Skip to content

Commit 542378f

Browse files
authored
Merge branch 'main' into dev_enable_le
2 parents 11034ee + 4eb5571 commit 542378f

File tree

537 files changed

+16973
-6613
lines changed

Some content is hidden

Large Commits have some content hidden by default. Use the searchbox below for content that may be hidden.

537 files changed

+16973
-6613
lines changed
Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -1 +1 @@
1-
44d8d54e38c0258357d4e92e1fefe21e845947a3
1+
467660923a5a25e4718e1d6697b93ff1bab4e807
Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -1 +1 @@
1-
53a2908a10f414a2f85caa06703a26a40e873869
1+
e6f766c7d750d40603eee3f66c5915bac606b3ea

.ci/docker/requirements-ci.txt

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -6,7 +6,7 @@ sympy==1.12
66
timm==0.6.13
77
tomli==2.0.1
88
torchsr==1.0.4
9-
transformers==4.47.1
9+
transformers==4.56.1
1010
zstd==1.5.5.1
1111
pandas>=2.2.2; python_version >= '3.10'
1212
pytest==7.2.0

.ci/scripts/setup-windows-msvc.ps1

Lines changed: 52 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,52 @@
1+
conda create --yes --quiet -n et python=3.12
2+
conda activate et
3+
4+
# Install cmake
5+
conda install -y cmake
6+
7+
# Activate the VS environment - this is required for MSVC to work
8+
# There are a bunch of environment variables that it requires.
9+
# See https://learn.microsoft.com/en-us/cpp/build/building-on-the-command-line.
10+
& "C:\Program Files (x86)\Microsoft Visual Studio\2022\BuildTools\Common7\Tools\Launch-VsDevShell.ps1" -Arch amd64
11+
12+
# Install CI requirements
13+
pip install -r .ci/docker/requirements-ci.txt
14+
15+
# Create build directory
16+
$buildDir = "cmake-out-msvc"
17+
if (Test-Path -Path $buildDir) {
18+
Remove-Item -Path $buildDir -Recurse -Force
19+
}
20+
New-Item -Path $buildDir -ItemType Directory
21+
22+
# Configure CMake with MSVC (not ClangCL) and disable custom/quantized ops
23+
cmake -S . -B $buildDir `
24+
-DCMAKE_BUILD_TYPE=Release `
25+
-DEXECUTORCH_BUILD_EXECUTOR_RUNNER=ON `
26+
-DEXECUTORCH_BUILD_EXTENSION_DATA_LOADER=ON `
27+
-DEXECUTORCH_BUILD_EXTENSION_MODULE=ON `
28+
-DEXECUTORCH_BUILD_EXTENSION_TENSOR=ON `
29+
-DEXECUTORCH_BUILD_EXTENSION_FLAT_TENSOR=ON `
30+
-DEXECUTORCH_BUILD_EXTENSION_NAMED_DATA_MAP=ON `
31+
-DEXECUTORCH_BUILD_KERNELS_OPTIMIZED=ON `
32+
-DEXECUTORCH_BUILD_KERNELS_CUSTOM=OFF `
33+
-DEXECUTORCH_BUILD_KERNELS_CUSTOM_AOT=OFF `
34+
-DEXECUTORCH_BUILD_KERNELS_QUANTIZED=OFF `
35+
-DEXECUTORCH_BUILD_XNNPACK=ON `
36+
-DEXECUTORCH_BUILD_EXTENSION_LLM=ON `
37+
-DEXECUTORCH_BUILD_EXTENSION_LLM_RUNNER=ON
38+
39+
if ($LASTEXITCODE -ne 0) {
40+
Write-Host "CMake configuration failed. Exit code: $LASTEXITCODE."
41+
exit $LASTEXITCODE
42+
}
43+
44+
# Build with MSVC
45+
cmake --build $buildDir --config Release -j16
46+
47+
if ($LASTEXITCODE -ne 0) {
48+
Write-Host "Build failed. Exit code: $LASTEXITCODE."
49+
exit $LASTEXITCODE
50+
}
51+
52+
Write-Host "MSVC build completed successfully!"

.ci/scripts/test_phi_3_mini.sh

Lines changed: 11 additions & 12 deletions
Original file line numberDiff line numberDiff line change
@@ -36,34 +36,33 @@ cmake_build_phi_3_mini() {
3636
cmake --build ${BUILD_DIR}/${MODEL_DIR} -j${NPROC} --config ${BUILD_TYPE}
3737
}
3838

39-
# Download and convert tokenizer.model
39+
# Download tokenizer.model
4040
prepare_tokenizer() {
41-
echo "Downloading and converting tokenizer.model"
42-
wget -O tokenizer.model "https://huggingface.co/microsoft/Phi-3-mini-128k-instruct/resolve/main/tokenizer.model?download=true"
43-
$PYTHON_EXECUTABLE -m pytorch_tokenizers.tools.llama2c.convert -t tokenizer.model -o tokenizer.bin
41+
echo "Downloading tokenizer.model"
42+
wget -O tokenizer.model "https://huggingface.co/microsoft/Phi-3-mini-4k-instruct/resolve/main/tokenizer.model?download=true"
4443
}
4544

4645
# Export phi-3-mini model to pte
4746
export_phi_3_mini () {
4847
echo "Exporting phi-3-mini. This will take a few minutes"
49-
$PYTHON_EXECUTABLE -m executorch.examples.models.phi-3-mini.export_phi-3-mini -c "4k" -s 128 -o phi-3-mini.pte
48+
optimum-cli export executorch --model microsoft/Phi-3-mini-4k-instruct --task text-generation --recipe xnnpack --output_dir ./
5049
}
5150

5251
run_and_verify() {
5352
NOW=$(date +"%H:%M:%S")
5453
echo "Starting to run phi-3-mini runner at ${NOW}"
55-
if [[ ! -f "phi-3-mini.pte" ]]; then
56-
echo "Export failed. Abort"
54+
if [[ ! -f "model.pte" ]]; then
55+
echo "Missing model artifact. Abort"
5756
exit 1
5857
fi
59-
if [[ ! -f "tokenizer.bin" ]]; then
60-
echo "tokenizer.bin is missing."
58+
if [[ ! -f "tokenizer.model" ]]; then
59+
echo "tokenizer.model is missing."
6160
exit 1
6261
fi
6362

6463
${BUILD_DIR}/${MODEL_DIR}/phi_3_mini_runner \
65-
--model_path=phi-3-mini.pte \
66-
--tokenizer_path=tokenizer.bin \
64+
--model_path=model.pte \
65+
--tokenizer_path=tokenizer.model \
6766
--seq_len=60 \
6867
--temperature=0 \
6968
--prompt="<|system|>
@@ -92,7 +91,7 @@ What is the capital of France?<|end|>
9291
cmake_install_executorch_libraries
9392
cmake_build_phi_3_mini
9493

95-
# Step 2. Export the tokenizer and model
94+
# Step 2. Export the model
9695
prepare_tokenizer
9796
export_phi_3_mini
9897

.ci/scripts/test_qnn_static_llama.sh

Lines changed: 0 additions & 69 deletions
This file was deleted.

.ci/scripts/test_qnn_static_llm.sh

Lines changed: 94 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,94 @@
1+
#!/bin/bash
2+
# Copyright (c) Qualcomm Innovation Center, Inc.
3+
# All rights reserved
4+
#
5+
# This source code is licensed under the BSD-style license found in the
6+
# LICENSE file in the root directory of this source tree.
7+
8+
set -euxo pipefail
9+
10+
source "$(dirname "${BASH_SOURCE[0]}")/utils.sh"
11+
12+
TASK_NAME=$1
13+
if [[ -z "${TASK_NAME:-}" ]]; then
14+
echo "Missing task name, exiting..."
15+
exit 1
16+
fi
17+
18+
19+
# Download QNN_SDK. If already downloaded, export environment path
20+
source "$(dirname "${BASH_SOURCE[0]}")/../../backends/qualcomm/scripts/install_qnn_sdk.sh"
21+
install_qnn
22+
23+
export EXECUTORCH_ROOT="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")/.." && pwd)"
24+
export LD_LIBRARY_PATH="${QNN_SDK_ROOT}/lib/x86_64-linux-clang"
25+
export PYTHONPATH=".."
26+
cp schema/program.fbs exir/_serialize/program.fbs
27+
cp schema/scalar_type.fbs exir/_serialize/scalar_type.fbs
28+
cp -f build-x86/backends/qualcomm/PyQnnManagerAdaptor.cpython-310-x86_64-linux-gnu.so backends/qualcomm/python
29+
cp -f build-x86/backends/qualcomm/PyQnnWrapperAdaptor.cpython-310-x86_64-linux-gnu.so backends/qualcomm/python
30+
31+
if [[ -z "${PYTHON_EXECUTABLE:-}" ]]; then
32+
PYTHON_EXECUTABLE=python3
33+
fi
34+
35+
which "${PYTHON_EXECUTABLE}"
36+
37+
# Although static llama CI does not require graphviz, it is required by test_qnn_delegate.py
38+
pip install graphviz
39+
40+
set +e
41+
42+
echo "Executing task: $TASK_NAME"
43+
if [[ "${TASK_NAME}" == "stories_110m" ]]; then
44+
# Download stories llama110m artifacts
45+
download_stories_model_artifacts
46+
echo "Creating tokenizer.bin"
47+
$PYTHON_EXECUTABLE -m pytorch_tokenizers.tools.llama2c.convert -t tokenizer.model -o tokenizer.bin
48+
49+
# Compile only as weight sharing is not applicable on x86.
50+
$PYTHON_EXECUTABLE backends/qualcomm/tests/test_qnn_delegate.py -k TestExampleLLMScript.test_llama_stories_110m --model SM8650 --build_folder build-android/ --executorch_root . --artifact_dir ./stories_110m_pte_size --llama_artifacts . --compile_only
51+
exit_code1=$?
52+
53+
# Checks accuracy with weight sharing disabled since x86 does not support weight sharing.
54+
$PYTHON_EXECUTABLE backends/qualcomm/tests/test_qnn_delegate.py -k TestExampleLLMScript.test_llama_stories_110m --model SM8650 --build_folder build-x86/ --executorch_root . --artifact_dir ./stories_110m_accuracy --llama_artifacts . --enable_x86_64
55+
exit_code2=$?
56+
57+
# Check the exit codes and print messages
58+
if [ $exit_code1 -ne 0 ]; then
59+
echo "Static Llama compile only with weight sharing test failed. $exit_code1."
60+
fi
61+
62+
if [ $exit_code2 -ne 0 ]; then
63+
echo "Static Llama accuracy test failed. $exit_code2."
64+
fi
65+
66+
if [ $exit_code1 -ne 0 ] || [ $exit_code2 -ne 0 ]; then
67+
exit 1
68+
else
69+
exit 0
70+
fi
71+
72+
elif [[ "${TASK_NAME}" == "stories_260k_bc" ]]; then
73+
74+
# Check BC
75+
bash backends/qualcomm/bc/test_qnn_static_llama_bc.sh
76+
exit_code1=$?
77+
if [ $exit_code1 -ne 0 ]; then
78+
exit 1
79+
else
80+
exit 0
81+
fi
82+
83+
elif [[ "${TASK_NAME}" == "smollm2_135m" ]]; then
84+
$PYTHON_EXECUTABLE backends/qualcomm/tests/test_qnn_delegate.py -k TestExampleLLMScript.test_static_smollm2 --model SM8650 --build_folder build-x86/ --executorch_root . --artifact_dir ./static_smollm2 --enable_x86_64
85+
exit_code1=$?
86+
if [ $exit_code1 -ne 0 ]; then
87+
exit 1
88+
else
89+
exit 0
90+
fi
91+
else
92+
echo "Unsupported task: $TASK_NAME"
93+
exit 1
94+
fi

.ci/scripts/utils.sh

Lines changed: 39 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -44,6 +44,44 @@ install_pip_dependencies() {
4444
popd || return
4545
}
4646

47+
dedupe_macos_loader_path_rpaths() {
48+
if [[ "$(uname)" != "Darwin" ]]; then
49+
return
50+
fi
51+
52+
local torch_lib_dir
53+
pushd ..
54+
torch_lib_dir=$(python -c "import importlib.util; print(importlib.util.find_spec('torch').submodule_search_locations[0])")/lib
55+
popd
56+
57+
if [[ -z "${torch_lib_dir}" || ! -d "${torch_lib_dir}" ]]; then
58+
return
59+
fi
60+
61+
local torch_libs=(
62+
"libtorch_cpu.dylib"
63+
"libtorch.dylib"
64+
"libc10.dylib"
65+
)
66+
67+
for lib_name in "${torch_libs[@]}"; do
68+
local lib_path="${torch_lib_dir}/${lib_name}"
69+
if [[ ! -f "${lib_path}" ]]; then
70+
continue
71+
fi
72+
73+
local removed=0
74+
# Repeatedly remove the @loader_path rpath entries until none remain.
75+
while install_name_tool -delete_rpath @loader_path "${lib_path}" 2>/dev/null; do
76+
removed=1
77+
done
78+
79+
if [[ "${removed}" == "1" ]]; then
80+
install_name_tool -add_rpath @loader_path "${lib_path}" || true
81+
fi
82+
done
83+
}
84+
4785
install_domains() {
4886
echo "Install torchvision and torchaudio"
4987
pip install --no-use-pep517 --user "git+https://github.com/pytorch/audio.git@${TORCHAUDIO_VERSION}"
@@ -101,6 +139,7 @@ install_pytorch_and_domains() {
101139
echo "Use cached wheel at ${cached_torch_wheel}"
102140
fi
103141

142+
dedupe_macos_loader_path_rpaths
104143
# Grab the pinned audio and vision commits from PyTorch
105144
TORCHAUDIO_VERSION=$(cat .github/ci_commit_pins/audio.txt)
106145
export TORCHAUDIO_VERSION

0 commit comments

Comments
 (0)