Skip to content

Arm backend: Install TOSA and VGF tooling from pip #18840

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
zingo merged 1 commit into from
Apr 14, 2026
Merged

Arm backend: Install TOSA and VGF tooling from pip #18840

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
5 changes: 2 additions & 3 deletions backends/arm/requirements-arm-tosa.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -7,6 +7,5 @@ ml_dtypes == 0.5.1
flatbuffers == 24.3.25
tosa-adapter-model-explorer == 0.1.0
ai-edge-model-explorer >= 0.1.16
# NOTE: Will be removed when tosa-tools is installed via pypi
pybind11 == 2.10.4
pytest-timeout == 2.4.0
pytest-timeout == 2.4.0
tosa-tools == 2026.2.1
6 changes: 3 additions & 3 deletions backends/arm/requirements-arm-vgf.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3,6 +3,6 @@
# This source code is licensed under the BSD-style license found in the
# LICENSE file in the root directory of this source tree.

ai_ml_emulation_layer_for_vulkan == 0.8.0
ai_ml_sdk_model_converter == 0.8.0
ai_ml_sdk_vgf_library == 0.8.0
ai_ml_emulation_layer_for_vulkan == 0.9.0
ai_ml_sdk_model_converter == 0.9.0
ai_ml_sdk_vgf_library == 0.9.0
3 changes: 2 additions & 1 deletion backends/arm/runtime/VGFBackend.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -157,7 +157,8 @@ class VGFBackend final : public ::executorch::runtime::BackendInterface {
new (repr) VgfRepr(
vk_instance, vk_physical_device, vk_device, vk_queue, vk_command_pool);

auto valid_vgf = repr->process_vgf(vgf_data, compile_specs);
auto valid_vgf =
repr->process_vgf(vgf_data, processed->size(), compile_specs);
if (!valid_vgf) {
ET_LOG(Error, "Failed to process VGF blob.");
return Error::Internal;
Expand Down
26 changes: 19 additions & 7 deletions backends/arm/runtime/VGFSetup.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,5 +1,5 @@
/*
* Copyright 2025 Arm Limited and/or its affiliates.
* Copyright 2025-2026 Arm Limited and/or its affiliates.
*
* This source code is licensed under the BSD-style license found in the
* LICENSE file in the root directory of this source tree.
Expand Down Expand Up @@ -324,26 +324,38 @@ static void debug_print_modules(
}
}

bool VgfRepr::process_vgf(const char* vgf_data, ArrayRef<CompileSpec> specs) {
bool VgfRepr::process_vgf(
const char* vgf_data,
size_t vgf_size,
ArrayRef<CompileSpec> specs) {
ET_LOG(Info, "Preparing VGF as Vulkan objects");

VkResult result;

// Prepare temporary decoders
unique_ptr<vgflib::HeaderDecoder> header_decoder =
vgflib::CreateHeaderDecoder(vgf_data);
vgflib::CreateHeaderDecoder(vgf_data, vgflib::HeaderSize(), vgf_size);
if (!header_decoder) {
ET_LOG(Error, "Failed to create VGF header decoder");
return false;
}

unique_ptr<vgflib::ModelSequenceTableDecoder> sequence_decoder =
vgflib::CreateModelSequenceTableDecoder(
vgf_data + header_decoder->GetModelSequenceTableOffset());
vgf_data + header_decoder->GetModelSequenceTableOffset(),
header_decoder->GetModelSequenceTableSize());
unique_ptr<vgflib::ModuleTableDecoder> module_decoder =
vgflib::CreateModuleTableDecoder(
vgf_data + header_decoder->GetModuleTableOffset());
vgf_data + header_decoder->GetModuleTableOffset(),
header_decoder->GetModuleTableSize());
Comment on lines 335 to +350
Copy link

Copilot AI Apr 13, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

header_decoder offsets/sizes are used to compute pointers and create the other decoders before verifying header_decoder->IsValid() / CheckVersion(). If the blob header is malformed, this can lead to out-of-bounds pointer arithmetic/reads. Validate the header (and ideally bounds-check offset+size against vgf_size) before constructing the table/constant decoders.

Copilot uses AI. Check for mistakes.
unique_ptr<vgflib::ModelResourceTableDecoder> resource_decoder =
vgflib::CreateModelResourceTableDecoder(
vgf_data + header_decoder->GetModelResourceTableOffset());
vgf_data + header_decoder->GetModelResourceTableOffset(),
header_decoder->GetModelResourceTableSize());
unique_ptr<vgflib::ConstantDecoder> constant_decoder =
vgflib::CreateConstantDecoder(
vgf_data + header_decoder->GetConstantsOffset());
vgf_data + header_decoder->GetConstantsOffset(),
header_decoder->GetConstantsSize());
// Check the VGF decoders
if (not(header_decoder && module_decoder && sequence_decoder &&
resource_decoder && constant_decoder && header_decoder->IsValid() &&
Expand Down
7 changes: 5 additions & 2 deletions backends/arm/runtime/VGFSetup.h
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,5 +1,5 @@
/*
* Copyright 2025 Arm Limited and/or its affiliates.
* Copyright 2025-2026 Arm Limited and/or its affiliates.
*
* This source code is licensed under the BSD-style license found in the
* LICENSE file in the root directory of this source tree.
Expand Down Expand Up @@ -58,7 +58,10 @@ class VgfRepr {
/*
* Process a VGF ready for execution, allocate necessary Vulkan objects.
*/
bool process_vgf(const char* vgf_data, ArrayRef<CompileSpec> specs);
bool process_vgf(
const char* vgf_data,
size_t vgf_size,
ArrayRef<CompileSpec> specs);

/*
* Execute the VGF we've previously processed.
Expand Down
10 changes: 10 additions & 0 deletions backends/arm/scripts/mlsdk_utils.sh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -291,6 +291,16 @@ PY
fi
fi

local vk_layer_line
vk_layer_line=$(echo "$exports" | grep 'VK_LAYER_PATH=' || true)
if [[ -n "${vk_layer_line}" ]]; then
local vk_layer_value=${vk_layer_line#export VK_LAYER_PATH=}
vk_layer_value=${vk_layer_value%%:\$VK_LAYER_PATH*}
if [[ -n "${vk_layer_value}" ]]; then
prepend_env_in_setup_path VK_LAYER_PATH "${vk_layer_value}"
fi
fi

local vk_add_line
vk_add_line=$(echo "$exports" | grep 'VK_ADD_LAYER_PATH=' || true)
if [[ -n "${vk_add_line}" ]]; then
Expand Down
30 changes: 0 additions & 30 deletions examples/arm/setup.sh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -335,36 +335,6 @@ if [[ $is_script_sourced -eq 0 ]]; then
CMAKE_POLICY_VERSION_MINIMUM=3.5 \
pip install --no-dependencies -r "$et_dir/backends/arm/requirements-arm-tosa.txt"
Copy link

Copilot AI Apr 13, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

pip install --no-dependencies will skip transitive deps for the newly added tosa-tools PyPI package. Either remove --no-dependencies for this install or explicitly list all tosa-tools runtime/build dependencies in requirements-arm-tosa.txt to avoid incomplete/fragile installs.

Suggested change
pip install --no-dependencies -r "$et_dir/backends/arm/requirements-arm-tosa.txt"
pip install -r "$et_dir/backends/arm/requirements-arm-tosa.txt"

Copilot uses AI. Check for mistakes.

pushd "$root_dir"
if [[ ! -d "tosa-tools" ]]; then
git clone https://git.gitlab.arm.com/tosa/tosa-tools.git
fi

pushd tosa-tools
git checkout v2025.11.2

if [[ ! -d "reference_model" ]]; then
log_step "main" "[error] Missing reference_model directory in tosa-tools repo."
exit 1
fi
if [[ ! -d "serialization" ]]; then
log_step "main" "[error] Missing serialization directory in tosa-tools repo."
exit 1
fi

export CMAKE_BUILD_PARALLEL_LEVEL="$(get_parallel_jobs)"

CMAKE_POLICY_VERSION_MINIMUM=3.5 \
BUILD_PYBIND=1 \
BUILD_TOSA_REFERENCE_MODEL_TESTS=0 \
pip install --no-dependencies ./reference_model

CMAKE_POLICY_VERSION_MINIMUM=3.5 \
BUILD_PYBIND=1 \
pip install --no-dependencies ./serialization
popd
popd

if [[ "${enable_vela}" -eq 1 ]]; then
log_step "deps" "Installing Ethos-U Vela compiler"
setup_ethos_u_tools
Expand Down
Loading