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

metal : reusing llama.cpp logging

Browse source
This commit is contained in:
Rickard Hallerbäck 2023-09-12 23:28:04 +02:00
parent 89e89599fd
commit 84471ffc73
4 changed files with 80 additions and 71 deletions
Download patch file Download diff file Expand all files Collapse all files

6
Makefile
View file

@ -427,7 +427,7 @@ endif
endif # LLAMA_METAL
ifdef LLAMA_METAL
ggml-metal.o: ggml-metal.m ggml-metal.h
ggml-metal.o: ggml-metal.m ggml-metal.h llama.h
$(CC) $(CFLAGS) -c $< -o $@
endif # LLAMA_METAL
@ -551,7 +551,7 @@ speculative: examples/speculative/speculative.cpp build-info.h ggml.o llama.o co
$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
ifdef LLAMA_METAL
metal: examples/metal/metal.cpp ggml.o $(OBJS)
metal: examples/metal/metal.cpp ggml.o llama.o $(OBJS)
$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
endif
@ -573,7 +573,7 @@ benchmark-matmult: examples/benchmark/benchmark-matmult.cpp build-info.h ggml.o
$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
./$@
vdot: pocs/vdot/vdot.cpp ggml.o $(OBJS)
vdot: pocs/vdot/vdot.cpp ggml.o llama.o $(OBJS)
$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
tests/test-llama-grammar: tests/test-llama-grammar.cpp build-info.h ggml.o common.o grammar-parser.o $(OBJS)

85
ggml-metal.m
View file

@ -11,11 +11,12 @@
#define MIN(a, b) ((a) < (b) ? (a) : (b))
#define MAX(a, b) ((a) > (b) ? (a) : (b))
// TODO: temporary - reuse llama.cpp logging
#ifdef GGML_METAL_NDEBUG
#define metal_printf(...)
#define LLAMA_LOG_INFO(...)
#define LLAMA_LOG_WARN(...)
#define LLAMA_LOG_ERROR(...)
#else
#define metal_printf(...) fprintf(stderr, __VA_ARGS__)
#import "llama.h"
#endif
#define UNUSED(x) (void)(x)
@ -118,7 +119,7 @@ static NSString * const msl_library_source = @"see metal.metal";
@end
struct ggml_metal_context * ggml_metal_init(int n_cb) {
metal_printf("%s: allocating\n", __func__);
LLAMA_LOG_INFO("%s: allocating\n", __func__);
id <MTLDevice> device;
NSString * s;
@ -128,14 +129,14 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
NSArray * devices = MTLCopyAllDevices();
for (device in devices) {
s = [device name];
metal_printf("%s: found device: %s\n", __func__, [s UTF8String]);
LLAMA_LOG_INFO("%s: found device: %s\n", __func__, [s UTF8String]);
}
#endif
// Pick and show default Metal device
device = MTLCreateSystemDefaultDevice();
s = [device name];
metal_printf("%s: picking default device: %s\n", __func__, [s UTF8String]);
LLAMA_LOG_INFO("%s: picking default device: %s\n", __func__, [s UTF8String]);
// Configure context
struct ggml_metal_context * ctx = malloc(sizeof(struct ggml_metal_context));
@ -162,7 +163,7 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
ctx->library = [ctx->device newLibraryWithURL:libURL error:&error];
if (error) {
metal_printf("%s: error: %s\n", __func__, [[error description] UTF8String]);
LLAMA_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
return NULL;
}
}
@ -176,11 +177,11 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
//NSString * path = [[NSBundle mainBundle] pathForResource:@"../../examples/metal/metal" ofType:@"metal"];
NSBundle * bundle = [NSBundle bundleForClass:[GGMLMetalClass class]];
NSString * path = [bundle pathForResource:@"ggml-metal" ofType:@"metal"];
metal_printf("%s: loading '%s'\n", __func__, [path UTF8String]);
LLAMA_LOG_INFO("%s: loading '%s'\n", __func__, [path UTF8String]);
NSString * src = [NSString stringWithContentsOfFile:path encoding:NSUTF8StringEncoding error:&error];
if (error) {
metal_printf("%s: error: %s\n", __func__, [[error description] UTF8String]);
LLAMA_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
return NULL;
}
@ -192,7 +193,7 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
ctx->library = [ctx->device newLibraryWithSource:src options:nil error:&error];
#endif
if (error) {
metal_printf("%s: error: %s\n", __func__, [[error description] UTF8String]);
LLAMA_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
return NULL;
}
}
@ -204,11 +205,11 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
#define GGML_METAL_ADD_KERNEL(name) \
ctx->function_##name = [ctx->library newFunctionWithName:@"kernel_"#name]; \
ctx->pipeline_##name = [ctx->device newComputePipelineStateWithFunction:ctx->function_##name error:&error]; \
metal_printf("%s: loaded %-32s %16p | th_max = %4d | th_width = %4d\n", __func__, "kernel_"#name, (void *) ctx->pipeline_##name, \
LLAMA_LOG_INFO("%s: loaded %-32s %16p | th_max = %4d | th_width = %4d\n", __func__, "kernel_"#name, (void *) ctx->pipeline_##name, \
(int) ctx->pipeline_##name.maxTotalThreadsPerThreadgroup, \
(int) ctx->pipeline_##name.threadExecutionWidth); \
if (error) { \
metal_printf("%s: load pipeline error: %s\n", __func__, [[error description] UTF8String]); \
LLAMA_LOG_ERROR("%s: error: load pipeline error: %s\n", __func__, [[error description] UTF8String]); \
return NULL; \
}
@ -264,13 +265,13 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
#undef GGML_METAL_ADD_KERNEL
}
metal_printf("%s: hasUnifiedMemory = %s\n", __func__, ctx->device.hasUnifiedMemory ? "true" : "false");
LLAMA_LOG_INFO("%s: hasUnifiedMemory = %s\n", __func__, ctx->device.hasUnifiedMemory ? "true" : "false");
#if TARGET_OS_OSX
metal_printf("%s: recommendedMaxWorkingSetSize = %8.2f MB\n", __func__, ctx->device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
LLAMA_LOG_INFO("%s: recommendedMaxWorkingSetSize = %8.2f MB\n", __func__, ctx->device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
if (ctx->device.maxTransferRate != 0) {
metal_printf("%s: maxTransferRate = %8.2f MB/s\n", __func__, ctx->device.maxTransferRate / 1024.0 / 1024.0);
LLAMA_LOG_INFO("%s: maxTransferRate = %8.2f MB/s\n", __func__, ctx->device.maxTransferRate / 1024.0 / 1024.0);
} else {
metal_printf("%s: maxTransferRate = built-in GPU\n", __func__);
LLAMA_LOG_INFO("%s: maxTransferRate = built-in GPU\n", __func__);
}
#endif
@ -278,7 +279,7 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
}
void ggml_metal_free(struct ggml_metal_context * ctx) {
metal_printf("%s: deallocating\n", __func__);
LLAMA_LOG_INFO("%s: deallocating\n", __func__);
#define GGML_METAL_DEL_KERNEL(name) \
[ctx->function_##name release]; \
[ctx->pipeline_##name release];
@ -350,7 +351,7 @@ void * ggml_metal_host_malloc(size_t n) {
void * data = NULL;
const int result = posix_memalign((void **) &data, sysconf(_SC_PAGESIZE), n);
if (result != 0) {
metal_printf("%s: error: posix_memalign failed\n", __func__);
LLAMA_LOG_ERROR("%s: error: posix_memalign failed\n", __func__);
return NULL;
}
@ -378,7 +379,7 @@ int * ggml_metal_get_concur_list(struct ggml_metal_context * ctx) {
// Metal buffer based on the host memory pointer
//
static id<MTLBuffer> ggml_metal_get_buffer(struct ggml_metal_context * ctx, struct ggml_tensor * t, size_t * offs) {
//metal_printf("%s: data tensor '%16s', offs_data = %8ld, offs_eval = %8ld, offs_cach = %8ld\n", __func__, t->name, offs_data, offs_eval, offs_cach);
//LLAMA_LOG_INFO("%s: data tensor '%16s', offs_data = %8ld, offs_eval = %8ld, offs_cach = %8ld\n", __func__, t->name, offs_data, offs_eval, offs_cach);
const int64_t tsize = ggml_nbytes(t);
@ -389,13 +390,13 @@ static id<MTLBuffer> ggml_metal_get_buffer(struct ggml_metal_context * ctx, stru
if (ioffs >= 0 && ioffs + tsize <= (int64_t) ctx->buffers[i].size) {
*offs = (size_t) ioffs;
//metal_printf("%s: '%s' tensor '%16s', offs = %8ld\n", __func__, ctx->buffers[i].name, t->name, *offs);
//LLAMA_LOG_INFO("%s: '%s' tensor '%16s', offs = %8ld\n", __func__, ctx->buffers[i].name, t->name, *offs);
return ctx->buffers[i].metal;
}
}
metal_printf("%s: error: buffer is nil\n", __func__);
LLAMA_LOG_ERROR("%s: error: buffer is nil\n", __func__);
return nil;
}
@ -407,7 +408,7 @@ bool ggml_metal_add_buffer(
size_t size,
size_t max_size) {
if (ctx->n_buffers >= GGML_METAL_MAX_BUFFERS) {
metal_printf("%s: too many buffers\n", __func__);
LLAMA_LOG_ERROR("%s: error: too many buffers\n", __func__);
return false;
}
@ -417,7 +418,7 @@ bool ggml_metal_add_buffer(
const int64_t ioffs = (int64_t) data - (int64_t) ctx->buffers[i].data;
if (ioffs >= 0 && ioffs < (int64_t) ctx->buffers[i].size) {
metal_printf("%s: error: buffer '%s' overlaps with '%s'\n", __func__, name, ctx->buffers[i].name);
LLAMA_LOG_ERROR("%s: error: buffer '%s' overlaps with '%s'\n", __func__, name, ctx->buffers[i].name);
return false;
}
}
@ -438,11 +439,11 @@ bool ggml_metal_add_buffer(
ctx->buffers[ctx->n_buffers].metal = [ctx->device newBufferWithBytesNoCopy:data length:size_aligned options:MTLResourceStorageModeShared deallocator:nil];
if (ctx->buffers[ctx->n_buffers].metal == nil) {
metal_printf("%s: failed to allocate '%-16s' buffer, size = %8.2f MB\n", __func__, name, size_aligned / 1024.0 / 1024.0);
LLAMA_LOG_ERROR("%s: error: failed to allocate '%-16s' buffer, size = %8.2f MB\n", __func__, name, size_aligned / 1024.0 / 1024.0);
return false;
}
metal_printf("%s: allocated '%-16s' buffer, size = %8.2f MB", __func__, name, size_aligned / 1024.0 / 1024.0);
LLAMA_LOG_INFO("%s: allocated '%-16s' buffer, size = %8.2f MB", __func__, name, size_aligned / 1024.0 / 1024.0);
++ctx->n_buffers;
} else {
@ -462,13 +463,13 @@ bool ggml_metal_add_buffer(
ctx->buffers[ctx->n_buffers].metal = [ctx->device newBufferWithBytesNoCopy:(void *) ((uint8_t *) data + i) length:size_step_aligned options:MTLResourceStorageModeShared deallocator:nil];
if (ctx->buffers[ctx->n_buffers].metal == nil) {
metal_printf("%s: failed to allocate '%-16s' buffer, size = %8.2f MB\n", __func__, name, size_step_aligned / 1024.0 / 1024.0);
LLAMA_LOG_ERROR("%s: error: failed to allocate '%-16s' buffer, size = %8.2f MB\n", __func__, name, size_step_aligned / 1024.0 / 1024.0);
return false;
}
metal_printf("%s: allocated '%-16s' buffer, size = %8.2f MB, offs = %12ld", __func__, name, size_step_aligned / 1024.0 / 1024.0, i);
LLAMA_LOG_INFO("%s: allocated '%-16s' buffer, size = %8.2f MB, offs = %12ld", __func__, name, size_step_aligned / 1024.0 / 1024.0, i);
if (i + size_step < size) {
metal_printf("\n");
LLAMA_LOG_INFO("\n");
}
++ctx->n_buffers;
@ -476,17 +477,17 @@ bool ggml_metal_add_buffer(
}
#if TARGET_OS_OSX
metal_printf(", (%8.2f / %8.2f)",
LLAMA_LOG_INFO(", (%8.2f / %8.2f)",
ctx->device.currentAllocatedSize / 1024.0 / 1024.0,
ctx->device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
if (ctx->device.currentAllocatedSize > ctx->device.recommendedMaxWorkingSetSize) {
metal_printf(", warning: current allocated size is greater than the recommended max working set size\n");
LLAMA_LOG_WARN("%s: warning: current allocated size is greater than the recommended max working set size\n", __func__);
} else {
metal_printf("\n");
LLAMA_LOG_INFO("\n");
}
#else
metal_printf(", (%8.2f)\n", ctx->device.currentAllocatedSize / 1024.0 / 1024.0);
LLAMA_LOG_INFO(", (%8.2f)\n", ctx->device.currentAllocatedSize / 1024.0 / 1024.0);
#endif
}
@ -599,7 +600,7 @@ void ggml_metal_graph_find_concurrency(
}
if (ctx->concur_list_len > GGML_MAX_CONCUR) {
metal_printf("%s: too many elements for metal ctx->concur_list!\n", __func__);
LLAMA_LOG_WARN("%s: too many elements for metal ctx->concur_list!\n", __func__);
}
}
@ -653,7 +654,7 @@ void ggml_metal_graph_compute(
continue;
}
//metal_printf("%s: encoding node %3d, op = %8s\n", __func__, i, ggml_op_name(gf->nodes[i]->op));
//LLAMA_LOG_INFO("%s: encoding node %3d, op = %8s\n", __func__, i, ggml_op_name(gf->nodes[i]->op));
struct ggml_tensor * src0 = gf->nodes[i]->src[0];
struct ggml_tensor * src1 = gf->nodes[i]->src[1];
@ -697,17 +698,17 @@ void ggml_metal_graph_compute(
id<MTLBuffer> id_src1 = src1 ? ggml_metal_get_buffer(ctx, src1, &offs_src1) : nil;
id<MTLBuffer> id_dst = dst ? ggml_metal_get_buffer(ctx, dst, &offs_dst) : nil;
//metal_printf("%s: op - %s\n", __func__, ggml_op_name(dst->op));
//LLAMA_LOG_INFO("%s: op - %s\n", __func__, ggml_op_name(dst->op));
//if (src0) {
// metal_printf("%s: src0 - %4s [%5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(src0t), ne00, ne01, ne02,
// LLAMA_LOG_INFO("%s: src0 - %4s [%5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(src0t), ne00, ne01, ne02,
// ggml_is_contiguous(src0), src0->name);
//}
//if (src1) {
// metal_printf("%s: src1 - %4s [%5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(src1t), ne10, ne11, ne12,
// LLAMA_LOG_INFO("%s: src1 - %4s [%5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(src1t), ne10, ne11, ne12,
// ggml_is_contiguous(src1), src1->name);
//}
//if (dst) {
// metal_printf("%s: dst - %4s [%5lld, %5lld, %5lld], 1, %s\n", __func__, ggml_type_name(dstt), ne0, ne1, ne2,
// LLAMA_LOG_INFO("%s: dst - %4s [%5lld, %5lld, %5lld], 1, %s\n", __func__, ggml_type_name(dstt), ne0, ne1, ne2,
// dst->name);
//}
@ -813,7 +814,7 @@ void ggml_metal_graph_compute(
} break;
default:
{
metal_printf("%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
LLAMA_LOG_WARN("%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
GGML_ASSERT(false);
}
} break;
@ -993,7 +994,7 @@ void ggml_metal_graph_compute(
} break;
default:
{
metal_printf("Asserting on type %d\n",(int)src0t);
LLAMA_LOG_ERROR("Asserting on type %d\n",(int)src0t);
GGML_ASSERT(false && "not implemented");
}
};
@ -1234,7 +1235,7 @@ void ggml_metal_graph_compute(
} break;
default:
{
metal_printf("%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
LLAMA_LOG_ERROR("%s: error: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
GGML_ASSERT(false);
}
}
@ -1259,7 +1260,7 @@ void ggml_metal_graph_compute(
MTLCommandBufferStatus status = (MTLCommandBufferStatus) [ctx->command_buffers[i] status];
if (status != MTLCommandBufferStatusCompleted) {
metal_printf("%s: command buffer %d failed with status %lu\n", __func__, i, status);
LLAMA_LOG_INFO("%s: command buffer %d failed with status %lu\n", __func__, i, status);
GGML_ASSERT(false);
}
}

30
llama.cpp
View file

@ -76,28 +76,6 @@
#pragma warning(disable: 4244 4267) // possible loss of data
#endif
#ifdef __GNUC__
#ifdef __MINGW32__
#define LLAMA_ATTRIBUTE_FORMAT(...) __attribute__((format(gnu_printf, __VA_ARGS__)))
#else
#define LLAMA_ATTRIBUTE_FORMAT(...) __attribute__((format(printf, __VA_ARGS__)))
#endif
#else
#define LLAMA_ATTRIBUTE_FORMAT(...)
#endif
//
// logging
//
LLAMA_ATTRIBUTE_FORMAT(2, 3)
static void llama_log_internal (llama_log_level level, const char* format, ...);
static void llama_log_callback_default(llama_log_level level, const char * text, void * user_data);
#define LLAMA_LOG_INFO(...) llama_log_internal(LLAMA_LOG_LEVEL_INFO , __VA_ARGS__)
#define LLAMA_LOG_WARN(...) llama_log_internal(LLAMA_LOG_LEVEL_WARN , __VA_ARGS__)
#define LLAMA_LOG_ERROR(...) llama_log_internal(LLAMA_LOG_LEVEL_ERROR, __VA_ARGS__)
//
// helpers
//
@ -6366,7 +6344,7 @@ void llama_log_set(llama_log_callback log_callback, void * user_data) {
g_state.log_callback_user_data = user_data;
}
static void llama_log_internal_v(llama_log_level level, const char * format, va_list args) {
void llama_log_v(llama_log_level level, const char * format, va_list args) {
va_list args_copy;
va_copy(args_copy, args);
char buffer[128];
@ -6383,14 +6361,14 @@ static void llama_log_internal_v(llama_log_level level, const char * format, va_
va_end(args_copy);
}
static void llama_log_internal(llama_log_level level, const char * format, ...) {
void llama_log(llama_log_level level, const char * format, ...) {
va_list args;
va_start(args, format);
llama_log_internal_v(level, format, args);
llama_log_v(level, format, args);
va_end(args);
}
static void llama_log_callback_default(llama_log_level level, const char * text, void * user_data) {
void llama_log_callback_default(llama_log_level level, const char * text, void * user_data) {
(void) level;
(void) user_data;
fputs(text, stderr);

30
llama.h
View file

@ -532,6 +532,36 @@ extern "C" {
}
#endif
#ifdef __GNUC__
#ifdef __MINGW32__
#define LLAMA_ATTRIBUTE_FORMAT(...) __attribute__((format(gnu_printf, __VA_ARGS__)))
#else
#define LLAMA_ATTRIBUTE_FORMAT(...) __attribute__((format(printf, __VA_ARGS__)))
#endif
#else
#define LLAMA_ATTRIBUTE_FORMAT(...)
#endif
//
// logging
//
#ifdef __cplusplus
extern "C" {
#endif
LLAMA_ATTRIBUTE_FORMAT(2, 3)
void llama_log (enum llama_log_level level, const char* format, ...);
void llama_log_callback_default(enum llama_log_level level, const char * text, void * user_data);
#ifdef __cplusplus
}
#endif
#define LLAMA_LOG_INFO(...) llama_log(LLAMA_LOG_LEVEL_INFO , __VA_ARGS__)
#define LLAMA_LOG_WARN(...) llama_log(LLAMA_LOG_LEVEL_WARN , __VA_ARGS__)
#define LLAMA_LOG_ERROR(...) llama_log(LLAMA_LOG_LEVEL_ERROR, __VA_ARGS__)
// Internal API to be implemented by llama.cpp and used by tests/benchmarks only
#ifdef LLAMA_API_INTERNAL