Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

metal : enable FA for MLA heads (#18950)

Browse Source
This commit is contained in:
Georgi Gerganov
2026-01-20 12:21:28 +02:00
committed by GitHub
parent 7dee9ff59a
commit 271191906c
3 changed files with 11 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-metal/ggml-metal-ops.cpp
+1 -1
View File
@@ -2520,7 +2520,7 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
// simdgroups per threadgroup (a.k.a. warps)
//nsg = ne01 <= nqptg ? MAX(4, MIN(nsgmax, MIN(ne11/ncpsg, (int64_t) pipeline.maxTotalThreadsPerThreadgroup/32))) : 4;
int32_t nsg = 4;
int32_t nsg = ne00 >= 512 ? 8 : 4;
const size_t smem = FATTN_SMEM(nsg);