From: Johannes Gäßler <redacted>
Date: Sun, 22 Mar 2026 16:53:33 +0000 (+0100)
Subject: CUDA: fix BF16 FA compilation (llama/20865)
X-Git-Tag: v0.9.9~24
X-Git-Url: https://git.djapps.eu/?a=commitdiff_plain;h=fd4e581b59290cc9ede8e87273580c92030da35d;p=pkg%2Fggml%2Fsources%2Fggml

CUDA: fix BF16 FA compilation (llama/20865)
---

diff --git a/src/ggml-cuda/convert.cuh b/src/ggml-cuda/convert.cuh
index b8caeacf..f5d37c7b 100644
--- a/src/ggml-cuda/convert.cuh
+++ b/src/ggml-cuda/convert.cuh
@@ -42,11 +42,15 @@ template<typename dst_t, typename src_t>
     } else if constexpr(std::is_same_v<src_t, float2> && std::is_same_v<dst_t, half2>) {
         return __float22half2_rn(x);
     } else if constexpr(std::is_same_v<src_t, nv_bfloat162> && std::is_same_v<dst_t, float2>) {
-#if !defined(GGML_USE_HIP) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 800
+#ifdef GGML_USE_HIP
+        return make_float2(__bfloat162float(__low2bfloat16(x)), __bfloat162float(__high2bfloat16(x)));
+#else
+#if __CUDA_ARCH__ >= 800
         return __bfloat1622float2(x);
 #else
-        return make_float2(__bfloat162float(__low2bfloat16(x)), __bfloat162float(__high2bfloat16(x)));
-#endif // !defined(GGML_USE_HIP) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 800
+        return make_float2(__bfloat162float(x.x), __bfloat162float(x.y));
+#endif // __CUDA_ARCH__ >= 800
+#endif // GGML_USE_HIP
     } else if constexpr(std::is_same_v<src_t, float2> && std::is_same_v<dst_t, nv_bfloat162>) {
         // bypass compile error on cuda 12.0.1
 #ifdef GGML_USE_HIP