From: Johannes Gäßler <redacted>
Date: Wed, 28 May 2025 11:33:37 +0000 (+0200)
Subject: CUDA: fix FA tg at long context for CC >= 8.9 (llama/13852)
X-Git-Url: https://git.djapps.eu/?a=commitdiff_plain;h=9a500394ad1ef1190dde3ac91d399d8cb05bac16;p=pkg%2Fggml%2Fsources%2Fwhisper.cpp

CUDA: fix FA tg at long context for CC >= 8.9 (llama/13852)
---

diff --git a/ggml/src/ggml-cuda/fattn-common.cuh b/ggml/src/ggml-cuda/fattn-common.cuh
index a4fbd823..cfab2b5e 100644
--- a/ggml/src/ggml-cuda/fattn-common.cuh
+++ b/ggml/src/ggml-cuda/fattn-common.cuh
@@ -623,8 +623,8 @@ static __global__ void flash_attn_combine_results(
     __builtin_assume(tid < D);
 
     extern __shared__ float2 meta[];
-    if (tid < 2*parallel_blocks) {
-        ((float *) meta)[threadIdx.x] = ((const float *)VKQ_meta) [blockIdx.z*(2*parallel_blocks) + tid];
+    for (int i = tid; i < 2*parallel_blocks; i += D) {
+        ((float *) meta)[i] = ((const float *)VKQ_meta) [blockIdx.z*(2*parallel_blocks) + i];
     }
 
     __syncthreads();