Flashinfer_CUTLASS_MOE fuses quantization for TP (#27223)

Signed-off-by: Shu Wang. <shuw@nvidia.com>
This commit is contained in:
Shu Wang
2025-10-31 17:54:29 +00:00
committed by GitHub
parent bc306fe5e9
commit fc16f1c477
4 changed files with 15 additions and 32 deletions
@@ -79,6 +79,7 @@ def select_nvfp4_gemm_impl(
ep_size=moe.moe_parallel_config.ep_size,
tp_rank=moe.moe_parallel_config.tp_rank,
tp_size=moe.moe_parallel_config.tp_size,
use_dp=moe.moe_parallel_config.dp_size > 1,
)
# native cutlass experts currently don't support DP; TP case won't call this