vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Skip GPUs without mem pool support.

Browse source
This commit is contained in:
Oleksii Maryshchenko 2023-11-04 17:50:59 +01:00
parent 81931b2ea7
commit 863166b4c3
1 changed files with 6 additions and 3 deletions
Download patch file Download diff file Expand all files Collapse all files

9
ggml-cuda.cu
View file

@ -5869,7 +5869,7 @@ void ggml_init_cublas() {
} }
} else { } else {
g_cudaMemPools[id] = nullptr; g_cudaMemPools[id] = nullptr;
fprintf(stderr, ", CUDA memory pool is not supported (cant load default pool)\n"); fprintf(stderr, ", CUDA memory pool is not supported (can't load default pool)\n");
} }
// test alloc/dealoc // test alloc/dealoc
if (err == cudaSuccess) { if (err == cudaSuccess) {
@ -5887,6 +5887,8 @@ void ggml_init_cublas() {
fprintf(stderr, ", CUDA memory pool is not supported (allocation failed)\n"); fprintf(stderr, ", CUDA memory pool is not supported (allocation failed)\n");
} }
} }
} else {
fprintf(stderr, ", CUDA memory pool is not supported\n");
} }
#endif #endif
g_tensor_split[id] = total_vram; g_tensor_split[id] = total_vram;
@ -5915,6 +5917,7 @@ void ggml_init_cublas() {
fprintf(stderr, fprintf(stderr,
"Warning: Device %d doesnt support CUDA memory pool, skipping pool access config\n", "Warning: Device %d doesnt support CUDA memory pool, skipping pool access config\n",
id); id);
continue;
} }
cudaMemAccessDesc desc_device = {}; cudaMemAccessDesc desc_device = {};
@ -5923,13 +5926,13 @@ void ggml_init_cublas() {
desc_device.flags = cudaMemAccessFlagsProtReadWrite; desc_device.flags = cudaMemAccessFlagsProtReadWrite;
cudaError_t err = cudaMemPoolSetAccess(main_device_pool, &desc_device, 1 /* numDescs */); cudaError_t err = cudaMemPoolSetAccess(main_device_pool, &desc_device, 1 /* numDescs */);
if (err != cudaSuccess) { if (err != cudaSuccess) {
fprintf(stderr, "Cant give access for main device memory pool to device %d\n", id); fprintf(stderr, "Can't give access for main device memory pool to device %d\n", id);
} }
cudaMemPool_t mempool; cudaMemPool_t mempool;
CUDA_CHECK(cudaDeviceGetDefaultMemPool(&mempool, id)); CUDA_CHECK(cudaDeviceGetDefaultMemPool(&mempool, id));
err = cudaMemPoolSetAccess(mempool, &desc_main_device, 1 /* numDescs */); err = cudaMemPoolSetAccess(mempool, &desc_main_device, 1 /* numDescs */);
if (err != cudaSuccess) { if (err != cudaSuccess) {
fprintf(stderr, "Cant give access for device %d memory pool to main device \n", id); fprintf(stderr, "Can't give access for device %d memory pool to main device \n", id);
} }
} }
} else { } else {