docs: add notes on CMAKE_CUDA_FLAGS

danbev · Jan 2, 2025 · bcee186 · bcee186
1 parent ed4e62f
commit bcee186
Showing 1 changed file with 9 additions and 0 deletions.
diff --git a/notes/cuda.md b/notes/cuda.md
@@ -188,3 +188,12 @@ When using streams we use the async version of memcpy, like `cudaMemcpyAsync`,
 and for launching a kernel we specify a stream as an argument.
 
 An example can be found in [streams.cu](../gpu/cuda/src/streams.cu).
+
+###  CMAKE_CUDA_FLAGS
+This is a cmake variable that is passed to `nvcc` when compiling CUDA code.
+
+#### FASTFP16_AVAILABLE
+This flag tells nvcc to enable the FP16 code paths in the CUDA code.
+```console
+$ cmake -S . -B build -DGGML_CUDA=ON -DCMAKE_CUDA_FLAGS="-DFASTFP16_AVAILABLE"
+```