vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

wip

Browse source
This commit is contained in:
Georgi Gerganov 2024-01-23 18:27:54 +02:00
parent 17720fad66
commit 06c2d0d117
No known key found for this signature in database
GPG key ID: 449E073F9DC10735
3 changed files with 136 additions and 73 deletions
Download patch file Download diff file Expand all files Collapse all files

9
ggml-metal.m
View file

@ -2252,15 +2252,16 @@ static bool ggml_metal_graph_compute(
[encoder setBytes:&ne3 length:sizeof( int64_t) atIndex:26];
[encoder setBytes:&scale length:sizeof( float) atIndex:27];
const int64_t nwarps = 32;
const int64_t nhptg = 2; // heads per threadgroup
const int64_t nwarps = 8;
const int64_t nhptg = 2; // heads per threadgroup !! sync with kernel template arguments !!
const int64_t nqptg = 4; // queries per threadgroup !! sync with kernel template arguments !!
const size_t smem = (nhptg*ne00 + nwarps*(nhptg*ne00 + 32))*(sizeof(float)/2);
const size_t smem = nqptg*(nhptg*ne00 + nwarps*(nhptg*ne00 + 32))*(sizeof(float)/2);
GGML_ASSERT(smem <= ctx->device.maxThreadgroupMemoryLength);
[encoder setThreadgroupMemoryLength:smem atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake(ne01, (ne02 + nhptg - 1)/(nhptg), ne03) threadsPerThreadgroup:MTLSizeMake(32, nwarps, 1)];
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + nqptg - 1)/nqptg, (ne02 + nhptg - 1)/(nhptg), ne03) threadsPerThreadgroup:MTLSizeMake(32, nwarps, 1)];
} break;
case GGML_OP_DUP:
case GGML_OP_CPY: