jdelony/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-06-27 23:50:20 -05:00
Code Issues Packages Projects Releases Wiki Activity

sycl : clamp softmax input to avoid underflow (#24941)

Browse Source
This commit is contained in:
Jassieluo 2026-06-26 15:02:42 +08:00 committed by GitHub
parent b11f7c16bc
commit e7e3f35090
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 2 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
ggml/src/ggml-sycl/softmax.cpp
View File

@ -126,7 +126,7 @@ static void soft_max_f32(const float * x,
break;
}
const float val = sycl::native::exp(vals[col] - max_val);
const float val = sycl::native::exp(sycl::max(vals[col] - max_val, -80.0f));
tmp += val;
vals[col] = val;
}
@ -154,7 +154,7 @@ static void soft_max_f32(const float * x,
tmp = warp_reduce_sum<WARP_SIZE>(tmp);
}
if (sinks) {
tmp += sycl::native::exp(sinks[i02] - max_val);
tmp += sycl::native::exp(sycl::max(sinks[i02] - max_val, -80.0f));
}
const float inv_sum = 1.0f / tmp;