Skip to content

[https://nvbugs/5597647][fix] Fix MNNVL Allreduce accuracy issue on Hopper #8612

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

Merged
LarryXFly merged 2 commits into from
Oct 27, 2025
Merged

[https://nvbugs/5597647][fix] Fix MNNVL Allreduce accuracy issue on Hopper #8612

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
86ce503
Fix hopper accuracy issue with wrong sem
timlee0212 Oct 23, 2025
f8b1fb0
Add back the unit test
timlee0212 Oct 23, 2025
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
4 changes: 2 additions & 2 deletions cpp/tensorrt_llm/kernels/communicationKernels/mnnvlTwoShotAllreduceKernels.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -117,8 +117,8 @@ __device__ struct __attribute__((aligned(32))) LamportFlags
{
#if (defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 1000))
asm volatile("red.async.release.global.gpu.add.u32 [%0], %1;" ::"l"(offset_access_ptr), "r"(1) : "memory");
#elif (defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 900))
asm volatile("red.global.gpu.add.u32 [%0], %1;" ::"l"(offset_access_ptr), "r"(1) : "memory");
#elif (defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 700))
asm volatile("red.release.global.gpu.add.u32 [%0], %1;" ::"l"(offset_access_ptr), "r"(1) : "memory");
Comment thread
timlee0212 marked this conversation as resolved.
#else
atomicAdd(offset_access_ptr, 1);
#endif
Expand Down
1 change: 0 additions & 1 deletion tests/unittest/_torch/multi_gpu/test_mnnvl_allreduce.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -183,7 +183,6 @@ def func(input, residual, norm_weight, eps, enable_fusion):
)
def test_row_linear_residual_norm_fusion(seq_len, hidden_size, dtype, strategy,
fusion):
pytest.skip("https://nvbugs/5597647")
torch.manual_seed(42)
tensor_parallel_size = 2

Expand Down