Public/koboldcpp
1
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-09-20 01:31:42 +02:00
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'upstream' into concedo_experimental

Browse Source 
# Conflicts:
#	.gitignore
#	CONTRIBUTING.md
#	Makefile
#	examples/llava/CMakeLists.txt
#	scripts/sync-ggml-am.sh
#	scripts/sync-ggml.last
#	scripts/sync-ggml.sh
#	src/llama-vocab.cpp
This commit is contained in:
Concedo
2024-08-10 11:42:32 +08:00
parent 6dd3d5515e b72942fac9
commit bdfe8526b8
44 changed files with 2241 additions and 439 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-aarch64.c
+2
View File
@@ -16,6 +16,8 @@
#if defined(__GNUC__)
#pragma GCC diagnostic ignored "-Woverlength-strings"
#elif defined(_MSC_VER)
#pragma warning(disable: 4244 4267) // possible loss of data
#endif
#define UNUSED GGML_UNUSED
ggml/src/ggml-backend.c
+15 -10
View File
@@ -351,15 +351,10 @@ void ggml_backend_tensor_copy_async(ggml_backend_t backend_src, ggml_backend_t b
}
// an async copy would normally happen after all the queued operations on both backends are completed
// sync src, set_async dst
if (ggml_backend_buffer_is_host(src->buffer)) {
ggml_backend_synchronize(backend_src);
ggml_backend_tensor_set_async(backend_dst, dst, src->data, 0, ggml_nbytes(src));
} else {
ggml_backend_synchronize(backend_src);
ggml_backend_tensor_copy(src, dst);
ggml_backend_synchronize(backend_dst);
}
// to simulate the same behavior, we need to synchronize both backends first, and do a blocking copy
ggml_backend_synchronize(backend_src);
ggml_backend_synchronize(backend_dst);
ggml_backend_tensor_copy(src, dst);
}
// events
@@ -1782,7 +1777,17 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
} else {
ggml_backend_synchronize(split_backend);
}
ggml_backend_tensor_copy_async(input_backend, split_backend, input, input_cpy);
// try async copy, but if not possible, we can still use a sync copy without synchronizing the dst backend, since we handle the synchronization here with multiple copies and events
// TODO: add public function to facilitate this, since applications do not have direct access to the backend interface
if (!split_backend->iface.cpy_tensor_async || !split_backend->iface.cpy_tensor_async(input_backend, split_backend, input, input_cpy)) {
ggml_backend_synchronize(input_backend);
if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
ggml_backend_event_synchronize(sched->events[split_backend_id][sched->cur_copy]);
} else {
ggml_backend_synchronize(split_backend);
}
ggml_backend_tensor_copy(input, input_cpy);
}
}
}
ggml/src/ggml-cuda.cu
+24 -21
View File
@@ -1501,7 +1501,7 @@ static void ggml_cuda_op_mul_mat(
}
// If src0 is on a temporary compute buffers (partial offloading) there may be some padding that needs to be cleared:
if (ne00 % MATRIX_ROW_PADDING != 0 && ggml_backend_buffer_get_usage(src0->buffer) == GGML_BACKEND_BUFFER_USAGE_COMPUTE && src0->view_src == nullptr) {
if (ne00 % MATRIX_ROW_PADDING != 0 && ggml_is_quantized(src0->type) && ggml_backend_buffer_get_usage(src0->buffer) == GGML_BACKEND_BUFFER_USAGE_COMPUTE && src0->view_src == nullptr) {
const int64_t nbytes_data = ggml_row_size(src0->type, (dev[id].row_high - dev[id].row_low)*ne00);
const int64_t nbytes_padding = ggml_row_size(src0->type, MATRIX_ROW_PADDING - ne00 % MATRIX_ROW_PADDING);
CUDA_CHECK(cudaMemsetAsync(dev[id].src0_dd + nbytes_data , 0, nbytes_padding, stream));
@@ -2362,33 +2362,35 @@ GGML_CALL static void ggml_backend_cuda_get_tensor_async(ggml_backend_t backend,
}
GGML_CALL static bool ggml_backend_cuda_cpy_tensor_async(ggml_backend_t backend_src, ggml_backend_t backend_dst, const ggml_tensor * src, ggml_tensor * dst) {
GGML_ASSERT(ggml_backend_is_cuda(backend_src) || ggml_backend_is_cuda(backend_dst));
ggml_backend_buffer_t buf_src = src->view_src ? src->view_src->buffer : src->buffer;
ggml_backend_buffer_t buf_dst = dst->view_src ? dst->view_src->buffer : dst->buffer;
if (!ggml_backend_buffer_is_cuda(src->buffer)) {
if (!ggml_backend_is_cuda(backend_src) || !ggml_backend_is_cuda(backend_dst)) {
return false;
}
if (!ggml_backend_buffer_is_cuda(dst->buffer)) {
if (!ggml_backend_buffer_is_cuda(src->buffer) || !ggml_backend_buffer_is_cuda(dst->buffer)) {
return false;
}
// device -> device
// device -> device copy
ggml_backend_cuda_context * cuda_ctx_src = (ggml_backend_cuda_context *)backend_src->context;
ggml_backend_cuda_context * cuda_ctx_dst = (ggml_backend_cuda_context *)backend_dst->context;
ggml_backend_cuda_buffer_context * buf_ctx_src = (ggml_backend_cuda_buffer_context *)buf_src->context;
ggml_backend_cuda_buffer_context * buf_ctx_dst = (ggml_backend_cuda_buffer_context *)buf_dst->context;
if (cuda_ctx_src->device != buf_ctx_src->device || cuda_ctx_dst->device != buf_ctx_dst->device) {
#ifndef NDEBUG
GGML_CUDA_LOG_WARN("%s: backend and buffer devices do not match\n", __func__);
#endif
return false;
}
if (backend_src != backend_dst) {
ggml_backend_cuda_buffer_context * buf_ctx_src = (ggml_backend_cuda_buffer_context *)buf_src->context;
ggml_backend_cuda_buffer_context * buf_ctx_dst = (ggml_backend_cuda_buffer_context *)buf_dst->context;
GGML_ASSERT(cuda_ctx_src->device == buf_ctx_src->device);
GGML_ASSERT(cuda_ctx_dst->device == buf_ctx_dst->device);
// copy on src stream
if (cuda_ctx_src->device == cuda_ctx_dst->device) {
CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(dst), cudaMemcpyDeviceToDevice, cuda_ctx_dst->stream()));
CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(dst), cudaMemcpyDeviceToDevice, cuda_ctx_src->stream()));
} else {
#ifdef GGML_CUDA_NO_PEER_COPY
return false;
@@ -2397,7 +2399,7 @@ GGML_CALL static bool ggml_backend_cuda_cpy_tensor_async(ggml_backend_t backend_
#endif
}
// record event on src stream
// record event on src stream after the copy
if (!cuda_ctx_src->copy_event) {
ggml_cuda_set_device(cuda_ctx_src->device);
CUDA_CHECK(cudaEventCreateWithFlags(&cuda_ctx_src->copy_event, cudaEventDisableTiming));
@@ -2409,7 +2411,7 @@ GGML_CALL static bool ggml_backend_cuda_cpy_tensor_async(ggml_backend_t backend_
CUDA_CHECK(cudaStreamWaitEvent(cuda_ctx_dst->stream(), cuda_ctx_src->copy_event, 0));
} else {
// src and dst are on the same backend
CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(dst), cudaMemcpyDeviceToDevice, cuda_ctx_dst->stream()));
CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(dst), cudaMemcpyDeviceToDevice, cuda_ctx_src->stream()));
}
return true;
}
@@ -2746,11 +2748,12 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
case GGML_OP_MUL_MAT_ID:
{
struct ggml_tensor * a = op->src[0];
if (op->op == GGML_OP_MUL_MAT) {
struct ggml_tensor * b = op->src[1];
if (a->ne[3] != b->ne[3]) {
return false;
}
struct ggml_tensor * b = op->src[1];
if (b->type == GGML_TYPE_F16 && a->type != GGML_TYPE_F16) {
return false;
}
if (op->op == GGML_OP_MUL_MAT && a->ne[3] != b->ne[3]) {
return false;
}
switch (a->type) {
case GGML_TYPE_F32:
@@ -2881,7 +2884,7 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
return true;
case GGML_OP_FLASH_ATTN_EXT:
#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
return op->src[0]->ne[0] == 64 || op->src[0]->ne[0] == 128;
return (op->src[0]->ne[0] == 64 && op->src[1]->type == GGML_TYPE_F16) || op->src[0]->ne[0] == 128;
#else
if (op->src[0]->ne[0] == 128) {
return true;
ggml/src/ggml-metal.m
+52 -17
View File
@@ -210,7 +210,7 @@ enum ggml_metal_kernel_type {
GGML_METAL_KERNEL_TYPE_COUNT
};
struct ggml_metal_context {
struct ggml_backend_metal_context {
int n_cb;
id<MTLDevice> device;
@@ -224,6 +224,10 @@ struct ggml_metal_context {
bool support_simdgroup_mm;
bool should_capture_next_compute;
// abort ggml_metal_graph_compute if callback returns true
ggml_abort_callback abort_callback;
void * abort_callback_data;
};
// MSL code
@@ -289,7 +293,7 @@ static void * ggml_metal_host_malloc(size_t n) {
return data;
}
static struct ggml_metal_context * ggml_metal_init(int n_cb) {
static struct ggml_backend_metal_context * ggml_metal_init(int n_cb) {
GGML_METAL_LOG_INFO("%s: allocating\n", __func__);
#if TARGET_OS_OSX && !GGML_METAL_NDEBUG
@@ -306,7 +310,7 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
GGML_METAL_LOG_INFO("%s: picking default device: %s\n", __func__, [[device name] UTF8String]);
// Configure context
struct ggml_metal_context * ctx = malloc(sizeof(struct ggml_metal_context));
struct ggml_backend_metal_context * ctx = malloc(sizeof(struct ggml_backend_metal_context));
ctx->device = device;
ctx->n_cb = MIN(n_cb, GGML_METAL_MAX_BUFFERS);
ctx->queue = [ctx->device newCommandQueue];
@@ -668,7 +672,7 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
return ctx;
}
static void ggml_metal_free(struct ggml_metal_context * ctx) {
static void ggml_metal_free(struct ggml_backend_metal_context * ctx) {
GGML_METAL_LOG_INFO("%s: deallocating\n", __func__);
for (int i = 0; i < GGML_METAL_KERNEL_TYPE_COUNT; ++i) {
@@ -734,7 +738,7 @@ static id<MTLBuffer> ggml_metal_get_buffer(struct ggml_tensor * t, size_t * offs
return nil;
}
static bool ggml_metal_supports_op(const struct ggml_metal_context * ctx, const struct ggml_tensor * op) {
static bool ggml_metal_supports_op(const struct ggml_backend_metal_context * ctx, const struct ggml_tensor * op) {
for (size_t i = 0, n = 3; i < n; ++i) {
if (op->src[i] != NULL && op->src[i]->type == GGML_TYPE_BF16) {
return false;
@@ -845,7 +849,7 @@ static bool ggml_metal_supports_op(const struct ggml_metal_context * ctx, const
}
static enum ggml_status ggml_metal_graph_compute(
struct ggml_metal_context * ctx,
struct ggml_backend_metal_context * ctx,
struct ggml_cgraph * gf) {
@autoreleasepool {
@@ -878,8 +882,11 @@ static enum ggml_status ggml_metal_graph_compute(
id<MTLCommandBuffer> command_buffer = [ctx->queue commandBufferWithUnretainedReferences];
command_buffer_builder[cb_idx] = command_buffer;
// enqueue the command buffers in order to specify their execution order
[command_buffer enqueue];
// always enqueue the first two command buffers
// enqueue all of the command buffers if we don't need to abort
if (cb_idx < 2 || ctx->abort_callback == NULL) {
[command_buffer enqueue];
}
}
const id<MTLCommandBuffer> *command_buffers = command_buffer_builder;
@@ -2827,7 +2834,9 @@ static enum ggml_status ggml_metal_graph_compute(
[encoder endEncoding];
[command_buffer commit];
if (cb_idx < 2 || ctx->abort_callback == NULL) {
[command_buffer commit];
}
});
// Wait for completion and check status of each command buffer
@@ -2847,6 +2856,23 @@ static enum ggml_status ggml_metal_graph_compute(
return GGML_STATUS_FAILED;
}
id<MTLCommandBuffer> next_buffer = (i + 1 < n_cb ? command_buffers[i + 1] : nil);
if (!next_buffer) {
continue;
}
bool next_queued = ([next_buffer status] != MTLCommandBufferStatusNotEnqueued);
if (next_queued) {
continue;
}
if (ctx->abort_callback && ctx->abort_callback(ctx->abort_callback_data)) {
GGML_METAL_LOG_INFO("%s: command buffer %d aborted", __func__, i);
return GGML_STATUS_ABORTED;
}
[next_buffer commit];
}
if (should_capture) {
@@ -3150,7 +3176,7 @@ GGML_CALL static const char * ggml_backend_metal_name(ggml_backend_t backend) {
}
GGML_CALL static void ggml_backend_metal_free(ggml_backend_t backend) {
struct ggml_metal_context * ctx = (struct ggml_metal_context *)backend->context;
struct ggml_backend_metal_context * ctx = (struct ggml_backend_metal_context *)backend->context;
ggml_metal_free(ctx);
free(backend);
}
@@ -3162,13 +3188,13 @@ GGML_CALL static ggml_backend_buffer_type_t ggml_backend_metal_get_default_buffe
}
GGML_CALL static enum ggml_status ggml_backend_metal_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
struct ggml_metal_context * metal_ctx = (struct ggml_metal_context *)backend->context;
struct ggml_backend_metal_context * metal_ctx = (struct ggml_backend_metal_context *)backend->context;
return ggml_metal_graph_compute(metal_ctx, cgraph);
}
GGML_CALL static bool ggml_backend_metal_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
struct ggml_metal_context * metal_ctx = (struct ggml_metal_context *)backend->context;
struct ggml_backend_metal_context * metal_ctx = (struct ggml_backend_metal_context *)backend->context;
return ggml_metal_supports_op(metal_ctx, op);
}
@@ -3213,9 +3239,9 @@ static ggml_guid_t ggml_backend_metal_guid(void) {
}
ggml_backend_t ggml_backend_metal_init(void) {
struct ggml_metal_context * ctx = ggml_metal_init(GGML_DEFAULT_N_THREADS);
struct ggml_backend_metal_context * ctx = ggml_metal_init(GGML_DEFAULT_N_THREADS);
if (ctx == NULL) {
GGML_METAL_LOG_ERROR("%s: error: failed to allocate context\n", __func__);
return NULL;
}
@@ -3237,15 +3263,24 @@ bool ggml_backend_is_metal(ggml_backend_t backend) {
void ggml_backend_metal_set_n_cb(ggml_backend_t backend, int n_cb) {
GGML_ASSERT(ggml_backend_is_metal(backend));
struct ggml_metal_context * ctx = (struct ggml_metal_context *)backend->context;
struct ggml_backend_metal_context * ctx = (struct ggml_backend_metal_context *)backend->context;
ctx->n_cb = MIN(n_cb, GGML_METAL_MAX_BUFFERS);
}
void ggml_backend_metal_set_abort_callback(ggml_backend_t backend, ggml_abort_callback abort_callback, void * user_data) {
GGML_ASSERT(ggml_backend_is_metal(backend));
struct ggml_backend_metal_context * ctx = (struct ggml_backend_metal_context *)backend->context;
ctx->abort_callback = abort_callback;
ctx->abort_callback_data = user_data;
}
bool ggml_backend_metal_supports_family(ggml_backend_t backend, int family) {
GGML_ASSERT(ggml_backend_is_metal(backend));
struct ggml_metal_context * ctx = (struct ggml_metal_context *)backend->context;
struct ggml_backend_metal_context * ctx = (struct ggml_backend_metal_context *)backend->context;
return [ctx->device supportsFamily:(MTLGPUFamilyApple1 + family - 1)];
}
@@ -3253,7 +3288,7 @@ bool ggml_backend_metal_supports_family(ggml_backend_t backend, int family) {
void ggml_backend_metal_capture_next_compute(ggml_backend_t backend) {
GGML_ASSERT(ggml_backend_is_metal(backend));
struct ggml_metal_context * ctx = (struct ggml_metal_context *)backend->context;
struct ggml_backend_metal_context * ctx = (struct ggml_backend_metal_context *)backend->context;
ctx->should_capture_next_compute = true;
}
ggml/src/ggml-rpc.cpp
+35 -1
View File
@@ -197,6 +197,10 @@ static std::shared_ptr<socket_t> create_server_socket(const char * host, int por
fprintf(stderr, "Failed to set SO_REUSEADDR\n");
return nullptr;
}
if (inet_addr(host) == INADDR_NONE) {
fprintf(stderr, "Invalid host address: %s\n", host);
return nullptr;
}
struct sockaddr_in serv_addr;
serv_addr.sin_family = AF_INET;
serv_addr.sin_addr.s_addr = inet_addr(host);
@@ -879,6 +883,14 @@ ggml_tensor * rpc_server::deserialize_tensor(struct ggml_context * ctx, const rp
if (result->buffer && buffers.find(result->buffer) == buffers.end()) {
return nullptr;
}
// require that the tensor data does not go beyond the buffer end
uint64_t tensor_size = (uint64_t) ggml_nbytes(result);
uint64_t buffer_start = (uint64_t) ggml_backend_buffer_get_base(result->buffer);
uint64_t buffer_size = (uint64_t) ggml_backend_buffer_get_size(result->buffer);
GGML_ASSERT(tensor->data + tensor_size >= tensor->data); // check for overflow
GGML_ASSERT(tensor->data >= buffer_start && tensor->data + tensor_size <= buffer_start + buffer_size);
result->op = (ggml_op) tensor->op;
for (uint32_t i = 0; i < GGML_MAX_OP_PARAMS / sizeof(int32_t); i++) {
result->op_params[i] = tensor->op_params[i];
@@ -898,7 +910,7 @@ bool rpc_server::set_tensor(const std::vector<uint8_t> & input) {
const rpc_tensor * in_tensor = (const rpc_tensor *)input.data();
uint64_t offset;
memcpy(&offset, input.data() + sizeof(rpc_tensor), sizeof(offset));
size_t size = input.size() - sizeof(rpc_tensor) - sizeof(offset);
const size_t size = input.size() - sizeof(rpc_tensor) - sizeof(offset);
struct ggml_init_params params {
/*.mem_size =*/ ggml_tensor_overhead(),
@@ -913,6 +925,17 @@ bool rpc_server::set_tensor(const std::vector<uint8_t> & input) {
return false;
}
GGML_PRINT_DEBUG("[%s] buffer: %p, data: %p, offset: %" PRIu64 ", size: %zu\n", __func__, (void*)tensor->buffer, tensor->data, offset, size);
// sanitize tensor->data
{
const size_t p0 = (size_t) ggml_backend_buffer_get_base(tensor->buffer);
const size_t p1 = p0 + ggml_backend_buffer_get_size(tensor->buffer);
if (in_tensor->data + offset < p0 || in_tensor->data + offset >= p1 || size > (p1 - in_tensor->data - offset)) {
GGML_ABORT("[%s] tensor->data out of bounds\n", __func__);
}
}
const void * data = input.data() + sizeof(rpc_tensor) + sizeof(offset);
ggml_backend_tensor_set(tensor, data, offset, size);
ggml_free(ctx);
@@ -943,6 +966,17 @@ bool rpc_server::get_tensor(const std::vector<uint8_t> & input, std::vector<uint
return false;
}
GGML_PRINT_DEBUG("[%s] buffer: %p, data: %p, offset: %" PRIu64 ", size: %" PRIu64 "\n", __func__, (void*)tensor->buffer, tensor->data, offset, size);
// sanitize tensor->data
{
const size_t p0 = (size_t) ggml_backend_buffer_get_base(tensor->buffer);
const size_t p1 = p0 + ggml_backend_buffer_get_size(tensor->buffer);
if (in_tensor->data + offset < p0 || in_tensor->data + offset >= p1 || size > (p1 - in_tensor->data - offset)) {
GGML_ABORT("[%s] tensor->data out of bounds\n", __func__);
}
}
// output serialization format: | data (size bytes) |
output.resize(size, 0);
ggml_backend_tensor_get(tensor, output.data(), offset, size);
ggml/src/ggml-sycl/dpct/helper.hpp
+16 -3
View File
@@ -874,7 +874,7 @@ namespace dpct
inline std::string get_preferred_gpu_platform_name() {
std::string result;
std::string filter = "level-zero";
std::string filter = "";
char* env = getenv("ONEAPI_DEVICE_SELECTOR");
if (env) {
if (std::strstr(env, "level_zero")) {
@@ -892,11 +892,24 @@ namespace dpct
else {
throw std::runtime_error("invalid device filter: " + std::string(env));
}
} else {
auto default_device = sycl::device(sycl::default_selector_v);
auto default_platform_name = default_device.get_platform().get_info<sycl::info::platform::name>();
if (std::strstr(default_platform_name.c_str(), "Level-Zero") || default_device.is_cpu()) {
filter = "level-zero";
}
else if (std::strstr(default_platform_name.c_str(), "CUDA")) {
filter = "cuda";
}
else if (std::strstr(default_platform_name.c_str(), "HIP")) {
filter = "hip";
}
}
auto plaform_list = sycl::platform::get_platforms();
auto platform_list = sycl::platform::get_platforms();
for (const auto& platform : plaform_list) {
for (const auto& platform : platform_list) {
auto devices = platform.get_devices();
auto gpu_dev = std::find_if(devices.begin(), devices.end(), [](const sycl::device& d) {
return d.is_gpu();
ggml/src/ggml-vulkan.cpp
+1 -1
View File
@@ -2108,9 +2108,9 @@ void ggml_vk_instance_init() {
}
static void ggml_vk_init(ggml_backend_vk_context * ctx, size_t idx) {
GGML_ASSERT(idx < vk_instance.device_indices.size());
VK_LOG_DEBUG("ggml_vk_init(" << ctx->name << ", " << idx << ")");
ggml_vk_instance_init();
GGML_ASSERT(idx < vk_instance.device_indices.size());
ctx->name = GGML_VK_NAME + std::to_string(idx);
ggml/src/ggml.c
+5 -1
View File
@@ -56,6 +56,9 @@ int ggml_sve_cnt_b = 0;
// disable POSIX deprecation warnings
// these functions are never going away, anyway
#pragma warning(disable: 4996)
// unreachable code because of multiple instances of code after GGML_ABORT
#pragma warning(disable: 4702)
#endif
#if defined(_WIN32)
@@ -3745,7 +3748,8 @@ static struct ggml_tensor * ggml_new_tensor_impl(
struct ggml_tensor * view_src,
size_t view_offs) {
assert(n_dims >= 1 && n_dims <= GGML_MAX_DIMS);
GGML_ASSERT(type >= 0 && type < GGML_TYPE_COUNT);
GGML_ASSERT(n_dims >= 1 && n_dims <= GGML_MAX_DIMS);
// find the base tensor and absolute offset
if (view_src != NULL && view_src->view_src != NULL) {
ggml/src/vulkan-shaders/CMakeLists.txt
+2
View File
@@ -1,5 +1,7 @@
find_package (Threads REQUIRED)
set(TARGET vulkan-shaders-gen)
add_executable(${TARGET} vulkan-shaders-gen.cpp)
install(TARGETS ${TARGET} RUNTIME)
target_compile_features(${TARGET} PRIVATE cxx_std_11)
target_link_libraries(vulkan-shaders-gen PUBLIC Threads::Threads)
ggml/src/vulkan-shaders/vulkan-shaders-gen.cpp
+1
View File
@@ -23,6 +23,7 @@
#ifdef _WIN32
#include <windows.h>
#include <direct.h> // For _mkdir on Windows
#include <algorithm> // For std::replace on w64devkit
#else
#include <unistd.h>
#include <sys/wait.h>