-
Notifications
You must be signed in to change notification settings - Fork 413
478 lines (437 loc) · 20 KB
/
android-perf.yml
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
name: android-perf
on:
schedule:
- cron: 0 0 * * *
pull_request:
paths:
- .github/workflows/android-perf.yml
- extension/benchmark/android/benchmark/android-llm-device-farm-test-spec.yml.j2
push:
branches:
- main
paths:
- .github/workflows/android-perf.yml
- extension/benchmark/android/benchmark/android-llm-device-farm-test-spec.yml.j2
# Note: GitHub has an upper limit of 10 inputs
workflow_dispatch:
inputs:
models:
description: Models to be benchmarked
required: false
type: string
default: stories110M
devices:
description: Target devices to run benchmark
required: false
type: string
default: samsung_galaxy_s22
benchmark_configs:
description: The list of configs used the benchmark
required: false
type: string
workflow_call:
inputs:
models:
description: Models to be benchmarked
required: false
type: string
default: stories110M
devices:
description: Target devices to run benchmark
required: false
type: string
default: samsung_galaxy_s22
benchmark_configs:
description: The list of configs used the benchmark
required: false
type: string
concurrency:
group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref_name }}-${{ github.ref_type == 'branch' && github.sha }}-${{ github.event_name == 'workflow_dispatch' }}-${{ github.event_name == 'schedule' }}
cancel-in-progress: true
jobs:
set-parameters:
runs-on: ubuntu-22.04
outputs:
benchmark_configs: ${{ steps.set-parameters.outputs.benchmark_configs }}
steps:
- uses: actions/checkout@v3
with:
submodules: 'false'
- uses: actions/setup-python@v4
with:
python-version: '3.10'
- name: Set parameters
id: set-parameters
shell: bash
env:
# Separate default values from the workflow dispatch. To ensure defaults are accessible
# during scheduled runs and to provide flexibility for different defaults between
# on-demand and periodic benchmarking.
CRON_DEFAULT_MODELS: ${{ github.event_name == 'schedule' && 'llama,mv3,mv2,ic4,ic3,resnet50,edsr,mobilebert,w2l,meta-llama/Llama-3.2-1B,meta-llama/Llama-3.2-1B-Instruct-SpinQuant_INT4_EO8,meta-llama/Llama-3.2-1B-Instruct-QLORA_INT4_EO8' || 'llama' }}
CRON_DEFAULT_DEVICES: samsung_galaxy_s22
run: |
set -eux
MODELS="${{ inputs.models }}"
if [ -z "$MODELS" ]; then
MODELS="$CRON_DEFAULT_MODELS"
fi
DEVICES="${{ inputs.devices }}"
if [ -z "$DEVICES" ]; then
DEVICES="$CRON_DEFAULT_DEVICES"
fi
PYTHONPATH="${PWD}" python .ci/scripts/gather_benchmark_configs.py \
--os "android" \
--models $MODELS \
--devices $DEVICES
prepare-test-specs:
runs-on: linux.2xlarge
needs: set-parameters
strategy:
matrix: ${{ fromJson(needs.set-parameters.outputs.benchmark_configs) }}
fail-fast: false
steps:
- uses: actions/checkout@v3
- name: Prepare the spec
id: prepare
shell: bash
env:
BENCHMARK_CONFIG: ${{ toJSON(matrix) }}
working-directory: extension/benchmark/android/benchmark
run: |
set -eux
# The model will be exported in the next step to this S3 path
MODEL_PATH="https://gha-artifacts.s3.amazonaws.com/${{ github.repository }}/${{ github.run_id }}/artifacts/${{ matrix.model }}_${{ matrix.config }}/model.zip"
# We could write a script to properly use jinja here, but there is only one variable,
# so let's just sed it
sed -i -e 's,{{ model_path }},'"${MODEL_PATH}"',g' android-llm-device-farm-test-spec.yml.j2
BENCHMARK_CONFIG_ID=$(echo "${{ matrix.model }}_${{ matrix.config }}" | sed -e 's/[^A-Za-z0-9._-]/_/g')
# The config for this benchmark runs, we save it in the test spec so that it can be fetched
# later by the upload script
sed -i -e 's,{{ benchmark_config_id }},'"${BENCHMARK_CONFIG_ID}"',g' android-llm-device-farm-test-spec.yml.j2
cp android-llm-device-farm-test-spec.yml.j2 android-llm-device-farm-test-spec.yml
# Just print the test spec for debugging
cat android-llm-device-farm-test-spec.yml
# Save the benchmark configs so that we can use it later in the dashboard
echo "${BENCHMARK_CONFIG}" > "${BENCHMARK_CONFIG_ID}.json"
echo "benchmark-config-id=${BENCHMARK_CONFIG_ID}" >> $GITHUB_OUTPUT
- name: Upload the spec
uses: seemethere/upload-artifact-s3@v5
with:
s3-bucket: gha-artifacts
s3-prefix: |
${{ github.repository }}/${{ github.run_id }}/artifacts/${{ matrix.model }}_${{ matrix.config }}
retention-days: 1
if-no-files-found: error
path: extension/benchmark/android/benchmark/android-llm-device-farm-test-spec.yml
- name: Update the benchmark configs
uses: seemethere/upload-artifact-s3@v5
with:
s3-bucket: gha-artifacts
s3-prefix: |
${{ github.repository }}/${{ github.run_id }}/artifacts/benchmark-configs/
retention-days: 1
if-no-files-found: error
path: extension/benchmark/android/benchmark/${{ steps.prepare.outputs.benchmark-config-id }}.json
export-models:
name: export-models
uses: pytorch/test-infra/.github/workflows/linux_job.yml@main
needs: set-parameters
secrets: inherit
strategy:
matrix: ${{ fromJson(needs.set-parameters.outputs.benchmark_configs) }}
fail-fast: false
with:
runner: linux.2xlarge.memory
docker-image: executorch-ubuntu-22.04-qnn-sdk
submodules: 'true'
timeout: 60
upload-artifact: android-models
upload-artifact-to-s3: true
secrets-env: EXECUTORCH_HF_TOKEN
script: |
# The generic Linux job chooses to use base env, not the one setup by the image
echo "::group::Setting up dev environment"
CONDA_ENV=$(conda env list --json | jq -r ".envs | .[-1]")
conda activate "${CONDA_ENV}"
if [[ ${{ matrix.config }} == *"qnn"* ]]; then
PYTHON_EXECUTABLE=python bash .ci/scripts/setup-qnn-deps.sh
PYTHON_EXECUTABLE=python bash .ci/scripts/build-qnn-sdk.sh
fi
PYTHON_EXECUTABLE=python bash .ci/scripts/setup-linux.sh "cmake"
# Install requirements for export_llama
PYTHON_EXECUTABLE=python bash examples/models/llama/install_requirements.sh
pip install -U "huggingface_hub[cli]"
huggingface-cli login --token $SECRET_EXECUTORCH_HF_TOKEN
pip install accelerate sentencepiece
pip list
ARTIFACTS_DIR_NAME=artifacts-to-be-uploaded/${{ matrix.model }}_${{ matrix.config }}
echo "::endgroup::"
echo "::group::Exporting ${{ matrix.config }} model: ${{ matrix.model }}"
BUILD_MODE="cmake"
if [[ ${{ matrix.model }} =~ ^[^/]+/[^/]+$ ]]; then
# HuggingFace model. Assume the pattern is always like "<org>/<repo>"
HF_MODEL_REPO=${{ matrix.model }}
OUT_ET_MODEL_NAME="$(echo "$HF_MODEL_REPO" | awk -F'/' '{print $2}' | sed 's/_/-/g' | tr '[:upper:]' '[:lower:]')_${{ matrix.config }}"
if [[ "$HF_MODEL_REPO" == meta-llama/* ]]; then
# Llama models on Hugging Face
if [[ ${{ matrix.config }} == "llama3_spinquant" ]]; then
# SpinQuant
# Download prequantized chceckpoint from Hugging Face
DOWNLOADED_PATH=$(
bash .ci/scripts/download_hf_hub.sh \
--model_id "${HF_MODEL_REPO}" \
--files "tokenizer.model" "params.json" "consolidated.00.pth"
)
# Export using ExecuTorch's model definition
python -m examples.models.llama.export_llama \
--model "llama3_2" \
--checkpoint "${DOWNLOADED_PATH}/consolidated.00.pth" \
--params "${DOWNLOADED_PATH}/params.json" \
--use_sdpa_with_kv_cache \
-X \
--xnnpack-extended-ops \
--preq_mode 8da4w_output_8da8w \
--preq_group_size 32 \
--max_seq_length 2048 \
--output_name "${OUT_ET_MODEL_NAME}.pte" \
-kv \
-d fp32 \
--preq_embedding_quantize 8,0 \
--use_spin_quant native \
--metadata '{"get_bos_id":128000, "get_eos_ids":[128009, 128001]}'
ls -lh "${OUT_ET_MODEL_NAME}.pte"
elif [[ ${{ matrix.config }} == "llama3_qlora" ]]; then
# QAT + LoRA
# Download prequantized chceckpoint from Hugging Face
DOWNLOADED_PATH=$(
bash .ci/scripts/download_hf_hub.sh \
--model_id "${HF_MODEL_REPO}" \
--files "tokenizer.model" "params.json" "consolidated.00.pth"
)
# Export using ExecuTorch's model definition
python -m examples.models.llama.export_llama \
--model "llama3_2" \
--checkpoint "${DOWNLOADED_PATH}/consolidated.00.pth" \
--params "${DOWNLOADED_PATH}/params.json" \
-qat \
-lora 16 \
--preq_mode 8da4w_output_8da8w \
--preq_group_size 32 \
--preq_embedding_quantize 8,0 \
--use_sdpa_with_kv_cache \
-kv \
-X \
--xnnpack-extended-ops \
-d fp32 \
--max_seq_length 2048 \
--output_name "${OUT_ET_MODEL_NAME}.pte" \
--metadata '{"get_bos_id":128000, "get_eos_ids":[128009, 128001]}'
ls -lh "${OUT_ET_MODEL_NAME}.pte"
elif [[ ${{ matrix.config }} == "llama3_fb16" ]]; then
# Original BF16 version, without any quantization
DOWNLOADED_PATH=$(bash .ci/scripts/download_hf_hub.sh --model_id "${HF_MODEL_REPO}" --subdir "original" --files "tokenizer.model" "params.json" "consolidated.00.pth")
python -m examples.models.llama.export_llama \
--model "llama3_2" \
--checkpoint "${DOWNLOADED_PATH}/consolidated.00.pth" \
--params "${DOWNLOADED_PATH}/params.json" \
-kv \
--use_sdpa_with_kv_cache \
-X \
-d bf16 \
--metadata '{"get_bos_id":128000, "get_eos_ids":[128009, 128001]}' \
--output_name="${OUT_ET_MODEL_NAME}.pte"
ls -lh "${OUT_ET_MODEL_NAME}.pte"
elif [[ ${{ matrix.config }} == "llama3_qnn_htp" ]]; then
export QNN_SDK_ROOT=/tmp/qnn/2.25.0.240728
export LD_LIBRARY_PATH=$QNN_SDK_ROOT/lib/x86_64-linux-clang/
export PYTHONPATH=$(pwd)/..
DOWNLOADED_PATH=$(bash .ci/scripts/download_hf_hub.sh --model_id "${HF_MODEL_REPO}" --subdir "original" --files "tokenizer.model" "params.json" "consolidated.00.pth")
python -m examples.qualcomm.oss_scripts.llama3_2.llama -- \
--checkpoint "${DOWNLOADED_PATH}/consolidated.00.pth" \
--params "${DOWNLOADED_PATH}/params.json" \
--tokenizer_model "${DOWNLOADED_PATH}/tokenizer.model" \
--compile_only \
--ptq 16a4w \
-m SM8650 \
--model_size 1B \
--model_mode kv \
--prompt "Once"
OUT_ET_MODEL_NAME="llama3_2_qnn" # Qualcomm hard-coded it in their script
find . -name "${OUT_ET_MODEL_NAME}.pte" -not -path "./${OUT_ET_MODEL_NAME}.pte" -exec mv {} ./ \;
ls -lh "${OUT_ET_MODEL_NAME}.pte"
else
# By default, test with the Hugging Face model and the xnnpack recipe
DOWNLOADED_PATH=$(bash .ci/scripts/download_hf_hub.sh --model_id "${HF_MODEL_REPO}" --subdir "original" --files "tokenizer.model")
python -m extension.export_util.export_hf_model -hfm="$HF_MODEL_REPO" -o "$OUT_ET_MODEL_NAME"
ls -lh "${OUT_ET_MODEL_NAME}.pte"
fi
else
echo "Unsupported model ${{ matrix.model }}"
exit 1
fi
zip -j model.zip "${OUT_ET_MODEL_NAME}.pte" "${DOWNLOADED_PATH}/tokenizer.model"
ls -lh model.zip
mkdir -p "${ARTIFACTS_DIR_NAME}"
mv model.zip "${ARTIFACTS_DIR_NAME}"
elif [[ ${{ matrix.model }} == "llama" ]]; then
# Install requirements for export_llama
PYTHON_EXECUTABLE=python bash examples/models/llama/install_requirements.sh
# Test llama2
if [[ ${{ matrix.config }} == *"xnnpack"* ]]; then
DELEGATE_CONFIG="xnnpack+custom+qe"
elif [[ ${{ matrix.config }} == *"qnn"* ]]; then
DELEGATE_CONFIG="qnn"
else
echo "Unsupported delegate ${{ matrix.config }}"
exit 1
fi
DTYPE="fp32"
PYTHON_EXECUTABLE=python bash .ci/scripts/test_llama.sh \
-model "${{ matrix.model }}" \
-build_tool "${BUILD_MODE}" \
-dtype "${DTYPE}" \
-mode "${DELEGATE_CONFIG}" \
-upload "${ARTIFACTS_DIR_NAME}"
else
PYTHON_EXECUTABLE=python bash .ci/scripts/test_model.sh \
"${{ matrix.model }}" \
"${BUILD_MODE}" \
"${{ matrix.config }}" \
"${ARTIFACTS_DIR_NAME}"
fi
echo "::endgroup::"
build-benchmark-app:
name: build-benchmark-app
uses: pytorch/test-infra/.github/workflows/linux_job.yml@main
needs: set-parameters
with:
runner: linux.2xlarge
docker-image: executorch-ubuntu-22.04-clang12-android
submodules: 'true'
ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
timeout: 90
upload-artifact: android-apps
upload-artifact-to-s3: true
script: |
set -eux
# The generic Linux job chooses to use base env, not the one setup by the image
CONDA_ENV=$(conda env list --json | jq -r ".envs | .[-1]")
conda activate "${CONDA_ENV}"
PYTHON_EXECUTABLE=python bash .ci/scripts/setup-linux.sh cmake
export ARTIFACTS_DIR_NAME=artifacts-to-be-uploaded
PYTHON_EXECUTABLE=python bash .ci/scripts/setup-qnn-deps.sh
PYTHON_EXECUTABLE=python bash .ci/scripts/build-qnn-sdk.sh
export ANDROID_ABIS="arm64-v8a"
PYTHON_EXECUTABLE=python EXECUTORCH_BUILD_QNN=ON QNN_SDK_ROOT=/tmp/qnn/2.25.0.240728 bash build/build_android_llm_demo.sh ${ARTIFACTS_DIR_NAME}
# Let's see how expensive this job is, we might want to tone it down by running it periodically
benchmark-on-device:
if: always()
permissions:
id-token: write
contents: read
uses: pytorch/test-infra/.github/workflows/mobile_job.yml@main
needs:
- set-parameters
- prepare-test-specs
- build-benchmark-app
- export-models
strategy:
matrix: ${{ fromJson(needs.set-parameters.outputs.benchmark_configs) }}
fail-fast: false
with:
# Due to scheduling a job may be pushed beyond the default 60m threshold
timeout: 120
device-type: android
runner: linux.2xlarge
test-infra-ref: ''
# This is the ARN of ExecuTorch project on AWS
project-arn: arn:aws:devicefarm:us-west-2:308535385114:project:02a2cf0f-6d9b-45ee-ba1a-a086587469e6
device-pool-arn: ${{ matrix.device_arn }}
android-app-archive: https://gha-artifacts.s3.amazonaws.com/${{ github.repository }}/${{ github.run_id }}/artifacts/minibench/app-debug.apk
android-test-archive: https://gha-artifacts.s3.amazonaws.com/${{ github.repository }}/${{ github.run_id }}/artifacts/minibench/app-debug-androidTest.apk
test-spec: https://gha-artifacts.s3.amazonaws.com/${{ github.repository }}/${{ github.run_id }}/artifacts/${{ matrix.model }}_${{ matrix.config }}/android-llm-device-farm-test-spec.yml
upload-benchmark-results:
needs:
- benchmark-on-device
if: always()
runs-on: linux.2xlarge
environment: upload-benchmark-results
permissions:
id-token: write
contents: read
steps:
- uses: actions/checkout@v3
with:
submodules: false
- name: Authenticate with AWS
uses: aws-actions/configure-aws-credentials@v4
with:
role-to-assume: arn:aws:iam::308535385114:role/gha_workflow_upload-benchmark-results
# The max duration enforced by the server side
role-duration-seconds: 18000
aws-region: us-east-1
- name: Setup conda
uses: pytorch/test-infra/.github/actions/setup-miniconda@main
with:
python-version: '3.10'
- name: Download the list of artifacts from S3
env:
ARTIFACTS_S3_DIR: s3://gha-artifacts/device_farm/${{ github.run_id }}/${{ github.run_attempt }}/artifacts/
shell: bash
run: |
set -eux
${CONDA_RUN} python -mpip install awscli==1.32.18
mkdir -p artifacts
pushd artifacts
${CONDA_RUN} aws s3 sync "${ARTIFACTS_S3_DIR}" .
popd
ls -lah artifacts
- name: Download the list of benchmark configs from S3
env:
BENCHMARK_CONFIGS_DIR: s3://gha-artifacts/${{ github.repository }}/${{ github.run_id }}/artifacts/benchmark-configs/
shell: bash
run: |
set -eux
mkdir -p benchmark-configs
pushd benchmark-configs
${CONDA_RUN} aws s3 sync "${BENCHMARK_CONFIGS_DIR}" .
popd
ls -lah benchmark-configs
- name: Extract the benchmark results JSON
shell: bash
run: |
set -eux
mkdir -p benchmark-results
for ARTIFACTS_BY_JOB in artifacts/*.json; do
[ -f "${ARTIFACTS_BY_JOB}" ] || break
echo "${ARTIFACTS_BY_JOB}"
${CONDA_RUN} python .github/scripts/extract_benchmark_results.py \
--artifacts "${ARTIFACTS_BY_JOB}" \
--output-dir benchmark-results \
--repo ${{ github.repository }} \
--head-branch ${{ github.head_ref || github.ref_name }} \
--workflow-name "${{ github.workflow }}" \
--workflow-run-id ${{ github.run_id }} \
--workflow-run-attempt ${{ github.run_attempt }} \
--benchmark-configs benchmark-configs
done
for SCHEMA in v2 v3; do
for BENCHMARK_RESULTS in benchmark-results/"${SCHEMA}"/*.json; do
cat "${BENCHMARK_RESULTS}"
echo
done
done
# TODO (huydhn): Remove v2 schema once the benchmark dashboard finishes the migration
- name: Upload the benchmark results (v2)
uses: pytorch/test-infra/.github/actions/upload-benchmark-results@main
with:
benchmark-results-dir: benchmark-results/v2
dry-run: false
schema-version: v2
- name: Upload the benchmark results (v3)
uses: pytorch/test-infra/.github/actions/upload-benchmark-results@main
with:
benchmark-results-dir: benchmark-results/v3
dry-run: false
schema-version: v3
github-token: ${{ secrets.GITHUB_TOKEN }}