From: 0cc4m <redacted>
Date: Tue, 18 Mar 2025 06:21:40 +0000 (+0100)
Subject: Vulkan: Default to 1GB allocations instead of 4GB to avoid fragmentation and driver... 
X-Git-Tag: upstream/0.0.5028~119
X-Git-Url: https://git.djapps.eu/?a=commitdiff_plain;h=fd123cfead49eb32e386e26b8ef7a6d41554dda5;p=pkg%2Fggml%2Fsources%2Fllama.cpp

Vulkan: Default to 1GB allocations instead of 4GB to avoid fragmentation and driver issues (#12434)
---

diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
index c0ee5dad..dd680aa5 100644
--- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
@@ -2524,13 +2524,9 @@ static vk_device ggml_vk_get_device(size_t idx) {
 
         if (GGML_VK_SUBALLOCATION_BLOCK_SIZE != nullptr) {
             device->suballocation_block_size = std::stoul(GGML_VK_SUBALLOCATION_BLOCK_SIZE);
-#if defined(_WIN32)
-        } else if (device->vendor_id == VK_VENDOR_ID_NVIDIA) {
+        } else {
             // Limit batching of allocations to 1GB by default to avoid fragmentation issues
             device->suballocation_block_size = 1024*1024*1024;
-#endif
-        } else {
-            device->suballocation_block_size = device->max_memory_allocation_size;
         }
         device->suballocation_block_size = std::min(device->suballocation_block_size, device->max_memory_allocation_size);