Skip to content

llama : stage mmap uploads on integrated GPUs #26023

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

Closed
liminfei-amd wants to merge 1 commit into from
+39 −4
Closed

llama : stage mmap uploads on integrated GPUs #26023

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
29 changes: 27 additions & 2 deletions src/llama-model-loader.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1416,6 +1416,7 @@ bool llama_model_loader::load_all_data(
struct ggml_context * ctx,
llama_buf_map & bufs,
llama_mlocks * lmlocks,
bool use_staged_mmap_uploads,
llama_progress_callback progress_callback,
void * progress_callback_user_data) {
if (files.empty()) {
Expand Down Expand Up @@ -1447,10 +1448,9 @@ bool llama_model_loader::load_all_data(
std::vector<void *> host_ptrs;
size_t buffer_idx = 0; // buffer to use for async loads
ggml_backend_t upload_backend = [&](const char * func) -> ggml_backend_t {
if (use_mmap || check_tensors) {
if (check_tensors || (use_mmap && !use_staged_mmap_uploads)) {
return nullptr;
}
// When not using mmaped io use async uploads from pinned memory to GPU memory.
// First determine if the backend supports the necessary features for async uploads.
auto * buf = bufs.count(0) ? bufs.at(0) : nullptr;
if (!buf) {
Expand All @@ -1465,6 +1465,11 @@ bool llama_model_loader::load_all_data(
ggml_backend_buft_name(buft));
return nullptr;
}
if (use_mmap && ggml_backend_dev_type(dev) != GGML_BACKEND_DEVICE_TYPE_IGPU) {
LLAMA_LOG_DEBUG("%s: device %s is not an integrated GPU for staged mmap uploads\n", func,
ggml_backend_dev_name(dev));
return nullptr;
}

if (buft != ggml_backend_dev_buffer_type(dev)) {
LLAMA_LOG_DEBUG("%s: buffer type %s is not the default buffer type for device %s for async uploads\n", func,
Expand Down Expand Up @@ -1520,6 +1525,10 @@ bool llama_model_loader::load_all_data(
return backend;
}(__func__);

if (use_mmap && use_staged_mmap_uploads && !upload_backend) {
LLAMA_LOG_DEBUG("%s: staged mmap uploads unavailable, using synchronous fallback\n", __func__);
}

if (upload_backend) {
LLAMA_LOG_DEBUG("%s: using async uploads for device %s, buffer type %s, backend %s\n", __func__,
ggml_backend_dev_name(ggml_backend_get_device(upload_backend)),
Expand Down Expand Up @@ -1567,6 +1576,22 @@ bool llama_model_loader::load_all_data(
auto & mmap_used = mmaps_used[weight->idx];
mmap_used.first = std::min(mmap_used.first, weight->offs);
mmap_used.second = std::max(mmap_used.second, weight->offs + n_size);
} else if (upload_backend && !ggml_backend_buffer_is_host(cur->buffer)) {
const auto & file = files.at(weight->idx);
file->seek(weight->offs, SEEK_SET);
size_t data_copied = 0;
while (data_copied < n_size) {
const size_t chunk_size = std::min(buffer_size, n_size - data_copied);

ggml_backend_event_synchronize(events[buffer_idx]);
file->read_raw_unsafe(host_ptrs[buffer_idx], chunk_size);
ggml_backend_tensor_set_async(
upload_backend, cur, host_ptrs[buffer_idx], data_copied, chunk_size);
ggml_backend_event_record(events[buffer_idx], upload_backend);

data_copied += chunk_size;
buffer_idx = (buffer_idx + 1) % n_buffers;
}
} else {
ggml_backend_tensor_set(cur, data, 0, n_size);
}
Expand Down
1 change: 1 addition & 0 deletions src/llama-model-loader.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -198,6 +198,7 @@ struct llama_model_loader {
struct ggml_context * ctx,
llama_buf_map & bufs,
llama_mlocks * lmlocks,
bool use_staged_mmap_uploads,
llama_progress_callback progress_callback,
void * progress_callback_user_data);

Expand Down
13 changes: 11 additions & 2 deletions src/llama-model.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1517,7 +1517,15 @@ bool llama_model_base::load_tensors(llama_model_loader & ml) {
}
}

ml.init_mappings(true, use_mlock ? &pimpl->mlock_mmaps : nullptr);
const auto can_stage_mmap = [](const auto & layer) {
ggml_backend_dev_props props;
ggml_backend_dev_get_props(layer.dev, &props);
return props.type == GGML_BACKEND_DEVICE_TYPE_IGPU &&
props.caps.async && props.caps.host_buffer && props.caps.events;
};
const bool use_staged_mmap_uploads = !use_mlock && !params.check_tensors && can_stage_mmap(pimpl->dev_output) &&
std::all_of(pimpl->dev_layer.begin(), pimpl->dev_layer.end(), can_stage_mmap);
ml.init_mappings(!use_staged_mmap_uploads, use_mlock ? &pimpl->mlock_mmaps : nullptr);
pimpl->mappings.reserve(ml.mappings.size());

// create the backend buffers
Expand Down Expand Up @@ -1641,7 +1649,8 @@ bool llama_model_base::load_tensors(llama_model_loader & ml) {

// load tensor data
for (auto & [ctx, buf_map] : ctx_buf_maps) {
if (!ml.load_all_data(ctx, buf_map, use_mlock ? &pimpl->mlock_mmaps : NULL, params.progress_callback, params.progress_callback_user_data)) {
if (!ml.load_all_data(ctx, buf_map, use_mlock ? &pimpl->mlock_mmaps : NULL, use_staged_mmap_uploads,
params.progress_callback, params.progress_callback_user_data)) {
return false;
}
}
Expand Down