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

llama : multi-threaded quantization (#1075)

Browse source 
* Multi-threading quantization.

Not much gain for simple quantizations, bit it will be important
for quantizations that require more CPU cycles.

* Multi-threading for quantize-stats

It now does the job in ~14 seconds on my Mac for
Q4_0, Q4_1 and Q4_2. Single-threaded it was taking
more than 2 minutes after adding the more elaborate
version of Q4_2.

* Reviewer comments

* Avoiding compiler confusion

After changing chunk_size to const int as suggested by
@ggerganov, clang and GCC starting to warn me that I don't
need to capture it in the lambda. So, I removed it from the
capture list. But that makes the MSVC build fail. So,
making it a constexpr to make every compiler happy.

* Still fighting with lambda captures in MSVC

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
This commit is contained in:
Kawrakow 2023-04-20 19:42:27 +02:00 • committed by GitHub
parent e0305ead3a
commit 38de86a711
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
6 changed files with 183 additions and 61 deletions
Download patch file Download diff file Expand all files Collapse all files

137
examples/quantize-stats/quantize-stats.cpp
View file

@ -15,6 +15,8 @@
#include <string>
#include <unordered_map>
#include <vector>
#include <thread>
#include <mutex>
struct quantize_stats_params {
std::string model = "models/7B/ggml-model-f16.bin";
@ -27,7 +29,6 @@ struct quantize_stats_params {
std::vector<enum ggml_type> include_types;
};
const int64_t SCRATCH_ELEMENTS = 32*32;
const size_t HISTOGRAM_BUCKETS = 150;
const double HISTOGRAM_RANGE = 0.03;
@ -90,6 +91,13 @@ void update_error_stats(int64_t nelements, const float * input, const float * ou
stats.num_samples += nelements;
}
void combine_error_stats(error_stats & into, const error_stats & from) {
into.num_samples += from.num_samples;
into.total_error += from.total_error;
if (from.max_error > into.max_error) into.max_error = from.max_error;
for (size_t i=0; i<HISTOGRAM_BUCKETS; ++i) into.error_histogram[i] += from.error_histogram[i];
}
double find_quantile(const error_stats & stats, double quantile) {
double sum = std::accumulate(std::begin(stats.error_histogram), std::end(stats.error_histogram), 0.0);
@ -130,6 +138,36 @@ static bool tensor_is_contiguous(const struct ggml_tensor * tensor) {
tensor->nb[3] == tensor->nb[2]*tensor->ne[2];
}
void test_roundtrip_on_chunk(
const ggml_tensor * layer,
int64_t offset,
int64_t chunk_size,
const quantize_fns_t & qfns,
bool use_reference,
float * input_scratch,
char * quantized_scratch,
float * output_scratch,
error_stats & stats) {
if (layer->type == GGML_TYPE_F16) {
for (int i = 0; i < chunk_size; i++) {
input_scratch[i] = ggml_get_f32_1d(layer, i + offset);
}
} else {
input_scratch = ggml_get_data_f32(layer) + offset;
}
if (use_reference) {
qfns.quantize_row_q_reference(input_scratch, quantized_scratch, chunk_size);
} else {
qfns.quantize_row_q(input_scratch, quantized_scratch, chunk_size);
}
qfns.dequantize_row_q(quantized_scratch, output_scratch, chunk_size);
update_error_stats(chunk_size, input_scratch, output_scratch, stats);
}
// Run quantization function for a single layer and update error stats
void test_roundtrip_on_layer(
std::string & name,
@ -137,40 +175,61 @@ void test_roundtrip_on_layer(
const quantize_fns_t & qfns,
bool use_reference,
const ggml_tensor * layer,
float * input_scratch,
char *quantized_scratch,
float * output_scratch,
error_stats & total_error) {
std::vector<float> & input_scratch,
std::vector<char> & quantized_scratch,
std::vector<float> & output_scratch,
error_stats & total_error,
int max_thread = 0) {
assert(tensor_is_contiguous(layer));
error_stats layer_error {};
int64_t nelements = ggml_nelements(layer);
uint64_t nelements = ggml_nelements(layer);
for (int64_t offset = 0; offset < nelements; offset += SCRATCH_ELEMENTS) {
int64_t chunk_size = std::min(SCRATCH_ELEMENTS, nelements - offset);
if (layer->type == GGML_TYPE_F16) {
for (int i = 0; i < chunk_size; i++) {
input_scratch[i] = ggml_get_f32_1d(layer, i + offset);
}
} else {
input_scratch = ggml_get_data_f32(layer) + offset;
}
if (use_reference) {
qfns.quantize_row_q_reference(input_scratch, quantized_scratch, chunk_size);
} else {
qfns.quantize_row_q(input_scratch, quantized_scratch, chunk_size);
}
qfns.dequantize_row_q(quantized_scratch, output_scratch, chunk_size);
update_error_stats(chunk_size, input_scratch, output_scratch, total_error);
if (print_layer_stats) {
update_error_stats(chunk_size, input_scratch, output_scratch, layer_error);
}
float* input_scratch_ptr = nullptr;
if (layer->type == GGML_TYPE_F16) {
if (input_scratch.size() < nelements) input_scratch.resize(nelements);
input_scratch_ptr = input_scratch.data();
}
if (quantized_scratch.size() < 4*nelements) quantized_scratch.resize(4*nelements);
if (output_scratch.size() < nelements) output_scratch.resize(nelements);
if (max_thread < 1) max_thread = std::thread::hardware_concurrency();
int chunk_size = 32*512;
int num_chunks = (nelements + chunk_size - 1)/chunk_size;
if (num_chunks < 2 || max_thread < 2) {
test_roundtrip_on_chunk(layer, 0, nelements, qfns, use_reference, input_scratch_ptr, quantized_scratch.data(),
output_scratch.data(), print_layer_stats ? layer_error : total_error);
} else {
auto & stats = print_layer_stats ? layer_error : total_error;
std::mutex mutex;
uint64_t counter = 0;
auto compute = [&mutex, &counter, &stats, &qfns, nelements, layer, use_reference, input_scratch_ptr,
&quantized_scratch, &output_scratch, chunk_size] () {
error_stats local_stats {};
while (true) {
std::unique_lock<std::mutex> lock(mutex);
uint64_t offset = counter; counter += chunk_size;
if (offset >= nelements) {
combine_error_stats(stats, local_stats);
break;
}
lock.unlock();
uint64_t chunk = offset + chunk_size < nelements ? chunk_size : nelements - offset;
test_roundtrip_on_chunk(layer, offset, chunk, qfns, use_reference, input_scratch_ptr + offset,
quantized_scratch.data() + 4*offset, output_scratch.data() + offset, local_stats);
}
};
int nthread = std::min(num_chunks, max_thread);
std::vector<std::thread> workers(nthread-1);
for (auto& w : workers) w = std::thread(compute);
compute();
for (auto& w : workers) w.join();
}
if (print_layer_stats) {
print_error_stats(name, layer_error, false);
combine_error_stats(total_error, layer_error);
}
}
@ -181,6 +240,7 @@ int main(int argc, char ** argv) {
// read command line
int max_thread = 0;
bool invalid_param = false;
std::string arg;
for (int i = 1; i < argc; i++) {
@ -230,6 +290,12 @@ int main(int argc, char ** argv) {
fprintf(stderr, "error: %s not in list of types\n", argv[i]);
invalid_param = true;
}
} else if (arg == "-n" || arg == "--num-threads") {
if (++i >= argc) {
invalid_param = true;
break;
}
max_thread = atoi(argv[i]);
} else {
fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
quantize_stats_print_usage(argc, argv);
@ -295,9 +361,9 @@ int main(int argc, char ** argv) {
}
printf("testing %d layers with max size %" PRId64 "\n", included_layers, max_nelements);
// allocate scratch space
std::vector<float> input_scratch(SCRATCH_ELEMENTS);
std::vector<char> quantized_scratch(SCRATCH_ELEMENTS*4);
std::vector<float> output_scratch(SCRATCH_ELEMENTS);
std::vector<float> input_scratch;
std::vector<char> quantized_scratch;
std::vector<float> output_scratch;
// loop throught quantization types
for (int i = 0; i < GGML_TYPE_COUNT; i++) {
@ -328,10 +394,11 @@ int main(int argc, char ** argv) {
qfns,
params.reference,
kv_tensor.second,
input_scratch.data(),
quantized_scratch.data(),
output_scratch.data(),
global_stats
input_scratch,
quantized_scratch,
output_scratch,
global_stats,
max_thread
);
}