From caa8fba0cc60e72e63e3510e2005e2f56416d670 Mon Sep 17 00:00:00 2001
From: Michael Wand <michael@baybridgeaquarium.com>
Date: Tue, 24 Mar 2026 02:04:55 -0700
Subject: [PATCH] Added check for dst_t to cuda_cast template for float

---
 ggml/src/ggml-cuda/convert.cuh | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/ggml/src/ggml-cuda/convert.cuh b/ggml/src/ggml-cuda/convert.cuh
index c429583661..c22282139f 100644
--- a/ggml/src/ggml-cuda/convert.cuh
+++ b/ggml/src/ggml-cuda/convert.cuh
@@ -62,8 +62,8 @@ template<typename dst_t, typename src_t>
 #else
         return {x.x, x.y};
 #endif // GGML_USE_HIP
-    } else if constexpr (std::is_same_v<src_t, ue4m3>) {
-#if defined(__CUDA_ARCH__)
+    } else if constexpr (std::is_same_v<src_t, ue4m3> && std::is_same_v<dst_t, float>) {
+    #if defined(__CUDA_ARCH__)
 #if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && defined(CUDART_VERSION) && CUDART_VERSION >= 12050 // This matches cuda_fp8.h's version gate.
         // This uses the same fp8 conversion that __nv_fp8_e4m3 uses internally.
         __half h = __half(__nv_cvt_fp8_to_halfraw((__nv_fp8_storage_t) x.x, __NV_E4M3));