ExternalVendorCode/whisper.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/whisper.cpp.git synced 2024-12-19 12:47:52 +00:00
Code Issues Actions 4 Packages Projects Releases Wiki Activity

metal : move mm_id indices to shared mem (llama/5982)

Browse Source
This commit is contained in:
Georgi Gerganov 2024-03-10 23:12:48 +02:00
parent 653d2e8ff9
commit 7bdb1de9ec
No known key found for this signature in database
GPG Key ID: BF970631944C16B7
2 changed files with 6 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
ggml-metal.m
View File

@ -1642,8 +1642,8 @@ static enum ggml_status ggml_metal_graph_compute(
// TODO: make this more general // TODO: make this more general
GGML_ASSERT(n_as <= 8); GGML_ASSERT(n_as <= 8);
// max size of the src1ids array in the kernel stack // max size of the src1ids array in the kernel shared buffer
GGML_ASSERT(ne11 <= 512); GGML_ASSERT(ne11 <= 4096);
const int64_t ne20 = src2 ? src2->ne[0] : 0; const int64_t ne20 = src2 ? src2->ne[0] : 0;
const int64_t ne21 = src2 ? src2->ne[1] : 0; const int64_t ne21 = src2 ? src2->ne[1] : 0;
@ -1741,7 +1741,7 @@ static enum ggml_status ggml_metal_graph_compute(
[encoder setBuffer:id_src_cur offset:offs_src_cur atIndex:19 + j]; [encoder setBuffer:id_src_cur offset:offs_src_cur atIndex:19 + j];
} }
[encoder setThreadgroupMemoryLength:8192 atIndex:0]; [encoder setThreadgroupMemoryLength:GGML_PAD(8192 + 2*ne11, 16) atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake((ne11 + 31)/32, (ne21 + 63)/64, n_as*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)]; [encoder dispatchThreadgroups:MTLSizeMake((ne11 + 31)/32, (ne21 + 63)/64, n_as*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)];
} else { } else {

6
ggml-metal.metal
View File

@ -5386,7 +5386,7 @@ template<typename block_q, short nl, void (*dequantize_func)(device const block_
void kernel_mul_mm_id_impl( void kernel_mul_mm_id_impl(
device const uchar * src0, device const uchar * src0,
device const uchar * src1, device const uchar * src1,
thread short * src1ids, threadgroup short * src1ids,
device float * dst, device float * dst,
constant int64_t & ne00, constant int64_t & ne00,
constant int64_t & ne02, constant int64_t & ne02,
@ -5589,9 +5589,9 @@ kernel void kernel_mul_mm_id(
tgpig.z = tgpig.z%(ne12*ne13); tgpig.z = tgpig.z%(ne12*ne13);
// row indices of src1 for expert id // row indices of src1 for expert id
int64_t _ne1 = 0; threadgroup short * src1ids = (threadgroup short *)(shared_memory + 8192);
short src1ids[512];
int64_t _ne1 = 0;
for (int64_t i1 = 0; i1 < ne1; i1++) { for (int64_t i1 = 0; i1 < ne1; i1++) {
if (((device int32_t *) (ids + i1*nbi1))[idx] == id) { if (((device int32_t *) (ids + i1*nbi1))[idx] == id) {
src1ids[_ne1++] = i1; src1ids[_ne1++] = i1;