Thanks to visit codestin.com
Credit goes to github.com

Skip to content

ggml-backend : add device and backend reg interfaces #9707

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
slaren merged 14 commits into from
Oct 2, 2024
Merged

ggml-backend : add device and backend reg interfaces #9707

Changes from 1 commit
Commits
Show all changes
14 commits
Select commit Hold shift + click to select a range
0cbdf13
ggml-backend : add device and backend reg interfaces
slaren Oct 1, 2024
805fea9
Update ggml/src/ggml-backend-impl.h
slaren Oct 1, 2024
6ff0e7a
add device props/caps, fully support async upload for all compatible …
slaren Oct 1, 2024
04ef648
update other backends
slaren Oct 2, 2024
db53f8e
fix pipeline parallelism check
slaren Oct 2, 2024
f9cab02
removed unused function, add missing statics
slaren Oct 2, 2024
2a60833
Merge remote-tracking branch 'origin/master' into sl/backend-registry-2
slaren Oct 2, 2024
6ff0d67
remove move unused reg_init functions from backends
slaren Oct 2, 2024
b5516aa
fix align [no ci]
slaren Oct 2, 2024
dc475c3
fix consistency issues with the usage of main_gpu
slaren Oct 2, 2024
d0c4954
move device backend_reg to the struct
slaren Oct 2, 2024
cfef355
fix some inconsistencies in the names of functions
slaren Oct 2, 2024
ffeca35
fix more naming inconsistencies, make interface structs const
slaren Oct 2, 2024
a9d172c
Merge remote-tracking branch 'origin/master' into sl/backend-registry-2
slaren Oct 2, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
Prev Previous commit
Next Next commit
fix pipeline parallelism check
  • Loading branch information
@slaren
slaren committed Oct 2, 2024
commit db53f8ef061ba463cc358caa2f4f6064e3905be2
30 changes: 24 additions & 6 deletions src/llama.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -19494,19 +19494,37 @@ struct llama_context * llama_new_context_with_model(
// buffer used to store the computation graph and the tensor meta data
ctx->buf_compute_meta.resize(ggml_tensor_overhead()*max_nodes + ggml_graph_overhead_custom(max_nodes, false));

// TODO: move these checks to ggml_backend_sched
// enabling pipeline parallelism in the scheduler increases memory usage, so it is only done when necessary
bool pipeline_parallel =
llama_get_device_count(*model) > 1 &&
model->n_gpu_layers > (int)model->hparams.n_layer &&
model->split_mode == LLAMA_SPLIT_MODE_LAYER &&
params.offload_kqv;

// FIXME
#if !defined(GGML_USE_CUDA) && false
// pipeline parallelism requires support for async compute and events
// currently this is only implemented in the CUDA backend
pipeline_parallel = false;
#endif
// pipeline parallelism requires support for async compute and events in all devices
if (pipeline_parallel) {
for (auto * backend : ctx->backends) {
if (ggml_backend_is_cpu(backend)) {
// ignore CPU backend
continue;
}
auto * dev = ggml_backend_get_device(backend);
if (!dev) {
// backend is using old interface, not supported
pipeline_parallel = false;
break;
}
ggml_backend_dev_props props;
ggml_backend_dev_get_props(dev, &props);
if (!props.caps.async || !props.caps.events) {
// device does not support async compute or events
pipeline_parallel = false;
break;
}
}
}

ctx->sched = ggml_backend_sched_new(ctx->backends.data(), backend_buft.data(), ctx->backends.size(), max_nodes, pipeline_parallel);

if (pipeline_parallel) {
Expand Down