EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

Revert "[1/N][Bug] Fix w4afp8 MoE NaN issue (sgl-kernel) (#9953)" (#10097)

Browse Source
This commit is contained in:
Yineng Zhang
2025-09-05 19:57:53 -07:00
committed by GitHub
parent 1a3d6f31da
commit 0e78c63c0e
1 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
sgl-kernel/csrc/moe/cutlass_moe/w4a8/w4a8_grouped_mm_c3x.cuh
View File

@@ -41,8 +41,8 @@ using MmaType = cutlass::float_e4m3_t; // FP8 e4m3 type
using QuantType = cutlass::int4b_t; // 4-bit integer type using QuantType = cutlass::int4b_t; // 4-bit integer type
using ElementAccumulator = float; // Accumulator type using ElementAccumulator = float; // Accumulator type
using ElementScale = cutlass::bfloat16_t; // Scale type using ElementScale = cutlass::bfloat16_t; // Scale type
using ElementC = cutlass::bfloat16_t; // Output type using ElementC = cutlass::half_t; // Default output type (FP16)
using ElementD = ElementC; // Output type using ElementD = ElementC; // Default output type (FP16)
using ProblemShape = cutlass::gemm::GroupProblemShape<Shape<int, int, int>>; using ProblemShape = cutlass::gemm::GroupProblemShape<Shape<int, int, int>>;
// Architecture-specific configurations // Architecture-specific configurations