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

ggml : fix BLAS with unsupported types (#9775)

Browse source 
* ggml : do not use BLAS with types without to_float

* ggml : return pointer from ggml_internal_get_type_traits to avoid unnecessary copies

* ggml : rename ggml_internal_get_type_traits -> ggml_get_type_traits

it's not really internal if everybody uses it
This commit is contained in:
Diego Devesa 2024-10-08 14:21:43 +02:00 • committed by GitHub
parent 458367a906
commit dca1d4b58a
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
13 changed files with 75 additions and 74 deletions
Download patch file Download diff file Expand all files Collapse all files

6
pocs/vdot/q8dot.cpp
View file

@ -136,7 +136,7 @@ int main(int argc, char** argv) {
auto ggml_type = type == 0 ? GGML_TYPE_Q4_0 : GGML_TYPE_Q4_1;
auto funcs = ggml_internal_get_type_traits(ggml_type);
const auto * funcs = ggml_get_type_traits(ggml_type);
Stat simple, ggml;
@ -156,8 +156,8 @@ int main(int argc, char** argv) {
t1 = std::chrono::high_resolution_clock::now();
float fs;
if (type == 0) funcs.vec_dot(kVecSize * QK4_1, &fs, 0, x40.data(), 0, y.data(), 0, 1);
else funcs.vec_dot(kVecSize * QK4_1, &fs, 0, x41.data(), 0, y.data(), 0, 1);
if (type == 0) funcs->vec_dot(kVecSize * QK4_1, &fs, 0, x40.data(), 0, y.data(), 0, 1);
else funcs->vec_dot(kVecSize * QK4_1, &fs, 0, x41.data(), 0, y.data(), 0, 1);
t2 = std::chrono::high_resolution_clock::now();
t = 1e-3*std::chrono::duration_cast<std::chrono::nanoseconds>(t2-t1).count();
if (iloop > 3) ggml.addResult(fs, t);

14
pocs/vdot/vdot.cpp
View file

@ -236,7 +236,7 @@ int main(int argc, char** argv) {
int n4 = useQ4_1 ? kVecSize / QK4_1 : kVecSize / QK4_0; n4 = 64*((n4 + 63)/64);
int n8 = kVecSize / QK8_0; n8 = 64*((n8 + 63)/64);
auto funcs = useQ4_1 ? ggml_internal_get_type_traits(GGML_TYPE_Q4_1) : ggml_internal_get_type_traits(GGML_TYPE_Q4_0);
const auto * funcs = useQ4_1 ? ggml_get_type_traits(GGML_TYPE_Q4_1) : ggml_get_type_traits(GGML_TYPE_Q4_0);
std::vector<block_q4_0> q40;
std::vector<block_q4_1> q41;
@ -261,9 +261,9 @@ int main(int argc, char** argv) {
// Note, we do not include this in the timing as in practical application
// we already have the quantized model weights.
if (useQ4_1) {
funcs.from_float(x1.data(), q41.data(), kVecSize);
funcs->from_float(x1.data(), q41.data(), kVecSize);
} else {
funcs.from_float(x1.data(), q40.data(), kVecSize);
funcs->from_float(x1.data(), q40.data(), kVecSize);
}
// Now measure time the dot product needs using the "scalar" version above
@ -282,10 +282,10 @@ int main(int argc, char** argv) {
dot_q4_q8(kVecSize, &result, q40.data(), q8.data());
}
else {
auto vdot = ggml_internal_get_type_traits(funcs.vec_dot_type);
vdot.from_float(y1.data(), q8.data(), kVecSize);
if (useQ4_1) funcs.vec_dot(kVecSize, &result, 0, q41.data(), 0, q8.data(), 0, 1);
else funcs.vec_dot(kVecSize, &result, 0, q40.data(), 0, q8.data(), 0, 1);
const auto * vdot = ggml_get_type_traits(funcs->vec_dot_type);
vdot->from_float(y1.data(), q8.data(), kVecSize);
if (useQ4_1) funcs->vec_dot(kVecSize, &result, 0, q41.data(), 0, q8.data(), 0, 1);
else funcs->vec_dot(kVecSize, &result, 0, q40.data(), 0, q8.data(), 0, 1);
}
sumq += result;
t2 = std::chrono::high_resolution_clock::now();