Flashinfer_CUTLASS_MOE fuses quantization for TP (#27223)
Signed-off-by: Shu Wang. <shuw@nvidia.com>
This commit is contained in:
@@ -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
|
||||
|
||||
Reference in New Issue
Block a user