Skip to content

[None][fix] Gate MoE BF16 to FP8 conversion on compiler support #15934

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

Open
waynehacking8 wants to merge 1 commit into
base: main
Choose a base branch
from
Open

[None][fix] Gate MoE BF16 to FP8 conversion on compiler support #15934

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
6 changes: 4 additions & 2 deletions cpp/tensorrt_llm/kernels/moe/communication/moeAlltoAllKernels.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1580,8 +1580,10 @@ __device__ __forceinline__ void vec_convert(
out[j] = DstT(static_cast<float>(in[j]));
}

// BF16 → FP8 e4m3: paired PTX cvt.rn.satfinite.e4m3x2.bf16x2 (SM100+, Blackwell).
#if TLLM_MOE_A2A_COMPILE_SM100
// Paired BF16 -> FP8 conversion requires PTX 9.1 (CUDA 13.1) and an
// architecture- or family-specific Blackwell target.
#if TLLM_MOE_A2A_COMPILE_SM100 && (__CUDACC_VER_MAJOR__ * 10000 + __CUDACC_VER_MINOR__ * 100 >= 130100) \
&& (TLLM_CUDA_HOST_PASS || defined(__CUDA_ARCH_SPECIFIC__) || defined(__CUDA_ARCH_FAMILY_SPECIFIC__))
template <size_t VEC_SIZE, std::enable_if_t<(VEC_SIZE % 2 == 0), int> = 0>
__device__ __forceinline__ void vec_convert(
flashinfer::vec_t<__nv_fp8_e4m3, VEC_SIZE>& out, flashinfer::vec_t<__nv_bfloat16, VEC_SIZE> const& in)
Expand Down