root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-26 03:14:35 +00:00
Code Issues Actions 7 Packages Projects Releases Wiki Activity

metal : correct fix of kernel_norm (#3060)

Browse Source 
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
This commit is contained in:
Kawrakow 2023-09-07 15:42:42 +02:00 committed by GitHub
parent c4f496648c
commit be6beeb8d7
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 5 additions and 25 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

30
ggml-metal.metal
View File

@ -220,29 +220,14 @@ kernel void kernel_norm(
} }
threadgroup_barrier(mem_flags::mem_threadgroup); threadgroup_barrier(mem_flags::mem_threadgroup);
} }
// broadcast const float mean = sum[0] / ne00;
if (tpitg == 0) {
sum[0] /= ne00; // recenter and VARIANCE
}
threadgroup_barrier(mem_flags::mem_threadgroup); threadgroup_barrier(mem_flags::mem_threadgroup);
const float mean = sum[0];
// recenter
device float * y = dst + tgpig*ne00; device float * y = dst + tgpig*ne00;
for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
y[i00] = x[i00] - mean;
}
// VARIANCE
// parallel sum
//
// WARNING: combining this loop with the one above will give you wrong results for nth == 256
// I have no idea why, so for now I am keeping them separate. But this behavior is very concerning.
// Tested with:
// ./perplexity -m ./falcon-7b/ggml-model-q4_0.gguf -f wiki.test.raw -ngl 1 -t 4
//
sum[tpitg] = 0.0f; sum[tpitg] = 0.0f;
for (int i00 = tpitg; i00 < ne00; i00 += ntg) { for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
y[i00] = x[i00] - mean;
sum[tpitg] += y[i00] * y[i00]; sum[tpitg] += y[i00] * y[i00];
} }
@ -254,12 +239,7 @@ kernel void kernel_norm(
} }
threadgroup_barrier(mem_flags::mem_threadgroup); threadgroup_barrier(mem_flags::mem_threadgroup);
} }
// broadcast const float variance = sum[0] / ne00;
if (tpitg == 0) {
sum[0] /= ne00;
}
threadgroup_barrier(mem_flags::mem_threadgroup);
const float variance = sum[0];
const float scale = 1.0f/sqrt(variance + eps); const float scale = 1.0f/sqrt(variance + eps);
for (int i00 = tpitg; i00 < ne00; i00 += ntg) { for (int i00 = tpitg; i00 < ne00; i00 += ntg) {