format * revert CI changes to main version * update readme conda env and ignore doc update in build and release * fix wildcard --------- Co-authored-by: xly <leyang.xue@ed.ac.uk> * feat: Introduce Local Server for OpenAI-Compatible APIs (#4) * update table format * improve table clarity * init code commit * add openai api support * add test scripts, update readme, update api * format and change to deepseek in example * fix format * remove unused files * fix api server token id device --------- Co-authored-by: xly <leyang.xue@ed.ac.uk> Co-authored-by: Yao <fuyao3860@gmail.com> * feat: set parameter to device before serving (#56) * update table format * improve table clarity * init code commit * add openai api support * add test scripts, update readme, update api * format and change to deepseek in example * fix format * remove unused files * fix api server token id device * fix gen broken * update readme links * cancel concurrent job * set dense node to device * sparse node set cpu * remove OS def * use update to date clang-format * fix setuptools version * fix setuptools version for python 3.8 * keep single cuda version in publish --------- Co-authored-by: xly <leyang.xue@ed.ac.uk> Co-authored-by: Yao <fuyao3860@gmail.com> * Chore(deps): Bump pyarrow from 12.0.0 to 14.0.1 (#69) Bumps [pyarrow](https://github.com/apache/arrow) from 12.0.0 to 14.0.1. - [Release notes](https://github.com/apache/arrow/releases) - [Commits](apache/arrow@go/v12.0.0...go/v14.0.1) --- updated-dependencies: - dependency-name: pyarrow dependency-version: 14.0.1 dependency-type: direct:production ... Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> --------- Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: xly <leyang.xue@ed.ac.uk> Co-authored-by: Zhan Lu <51200935+lausannel@users.noreply.github.com> Co-authored-by: Yao Fu <yao.fu.aisys@gmail.com> Co-authored-by: Yao <fuyao3860@gmail.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> * Audit repository for stale code indicators (#71) * Fix: Undefined Symbol Compilation Error (#37) * reformat code vllm style * add threadsafe queues * fix compilation error --------- Co-authored-by: xly <leyang.xue@ed.ac.uk> * Refactor code for better performance (#38) * reformat code vllm style * add threadsafe queues * fix compilation error * split files and remove queuing * performance improvement * remove error dependency * add try lock return check * fix header dependency * fix hard coded number --------- Co-authored-by: xly <leyang.xue@ed.ac.uk> * CI: add pre commit format ci (#40) * ci: add pre commit format ci * fix: add requirements for linting * fix: format code before merge * fix: update local clang format version * Chore: rename organization name & optimize CI (#41) * reformat code vllm style * add threadsafe queues * fix compilation error * split files and remove queuing * performance improvement * remove error dependency * add try lock return check * fix header dependency * fix hard coded number * update CI using cuda docker image * repo consistency * pr template fix * format doc * delete gpu option, add --no-install-recommends * add cuda matrix and remove cuda full package install * remove publish container * change team name to efficient moe --------- Co-authored-by: xly <leyang.xue@ed.ac.uk> * CI: fix not a git repository in CI (#43) * CI: fix missing sudo in apt install (#44) * CI: fix missing sudo (#45) * CI: revert os matrix in CI (#46) * CI: add missing apt update after installing deb file (#47) * Doc: Update README example to DeepSeek and Suppress Warning (#49) * reformat code vllm style * add threadsafe queues * fix compilation error * split files and remove queuing * performance improvement * remove error dependency * add try lock return check * fix header dependency * fix hard coded number * update CI using cuda docker image * repo consistency * pr template fix * format doc * delete gpu option, add --no-install-recommends * add cuda matrix and remove cuda full package install * remove publish container * change team name to efficient moe * update readme example to deepseek and supress warning * format * revert CI changes to main version --------- Co-authored-by: xly <leyang.xue@ed.ac.uk> * CI: do not build test if document update (#52) * reformat code vllm style * add threadsafe queues * fix compilation error * split files and remove queuing * performance improvement * remove error dependency * add try lock return check * fix header dependency * fix hard coded number * update CI using cuda docker image * repo consistency * pr template fix * format doc * delete gpu option, add --no-install-recommends * add cuda matrix and remove cuda full package install * remove publish container * change team name to efficient moe * update readme example to deepseek and supress warning * format * revert CI changes to main version * update readme conda env and ignore doc update in build and release * fix wildcard --------- Co-authored-by: xly <leyang.xue@ed.ac.uk> * feat: Introduce Local Server for OpenAI-Compatible APIs (#4) * update table format * improve table clarity * init code commit * add openai api support * add test scripts, update readme, update api * format and change to deepseek in example * fix format * remove unused files * fix api server token id device --------- Co-authored-by: xly <leyang.xue@ed.ac.uk> Co-authored-by: Yao <fuyao3860@gmail.com> * feat: set parameter to device before serving (#56) * update table format * improve table clarity * init code commit * add openai api support * add test scripts, update readme, update api * format and change to deepseek in example * fix format * remove unused files * fix api server token id device * fix gen broken * update readme links * cancel concurrent job * set dense node to device * sparse node set cpu * remove OS def * use update to date clang-format * fix setuptools version * fix setuptools version for python 3.8 * keep single cuda version in publish --------- Co-authored-by: xly <leyang.xue@ed.ac.uk> Co-authored-by: Yao <fuyao3860@gmail.com> * Initial plan * Add mypy lint hook Co-authored-by: drunkcoding <14305648+drunkcoding@users.noreply.github.com> * Configure mypy settings Co-authored-by: drunkcoding <14305648+drunkcoding@users.noreply.github.com> * Adjust mypy scope Co-authored-by: drunkcoding <14305648+drunkcoding@users.noreply.github.com> * Scope mypy checks Co-authored-by: drunkcoding <14305648+drunkcoding@users.noreply.github.com> --------- Co-authored-by: Leyang Xue <s2062808@ed.ac.uk> Co-authored-by: xly <leyang.xue@ed.ac.uk> Co-authored-by: Zhan Lu <51200935+lausannel@users.noreply.github.com> Co-authored-by: Yao Fu <yao.fu.aisys@gmail.com> Co-authored-by: Yao <fuyao3860@gmail.com> Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com> Co-authored-by: drunkcoding <14305648+drunkcoding@users.noreply.github.com> * feat: performance improvement and Qwen3 support (#60) * add openai api support * add test scripts, update readme, update api * Fix: Undefined Symbol Compilation Error (#37) * reformat code vllm style * add threadsafe queues * fix compilation error --------- Co-authored-by: xly <leyang.xue@ed.ac.uk> * Refactor code for better performance (#38) * reformat code vllm style * add threadsafe queues * fix compilation error * split files and remove queuing * performance improvement * remove error dependency * add try lock return check * fix header dependency * fix hard coded number --------- Co-authored-by: xly <leyang.xue@ed.ac.uk> * CI: add pre commit format ci (#40) * ci: add pre commit format ci * fix: add requirements for linting * fix: format code before merge * fix: update local clang format version * Chore: rename organization name & optimize CI (#41) * reformat code vllm style * add threadsafe queues * fix compilation error * split files and remove queuing * performance improvement * remove error dependency * add try lock return check * fix header dependency * fix hard coded number * update CI using cuda docker image * repo consistency * pr template fix * format doc * delete gpu option, add --no-install-recommends * add cuda matrix and remove cuda full package install * remove publish container * change team name to efficient moe --------- Co-authored-by: xly <leyang.xue@ed.ac.uk> * CI: fix not a git repository in CI (#43) * CI: fix missing sudo in apt install (#44) * CI: fix missing sudo (#45) * CI: revert os matrix in CI (#46) * CI: add missing apt update after installing deb file (#47) * Doc: Update README example to DeepSeek and Suppress Warning (#49) * reformat code vllm style * add threadsafe queues * fix compilation error * split files and remove queuing * performance improvement * remove error dependency * add try lock return check * fix header dependency * fix hard coded number * update CI using cuda docker image * repo consistency * pr template fix * format doc * delete gpu option, add --no-install-recommends * add cuda matrix and remove cuda full package install * remove publish container * change team name to efficient moe * update readme example to deepseek and supress warning * format * revert CI changes to main version --------- Co-authored-by: xly <leyang.xue@ed.ac.uk> * CI: do not build test if document update (#52) * reformat code vllm style * add threadsafe queues * fix compilation error * split files and remove queuing * performance improvement * remove error dependency * add try lock return check * fix header dependency * fix hard coded number * update CI using cuda docker image * repo consistency * pr template fix * format doc * delete gpu option, add --no-install-recommends * add cuda matrix and remove cuda full package install * remove publish container * change team name to efficient moe * update readme example to deepseek and supress warning * format * revert CI changes to main version * update readme conda env and ignore doc update in build and release * fix wildcard --------- Co-authored-by: xly <leyang.xue@ed.ac.uk> * format and change to deepseek in example * fix format * remove unused files * fix api server token id device * feat: Introduce Local Server for OpenAI-Compatible APIs (#4) * update table format * improve table clarity * init code commit * add openai api support * add test scripts, update readme, update api * format and change to deepseek in example * fix format * remove unused files * fix api server token id device --------- Co-authored-by: xly <leyang.xue@ed.ac.uk> Co-authored-by: Yao <fuyao3860@gmail.com> * fix gen broken * update readme links * cancel concurrent job * set dense node to device * sparse node set cpu * remove OS def * use update to date clang-format * fix setuptools version * fix setuptools version for python 3.8 * keep single cuda version in publish * feat: set parameter to device before serving (#56) * update table format * improve table clarity * init code commit * add openai api support * add test scripts, update readme, update api * format and change to deepseek in example * fix format * remove unused files * fix api server token id device * fix gen broken * update readme links * cancel concurrent job * set dense node to device * sparse node set cpu * remove OS def * use update to date clang-format * fix setuptools version * fix setuptools version for python 3.8 * keep single cuda version in publish --------- Co-authored-by: xly <leyang.xue@ed.ac.uk> Co-authored-by: Yao <fuyao3860@gmail.com> * add max length in gen openai * fix cache race condition * all param init at host * add qwen3 * ubuntu lts and build * pre-commit ubuntu version * router weights update overlap * rename deepseek_v2 and reduce torch kernel launch * fix import * fix build and fix bug * fix citation linebreak * fix typo * fix dtype size * remove comments * fix example * pr update init * remove comment and unify deepseek preroute * feat: Merge kernels from vLLM and FlashInfer (#63) * new allocator * add kernel compilation * stable topk --------- Co-authored-by: xly <leyang.xue@ed.ac.uk> --------- Co-authored-by: Yao <fuyao3860@gmail.com> Co-authored-by: xly <leyang.xue@ed.ac.uk> Co-authored-by: Zhan Lu <51200935+lausannel@users.noreply.github.com> Co-authored-by: Yao Fu <yao.fu.aisys@gmail.com> * Add Claude Code GitHub Workflow (#73) * "Claude PR Assistant workflow" * "Claude Code Review workflow" * Xly/code clean (#74) * add openai api support * add test scripts, update readme, update api * format and change to deepseek in example * fix format * remove unused files * fix api server token id device * fix gen broken * update readme links * cancel concurrent job * set dense node to device * sparse node set cpu * remove OS def * use update to date clang-format * fix setuptools version * fix setuptools version for python 3.8 * keep single cuda version in publish * add max length in gen openai * fix cache race condition * all param init at host * add docker and sllm style read * wrap docker and test coverage * test * Clean up symlinks: remove unused op_builder, core/core, and move test_io to extensions * Replace core/kernel directory with symlink to extensions/kernel * seperations * remove ops dependency * Add CUTLASS fused MoE FFN kernel and supporting infrastructure - Add extensions/kernel/fused_moe_mlp.cu/h: BF16 CUTLASS 3-GEMM fused path (gate → up w/ SiLU-mul epilogue → down) with small-M and large-K tile dispatch - Add tests/cuda/test_fused_mlp_cutlass.cu: BF16 CUTLASS vs Torch-native benchmark - Integrate fused kernel into core/parallel/expert_module.cpp via ForwardHelper() - Update core/model/fused_mlp.cu/h and extensions/kernel/epilogue_utils.h - Improve core/utils: cache.h, lockfree_queue.h, simple_object_pool.h - Update tests/cuda/CMakeLists.txt with KERNEL_SRC pattern for CUTLASS tests - Update CLAUDE.md docs and setup.py build config Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * Add prefill-decode collocation benchmark with throughput analysis Benchmarks five attention colocation strategies for serving decode and prefill requests on the same GPU time-slice: 0 serial — sequential on default stream 1 varlen-fused — single flash_attn_varlen_func (continuous batching) 2 dual-stream — two CUDA streams, no SM partition 3 green-ctx-sm — SM-partitioned green contexts (CUDA ≥ 12.4) 4 green-ctx-sm-wq — SM + work-queue balanced scope (CUDA 13.1+) Throughput analysis includes: - Separate decode-only / prefill-only baselines with TFLOPS and tok/s - Ideal-overlap bound (perfect concurrency = max(dec, pre)) - Per-mode: TFLOPS, decode tok/s, prefill tok/s, overlap efficiency - Generation-projection table: decode overhead and Δ vs serial per mode CUDA 13.1 green context API notes (driver 590.x): - CUdevResourceDesc is a pointer typedef (c_void_p), not a struct - cuGreenCtxStreamCreate requires CU_STREAM_NON_BLOCKING flag - CU_DEV_RESOURCE_TYPE_WORKQUEUE_CONFIG = 1000; configure sharingScope to CU_WORKQUEUE_SCOPE_GREEN_CTX_BALANCED for WQ isolation Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * Bump pydantic and transformers to resolve Dependabot alerts - pydantic 1.10.12 → 1.10.13: fixes ReDoS in email validation (GHSA-mr82-8j83-vxmv) - transformers 4.51.3 → 4.53.0: fixes 14 alerts including 3 HIGH RCE (GHSA-wrfc-pvp9-mr9g, GHSA-hxxf-235m-72v3, GHSA-qxrp-vhvm-j765) and 11 MEDIUM/LOW ReDoS vulnerabilities - Remove torch==2.3.1 pin (managed by conda env / base image) - Add flash-attn to requirements Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * fix build * Make readme_example.py testable via --help Add argparse to readme_example.py so that model-loading code runs only after parse_args(), allowing `--help` to exit 0 without a GPU or model. Replace the AST-only test_readme_example_syntax with test_readme_example_help, which mirrors the existing test_interface_example_help pattern and is verified passing in Docker. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * Fix format CI and build-test CI - requirements.txt: sort flash-attn alphabetically (between fastapi and hjson) so requirements-txt-fixer pre-commit hook passes - build-test.yml: replace Ubuntu 20.04 CUDA container (Python 3.8, broken PyTorch wheel) with actions/setup-python Python 3.10 + CPU-only torch; switch from full wheel build to sdist-only (--no-isolation) to avoid CUTLASS dependency and 20+ min compile time Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * green ctx bench * Fix CI: guard CUDA extensions behind cuda_available, add statics to codespell ignore - setup.py: only build CUDAExtension when torch.version.cuda is set; the build-test CI installs CPU-only torch and lacks CUDA_HOME, causing CUDAExtension to abort with OSError - .pre-commit-config.yaml: add 'statics' to codespell ignore-words-list; the term is valid C++ (module-level static variables) but was flagged as a misspelling of 'statistics' Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * tests update --------- Co-authored-by: Yao <fuyao3860@gmail.com> Co-authored-by: xly <leyang.xue@ed.ac.uk> Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com> * remove claude * format * resolve review * resolve reviews --------- Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: xly <leyang.xue@ed.ac.uk> Co-authored-by: Zhan Lu <51200935+lausannel@users.noreply.github.com> Co-authored-by: lausannel <513964121@qq.com> Co-authored-by: Yao Fu <fuyao3860@gmail.com> Co-authored-by: Yao Fu <yao.fu.aisys@gmail.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Co-authored-by: Copilot <198982749+Copilot@users.noreply.github.com> Co-authored-by: drunkcoding <14305648+drunkcoding@users.noreply.github.com> Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
seanlinmt
pushed a commit
to seanlinmt/MoE-Infinity
that referenced
this pull request
Apr 2, 2026
…cientMoE#75) * update table format * improve table clarity * init code commit * doc: add flashattention installation guide and change toc * feat: remove libaio dependency * remove spdlog dependency * misc: remove unused code and dependencies * misc: remove commented-out code and unused imports * fix: cuda oom due to safe tensors open * remove gcc-12 requirement * gptq disable exllama * fix: key error in offload set * add forward and call (EfficientMoE#7) * add forward and call * fix a bug * feat: support grok-1 model * update API note and install * Feature/expert parallel (EfficientMoE#9) * add back expert parallel by id hash * add grok ep * fix mistral typo * accom cuda copy bug * sync after compute * fix:sync to make sure that input is ready --------- Co-authored-by: xly <leyang.xue@ed.ac.uk> Co-authored-by: luzhan <513964121@qq.com> * fix tokenizer in example * Xly/deepseek (EfficientMoE#34) * add override QuantLinear (EfficientMoE#29) Co-authored-by: xly <leyang.xue@ed.ac.uk> * use torch streampool * format * working deepspeed backend * fix: revert apply_rotary_pos_emb in deepseek * fix busy waiting * fix deepseek flashattn * add deepseek v3 * format and fix multigpu deepseek bug * with device caching allocator * add on-demand lock cache --------- Co-authored-by: xly <leyang.xue@ed.ac.uk> Co-authored-by: lausannel <513964121@qq.com> * Upstream (EfficientMoE#72) * Fix: Undefined Symbol Compilation Error (EfficientMoE#37) * reformat code vllm style * add threadsafe queues * fix compilation error --------- Co-authored-by: xly <leyang.xue@ed.ac.uk> * Refactor code for better performance (EfficientMoE#38) * reformat code vllm style * add threadsafe queues * fix compilation error * split files and remove queuing * performance improvement * remove error dependency * add try lock return check * fix header dependency * fix hard coded number --------- Co-authored-by: xly <leyang.xue@ed.ac.uk> * CI: add pre commit format ci (EfficientMoE#40) * ci: add pre commit format ci * fix: add requirements for linting * fix: format code before merge * fix: update local clang format version * Chore: rename organization name & optimize CI (EfficientMoE#41) * reformat code vllm style * add threadsafe queues * fix compilation error * split files and remove queuing * performance improvement * remove error dependency * add try lock return check * fix header dependency * fix hard coded number * update CI using cuda docker image * repo consistency * pr template fix * format doc * delete gpu option, add --no-install-recommends * add cuda matrix and remove cuda full package install * remove publish container * change team name to efficient moe --------- Co-authored-by: xly <leyang.xue@ed.ac.uk> * CI: fix not a git repository in CI (EfficientMoE#43) * CI: fix missing sudo in apt install (EfficientMoE#44) * CI: fix missing sudo (EfficientMoE#45) * CI: revert os matrix in CI (EfficientMoE#46) * CI: add missing apt update after installing deb file (EfficientMoE#47) * Doc: Update README example to DeepSeek and Suppress Warning (EfficientMoE#49) * reformat code vllm style * add threadsafe queues * fix compilation error * split files and remove queuing * performance improvement * remove error dependency * add try lock return check * fix header dependency * fix hard coded number * update CI using cuda docker image * repo consistency * pr template fix * format doc * delete gpu option, add --no-install-recommends * add cuda matrix and remove cuda full package install * remove publish container * change team name to efficient moe * update readme example to deepseek and supress warning * format * revert CI changes to main version --------- Co-authored-by: xly <leyang.xue@ed.ac.uk> * CI: do not build test if document update (EfficientMoE#52) * reformat code vllm style * add threadsafe queues * fix compilation error * split files and remove queuing * performance improvement * remove error dependency * add try lock return check * fix header dependency * fix hard coded number * update CI using cuda docker image * repo consistency * pr template fix * format doc * delete gpu option, add --no-install-recommends * add cuda matrix and remove cuda full package install * remove publish container * change team name to efficient moe * update readme example to deepseek and supress warning * format * revert CI changes to main version * update readme conda env and ignore doc update in build and release * fix wildcard --------- Co-authored-by: xly <leyang.xue@ed.ac.uk> * feat: Introduce Local Server for OpenAI-Compatible APIs (EfficientMoE#4) * update table format * improve table clarity * init code commit * add openai api support * add test scripts, update readme, update api * format and change to deepseek in example * fix format * remove unused files * fix api server token id device --------- Co-authored-by: xly <leyang.xue@ed.ac.uk> Co-authored-by: Yao <fuyao3860@gmail.com> * feat: set parameter to device before serving (EfficientMoE#56) * update table format * improve table clarity * init code commit * add openai api support * add test scripts, update readme, update api * format and change to deepseek in example * fix format * remove unused files * fix api server token id device * fix gen broken * update readme links * cancel concurrent job * set dense node to device * sparse node set cpu * remove OS def * use update to date clang-format * fix setuptools version * fix setuptools version for python 3.8 * keep single cuda version in publish --------- Co-authored-by: xly <leyang.xue@ed.ac.uk> Co-authored-by: Yao <fuyao3860@gmail.com> * Chore(deps): Bump pyarrow from 12.0.0 to 14.0.1 (EfficientMoE#69) Bumps [pyarrow](https://github.com/apache/arrow) from 12.0.0 to 14.0.1. - [Release notes](https://github.com/apache/arrow/releases) - [Commits](apache/arrow@go/v12.0.0...go/v14.0.1) --- updated-dependencies: - dependency-name: pyarrow dependency-version: 14.0.1 dependency-type: direct:production ... Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> --------- Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: xly <leyang.xue@ed.ac.uk> Co-authored-by: Zhan Lu <51200935+lausannel@users.noreply.github.com> Co-authored-by: Yao Fu <yao.fu.aisys@gmail.com> Co-authored-by: Yao <fuyao3860@gmail.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> * Audit repository for stale code indicators (EfficientMoE#71) * Fix: Undefined Symbol Compilation Error (EfficientMoE#37) * reformat code vllm style * add threadsafe queues * fix compilation error --------- Co-authored-by: xly <leyang.xue@ed.ac.uk> * Refactor code for better performance (EfficientMoE#38) * reformat code vllm style * add threadsafe queues * fix compilation error * split files and remove queuing * performance improvement * remove error dependency * add try lock return check * fix header dependency * fix hard coded number --------- Co-authored-by: xly <leyang.xue@ed.ac.uk> * CI: add pre commit format ci (EfficientMoE#40) * ci: add pre commit format ci * fix: add requirements for linting * fix: format code before merge * fix: update local clang format version * Chore: rename organization name & optimize CI (EfficientMoE#41) * reformat code vllm style * add threadsafe queues * fix compilation error * split files and remove queuing * performance improvement * remove error dependency * add try lock return check * fix header dependency * fix hard coded number * update CI using cuda docker image * repo consistency * pr template fix * format doc * delete gpu option, add --no-install-recommends * add cuda matrix and remove cuda full package install * remove publish container * change team name to efficient moe --------- Co-authored-by: xly <leyang.xue@ed.ac.uk> * CI: fix not a git repository in CI (EfficientMoE#43) * CI: fix missing sudo in apt install (EfficientMoE#44) * CI: fix missing sudo (EfficientMoE#45) * CI: revert os matrix in CI (EfficientMoE#46) * CI: add missing apt update after installing deb file (EfficientMoE#47) * Doc: Update README example to DeepSeek and Suppress Warning (EfficientMoE#49) * reformat code vllm style * add threadsafe queues * fix compilation error * split files and remove queuing * performance improvement * remove error dependency * add try lock return check * fix header dependency * fix hard coded number * update CI using cuda docker image * repo consistency * pr template fix * format doc * delete gpu option, add --no-install-recommends * add cuda matrix and remove cuda full package install * remove publish container * change team name to efficient moe * update readme example to deepseek and supress warning * format * revert CI changes to main version --------- Co-authored-by: xly <leyang.xue@ed.ac.uk> * CI: do not build test if document update (EfficientMoE#52) * reformat code vllm style * add threadsafe queues * fix compilation error * split files and remove queuing * performance improvement * remove error dependency * add try lock return check * fix header dependency * fix hard coded number * update CI using cuda docker image * repo consistency * pr template fix * format doc * delete gpu option, add --no-install-recommends * add cuda matrix and remove cuda full package install * remove publish container * change team name to efficient moe * update readme example to deepseek and supress warning * format * revert CI changes to main version * update readme conda env and ignore doc update in build and release * fix wildcard --------- Co-authored-by: xly <leyang.xue@ed.ac.uk> * feat: Introduce Local Server for OpenAI-Compatible APIs (EfficientMoE#4) * update table format * improve table clarity * init code commit * add openai api support * add test scripts, update readme, update api * format and change to deepseek in example * fix format * remove unused files * fix api server token id device --------- Co-authored-by: xly <leyang.xue@ed.ac.uk> Co-authored-by: Yao <fuyao3860@gmail.com> * feat: set parameter to device before serving (EfficientMoE#56) * update table format * improve table clarity * init code commit * add openai api support * add test scripts, update readme, update api * format and change to deepseek in example * fix format * remove unused files * fix api server token id device * fix gen broken * update readme links * cancel concurrent job * set dense node to device * sparse node set cpu * remove OS def * use update to date clang-format * fix setuptools version * fix setuptools version for python 3.8 * keep single cuda version in publish --------- Co-authored-by: xly <leyang.xue@ed.ac.uk> Co-authored-by: Yao <fuyao3860@gmail.com> * Initial plan * Add mypy lint hook Co-authored-by: drunkcoding <14305648+drunkcoding@users.noreply.github.com> * Configure mypy settings Co-authored-by: drunkcoding <14305648+drunkcoding@users.noreply.github.com> * Adjust mypy scope Co-authored-by: drunkcoding <14305648+drunkcoding@users.noreply.github.com> * Scope mypy checks Co-authored-by: drunkcoding <14305648+drunkcoding@users.noreply.github.com> --------- Co-authored-by: Leyang Xue <s2062808@ed.ac.uk> Co-authored-by: xly <leyang.xue@ed.ac.uk> Co-authored-by: Zhan Lu <51200935+lausannel@users.noreply.github.com> Co-authored-by: Yao Fu <yao.fu.aisys@gmail.com> Co-authored-by: Yao <fuyao3860@gmail.com> Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com> Co-authored-by: drunkcoding <14305648+drunkcoding@users.noreply.github.com> * feat: performance improvement and Qwen3 support (EfficientMoE#60) * add openai api support * add test scripts, update readme, update api * Fix: Undefined Symbol Compilation Error (EfficientMoE#37) * reformat code vllm style * add threadsafe queues * fix compilation error --------- Co-authored-by: xly <leyang.xue@ed.ac.uk> * Refactor code for better performance (EfficientMoE#38) * reformat code vllm style * add threadsafe queues * fix compilation error * split files and remove queuing * performance improvement * remove error dependency * add try lock return check * fix header dependency * fix hard coded number --------- Co-authored-by: xly <leyang.xue@ed.ac.uk> * CI: add pre commit format ci (EfficientMoE#40) * ci: add pre commit format ci * fix: add requirements for linting * fix: format code before merge * fix: update local clang format version * Chore: rename organization name & optimize CI (EfficientMoE#41) * reformat code vllm style * add threadsafe queues * fix compilation error * split files and remove queuing * performance improvement * remove error dependency * add try lock return check * fix header dependency * fix hard coded number * update CI using cuda docker image * repo consistency * pr template fix * format doc * delete gpu option, add --no-install-recommends * add cuda matrix and remove cuda full package install * remove publish container * change team name to efficient moe --------- Co-authored-by: xly <leyang.xue@ed.ac.uk> * CI: fix not a git repository in CI (EfficientMoE#43) * CI: fix missing sudo in apt install (EfficientMoE#44) * CI: fix missing sudo (EfficientMoE#45) * CI: revert os matrix in CI (EfficientMoE#46) * CI: add missing apt update after installing deb file (EfficientMoE#47) * Doc: Update README example to DeepSeek and Suppress Warning (EfficientMoE#49) * reformat code vllm style * add threadsafe queues * fix compilation error * split files and remove queuing * performance improvement * remove error dependency * add try lock return check * fix header dependency * fix hard coded number * update CI using cuda docker image * repo consistency * pr template fix * format doc * delete gpu option, add --no-install-recommends * add cuda matrix and remove cuda full package install * remove publish container * change team name to efficient moe * update readme example to deepseek and supress warning * format * revert CI changes to main version --------- Co-authored-by: xly <leyang.xue@ed.ac.uk> * CI: do not build test if document update (EfficientMoE#52) * reformat code vllm style * add threadsafe queues * fix compilation error * split files and remove queuing * performance improvement * remove error dependency * add try lock return check * fix header dependency * fix hard coded number * update CI using cuda docker image * repo consistency * pr template fix * format doc * delete gpu option, add --no-install-recommends * add cuda matrix and remove cuda full package install * remove publish container * change team name to efficient moe * update readme example to deepseek and supress warning * format * revert CI changes to main version * update readme conda env and ignore doc update in build and release * fix wildcard --------- Co-authored-by: xly <leyang.xue@ed.ac.uk> * format and change to deepseek in example * fix format * remove unused files * fix api server token id device * feat: Introduce Local Server for OpenAI-Compatible APIs (EfficientMoE#4) * update table format * improve table clarity * init code commit * add openai api support * add test scripts, update readme, update api * format and change to deepseek in example * fix format * remove unused files * fix api server token id device --------- Co-authored-by: xly <leyang.xue@ed.ac.uk> Co-authored-by: Yao <fuyao3860@gmail.com> * fix gen broken * update readme links * cancel concurrent job * set dense node to device * sparse node set cpu * remove OS def * use update to date clang-format * fix setuptools version * fix setuptools version for python 3.8 * keep single cuda version in publish * feat: set parameter to device before serving (EfficientMoE#56) * update table format * improve table clarity * init code commit * add openai api support * add test scripts, update readme, update api * format and change to deepseek in example * fix format * remove unused files * fix api server token id device * fix gen broken * update readme links * cancel concurrent job * set dense node to device * sparse node set cpu * remove OS def * use update to date clang-format * fix setuptools version * fix setuptools version for python 3.8 * keep single cuda version in publish --------- Co-authored-by: xly <leyang.xue@ed.ac.uk> Co-authored-by: Yao <fuyao3860@gmail.com> * add max length in gen openai * fix cache race condition * all param init at host * add qwen3 * ubuntu lts and build * pre-commit ubuntu version * router weights update overlap * rename deepseek_v2 and reduce torch kernel launch * fix import * fix build and fix bug * fix citation linebreak * fix typo * fix dtype size * remove comments * fix example * pr update init * remove comment and unify deepseek preroute * feat: Merge kernels from vLLM and FlashInfer (EfficientMoE#63) * new allocator * add kernel compilation * stable topk --------- Co-authored-by: xly <leyang.xue@ed.ac.uk> --------- Co-authored-by: Yao <fuyao3860@gmail.com> Co-authored-by: xly <leyang.xue@ed.ac.uk> Co-authored-by: Zhan Lu <51200935+lausannel@users.noreply.github.com> Co-authored-by: Yao Fu <yao.fu.aisys@gmail.com> * Add Claude Code GitHub Workflow (EfficientMoE#73) * "Claude PR Assistant workflow" * "Claude Code Review workflow" * Xly/code clean (EfficientMoE#74) * add openai api support * add test scripts, update readme, update api * format and change to deepseek in example * fix format * remove unused files * fix api server token id device * fix gen broken * update readme links * cancel concurrent job * set dense node to device * sparse node set cpu * remove OS def * use update to date clang-format * fix setuptools version * fix setuptools version for python 3.8 * keep single cuda version in publish * add max length in gen openai * fix cache race condition * all param init at host * add docker and sllm style read * wrap docker and test coverage * test * Clean up symlinks: remove unused op_builder, core/core, and move test_io to extensions * Replace core/kernel directory with symlink to extensions/kernel * seperations * remove ops dependency * Add CUTLASS fused MoE FFN kernel and supporting infrastructure - Add extensions/kernel/fused_moe_mlp.cu/h: BF16 CUTLASS 3-GEMM fused path (gate → up w/ SiLU-mul epilogue → down) with small-M and large-K tile dispatch - Add tests/cuda/test_fused_mlp_cutlass.cu: BF16 CUTLASS vs Torch-native benchmark - Integrate fused kernel into core/parallel/expert_module.cpp via ForwardHelper() - Update core/model/fused_mlp.cu/h and extensions/kernel/epilogue_utils.h - Improve core/utils: cache.h, lockfree_queue.h, simple_object_pool.h - Update tests/cuda/CMakeLists.txt with KERNEL_SRC pattern for CUTLASS tests - Update CLAUDE.md docs and setup.py build config Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * Add prefill-decode collocation benchmark with throughput analysis Benchmarks five attention colocation strategies for serving decode and prefill requests on the same GPU time-slice: 0 serial — sequential on default stream 1 varlen-fused — single flash_attn_varlen_func (continuous batching) 2 dual-stream — two CUDA streams, no SM partition 3 green-ctx-sm — SM-partitioned green contexts (CUDA ≥ 12.4) 4 green-ctx-sm-wq — SM + work-queue balanced scope (CUDA 13.1+) Throughput analysis includes: - Separate decode-only / prefill-only baselines with TFLOPS and tok/s - Ideal-overlap bound (perfect concurrency = max(dec, pre)) - Per-mode: TFLOPS, decode tok/s, prefill tok/s, overlap efficiency - Generation-projection table: decode overhead and Δ vs serial per mode CUDA 13.1 green context API notes (driver 590.x): - CUdevResourceDesc is a pointer typedef (c_void_p), not a struct - cuGreenCtxStreamCreate requires CU_STREAM_NON_BLOCKING flag - CU_DEV_RESOURCE_TYPE_WORKQUEUE_CONFIG = 1000; configure sharingScope to CU_WORKQUEUE_SCOPE_GREEN_CTX_BALANCED for WQ isolation Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * Bump pydantic and transformers to resolve Dependabot alerts - pydantic 1.10.12 → 1.10.13: fixes ReDoS in email validation (GHSA-mr82-8j83-vxmv) - transformers 4.51.3 → 4.53.0: fixes 14 alerts including 3 HIGH RCE (GHSA-wrfc-pvp9-mr9g, GHSA-hxxf-235m-72v3, GHSA-qxrp-vhvm-j765) and 11 MEDIUM/LOW ReDoS vulnerabilities - Remove torch==2.3.1 pin (managed by conda env / base image) - Add flash-attn to requirements Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * fix build * Make readme_example.py testable via --help Add argparse to readme_example.py so that model-loading code runs only after parse_args(), allowing `--help` to exit 0 without a GPU or model. Replace the AST-only test_readme_example_syntax with test_readme_example_help, which mirrors the existing test_interface_example_help pattern and is verified passing in Docker. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * Fix format CI and build-test CI - requirements.txt: sort flash-attn alphabetically (between fastapi and hjson) so requirements-txt-fixer pre-commit hook passes - build-test.yml: replace Ubuntu 20.04 CUDA container (Python 3.8, broken PyTorch wheel) with actions/setup-python Python 3.10 + CPU-only torch; switch from full wheel build to sdist-only (--no-isolation) to avoid CUTLASS dependency and 20+ min compile time Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * green ctx bench * Fix CI: guard CUDA extensions behind cuda_available, add statics to codespell ignore - setup.py: only build CUDAExtension when torch.version.cuda is set; the build-test CI installs CPU-only torch and lacks CUDA_HOME, causing CUDAExtension to abort with OSError - .pre-commit-config.yaml: add 'statics' to codespell ignore-words-list; the term is valid C++ (module-level static variables) but was flagged as a misspelling of 'statistics' Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * tests update --------- Co-authored-by: Yao <fuyao3860@gmail.com> Co-authored-by: xly <leyang.xue@ed.ac.uk> Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com> * remove claude * format * resolve review * resolve reviews --------- Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: xly <leyang.xue@ed.ac.uk> Co-authored-by: Zhan Lu <51200935+lausannel@users.noreply.github.com> Co-authored-by: lausannel <513964121@qq.com> Co-authored-by: Yao Fu <fuyao3860@gmail.com> Co-authored-by: Yao Fu <yao.fu.aisys@gmail.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Co-authored-by: Copilot <198982749+Copilot@users.noreply.github.com> Co-authored-by: drunkcoding <14305648+drunkcoding@users.noreply.github.com> Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Description
Fuse MoE layer kernels
Motivation
Kernel launch overhead too large
Type of Change
Checklist