[Bugfix][kernels] Fix half2float conversion in gguf kernels (#15995)

Signed-off-by: Isotr0py <2037008807@qq.com>

[Bugfix][kernels] Fix half2float conversion in gguf kernels (#15995)
Signed-off-by: Isotr0py <2037008807@qq.com>
230b131b · Isotr0py · GitHub · 0812d8dd · 230b131b
Unverified Commit 230b131b authored Apr 05, 2025 by Isotr0py Committed by GitHub Apr 04, 2025
Show whitespace changes
Inline Side-by-side

Showing with 5 additions and 0 deletions

csrc/quantization/gguf/ggml-common.h csrc/quantization/gguf/ggml-common.h +5 -0

No files found.
--- a/csrc/quantization/gguf/ggml-common.h
+++ b/csrc/quantization/gguf/ggml-common.h
@@ -1090,6 +1090,11 @@ __device__ __forceinline__ c10::BFloat16 convert_from_half<c10::BFloat16>(half v
 #endif  // defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 800
 }

+template<>
+__device__ __forceinline__ float convert_from_half<float>(half val) {
+    return __half2float(val);
+}
+
 #if defined(USE_ROCM)

 #ifndef __has_builtin