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

metal : f16 precision

Browse source
This commit is contained in:
Georgi Gerganov 2024-01-21 11:13:24 +02:00
parent 1173f49c3b
commit 528da7515e
No known key found for this signature in database
GPG key ID: 449E073F9DC10735
2 changed files with 26 additions and 20 deletions
Download patch file Download diff file Expand all files Collapse all files

6
ggml-metal.m
View file

@ -2237,8 +2237,10 @@ static bool ggml_metal_graph_compute(
const int nwarps = 1;
GGML_ASSERT(2*32*nwarps*ne00*sizeof(float) <= ctx->device.maxThreadgroupMemoryLength);
[encoder setThreadgroupMemoryLength:2*32*nwarps*ne00*sizeof(float) atIndex:0];
const size_t shalf = sizeof(float)/2;
GGML_ASSERT(2*32*nwarps*ne00*shalf <= ctx->device.maxThreadgroupMemoryLength);
[encoder setThreadgroupMemoryLength:2*32*nwarps*ne00*shalf atIndex:0];
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 31)/32, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(32, 1, 1)];
} break;