ggml/ex: calculate accuracy in graph, adapt MNIST (ggml/980)

2024-10-03 17:29:59 +02:00
parent eee39bdc96
commit fabdc3bda3
11 changed files with 389 additions and 8 deletions
--- a/ggml/src/ggml-cuda/fattn-tile-f16.cu
+++ b/ggml/src/ggml-cuda/fattn-tile-f16.cu
@@ -259,7 +259,7 @@ static __global__ void flash_attn_tile_ext_f16(
        }

        half kqsum_j = __low2half(kqsum[j_VKQ_0/nwarps]) + __high2half(kqsum[j_VKQ_0/nwarps]);
-        kqsum_j = warp_reduce_sum(kqsum_j);
+        kqsum_j = warp_reduce_sum((float)kqsum_j);

 #pragma unroll
        for (int i00 = 0; i00 < D; i00 += 2*WARP_SIZE) {