mirrors/ollama-for-amd
1
0
Fork 0
mirror of https://github.com/likelovewant/ollama-for-amd.git synced 2025-12-21 22:33:56 +00:00
Code Issues Packages Projects Releases Wiki Activity

discovery: only retry AMD GPUs (#12894)

Browse Source 
* discovery: only retry AMD GPUs

CUDA and Vulkan don't crash on unsupported devices, so retry isn't necessary.
This also refactors the code to shift the Library specific logic into the ml
package.

* review comments
This commit is contained in:
Daniel Hiltgen
2025-11-04 15:33:46 -08:00
committed by GitHub
parent 220e133fca
commit 27f1fde413
9 changed files with 96 additions and 137 deletions
Download Patch File Download Diff File Expand all files Collapse all files

53
llama/patches/0026-GPU-discovery-enhancements.patch
View File

@@ -14,24 +14,24 @@ Vulkan PCI and Memory
fix vulkan PCI ID and ID handling
---
ggml/include/ggml-backend.h | 8 +
ggml/include/ggml-backend.h | 6 +
ggml/src/CMakeLists.txt | 2 +
ggml/src/ggml-cuda/ggml-cuda.cu | 65 ++++
ggml/src/ggml-cuda/vendors/hip.h | 3 +
ggml/src/ggml-impl.h | 8 +
ggml/src/ggml-metal/ggml-metal.cpp | 2 +
ggml/src/ggml-vulkan/ggml-vulkan.cpp | 212 +++++++++++--
ggml/src/ggml-vulkan/ggml-vulkan.cpp | 209 +++++++++++--
ggml/src/mem_hip.cpp | 452 +++++++++++++++++++++++++++
ggml/src/mem_nvml.cpp | 209 +++++++++++++
9 files changed, 931 insertions(+), 30 deletions(-)
9 files changed, 926 insertions(+), 30 deletions(-)
create mode 100644 ggml/src/mem_hip.cpp
create mode 100644 ggml/src/mem_nvml.cpp
diff --git a/ggml/include/ggml-backend.h b/ggml/include/ggml-backend.h
index ba181d09d..809835243 100644
index 69223c488..6510e0cba 100644
--- a/ggml/include/ggml-backend.h
+++ b/ggml/include/ggml-backend.h
@@ -169,6 +169,14 @@ extern "C" {
@@ -169,6 +169,12 @@ extern "C" {
const char * device_id;
// device capabilities
struct ggml_backend_dev_caps caps;
@@ -41,8 +41,6 @@ index ba181d09d..809835243 100644
+ int compute_minor;
+ int integrated;
+ const char *library;
+ // number with which the devices are accessed (Vulkan)
+ const char *numeric_id;
};
GGML_API const char * ggml_backend_dev_name(ggml_backend_dev_t device);
@@ -60,7 +58,7 @@ index 0609c6503..aefe43bdd 100644
target_include_directories(ggml-base PRIVATE .)
diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
index 87c6c34a4..b075a18be 100644
index 5787e8cd5..d232bf828 100644
--- a/ggml/src/ggml-cuda/ggml-cuda.cu
+++ b/ggml/src/ggml-cuda/ggml-cuda.cu
@@ -261,6 +261,16 @@ static ggml_cuda_device_info ggml_cuda_init() {
@@ -92,7 +90,7 @@ index 87c6c34a4..b075a18be 100644
GGML_LOG_INFO(" Device %d: %s, compute capability %d.%d, VMM: %s, ID: %s\n",
id, prop.name, prop.major, prop.minor, device_vmm ? "yes" : "no",
ggml_cuda_parse_uuid(prop, id).c_str());
@@ -3484,6 +3499,11 @@ struct ggml_backend_cuda_device_context {
@@ -3476,6 +3491,11 @@ struct ggml_backend_cuda_device_context {
std::string description;
std::string pci_bus_id;
std::string id;
@@ -104,7 +102,7 @@ index 87c6c34a4..b075a18be 100644
};
static const char * ggml_backend_cuda_device_get_name(ggml_backend_dev_t dev) {
@@ -3504,6 +3524,28 @@ static const char * ggml_backend_cuda_device_get_id(ggml_backend_dev_t dev) {
@@ -3496,6 +3516,28 @@ static const char * ggml_backend_cuda_device_get_id(ggml_backend_dev_t dev) {
static void ggml_backend_cuda_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
ggml_backend_cuda_device_context * ctx = (ggml_backend_cuda_device_context *)dev->context;
ggml_cuda_set_device(ctx->device);
@@ -133,7 +131,7 @@ index 87c6c34a4..b075a18be 100644
CUDA_CHECK(cudaMemGetInfo(free, total));
}
@@ -3512,6 +3554,7 @@ static enum ggml_backend_dev_type ggml_backend_cuda_device_get_type(ggml_backend
@@ -3504,6 +3546,7 @@ static enum ggml_backend_dev_type ggml_backend_cuda_device_get_type(ggml_backend
return GGML_BACKEND_DEVICE_TYPE_GPU;
}
@@ -141,7 +139,7 @@ index 87c6c34a4..b075a18be 100644
static void ggml_backend_cuda_device_get_props(ggml_backend_dev_t dev, ggml_backend_dev_props * props) {
ggml_backend_cuda_device_context * ctx = (ggml_backend_cuda_device_context *)dev->context;
@@ -3525,6 +3568,19 @@ static void ggml_backend_cuda_device_get_props(ggml_backend_dev_t dev, ggml_back
@@ -3517,6 +3560,19 @@ static void ggml_backend_cuda_device_get_props(ggml_backend_dev_t dev, ggml_back
// If you need the memory data, call ggml_backend_dev_memory() explicitly.
props->memory_total = props->memory_free = 0;
@@ -161,7 +159,7 @@ index 87c6c34a4..b075a18be 100644
bool host_buffer = getenv("GGML_CUDA_NO_PINNED") == nullptr;
#ifdef GGML_CUDA_NO_PEER_COPY
bool events = false;
@@ -4087,6 +4143,7 @@ ggml_backend_reg_t ggml_backend_cuda_reg() {
@@ -4079,6 +4135,7 @@ ggml_backend_reg_t ggml_backend_cuda_reg() {
std::lock_guard<std::mutex> lock(mutex);
if (!initialized) {
ggml_backend_cuda_reg_context * ctx = new ggml_backend_cuda_reg_context;
@@ -169,7 +167,7 @@ index 87c6c34a4..b075a18be 100644
for (int i = 0; i < ggml_cuda_info().device_count; i++) {
ggml_backend_cuda_device_context * dev_ctx = new ggml_backend_cuda_device_context;
@@ -4102,6 +4159,14 @@ ggml_backend_reg_t ggml_backend_cuda_reg() {
@@ -4094,6 +4151,14 @@ ggml_backend_reg_t ggml_backend_cuda_reg() {
snprintf(pci_bus_id, sizeof(pci_bus_id), "%04x:%02x:%02x.0", prop.pciDomainID, prop.pciBusID, prop.pciDeviceID);
dev_ctx->pci_bus_id = pci_bus_id;
@@ -225,10 +223,10 @@ index d0fb3bcca..b63edd0c1 100644
}
#endif
diff --git a/ggml/src/ggml-metal/ggml-metal.cpp b/ggml/src/ggml-metal/ggml-metal.cpp
index f2ff9f322..f356e4a0a 100644
index 05ff6a5a6..032dee76d 100644
--- a/ggml/src/ggml-metal/ggml-metal.cpp
+++ b/ggml/src/ggml-metal/ggml-metal.cpp
@@ -535,6 +535,7 @@ static enum ggml_backend_dev_type ggml_backend_metal_device_get_type(ggml_backen
@@ -537,6 +537,7 @@ static enum ggml_backend_dev_type ggml_backend_metal_device_get_type(ggml_backen
GGML_UNUSED(dev);
}
@@ -236,7 +234,7 @@ index f2ff9f322..f356e4a0a 100644
static void ggml_backend_metal_device_get_props(ggml_backend_dev_t dev, ggml_backend_dev_props * props) {
props->name = ggml_backend_metal_device_get_name(dev);
props->description = ggml_backend_metal_device_get_description(dev);
@@ -543,6 +544,7 @@ static void ggml_backend_metal_device_get_props(ggml_backend_dev_t dev, ggml_bac
@@ -545,6 +546,7 @@ static void ggml_backend_metal_device_get_props(ggml_backend_dev_t dev, ggml_bac
ggml_backend_metal_device_get_memory(dev, &props->memory_free, &props->memory_total);
@@ -245,7 +243,7 @@ index f2ff9f322..f356e4a0a 100644
/* .async = */ true,
/* .host_buffer = */ false,
diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
index ed83236f4..0bbcecd01 100644
index bd3ece516..7cfb14a54 100644
--- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
@@ -231,6 +231,7 @@ class vk_memory_logger;
@@ -286,7 +284,7 @@ index ed83236f4..0bbcecd01 100644
// backend interface
#define UNUSED GGML_UNUSED
@@ -12391,31 +12415,103 @@ void ggml_backend_vk_get_device_description(int device, char * description, size
@@ -12392,31 +12416,102 @@ void ggml_backend_vk_get_device_description(int device, char * description, size
ggml_vk_get_device_description(dev_idx, description, description_size);
}
@@ -309,7 +307,6 @@ index ed83236f4..0bbcecd01 100644
+ std::string pci_id;
+ std::string id;
+ std::string uuid;
+ std::string numeric_id;
+ int major;
+ int minor;
+ int driver_major;
@@ -407,7 +404,7 @@ index ed83236f4..0bbcecd01 100644
break;
}
}
@@ -12448,8 +12544,13 @@ static std::string ggml_backend_vk_get_device_pci_id(int device_idx) {
@@ -12449,8 +12544,13 @@ static std::string ggml_backend_vk_get_device_pci_id(int device_idx) {
}
}
@@ -422,7 +419,7 @@ index ed83236f4..0bbcecd01 100644
}
vk::PhysicalDeviceProperties2 props = {};
@@ -12466,19 +12567,24 @@ static std::string ggml_backend_vk_get_device_pci_id(int device_idx) {
@@ -12467,19 +12567,24 @@ static std::string ggml_backend_vk_get_device_pci_id(int device_idx) {
char pci_bus_id[16] = {};
snprintf(pci_bus_id, sizeof(pci_bus_id), "%04x:%02x:%02x.%x", pci_domain, pci_bus, pci_device, pci_function);
@@ -456,7 +453,7 @@ index ed83236f4..0bbcecd01 100644
static const char * ggml_backend_vk_device_get_name(ggml_backend_dev_t dev) {
ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)dev->context;
@@ -12490,9 +12596,14 @@ static const char * ggml_backend_vk_device_get_description(ggml_backend_dev_t de
@@ -12491,9 +12596,14 @@ static const char * ggml_backend_vk_device_get_description(ggml_backend_dev_t de
return ctx->description.c_str();
}
@@ -472,7 +469,7 @@ index ed83236f4..0bbcecd01 100644
}
static ggml_backend_buffer_type_t ggml_backend_vk_device_get_buffer_type(ggml_backend_dev_t dev) {
@@ -12516,8 +12627,9 @@ static void ggml_backend_vk_device_get_props(ggml_backend_dev_t dev, struct ggml
@@ -12517,8 +12627,9 @@ static void ggml_backend_vk_device_get_props(ggml_backend_dev_t dev, struct ggml
props->name = ggml_backend_vk_device_get_name(dev);
props->description = ggml_backend_vk_device_get_description(dev);
@@ -483,7 +480,7 @@ index ed83236f4..0bbcecd01 100644
ggml_backend_vk_device_get_memory(dev, &props->memory_free, &props->memory_total);
props->caps = {
/* .async = */ false,
@@ -12525,6 +12637,14 @@ static void ggml_backend_vk_device_get_props(ggml_backend_dev_t dev, struct ggml
@@ -12526,6 +12637,13 @@ static void ggml_backend_vk_device_get_props(ggml_backend_dev_t dev, struct ggml
/* .buffer_from_host_ptr = */ false,
/* .events = */ false,
};
@@ -494,11 +491,10 @@ index ed83236f4..0bbcecd01 100644
+ props->driver_minor = ctx->driver_minor;
+ props->integrated = ctx->is_integrated_gpu;
+ props->library = GGML_VK_NAME;
+ props->numeric_id = ctx->numeric_id.c_str();
}
static ggml_backend_t ggml_backend_vk_device_init(ggml_backend_dev_t dev, const char * params) {
@@ -12953,6 +13073,8 @@ static ggml_backend_dev_t ggml_backend_vk_reg_get_device(ggml_backend_reg_t reg,
@@ -12954,6 +13072,8 @@ static ggml_backend_dev_t ggml_backend_vk_reg_get_device(ggml_backend_reg_t reg,
static std::mutex mutex;
std::lock_guard<std::mutex> lock(mutex);
if (!initialized) {
@@ -507,7 +503,7 @@ index ed83236f4..0bbcecd01 100644
for (int i = 0; i < ggml_backend_vk_get_device_count(); i++) {
ggml_backend_vk_device_context * ctx = new ggml_backend_vk_device_context;
char desc[256];
@@ -12961,12 +13083,42 @@ static ggml_backend_dev_t ggml_backend_vk_reg_get_device(ggml_backend_reg_t reg,
@@ -12962,12 +13082,41 @@ static ggml_backend_dev_t ggml_backend_vk_reg_get_device(ggml_backend_reg_t reg,
ctx->name = GGML_VK_NAME + std::to_string(i);
ctx->description = desc;
ctx->is_integrated_gpu = ggml_backend_vk_get_device_type(i) == vk::PhysicalDeviceType::eIntegratedGpu;
@@ -547,7 +543,6 @@ index ed83236f4..0bbcecd01 100644
+ // TODO regex parse driver_props.driverInfo for a X.Y or X.Y.Z version string
+ ctx->driver_major = 0;
+ ctx->driver_minor = 0;
+ ctx->numeric_id = std::to_string(i);
}
initialized = true;
}

12
llama/patches/0031-Add-memory-detection-using-DXGI-PDH.patch → llama/patches/0030-Add-memory-detection-using-DXGI-PDH.patch
View File

@@ -38,7 +38,7 @@ index b63edd0c1..81cad8cf3 100644
#ifdef __cplusplus
}
diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
index cc68e7968..27d6574da 100644
index 7cfb14a54..a1c46d0b3 100644
--- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
@@ -73,6 +73,7 @@ DispatchLoaderDynamic & ggml_vk_default_dispatcher();
@@ -54,10 +54,10 @@ index cc68e7968..27d6574da 100644
std::string id;
std::string uuid;
+ std::string luid;
std::string numeric_id;
int major;
int minor;
@@ -12449,8 +12451,22 @@ void ggml_backend_vk_get_device_memory(ggml_backend_vk_device_context *ctx, size
int driver_major;
@@ -12448,8 +12450,22 @@ void ggml_backend_vk_get_device_memory(ggml_backend_vk_device_context *ctx, size
vk::PhysicalDeviceMemoryProperties memprops = vkdev.getMemoryProperties();
vk::PhysicalDeviceProperties2 props2;
vkdev.getProperties2(&props2);
@@ -81,7 +81,7 @@ index cc68e7968..27d6574da 100644
{
// Use vendor specific management libraries for best VRAM reporting if available
switch (props2.properties.vendorID) {
@@ -12478,8 +12494,8 @@ void ggml_backend_vk_get_device_memory(ggml_backend_vk_device_context *ctx, size
@@ -12477,8 +12493,8 @@ void ggml_backend_vk_get_device_memory(ggml_backend_vk_device_context *ctx, size
break;
}
}
@@ -91,7 +91,7 @@ index cc68e7968..27d6574da 100644
*total = 0;
*free = 0;
vk::PhysicalDeviceMemoryBudgetPropertiesEXT mem_budget_props;
@@ -13091,7 +13107,6 @@ static ggml_backend_dev_t ggml_backend_vk_reg_get_device(ggml_backend_reg_t reg,
@@ -13089,7 +13105,6 @@ static ggml_backend_dev_t ggml_backend_vk_reg_get_device(ggml_backend_reg_t reg,
/* .reg = */ reg,
/* .context = */ ctx,
});
@@ -99,7 +99,7 @@ index cc68e7968..27d6574da 100644
// Gather additional information about the device
int dev_idx = vk_instance.device_indices[i];
vk::PhysicalDeviceProperties props1;
@@ -13114,6 +13129,14 @@ static ggml_backend_dev_t ggml_backend_vk_reg_get_device(ggml_backend_reg_t reg,
@@ -13112,6 +13127,14 @@ static ggml_backend_dev_t ggml_backend_vk_reg_get_device(ggml_backend_reg_t reg,
}
}
ctx->uuid = oss.str();

4
llama/patches/0032-interleave-multi-rope.patch → llama/patches/0031-interleave-multi-rope.patch
View File

@@ -1,6 +1,6 @@
From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
From: Michael Yang <git@mxy.ng>
Date: Web, 16 Oct 2025 20:37:19 -0700
Date: Thu, 16 Oct 2025 20:37:19 -0700
Subject: [PATCH] interleave multi rope
since ollama doesn't use mrope for anything else, change it to mean the
@@ -85,7 +85,7 @@ index 375a0c7fd..9866c96b4 100644
// end of mrope
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/rope_multi.comp b/ggml/src/ggml-vulkan/vulkan-shaders/rope_multi.comp
index 111286b49..6fc2b42f8 100644
index 111286b49..633dc20ff 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/rope_multi.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/rope_multi.comp
@@ -31,19 +31,13 @@ void main() {