electron-rare/prima-cpp
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

count model flops for f32xf32, f16xf32, q4kxf32, q6kxf32

Browse Source
This commit is contained in:
Lizonghang
2024-11-24 13:13:32 +04:00
parent a5ba34169a
commit 3fe00a16a0
4 changed files with 188 additions and 119 deletions
Download Patch File Download Diff File Expand all files Collapse all files
common/profiler.cpp
+47 -35
View File
@@ -465,15 +465,15 @@ void device_print_props(struct device_info * dev_info_set, int n, struct llama_m
}
LOG_INF("\n");
LOG_INF("| CPU flops (F32 x F32, GFLOPS)");
LOG_INF("| CPU flops (F32xF32, GFLOPS) ");
for (int i = 0; i < n; ++i) {
LOG_INF("| %-10.1f ", dev_info_set[i].cpu_props.flops_f32);
LOG_INF("| %-10.1f ", dev_info_set[i].cpu_props.flops_f32_f32);
}
LOG_INF("\n");
LOG_INF("| CPU flops (F16 x F16, GFLOPS)");
LOG_INF("| CPU flops (F16xF32, GFLOPS) ");
for (int i = 0; i < n; ++i) {
LOG_INF("| %-10.1f ", dev_info_set[i].cpu_props.flops_f16);
LOG_INF("| %-10.1f ", dev_info_set[i].cpu_props.flops_f16_f32);
}
LOG_INF("\n");
@@ -593,13 +593,13 @@ void device_print_props(struct device_info * dev_info_set, int n, struct llama_m
LOG_INF("| Metal flops (F32xF32, GFLOPS)");
for (int i = 0; i < n; ++i) {
LOG_INF("| %-10.1f ", dev_info_set[i].gpu_props.metal_flops_f32);
LOG_INF("| %-10.1f ", dev_info_set[i].gpu_props.metal_flops_f32_f32);
}
LOG_INF("\n");
LOG_INF("| Metal flops (F16xF16, GFLOPS)");
LOG_INF("| Metal flops (F16xF32, GFLOPS)");
for (int i = 0; i < n; ++i) {
LOG_INF("| %-10.1f ", dev_info_set[i].gpu_props.metal_flops_f16);
LOG_INF("| %-10.1f ", dev_info_set[i].gpu_props.metal_flops_f16_f32);
}
LOG_INF("\n");
@@ -617,13 +617,13 @@ void device_print_props(struct device_info * dev_info_set, int n, struct llama_m
LOG_INF("| CUDA flops (F32xF32, GFLOPS)");
for (int i = 0; i < n; ++i) {
LOG_INF("| %-10.1f ", dev_info_set[i].gpu_props.cuda_flops_f32);
LOG_INF("| %-10.1f ", dev_info_set[i].gpu_props.cuda_flops_f32_f32);
}
LOG_INF("\n");
LOG_INF("| CUDA flops (F16xF16, GFLOPS)");
LOG_INF("| CUDA flops (F16xF32, GFLOPS)");
for (int i = 0; i < n; ++i) {
LOG_INF("| %-10.1f ", dev_info_set[i].gpu_props.cuda_flops_f16);
LOG_INF("| %-10.1f ", dev_info_set[i].gpu_props.cuda_flops_f16_f32);
}
LOG_INF("\n");
@@ -639,33 +639,45 @@ void device_print_props(struct device_info * dev_info_set, int n, struct llama_m
}
LOG_INF("\n");
LOG_INF("| Model flops (input) ");
LOG_INF("| %-10lu ", dev_info_set[0].model_flops.input_flops);
LOG_INF("| Model flops (output F32xF32) ");
LOG_INF("| %-10lu ", dev_info_set[0].model_flops.output_f32_f32);
LOG_INF("\n");
LOG_INF("| Model flops (each layer) ");
LOG_INF("| %-10lu ", dev_info_set[0].model_flops.layer_flops);
LOG_INF("| Model flops (output Q6KxF32) ");
LOG_INF("| %-10lu ", dev_info_set[0].model_flops.output_q6k_f32);
LOG_INF("\n");
LOG_INF("| Model flops (output) ");
LOG_INF("| %-10lu ", dev_info_set[0].model_flops.output_flops);
LOG_INF("| Model flops (layer F32xF32) ");
LOG_INF("| %-10lu ", dev_info_set[0].model_flops.layer_f32_f32);
LOG_INF("\n");
LOG_INF("| Model flops (layer F16xF32) ");
LOG_INF("| %-10lu ", dev_info_set[0].model_flops.layer_f16_f32);
LOG_INF("\n");
LOG_INF("| Model flops (layer Q4KxF32) ");
LOG_INF("| %-10lu ", dev_info_set[0].model_flops.layer_q4k_f32);
LOG_INF("\n");
LOG_INF("| Model flops (layer Q6KxF32) ");
LOG_INF("| %-10lu ", dev_info_set[0].model_flops.layer_q6k_f32);
LOG_INF("\n");
LOG_INF("| Model params (input) ");
LOG_INF("| %-10lu ", dev_info_set[0].model_flops.input_params);
LOG_INF("| %-10lu ", dev_info_set[0].model_params.input_params);
LOG_INF("\n");
LOG_INF("| Model params (each layer) ");
LOG_INF("| %-10lu ", dev_info_set[0].model_flops.layer_params);
LOG_INF("| %-10lu ", dev_info_set[0].model_params.layer_params);
LOG_INF("\n");
LOG_INF("| Model params (output) ");
LOG_INF("| %-10lu ", dev_info_set[0].model_flops.output_params);
LOG_INF("| %-10lu ", dev_info_set[0].model_params.output_params);
LOG_INF("\n");
model_flops ffo = dev_info_set[0].model_flops;
int64_t total_flops = ffo.input_flops + ffo.output_flops + (ffo.layer_flops * llama_model_n_layers(model));
double cpu_flops_f16 = dev_info_set[0].cpu_props.flops_f16 * 1e9;
int64_t total_flops = ffo.output_f32_f32 + (ffo.layer_f32_f32 * llama_model_n_layers(model)); // todo
double cpu_flops_f16 = dev_info_set[0].cpu_props.flops_f16_f32 * 1e9;
LOG_INF("| Token latency (ms) ");
LOG_INF("| %-10.2f ", total_flops / cpu_flops_f16 * 1000);
@@ -739,10 +751,10 @@ size_t serialize(const struct device_info * dev_info, char ** buffer) {
memcpy(ptr, &dev_info->cpu_props.cores, sizeof(uint32_t));
ptr += sizeof(uint32_t);
memcpy(ptr, &dev_info->cpu_props.flops_f32, sizeof(float));
memcpy(ptr, &dev_info->cpu_props.flops_f32_f32, sizeof(float));
ptr += sizeof(float);
memcpy(ptr, &dev_info->cpu_props.flops_f16, sizeof(float));
memcpy(ptr, &dev_info->cpu_props.flops_f16_f32, sizeof(float));
ptr += sizeof(float);
memcpy(ptr, &dev_info->cpu_props.flops_q4k_f32, sizeof(float));
@@ -763,10 +775,10 @@ size_t serialize(const struct device_info * dev_info, char ** buffer) {
memcpy(ptr, &dev_info->gpu_props.memory_total, sizeof(float));
ptr += sizeof(float);
memcpy(ptr, &dev_info->gpu_props.metal_flops_f32, sizeof(float));
memcpy(ptr, &dev_info->gpu_props.metal_flops_f32_f32, sizeof(float));
ptr += sizeof(float);
memcpy(ptr, &dev_info->gpu_props.metal_flops_f16, sizeof(float));
memcpy(ptr, &dev_info->gpu_props.metal_flops_f16_f32, sizeof(float));
ptr += sizeof(float);
memcpy(ptr, &dev_info->gpu_props.metal_flops_q4k_f32, sizeof(float));
@@ -775,10 +787,10 @@ size_t serialize(const struct device_info * dev_info, char ** buffer) {
memcpy(ptr, &dev_info->gpu_props.metal_flops_q6k_f32, sizeof(float));
ptr += sizeof(float);
memcpy(ptr, &dev_info->gpu_props.cuda_flops_f32, sizeof(float));
memcpy(ptr, &dev_info->gpu_props.cuda_flops_f32_f32, sizeof(float));
ptr += sizeof(float);
memcpy(ptr, &dev_info->gpu_props.cuda_flops_f16, sizeof(float));
memcpy(ptr, &dev_info->gpu_props.cuda_flops_f16_f32, sizeof(float));
ptr += sizeof(float);
memcpy(ptr, &dev_info->gpu_props.cuda_flops_q4k_f32, sizeof(float));
@@ -786,7 +798,7 @@ size_t serialize(const struct device_info * dev_info, char ** buffer) {
memcpy(ptr, &dev_info->gpu_props.cuda_flops_q6k_f32, sizeof(float));
// no need to synchronize model flops
// no need to synchronize model flops and model params
return total_size;
}
@@ -844,10 +856,10 @@ void deserialize(const char * buffer, struct device_info * dev_info) {
memcpy(&dev_info->cpu_props.cores, ptr, sizeof(uint32_t));
ptr += sizeof(uint32_t);
memcpy(&dev_info->cpu_props.flops_f32, ptr, sizeof(float));
memcpy(&dev_info->cpu_props.flops_f32_f32, ptr, sizeof(float));
ptr += sizeof(float);
memcpy(&dev_info->cpu_props.flops_f16, ptr, sizeof(float));
memcpy(&dev_info->cpu_props.flops_f16_f32, ptr, sizeof(float));
ptr += sizeof(float);
memcpy(&dev_info->cpu_props.flops_q4k_f32, ptr, sizeof(float));
@@ -868,10 +880,10 @@ void deserialize(const char * buffer, struct device_info * dev_info) {
memcpy(&dev_info->gpu_props.memory_total, ptr, sizeof(float));
ptr += sizeof(float);
memcpy(&dev_info->gpu_props.metal_flops_f32, ptr, sizeof(float));
memcpy(&dev_info->gpu_props.metal_flops_f32_f32, ptr, sizeof(float));
ptr += sizeof(float);
memcpy(&dev_info->gpu_props.metal_flops_f16, ptr, sizeof(float));
memcpy(&dev_info->gpu_props.metal_flops_f16_f32, ptr, sizeof(float));
ptr += sizeof(float);
memcpy(&dev_info->gpu_props.metal_flops_q4k_f32, ptr, sizeof(float));
@@ -880,10 +892,10 @@ void deserialize(const char * buffer, struct device_info * dev_info) {
memcpy(&dev_info->gpu_props.metal_flops_q6k_f32, ptr, sizeof(float));
ptr += sizeof(float);
memcpy(&dev_info->gpu_props.cuda_flops_f32, ptr, sizeof(float));
memcpy(&dev_info->gpu_props.cuda_flops_f32_f32, ptr, sizeof(float));
ptr += sizeof(float);
memcpy(&dev_info->gpu_props.cuda_flops_f16, ptr, sizeof(float));
memcpy(&dev_info->gpu_props.cuda_flops_f16_f32, ptr, sizeof(float));
ptr += sizeof(float);
memcpy(&dev_info->gpu_props.cuda_flops_q4k_f32, ptr, sizeof(float));
@@ -891,5 +903,5 @@ void deserialize(const char * buffer, struct device_info * dev_info) {
memcpy(&dev_info->gpu_props.cuda_flops_q6k_f32, ptr, sizeof(float));
// no need to synchronize model flops
// no need to synchronize model flops and model params
}
common/profiler.h
+49 -33
View File
@@ -8,8 +8,8 @@ struct cpu_props {
const char * name;
const char * description;
uint32_t cores;
float flops_f32; // in GFLOPS
float flops_f16; // in GFLOPS
float flops_f32_f32; // in GFLOPS
float flops_f16_f32; // in GFLOPS
float flops_q4k_f32; // in GFLOPS
float flops_q6k_f32; // in GFLOPS
@@ -17,8 +17,8 @@ struct cpu_props {
name(""),
description(""),
cores(0),
flops_f32 (0.0f),
flops_f16 (0.0f),
flops_f32_f32(0.0f),
flops_f16_f32(0.0f),
flops_q4k_f32(0.0f),
flops_q6k_f32(0.0f) {}
};
@@ -62,12 +62,12 @@ struct gpu_props {
const char * description;
float memory_free; // in GB
float memory_total; // in GB
float metal_flops_f32; // in GFLOPS
float metal_flops_f16; // in GFLOPS
float metal_flops_f32_f32; // in GFLOPS
float metal_flops_f16_f32; // in GFLOPS
float metal_flops_q4k_f32; // in GFLOPS
float metal_flops_q6k_f32; // in GFLOPS
float cuda_flops_f32; // in GFLOPS
float cuda_flops_f16; // in GFLOPS
float cuda_flops_f32_f32; // in GFLOPS
float cuda_flops_f16_f32; // in GFLOPS
float cuda_flops_q4k_f32; // in GFLOPS
float cuda_flops_q6k_f32; // in GFLOPS
@@ -76,45 +76,54 @@ struct gpu_props {
description(""),
memory_free (0.0f),
memory_total (0.0f),
metal_flops_f32 (0.0f),
metal_flops_f16 (0.0f),
metal_flops_f32_f32(0.0f),
metal_flops_f16_f32(0.0f),
metal_flops_q4k_f32(0.0f),
metal_flops_q6k_f32(0.0f),
cuda_flops_f32 (0.0f),
cuda_flops_f16 (0.0f),
cuda_flops_f32_f32 (0.0f),
cuda_flops_f16_f32 (0.0f),
cuda_flops_q4k_f32 (0.0f),
cuda_flops_q6k_f32 (0.0f) {}
};
struct model_flops {
// model flops
int64_t input_flops;
int64_t output_flops;
int64_t layer_flops;
// model params
int64_t output_f32_f32;
int64_t output_q6k_f32;
int64_t layer_f32_f32;
int64_t layer_f16_f32;
int64_t layer_q4k_f32;
int64_t layer_q6k_f32;
model_flops() :
output_f32_f32(0),
output_q6k_f32(0),
layer_f32_f32 (0),
layer_f16_f32 (0),
layer_q4k_f32 (0),
layer_q6k_f32 (0) {}
};
struct model_params {
int64_t input_params;
int64_t output_params;
int64_t layer_params;
model_flops() :
input_flops (0),
output_flops (0),
layer_flops (0),
input_params (0),
output_params(0),
model_params() :
input_params (0),
output_params(0),
layer_params (0) {}
};
struct device_info {
uint32_t rank;
const char * device_name;
float disk_read_bandwidth; // in GB/s
struct cpu_props cpu_props;
struct memory_info memory;
struct gpu_support gpu_support;
struct gpu_props gpu_props;
struct model_flops model_flops;
uint32_t rank;
const char * device_name;
float disk_read_bandwidth; // in GB/s
struct cpu_props cpu_props;
struct memory_info memory;
struct gpu_support gpu_support;
struct gpu_props gpu_props;
struct model_flops model_flops;
struct model_params model_params;
device_info() :
rank(0),
@@ -124,7 +133,8 @@ struct device_info {
memory(),
gpu_support(),
gpu_props(),
model_flops() {}
model_flops(),
model_params() {}
};
enum profiler_backend_type {
@@ -133,6 +143,12 @@ enum profiler_backend_type {
PROFILER_BACKEND_TYPE_CUDA = 2,
};
enum profiler_layer_type {
PROFILER_LAYER_INPUT = 0,
PROFILER_LAYER_OUTPUT = 1,
PROFILER_LAYER_BACKEND = 2,
};
const char * device_name(void);
uint32_t device_cpu_cores (void);