biondizzle/nvfp4-megamoe-kernel
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix: HD_val variable name in test

Browse Source
This commit is contained in:
biondizzle
2026-05-28 06:32:01 +00:00
parent e173295a3a
commit e940786fd5
1 changed files with 5 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
tests/unit/test_fmha_sm100_standalone.cu
View File

@@ -92,13 +92,13 @@ int test_kernel(const char* name, int HD_val, int sk, float scale,
}
// Copy result and compare
uint16_t* hob = (uint16_t*)malloc(B*H*HD*sizeof(uint16_t));
cudaMemcpy(hob, do_gpu, B*H*HD*sizeof(uint16_t), cudaMemcpyDeviceToHost);
uint16_t* hob = (uint16_t*)malloc(B*H*HD_val*sizeof(uint16_t));
cudaMemcpy(hob, do_gpu, B*H*HD_val*sizeof(uint16_t), cudaMemcpyDeviceToHost);
float* ho_gpu = (float*)malloc(B*H*HD*sizeof(float));
for (int i = 0; i < B*H*HD; i++) ho_gpu[i] = bf16_to_f32_cpu(hob[i]);
float* ho_gpu = (float*)malloc(B*H*HD_val*sizeof(float));
for (int i = 0; i < B*H*HD_val; i++) ho_gpu[i] = bf16_to_f32_cpu(hob[i]);
float cos = cosine_sim(ho_gpu, ho_ref, B*H*HD);
float cos = cosine_sim(ho_gpu, ho_ref, B*H*HD_val);
int pass = cos > 0.999f;
printf(" %s hd=%d s_k=%d: cos %.6f %s\n", name, HD, sk, cos, pass ? "✅" : "❌");