cuda : fix LLAMA_CUDA_F16 build (#6298)

diff --git a/ggml-cuda/dmmv.cu b/ggml-cuda/dmmv.cu
index f91732df56143f4b5319c5d0e32228ca20336c14..7156c9219e31eabe992645ff2da6b4b5f5ef6b0b 100644 (file)
--- a/ggml-cuda/dmmv.cu
+++ b/ggml-cuda/dmmv.cu
@@ -1,5 +1,6 @@
 #include "dmmv.cuh"
 #include "dequantize.cuh"
+#include "convert.cuh"
 
 // dmmv = dequantize_mul_mat_vec
 #ifndef GGML_CUDA_DMMV_X