Skip to content

llama : keep explicit host buffer overrides when using mmap #28223

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

Open
Inovello wants to merge 2 commits into
base: master
Choose a base branch
from
+18 −2
Open

llama : keep explicit host buffer overrides when using mmap #28223

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
5cfa6a8
llama : keep explicit host buffer overrides when using mmap
Inovello Sep 2, 2026
c59754b
llama : loading issue caused by copying tensors through mmap resolved
Inovello Sep 3, 2026
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
5 changes: 5 additions & 0 deletions common/arg.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -259,6 +259,11 @@ static void parse_tensor_buffer_overrides(const std::string & value, std::vector
if (buft) {
buft_list[ggml_backend_buft_name(buft)] = buft;
}
// host buffer types (e.g. CUDA_Host) so CPU-resident tensors can use pinned memory
auto * host_buft = ggml_backend_dev_host_buffer_type(dev);
if (host_buft) {
buft_list[ggml_backend_buft_name(host_buft)] = host_buft;
}
}

for (const auto & override : string_split<std::string>(value, ',')) {
Expand Down
15 changes: 13 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 @@ -1223,6 +1223,7 @@ struct ggml_tensor * llama_model_loader::create_tensor(
}

ggml_backend_buffer_type_t buft = nullptr;
bool buft_overridden = false;

// check overrides
if (tensor_buft_overrides) {
Expand All @@ -1241,6 +1242,7 @@ struct ggml_tensor * llama_model_loader::create_tensor(
}
} else {
buft = overrides->buft;
buft_overridden = true;
}

LLAMA_LOG_DEBUG("tensor %s (%zu MiB %s) buffer type overridden to %s\n",
Expand All @@ -1259,9 +1261,9 @@ struct ggml_tensor * llama_model_loader::create_tensor(
}
}

// avoid using a host buffer when using mmap
// avoid using a host buffer when using mmap, unless an override asks for it: the tensor is then copied into pinned memory
auto * buft_dev = ggml_backend_buft_get_device(buft);
if (use_mmap && buft_dev && buft == ggml_backend_dev_host_buffer_type(buft_dev)) {
if (use_mmap && !buft_overridden && buft_dev && buft == ggml_backend_dev_host_buffer_type(buft_dev)) {
auto * cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
if (!cpu_dev) {
throw std::runtime_error("no CPU backend found");
Expand Down Expand Up @@ -1642,6 +1644,15 @@ 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 (ggml_backend_buffer_is_host(cur->buffer)) {
// The destination is host memory (e.g. a pinned CUDA_Host buffer selected by -ot), so it can be
// filled directly from the file. Copying from the mapping instead would fault in every page one
// by one at queue depth 1 - and with --numa distribute the whole mapping carries POSIX_MADV_RANDOM,
// which disables readahead entirely. A plain read gets full readahead and is an order of magnitude
// faster. Non-host destinations keep the memcpy from the mapping below.
const auto & file = files.at(weight->idx);
file->seek(weight->offs, SEEK_SET);
file->read_raw(cur->data, n_size);
} else {
ggml_backend_tensor_set(cur, data, 0, n_size);
}
Expand Down