fix: add cuda_runtime.h to hgemm_bind.cpp for cudaStream_t
This commit is contained in:
@@ -5,6 +5,7 @@
|
|||||||
// moe_expert_gemm(input, weights, expert_counts) → output
|
// moe_expert_gemm(input, weights, expert_counts) → output
|
||||||
|
|
||||||
#include <torch/extension.h>
|
#include <torch/extension.h>
|
||||||
|
#include <cuda_runtime.h>
|
||||||
#include <cuda_fp16.h>
|
#include <cuda_fp16.h>
|
||||||
#include <vector>
|
#include <vector>
|
||||||
|
|
||||||
|
|||||||
Reference in New Issue
Block a user