CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-10-31 08:51:55 +00:00
Code Issues Packages Projects Releases Wiki Activity

fix: Correctly size the shared memory bufer and assert expected size relationships

Browse Source 
Branch: GraniteFourPerf

Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
This commit is contained in:
Gabe Goodhart
2025-07-18 11:05:31 -06:00
parent e16e24bebd
commit 21db0b598a
1 changed files with 9 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
ggml/src/ggml-metal/ggml-metal.m
View File

@@ -3017,7 +3017,15 @@ static bool ggml_metal_encode_node(
if (ne30 == 1) {
// Mamba-2
[encoder setThreadgroupMemoryLength:32*sizeof(float) atIndex:0]; // SIMD size
// One shared memory bucket for each simd group in the threadgroup
const int64_t shmem_size = d_state / 32;
GGML_ASSERT(shmem_size * 32 == d_state);
// One thread pre element in d_state
GGML_ASSERT(d_state <= (int64_t)pipeline.maxTotalThreadsPerThreadgroup);
[encoder setThreadgroupMemoryLength:(shmem_size)*sizeof(float) atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake(d_inner, n_head, n_seqs) threadsPerThreadgroup:MTLSizeMake(d_state, 1, 1)];
} else {
GGML_ASSERT(d_inner == 1);