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

quantize-stats: add option to test against reference quantization

Browse source 
Expose reference quantization implementation and add option to use it
for tests.
This commit is contained in:
Håkon H. Hitland 2023-04-05 03:30:23 +02:00
parent d4915074c4
commit 63cfa43200
3 changed files with 21 additions and 7 deletions
Download patch file Download diff file Expand all files Collapse all files

13
examples/quantize-stats/quantize-stats.cpp
View file

@ -24,6 +24,7 @@ struct quantize_stats_params {
bool verbose = false; bool verbose = false;
bool per_layer_stats = false; bool per_layer_stats = false;
bool print_histogram = false; bool print_histogram = false;
bool reference = false;
std::vector<std::string> include_layers; std::vector<std::string> include_layers;
std::vector<std::string> exclude_layers; std::vector<std::string> exclude_layers;
std::vector<enum ggml_type> include_types; std::vector<enum ggml_type> include_types;
@ -49,6 +50,8 @@ void quantize_stats_print_usage(int /*argc*/, char ** argv) {
fprintf(stderr, " -h, --help show this help message and exit\n"); fprintf(stderr, " -h, --help show this help message and exit\n");
fprintf(stderr, " -m FNAME, --model FNAME\n"); fprintf(stderr, " -m FNAME, --model FNAME\n");
fprintf(stderr, " model path (default: %s)\n", params.model.c_str()); fprintf(stderr, " model path (default: %s)\n", params.model.c_str());
fprintf(stderr, " -r, --reference\n");
fprintf(stderr, " use reference implementation (default: false)\n");
fprintf(stderr, " -v, --verbose\n"); fprintf(stderr, " -v, --verbose\n");
fprintf(stderr, " verbose output (default: false)\n"); fprintf(stderr, " verbose output (default: false)\n");
fprintf(stderr, " -p, --per-layer-stats\n"); fprintf(stderr, " -p, --per-layer-stats\n");
@ -135,6 +138,7 @@ void test_roundtrip_on_layer(
std::string & name, std::string & name,
bool print_layer_stats, bool print_layer_stats,
const quantize_fns_t & qfns, const quantize_fns_t & qfns,
bool use_reference,
const ggml_tensor * layer, const ggml_tensor * layer,
float * input_scratch, float * input_scratch,
char *quantized_scratch, char *quantized_scratch,
@ -156,7 +160,11 @@ void test_roundtrip_on_layer(
input_scratch = ggml_get_data_f32(layer) + offset; input_scratch = ggml_get_data_f32(layer) + offset;
} }
qfns.quantize_row_q(input_scratch, quantized_scratch, chunk_size); 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); qfns.dequantize_row_q(quantized_scratch, output_scratch, chunk_size);
update_error_stats(chunk_size, input_scratch, output_scratch, total_error); update_error_stats(chunk_size, input_scratch, output_scratch, total_error);
@ -184,6 +192,8 @@ int main(int argc, char ** argv) {
if (arg == "-h" || arg == "--help") { if (arg == "-h" || arg == "--help") {
quantize_stats_print_usage(argc, argv); quantize_stats_print_usage(argc, argv);
exit(0); exit(0);
} else if (arg == "-r" || arg == "--reference") {
params.reference = true;
} else if (arg == "-v") { } else if (arg == "-v") {
params.verbose = true; params.verbose = true;
} else if (arg == "-p" || arg == "--per-layer-stats") { } else if (arg == "-p" || arg == "--per-layer-stats") {
@ -320,6 +330,7 @@ int main(int argc, char ** argv) {
layer_name, layer_name,
params.per_layer_stats, params.per_layer_stats,
qfns, qfns,
params.reference,
kv_tensor.second, kv_tensor.second,
input_scratch.data(), input_scratch.data(),
quantized_scratch.data(), quantized_scratch.data(),

14
ggml.c
View file

@ -6499,14 +6499,16 @@ static void ggml_compute_forward_mul_mat_f16_f32(
static const quantize_fns_t quantize_fns[GGML_TYPE_COUNT] = { static const quantize_fns_t quantize_fns[GGML_TYPE_COUNT] = {
[GGML_TYPE_Q4_0] = { [GGML_TYPE_Q4_0] = {
.dequantize_row_q = dequantize_row_q4_0, .dequantize_row_q = dequantize_row_q4_0,
.quantize_row_q = quantize_row_q4_0, .quantize_row_q = quantize_row_q4_0,
.vec_dot_q = ggml_vec_dot_q4_0, .quantize_row_q_reference = (quantize_row_q_t) quantize_row_q4_0_reference,
.vec_dot_q = ggml_vec_dot_q4_0,
}, },
[GGML_TYPE_Q4_1] = { [GGML_TYPE_Q4_1] = {
.dequantize_row_q = dequantize_row_q4_1, .dequantize_row_q = dequantize_row_q4_1,
.quantize_row_q = quantize_row_q4_1, .quantize_row_q = quantize_row_q4_1,
.vec_dot_q = ggml_vec_dot_q4_1, .quantize_row_q_reference = (quantize_row_q_t) quantize_row_q4_1_reference,
.vec_dot_q = ggml_vec_dot_q4_1,
}, },
}; };

1
ggml_internal.h
View file

@ -15,6 +15,7 @@ typedef void (*vec_dot_q_t)(const int n, float * restrict s, const void * restri
typedef struct { typedef struct {
dequantize_row_q_t dequantize_row_q; dequantize_row_q_t dequantize_row_q;
quantize_row_q_t quantize_row_q; quantize_row_q_t quantize_row_q;
quantize_row_q_t quantize_row_q_reference;
vec_dot_q_t vec_dot_q; vec_dot_q_t vec_dot_q;
} quantize_fns_t; } quantize_fns_t;