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

Merge e9e9d736bc into 2634ed207a

This commit is contained in:
Kevin Pouget 2026-02-01 09:20:22 -03:00 committed by GitHub
parent 2634ed207a e9e9d736bc
commit 5ce42066bd
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
15 changed files with 235 additions and 135 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
ggml/src/ggml-virtgpu/backend/backend-dispatched-buffer-type.cpp
View File

@ -36,18 +36,22 @@ uint32_t backend_buffer_type_get_max_size(apir_encoder * enc, apir_decoder * dec
ggml_backend_buffer_type_t buft;
buft = apir_decode_ggml_buffer_type(dec);
size_t value = buft->iface.get_max_size(buft);
size_t value = SIZE_MAX;
if (buft->iface.get_max_size) {
value = buft->iface.get_max_size(buft);
}
apir_encode_size_t(enc, &value);
return 0;
}
/* APIR_COMMAND_TYPE_BUFFER_TYPE_IS_HOST is deprecated. Keeping the handler for backward compatibility. */
uint32_t backend_buffer_type_is_host(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx) {
GGML_UNUSED(ctx);
ggml_backend_buffer_type_t buft;
buft = apir_decode_ggml_buffer_type(dec);
GGML_UNUSED(dec);
const bool is_host = false;
bool is_host = buft->iface.is_host(buft);
apir_encode_bool_t(enc, &is_host);
return 0;

5
ggml/src/ggml-virtgpu/backend/backend-dispatched.gen.h
View File

@ -16,6 +16,7 @@ uint32_t backend_device_buffer_from_ptr(apir_encoder * enc, apir_decoder * dec,
uint32_t backend_buffer_type_get_name(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
uint32_t backend_buffer_type_get_alignment(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
uint32_t backend_buffer_type_get_max_size(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
/* APIR_COMMAND_TYPE_BUFFER_TYPE_IS_HOST is deprecated. Keeping the handler for backward compatibility. */
uint32_t backend_buffer_type_is_host(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
uint32_t backend_buffer_type_alloc_buffer(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
uint32_t backend_buffer_type_get_alloc_size(apir_encoder * enc, apir_decoder * dec, virgl_apir_context * ctx);
@ -62,7 +63,7 @@ static inline const char * backend_dispatch_command_name(ApirBackendCommandType
case APIR_COMMAND_TYPE_BUFFER_TYPE_GET_MAX_SIZE:
return "backend_buffer_type_get_max_size";
case APIR_COMMAND_TYPE_BUFFER_TYPE_IS_HOST:
return "backend_buffer_type_is_host";
return "backend_buffer_type_is_host (DEPRECATED)";
case APIR_COMMAND_TYPE_BUFFER_TYPE_ALLOC_BUFFER:
return "backend_buffer_type_alloc_buffer";
case APIR_COMMAND_TYPE_BUFFER_TYPE_GET_ALLOC_SIZE:
@ -110,7 +111,7 @@ static const backend_dispatch_t apir_backend_dispatch_table[APIR_BACKEND_DISPATC
/* APIR_COMMAND_TYPE_BUFFER_TYPE_GET_NAME = */ backend_buffer_type_get_name,
/* APIR_COMMAND_TYPE_BUFFER_TYPE_GET_ALIGNMENT = */ backend_buffer_type_get_alignment,
/* APIR_COMMAND_TYPE_BUFFER_TYPE_GET_MAX_SIZE = */ backend_buffer_type_get_max_size,
/* APIR_COMMAND_TYPE_BUFFER_TYPE_IS_HOST = */ backend_buffer_type_is_host,
/* APIR_COMMAND_TYPE_BUFFER_TYPE_IS_HOST = */ backend_buffer_type_is_host /* DEPRECATED */,
/* APIR_COMMAND_TYPE_BUFFER_TYPE_ALLOC_BUFFER = */ backend_buffer_type_alloc_buffer,
/* APIR_COMMAND_TYPE_BUFFER_TYPE_GET_ALLOC_SIZE = */ backend_buffer_type_get_alloc_size,

10
ggml/src/ggml-virtgpu/backend/shared/apir_cs_ggml.h
View File

@ -39,11 +39,17 @@ static inline void apir_encode_ggml_tensor(apir_encoder * enc, const ggml_tensor
static inline const ggml_tensor * apir_decode_ggml_tensor(apir_decoder * dec) {
const apir_rpc_tensor * apir_rpc_tensor = apir_decode_apir_rpc_tensor_inplace(dec);
if (!apir_rpc_tensor) {
return NULL;
}
ggml_init_params params{
/*.mem_size =*/ ggml_tensor_overhead(),
/*.mem_buffer =*/ NULL,
/*.no_alloc =*/ true,
};
ggml_context * ctx = ggml_init(params);
const ggml_tensor * tensor = apir_deserialize_tensor(ctx, apir_rpc_tensor);
@ -71,6 +77,10 @@ static inline ggml_backend_buffer_type_t apir_decode_ggml_buffer_type(apir_decod
return (ggml_backend_buffer_type_t) handle;
}
static inline void apir_encode_apir_buffer_type_host_handle(apir_encoder * enc, apir_buffer_type_host_handle_t handle) {
apir_encoder_write(enc, sizeof(handle), &handle, sizeof(handle));
}
static inline apir_buffer_type_host_handle_t apir_decode_apir_buffer_type_host_handle(apir_decoder * dec) {
apir_buffer_type_host_handle_t handle;

27
ggml/src/ggml-virtgpu/ggml-backend-buffer-type.cpp
View File

@ -20,7 +20,7 @@ static ggml_backend_buffer_t ggml_backend_remoting_buffer_type_alloc_buffer(ggml
context->base = context->apir_context.shmem.mmap_ptr;
context->is_from_ptr = true;
} else {
context->apir_context = apir_buffer_type_alloc_buffer(gpu, buft, size);
context->apir_context = apir_buffer_type_alloc_buffer(gpu, gpu->cached_buffer_type.host_handle, size);
context->is_from_ptr = false;
context->base = NULL;
}
@ -34,36 +34,19 @@ static ggml_backend_buffer_t ggml_backend_remoting_buffer_type_alloc_buffer(ggml
static const char * ggml_backend_remoting_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
virtgpu * gpu = BUFT_TO_GPU(buft);
return apir_buffer_type_get_name(gpu, buft);
return gpu->cached_buffer_type.name;
}
static size_t ggml_backend_remoting_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
virtgpu * gpu = BUFT_TO_GPU(buft);
static size_t align = 0;
if (align == 0) {
align = apir_buffer_type_get_alignment(gpu, buft);
}
return align;
return gpu->cached_buffer_type.alignment;
}
static size_t ggml_backend_remoting_buffer_type_get_max_size(ggml_backend_buffer_type_t buft) {
virtgpu * gpu = BUFT_TO_GPU(buft);
static size_t max_size = 0;
if (max_size == 0) {
max_size = apir_buffer_type_get_max_size(gpu, buft);
}
return max_size;
}
static bool ggml_backend_remoting_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
virtgpu * gpu = BUFT_TO_GPU(buft);
return apir_buffer_type_is_host(gpu, buft);
return gpu->cached_buffer_type.max_size;
}
static size_t ggml_backend_remoting_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft,
@ -76,7 +59,7 @@ static size_t ggml_backend_remoting_buffer_type_get_alloc_size(ggml_backend_buff
return ggml_nbytes(tensor);
}
return apir_buffer_type_get_alloc_size(gpu, buft, tensor);
return apir_buffer_type_get_alloc_size(gpu, gpu->cached_buffer_type.host_handle, tensor);
}
const ggml_backend_buffer_type_i ggml_backend_remoting_buffer_type_interface = {

59
ggml/src/ggml-virtgpu/ggml-backend-device.cpp
View File

@ -3,32 +3,27 @@
static const char * ggml_backend_remoting_device_get_name(ggml_backend_dev_t dev) {
virtgpu * gpu = DEV_TO_GPU(dev);
return apir_device_get_name(gpu);
return gpu->cached_device_info.name;
}
static const char * ggml_backend_remoting_device_get_description(ggml_backend_dev_t dev) {
virtgpu * gpu = DEV_TO_GPU(dev);
return apir_device_get_description(gpu);
// Return the pre-cached description from the virtgpu structure
return gpu->cached_device_info.description;
}
static enum ggml_backend_dev_type ggml_backend_remoting_device_get_type(ggml_backend_dev_t dev) {
virtgpu * gpu = DEV_TO_GPU(dev);
static enum ggml_backend_dev_type type;
static bool has_type = false;
if (!has_type) {
has_type = true;
type = (enum ggml_backend_dev_type) apir_device_get_type(gpu);
}
return type;
return (enum ggml_backend_dev_type) gpu->cached_device_info.type;
}
static void ggml_backend_remoting_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
virtgpu * gpu = DEV_TO_GPU(dev);
return apir_device_get_memory(gpu, free, total);
*free = gpu->cached_device_info.memory_free;
*total = gpu->cached_device_info.memory_total;
}
static bool ggml_backend_remoting_device_supports_op(ggml_backend_dev_t dev, const ggml_tensor * op) {
@ -77,13 +72,22 @@ static void ggml_backend_remoting_device_get_props(ggml_backend_dev_t dev, ggml_
ggml_backend_buffer_type_t ggml_backend_remoting_device_get_buffer_type(ggml_backend_dev_t dev) {
virtgpu * gpu = DEV_TO_GPU(dev);
apir_buffer_type_host_handle_t ctx = apir_device_get_buffer_type(gpu);
static std::atomic<bool> initialized = false;
static ggml_backend_buffer_type buft;
static ggml_backend_buffer_type buft{
/* .iface = */ ggml_backend_remoting_buffer_type_interface,
/* .device = */ dev,
/* .context = */ (void *) ctx,
};
if (!initialized) {
static std::mutex mutex;
std::lock_guard<std::mutex> lock(mutex);
if (!initialized) {
buft = {
/* .iface = */ ggml_backend_remoting_buffer_type_interface,
/* .device = */ dev,
/* .context = */ (void *) gpu->cached_buffer_type.host_handle,
};
initialized = true;
}
}
return &buft;
}
@ -91,13 +95,22 @@ ggml_backend_buffer_type_t ggml_backend_remoting_device_get_buffer_type(ggml_bac
static ggml_backend_buffer_type_t ggml_backend_remoting_device_get_buffer_from_ptr_type(ggml_backend_dev_t dev) {
virtgpu * gpu = DEV_TO_GPU(dev);
apir_buffer_type_host_handle_t ctx = apir_device_get_buffer_type(gpu);
static std::atomic<bool> initialized = false;
static ggml_backend_buffer_type buft;
static ggml_backend_buffer_type buft{
/* .iface = */ ggml_backend_remoting_buffer_from_ptr_type_interface,
/* .device = */ dev,
/* .context = */ (void *) ctx,
};
if (!initialized) {
static std::mutex mutex;
std::lock_guard<std::mutex> lock(mutex);
if (!initialized) {
buft = {
/* .iface = */ ggml_backend_remoting_buffer_from_ptr_type_interface,
/* .device = */ dev,
/* .context = */ (void *) gpu->cached_buffer_type.host_handle,
};
initialized = true;
}
}
return &buft;
}

75
ggml/src/ggml-virtgpu/ggml-backend-reg.cpp
View File

@ -4,27 +4,60 @@
#include <iostream>
#include <mutex>
void ggml_virtgpu_cleanup(virtgpu * gpu);
static virtgpu * apir_initialize() {
static virtgpu * apir_gpu_instance = NULL;
static bool apir_initialized = false;
static virtgpu * gpu = NULL;
static std::atomic<bool> initialized = false;
if (initialized) {
// fast track
return gpu;
}
{
static std::mutex mutex;
std::lock_guard<std::mutex> lock(mutex);
if (apir_initialized) {
return apir_gpu_instance;
if (initialized) {
// thread safe
return gpu;
}
apir_gpu_instance = create_virtgpu();
if (!apir_gpu_instance) {
gpu = create_virtgpu();
if (!gpu) {
GGML_ABORT("failed to initialize the virtgpu");
}
apir_initialized = true;
// Pre-fetch and cache all device information, it will not change
gpu->cached_device_info.description = apir_device_get_description(gpu);
if (!gpu->cached_device_info.description) {
GGML_ABORT("failed to initialize the virtgpu device description");
}
gpu->cached_device_info.name = apir_device_get_name(gpu);
if (!gpu->cached_device_info.name) {
GGML_ABORT("failed to initialize the virtgpu device name");
}
gpu->cached_device_info.device_count = apir_device_get_count(gpu);
gpu->cached_device_info.type = apir_device_get_type(gpu);
apir_device_get_memory(gpu,
&gpu->cached_device_info.memory_free,
&gpu->cached_device_info.memory_total);
apir_buffer_type_host_handle_t buft_host_handle = apir_device_get_buffer_type(gpu);
gpu->cached_buffer_type.host_handle = buft_host_handle;
gpu->cached_buffer_type.name = apir_buffer_type_get_name(gpu, buft_host_handle);
if (!gpu->cached_buffer_type.name) {
GGML_ABORT("failed to initialize the virtgpu buffer type name");
}
gpu->cached_buffer_type.alignment = apir_buffer_type_get_alignment(gpu, buft_host_handle);
gpu->cached_buffer_type.max_size = apir_buffer_type_get_max_size(gpu, buft_host_handle);
initialized = true;
}
return apir_gpu_instance;
return gpu;
}
static int ggml_backend_remoting_get_device_count() {
@ -34,7 +67,7 @@ static int ggml_backend_remoting_get_device_count() {
return 0;
}
return apir_device_get_count(gpu);
return gpu->cached_device_info.device_count;
}
static size_t ggml_backend_remoting_reg_get_device_count(ggml_backend_reg_t reg) {
@ -62,7 +95,11 @@ static void ggml_backend_remoting_reg_init_devices(ggml_backend_reg_t reg) {
return;
}
static bool initialized = false;
static std::atomic<bool> initialized = false;
if (initialized) {
return; // fast track
}
{
static std::mutex mutex;
@ -134,4 +171,22 @@ ggml_backend_reg_t ggml_backend_virtgpu_reg() {
return &reg;
}
// public function, not exposed in the GGML interface at the moment
void ggml_virtgpu_cleanup(virtgpu * gpu) {
if (gpu->cached_device_info.name) {
free(gpu->cached_device_info.name);
gpu->cached_device_info.name = NULL;
}
if (gpu->cached_device_info.description) {
free(gpu->cached_device_info.description);
gpu->cached_device_info.description = NULL;
}
if (gpu->cached_buffer_type.name) {
free(gpu->cached_buffer_type.name);
gpu->cached_buffer_type.name = NULL;
}
mtx_destroy(&gpu->data_shmem_mutex);
}
GGML_BACKEND_DL_IMPL(ggml_backend_virtgpu_reg)

20
ggml/src/ggml-virtgpu/ggmlremoting_functions.yaml
View File

@ -24,10 +24,10 @@ functions:
frontend_return: "int"
get_name:
frontend_return: "const char *"
frontend_return: "char *"
get_description:
frontend_return: "const char *"
frontend_return: "char *"
get_type:
frontend_return: "uint32_t"
@ -64,35 +64,33 @@ functions:
group_description: "buffer-type"
functions:
get_name:
frontend_return: "const char *"
frontend_return: "char *"
frontend_extra_params:
- "ggml_backend_buffer_type_t buft"
- "apir_buffer_type_host_handle_t host_handle"
get_alignment:
frontend_return: "size_t"
frontend_extra_params:
- "ggml_backend_buffer_type_t buft"
- "apir_buffer_type_host_handle_t host_handle"
get_max_size:
frontend_return: "size_t"
frontend_extra_params:
- "ggml_backend_buffer_type_t buft"
- "apir_buffer_type_host_handle_t host_handle"
is_host:
frontend_return: "bool"
frontend_extra_params:
- "ggml_backend_buffer_type_t buft"
deprecated: true
alloc_buffer:
frontend_return: "apir_buffer_context_t"
frontend_extra_params:
- "ggml_backend_buffer_type_t buffer_buft"
- "apir_buffer_type_host_handle_t host_handle"
- "size_t size"
get_alloc_size:
frontend_return: "size_t"
frontend_extra_params:
- "ggml_backend_buffer_type_t buft"
- "apir_buffer_type_host_handle_t host_handle"
- "const ggml_tensor *op"
buffer:

18
ggml/src/ggml-virtgpu/regenerate_remoting.py
View File

@ -116,7 +116,7 @@ class RemotingCodebaseGenerator:
'frontend_return': func_metadata.get('frontend_return', 'void'),
'frontend_extra_params': func_metadata.get('frontend_extra_params', []),
'group_description': group_description,
'newly_added': func_metadata.get('newly_added', False)
'deprecated': func_metadata.get('deprecated', False),
})
enum_value += 1
@ -165,6 +165,9 @@ class RemotingCodebaseGenerator:
signature = "uint32_t"
params = "apir_encoder *enc, apir_decoder *dec, virgl_apir_context *ctx"
if func['deprecated']:
decl_lines.append(f"/* {func['enum_name']} is deprecated. Keeping the handler for backward compatibility. */")
decl_lines.append(f"{signature} {func['backend_function']}({params});")
# Switch cases
@ -176,7 +179,9 @@ class RemotingCodebaseGenerator:
switch_lines.append(f" /* {func['group_description']} */")
current_group = func['group_name']
switch_lines.append(f" case {func['enum_name']}: return \"{func['backend_function']}\";")
deprecated = " (DEPRECATED)" if func['deprecated'] else ""
switch_lines.append(f" case {func['enum_name']}: return \"{func['backend_function']}{deprecated}\";")
# Dispatch table
table_lines = []
@ -188,7 +193,8 @@ class RemotingCodebaseGenerator:
table_lines.append("")
current_group = func['group_name']
table_lines.append(f" /* {func['enum_name']} = */ {func['backend_function']},")
deprecated = " /* DEPRECATED */" if func['deprecated'] else ""
table_lines.append(f" /* {func['enum_name']} = */ {func['backend_function']}{deprecated},")
header_content = f'''\
#pragma once
@ -225,6 +231,10 @@ static const backend_dispatch_t apir_backend_dispatch_table[APIR_BACKEND_DISPATC
decl_lines.append(f"/* {func['group_description']} */")
current_group = func['group_name']
if func['deprecated']:
decl_lines.append(f"/* {func['frontend_function']} is deprecated. */")
continue
# Build parameter list
params = [self.naming_patterns['frontend_base_param']]
params.extend(func['frontend_extra_params'])
@ -287,7 +297,7 @@ static const backend_dispatch_t apir_backend_dispatch_table[APIR_BACKEND_DISPATC
generated_files = [apir_backend_path, backend_dispatched_path, virtgpu_forward_path]
if not self.clang_format_available:
logging.warning("\n⚠️clang-format not found in PATH. Generated files will not be formatted."
logging.warning("\n⚠️clang-format not found in PATH. Generated files will not be formatted.\n"
" Install clang-format to enable automatic code formatting.")
else:
logging.info("\n🎨 Formatting files with clang-format...")

12
ggml/src/ggml-virtgpu/virtgpu-forward-backend.cpp
View File

@ -18,9 +18,14 @@ ggml_status apir_backend_graph_compute(virtgpu * gpu, ggml_cgraph * cgraph) {
virtgpu_shmem temp_shmem; // Local storage for large buffers
virtgpu_shmem * shmem = &temp_shmem;
bool using_shared_shmem = false;
if (cgraph_size <= gpu->data_shmem.mmap_size) {
// prefer the init-time allocated page, if large enough
// Lock mutex before using shared data_shmem buffer
if (mtx_lock(&gpu->data_shmem_mutex) != thrd_success) {
GGML_ABORT("Failed to lock data_shmem mutex");
}
using_shared_shmem = true;
shmem = &gpu->data_shmem;
} else if (virtgpu_shmem_create(gpu, cgraph_size, shmem)) {
GGML_ABORT("Couldn't allocate the guest-host shared buffer");
@ -42,7 +47,10 @@ ggml_status apir_backend_graph_compute(virtgpu * gpu, ggml_cgraph * cgraph) {
remote_call_finish(gpu, encoder, decoder);
if (shmem != &gpu->data_shmem) {
// Unlock mutex before cleanup
if (using_shared_shmem) {
mtx_unlock(&gpu->data_shmem_mutex);
} else {
virtgpu_shmem_destroy(gpu, shmem);
}

39
ggml/src/ggml-virtgpu/virtgpu-forward-buffer-type.cpp
View File

@ -1,13 +1,13 @@
#include "virtgpu-forward-impl.h"
const char * apir_buffer_type_get_name(virtgpu * gpu, ggml_backend_buffer_type_t buft) {
char * apir_buffer_type_get_name(virtgpu * gpu, apir_buffer_type_host_handle_t host_handle) {
apir_encoder * encoder;
apir_decoder * decoder;
ApirForwardReturnCode ret;
REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_BUFFER_TYPE_GET_NAME);
apir_encode_ggml_buffer_type(encoder, buft);
apir_encode_apir_buffer_type_host_handle(encoder, host_handle);
REMOTE_CALL(gpu, encoder, decoder, ret);
@ -24,14 +24,14 @@ const char * apir_buffer_type_get_name(virtgpu * gpu, ggml_backend_buffer_type_t
return string;
}
size_t apir_buffer_type_get_alignment(virtgpu * gpu, ggml_backend_buffer_type_t buft) {
size_t apir_buffer_type_get_alignment(virtgpu * gpu, apir_buffer_type_host_handle_t host_handle) {
apir_encoder * encoder;
apir_decoder * decoder;
ApirForwardReturnCode ret;
REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_BUFFER_TYPE_GET_ALIGNMENT);
apir_encode_ggml_buffer_type(encoder, buft);
apir_encode_apir_buffer_type_host_handle(encoder, host_handle);
REMOTE_CALL(gpu, encoder, decoder, ret);
@ -43,14 +43,14 @@ size_t apir_buffer_type_get_alignment(virtgpu * gpu, ggml_backend_buffer_type_t
return alignment;
}
size_t apir_buffer_type_get_max_size(virtgpu * gpu, ggml_backend_buffer_type_t buft) {
size_t apir_buffer_type_get_max_size(virtgpu * gpu, apir_buffer_type_host_handle_t host_handle) {
apir_encoder * encoder;
apir_decoder * decoder;
ApirForwardReturnCode ret;
REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_BUFFER_TYPE_GET_MAX_SIZE);
apir_encode_ggml_buffer_type(encoder, buft);
apir_encode_apir_buffer_type_host_handle(encoder, host_handle);
REMOTE_CALL(gpu, encoder, decoder, ret);
@ -62,26 +62,7 @@ size_t apir_buffer_type_get_max_size(virtgpu * gpu, ggml_backend_buffer_type_t b
return max_size;
}
bool apir_buffer_type_is_host(virtgpu * gpu, ggml_backend_buffer_type_t buft) {
apir_encoder * encoder;
apir_decoder * decoder;
ApirForwardReturnCode ret;
REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_BUFFER_TYPE_IS_HOST);
apir_encode_ggml_buffer_type(encoder, buft);
REMOTE_CALL(gpu, encoder, decoder, ret);
bool is_host;
apir_decode_bool_t(decoder, &is_host);
remote_call_finish(gpu, encoder, decoder);
return is_host;
}
apir_buffer_context_t apir_buffer_type_alloc_buffer(virtgpu * gpu, ggml_backend_buffer_type_t buft, size_t size) {
apir_buffer_context_t apir_buffer_type_alloc_buffer(virtgpu * gpu, apir_buffer_type_host_handle_t host_handle, size_t size) {
apir_encoder * encoder;
apir_decoder * decoder;
ApirForwardReturnCode ret;
@ -90,7 +71,7 @@ apir_buffer_context_t apir_buffer_type_alloc_buffer(virtgpu * gpu, ggml_backend_
REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_BUFFER_TYPE_ALLOC_BUFFER);
apir_encode_ggml_buffer_type(encoder, buft);
apir_encode_apir_buffer_type_host_handle(encoder, host_handle);
apir_encode_size_t(encoder, &size);
@ -103,14 +84,14 @@ apir_buffer_context_t apir_buffer_type_alloc_buffer(virtgpu * gpu, ggml_backend_
return buffer_context;
}
size_t apir_buffer_type_get_alloc_size(virtgpu * gpu, ggml_backend_buffer_type_t buft, const ggml_tensor * op) {
size_t apir_buffer_type_get_alloc_size(virtgpu * gpu, apir_buffer_type_host_handle_t host_handle, const ggml_tensor * op) {
apir_encoder * encoder;
apir_decoder * decoder;
ApirForwardReturnCode ret;
REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_BUFFER_TYPE_GET_ALLOC_SIZE);
apir_encode_ggml_buffer_type(encoder, buft);
apir_encode_apir_buffer_type_host_handle(encoder, host_handle);
apir_encode_ggml_tensor_inline(encoder, op);

24
ggml/src/ggml-virtgpu/virtgpu-forward-buffer.cpp
View File

@ -36,9 +36,14 @@ void apir_buffer_set_tensor(virtgpu * gpu,
virtgpu_shmem temp_shmem; // Local storage for large buffers
virtgpu_shmem * shmem = &temp_shmem;
bool using_shared_shmem = false;
if (size <= gpu->data_shmem.mmap_size) {
// prefer the init-time allocated page, if large enough
// Lock mutex before using shared data_shmem buffer
if (mtx_lock(&gpu->data_shmem_mutex) != thrd_success) {
GGML_ABORT("Failed to lock data_shmem mutex");
}
using_shared_shmem = true;
shmem = &gpu->data_shmem;
} else if (virtgpu_shmem_create(gpu, size, shmem)) {
@ -55,7 +60,10 @@ void apir_buffer_set_tensor(virtgpu * gpu,
remote_call_finish(gpu, encoder, decoder);
if (shmem != &gpu->data_shmem) {
// Unlock mutex before cleanup
if (using_shared_shmem) {
mtx_unlock(&gpu->data_shmem_mutex);
} else {
virtgpu_shmem_destroy(gpu, shmem);
}
@ -79,9 +87,14 @@ void apir_buffer_get_tensor(virtgpu * gpu,
virtgpu_shmem temp_shmem; // Local storage for large buffers
virtgpu_shmem * shmem = &temp_shmem;
bool using_shared_shmem = false;
if (size <= gpu->data_shmem.mmap_size) {
// prefer the init-time allocated page, if large enough
// Lock mutex before using shared data_shmem buffer
if (mtx_lock(&gpu->data_shmem_mutex) != thrd_success) {
GGML_ABORT("Failed to lock data_shmem mutex");
}
using_shared_shmem = true;
shmem = &gpu->data_shmem;
} else if (virtgpu_shmem_create(gpu, size, shmem)) {
@ -98,7 +111,10 @@ void apir_buffer_get_tensor(virtgpu * gpu,
remote_call_finish(gpu, encoder, decoder);
if (shmem != &gpu->data_shmem) {
// Unlock mutex before cleanup
if (using_shared_shmem) {
mtx_unlock(&gpu->data_shmem_mutex);
} else {
virtgpu_shmem_destroy(gpu, shmem);
}
}

16
ggml/src/ggml-virtgpu/virtgpu-forward-device.cpp
View File

@ -2,11 +2,6 @@
#include "virtgpu-shm.h"
int apir_device_get_count(virtgpu * gpu) {
static int32_t dev_count = -1;
if (dev_count != -1) {
return dev_count;
}
apir_encoder * encoder;
apir_decoder * decoder;
ApirForwardReturnCode ret;
@ -14,6 +9,7 @@ int apir_device_get_count(virtgpu * gpu) {
REMOTE_CALL_PREPARE(gpu, encoder, APIR_COMMAND_TYPE_DEVICE_GET_COUNT);
REMOTE_CALL(gpu, encoder, decoder, ret);
int32_t dev_count = -1;
apir_decode_int32_t(decoder, &dev_count);
remote_call_finish(gpu, encoder, decoder);
@ -21,11 +17,7 @@ int apir_device_get_count(virtgpu * gpu) {
return dev_count;
}
const char * apir_device_get_name(virtgpu * gpu) {
static char * string = nullptr;
if (string) {
return string;
}
char * apir_device_get_name(virtgpu * gpu) {
apir_encoder * encoder;
apir_decoder * decoder;
ApirForwardReturnCode ret;
@ -34,7 +26,7 @@ const char * apir_device_get_name(virtgpu * gpu) {
REMOTE_CALL(gpu, encoder, decoder, ret);
const size_t string_size = apir_decode_array_size_unchecked(decoder);
string = (char *) apir_decoder_alloc_array(sizeof(char), string_size);
char * string = (char *) apir_decoder_alloc_array(sizeof(char), string_size);
if (!string) {
GGML_LOG_ERROR("%s: Could not allocate the device name buffer\n", __func__);
return NULL;
@ -46,7 +38,7 @@ const char * apir_device_get_name(virtgpu * gpu) {
return string;
}
const char * apir_device_get_description(virtgpu * gpu) {
char * apir_device_get_description(virtgpu * gpu) {
apir_encoder * encoder;
apir_decoder * decoder;
ApirForwardReturnCode ret;

21
ggml/src/ggml-virtgpu/virtgpu-forward.gen.h
View File

@ -3,8 +3,8 @@
/* device */
void apir_device_get_device_count(struct virtgpu * gpu);
int apir_device_get_count(struct virtgpu * gpu);
const char * apir_device_get_name(struct virtgpu * gpu);
const char * apir_device_get_description(struct virtgpu * gpu);
char * apir_device_get_name(struct virtgpu * gpu);
char * apir_device_get_description(struct virtgpu * gpu);
uint32_t apir_device_get_type(struct virtgpu * gpu);
void apir_device_get_memory(struct virtgpu * gpu, size_t * free, size_t * total);
bool apir_device_supports_op(struct virtgpu * gpu, const ggml_tensor * op);
@ -17,14 +17,15 @@ void apir_device_get_props(struct virtgpu * gpu,
apir_buffer_context_t apir_device_buffer_from_ptr(struct virtgpu * gpu, size_t size, size_t max_tensor_size);
/* buffer-type */
const char * apir_buffer_type_get_name(struct virtgpu * gpu, ggml_backend_buffer_type_t buft);
size_t apir_buffer_type_get_alignment(struct virtgpu * gpu, ggml_backend_buffer_type_t buft);
size_t apir_buffer_type_get_max_size(struct virtgpu * gpu, ggml_backend_buffer_type_t buft);
bool apir_buffer_type_is_host(struct virtgpu * gpu, ggml_backend_buffer_type_t buft);
apir_buffer_context_t apir_buffer_type_alloc_buffer(struct virtgpu * gpu,
ggml_backend_buffer_type_t buffer_buft,
size_t size);
size_t apir_buffer_type_get_alloc_size(struct virtgpu * gpu, ggml_backend_buffer_type_t buft, const ggml_tensor * op);
char * apir_buffer_type_get_name(struct virtgpu * gpu, apir_buffer_type_host_handle_t host_handle);
size_t apir_buffer_type_get_alignment(struct virtgpu * gpu, apir_buffer_type_host_handle_t host_handle);
size_t apir_buffer_type_get_max_size(struct virtgpu * gpu, apir_buffer_type_host_handle_t host_handle);
apir_buffer_context_t apir_buffer_type_alloc_buffer(struct virtgpu * gpu,
apir_buffer_type_host_handle_t host_handle,
size_t size);
size_t apir_buffer_type_get_alloc_size(struct virtgpu * gpu,
apir_buffer_type_host_handle_t host_handle,
const ggml_tensor * op);
/* buffer */
void * apir_buffer_get_base(struct virtgpu * gpu, apir_buffer_context_t * buffer_context);

11
ggml/src/ggml-virtgpu/virtgpu.cpp
View File

@ -149,6 +149,13 @@ virtgpu * create_virtgpu() {
gpu->use_apir_capset = getenv("GGML_REMOTING_USE_APIR_CAPSET") != nullptr;
util_sparse_array_init(&gpu->shmem_array, sizeof(virtgpu_shmem), 1024);
// Initialize mutex to protect shared data_shmem buffer
if (mtx_init(&gpu->data_shmem_mutex, mtx_plain) != thrd_success) {
delete gpu;
GGML_ABORT("%s: failed to initialize data_shmem mutex", __func__);
return NULL;
}
if (virtgpu_open(gpu) != APIR_SUCCESS) {
GGML_ABORT("%s: failed to open the virtgpu device", __func__);
return NULL;
@ -333,9 +340,9 @@ apir_encoder * remote_call_prepare(virtgpu * gpu, ApirCommandType apir_cmd_type,
* Prepare the command encoder and its buffer
*/
static char encoder_buffer[4096];
thread_local char encoder_buffer[4096];
static apir_encoder enc;
thread_local apir_encoder enc;
enc = {
.cur = encoder_buffer,
.start = encoder_buffer,

21
ggml/src/ggml-virtgpu/virtgpu.h
View File

@ -73,6 +73,27 @@ struct virtgpu {
/* APIR communication pages */
virtgpu_shmem reply_shmem;
virtgpu_shmem data_shmem;
/* Mutex to protect shared data_shmem buffer from concurrent access */
mtx_t data_shmem_mutex;
/* Cached device information to prevent memory leaks and race conditions */
struct {
char * description;
char * name;
int32_t device_count;
uint32_t type;
size_t memory_free;
size_t memory_total;
} cached_device_info;
/* Cached buffer type information to prevent memory leaks and race conditions */
struct {
apir_buffer_type_host_handle_t host_handle;
char * name;
size_t alignment;
size_t max_size;
} cached_buffer_type;
};
static inline int virtgpu_ioctl(virtgpu * gpu, unsigned long request, void * args) {