happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

vulkan: change memory_logger to be controlled by an env var (#18769)

This commit is contained in:
Jeff Bolz 2026-01-12 06:32:55 -06:00 committed by GitHub
parent ce3bf9b1a4
commit 8e2da778da
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 11 additions and 20 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

31
ggml/src/ggml-vulkan/ggml-vulkan.cpp
View File

@ -238,9 +238,7 @@ static ggml_backend_buffer_type_i ggml_backend_vk_buffer_type_interface = {
/* .is_host = */ NULL,
};
#ifdef GGML_VULKAN_MEMORY_DEBUG
class vk_memory_logger;
#endif
class vk_perf_logger;
static void ggml_vk_destroy_buffer(vk_buffer& buf);
static void ggml_vk_synchronize(ggml_backend_vk_context * ctx);
@ -825,9 +823,7 @@ struct vk_device_struct {
bool allow_sysmem_fallback;
bool disable_graph_optimize;
#ifdef GGML_VULKAN_MEMORY_DEBUG
std::unique_ptr<vk_memory_logger> memory_logger;
#endif
~vk_device_struct() {
VK_LOG_DEBUG("destroy device " << name);
@ -1563,8 +1559,9 @@ static void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx, vk_contex
static void ggml_vk_load_shaders(vk_device& device);
static void ggml_pipeline_allocate_descriptor_sets(ggml_backend_vk_context * ctx);
#if defined(GGML_VULKAN_MEMORY_DEBUG) || defined(GGML_VULKAN_DEBUG)
#define VK_LOG_MEMORY(msg) std::cerr << "ggml_vulkan memory: " << msg << std::endl
static bool vk_memory_logger_enabled = false;
#define VK_LOG_MEMORY(msg) if (vk_memory_logger_enabled) { std::cerr << "ggml_vulkan memory: " << msg << std::endl; }
static std::string format_size(size_t size) {
const size_t kib = 1024;
@ -1597,10 +1594,10 @@ private:
std::map<vk::Buffer, size_t> allocations; // Track allocations
size_t total_device;
size_t total_host;
static std::mutex log_mutex;
};
#else
#define VK_LOG_MEMORY(msg) ((void) 0)
#endif // GGML_VULKAN_MEMORY_DEBUG
std::mutex vk_memory_logger::log_mutex;
static bool vk_perf_logger_enabled = false;
static bool vk_perf_logger_concurrent = false;
@ -1907,10 +1904,10 @@ struct ggml_backend_vk_buffer_context {
}
};
#ifdef GGML_VULKAN_MEMORY_DEBUG
static std::mutex log_mutex;
void vk_memory_logger::log_allocation(vk_buffer_ref buf_ref, size_t size) {
if (!vk_memory_logger_enabled) {
return;
}
std::lock_guard<std::mutex> guard(log_mutex);
vk_buffer buf = buf_ref.lock();
const bool device = bool(buf->memory_property_flags & vk::MemoryPropertyFlagBits::eDeviceLocal);
@ -1922,7 +1919,7 @@ void vk_memory_logger::log_allocation(vk_buffer_ref buf_ref, size_t size) {
}
void vk_memory_logger::log_deallocation(vk_buffer_ref buf_ref) {
if (buf_ref.expired() || buf_ref.lock()->size == 0) {
if (buf_ref.expired() || buf_ref.lock()->size == 0 || !vk_memory_logger_enabled) {
return;
}
@ -1940,7 +1937,6 @@ void vk_memory_logger::log_deallocation(vk_buffer_ref buf_ref) {
VK_LOG_MEMORY("ERROR " << buf->device->name << ": Attempted to deallocate unknown " << type << " memory at " << buf->buffer);
}
}
#endif // GGML_VULKAN_MEMORY_DEBUG
struct vk_instance_t {
vk::Instance instance;
@ -2593,9 +2589,7 @@ static vk_buffer ggml_vk_create_buffer(vk_device& device, size_t size, const std
buf->bda_addr = device->device.getBufferAddress(addressInfo);
}
#ifdef GGML_VULKAN_MEMORY_DEBUG
device->memory_logger->log_allocation(buf, size);
#endif
return buf;
}
@ -2652,11 +2646,9 @@ static void ggml_vk_destroy_buffer(vk_buffer& buf) {
return;
}
#ifdef GGML_VULKAN_MEMORY_DEBUG
if (buf->device != nullptr) {
buf->device->memory_logger->log_deallocation(buf);
}
#endif
buf.reset();
}
@ -4477,9 +4469,7 @@ static vk_device ggml_vk_get_device(size_t idx) {
vk_device device = std::make_shared<vk_device_struct>();
vk_instance.devices[idx] = device;
#ifdef GGML_VULKAN_MEMORY_DEBUG
device->memory_logger = std::unique_ptr<vk_memory_logger>(new vk_memory_logger());
#endif
size_t dev_num = vk_instance.device_indices[idx];
@ -5476,6 +5466,7 @@ static void ggml_vk_instance_init() {
vk_perf_logger_enabled = getenv("GGML_VK_PERF_LOGGER") != nullptr;
vk_perf_logger_concurrent = getenv("GGML_VK_PERF_LOGGER_CONCURRENT") != nullptr;
vk_enable_sync_logger = getenv("GGML_VK_SYNC_LOGGER") != nullptr;
vk_memory_logger_enabled = getenv("GGML_VK_MEMORY_LOGGER") != nullptr;
const char* GGML_VK_PERF_LOGGER_FREQUENCY = getenv("GGML_VK_PERF_LOGGER_FREQUENCY");
if (GGML_VK_PERF_LOGGER_FREQUENCY != nullptr) {