CUDA: fix MMQ writeback for int8 tensor cores ggerganov#8100

Co-Authored-By: Johannes Gäßler <johannesg@5d6.de>
Nexesenex · Jun 25, 2024 · 0d5cc68 · 0d5cc68
1 parent 54e1ff4
commit 0d5cc68
Showing 1 changed file with 1 addition and 3 deletions.
diff --git a/ggml-cuda/mmq.cuh b/ggml-cuda/mmq.cuh
@@ -2035,15 +2035,13 @@ static __device__ __forceinline__ void mmq_write_back_mma(
     static_assert(nwarps*mma_C::I == mmq_y, "nwarps*mma_C::I != mmq_y");
 #endif // INT8_MMA_AVAILABLE
 
-    dst += (threadIdx.y % ntx) * mma_C::J*stride;
-
 #pragma unroll
     for (int j0 = 0; j0 < mmq_x; j0 += ntx*mma_C::J) {
 #pragma unroll
         for (int n = 0; n < ntx; ++n) {
 #pragma unroll
             for (int l = 0; l < mma_C::ne; ++l) {
-                const int j = j0 + mma_C::get_j(l);
+                const int j = j0 + (threadIdx.y % ntx) * mma_C::J + mma_C::get_j(l);
 
                 if (j > j_max) {
                     continue;