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

test: force 1 K-tile for HD=64 debug

Browse Source
This commit is contained in:
biondizzle
2026-05-28 11:46:12 +00:00
parent abe1870429
commit 73bd21ce01
1 changed files with 1 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
tests/unit/test_umma_qk_hd64.cu
View File

@@ -29,7 +29,7 @@ test_umma_qk_hd64(const bf16_t* q, const bf16_t* k,
{
const int tid = threadIdx.x;
const int wid = tid / 32, lane = tid % 32;
const int n_ktiles = hd / MMA_K; // 4 for hd=64
const int n_ktiles = 1; // hd / MMA_K; // Force 1 K-tile for debugging
// SMEM: sQ (128, HD) canonical + sK (128, HD) canonical
// Each K-tile of (128, 16) = 4096 bytes