diff options
| author | hydai <z54981220@gmail.com> | 2023-12-30 00:31:19 +0800 |
|---|---|---|
| committer | GitHub <noreply@github.com> | 2023-12-29 17:31:19 +0100 |
| commit | 91bb39cec7e4dfb9e2293509ef60298a67f0b1b7 (patch) | |
| tree | 9a0085ff74960585798ca43e70be5e65a14bf8c8 /ggml-cuda.cu | |
| parent | 04ac0607e913ab91234dfb240e12a76509e30982 (diff) | |
cuda: fix vmm oom issue on NVIDIA AGX Orin (#4687)
Signed-off-by: hydai <hydai@secondstate.io>
Diffstat (limited to 'ggml-cuda.cu')
| -rw-r--r-- | ggml-cuda.cu | 2 |
1 files changed, 1 insertions, 1 deletions
diff --git a/ggml-cuda.cu b/ggml-cuda.cu index 9a9effcf..09585b07 100644 --- a/ggml-cuda.cu +++ b/ggml-cuda.cu @@ -6662,7 +6662,7 @@ static void ggml_cuda_pool_free_leg(int device, void * ptr, size_t size) { // pool with virtual memory static CUdeviceptr g_cuda_pool_addr[GGML_CUDA_MAX_DEVICES] = {0}; static size_t g_cuda_pool_used[GGML_CUDA_MAX_DEVICES] = {0}; -static const size_t CUDA_POOL_VMM_MAX_SIZE = 1ull << 36; // 64 GB +static const size_t CUDA_POOL_VMM_MAX_SIZE = 1ull << 35; // 32 GB static void * ggml_cuda_pool_malloc_vmm(int device, size_t size, size_t * actual_size) { scoped_spin_lock lock(g_cuda_pool_lock); |