diff --git a/ggml/src/ggml-cuda.cu b/ggml/src/ggml-cuda.cu
index 8f07eb5f..b8a8ebb6 100644
--- a/ggml/src/ggml-cuda.cu
+++ b/ggml/src/ggml-cuda.cu
@@ -4536,7 +4536,7 @@ struct cuda_params {
     int  fusion = GGML_CUDA_FUSION;
     int  offload_batch_size = GGML_CUDA_MIN_BATCH_OFFLOAD;
     int  mmq_id_thresh = 32;
-    float fa_offset = 0;
+    float fa_offset = 0.6931f;
 #ifdef USE_CUDA_GRAPH
     bool use_cuda_graph = true;
 #else
diff --git a/ggml/src/ggml-cuda/common.cuh b/ggml/src/ggml-cuda/common.cuh
index f4209197..db67d1cc 100644
--- a/ggml/src/ggml-cuda/common.cuh
+++ b/ggml/src/ggml-cuda/common.cuh
@@ -853,7 +853,7 @@ struct ggml_backend_cuda_context {
     int   fusion = GGML_CUDA_FUSION;
     int   offload_batch_size = GGML_CUDA_MIN_BATCH_OFFLOAD;
     int   mmq_id_thresh = 32;
-    float fa_offset = 0.0f;
+    float fa_offset = 0.6931f; // ln(2)
 #ifdef USE_CUDA_GRAPH
     bool use_cuda_graph = true;