Skip to content

Commit d80dfa3

Browse files
pytorchbotpytorchbot
authored
CI test for lora (#13068)
Pull Request resolved: #12724 Add CI test for LoRA: note, not program-data separated yet. ghstack-source-id: 299976063 Differential Revision: [D78751767](https://our.internmc.facebook.com/intern/diff/D78751767/)
1 parent e38ca0d commit d80dfa3

File tree

3 files changed

+131
-1
lines changed

3 files changed

+131
-1
lines changed

‎.ci/scripts/test_llama_lora.sh‎

Lines changed: 96 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,96 @@
1+
#!/bin/bash
2+
# Copyright (c) Meta Platforms, Inc. and affiliates.
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 -exu
9+
# shellcheck source=/dev/null
10+
source "$(dirname "${BASH_SOURCE[0]}")/utils.sh"
11+
12+
cmake_install_executorch_libraries() {
13+
echo "Installing libexecutorch.a, libextension_module.so, libportable_ops_lib.a"
14+
rm -rf cmake-out
15+
retry cmake --preset llm \
16+
-DCMAKE_INSTALL_PREFIX=cmake-out \
17+
-DCMAKE_BUILD_TYPE=Release
18+
cmake --build cmake-out -j9 --target install --config Release
19+
}
20+
21+
cmake_build_llama_runner() {
22+
echo "Building llama runner"
23+
pushd extension/llm/tokenizers
24+
echo "Updating tokenizers submodule"
25+
git submodule update --init
26+
popd
27+
dir="examples/models/llama"
28+
retry cmake \
29+
-DBUILD_TESTING=OFF \
30+
-DCMAKE_INSTALL_PREFIX=cmake-out \
31+
-DCMAKE_BUILD_TYPE=Release \
32+
-Bcmake-out/${dir} \
33+
${dir}
34+
cmake --build cmake-out/${dir} -j9 --config Release
35+
}
36+
37+
cleanup_files() {
38+
echo "Deleting downloaded and generated files"
39+
rm -rf "${DOWNLOADED_PATH}/"
40+
rm result.txt
41+
}
42+
43+
# Download model artifacts from HF Hub.
44+
# Hosting in personal repo for now.
45+
HF_MODEL_REPO="lucylq/llama3_1B_lora"
46+
DOWNLOADED_PATH=$(
47+
bash "$(dirname "${BASH_SOURCE[0]}")/download_hf_hub.sh" \
48+
--model_id "${HF_MODEL_REPO}" \
49+
--files "adapter_config.json" "adapter_model.pt" "consolidated.00.pth" "params.json" "tokenizer.model"
50+
)
51+
EXPORTED_MODEL_NAME="llama_3_2_1B_lora.pte"
52+
# Export model.
53+
$PYTHON_EXECUTABLE -m extension.llm.export.export_llm \
54+
base.checkpoint="${DOWNLOADED_PATH}/consolidated.00.pth" \
55+
base.params="${DOWNLOADED_PATH}/params.json" \
56+
base.adapter_checkpoint="${DOWNLOADED_PATH}/adapter_model.pt" \
57+
base.adapter_config="${DOWNLOADED_PATH}/adapter_config.json" \
58+
base.tokenizer_path="${DOWNLOADED_PATH}/tokenizer.model" \
59+
model.use_kv_cache=true \
60+
model.use_sdpa_with_kv_cache=true \
61+
model.dtype_override="fp32" \
62+
backend.xnnpack.enabled=true \
63+
backend.xnnpack.extended_ops=true \
64+
export.output_name="${EXPORTED_MODEL_NAME}"
65+
66+
# Build llama runner.
67+
cmake_install_executorch_libraries
68+
cmake_build_llama_runner
69+
70+
PROMPT="What happens if you eat watermelon seeds?"
71+
# Run llama runner
72+
RUNTIME_ARGS="--model_path=${EXPORTED_MODEL_NAME} --tokenizer_path=${DOWNLOADED_PATH}/tokenizer.model --temperature=0 --seq_len=20 --warmup=1"
73+
74+
NOW=$(date +"%H:%M:%S")
75+
echo "Starting to run llama runner at ${NOW}"
76+
# shellcheck source=/dev/null
77+
cmake-out/examples/models/llama/llama_main --prompt="${PROMPT}" ${RUNTIME_ARGS} > result.txt
78+
NOW=$(date +"%H:%M:%S")
79+
echo "Finished at ${NOW}"
80+
81+
RESULT=$(cat result.txt)
82+
EXPECTED_PREFIX="What happens if you eat watermelon seeds? Watermelon seeds are a good source of vitamin C,"
83+
84+
if [[ "${RESULT}" == "${EXPECTED_PREFIX}"* ]]; then
85+
echo "Expected result prefix: ${EXPECTED_PREFIX}"
86+
echo "Actual result: ${RESULT}"
87+
echo "Success"
88+
cleanup_files
89+
else
90+
echo "Expected result prefix: ${EXPECTED_PREFIX}"
91+
echo "Actual result: ${RESULT}"
92+
echo "Failure; results not the same"
93+
94+
cleanup_files
95+
exit 1
96+
fi

‎.github/workflows/pull.yml‎

Lines changed: 30 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -687,6 +687,36 @@ jobs:
687687
# run llama runner in eager mode
688688
PYTHON_EXECUTABLE=python bash .ci/scripts/test_llama_runner_eager.sh
689689
690+
test-llama-lora-linux:
691+
name: test-llama-lora-linux
692+
uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
693+
permissions:
694+
id-token: write
695+
contents: read
696+
strategy:
697+
fail-fast: false
698+
with:
699+
runner: linux.24xlarge
700+
docker-image: ci-image:executorch-ubuntu-22.04-clang12
701+
submodules: 'recursive'
702+
ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
703+
timeout: 90
704+
script: |
705+
# The generic Linux job chooses to use base env, not the one setup by the image
706+
CONDA_ENV=$(conda env list --json | jq -r ".envs | .[-1]")
707+
conda activate "${CONDA_ENV}"
708+
709+
PYTHON_EXECUTABLE=python bash .ci/scripts/setup-linux.sh --build-tool "cmake"
710+
711+
# Install llama requirements
712+
bash examples/models/llama/install_requirements.sh
713+
714+
# install a recent version of torchtune.
715+
PYTHON_EXECUTABLE=python python -m pip install torchtune==0.7.0.dev20250730 --extra-index-url https://download.pytorch.org/whl/nightly/cpu
716+
717+
# run llama runner in eager mode
718+
PYTHON_EXECUTABLE=python bash .ci/scripts/test_llama_lora.sh
719+
690720
test-mediatek-models-linux:
691721
name: test-mediatek-models-linux
692722
uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main

‎extension/llm/export/config/llm_config.py‎

Lines changed: 5 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -60,7 +60,7 @@ class PreqMode(str, Enum):
6060
@dataclass
6161
class BaseConfig:
6262
"""
63-
Configurations specific to the model, e.g. whether it’s Qwen3 or Phi-4-mini,
63+
Configurations specific to the model, e.g. whether it's Qwen3 or Phi-4-mini,
6464
and are the minimal set of parameters needed to load the pretrained
6565
eager model and its weights.
6666
@@ -487,6 +487,10 @@ def from_args(cls, args: argparse.Namespace) -> "LlmConfig": # noqa: C901
487487
llm_config.base.checkpoint = args.checkpoint
488488
if hasattr(args, "checkpoint_dir"):
489489
llm_config.base.checkpoint_dir = args.checkpoint_dir
490+
if hasattr(args, "adapter_checkpoint"):
491+
llm_config.base.adapter_checkpoint = args.adapter_checkpoint
492+
if hasattr(args, "adapter_config"):
493+
llm_config.base.adapter_config = args.adapter_config
490494
if hasattr(args, "tokenizer_path"):
491495
llm_config.base.tokenizer_path = args.tokenizer_path
492496
if hasattr(args, "metadata"):

0 commit comments

Comments
 (0)