From 71ee33f62cd467166281d07cdeec65afa6408412 Mon Sep 17 00:00:00 2001
From: Aroun Demeure <arundemeure@gmail.com>
Date: Sun, 24 Aug 2025 10:17:07 +0100
Subject: [PATCH] Fix for CUDA 13

---
 llmc/cuda_utils.cuh | 9 ++++++++-
 1 file changed, 8 insertions(+), 1 deletion(-)

diff --git a/llmc/cuda_utils.cuh b/llmc/cuda_utils.cuh
index 030ec07..976ec94 100644
--- a/llmc/cuda_utils.cuh
+++ b/llmc/cuda_utils.cuh
@@ -217,7 +217,14 @@ int cudaMallocConditionallyManaged(void** out, size_t bytes, const char *file, i
         // if we OOM, fallback to a managed allocation. slower but at least won't crash.
         cudaGetLastError(); // reset the error before the next API call
         cudaCheck_(cudaMallocManaged(out, bytes), file, line);
-        cudaCheck_(cudaMemAdvise(*out, bytes, cudaMemAdviseSetPreferredLocation, cudaCpuDeviceId), file, line);
+        // Prefer host (CPU) location when falling back to managed memory
+        // Inline compatibility for CUDA 12+ vs older versions
+        #if defined(CUDART_VERSION) && (CUDART_VERSION >= 12000)
+            cudaMemLocation loc{};
+            loc.type = cudaMemLocationTypeHost;
+            loc.id = cudaCpuDeviceId;
+            cudaCheck_(cudaMemAdvise(*out, bytes, cudaMemAdviseSetPreferredLocation, loc), file, line);
+        #endif
         return 1;
     } else {
         cudaCheck_(err, file, line);