vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Cann backend now uses GGML logging

Browse source
This commit is contained in:
Mason M 2024-10-01 13:51:20 -03:00
parent 15817d0f83
commit c6b6540cb3
2 changed files with 13 additions and 87 deletions
Download patch file Download diff file Expand all files Collapse all files

11
ggml/include/ggml-cann.h
View file

@ -116,17 +116,6 @@ GGML_API GGML_CALL void ggml_backend_cann_get_device_memory(int32_t device,
size_t* free, size_t* free,
size_t* total); size_t* total);
/**
* @brief Set the logging callback for GGML.
*
* This function sets the logging callback and user data for logging.
*
* @param log_callback The logging callback to set.
* @param user_data User data to pass to the logging callback.
*/
GGML_API void ggml_backend_cann_log_set_callback(ggml_log_callback log_callback,
void* user_data);
#ifdef __cplusplus #ifdef __cplusplus
} }
#endif #endif

89
ggml/src/ggml-cann.cpp
View file

@ -39,69 +39,6 @@
#include "ggml-common.h" #include "ggml-common.h"
/**
* @brief Default logging callback for GGML.
*
* This function is the default logging callback that logs messages to stderr.
*
* @param level The log level.
* @param msg The log message.
* @param user_data User data passed to the callback.
*/
static void ggml_cann_default_log_callback(enum ggml_log_level level,
const char* msg, void* user_data) {
GGML_UNUSED(level);
GGML_UNUSED(user_data);
fprintf(stderr, "%s", msg);
}
ggml_log_callback ggml_cann_log_callback = ggml_cann_default_log_callback;
void* ggml_cann_log_user_data = NULL;
GGML_API void ggml_backend_cann_log_set_callback(ggml_log_callback log_callback,
void* user_data) {
ggml_cann_log_callback = log_callback;
ggml_cann_log_user_data = user_data;
}
#define GGML_CANN_LOG_INFO(...) ggml_cann_log(GGML_LOG_LEVEL_INFO, __VA_ARGS__)
#define GGML_CANN_LOG_WARN(...) ggml_cann_log(GGML_LOG_LEVEL_WARN, __VA_ARGS__)
#define GGML_CANN_LOG_ERROR(...) \
ggml_cann_log(GGML_LOG_LEVEL_ERROR, __VA_ARGS__)
GGML_ATTRIBUTE_FORMAT(2, 3)
/**
* @brief Log a message using the current logging callback.
*
* This function formats a log message and passes it to the current logging
* callback.
*
* @param level The log level.
* @param format The format string for the log message.
* @param ... The arguments for the format string.
*/
static void ggml_cann_log(enum ggml_log_level level, const char* format, ...) {
if (ggml_cann_log_callback != NULL) {
va_list args;
va_start(args, format);
char buffer[128];
int len = vsnprintf(buffer, 128, format, args);
if (len < 128) {
ggml_cann_log_callback(level, buffer, ggml_cann_log_user_data);
} else {
// vsnprintf adds a null terminator
std::vector<char> buffer2(len + 1);
va_end(args);
va_start(args, format);
vsnprintf(&buffer2[0], buffer2.size(), format, args);
ggml_cann_log_callback(level, buffer2.data(),
ggml_cann_log_user_data);
}
va_end(args);
}
}
/** /**
* @brief Handles CANN errors by printing an error message and aborting. * @brief Handles CANN errors by printing an error message and aborting.
* *
@ -116,10 +53,10 @@ static void ggml_cann_log(enum ggml_log_level level, const char* format, ...) {
int32_t id = -1; int32_t id = -1;
aclrtGetDevice(&id); aclrtGetDevice(&id);
GGML_CANN_LOG_ERROR("CANN error: %s\n", msg); GGML_LOG_ERROR("CANN error: %s\n", msg);
GGML_CANN_LOG_ERROR(" current device: %d, in function %s at %s:%d\n", id, func, GGML_LOG_ERROR(" current device: %d, in function %s at %s:%d\n", id, func,
file, line); file, line);
GGML_CANN_LOG_ERROR(" %s\n", stmt); GGML_LOG_ERROR(" %s\n", stmt);
// abort with GGML_ASSERT to get a stack trace // abort with GGML_ASSERT to get a stack trace
GGML_ABORT("CANN error"); GGML_ABORT("CANN error");
} }
@ -165,7 +102,7 @@ static ggml_cann_device_info ggml_cann_init() {
aclError err = aclrtGetDeviceCount((uint32_t*)&info.device_count); aclError err = aclrtGetDeviceCount((uint32_t*)&info.device_count);
if (err != ACL_SUCCESS) { if (err != ACL_SUCCESS) {
GGML_CANN_LOG_ERROR("%s: failed to initialize CANN: %s\n", GGML_LOG_ERROR("%s: failed to initialize CANN: %s\n",
__func__, aclGetRecentErrMsg()); __func__, aclGetRecentErrMsg());
return info; return info;
} }
@ -315,7 +252,7 @@ struct ggml_cann_pool_leg : public ggml_cann_pool {
*actual_size = look_ahead_size; *actual_size = look_ahead_size;
pool_size += look_ahead_size; pool_size += look_ahead_size;
#ifdef DEBUG_CANN_MALLOC #ifdef DEBUG_CANN_MALLOC
GGML_CANN_LOG_INFO( GGML_LOG_INFO(
"%s[%d]: %d buffers, max_size = %u MB, pool_size = %u MB, " "%s[%d]: %d buffers, max_size = %u MB, pool_size = %u MB, "
"requested %u MB\n", "requested %u MB\n",
__func__, device, nnz, (uint32_t)(max_size / 1024 / 1024), __func__, device, nnz, (uint32_t)(max_size / 1024 / 1024),
@ -470,7 +407,7 @@ struct ggml_cann_pool_vmm : public ggml_cann_pool {
// add to the pool // add to the pool
pool_size += reserve_size; pool_size += reserve_size;
// GGML_CANN_LOG_INFO("cann pool[%d]: size increased to %llu MB ( // GGML_LOG_INFO("cann pool[%d]: size increased to %llu MB (
// reserved %llu MB)\n", // reserved %llu MB)\n",
// device, (unsigned long long) (pool_size/1024/1024), // device, (unsigned long long) (pool_size/1024/1024),
// (unsigned long long) (reserve_size/1024/1024)); // (unsigned long long) (reserve_size/1024/1024));
@ -483,7 +420,7 @@ struct ggml_cann_pool_vmm : public ggml_cann_pool {
pool_used += size; pool_used += size;
#ifdef DEBUG_CANN_MALLOC #ifdef DEBUG_CANN_MALLOC
GGML_CANN_LOG_INFO("cann pool[%d]: allocated %llu bytes at %llx\n", device, GGML_LOG_INFO("cann pool[%d]: allocated %llu bytes at %llx\n", device,
(unsigned long long)size, (unsigned long long)ptr); (unsigned long long)size, (unsigned long long)ptr);
#endif #endif
return ptr; return ptr;
@ -497,7 +434,7 @@ struct ggml_cann_pool_vmm : public ggml_cann_pool {
*/ */
void free(void* ptr, size_t size) override { void free(void* ptr, size_t size) override {
#ifdef DEBUG_CANN_MALLOC #ifdef DEBUG_CANN_MALLOC
GGML_CANN_LOG_INFO("cann pool[%d]: freed %llu bytes at %llx\n", device, GGML_LOG_INFO("cann pool[%d]: freed %llu bytes at %llx\n", device,
(unsigned long long)size, (unsigned long long)ptr); (unsigned long long)size, (unsigned long long)ptr);
#endif #endif
@ -1095,7 +1032,7 @@ ggml_backend_cann_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft,
void* dev_ptr; void* dev_ptr;
aclError err = aclrtMalloc(&dev_ptr, size, ACL_MEM_MALLOC_HUGE_FIRST); aclError err = aclrtMalloc(&dev_ptr, size, ACL_MEM_MALLOC_HUGE_FIRST);
if (err != ACL_SUCCESS) { if (err != ACL_SUCCESS) {
GGML_CANN_LOG_ERROR( GGML_LOG_ERROR(
"%s: allocating %.2f MiB on device %d: aclrtMalloc failed: %s\n", "%s: allocating %.2f MiB on device %d: aclrtMalloc failed: %s\n",
__func__, size / 1024.0 / 1024.0, buft_ctx->device, __func__, size / 1024.0 / 1024.0, buft_ctx->device,
aclGetRecentErrMsg()); aclGetRecentErrMsg());
@ -1280,7 +1217,7 @@ static void * ggml_cann_host_malloc(size_t size) {
aclError err = aclrtMallocHost((void **) &hostPtr, size); aclError err = aclrtMallocHost((void **) &hostPtr, size);
if (err != ACL_SUCCESS) { if (err != ACL_SUCCESS) {
GGML_CANN_LOG_WARN("%s: failed to allocate %.2f MiB of pinned memory: %s\n", __func__, GGML_LOG_WARN("%s: failed to allocate %.2f MiB of pinned memory: %s\n", __func__,
size / 1024.0 / 1024.0, aclGetRecentErrMsg()); size / 1024.0 / 1024.0, aclGetRecentErrMsg());
return nullptr; return nullptr;
} }
@ -1732,7 +1669,7 @@ GGML_CALL static enum ggml_status ggml_backend_cann_graph_compute(
bool ok = ggml_cann_compute_forward(*cann_ctx, node); bool ok = ggml_cann_compute_forward(*cann_ctx, node);
if (!ok) { if (!ok) {
GGML_CANN_LOG_ERROR("%s: error: op not supported %s (%s)\n", __func__, GGML_LOG_ERROR("%s: error: op not supported %s (%s)\n", __func__,
node->name, ggml_op_name(node->op)); node->name, ggml_op_name(node->op));
} }
GGML_ASSERT(ok); GGML_ASSERT(ok);
@ -2045,13 +1982,13 @@ static ggml_guid_t ggml_backend_cann_guid() {
GGML_CALL ggml_backend_t ggml_backend_cann_init(int32_t device) { GGML_CALL ggml_backend_t ggml_backend_cann_init(int32_t device) {
aclInit(nullptr); aclInit(nullptr);
if (device < 0 || device >= ggml_backend_cann_get_device_count()) { if (device < 0 || device >= ggml_backend_cann_get_device_count()) {
GGML_CANN_LOG_ERROR("%s: error: invalid device %d\n", __func__, device); GGML_LOG_ERROR("%s: error: invalid device %d\n", __func__, device);
return nullptr; return nullptr;
} }
ggml_backend_cann_context* ctx = new ggml_backend_cann_context(device); ggml_backend_cann_context* ctx = new ggml_backend_cann_context(device);
if (ctx == nullptr) { if (ctx == nullptr) {
GGML_CANN_LOG_ERROR("%s: error: failed to allocate context\n", __func__); GGML_LOG_ERROR("%s: error: failed to allocate context\n", __func__);
return nullptr; return nullptr;
} }
ggml_cann_set_device(ctx->device); ggml_cann_set_device(ctx->device);