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

ggml : generalize quantize_fns for simpler FP16 handling (#1237)

Browse source 
* Generalize quantize_fns for simpler FP16 handling

* Remove call to ggml_cuda_mul_mat_get_wsize

* ci : disable FMA for mac os actions

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
This commit is contained in:
Stephan Walter 2023-07-05 16:13:06 +00:00 • committed by GitHub
parent 8567c76b53
commit 1b107b8550
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
9 changed files with 172 additions and 548 deletions
Download patch file Download diff file Expand all files Collapse all files

30
tests/test-quantize-fns.cpp
View file

@ -40,26 +40,26 @@ float array_rmse(const float * a1, const float * a2, size_t n) {
}
// Total quantization error on test data
float total_quantization_error(quantize_fns_t & qfns, size_t test_size, const float * test_data) {
float total_quantization_error(ggml_type_traits_t & qfns, size_t test_size, const float * test_data) {
std::vector<uint8_t> tmp_q(2*test_size);
std::vector<float> tmp_out(test_size);
qfns.quantize_row_q(test_data, tmp_q.data(), test_size);
qfns.dequantize_row_q(tmp_q.data(), tmp_out.data(), test_size);
qfns.from_float(test_data, tmp_q.data(), test_size);
qfns.to_float(tmp_q.data(), tmp_out.data(), test_size);
return array_rmse(test_data, tmp_out.data(), test_size);
}
// Total quantization error on test data
float reference_quantization_error(quantize_fns_t & qfns, size_t test_size, const float * test_data) {
float reference_quantization_error(ggml_type_traits_t & qfns, size_t test_size, const float * test_data) {
std::vector<uint8_t> tmp_q(2*test_size);
std::vector<float> tmp_out(test_size);
std::vector<float> tmp_out_ref(test_size);
qfns.quantize_row_q(test_data, tmp_q.data(), test_size);
qfns.dequantize_row_q(tmp_q.data(), tmp_out.data(), test_size);
qfns.from_float(test_data, tmp_q.data(), test_size);
qfns.to_float(tmp_q.data(), tmp_out.data(), test_size);
qfns.quantize_row_q_reference(test_data, tmp_q.data(), test_size);
qfns.dequantize_row_q(tmp_q.data(), tmp_out_ref.data(), test_size);
qfns.from_float_reference(test_data, tmp_q.data(), test_size);
qfns.to_float(tmp_q.data(), tmp_out_ref.data(), test_size);
return array_rmse(tmp_out.data(), tmp_out_ref.data(), test_size);
}
@ -73,15 +73,17 @@ float dot_product(const float * a1, const float * a2, size_t test_size) {
}
// Total dot product error
float dot_product_error(quantize_fns_t & qfns, size_t test_size, const float * test_data1, const float *test_data2) {
float dot_product_error(ggml_type_traits_t & qfns, size_t test_size, const float * test_data1, const float *test_data2) {
std::vector<uint8_t> tmp_q1(2*test_size);
std::vector<uint8_t> tmp_q2(2*test_size);
qfns.quantize_row_q (test_data1, tmp_q1.data(), test_size);
qfns.quantize_row_q_dot(test_data2, tmp_q2.data(), test_size);
auto vdot = ggml_internal_get_type_traits(qfns.vec_dot_type);
qfns.from_float(test_data1, tmp_q1.data(), test_size);
vdot.from_float(test_data2, tmp_q2.data(), test_size);
float result = INFINITY;
qfns.vec_dot_q(test_size, &result, tmp_q1.data(), tmp_q2.data());
qfns.vec_dot(test_size, &result, tmp_q1.data(), tmp_q2.data());
const float dot_ref = dot_product(test_data1, test_data2, test_size);
@ -123,9 +125,9 @@ int main(int argc, char * argv[]) {
for (int i = 0; i < GGML_TYPE_COUNT; i++) {
ggml_type type = (ggml_type) i;
quantize_fns_t qfns = ggml_internal_get_quantize_fn(i);
ggml_type_traits_t qfns = ggml_internal_get_type_traits(type);
if (qfns.quantize_row_q && qfns.dequantize_row_q) {
if (qfns.from_float && qfns.to_float) {
const float total_error = total_quantization_error(qfns, test_size, test_data.data());
const float max_quantization_error =
type == GGML_TYPE_Q2_K ? MAX_QUANTIZATION_TOTAL_ERROR_2BITS :