This commit is contained in:
yadaish
2025-12-04 18:28:17 +00:00
parent 7059613404
commit bd6897d432

View File

@@ -639,7 +639,8 @@ struct MoeFlatmmKernel
index_t FlatScaleN = kargs.N / N_Pack / BlockGemmShape::WarpTile::at(I1);
// using ScaleType = std::conditional_t<MXFP4_Pipeline, e8m0_t, float>;
using ScaleType = std::conditional_t<MXFP4_Pipeline, bfloat16_t, float>;
static constexpr bool IsInt4 = std::is_same_v<BDataType, pk_int4_t>;
using ScaleType = std::conditional_t<MXFP4_Pipeline, std::conditional_t<IsInt4, bfloat16, e8m0_t>, float>;
const auto scale_b_flat_view = make_naive_tensor_view<address_space_enum::global>(
reinterpret_cast<const ScaleType*>(scale_n.ptr) + expert_id * kargs.N * scale_k,