Skip to content

feat(moe-a2a): Update nvlink onesided all-to-all #3139

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
trevor-m wants to merge 4 commits into from
Closed

feat(moe-a2a): Update nvlink onesided all-to-all #3139

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
3a89648
comm: port TRT-LLM b802a3818 to FlashInfer MoE A2A kernels
trevor-m Apr 20, 2026
752476d
comm: add PDL (Programmatic Dependent Launch) support to MoE A2A kernels
trevor-m Apr 20, 2026
d1810b2
Use bool flag for PDL
trevor-m Apr 21, 2026
bfcc10f
Use cuda intrinistic instead of PTX
trevor-m Apr 22, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
23 changes: 23 additions & 0 deletions csrc/nv_internal/tensorrt_llm/common/envUtils.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -19,6 +19,10 @@
#include <cstdint>
#include <optional>
#include <string>
#include <utility>

#include <cuda_runtime_api.h>
#include "flashinfer/exception.h"

namespace tensorrt_llm::common {
// Useful when you want to inject some debug code controllable with env var.
Expand Down Expand Up @@ -101,4 +105,23 @@ int getEnvMoeA2ADispatchBlockSize();
// Block size (threads per block) for MoE A2A Combine kernels (default 256 if unset or invalid)
int getEnvMoeA2ACombineBlockSize();

template <typename KernelFn, typename... Args>
inline void launchWithPdlWhenEnabled(char const* name, bool enable_pdl, KernelFn kernelFn,
dim3 grid, dim3 block, size_t dynamicShmSize, cudaStream_t stream, Args&&... args)
{
cudaLaunchConfig_t kernelConfig;
kernelConfig.gridDim = grid;
kernelConfig.blockDim = block;
kernelConfig.dynamicSmemBytes = dynamicShmSize;
kernelConfig.stream = stream;
cudaLaunchAttribute attrs[1];
attrs[0].id = cudaLaunchAttributeProgrammaticStreamSerialization;
attrs[0].val.programmaticStreamSerializationAllowed = enable_pdl;
kernelConfig.attrs = attrs;
kernelConfig.numAttrs = 1;
cudaError_t e = cudaLaunchKernelEx(&kernelConfig, kernelFn, std::forward<Args>(args)...);
FLASHINFER_CHECK(e == cudaSuccess, "cudaLaunchKernelEx (", name, ") failed: ",
cudaGetErrorString(e));
}
Comment on lines +108 to +125

@coderabbitai coderabbitai Bot Apr 21, 2026

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

⚠️ Potential issue | 🟡 Minor

Commit the clang-format output for this helper.

Pre-commit is failing because clang-format rewrote this file’s include order/signature/body wrapping. Please run formatting and commit the result so CI can pass.

🤖 Prompt for AI Agents 
Verify each finding against the current code and only fix it if needed.

In `@csrc/nv_internal/tensorrt_llm/common/envUtils.h` around lines 108 - 125, Run
clang-format and commit the formatted version of the helper function
launchWithPdlWhenEnabled in envUtils.h: apply your project's clang-format
configuration to reformat the include order, signature/wrapping and function
body (the cudaLaunchConfig_t setup, attrs array, and cudaLaunchKernelEx call),
then stage and commit the changed file so the pre-commit CI hook stops failing.


} // namespace tensorrt_llm::common
Loading
Loading