From: hydai <redacted>
Date: Fri, 29 Dec 2023 16:31:19 +0000 (+0800)
Subject: cuda: fix vmm oom issue on NVIDIA AGX Orin (#4687)
X-Git-Tag: upstream/0.0.4488~2762
X-Git-Url: https://git.djapps.eu/?a=commitdiff_plain;h=91bb39cec7e4dfb9e2293509ef60298a67f0b1b7;p=pkg%2Fggml%2Fsources%2Fllama.cpp

cuda: fix vmm oom issue on NVIDIA AGX Orin (#4687)

Signed-off-by: hydai <redacted>
---

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);