Skip to content

LLM inference in C/C++

C++ ggml

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

Jassieluoe7e3f350904c
diff --git a/ggml/src/ggml-sycl/softmax.cpp b/ggml/src/ggml-sycl/softmax.cpp
index 18bf379bbe..67ea282b4b 100644
--- a/ggml/src/ggml-sycl/softmax.cpp
+++ b/ggml/src/ggml-sycl/softmax.cpp
@@ -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;