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

ggml : change ggml_graph_compute() API to not require context (#1999)

Browse source 
* ggml_graph_compute: deprecate using ggml_context, try resolve issue #287

* rewrite: no longer consider backward compitability; plan and make_plan

* minor: rename ctx as plan; const

* remove ggml_graph_compute from tests/test-grad0.c, but current change breaks backward

* add static ggml_graph_compute_sugar()

* minor: update comments

* reusable buffers

* ggml : more consistent naming + metal fixes

* ggml : fix docs

* tests : disable grad / opt + minor naming changes

* ggml : add ggml_graph_compute_with_ctx()

- backwards compatible API
- deduplicates a lot of copy-paste

* ci : enable test-grad0

* examples : factor out plan allocation into a helper function

* llama : factor out plan stuff into a helper function

* ci : fix env

* llama : fix duplicate symbols + refactor example benchmark

* ggml : remove obsolete assert + refactor n_tasks section

* ggml : fix indentation in switch

* llama : avoid unnecessary bool

* ggml : remove comments from source file and match order in header

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
This commit is contained in:
Qingyou Meng 2023-07-08 00:24:01 +08:00 • committed by GitHub
parent 7242140283
commit 1d656d6360
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
13 changed files with 571 additions and 449 deletions
Download patch file Download diff file Expand all files Collapse all files

29
examples/benchmark/benchmark-matmult.cpp
View file

@ -20,6 +20,17 @@
#pragma warning(disable: 4244 4267) // possible loss of data
#endif
void ggml_graph_compute_helper(std::vector<uint8_t> & buf, ggml_cgraph * graph, int n_threads) {
struct ggml_cplan plan = ggml_graph_plan(graph, n_threads);
if (plan.work_size > 0) {
buf.resize(plan.work_size);
plan.work_data = buf.data();
}
ggml_graph_compute(graph, &plan);
}
float tensor_sum_elements(const ggml_tensor * tensor) {
float sum = 0;
if (tensor->type==GGML_TYPE_F32) {
@ -159,13 +170,14 @@ int main(int argc, char ** argv) {
// printf("Creating compute graph\n");
struct ggml_cgraph gf = ggml_build_forward(m11xm2);
gf.n_threads=benchmark_params.n_threads;
printf("cgraph->n_threads=%i\n",gf.n_threads);
printf("n_threads=%i\n", benchmark_params.n_threads);
TENSOR_DUMP(m11);
TENSOR_DUMP(m2);
ggml_graph_compute(ctx, &gf);
std::vector<uint8_t> work_buffer;
ggml_graph_compute_helper(work_buffer, &gf, benchmark_params.n_threads);
TENSOR_DUMP(gf.nodes[0]);
@ -187,7 +199,6 @@ int main(int argc, char ** argv) {
// printf("Creating compute graph\n");
struct ggml_cgraph gf31 = ggml_build_forward(q31);
gf31.n_threads=benchmark_params.n_threads;
// Set up a second graph computation to make sure we override the CPU cache lines
// printf("Creating new tensor q12 & Running quantize\n");
@ -199,8 +210,7 @@ int main(int argc, char ** argv) {
//printf("Creating compute graph\n");
struct ggml_cgraph gf32 = ggml_build_forward(q32);
gf32.n_threads=benchmark_params.n_threads;
printf("cgraph->n_threads=%i\n",gf31.n_threads);
printf("n_threads=%i\n", benchmark_params.n_threads);
const int dimx = sizex;
const int dimy = sizey;
@ -221,14 +231,15 @@ int main(int argc, char ** argv) {
long long int start = ggml_time_us();
//printf("Running ggml_graph_compute\n");
ggml_graph_compute(ctx, &gf31);
ggml_graph_compute_helper(work_buffer, &gf31, benchmark_params.n_threads);
long long int stop = ggml_time_us();
long long int usec = stop-start;
double gflops = (double)(flops_per_matrix)/usec/1000.0;
gflops_sum += gflops;
printf("%9i;%8i;%6i;%6i;%6i;%15lli;%18lli;%10.2f\n",
i,
gf31.n_threads,
benchmark_params.n_threads,
sizex, sizey, sizez, flops_per_matrix,
usec,gflops);
@ -253,7 +264,7 @@ int main(int argc, char ** argv) {
}
// Running a different graph computation to make sure we override the CPU cache lines
ggml_graph_compute(ctx, &gf32);
ggml_graph_compute_helper(work_buffer, &gf32, benchmark_params.n_threads);
}
printf("\n");
printf("Average%78.2f\n",gflops_sum/((double)benchmark_params.n_iterations));