fallback to CPU buffer if host buffer alloc fails

ggml-cuda.cu

@@ -6729,8 +6729,7 @@ void * ggml_cuda_host_malloc(size_t size) {
     void * ptr = nullptr;
     cudaError_t err = cudaMallocHost((void **) &ptr, size);
     if (err != cudaSuccess) {
-        // The allocation error can be bypassed. A null ptr will assigned out of this function.
-        // This can fixed the OOM error in WSL.
+        // clear the error
         cudaGetLastError();
         fprintf(stderr, "WARNING: failed to allocate %.2f MB of pinned memory: %s\n",
             size/1024.0/1024.0, cudaGetErrorString(err));
@@ -9674,12 +9673,14 @@ ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device) {
 // host buffer type
 
 static void ggml_backend_cuda_host_buffer_free_buffer(ggml_backend_buffer_t buffer) {
-    CUDA_CHECK(cudaFreeHost(buffer->context));
+    ggml_cuda_host_free(buffer->context);
 }
 
 static ggml_backend_buffer_t ggml_backend_cuda_host_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
-    void * ptr;
-    CUDA_CHECK(cudaMallocHost(&ptr, size));
+    void * ptr = ggml_cuda_host_malloc(size);
+    if (ptr == nullptr) {
+        return nullptr;
+    }
 
     // FIXME: this is a hack to avoid having to implement a new buffer type
     ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(ptr, size);

llama.cpp

@@ -1177,21 +1177,27 @@ static std::string llama_token_to_piece(const struct llama_context * ctx, llama_
 }
 
 static ggml_backend_buffer_type_t llama_default_buffer_type(int n_gpu_layers) {
+    ggml_backend_buffer_type_t buft = nullptr;
+
 #ifdef GGML_USE_METAL
     if (n_gpu_layers > 0) {
-        return ggml_backend_metal_buffer_type();
+        buft = ggml_backend_metal_buffer_type();
     }
 #elif defined(GGML_USE_CUBLAS) && defined(LLAMA_GGML_BACKEND_CUDA_TEST)
     if (n_gpu_layers > 0) {
-        return ggml_backend_cuda_buffer_type(0);
+        buft = ggml_backend_cuda_buffer_type(0);
     }
 #elif defined(GGML_USE_CUBLAS)
-    return ggml_backend_cuda_host_buffer_type();
+    buft = ggml_backend_cuda_host_buffer_type();
 #elif defined(GGML_USE_CPU_HBM)
-    return ggml_backend_cpu_hbm_buffer_type();
+    buft = ggml_backend_cpu_hbm_buffer_type();
 #endif
 
-    return ggml_backend_cpu_buffer_type();
+    if (buft == nullptr) {
+        buft = ggml_backend_cpu_buffer_type();
+    }
+
+    return buft;
 
     GGML_UNUSED(n_gpu_layers);
 }