update for fp8

Signed-off-by: yewentao256 <zhyanwentao@126.com>
2025-08-27 21:36:03 +00:00
parent 57f2f26a05
commit e92676ef4e
12 changed files with 13 additions and 13 deletions
--- a/csrc/attention/attention_kernels.cuh
+++ b/csrc/attention/attention_kernels.cuh
@@ -28,10 +28,10 @@

 #ifdef USE_ROCM
  #include <hip/hip_bf16.h>
-  #include "../quantization/fp8/amd/quant_utils.cuh"
+  #include "../quantization/w8a8/fp8/amd/quant_utils.cuh"
 typedef __hip_bfloat16 __nv_bfloat16;
 #else
-  #include "../quantization/fp8/nvidia/quant_utils.cuh"
+  #include "../quantization/w8a8/fp8/nvidia/quant_utils.cuh"
 #endif

 #define MAX(a, b) ((a) > (b) ? (a) : (b))