mirror of
https://github.com/huggingface/transformers.git
synced 2025-11-19 17:54:41 +08:00
Compare commits
2 Commits
flaky_gene
...
v4.57.0
| Author | SHA1 | Date | |
|---|---|---|---|
| 8ac2b916b0 | |||
| 2ccc6cae21 |
7
.github/ISSUE_TEMPLATE/bug-report.yml
vendored
7
.github/ISSUE_TEMPLATE/bug-report.yml
vendored
@ -48,17 +48,18 @@ body:
|
||||
- continuous batching: @remi-or @ArthurZucker @McPatate
|
||||
- pipelines: @Rocketknight1
|
||||
- tokenizers: @ArthurZucker and @itazap
|
||||
- trainer: @SunMarc
|
||||
- trainer: @zach-huggingface @SunMarc
|
||||
- attention: @vasqu @ArthurZucker @CyrilVallez
|
||||
- model loading (from pretrained, etc): @CyrilVallez
|
||||
- distributed: @3outeille @ArthurZucker
|
||||
- distributed: @3outeille @ArthurZucker @S1ro1
|
||||
- CIs: @ydshieh
|
||||
|
||||
Integrations:
|
||||
|
||||
- deepspeed: HF Trainer/Accelerate: @SunMarc @zach-huggingface
|
||||
- ray/raytune: @richardliaw, @amogkam
|
||||
- Big Model Inference: @SunMarc
|
||||
- quantization: @SunMarc @MekkCyber
|
||||
- quantization (bitsandbytes, autogpt): @SunMarc @MekkCyber
|
||||
- kernels: @MekkCyber @drbh
|
||||
- peft: @BenjaminBossan @githubnemo
|
||||
|
||||
|
||||
7
.github/PULL_REQUEST_TEMPLATE.md
vendored
7
.github/PULL_REQUEST_TEMPLATE.md
vendored
@ -51,17 +51,18 @@ Library:
|
||||
- continuous batching: @remi-or @ArthurZucker @McPatate
|
||||
- pipelines: @Rocketknight1
|
||||
- tokenizers: @ArthurZucker and @itazap
|
||||
- trainer: @SunMarc
|
||||
- trainer: @zach-huggingface @SunMarc
|
||||
- attention: @vasqu @ArthurZucker @CyrilVallez
|
||||
- model loading (from pretrained, etc): @CyrilVallez
|
||||
- distributed: @3outeille @ArthurZucker
|
||||
- distributed: @3outeille @ArthurZucker @S1ro1
|
||||
- CIs: @ydshieh
|
||||
|
||||
Integrations:
|
||||
|
||||
- deepspeed: HF Trainer/Accelerate: @SunMarc @zach-huggingface
|
||||
- ray/raytune: @richardliaw, @amogkam
|
||||
- Big Model Inference: @SunMarc
|
||||
- quantization: @SunMarc @MekkCyber
|
||||
- quantization (bitsandbytes, autogpt): @SunMarc @MekkCyber
|
||||
- kernels: @MekkCyber @drbh
|
||||
- peft: @BenjaminBossan @githubnemo
|
||||
|
||||
|
||||
1
.github/scripts/codeowners_for_review_action
vendored
1
.github/scripts/codeowners_for_review_action
vendored
@ -22,6 +22,7 @@ tests/generation/ @gante
|
||||
/src/transformers/models/auto/ @ArthurZucker
|
||||
/src/transformers/utils/ @ArthurZucker @Rocketknight1
|
||||
/src/transformers/loss/ @ArthurZucker
|
||||
/src/transformers/onnx/ @michaelbenayoun
|
||||
|
||||
# Specific files come after the sections/globs, so they take priority
|
||||
/.circleci/config.yml @ArthurZucker @ydshieh
|
||||
|
||||
31
.github/workflows/benchmark.yml
vendored
31
.github/workflows/benchmark.yml
vendored
@ -12,8 +12,6 @@ concurrency:
|
||||
|
||||
env:
|
||||
HF_HOME: /mnt/cache
|
||||
DATASET_ID: hf-benchmarks/transformers
|
||||
MODEL_ID: meta-llama/Llama-3.1-8B-Instruct
|
||||
|
||||
jobs:
|
||||
benchmark:
|
||||
@ -28,20 +26,35 @@ jobs:
|
||||
(github.event_name == 'pull_request' && contains( github.event.pull_request.labels.*.name, 'run-benchmark') )||
|
||||
(github.event_name == 'push' && github.ref == 'refs/heads/main')
|
||||
container:
|
||||
image: huggingface/transformers-all-latest-gpu
|
||||
image: huggingface/transformers-pytorch-gpu
|
||||
options: --gpus all --privileged --ipc host
|
||||
steps:
|
||||
- name: Get repo
|
||||
uses: actions/checkout@v5
|
||||
uses: actions/checkout@v4
|
||||
with:
|
||||
fetch-depth: 1
|
||||
ref: ${{ github.event.pull_request.head.sha || github.sha }}
|
||||
|
||||
- name: Install libpq-dev & psql
|
||||
run: |
|
||||
apt update
|
||||
apt install -y libpq-dev postgresql-client
|
||||
|
||||
- name: Install benchmark script dependencies
|
||||
run: python3 -m pip install -r benchmark_v2/requirements.txt kernels
|
||||
run: python3 -m pip install -r benchmark/requirements.txt
|
||||
|
||||
- name: Reinstall transformers in edit mode (remove the one installed during docker image build)
|
||||
working-directory: /transformers
|
||||
run: python3 -m pip uninstall -y transformers && python3 -m pip install -e ".[torch]"
|
||||
|
||||
- name: Run database init script
|
||||
run: |
|
||||
psql -f benchmark/utils/init_db.sql
|
||||
env:
|
||||
PGDATABASE: metrics
|
||||
PGHOST: ${{ secrets.TRANSFORMERS_BENCHMARKS_PGHOST }}
|
||||
PGUSER: transformers_benchmarks
|
||||
PGPASSWORD: ${{ secrets.TRANSFORMERS_BENCHMARKS_PGPASSWORD }}
|
||||
|
||||
- name: Run benchmark
|
||||
run: |
|
||||
git config --global --add safe.directory /__w/transformers/transformers
|
||||
@ -51,11 +64,13 @@ jobs:
|
||||
commit_id=$GITHUB_SHA
|
||||
fi
|
||||
commit_msg=$(git show -s --format=%s | cut -c1-70)
|
||||
python3 benchmark_v2/run_benchmarks.py -b 32 -s 128 -n 256 --level 2 --branch-name "$BRANCH_NAME" --commit-id "$commit_id" --commit-message "$commit_msg" --model-id "$MODEL_ID" --log-level INFO --push-result-to-dataset "$DATASET_ID"
|
||||
python3 benchmark/benchmarks_entrypoint.py "huggingface/transformers" "$BRANCH_NAME" "$commit_id" "$commit_msg"
|
||||
env:
|
||||
HF_TOKEN: ${{ secrets.HF_HUB_READ_TOKEN }}
|
||||
PUSH_TO_HUB_TOKEN: ${{ secrets.PUSH_TO_HUB_TOKEN }}
|
||||
# Enable this to see debug logs
|
||||
# HF_HUB_VERBOSITY: debug
|
||||
# TRANSFORMERS_VERBOSITY: debug
|
||||
PGHOST: ${{ secrets.TRANSFORMERS_BENCHMARKS_PGHOST }}
|
||||
PGUSER: transformers_benchmarks
|
||||
PGPASSWORD: ${{ secrets.TRANSFORMERS_BENCHMARKS_PGPASSWORD }}
|
||||
BRANCH_NAME: ${{ github.head_ref || github.ref_name }}
|
||||
|
||||
32
.github/workflows/benchmark_v2.yml
vendored
32
.github/workflows/benchmark_v2.yml
vendored
@ -1,7 +1,35 @@
|
||||
name: Benchmark v2 Framework
|
||||
|
||||
on:
|
||||
workflow_dispatch:
|
||||
workflow_call:
|
||||
inputs:
|
||||
runner:
|
||||
description: 'GH Actions runner group to use'
|
||||
required: true
|
||||
type: string
|
||||
container_image:
|
||||
description: 'Docker image to use'
|
||||
required: true
|
||||
type: string
|
||||
container_options:
|
||||
description: 'Container options to use'
|
||||
required: true
|
||||
type: string
|
||||
commit_sha:
|
||||
description: 'Commit SHA to benchmark'
|
||||
required: false
|
||||
type: string
|
||||
default: ''
|
||||
run_id:
|
||||
description: 'Custom run ID for organizing results (auto-generated if not provided)'
|
||||
required: false
|
||||
type: string
|
||||
default: ''
|
||||
benchmark_repo_id:
|
||||
description: 'HuggingFace Dataset to upload results to (e.g., "org/benchmark-results")'
|
||||
required: false
|
||||
type: string
|
||||
default: ''
|
||||
|
||||
env:
|
||||
HF_HOME: /mnt/cache
|
||||
@ -54,4 +82,4 @@ jobs:
|
||||
--token '${{ secrets.TRANSFORMERS_CI_RESULTS_UPLOAD_TOKEN }}' \
|
||||
--log-level INFO
|
||||
env:
|
||||
HF_TOKEN: ${{ secrets.HF_HUB_READ_TOKEN }}
|
||||
HF_TOKEN: ${{ secrets.HF_HUB_READ_TOKEN }}
|
||||
10
.github/workflows/benchmark_v2_a10_caller.yml
vendored
10
.github/workflows/benchmark_v2_a10_caller.yml
vendored
@ -1,7 +1,11 @@
|
||||
name: Benchmark v2 Scheduled Runner - A10 Single-GPU
|
||||
|
||||
on:
|
||||
workflow_dispatch:
|
||||
schedule:
|
||||
# Run daily at 16:30 UTC
|
||||
- cron: "30 16 * * *"
|
||||
pull_request:
|
||||
types: [ opened, labeled, reopened, synchronize ]
|
||||
|
||||
jobs:
|
||||
benchmark-v2-default:
|
||||
@ -9,9 +13,9 @@ jobs:
|
||||
uses: ./.github/workflows/benchmark_v2.yml
|
||||
with:
|
||||
runner: aws-g5-4xlarge-cache-use1-public-80
|
||||
container_image: huggingface/transformers-all-latest-gpu
|
||||
container_image: huggingface/transformers-pytorch-gpu
|
||||
container_options: --gpus all --privileged --ipc host --shm-size "16gb"
|
||||
commit_sha: ${{ github.sha }}
|
||||
run_id: ${{ github.run_id }}
|
||||
benchmark_repo_id: hf-internal-testing/transformers-daily-benchmarks
|
||||
secrets: inherit
|
||||
secrets: inherit
|
||||
@ -1,7 +1,11 @@
|
||||
name: Benchmark v2 Scheduled Runner - MI325 Single-GPU
|
||||
|
||||
on:
|
||||
workflow_dispatch:
|
||||
schedule:
|
||||
# Run daily at 16:30 UTC
|
||||
- cron: "30 16 * * *"
|
||||
pull_request:
|
||||
types: [ opened, labeled, reopened, synchronize ]
|
||||
|
||||
jobs:
|
||||
benchmark-v2-default:
|
||||
@ -14,4 +18,4 @@ jobs:
|
||||
commit_sha: ${{ github.sha }}
|
||||
run_id: ${{ github.run_id }}
|
||||
benchmark_repo_id: hf-internal-testing/transformers-daily-benchmarks
|
||||
secrets: inherit
|
||||
secrets: inherit
|
||||
184
.github/workflows/build-docker-images.yml
vendored
184
.github/workflows/build-docker-images.yml
vendored
@ -45,59 +45,33 @@ jobs:
|
||||
REF=main
|
||||
push: true
|
||||
tags: huggingface/transformers-all-latest-gpu${{ inputs.image_postfix }}
|
||||
|
||||
- name: Post to Slack
|
||||
if: always()
|
||||
uses: huggingface/hf-workflows/.github/actions/post-slack@main
|
||||
with:
|
||||
slack_channel: ${{ secrets.CI_SLACK_CHANNEL_DOCKER }}
|
||||
title: 🤗 Results of the transformers-all-latest-gpu docker build
|
||||
status: ${{ job.status }}
|
||||
slack_token: ${{ secrets.SLACK_CIFEEDBACK_BOT_TOKEN }}
|
||||
|
||||
flash-attn-ci-image:
|
||||
name: "PyTorch with Flash Attn [dev]"
|
||||
runs-on:
|
||||
group: aws-general-8-plus
|
||||
steps:
|
||||
# Push CI images still need to be re-built daily
|
||||
-
|
||||
name: Set up Docker Buildx
|
||||
uses: docker/setup-buildx-action@v3
|
||||
-
|
||||
name: Check out code
|
||||
uses: actions/checkout@v4
|
||||
-
|
||||
name: Login to DockerHub
|
||||
uses: docker/login-action@v3
|
||||
with:
|
||||
username: ${{ secrets.DOCKERHUB_USERNAME }}
|
||||
password: ${{ secrets.DOCKERHUB_PASSWORD }}
|
||||
-
|
||||
name: Build and push
|
||||
name: Build and push (for Push CI) in a daily basis
|
||||
# This condition allows `schedule` events, or `push` events that trigger this workflow NOT via `workflow_call`.
|
||||
# The later case is useful for manual image building for debugging purpose. Use another tag in this case!
|
||||
if: inputs.image_postfix != '-push-ci'
|
||||
uses: docker/build-push-action@v5
|
||||
with:
|
||||
context: ./docker/transformers-all-latest-gpu
|
||||
build-args: |
|
||||
REF=main
|
||||
PYTORCH=2.8.0
|
||||
TORCHCODEC=0.7.0
|
||||
FLASH_ATTN=yes
|
||||
push: true
|
||||
tags: huggingface/transformers-all-latest-gpu${{ inputs.image_postfix }}:flash-attn
|
||||
tags: huggingface/transformers-all-latest-gpu-push-ci
|
||||
|
||||
- name: Post to Slack
|
||||
if: always()
|
||||
uses: huggingface/hf-workflows/.github/actions/post-slack@main
|
||||
with:
|
||||
slack_channel: ${{ secrets.CI_SLACK_CHANNEL_DOCKER }}
|
||||
title: 🤗 Results of the transformers-all-latest-gpu docker build
|
||||
title: 🤗 Results of the transformers-all-latest-gpu-push-ci docker build
|
||||
status: ${{ job.status }}
|
||||
slack_token: ${{ secrets.SLACK_CIFEEDBACK_BOT_TOKEN }}
|
||||
|
||||
latest-torch-deepspeed-docker:
|
||||
name: "Latest PyTorch + DeepSpeed"
|
||||
runs-on:
|
||||
group: aws-general-8-plus
|
||||
group: aws-g4dn-2xlarge-cache
|
||||
steps:
|
||||
-
|
||||
name: Set up Docker Buildx
|
||||
@ -130,8 +104,51 @@ jobs:
|
||||
status: ${{ job.status }}
|
||||
slack_token: ${{ secrets.SLACK_CIFEEDBACK_BOT_TOKEN }}
|
||||
|
||||
# Can't build 2 images in a single job `latest-torch-deepspeed-docker` (for `nvcr.io/nvidia`)
|
||||
latest-torch-deepspeed-docker-for-push-ci-daily-build:
|
||||
name: "Latest PyTorch + DeepSpeed (Push CI - Daily Build)"
|
||||
runs-on:
|
||||
group: aws-general-8-plus
|
||||
steps:
|
||||
-
|
||||
name: Set up Docker Buildx
|
||||
uses: docker/setup-buildx-action@v3
|
||||
-
|
||||
name: Check out code
|
||||
uses: actions/checkout@v4
|
||||
-
|
||||
name: Login to DockerHub
|
||||
uses: docker/login-action@v3
|
||||
with:
|
||||
username: ${{ secrets.DOCKERHUB_USERNAME }}
|
||||
password: ${{ secrets.DOCKERHUB_PASSWORD }}
|
||||
# Push CI images still need to be re-built daily
|
||||
-
|
||||
name: Build and push (for Push CI) in a daily basis
|
||||
# This condition allows `schedule` events, or `push` events that trigger this workflow NOT via `workflow_call`.
|
||||
# The later case is useful for manual image building for debugging purpose. Use another tag in this case!
|
||||
if: inputs.image_postfix != '-push-ci'
|
||||
uses: docker/build-push-action@v5
|
||||
with:
|
||||
context: ./docker/transformers-pytorch-deepspeed-latest-gpu
|
||||
build-args: |
|
||||
REF=main
|
||||
push: true
|
||||
tags: huggingface/transformers-pytorch-deepspeed-latest-gpu-push-ci
|
||||
|
||||
- name: Post to Slack
|
||||
if: always()
|
||||
uses: huggingface/hf-workflows/.github/actions/post-slack@main
|
||||
with:
|
||||
slack_channel: ${{ secrets.CI_SLACK_CHANNEL_DOCKER }}
|
||||
title: 🤗 Results of the transformers-pytorch-deepspeed-latest-gpu-push-ci docker build
|
||||
status: ${{ job.status }}
|
||||
slack_token: ${{ secrets.SLACK_CIFEEDBACK_BOT_TOKEN }}
|
||||
|
||||
doc-builder:
|
||||
name: "Doc builder"
|
||||
# Push CI doesn't need this image
|
||||
if: inputs.image_postfix != '-push-ci'
|
||||
runs-on:
|
||||
group: aws-general-8-plus
|
||||
steps:
|
||||
@ -164,6 +181,44 @@ jobs:
|
||||
status: ${{ job.status }}
|
||||
slack_token: ${{ secrets.SLACK_CIFEEDBACK_BOT_TOKEN }}
|
||||
|
||||
latest-pytorch:
|
||||
name: "Latest PyTorch [dev]"
|
||||
# Push CI doesn't need this image
|
||||
if: inputs.image_postfix != '-push-ci'
|
||||
runs-on:
|
||||
group: aws-general-8-plus
|
||||
steps:
|
||||
-
|
||||
name: Set up Docker Buildx
|
||||
uses: docker/setup-buildx-action@v3
|
||||
-
|
||||
name: Check out code
|
||||
uses: actions/checkout@v4
|
||||
-
|
||||
name: Login to DockerHub
|
||||
uses: docker/login-action@v3
|
||||
with:
|
||||
username: ${{ secrets.DOCKERHUB_USERNAME }}
|
||||
password: ${{ secrets.DOCKERHUB_PASSWORD }}
|
||||
-
|
||||
name: Build and push
|
||||
uses: docker/build-push-action@v5
|
||||
with:
|
||||
context: ./docker/transformers-pytorch-gpu
|
||||
build-args: |
|
||||
REF=main
|
||||
push: true
|
||||
tags: huggingface/transformers-pytorch-gpu
|
||||
|
||||
- name: Post to Slack
|
||||
if: always()
|
||||
uses: huggingface/hf-workflows/.github/actions/post-slack@main
|
||||
with:
|
||||
slack_channel: ${{ secrets.CI_SLACK_CHANNEL_DOCKER }}
|
||||
title: 🤗 Results of the huggingface/transformers-pytorch-gpudocker build
|
||||
status: ${{ job.status }}
|
||||
slack_token: ${{ secrets.SLACK_CIFEEDBACK_BOT_TOKEN }}
|
||||
|
||||
latest-pytorch-amd:
|
||||
name: "Latest PyTorch (AMD) [dev]"
|
||||
runs-on:
|
||||
@ -190,47 +245,29 @@ jobs:
|
||||
REF=main
|
||||
push: true
|
||||
tags: huggingface/transformers-pytorch-amd-gpu${{ inputs.image_postfix }}
|
||||
# Push CI images still need to be re-built daily
|
||||
-
|
||||
name: Build and push (for Push CI) in a daily basis
|
||||
# This condition allows `schedule` events, or `push` events that trigger this workflow NOT via `workflow_call`.
|
||||
# The later case is useful for manual image building for debugging purpose. Use another tag in this case!
|
||||
if: inputs.image_postfix != '-push-ci'
|
||||
uses: docker/build-push-action@v5
|
||||
with:
|
||||
context: ./docker/transformers-pytorch-amd-gpu
|
||||
build-args: |
|
||||
REF=main
|
||||
push: true
|
||||
tags: huggingface/transformers-pytorch-amd-gpu-push-ci
|
||||
|
||||
- name: Post to Slack
|
||||
if: always()
|
||||
uses: huggingface/hf-workflows/.github/actions/post-slack@main
|
||||
with:
|
||||
slack_channel: ${{ secrets.CI_SLACK_CHANNEL_DOCKER }}
|
||||
title: 🤗 Results of the huggingface/transformers-pytorch-amd-gpu build
|
||||
title: 🤗 Results of the huggingface/transformers-pytorch-amd-gpu-push-ci build
|
||||
status: ${{ job.status }}
|
||||
slack_token: ${{ secrets.SLACK_CIFEEDBACK_BOT_TOKEN }}
|
||||
|
||||
cache-latest-pytorch-amd:
|
||||
name: "Cache Latest Pytorch (AMD) Image"
|
||||
needs: latest-pytorch-amd
|
||||
runs-on:
|
||||
group: amd-mi325-1gpu
|
||||
steps:
|
||||
-
|
||||
name: Login to DockerHub
|
||||
uses: docker/login-action@v3
|
||||
with:
|
||||
username: ${{ secrets.DOCKERHUB_USERNAME }}
|
||||
password: ${{ secrets.DOCKERHUB_PASSWORD }}
|
||||
|
||||
-
|
||||
name: Pull and save docker image to cache
|
||||
run: |
|
||||
image="huggingface/transformers-pytorch-amd-gpu"
|
||||
final_path="/mnt/image-cache/transformers-pytorch-amd-gpu.tar"
|
||||
tmp_path="${final_path}.tmp"
|
||||
|
||||
echo "Pulling image: ${image}"
|
||||
docker pull "${image}"
|
||||
|
||||
echo "Saving to temp file: ${tmp_path}"
|
||||
docker save "${image}" -o "${tmp_path}"
|
||||
|
||||
echo "Moving to final path: ${final_path}"
|
||||
mv -f "${tmp_path}" "${final_path}"
|
||||
|
||||
echo "Cache populated successfully at ${final_path}"
|
||||
|
||||
latest-pytorch-deepspeed-amd:
|
||||
name: "PyTorch + DeepSpeed (AMD) [dev]"
|
||||
runs-on:
|
||||
@ -257,6 +294,19 @@ jobs:
|
||||
REF=main
|
||||
push: true
|
||||
tags: huggingface/transformers-pytorch-deepspeed-amd-gpu${{ inputs.image_postfix }}
|
||||
# Push CI images still need to be re-built daily
|
||||
-
|
||||
name: Build and push (for Push CI) in a daily basis
|
||||
# This condition allows `schedule` events, or `push` events that trigger this workflow NOT via `workflow_call`.
|
||||
# The later case is useful for manual image building for debugging purpose. Use another tag in this case!
|
||||
if: inputs.image_postfix != '-push-ci'
|
||||
uses: docker/build-push-action@v5
|
||||
with:
|
||||
context: ./docker/transformers-pytorch-deepspeed-amd-gpu
|
||||
build-args: |
|
||||
REF=main
|
||||
push: true
|
||||
tags: huggingface/transformers-pytorch-deepspeed-amd-gpu-push-ci
|
||||
|
||||
- name: Post to Slack
|
||||
if: always()
|
||||
@ -269,6 +319,8 @@ jobs:
|
||||
|
||||
latest-quantization-torch-docker:
|
||||
name: "Latest Pytorch + Quantization [dev]"
|
||||
# Push CI doesn't need this image
|
||||
if: inputs.image_postfix != '-push-ci'
|
||||
runs-on:
|
||||
group: aws-general-8-plus
|
||||
steps:
|
||||
|
||||
23
.github/workflows/check-workflow-permissions.yml
vendored
23
.github/workflows/check-workflow-permissions.yml
vendored
@ -1,23 +0,0 @@
|
||||
---
|
||||
name: Check Permissions Advisor
|
||||
|
||||
on:
|
||||
workflow_dispatch:
|
||||
inputs:
|
||||
workflow_name:
|
||||
description: 'Workflow file name'
|
||||
type: string
|
||||
run_count:
|
||||
description: 'Number of runs to analyze'
|
||||
type: string
|
||||
default: "10"
|
||||
|
||||
jobs:
|
||||
advisor:
|
||||
uses: huggingface/security-workflows/.github/workflows/permissions-advisor-reusable.yml@main
|
||||
permissions:
|
||||
actions: read
|
||||
contents: read
|
||||
with:
|
||||
workflow_name: ${{ inputs.workflow_name }}
|
||||
run_count: ${{ fromJSON(inputs.run_count) }}
|
||||
205
.github/workflows/check_failed_tests.yml
vendored
205
.github/workflows/check_failed_tests.yml
vendored
@ -6,6 +6,9 @@ on:
|
||||
docker:
|
||||
required: true
|
||||
type: string
|
||||
start_sha:
|
||||
required: true
|
||||
type: string
|
||||
job:
|
||||
required: true
|
||||
type: string
|
||||
@ -21,13 +24,7 @@ on:
|
||||
commit_sha:
|
||||
required: false
|
||||
type: string
|
||||
pr_number:
|
||||
required: false
|
||||
type: string
|
||||
outputs:
|
||||
report:
|
||||
description: "Content of the report of new failures"
|
||||
value: ${{ jobs.process_new_failures_with_commit_info.outputs.report }}
|
||||
|
||||
|
||||
env:
|
||||
HF_HOME: /mnt/cache
|
||||
@ -38,20 +35,16 @@ env:
|
||||
# For gated repositories, we still need to agree to share information on the Hub repo. page in order to get access.
|
||||
# This token is created under the bot `hf-transformers-bot`.
|
||||
HF_HUB_READ_TOKEN: ${{ secrets.HF_HUB_READ_TOKEN }}
|
||||
SIGOPT_API_TOKEN: ${{ secrets.SIGOPT_API_TOKEN }}
|
||||
TF_FORCE_GPU_ALLOW_GROWTH: true
|
||||
CUDA_VISIBLE_DEVICES: 0,1
|
||||
|
||||
|
||||
jobs:
|
||||
check_new_failures:
|
||||
name: "Find commits for new failing tests"
|
||||
strategy:
|
||||
matrix:
|
||||
run_idx: [1]
|
||||
name: " "
|
||||
runs-on:
|
||||
group: aws-g5-4xlarge-cache
|
||||
outputs:
|
||||
process: ${{ steps.check_file.outputs.process }}
|
||||
container:
|
||||
image: ${{ inputs.docker }}
|
||||
options: --gpus all --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
|
||||
@ -62,19 +55,14 @@ jobs:
|
||||
path: /transformers/ci_results_${{ inputs.job }}
|
||||
|
||||
- name: Check file
|
||||
id: check_file
|
||||
working-directory: /transformers
|
||||
env:
|
||||
job: ${{ inputs.job }}
|
||||
run: |
|
||||
if [ -f "ci_results_${job}/new_failures.json" ]; then
|
||||
echo "\`ci_results_${job}/new_failures.json\` exists, continue ..."
|
||||
if [ -f ci_results_${{ inputs.job }}/new_failures.json ]; then
|
||||
echo "`ci_results_${{ inputs.job }}/new_failures.json` exists, continue ..."
|
||||
echo "process=true" >> $GITHUB_ENV
|
||||
echo "process=true" >> $GITHUB_OUTPUT
|
||||
else
|
||||
echo "\`ci_results_${job}/new_failures.json\` doesn't exist, abort."
|
||||
echo "`ci_results_${{ inputs.job }}/new_failures.json` doesn't exist, abort."
|
||||
echo "process=false" >> $GITHUB_ENV
|
||||
echo "process=false" >> $GITHUB_OUTPUT
|
||||
fi
|
||||
|
||||
- uses: actions/download-artifact@v4
|
||||
@ -93,62 +81,27 @@ jobs:
|
||||
echo "PREV_WORKFLOW_RUN_ID=" >> $GITHUB_ENV
|
||||
fi
|
||||
|
||||
if [ -f setup_values/other_workflow_run_id.txt ]; then
|
||||
echo "OTHER_WORKFLOW_RUN_ID=$(cat setup_values/other_workflow_run_id.txt)" >> $GITHUB_ENV
|
||||
else
|
||||
echo "OTHER_WORKFLOW_RUN_ID=" >> $GITHUB_ENV
|
||||
fi
|
||||
|
||||
- name: Update clone
|
||||
working-directory: /transformers
|
||||
if: ${{ env.process == 'true' }}
|
||||
env:
|
||||
commit_sha: ${{ inputs.commit_sha || github.sha }}
|
||||
run: |
|
||||
git fetch origin "$commit_sha" && git checkout "$commit_sha"
|
||||
run: git fetch && git checkout ${{ inputs.commit_sha || github.sha }}
|
||||
|
||||
- name: Get `START_SHA`
|
||||
- name: Get target commit
|
||||
working-directory: /transformers/utils
|
||||
if: ${{ env.process == 'true' }}
|
||||
env:
|
||||
commit_sha: ${{ inputs.commit_sha || github.sha }}
|
||||
run: |
|
||||
echo "START_SHA=$commit_sha" >> $GITHUB_ENV
|
||||
echo "END_SHA=$(TOKEN=${{ secrets.ACCESS_REPO_INFO_TOKEN }} python3 -c 'import os; from get_previous_daily_ci import get_last_daily_ci_run_commit; commit=get_last_daily_ci_run_commit(token=os.environ["TOKEN"], workflow_run_id=os.environ["PREV_WORKFLOW_RUN_ID"]); print(commit)')" >> $GITHUB_ENV
|
||||
|
||||
# This is used if the CI is triggered from a pull request `self-comment-ci.yml` (after security check is verified)
|
||||
- name: Extract the base commit on `main` (of the merge commit created by Github) if it is a PR
|
||||
id: pr_info
|
||||
if: ${{ env.process == 'true' && inputs.pr_number != '' }}
|
||||
uses: actions/github-script@v6
|
||||
with:
|
||||
script: |
|
||||
const { data: pr } = await github.rest.pulls.get({
|
||||
owner: context.repo.owner,
|
||||
repo: context.repo.repo,
|
||||
pull_number: ${{ inputs.pr_number }}
|
||||
});
|
||||
|
||||
const { data: merge_commit } = await github.rest.repos.getCommit({
|
||||
owner: pr.base.repo.owner.login,
|
||||
repo: pr.base.repo.name,
|
||||
ref: '${{ inputs.commit_sha }}',
|
||||
});
|
||||
|
||||
core.setOutput('merge_commit_base_sha', merge_commit.parents[0].sha);
|
||||
|
||||
# Usually, `END_SHA` should be the commit of the last previous workflow run of the **SAME** (scheduled) workflow.
|
||||
# (This is why we don't need to specify `workflow_id` which would be fetched automatically in the python script.)
|
||||
- name: Get `END_SHA` from previous CI runs of the same workflow
|
||||
working-directory: /transformers/utils
|
||||
if: ${{ env.process == 'true' && inputs.pr_number == '' }}
|
||||
env:
|
||||
ACCESS_TOKEN: ${{ secrets.ACCESS_REPO_INFO_TOKEN }}
|
||||
run: |
|
||||
echo "END_SHA=$(TOKEN="$ACCESS_TOKEN" python3 -c 'import os; from get_previous_daily_ci import get_last_daily_ci_run_commit; commit=get_last_daily_ci_run_commit(token=os.environ["TOKEN"], workflow_run_id=os.environ["PREV_WORKFLOW_RUN_ID"]); print(commit)')" >> $GITHUB_ENV
|
||||
|
||||
# However, for workflow runs triggered by `issue_comment` (for pull requests), we want to check against the
|
||||
# parent commit (on `main`) of the `merge_commit` (dynamically created by GitHub). In this case, the goal is to
|
||||
# see if a reported failing test is actually ONLY failing on the `merge_commit`.
|
||||
- name: Set `END_SHA`
|
||||
if: ${{ env.process == 'true' && inputs.pr_number != '' }}
|
||||
env:
|
||||
merge_commit_base_sha: ${{ steps.pr_info.outputs.merge_commit_base_sha }}
|
||||
run: |
|
||||
echo "END_SHA=$merge_commit_base_sha" >> $GITHUB_ENV
|
||||
- name: Checkout to `start_sha`
|
||||
working-directory: /transformers
|
||||
if: ${{ env.process == 'true' }}
|
||||
run: git fetch && git checkout ${{ inputs.start_sha }}
|
||||
|
||||
- name: Reinstall transformers in edit mode (remove the one installed during docker image build)
|
||||
working-directory: /transformers
|
||||
@ -166,10 +119,6 @@ jobs:
|
||||
run: |
|
||||
python3 utils/print_env.py
|
||||
|
||||
- name: Install pytest-flakefinder
|
||||
if: ${{ env.process == 'true' }}
|
||||
run: python3 -m pip install pytest-flakefinder
|
||||
|
||||
- name: Show installed libraries and their versions
|
||||
working-directory: /transformers
|
||||
if: ${{ env.process == 'true' }}
|
||||
@ -178,78 +127,37 @@ jobs:
|
||||
- name: Check failed tests
|
||||
working-directory: /transformers
|
||||
if: ${{ env.process == 'true' }}
|
||||
env:
|
||||
job: ${{ inputs.job }}
|
||||
run_idx: ${{ matrix.run_idx }}
|
||||
run: python3 utils/check_bad_commit.py --start_commit "$START_SHA" --end_commit "$END_SHA" --file "ci_results_${job}/new_failures.json" --output_file "new_failures_with_bad_commit_${job}_${run_idx}.json"
|
||||
run: python3 utils/check_bad_commit.py --start_commit ${{ inputs.start_sha }} --end_commit ${{ env.END_SHA }} --file ci_results_${{ inputs.job }}/new_failures.json --output_file new_failures_with_bad_commit.json
|
||||
|
||||
- name: Show results
|
||||
working-directory: /transformers
|
||||
if: ${{ env.process == 'true' }}
|
||||
env:
|
||||
job: ${{ inputs.job }}
|
||||
run_idx: ${{ matrix.run_idx }}
|
||||
run: |
|
||||
ls -l "new_failures_with_bad_commit_${job}_${run_idx}.json"
|
||||
cat "new_failures_with_bad_commit_${job}_${run_idx}.json"
|
||||
ls -l new_failures_with_bad_commit.json
|
||||
cat new_failures_with_bad_commit.json
|
||||
|
||||
- name: Upload artifacts
|
||||
uses: actions/upload-artifact@v4
|
||||
with:
|
||||
name: new_failures_with_bad_commit_${{ inputs.job }}_${{ matrix.run_idx }}
|
||||
path: /transformers/new_failures_with_bad_commit_${{ inputs.job }}_${{ matrix.run_idx }}.json
|
||||
|
||||
process_new_failures_with_commit_info:
|
||||
name: "process bad commit reports"
|
||||
needs: check_new_failures
|
||||
if: needs.check_new_failures.outputs.process == 'true'
|
||||
runs-on:
|
||||
group: aws-g5-4xlarge-cache
|
||||
outputs:
|
||||
report: ${{ steps.set_output.outputs.report }}
|
||||
container:
|
||||
image: ${{ inputs.docker }}
|
||||
options: --gpus all --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
|
||||
steps:
|
||||
- uses: actions/download-artifact@v4
|
||||
with:
|
||||
name: ci_results_${{ inputs.job }}
|
||||
path: /transformers/ci_results_${{ inputs.job }}
|
||||
|
||||
- uses: actions/download-artifact@v4
|
||||
with:
|
||||
pattern: new_failures_with_bad_commit_${{ inputs.job }}*
|
||||
path: /transformers/new_failures_with_bad_commit_${{ inputs.job }}
|
||||
merge-multiple: true
|
||||
|
||||
- name: Check files
|
||||
- name: Checkout back
|
||||
working-directory: /transformers
|
||||
env:
|
||||
job: ${{ inputs.job }}
|
||||
if: ${{ env.process == 'true' }}
|
||||
run: |
|
||||
ls -la /transformers
|
||||
ls -la "/transformers/new_failures_with_bad_commit_${job}"
|
||||
|
||||
# Currently, we only run with a single runner by using `run_idx: [1]`. We might try to run with multiple runners
|
||||
# to further reduce the false positive caused by flaky tests, which requires further processing to merge reports.
|
||||
- name: Merge files
|
||||
shell: bash
|
||||
working-directory: /transformers
|
||||
env:
|
||||
job: ${{ inputs.job }}
|
||||
run: |
|
||||
cp "/transformers/new_failures_with_bad_commit_${job}/new_failures_with_bad_commit_${job}_1.json" new_failures_with_bad_commit.json
|
||||
|
||||
- name: Update clone
|
||||
working-directory: /transformers
|
||||
env:
|
||||
commit_sha: ${{ inputs.commit_sha || github.sha }}
|
||||
run: |
|
||||
git fetch origin "$commit_sha" && git checkout "$commit_sha"
|
||||
git checkout ${{ inputs.start_sha }}
|
||||
|
||||
- name: Process report
|
||||
shell: bash
|
||||
working-directory: /transformers
|
||||
if: ${{ env.process == 'true' }}
|
||||
env:
|
||||
ACCESS_REPO_INFO_TOKEN: ${{ secrets.ACCESS_REPO_INFO_TOKEN }}
|
||||
TRANSFORMERS_CI_RESULTS_UPLOAD_TOKEN: ${{ secrets.TRANSFORMERS_CI_RESULTS_UPLOAD_TOKEN }}
|
||||
JOB_NAME: ${{ inputs.job }}
|
||||
REPORT_REPO_ID: ${{ inputs.report_repo_id }}
|
||||
run: |
|
||||
python3 utils/process_bad_commit_report.py
|
||||
|
||||
- name: Process report
|
||||
shell: bash
|
||||
working-directory: /transformers
|
||||
if: ${{ env.process == 'true' }}
|
||||
env:
|
||||
ACCESS_REPO_INFO_TOKEN: ${{ secrets.ACCESS_REPO_INFO_TOKEN }}
|
||||
TRANSFORMERS_CI_RESULTS_UPLOAD_TOKEN: ${{ secrets.TRANSFORMERS_CI_RESULTS_UPLOAD_TOKEN }}
|
||||
@ -262,40 +170,15 @@ jobs:
|
||||
echo EOF
|
||||
} >> "$GITHUB_ENV"
|
||||
|
||||
# The output is useful if a caller needs more processing, for example, we have a chain
|
||||
# self-comment-ci.yml -> self-scheduled.yml -> this one (check_failed_tests.yml),
|
||||
# and `self-comment-ci.yml` needs further processing before sending a GitHub comment to the pull request page.
|
||||
- name: Show results & Set outputs
|
||||
id: set_output
|
||||
working-directory: /transformers
|
||||
run: |
|
||||
ls -l new_failures_with_bad_commit.json
|
||||
cat new_failures_with_bad_commit.json
|
||||
|
||||
{
|
||||
echo 'report<<EOF'
|
||||
cat new_failures_with_bad_commit.json
|
||||
echo '' # Force a newline
|
||||
echo EOF
|
||||
} >> "$GITHUB_OUTPUT"
|
||||
|
||||
- name: Upload artifacts
|
||||
uses: actions/upload-artifact@v4
|
||||
with:
|
||||
name: new_failures_with_bad_commit_${{ inputs.job }}
|
||||
path: /transformers/new_failures_with_bad_commit.json
|
||||
|
||||
- name: Prepare Slack report title
|
||||
working-directory: /transformers
|
||||
env:
|
||||
ci_event: ${{ inputs.ci_event }}
|
||||
job: ${{ inputs.job }}
|
||||
if: ${{ env.process == 'true' }}
|
||||
run: |
|
||||
pip install slack_sdk
|
||||
echo "title=$(python3 -c 'import sys; import os; sys.path.append("utils"); from utils.notification_service import job_to_test_map; ci_event = os.environ["ci_event"]; job = os.environ["job"]; test_name = job_to_test_map[job]; title = f"New failed tests of {ci_event}" + ":" + f" {test_name}"; print(title)')" >> $GITHUB_ENV
|
||||
echo "title=$(python3 -c 'import sys; sys.path.append("utils"); from utils.notification_service import job_to_test_map; ci_event = "${{ inputs.ci_event }}"; job = "${{ inputs.job }}"; test_name = job_to_test_map[job]; title = f"New failed tests of {ci_event}" + ":" + f" {test_name}"; print(title)')" >> $GITHUB_ENV
|
||||
|
||||
- name: Send processed report
|
||||
if: ${{ !endsWith(env.REPORT_TEXT, '{}') }}
|
||||
if: ${{ env.process == 'true' && !endsWith(env.REPORT_TEXT, '{}') }}
|
||||
uses: slackapi/slack-github-action@6c661ce58804a1a20f6dc5fbee7f0381b469e001
|
||||
with:
|
||||
# Slack channel id, channel name, or user id to post message.
|
||||
|
||||
22
.github/workflows/codeql.yml
vendored
22
.github/workflows/codeql.yml
vendored
@ -1,22 +0,0 @@
|
||||
---
|
||||
name: CodeQL Security Analysis
|
||||
|
||||
on:
|
||||
push:
|
||||
branches: ["main", "fix_security_issue_*"]
|
||||
# pull_request:
|
||||
# branches: ["main"]
|
||||
workflow_dispatch:
|
||||
|
||||
jobs:
|
||||
codeql:
|
||||
name: CodeQL Analysis
|
||||
uses: huggingface/security-workflows/.github/workflows/codeql-reusable.yml@main
|
||||
permissions:
|
||||
security-events: write
|
||||
packages: read
|
||||
actions: read
|
||||
contents: read
|
||||
with:
|
||||
languages: '["actions"]'
|
||||
queries: 'security-extended,security-and-quality'
|
||||
1
.github/workflows/doctest_job.yml
vendored
1
.github/workflows/doctest_job.yml
vendored
@ -16,6 +16,7 @@ env:
|
||||
RUN_SLOW: yes
|
||||
OMP_NUM_THREADS: 16
|
||||
MKL_NUM_THREADS: 16
|
||||
SIGOPT_API_TOKEN: ${{ secrets.SIGOPT_API_TOKEN }}
|
||||
TF_FORCE_GPU_ALLOW_GROWTH: true
|
||||
|
||||
jobs:
|
||||
|
||||
24
.github/workflows/get-pr-info.yml
vendored
24
.github/workflows/get-pr-info.yml
vendored
@ -39,9 +39,6 @@ on:
|
||||
PR_MERGE_COMMIT_SHA:
|
||||
description: "The sha of the merge commit for the pull request (created by GitHub) in the base repository"
|
||||
value: ${{ jobs.get-pr-info.outputs.PR_MERGE_COMMIT_SHA }}
|
||||
PR_MERGE_COMMIT_BASE_SHA:
|
||||
description: "The sha of the parent commit of the the merge commit on the target branch in the base repository"
|
||||
value: ${{ jobs.get-pr-info.outputs.PR_MERGE_COMMIT_BASE_SHA }}
|
||||
PR_HEAD_COMMIT_DATE:
|
||||
description: "The date of the head sha of the pull request branch in the head repository"
|
||||
value: ${{ jobs.get-pr-info.outputs.PR_HEAD_COMMIT_DATE }}
|
||||
@ -77,7 +74,6 @@ jobs:
|
||||
PR_BASE_REF: ${{ steps.pr_info.outputs.base_ref }}
|
||||
PR_HEAD_SHA: ${{ steps.pr_info.outputs.head_sha }}
|
||||
PR_BASE_SHA: ${{ steps.pr_info.outputs.base_sha }}
|
||||
PR_MERGE_COMMIT_BASE_SHA: ${{ steps.pr_info.outputs.merge_commit_base_sha }}
|
||||
PR_MERGE_COMMIT_SHA: ${{ steps.pr_info.outputs.merge_commit_sha }}
|
||||
PR_HEAD_COMMIT_DATE: ${{ steps.pr_info.outputs.head_commit_date }}
|
||||
PR_MERGE_COMMIT_DATE: ${{ steps.pr_info.outputs.merge_commit_date }}
|
||||
@ -126,7 +122,6 @@ jobs:
|
||||
core.setOutput('base_ref', pr.base.ref);
|
||||
core.setOutput('head_sha', pr.head.sha);
|
||||
core.setOutput('base_sha', pr.base.sha);
|
||||
core.setOutput('merge_commit_base_sha', merge_commit.parents[0].sha);
|
||||
core.setOutput('merge_commit_sha', pr.merge_commit_sha);
|
||||
core.setOutput('pr', pr);
|
||||
|
||||
@ -147,21 +142,16 @@ jobs:
|
||||
date: merge_commit.commit.committer.date
|
||||
});
|
||||
|
||||
console.log('PR Info:', {
|
||||
pr_info: pr
|
||||
});
|
||||
|
||||
- name: Convert dates to timestamps
|
||||
id: get_timestamps
|
||||
env:
|
||||
head_commit_date: ${{ steps.pr_info.outputs.head_commit_date }}
|
||||
merge_commit_date: ${{ steps.pr_info.outputs.merge_commit_date }}
|
||||
run: |
|
||||
echo "$head_commit_date"
|
||||
echo "$merge_commit_date"
|
||||
head_commit_date=${{ steps.pr_info.outputs.head_commit_date }}
|
||||
merge_commit_date=${{ steps.pr_info.outputs.merge_commit_date }}
|
||||
echo $head_commit_date
|
||||
echo $merge_commit_date
|
||||
head_commit_timestamp=$(date -d "$head_commit_date" +%s)
|
||||
merge_commit_timestamp=$(date -d "$merge_commit_date" +%s)
|
||||
echo "$head_commit_timestamp"
|
||||
echo "$merge_commit_timestamp"
|
||||
echo $head_commit_timestamp
|
||||
echo $merge_commit_timestamp
|
||||
echo "head_commit_timestamp=$head_commit_timestamp" >> $GITHUB_OUTPUT
|
||||
echo "merge_commit_timestamp=$merge_commit_timestamp" >> $GITHUB_OUTPUT
|
||||
echo "merge_commit_timestamp=$merge_commit_timestamp" >> $GITHUB_OUTPUT
|
||||
|
||||
22
.github/workflows/get-pr-number.yml
vendored
22
.github/workflows/get-pr-number.yml
vendored
@ -15,19 +15,13 @@ jobs:
|
||||
steps:
|
||||
- name: Get PR number
|
||||
shell: bash
|
||||
env:
|
||||
issue_number: ${{ github.event.issue.number }}
|
||||
is_pull_request_issue: ${{ github.event.issue.pull_request != null }}
|
||||
pr_number: ${{ github.event.pull_request.number }}
|
||||
is_pull_request: ${{ github.event.pull_request != null }}
|
||||
event_number: ${{ github.event.number }}
|
||||
run: |
|
||||
if [[ "$issue_number" != "" && "$is_pull_request_issue" == "true" ]]; then
|
||||
echo "PR_NUMBER=$issue_number" >> $GITHUB_ENV
|
||||
elif [[ "$pr_number" != "" ]]; then
|
||||
echo "PR_NUMBER=$pr_number" >> $GITHUB_ENV
|
||||
elif [[ "$is_pull_request" == "true" ]]; then
|
||||
echo "PR_NUMBER=$event_number" >> $GITHUB_ENV
|
||||
if [[ "${{ github.event.issue.number }}" != "" && "${{ github.event.issue.pull_request }}" != "" ]]; then
|
||||
echo "PR_NUMBER=${{ github.event.issue.number }}" >> $GITHUB_ENV
|
||||
elif [[ "${{ github.event.pull_request.number }}" != "" ]]; then
|
||||
echo "PR_NUMBER=${{ github.event.pull_request.number }}" >> $GITHUB_ENV
|
||||
elif [[ "${{ github.event.pull_request }}" != "" ]]; then
|
||||
echo "PR_NUMBER=${{ github.event.number }}" >> $GITHUB_ENV
|
||||
else
|
||||
echo "PR_NUMBER=" >> $GITHUB_ENV
|
||||
fi
|
||||
@ -35,8 +29,8 @@ jobs:
|
||||
- name: Check PR number
|
||||
shell: bash
|
||||
run: |
|
||||
echo "$PR_NUMBER"
|
||||
echo "${{ env.PR_NUMBER }}"
|
||||
|
||||
- name: Set PR number
|
||||
id: set_pr_number
|
||||
run: echo "PR_NUMBER=$PR_NUMBER" >> "$GITHUB_OUTPUT"
|
||||
run: echo "PR_NUMBER=${{ env.PR_NUMBER }}" >> "$GITHUB_OUTPUT"
|
||||
|
||||
64
.github/workflows/model_jobs.yml
vendored
64
.github/workflows/model_jobs.yml
vendored
@ -28,9 +28,6 @@ on:
|
||||
report_repo_id:
|
||||
required: false
|
||||
type: string
|
||||
pytest_marker:
|
||||
required: false
|
||||
type: string
|
||||
|
||||
env:
|
||||
HF_HOME: /mnt/cache
|
||||
@ -41,6 +38,7 @@ env:
|
||||
# For gated repositories, we still need to agree to share information on the Hub repo. page in order to get access.
|
||||
# This token is created under the bot `hf-transformers-bot`.
|
||||
HF_HUB_READ_TOKEN: ${{ secrets.HF_HUB_READ_TOKEN }}
|
||||
SIGOPT_API_TOKEN: ${{ secrets.SIGOPT_API_TOKEN }}
|
||||
TF_FORCE_GPU_ALLOW_GROWTH: true
|
||||
CUDA_VISIBLE_DEVICES: 0,1
|
||||
|
||||
@ -62,33 +60,25 @@ jobs:
|
||||
steps:
|
||||
- name: Echo input and matrix info
|
||||
shell: bash
|
||||
env:
|
||||
folder_slices: ${{ inputs.folder_slices }}
|
||||
matrix_folders: ${{ matrix.folders }}
|
||||
slice_data: ${{ toJson(fromJson(inputs.folder_slices)[inputs.slice_id]) }}
|
||||
run: |
|
||||
echo "$folder_slices"
|
||||
echo "$matrix_folders"
|
||||
echo "$slice_data"
|
||||
echo "${{ inputs.folder_slices }}"
|
||||
echo "${{ matrix.folders }}"
|
||||
echo "${{ toJson(fromJson(inputs.folder_slices)[inputs.slice_id]) }}"
|
||||
|
||||
- name: Echo folder ${{ matrix.folders }}
|
||||
shell: bash
|
||||
# For folders like `models/bert`, set an env. var. (`matrix_folders`) to `models_bert`, which will be used to
|
||||
# set the artifact folder names (because the character `/` is not allowed).
|
||||
env:
|
||||
matrix_folders_raw: ${{ matrix.folders }}
|
||||
run: |
|
||||
echo "$matrix_folders_raw"
|
||||
matrix_folders="${matrix_folders_raw/'models/'/'models_'}"
|
||||
echo "${{ matrix.folders }}"
|
||||
matrix_folders=${{ matrix.folders }}
|
||||
matrix_folders=${matrix_folders/'models/'/'models_'}
|
||||
echo "$matrix_folders"
|
||||
echo "matrix_folders=$matrix_folders" >> $GITHUB_ENV
|
||||
|
||||
- name: Update clone
|
||||
working-directory: /transformers
|
||||
env:
|
||||
commit_sha: ${{ inputs.commit_sha || github.sha }}
|
||||
run: |
|
||||
git fetch origin "$commit_sha" && git checkout "$commit_sha"
|
||||
run: git fetch && git checkout ${{ inputs.commit_sha || github.sha }}
|
||||
|
||||
- name: Reinstall transformers in edit mode (remove the one installed during docker image build)
|
||||
working-directory: /transformers
|
||||
@ -123,17 +113,15 @@ jobs:
|
||||
id: set_machine_type
|
||||
working-directory: /transformers
|
||||
shell: bash
|
||||
env:
|
||||
input_machine_type: ${{ inputs.machine_type }}
|
||||
run: |
|
||||
echo "$input_machine_type"
|
||||
echo "${{ inputs.machine_type }}"
|
||||
|
||||
if [ "$input_machine_type" = "aws-g5-4xlarge-cache" ]; then
|
||||
if [ "${{ inputs.machine_type }}" = "aws-g5-4xlarge-cache" ]; then
|
||||
machine_type=single-gpu
|
||||
elif [ "$input_machine_type" = "aws-g5-12xlarge-cache" ]; then
|
||||
elif [ "${{ inputs.machine_type }}" = "aws-g5-12xlarge-cache" ]; then
|
||||
machine_type=multi-gpu
|
||||
else
|
||||
machine_type="$input_machine_type"
|
||||
machine_type=${{ inputs.machine_type }}
|
||||
fi
|
||||
|
||||
echo "$machine_type"
|
||||
@ -142,21 +130,15 @@ jobs:
|
||||
|
||||
- name: Create report directory if it doesn't exist
|
||||
shell: bash
|
||||
env:
|
||||
report_name_prefix: ${{ inputs.report_name_prefix }}
|
||||
run: |
|
||||
mkdir -p "/transformers/reports/${machine_type}_${report_name_prefix}_${matrix_folders}_test_reports"
|
||||
echo "dummy" > "/transformers/reports/${machine_type}_${report_name_prefix}_${matrix_folders}_test_reports/dummy.txt"
|
||||
ls -la "/transformers/reports/${machine_type}_${report_name_prefix}_${matrix_folders}_test_reports"
|
||||
mkdir -p /transformers/reports/${{ env.machine_type }}_${{ inputs.report_name_prefix }}_${{ env.matrix_folders }}_test_reports
|
||||
echo "dummy" > /transformers/reports/${{ env.machine_type }}_${{ inputs.report_name_prefix }}_${{ env.matrix_folders }}_test_reports/dummy.txt
|
||||
ls -la /transformers/reports/${{ env.machine_type }}_${{ inputs.report_name_prefix }}_${{ env.matrix_folders }}_test_reports
|
||||
|
||||
- name: Run all tests on GPU
|
||||
working-directory: /transformers
|
||||
env:
|
||||
report_name_prefix: ${{ inputs.report_name_prefix }}
|
||||
pytest_marker: ${{ inputs.pytest_marker }}
|
||||
model: ${{ matrix.folders }}
|
||||
run: |
|
||||
script -q -c "PATCH_TESTING_METHODS_TO_COLLECT_OUTPUTS=yes _PATCHED_TESTING_METHODS_OUTPUT_DIR=/transformers/reports/${machine_type}_${report_name_prefix}_${matrix_folders}_test_reports python3 -m pytest -rsfE -v -m '${pytest_marker}' --make-reports=${machine_type}_${report_name_prefix}_${matrix_folders}_test_reports tests/${model}" test_outputs.txt
|
||||
script -q -c "PATCH_TESTING_METHODS_TO_COLLECT_OUTPUTS=yes _PATCHED_TESTING_METHODS_OUTPUT_DIR=/transformers/reports/${{ env.machine_type }}_${{ inputs.report_name_prefix }}_${{ env.matrix_folders }}_test_reports python3 -m pytest -rsfE -v --make-reports=${{ env.machine_type }}_${{ inputs.report_name_prefix }}_${{ env.matrix_folders }}_test_reports tests/${{ matrix.folders }}" test_outputs.txt
|
||||
ls -la
|
||||
# Extract the exit code from the output file
|
||||
EXIT_CODE=$(tail -1 test_outputs.txt | grep -o 'COMMAND_EXIT_CODE="[0-9]*"' | cut -d'"' -f2)
|
||||
@ -167,25 +149,19 @@ jobs:
|
||||
# This step is only to show information on Github Actions log.
|
||||
# Always mark this step as successful, even if the report directory or the file `failures_short.txt` in it doesn't exist
|
||||
continue-on-error: true
|
||||
env:
|
||||
report_name_prefix: ${{ inputs.report_name_prefix }}
|
||||
run: cat "/transformers/reports/${machine_type}_${report_name_prefix}_${matrix_folders}_test_reports/failures_short.txt"
|
||||
run: cat /transformers/reports/${{ env.machine_type }}_${{ inputs.report_name_prefix }}_${{ env.matrix_folders }}_test_reports/failures_short.txt
|
||||
|
||||
- name: Captured information
|
||||
if: ${{ failure() }}
|
||||
continue-on-error: true
|
||||
env:
|
||||
report_name_prefix: ${{ inputs.report_name_prefix }}
|
||||
run: |
|
||||
cat "/transformers/reports/${machine_type}_${report_name_prefix}_${matrix_folders}_test_reports/captured_info.txt"
|
||||
cat /transformers/reports/${{ env.machine_type }}_${{ inputs.report_name_prefix }}_${{ env.matrix_folders }}_test_reports/captured_info.txt
|
||||
|
||||
- name: Copy test_outputs.txt
|
||||
if: ${{ always() }}
|
||||
continue-on-error: true
|
||||
env:
|
||||
report_name_prefix: ${{ inputs.report_name_prefix }}
|
||||
run: |
|
||||
cp /transformers/test_outputs.txt "/transformers/reports/${machine_type}_${report_name_prefix}_${matrix_folders}_test_reports"
|
||||
cp /transformers/test_outputs.txt /transformers/reports/${{ env.machine_type }}_${{ inputs.report_name_prefix }}_${{ env.matrix_folders }}_test_reports
|
||||
|
||||
- name: "Test suite reports artifacts: ${{ env.machine_type }}_${{ inputs.report_name_prefix }}_${{ env.matrix_folders }}_test_reports"
|
||||
if: ${{ always() }}
|
||||
@ -196,7 +172,7 @@ jobs:
|
||||
|
||||
collated_reports:
|
||||
name: Collated Reports
|
||||
if: ${{ always() && inputs.runner_type != '' }}
|
||||
if: ${{ always() }}
|
||||
needs: run_models_gpu
|
||||
uses: huggingface/transformers/.github/workflows/collated-reports.yml@main
|
||||
with:
|
||||
|
||||
1
.github/workflows/model_jobs_intel_gaudi.yml
vendored
1
.github/workflows/model_jobs_intel_gaudi.yml
vendored
@ -26,6 +26,7 @@ env:
|
||||
TRANSFORMERS_IS_CI: yes
|
||||
PT_ENABLE_INT64_SUPPORT: 1
|
||||
HF_HUB_READ_TOKEN: ${{ secrets.HF_HUB_READ_TOKEN }}
|
||||
SIGOPT_API_TOKEN: ${{ secrets.SIGOPT_API_TOKEN }}
|
||||
HF_HOME: /mnt/cache/.cache/huggingface
|
||||
|
||||
jobs:
|
||||
|
||||
@ -98,7 +98,7 @@ jobs:
|
||||
commit_sha: ${{ needs.get-pr-info.outputs.PR_HEAD_SHA }}
|
||||
pr_number: ${{ needs.get-pr-number.outputs.PR_NUMBER }}
|
||||
package: transformers
|
||||
languages: ar de en es fr hi it ja ko pt zh
|
||||
languages: ar de en es fr hi it ko pt tr zh ja te
|
||||
|
||||
update_run_status:
|
||||
name: Update Check Run Status
|
||||
|
||||
@ -1,4 +1,4 @@
|
||||
name: PR slow CI - Suggestion
|
||||
name: PR slow CI
|
||||
on:
|
||||
pull_request_target:
|
||||
types: [opened, synchronize, reopened]
|
||||
@ -23,28 +23,11 @@ jobs:
|
||||
outputs:
|
||||
jobs: ${{ steps.get_jobs.outputs.jobs_to_run }}
|
||||
steps:
|
||||
# This checkout to the main branch
|
||||
- uses: actions/checkout@v4
|
||||
with:
|
||||
fetch-depth: "0"
|
||||
|
||||
# We need to use `${{ ... }}` here to avoid `Argument list too long` error when a PR changes a lot of files.
|
||||
# (We could also try to use artifact approach, but it's more involved).
|
||||
# `CodeQL` doesn't identify any security issue here. Also `PR_FILES` is from `get-pr-info.yml` by using an api
|
||||
# `github.rest.pulls.listFiles`, which is fine.
|
||||
- name: Write pr_files file
|
||||
run: |
|
||||
cat > pr_files.txt << 'EOF'
|
||||
${{ needs.get-pr-info.outputs.PR_FILES }}
|
||||
EOF
|
||||
|
||||
- name: Get repository content
|
||||
id: repo_content
|
||||
uses: actions/github-script@v6
|
||||
with:
|
||||
script: |
|
||||
const fs = require('node:fs');
|
||||
|
||||
const { data: tests_dir } = await github.rest.repos.getContent({
|
||||
owner: '${{ needs.get-pr-info.outputs.PR_HEAD_REPO_OWNER }}',
|
||||
repo: '${{ needs.get-pr-info.outputs.PR_HEAD_REPO_NAME }}',
|
||||
@ -66,10 +49,38 @@ jobs:
|
||||
ref: '${{ needs.get-pr-info.outputs.PR_HEAD_SHA }}',
|
||||
});
|
||||
|
||||
// Write to files instead of outputs
|
||||
fs.writeFileSync('tests_dir.txt', JSON.stringify(tests_dir, null, 2));
|
||||
fs.writeFileSync('tests_models_dir.txt', JSON.stringify(tests_models_dir, null, 2));
|
||||
fs.writeFileSync('tests_quantization_dir.txt', JSON.stringify(tests_quantization_dir, null, 2));
|
||||
core.setOutput('tests_dir', tests_dir);
|
||||
core.setOutput('tests_models_dir', tests_models_dir);
|
||||
core.setOutput('tests_quantization_dir', tests_quantization_dir);
|
||||
|
||||
# This checkout to the main branch
|
||||
- uses: actions/checkout@v4
|
||||
with:
|
||||
fetch-depth: "0"
|
||||
|
||||
- name: Write pr_files file
|
||||
run: |
|
||||
cat > pr_files.txt << 'EOF'
|
||||
${{ needs.get-pr-info.outputs.PR_FILES }}
|
||||
EOF
|
||||
|
||||
- name: Write tests_dir file
|
||||
run: |
|
||||
cat > tests_dir.txt << 'EOF'
|
||||
${{ steps.repo_content.outputs.tests_dir }}
|
||||
EOF
|
||||
|
||||
- name: Write tests_models_dir file
|
||||
run: |
|
||||
cat > tests_models_dir.txt << 'EOF'
|
||||
${{ steps.repo_content.outputs.tests_models_dir }}
|
||||
EOF
|
||||
|
||||
- name: Write tests_quantization_dir file
|
||||
run: |
|
||||
cat > tests_quantization_dir.txt << 'EOF'
|
||||
${{ steps.repo_content.outputs.tests_quantization_dir }}
|
||||
EOF
|
||||
|
||||
- name: Run script to get jobs to run
|
||||
id: get_jobs
|
||||
4
.github/workflows/push-important-models.yml
vendored
4
.github/workflows/push-important-models.yml
vendored
@ -149,9 +149,9 @@ jobs:
|
||||
with:
|
||||
job: run_models_gpu
|
||||
slack_report_channel: "#transformers-ci-push"
|
||||
docker: huggingface/transformers-all-latest-gpu:flash-attn
|
||||
docker: huggingface/transformers-all-latest-gpu
|
||||
ci_event: push
|
||||
report_repo_id: hf-internal-testing/transformers_ci_push
|
||||
commit_sha: ${{ github.sha }}
|
||||
subdirs: ${{ needs.get_modified_models.outputs.matrix }}
|
||||
models: ${{ needs.get_modified_models.outputs.matrix }}
|
||||
secrets: inherit
|
||||
|
||||
506
.github/workflows/self-comment-ci.yml
vendored
506
.github/workflows/self-comment-ci.yml
vendored
@ -20,37 +20,66 @@ env:
|
||||
# For gated repositories, we still need to agree to share information on the Hub repo. page in order to get access.
|
||||
# This token is created under the bot `hf-transformers-bot`.
|
||||
HF_HUB_READ_TOKEN: ${{ secrets.HF_HUB_READ_TOKEN }}
|
||||
SIGOPT_API_TOKEN: ${{ secrets.SIGOPT_API_TOKEN }}
|
||||
TF_FORCE_GPU_ALLOW_GROWTH: true
|
||||
CUDA_VISIBLE_DEVICES: 0,1
|
||||
|
||||
|
||||
jobs:
|
||||
get-pr-number:
|
||||
runs-on: ubuntu-22.04
|
||||
name: Get PR number
|
||||
# For security: only allow team members to run
|
||||
if: ${{ github.event.issue.state == 'open' && contains(fromJSON('["ydshieh", "ArthurZucker", "zucchini-nlp", "molbap", "gante", "LysandreJik", "Cyrilvallez", "Rocketknight1", "SunMarc", "eustlb", "MekkCyber", "vasqu", "ivarflakstad", "stevhliu", "ebezzam", "remi-or", "itazap"]'), github.actor) && (startsWith(github.event.comment.body, 'run-slow') || startsWith(github.event.comment.body, 'run slow') || startsWith(github.event.comment.body, 'run_slow')) }}
|
||||
uses: ./.github/workflows/get-pr-number.yml
|
||||
outputs:
|
||||
PR_NUMBER: ${{ steps.set_pr_number.outputs.PR_NUMBER }}
|
||||
steps:
|
||||
- name: Get PR number
|
||||
shell: bash
|
||||
run: |
|
||||
if [[ "${{ github.event.issue.number }}" != "" && "${{ github.event.issue.pull_request }}" != "" ]]; then
|
||||
echo "PR_NUMBER=${{ github.event.issue.number }}" >> $GITHUB_ENV
|
||||
else
|
||||
echo "PR_NUMBER=" >> $GITHUB_ENV
|
||||
fi
|
||||
|
||||
get-pr-info:
|
||||
name: Get PR commit SHA
|
||||
- name: Check PR number
|
||||
shell: bash
|
||||
run: |
|
||||
echo "${{ env.PR_NUMBER }}"
|
||||
|
||||
- name: Set PR number
|
||||
id: set_pr_number
|
||||
run: echo "PR_NUMBER=${{ env.PR_NUMBER }}" >> "$GITHUB_OUTPUT"
|
||||
|
||||
get-sha:
|
||||
runs-on: ubuntu-22.04
|
||||
needs: get-pr-number
|
||||
if: ${{ needs.get-pr-number.outputs.PR_NUMBER != ''}}
|
||||
uses: ./.github/workflows/get-pr-info.yml
|
||||
with:
|
||||
pr_number: ${{ needs.get-pr-number.outputs.PR_NUMBER }}
|
||||
|
||||
check-timestamps:
|
||||
name: Check timestamps (security check)
|
||||
runs-on: ubuntu-22.04
|
||||
needs: get-pr-info
|
||||
outputs:
|
||||
PR_HEAD_SHA: ${{ needs.get-pr-info.outputs.PR_HEAD_SHA }}
|
||||
PR_MERGE_SHA: ${{ needs.get-pr-info.outputs.PR_MERGE_COMMIT_SHA }}
|
||||
PR_HEAD_SHA: ${{ steps.get_sha.outputs.PR_HEAD_SHA }}
|
||||
PR_MERGE_SHA: ${{ steps.get_sha.outputs.PR_MERGE_SHA }}
|
||||
steps:
|
||||
- name: Verify `merge_commit` timestamp is older than the issue comment timestamp
|
||||
- uses: actions/checkout@v4
|
||||
with:
|
||||
fetch-depth: "0"
|
||||
ref: "refs/pull/${{needs.get-pr-number.outputs.PR_NUMBER}}/merge"
|
||||
|
||||
- name: Get SHA (and verify timestamps against the issue comment date)
|
||||
id: get_sha
|
||||
env:
|
||||
PR_NUMBER: ${{ needs.get-pr-number.outputs.PR_NUMBER }}
|
||||
COMMENT_DATE: ${{ github.event.comment.created_at }}
|
||||
PR_MERGE_COMMIT_TIMESTAMP: ${{ needs.get-pr-info.outputs.PR_MERGE_COMMIT_TIMESTAMP }}
|
||||
run: |
|
||||
git fetch origin refs/pull/$PR_NUMBER/head:refs/remotes/pull/$PR_NUMBER/head
|
||||
git checkout refs/remotes/pull/$PR_NUMBER/head
|
||||
echo "PR_HEAD_SHA: $(git log -1 --format=%H)"
|
||||
echo "PR_HEAD_SHA=$(git log -1 --format=%H)" >> "$GITHUB_OUTPUT"
|
||||
git fetch origin refs/pull/$PR_NUMBER/merge:refs/remotes/pull/$PR_NUMBER/merge
|
||||
git checkout refs/remotes/pull/$PR_NUMBER/merge
|
||||
echo "PR_MERGE_SHA: $(git log -1 --format=%H)"
|
||||
echo "PR_MERGE_SHA=$(git log -1 --format=%H)" >> "$GITHUB_OUTPUT"
|
||||
PR_MERGE_COMMIT_TIMESTAMP=$(git log -1 --date=unix --format=%cd)
|
||||
echo "PR_MERGE_COMMIT_TIMESTAMP: $PR_MERGE_COMMIT_TIMESTAMP"
|
||||
COMMENT_TIMESTAMP=$(date -d "${COMMENT_DATE}" +"%s")
|
||||
echo "COMMENT_DATE: $COMMENT_DATE"
|
||||
echo "COMMENT_TIMESTAMP: $COMMENT_TIMESTAMP"
|
||||
@ -59,10 +88,13 @@ jobs:
|
||||
exit -1;
|
||||
fi
|
||||
|
||||
# use a python script to handle this complex logic.
|
||||
# use a python script to handle this complex logic
|
||||
# case 1: `run-slow` (auto. infer with limited number of models, but in particular, new model)
|
||||
# case 2: `run-slow model_1, model_2`
|
||||
get-tests:
|
||||
runs-on: ubuntu-22.04
|
||||
needs: [get-pr-number, check-timestamps]
|
||||
needs: [get-pr-number, get-sha]
|
||||
if: ${{ needs.get-pr-number.outputs.PR_NUMBER != ''}}
|
||||
outputs:
|
||||
models: ${{ steps.models_to_run.outputs.models }}
|
||||
quantizations: ${{ steps.models_to_run.outputs.quantizations }}
|
||||
@ -70,11 +102,11 @@ jobs:
|
||||
- uses: actions/checkout@v4
|
||||
with:
|
||||
fetch-depth: "0"
|
||||
ref: "refs/pull/${{ needs.get-pr-number.outputs.PR_NUMBER }}/merge"
|
||||
ref: "refs/pull/${{needs.get-pr-number.outputs.PR_NUMBER}}/merge"
|
||||
|
||||
- name: Verify merge commit SHA
|
||||
env:
|
||||
VERIFIED_PR_MERGE_SHA: ${{ needs.check-timestamps.outputs.PR_MERGE_SHA }}
|
||||
VERIFIED_PR_MERGE_SHA: ${{ needs.get-sha.outputs.PR_MERGE_SHA }}
|
||||
run: |
|
||||
PR_MERGE_SHA=$(git log -1 --format=%H)
|
||||
if [ $PR_MERGE_SHA != $VERIFIED_PR_MERGE_SHA ]; then
|
||||
@ -95,33 +127,11 @@ jobs:
|
||||
- name: Show models to test
|
||||
id: models_to_run
|
||||
run: |
|
||||
echo "$models"
|
||||
echo "models=$models" >> $GITHUB_OUTPUT
|
||||
echo "$quantizations"
|
||||
echo "quantizations=$quantizations" >> $GITHUB_OUTPUT
|
||||
|
||||
# Report back if we are not able to get the tests (for example, security check is failing)
|
||||
report_error_earlier:
|
||||
name: Report error earlier
|
||||
if: ${{ always() && needs.get-pr-info.result == 'success' && needs.get-tests.result != 'success' }}
|
||||
needs: [get-pr-number, get-pr-info, get-tests]
|
||||
permissions:
|
||||
pull-requests: write
|
||||
runs-on: ubuntu-22.04
|
||||
steps:
|
||||
- name: Reply to the comment
|
||||
env:
|
||||
GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
|
||||
GITHUB_RUN_URL: https://github.com/${{ github.repository }}/actions/runs/${{ github.run_id }}
|
||||
github_repository: ${{ github.repository }}
|
||||
pr_number: ${{ needs.get-pr-number.outputs.PR_NUMBER }}
|
||||
run: |
|
||||
gh api \
|
||||
--method POST \
|
||||
-H "Accept: application/vnd.github+json" \
|
||||
-H "X-GitHub-Api-Version: 2022-11-28" \
|
||||
"repos/${github_repository}/issues/${pr_number}/comments" \
|
||||
-f body="💔 This comment contains \`run-slow\`, but unknown error occurred and [the workflow run]($GITHUB_RUN_URL) aborted!"
|
||||
echo "${{ env.models }}"
|
||||
echo "models=${{ env.models }}" >> $GITHUB_ENV
|
||||
echo "models=${{ env.models }}" >> $GITHUB_OUTPUT
|
||||
echo "${{ env.quantizations }}"
|
||||
echo "quantizations=${{ env.quantizations }}" >> $GITHUB_OUTPUT
|
||||
|
||||
reply_to_comment:
|
||||
name: Reply to the comment
|
||||
@ -134,20 +144,20 @@ jobs:
|
||||
- name: Reply to the comment
|
||||
env:
|
||||
GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
|
||||
BODY: '\n\nmodels: ${{ needs.get-tests.outputs.models }}\nquantizations: ${{ needs.get-tests.outputs.quantizations }}'
|
||||
github_repository: ${{ github.repository }}
|
||||
pr_number: ${{ needs.get-pr-number.outputs.PR_NUMBER }}
|
||||
MODELS: ${{ needs.get-tests.outputs.models }}
|
||||
BODY: "\n\nmodels: ${{ needs.get-tests.outputs.models }}\nquantizations: ${{ needs.get-tests.outputs.quantizations }}"
|
||||
run: |
|
||||
gh api \
|
||||
--method POST \
|
||||
-H "Accept: application/vnd.github+json" \
|
||||
-H "X-GitHub-Api-Version: 2022-11-28" \
|
||||
"repos/${github_repository}/issues/${pr_number}/comments" \
|
||||
-f body="This comment contains \`run-slow\`, running the specified jobs: $(echo -e "$BODY")"
|
||||
repos/${{ github.repository }}/issues/${{ needs.get-pr-number.outputs.PR_NUMBER }}/comments \
|
||||
-f "body=This comment contains run-slow, running the specified jobs: ${{ env.BODY }} ..."
|
||||
|
||||
create_run:
|
||||
name: Create run
|
||||
needs: [check-timestamps, reply_to_comment]
|
||||
if: ${{ needs.get-tests.outputs.models != '[]' || needs.get-tests.outputs.quantizations != '[]' }}
|
||||
needs: [get-sha, get-tests, reply_to_comment]
|
||||
permissions:
|
||||
statuses: write
|
||||
runs-on: ubuntu-22.04
|
||||
@ -159,196 +169,248 @@ jobs:
|
||||
# Create a commit status (pending) for a run of this workflow. The status has to be updated later in `update_run_status`.
|
||||
# See https://docs.github.com/en/rest/commits/statuses?apiVersion=2022-11-28#create-a-commit-status
|
||||
GITHUB_RUN_URL: https://github.com/${{ github.repository }}/actions/runs/${{ github.run_id }}
|
||||
github_repository: ${{ github.repository }}
|
||||
pr_head_sha: ${{ needs.check-timestamps.outputs.PR_HEAD_SHA }}
|
||||
run: |
|
||||
gh api \
|
||||
--method POST \
|
||||
-H "Accept: application/vnd.github+json" \
|
||||
-H "X-GitHub-Api-Version: 2022-11-28" \
|
||||
"repos/${github_repository}/statuses/${pr_head_sha}" \
|
||||
repos/${{ github.repository }}/statuses/${{ needs.get-sha.outputs.PR_HEAD_SHA }} \
|
||||
-f "target_url=$GITHUB_RUN_URL" -f "state=pending" -f "description=Slow CI job" -f "context=pytest/custom-tests"
|
||||
|
||||
model-ci:
|
||||
name: Model CI
|
||||
run_models_gpu:
|
||||
name: Run all tests for the model
|
||||
if: ${{ needs.get-tests.outputs.models != '[]' }}
|
||||
uses: ./.github/workflows/self-scheduled.yml
|
||||
needs: [get-pr-number, check-timestamps, get-tests, create_run]
|
||||
with:
|
||||
job: run_models_gpu
|
||||
slack_report_channel: "#transformers-ci-pr"
|
||||
docker: huggingface/transformers-all-latest-gpu
|
||||
ci_event: PR Comment CI
|
||||
report_repo_id: hf-internal-testing/transformers_pr_ci
|
||||
commit_sha: ${{ needs.check-timestamps.outputs.PR_MERGE_SHA }}
|
||||
subdirs: ${{ needs.get-tests.outputs.models }}
|
||||
pr_number: ${{ needs.get-pr-number.outputs.PR_NUMBER }}
|
||||
secrets: inherit
|
||||
needs: [get-pr-number, get-sha, get-tests, create_run]
|
||||
strategy:
|
||||
fail-fast: false
|
||||
matrix:
|
||||
folders: ${{ fromJson(needs.get-tests.outputs.models) }}
|
||||
machine_type: [aws-g5-4xlarge-cache, aws-g5-12xlarge-cache]
|
||||
runs-on:
|
||||
group: '${{ matrix.machine_type }}'
|
||||
container:
|
||||
image: huggingface/transformers-all-latest-gpu
|
||||
options: --gpus all --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
|
||||
steps:
|
||||
- name: Echo input and matrix info
|
||||
shell: bash
|
||||
run: |
|
||||
echo "${{ matrix.folders }}"
|
||||
|
||||
quantization-ci:
|
||||
name: Quantization CI
|
||||
- name: Echo folder ${{ matrix.folders }}
|
||||
shell: bash
|
||||
# For folders like `models/bert`, set an env. var. (`matrix_folders`) to `models_bert`, which will be used to
|
||||
# set the artifact folder names (because the character `/` is not allowed).
|
||||
run: |
|
||||
echo "${{ matrix.folders }}"
|
||||
matrix_folders=${{ matrix.folders }}
|
||||
matrix_folders=${matrix_folders/'models/'/'models_'}
|
||||
echo "$matrix_folders"
|
||||
echo "matrix_folders=$matrix_folders" >> $GITHUB_ENV
|
||||
|
||||
- name: Checkout to PR merge commit
|
||||
working-directory: /transformers
|
||||
run: |
|
||||
git fetch origin refs/pull/${{ needs.get-pr-number.outputs.PR_NUMBER }}/merge:refs/remotes/pull/${{ needs.get-pr-number.outputs.PR_NUMBER }}/merge
|
||||
git checkout refs/remotes/pull/${{ needs.get-pr-number.outputs.PR_NUMBER }}/merge
|
||||
git log -1 --format=%H
|
||||
|
||||
- name: Verify merge commit SHA
|
||||
env:
|
||||
VERIFIED_PR_MERGE_SHA: ${{ needs.get-sha.outputs.PR_MERGE_SHA }}
|
||||
working-directory: /transformers
|
||||
run: |
|
||||
PR_MERGE_SHA=$(git log -1 --format=%H)
|
||||
if [ $PR_MERGE_SHA != $VERIFIED_PR_MERGE_SHA ]; then
|
||||
echo "The merged commit SHA is not the same as the verified one! Security issue detected, abort the workflow!";
|
||||
exit -1;
|
||||
fi
|
||||
|
||||
- name: Reinstall transformers in edit mode (remove the one installed during docker image build)
|
||||
working-directory: /transformers
|
||||
run: python3 -m pip uninstall -y transformers && python3 -m pip install -e .
|
||||
|
||||
- name: NVIDIA-SMI
|
||||
run: |
|
||||
nvidia-smi
|
||||
|
||||
- name: Set `machine_type` for report and artifact names
|
||||
working-directory: /transformers
|
||||
shell: bash
|
||||
run: |
|
||||
echo "${{ matrix.machine_type }}"
|
||||
if [ "${{ matrix.machine_type }}" = "aws-g5-4xlarge-cache" ]; then
|
||||
machine_type=single-gpu
|
||||
elif [ "${{ matrix.machine_type }}" = "aws-g5-12xlarge-cache" ]; then
|
||||
machine_type=multi-gpu
|
||||
else
|
||||
machine_type=${{ matrix.machine_type }}
|
||||
fi
|
||||
echo "$machine_type"
|
||||
echo "machine_type=$machine_type" >> $GITHUB_ENV
|
||||
|
||||
- name: Environment
|
||||
working-directory: /transformers
|
||||
run: |
|
||||
python3 utils/print_env.py
|
||||
|
||||
- name: Show installed libraries and their versions
|
||||
working-directory: /transformers
|
||||
run: pip freeze
|
||||
|
||||
- name: Run all tests on GPU
|
||||
working-directory: /transformers
|
||||
run: |
|
||||
export CUDA_VISIBLE_DEVICES="$(python3 utils/set_cuda_devices_for_ci.py --test_folder ${{ matrix.folders }})"
|
||||
echo $CUDA_VISIBLE_DEVICES
|
||||
python3 -m pytest -v -rsfE --make-reports=${{ env.machine_type }}_run_models_gpu_${{ matrix.folders }}_test_reports tests/${{ matrix.folders }}
|
||||
|
||||
- name: Failure short reports
|
||||
if: ${{ failure() }}
|
||||
continue-on-error: true
|
||||
run: cat /transformers/reports/${{ env.machine_type }}_run_models_gpu_${{ matrix.folders }}_test_reports/failures_short.txt
|
||||
|
||||
- name: Make sure report directory exists
|
||||
shell: bash
|
||||
run: |
|
||||
mkdir -p /transformers/reports/${{ env.machine_type }}_run_models_gpu_${{ matrix.folders }}_test_reports
|
||||
echo "hello" > /transformers/reports/${{ env.machine_type }}_run_models_gpu_${{ matrix.folders }}_test_reports/hello.txt
|
||||
echo "${{ env.machine_type }}_run_models_gpu_${{ matrix.folders }}_test_reports"
|
||||
|
||||
- name: "Test suite reports artifacts: ${{ env.machine_type }}_run_models_gpu_${{ env.matrix_folders }}_test_reports"
|
||||
if: ${{ always() }}
|
||||
uses: actions/upload-artifact@v4
|
||||
with:
|
||||
name: ${{ env.machine_type }}_run_models_gpu_${{ env.matrix_folders }}_test_reports
|
||||
path: /transformers/reports/${{ env.machine_type }}_run_models_gpu_${{ matrix.folders }}_test_reports
|
||||
|
||||
run_quantization_torch_gpu:
|
||||
name: Run all tests for a quantization
|
||||
if: ${{ needs.get-tests.outputs.quantizations != '[]' }}
|
||||
uses: ./.github/workflows/self-scheduled.yml
|
||||
needs: [get-pr-number, check-timestamps, get-tests, create_run]
|
||||
with:
|
||||
job: run_quantization_torch_gpu
|
||||
slack_report_channel: "#transformers-ci-pr"
|
||||
docker: huggingface/transformers-quantization-latest-gpu
|
||||
ci_event: PR Comment CI
|
||||
report_repo_id: hf-internal-testing/transformers_pr_ci
|
||||
commit_sha: ${{ needs.check-timestamps.outputs.PR_MERGE_SHA }}
|
||||
subdirs: ${{ needs.get-tests.outputs.quantizations }}
|
||||
pr_number: ${{ needs.get-pr-number.outputs.PR_NUMBER }}
|
||||
secrets: inherit
|
||||
needs: [get-pr-number, get-sha, get-tests, create_run]
|
||||
strategy:
|
||||
fail-fast: false
|
||||
matrix:
|
||||
folders: ${{ fromJson(needs.get-tests.outputs.quantizations) }}
|
||||
machine_type: [aws-g5-4xlarge-cache, aws-g5-12xlarge-cache]
|
||||
runs-on:
|
||||
group: '${{ matrix.machine_type }}'
|
||||
container:
|
||||
image: huggingface/transformers-quantization-latest-gpu
|
||||
options: --gpus all --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
|
||||
steps:
|
||||
- name: Echo folder ${{ matrix.folders }}
|
||||
shell: bash
|
||||
run: |
|
||||
echo "${{ matrix.folders }}"
|
||||
matrix_folders=${{ matrix.folders }}
|
||||
matrix_folders=${matrix_folders/'quantization/'/'quantization_'}
|
||||
echo "$matrix_folders"
|
||||
echo "matrix_folders=$matrix_folders" >> $GITHUB_ENV
|
||||
|
||||
report:
|
||||
name: Check & Report
|
||||
needs: [get-pr-number, check-timestamps, create_run, model-ci, quantization-ci]
|
||||
- name: Checkout to PR merge commit
|
||||
working-directory: /transformers
|
||||
run: |
|
||||
git fetch origin refs/pull/${{ needs.get-pr-number.outputs.PR_NUMBER }}/merge:refs/remotes/pull/${{ needs.get-pr-number.outputs.PR_NUMBER }}/merge
|
||||
git checkout refs/remotes/pull/${{ needs.get-pr-number.outputs.PR_NUMBER }}/merge
|
||||
git log -1 --format=%H
|
||||
|
||||
- name: Verify merge commit SHA
|
||||
env:
|
||||
VERIFIED_PR_MERGE_SHA: ${{ needs.get-sha.outputs.PR_MERGE_SHA }}
|
||||
working-directory: /transformers
|
||||
run: |
|
||||
PR_MERGE_SHA=$(git log -1 --format=%H)
|
||||
if [ $PR_MERGE_SHA != $VERIFIED_PR_MERGE_SHA ]; then
|
||||
echo "The merged commit SHA is not the same as the verified one! Security issue detected, abort the workflow!";
|
||||
exit -1;
|
||||
fi
|
||||
|
||||
- name: Reinstall transformers in edit mode (remove the one installed during docker image build)
|
||||
working-directory: /transformers
|
||||
run: python3 -m pip uninstall -y transformers && python3 -m pip install -e .
|
||||
- name: NVIDIA-SMI
|
||||
run: |
|
||||
nvidia-smi
|
||||
|
||||
- name: Set `machine_type` for report and artifact names
|
||||
working-directory: /transformers
|
||||
shell: bash
|
||||
run: |
|
||||
echo "${{ matrix.machine_type }}"
|
||||
if [ "${{ matrix.machine_type }}" = "aws-g5-4xlarge-cache" ]; then
|
||||
machine_type=single-gpu
|
||||
elif [ "${{ matrix.machine_type }}" = "aws-g5-12xlarge-cache" ]; then
|
||||
machine_type=multi-gpu
|
||||
else
|
||||
machine_type=${{ matrix.machine_type }}
|
||||
fi
|
||||
echo "$machine_type"
|
||||
echo "machine_type=$machine_type" >> $GITHUB_ENV
|
||||
|
||||
- name: Environment
|
||||
working-directory: /transformers
|
||||
run: |
|
||||
python3 utils/print_env.py
|
||||
|
||||
- name: Show installed libraries and their versions
|
||||
working-directory: /transformers
|
||||
run: pip freeze
|
||||
|
||||
- name: Run quantization tests on GPU
|
||||
working-directory: /transformers
|
||||
run: |
|
||||
python3 -m pytest -v --make-reports=${{ env.machine_type }}_run_quantization_torch_gpu_${{ matrix.folders }}_test_reports tests/${{ matrix.folders }}
|
||||
|
||||
- name: Failure short reports
|
||||
if: ${{ failure() }}
|
||||
continue-on-error: true
|
||||
run: cat /transformers/reports/${{ env.machine_type }}_run_quantization_torch_gpu_${{ matrix.folders }}_test_reports/failures_short.txt
|
||||
|
||||
- name: Make sure report directory exists
|
||||
shell: bash
|
||||
run: |
|
||||
mkdir -p /transformers/reports/${{ env.machine_type }}_run_quantization_gpu_${{ matrix.folders }}_test_reports
|
||||
echo "hello" > /transformers/reports/${{ env.machine_type }}_run_quantization_gpu_${{ matrix.folders }}_test_reports/hello.txt
|
||||
echo "${{ env.machine_type }}_run_quantization_gpu_${{ matrix.folders }}_test_reports"
|
||||
|
||||
- name: "Test suite reports artifacts: ${{ env.machine_type }}_run_quantization_torch_gpu_${{ env.matrix_folders }}_test_reports"
|
||||
if: ${{ always() }}
|
||||
uses: actions/upload-artifact@v4
|
||||
with:
|
||||
name: ${{ env.machine_type }}_run_quantization_torch_gpu_${{ env.matrix_folders }}_test_reports
|
||||
path: /transformers/reports/${{ env.machine_type }}_run_quantization_torch_gpu_${{ matrix.folders }}_test_reports
|
||||
|
||||
update_run_status:
|
||||
name: Update Check Run Status
|
||||
needs: [get-sha, create_run, run_models_gpu, run_quantization_torch_gpu]
|
||||
permissions:
|
||||
pull-requests: write
|
||||
statuses: write
|
||||
if: ${{ always() && needs.create_run.result == 'success' }}
|
||||
runs-on: ubuntu-22.04
|
||||
env:
|
||||
GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
|
||||
GITHUB_RUN_URL: https://github.com/${{ github.repository }}/actions/runs/${{ github.run_id }}
|
||||
STATUS_OK: ${{ contains(fromJSON('["skipped", "success"]'), needs.run_models_gpu.result) && contains(fromJSON('["skipped", "success"]'), needs.run_quantization_torch_gpu.result) }}
|
||||
steps:
|
||||
- name: Show reports from jobs
|
||||
env:
|
||||
MODEL_REPORT: ${{ needs.model-ci.outputs.report }}
|
||||
QUANT_REPORT: ${{ needs.quantization-ci.outputs.report }}
|
||||
- name: Get `run_models_gpu` job status
|
||||
run: |
|
||||
echo "$MODEL_REPORT"
|
||||
echo "$QUANT_REPORT"
|
||||
|
||||
- name: Process and filter reports
|
||||
env:
|
||||
MODEL_REPORT: ${{ needs.model-ci.outputs.report }}
|
||||
QUANT_REPORT: ${{ needs.quantization-ci.outputs.report }}
|
||||
run: |
|
||||
# Preprocess with Python
|
||||
python3 << 'PYTHON_SCRIPT'
|
||||
import json
|
||||
import os
|
||||
|
||||
def filter_and_format_report(data):
|
||||
"""
|
||||
Filter out entries where commit is `None` (failing tests who status is not certain) and format as text
|
||||
"""
|
||||
lines = []
|
||||
|
||||
for model, model_result in data.items():
|
||||
model_lines = []
|
||||
for device, failures in model_result.items():
|
||||
|
||||
# Filter out None commits and extract just the test names
|
||||
test_names = [
|
||||
failure['test']
|
||||
for failure in failures
|
||||
if isinstance(failure, dict) and failure.get('commit') is not None
|
||||
]
|
||||
|
||||
# Add tests to model lines
|
||||
for idx, test_name in enumerate(test_names):
|
||||
if idx == 0:
|
||||
job_link = failures[idx]['job_link']
|
||||
model_lines.append(f"- [{model}]({job_link}):")
|
||||
|
||||
model_lines.append(f" {test_name}")
|
||||
|
||||
# Only add model section if it has tests
|
||||
if len(model_lines) > 0:
|
||||
lines.extend(model_lines)
|
||||
lines.append("") # Empty line between models
|
||||
|
||||
return "\n".join(lines).strip()
|
||||
|
||||
# Load and filter reports
|
||||
model_report_str = os.environ.get('MODEL_REPORT', '{}')
|
||||
quant_report_str = os.environ.get('QUANT_REPORT', '{}')
|
||||
|
||||
model_report = json.loads(model_report_str) if model_report_str else {}
|
||||
quant_report = json.loads(quant_report_str) if quant_report_str else {}
|
||||
|
||||
formatted_model = filter_and_format_report(model_report)
|
||||
formatted_quant = filter_and_format_report(quant_report)
|
||||
|
||||
# Write to files
|
||||
with open('model_ci.txt', 'w') as f:
|
||||
f.write(formatted_model)
|
||||
if formatted_model:
|
||||
f.write('\n')
|
||||
|
||||
with open('quantization_ci.txt', 'w') as f:
|
||||
f.write(formatted_quant)
|
||||
if formatted_quant:
|
||||
f.write('\n')
|
||||
PYTHON_SCRIPT
|
||||
|
||||
- name: Post results as PR comment
|
||||
env:
|
||||
GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
|
||||
GITHUB_RUN_URL: https://github.com/${{ github.repository }}/actions/runs/${{ github.run_id }}
|
||||
github_repository: ${{ github.repository }}
|
||||
pr_number: ${{ needs.get-pr-number.outputs.PR_NUMBER }}
|
||||
model_ci_result: ${{ needs.model-ci.result }}
|
||||
quantization_ci_result: ${{ needs.quantization-ci.result }}
|
||||
run: |
|
||||
{
|
||||
echo '## CI Results'
|
||||
echo "[Workflow Run ⚙️]($GITHUB_RUN_URL)"
|
||||
echo ''
|
||||
|
||||
# Check if both jobs were skipped or cancelled
|
||||
if [[ "$model_ci_result" == "skipped" || "$model_ci_result" == "cancelled" ]] && \
|
||||
[[ "$quantization_ci_result" == "skipped" || "$quantization_ci_result" == "cancelled" ]]; then
|
||||
echo '⚠️ No test being reported (jobs are skipped or cancelled)!'
|
||||
echo "STATUS=error" >> $GITHUB_ENV
|
||||
|
||||
# Check if either file has content
|
||||
elif [ -s model_ci.txt ] || [ -s quantization_ci.txt ]; then
|
||||
echo "STATUS=failure" >> $GITHUB_ENV
|
||||
|
||||
# Check if model_ci.txt has content
|
||||
if [ -s model_ci.txt ]; then
|
||||
echo '### Model CI Report'
|
||||
echo ''
|
||||
echo '#### ❌ Failed tests'
|
||||
echo ''
|
||||
cat model_ci.txt
|
||||
echo ''
|
||||
fi
|
||||
|
||||
# Check if quantization_ci.txt has content
|
||||
if [ -s quantization_ci.txt ]; then
|
||||
echo '### Quantization CI Report'
|
||||
echo ''
|
||||
echo '#### ❌ Failed tests'
|
||||
echo ''
|
||||
cat quantization_ci.txt
|
||||
echo ''
|
||||
fi
|
||||
else
|
||||
echo "STATUS=success" >> $GITHUB_ENV
|
||||
echo '✅ No failing test specific to this PR 🎉 !'
|
||||
fi
|
||||
} > comment_body.txt
|
||||
|
||||
gh api \
|
||||
--method POST \
|
||||
-H "Accept: application/vnd.github+json" \
|
||||
-H "X-GitHub-Api-Version: 2022-11-28" \
|
||||
"repos/${github_repository}/issues/${pr_number}/comments" \
|
||||
-F body=@comment_body.txt
|
||||
echo "${{ needs.run_models_gpu.result }}"
|
||||
echo "${{ needs.run_quantization_torch_gpu.result }}"
|
||||
echo $STATUS_OK
|
||||
if [ "$STATUS_OK" = "true" ]; then
|
||||
echo "STATUS=success" >> $GITHUB_ENV
|
||||
else
|
||||
echo "STATUS=failure" >> $GITHUB_ENV
|
||||
fi
|
||||
|
||||
- name: Update PR commit statuses
|
||||
env:
|
||||
GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
|
||||
GITHUB_RUN_URL: https://github.com/${{ github.repository }}/actions/runs/${{ github.run_id }}
|
||||
github_repository: ${{ github.repository }}
|
||||
pr_head_sha: ${{ needs.check-timestamps.outputs.PR_HEAD_SHA }}
|
||||
# The env. variable `STATUS` used here is set in the previous step
|
||||
run: |
|
||||
echo "${{ needs.run_models_gpu.result }}"
|
||||
echo "${{ env.STATUS }}"
|
||||
gh api \
|
||||
--method POST \
|
||||
-H "Accept: application/vnd.github+json" \
|
||||
-H "X-GitHub-Api-Version: 2022-11-28" \
|
||||
"repos/${github_repository}/statuses/${pr_head_sha}" \
|
||||
-f "target_url=$GITHUB_RUN_URL" -f "state=$STATUS" -f "description=Slow CI job" -f "context=pytest/custom-tests"
|
||||
repos/${{ github.repository }}/statuses/${{ needs.get-sha.outputs.PR_HEAD_SHA }} \
|
||||
-f "target_url=$GITHUB_RUN_URL" -f "state=${{ env.STATUS }}" -f "description=Slow CI job" -f "context=pytest/custom-tests"
|
||||
|
||||
1
.github/workflows/self-nightly-caller.yml
vendored
1
.github/workflows/self-nightly-caller.yml
vendored
@ -51,7 +51,6 @@ jobs:
|
||||
slack_report_channel: "#transformers-ci-past-future"
|
||||
docker: huggingface/transformers-all-latest-torch-nightly-gpu
|
||||
ci_event: Nightly CI
|
||||
runner_type: "a10"
|
||||
report_repo_id: hf-internal-testing/transformers_daily_ci_with_torch_nightly
|
||||
commit_sha: ${{ github.event.workflow_run.head_sha || github.sha }}
|
||||
secrets: inherit
|
||||
|
||||
25
.github/workflows/self-push-amd-mi210-caller.yml
vendored
Normal file
25
.github/workflows/self-push-amd-mi210-caller.yml
vendored
Normal file
@ -0,0 +1,25 @@
|
||||
name: Self-hosted runner (AMD mi210 CI caller)
|
||||
|
||||
on:
|
||||
#workflow_run:
|
||||
# workflows: ["Self-hosted runner (push-caller)"]
|
||||
# branches: ["main"]
|
||||
# types: [completed]
|
||||
push:
|
||||
branches:
|
||||
- run_amd_push_ci_caller*
|
||||
paths:
|
||||
- "src/**"
|
||||
- "tests/**"
|
||||
- ".github/**"
|
||||
- "templates/**"
|
||||
- "utils/**"
|
||||
|
||||
jobs:
|
||||
run_amd_ci:
|
||||
name: AMD mi210
|
||||
if: (cancelled() != true) && ((github.event_name == 'workflow_run') || ((github.event_name == 'push') && startsWith(github.ref_name, 'run_amd_push_ci_caller')))
|
||||
uses: ./.github/workflows/self-push-amd.yml
|
||||
with:
|
||||
gpu_flavor: mi210
|
||||
secrets: inherit
|
||||
25
.github/workflows/self-push-amd-mi250-caller.yml
vendored
Normal file
25
.github/workflows/self-push-amd-mi250-caller.yml
vendored
Normal file
@ -0,0 +1,25 @@
|
||||
name: Self-hosted runner (AMD mi250 CI caller)
|
||||
|
||||
on:
|
||||
#workflow_run:
|
||||
# workflows: ["Self-hosted runner (push-caller)"]
|
||||
# branches: ["main"]
|
||||
# types: [completed]
|
||||
push:
|
||||
branches:
|
||||
- run_amd_push_ci_caller*
|
||||
paths:
|
||||
- "src/**"
|
||||
- "tests/**"
|
||||
- ".github/**"
|
||||
- "templates/**"
|
||||
- "utils/**"
|
||||
|
||||
jobs:
|
||||
run_amd_ci:
|
||||
name: AMD mi250
|
||||
if: (cancelled() != true) && ((github.event_name == 'workflow_run') || ((github.event_name == 'push') && startsWith(github.ref_name, 'run_amd_push_ci_caller')))
|
||||
uses: ./.github/workflows/self-push-amd.yml
|
||||
with:
|
||||
gpu_flavor: mi250
|
||||
secrets: inherit
|
||||
334
.github/workflows/self-push-amd.yml
vendored
Normal file
334
.github/workflows/self-push-amd.yml
vendored
Normal file
@ -0,0 +1,334 @@
|
||||
name: Self-hosted runner AMD GPU (push)
|
||||
|
||||
on:
|
||||
workflow_call:
|
||||
inputs:
|
||||
gpu_flavor:
|
||||
required: true
|
||||
type: string
|
||||
|
||||
env:
|
||||
HF_HOME: /mnt/cache
|
||||
TRANSFORMERS_IS_CI: yes
|
||||
OMP_NUM_THREADS: 8
|
||||
MKL_NUM_THREADS: 8
|
||||
PYTEST_TIMEOUT: 60
|
||||
TF_FORCE_GPU_ALLOW_GROWTH: true
|
||||
HF_HUB_READ_TOKEN: ${{ secrets.HF_HUB_READ_TOKEN }}
|
||||
|
||||
jobs:
|
||||
check_runner_status:
|
||||
name: Check Runner Status
|
||||
runs-on: ubuntu-22.04
|
||||
steps:
|
||||
- name: Checkout transformers
|
||||
uses: actions/checkout@v4
|
||||
with:
|
||||
fetch-depth: 2
|
||||
|
||||
- name: Check Runner Status
|
||||
run: python utils/check_self_hosted_runner.py --target_runners amd-mi210-single-gpu-ci-runner-docker --token ${{ secrets.ACCESS_REPO_INFO_TOKEN }}
|
||||
|
||||
check_runners:
|
||||
name: Check Runners
|
||||
needs: check_runner_status
|
||||
strategy:
|
||||
matrix:
|
||||
machine_type: [single-gpu, multi-gpu]
|
||||
runs-on: [self-hosted, amd-gpu, '${{ matrix.machine_type }}', '${{ inputs.gpu_flavor }}']
|
||||
container:
|
||||
image: huggingface/transformers-pytorch-amd-gpu-push-ci # <--- We test only for PyTorch for now
|
||||
options: --device /dev/kfd --device /dev/dri --env ROCR_VISIBLE_DEVICES --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
|
||||
steps:
|
||||
- name: ROCM-SMI
|
||||
run: |
|
||||
rocm-smi
|
||||
- name: ROCM-INFO
|
||||
run: |
|
||||
rocminfo | grep "Agent" -A 14
|
||||
- name: Show ROCR environment
|
||||
run: |
|
||||
echo "ROCR: $ROCR_VISIBLE_DEVICES"
|
||||
|
||||
setup_gpu:
|
||||
name: Setup
|
||||
needs: check_runners
|
||||
strategy:
|
||||
matrix:
|
||||
machine_type: [single-gpu, multi-gpu]
|
||||
runs-on: [self-hosted, amd-gpu, '${{ matrix.machine_type }}', '${{ inputs.gpu_flavor }}']
|
||||
container:
|
||||
image: huggingface/transformers-pytorch-amd-gpu-push-ci # <--- We test only for PyTorch for now
|
||||
options: --device /dev/kfd --device /dev/dri --env ROCR_VISIBLE_DEVICES --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
|
||||
outputs:
|
||||
matrix: ${{ steps.set-matrix.outputs.matrix }}
|
||||
test_map: ${{ steps.set-matrix.outputs.test_map }}
|
||||
env:
|
||||
# `CI_BRANCH_PUSH`: The branch name from the push event
|
||||
# `CI_BRANCH_WORKFLOW_RUN`: The name of the branch on which this workflow is triggered by `workflow_run` event
|
||||
# `CI_SHA_PUSH`: The commit SHA from the push event
|
||||
# `CI_SHA_WORKFLOW_RUN`: The commit SHA that triggers this workflow by `workflow_run` event
|
||||
CI_BRANCH_PUSH: ${{ github.event.ref }}
|
||||
CI_BRANCH_WORKFLOW_RUN: ${{ github.event.workflow_run.head_branch }}
|
||||
CI_SHA_PUSH: ${{ github.event.head_commit.id }}
|
||||
CI_SHA_WORKFLOW_RUN: ${{ github.event.workflow_run.head_sha }}
|
||||
steps:
|
||||
# Necessary to get the correct branch name and commit SHA for `workflow_run` event
|
||||
# We also take into account the `push` event (we might want to test some changes in a branch)
|
||||
- name: Prepare custom environment variables
|
||||
shell: bash
|
||||
# `CI_BRANCH`: The non-empty branch name from the above two (one and only one of them is empty)
|
||||
# `CI_SHA`: The non-empty commit SHA from the above two (one and only one of them is empty)
|
||||
run: |
|
||||
CI_BRANCH_PUSH=${CI_BRANCH_PUSH/'refs/heads/'/''}
|
||||
echo $CI_BRANCH_PUSH
|
||||
echo $CI_BRANCH_WORKFLOW_RUN
|
||||
echo $CI_SHA_PUSH
|
||||
echo $CI_SHA_WORKFLOW_RUN
|
||||
[[ ! -z "$CI_BRANCH_PUSH" ]] && echo "CI_BRANCH=$CI_BRANCH_PUSH" >> $GITHUB_ENV || echo "CI_BRANCH=$CI_BRANCH_WORKFLOW_RUN" >> $GITHUB_ENV
|
||||
[[ ! -z "$CI_SHA_PUSH" ]] && echo "CI_SHA=$CI_SHA_PUSH" >> $GITHUB_ENV || echo "CI_SHA=$CI_SHA_WORKFLOW_RUN" >> $GITHUB_ENV
|
||||
|
||||
- name: print environment variables
|
||||
run: |
|
||||
echo "env.CI_BRANCH = ${{ env.CI_BRANCH }}"
|
||||
echo "env.CI_SHA = ${{ env.CI_SHA }}"
|
||||
|
||||
- name: Update clone using environment variables
|
||||
working-directory: /transformers
|
||||
run: |
|
||||
echo "original branch = $(git branch --show-current)"
|
||||
git fetch && git checkout ${{ env.CI_BRANCH }}
|
||||
echo "updated branch = $(git branch --show-current)"
|
||||
git checkout ${{ env.CI_SHA }}
|
||||
echo "log = $(git log -n 1)"
|
||||
|
||||
- name: Cleanup
|
||||
working-directory: /transformers
|
||||
run: |
|
||||
rm -rf tests/__pycache__
|
||||
rm -rf tests/models/__pycache__
|
||||
rm -rf reports
|
||||
|
||||
- name: Show installed libraries and their versions
|
||||
working-directory: /transformers
|
||||
run: pip freeze
|
||||
|
||||
- name: Fetch the tests to run
|
||||
working-directory: /transformers
|
||||
# TODO: add `git-python` in the docker images
|
||||
run: |
|
||||
pip install --upgrade git-python
|
||||
python3 utils/tests_fetcher.py --diff_with_last_commit | tee test_preparation.txt
|
||||
|
||||
- name: Report fetched tests
|
||||
uses: actions/upload-artifact@v4
|
||||
with:
|
||||
name: test_fetched
|
||||
path: /transformers/test_preparation.txt
|
||||
|
||||
- id: set-matrix
|
||||
name: Organize tests into models
|
||||
working-directory: /transformers
|
||||
# The `keys` is used as GitHub actions matrix for jobs, i.e. `models/bert`, `tokenization`, `pipeline`, etc.
|
||||
# The `test_map` is used to get the actual identified test files under each key.
|
||||
# If no test to run (so no `test_map.json` file), create a dummy map (empty matrix will fail)
|
||||
run: |
|
||||
if [ -f test_map.json ]; then
|
||||
keys=$(python3 -c 'import json; fp = open("test_map.json"); test_map = json.load(fp); fp.close(); d = list(test_map.keys()); print(d)')
|
||||
test_map=$(python3 -c 'import json; fp = open("test_map.json"); test_map = json.load(fp); fp.close(); print(test_map)')
|
||||
else
|
||||
keys=$(python3 -c 'keys = ["dummy"]; print(keys)')
|
||||
test_map=$(python3 -c 'test_map = {"dummy": []}; print(test_map)')
|
||||
fi
|
||||
echo $keys
|
||||
echo $test_map
|
||||
echo "matrix=$keys" >> $GITHUB_OUTPUT
|
||||
echo "test_map=$test_map" >> $GITHUB_OUTPUT
|
||||
|
||||
run_models_gpu:
|
||||
name: Model tests
|
||||
needs: setup_gpu
|
||||
# `dummy` means there is no test to run
|
||||
if: contains(fromJson(needs.setup_gpu.outputs.matrix), 'dummy') != true
|
||||
strategy:
|
||||
fail-fast: false
|
||||
matrix:
|
||||
folders: ${{ fromJson(needs.setup_gpu.outputs.matrix) }}
|
||||
machine_type: [single-gpu, multi-gpu]
|
||||
runs-on: [self-hosted, amd-gpu, '${{ matrix.machine_type }}', '${{ inputs.gpu_flavor }}']
|
||||
container:
|
||||
image: huggingface/transformers-pytorch-amd-gpu-push-ci # <--- We test only for PyTorch for now
|
||||
options: --device /dev/kfd --device /dev/dri --env ROCR_VISIBLE_DEVICES --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
|
||||
env:
|
||||
# For the meaning of these environment variables, see the job `Setup`
|
||||
CI_BRANCH_PUSH: ${{ github.event.ref }}
|
||||
CI_BRANCH_WORKFLOW_RUN: ${{ github.event.workflow_run.head_branch }}
|
||||
CI_SHA_PUSH: ${{ github.event.head_commit.id }}
|
||||
CI_SHA_WORKFLOW_RUN: ${{ github.event.workflow_run.head_sha }}
|
||||
steps:
|
||||
# Necessary to get the correct branch name and commit SHA for `workflow_run` event
|
||||
# We also take into account the `push` event (we might want to test some changes in a branch)
|
||||
- name: Prepare custom environment variables
|
||||
shell: bash
|
||||
# For the meaning of these environment variables, see the job `Setup`
|
||||
run: |
|
||||
CI_BRANCH_PUSH=${CI_BRANCH_PUSH/'refs/heads/'/''}
|
||||
echo $CI_BRANCH_PUSH
|
||||
echo $CI_BRANCH_WORKFLOW_RUN
|
||||
echo $CI_SHA_PUSH
|
||||
echo $CI_SHA_WORKFLOW_RUN
|
||||
[[ ! -z "$CI_BRANCH_PUSH" ]] && echo "CI_BRANCH=$CI_BRANCH_PUSH" >> $GITHUB_ENV || echo "CI_BRANCH=$CI_BRANCH_WORKFLOW_RUN" >> $GITHUB_ENV
|
||||
[[ ! -z "$CI_SHA_PUSH" ]] && echo "CI_SHA=$CI_SHA_PUSH" >> $GITHUB_ENV || echo "CI_SHA=$CI_SHA_WORKFLOW_RUN" >> $GITHUB_ENV
|
||||
|
||||
- name: print environment variables
|
||||
run: |
|
||||
echo "env.CI_BRANCH = ${{ env.CI_BRANCH }}"
|
||||
echo "env.CI_SHA = ${{ env.CI_SHA }}"
|
||||
|
||||
- name: Update clone using environment variables
|
||||
working-directory: /transformers
|
||||
run: |
|
||||
echo "original branch = $(git branch --show-current)"
|
||||
git fetch && git checkout ${{ env.CI_BRANCH }}
|
||||
echo "updated branch = $(git branch --show-current)"
|
||||
git checkout ${{ env.CI_SHA }}
|
||||
echo "log = $(git log -n 1)"
|
||||
|
||||
- name: Reinstall transformers in edit mode (remove the one installed during docker image build)
|
||||
working-directory: /transformers
|
||||
run: python3 -m pip uninstall -y transformers && python3 -m pip install -e .
|
||||
|
||||
- name: Echo folder ${{ matrix.folders }}
|
||||
shell: bash
|
||||
# For folders like `models/bert`, set an env. var. (`matrix_folders`) to `models_bert`, which will be used to
|
||||
# set the artifact folder names (because the character `/` is not allowed).
|
||||
run: |
|
||||
echo "${{ matrix.folders }}"
|
||||
echo "${{ fromJson(needs.setup_gpu.outputs.test_map)[matrix.folders] }}"
|
||||
matrix_folders=${{ matrix.folders }}
|
||||
matrix_folders=${matrix_folders/'models/'/'models_'}
|
||||
echo "$matrix_folders"
|
||||
echo "matrix_folders=$matrix_folders" >> $GITHUB_ENV
|
||||
|
||||
- name: ROCM-SMI
|
||||
run: |
|
||||
rocm-smi
|
||||
- name: ROCM-INFO
|
||||
run: |
|
||||
rocminfo | grep "Agent" -A 14
|
||||
- name: Show ROCR environment
|
||||
run: |
|
||||
echo "ROCR: $ROCR_VISIBLE_DEVICES"
|
||||
|
||||
- name: Environment
|
||||
working-directory: /transformers
|
||||
run: |
|
||||
python3 utils/print_env.py
|
||||
|
||||
- name: Show installed libraries and their versions
|
||||
working-directory: /transformers
|
||||
run: pip freeze
|
||||
|
||||
- name: Run all non-slow selected tests on GPU
|
||||
working-directory: /transformers
|
||||
run: |
|
||||
python3 -m pytest -n 2 --dist=loadfile -v --make-reports=${{ matrix.machine_type }}_run_models_gpu_${{ matrix.folders }}_test_reports ${{ fromJson(needs.setup_gpu.outputs.test_map)[matrix.folders] }} -m "not not_device_test"
|
||||
|
||||
- name: Failure short reports
|
||||
if: ${{ failure() }}
|
||||
continue-on-error: true
|
||||
run: cat /transformers/reports/${{ matrix.machine_type }}_run_models_gpu_${{ matrix.folders }}_test_reports/failures_short.txt
|
||||
|
||||
- name: "Test suite reports artifacts: ${{ matrix.machine_type }}_run_models_gpu_${{ env.matrix_folders }}_test_reports"
|
||||
if: ${{ always() }}
|
||||
uses: actions/upload-artifact@v4
|
||||
with:
|
||||
name: ${{ matrix.machine_type }}_run_models_gpu_${{ env.matrix_folders }}_test_reports
|
||||
path: /transformers/reports/${{ matrix.machine_type }}_run_models_gpu_${{ matrix.folders }}_test_reports
|
||||
|
||||
send_results:
|
||||
name: Send results to webhook
|
||||
runs-on: ubuntu-22.04
|
||||
if: always()
|
||||
needs: [
|
||||
check_runner_status,
|
||||
check_runners,
|
||||
setup_gpu,
|
||||
run_models_gpu,
|
||||
# run_tests_torch_cuda_extensions_single_gpu,
|
||||
# run_tests_torch_cuda_extensions_multi_gpu
|
||||
]
|
||||
env:
|
||||
# For the meaning of these environment variables, see the job `Setup`
|
||||
CI_BRANCH_PUSH: ${{ github.event.ref }}
|
||||
CI_BRANCH_WORKFLOW_RUN: ${{ github.event.workflow_run.head_branch }}
|
||||
CI_SHA_PUSH: ${{ github.event.head_commit.id }}
|
||||
CI_SHA_WORKFLOW_RUN: ${{ github.event.workflow_run.head_sha }}
|
||||
steps:
|
||||
- name: Preliminary job status
|
||||
shell: bash
|
||||
# For the meaning of these environment variables, see the job `Setup`
|
||||
run: |
|
||||
echo "Runner availability: ${{ needs.check_runner_status.result }}"
|
||||
echo "Setup status: ${{ needs.setup_gpu.result }}"
|
||||
echo "Runner status: ${{ needs.check_runners.result }}"
|
||||
|
||||
# Necessary to get the correct branch name and commit SHA for `workflow_run` event
|
||||
# We also take into account the `push` event (we might want to test some changes in a branch)
|
||||
- name: Prepare custom environment variables
|
||||
shell: bash
|
||||
# For the meaning of these environment variables, see the job `Setup`
|
||||
run: |
|
||||
CI_BRANCH_PUSH=${CI_BRANCH_PUSH/'refs/heads/'/''}
|
||||
echo $CI_BRANCH_PUSH
|
||||
echo $CI_BRANCH_WORKFLOW_RUN
|
||||
echo $CI_SHA_PUSH
|
||||
echo $CI_SHA_WORKFLOW_RUN
|
||||
[[ ! -z "$CI_BRANCH_PUSH" ]] && echo "CI_BRANCH=$CI_BRANCH_PUSH" >> $GITHUB_ENV || echo "CI_BRANCH=$CI_BRANCH_WORKFLOW_RUN" >> $GITHUB_ENV
|
||||
[[ ! -z "$CI_SHA_PUSH" ]] && echo "CI_SHA=$CI_SHA_PUSH" >> $GITHUB_ENV || echo "CI_SHA=$CI_SHA_WORKFLOW_RUN" >> $GITHUB_ENV
|
||||
|
||||
- name: print environment variables
|
||||
run: |
|
||||
echo "env.CI_BRANCH = ${{ env.CI_BRANCH }}"
|
||||
echo "env.CI_SHA = ${{ env.CI_SHA }}"
|
||||
|
||||
- uses: actions/checkout@v4
|
||||
# To avoid failure when multiple commits are merged into `main` in a short period of time.
|
||||
# Checking out to an old commit beyond the fetch depth will get an error `fatal: reference is not a tree: ...
|
||||
# (Only required for `workflow_run` event, where we get the latest HEAD on `main` instead of the event commit)
|
||||
with:
|
||||
fetch-depth: 20
|
||||
|
||||
- name: Update clone using environment variables
|
||||
run: |
|
||||
echo "original branch = $(git branch --show-current)"
|
||||
git fetch && git checkout ${{ env.CI_BRANCH }}
|
||||
echo "updated branch = $(git branch --show-current)"
|
||||
git checkout ${{ env.CI_SHA }}
|
||||
echo "log = $(git log -n 1)"
|
||||
|
||||
- uses: actions/download-artifact@v4
|
||||
- name: Send message to Slack
|
||||
env:
|
||||
CI_SLACK_BOT_TOKEN: ${{ secrets.CI_SLACK_BOT_TOKEN }}
|
||||
CI_SLACK_CHANNEL_ID: ${{ secrets.CI_SLACK_CHANNEL_ID }}
|
||||
CI_SLACK_CHANNEL_ID_DAILY: ${{ secrets.CI_SLACK_CHANNEL_ID_DAILY }}
|
||||
CI_SLACK_CHANNEL_ID_AMD: ${{ secrets.CI_SLACK_CHANNEL_ID_AMD }}
|
||||
CI_SLACK_CHANNEL_DUMMY_TESTS: ${{ secrets.CI_SLACK_CHANNEL_DUMMY_TESTS }}
|
||||
CI_SLACK_REPORT_CHANNEL_ID: ${{ secrets.CI_SLACK_CHANNEL_ID_AMD }}
|
||||
ACCESS_REPO_INFO_TOKEN: ${{ secrets.ACCESS_REPO_INFO_TOKEN }}
|
||||
CI_EVENT: Push CI (AMD) - ${{ inputs.gpu_flavor }}
|
||||
CI_TITLE_PUSH: ${{ github.event.head_commit.message }}
|
||||
CI_TITLE_WORKFLOW_RUN: ${{ github.event.workflow_run.head_commit.message }}
|
||||
CI_SHA: ${{ env.CI_SHA }}
|
||||
RUNNER_STATUS: ${{ needs.check_runner_status.result }}
|
||||
RUNNER_ENV_STATUS: ${{ needs.check_runners.result }}
|
||||
SETUP_STATUS: ${{ needs.setup_gpu.result }}
|
||||
|
||||
# We pass `needs.setup_gpu.outputs.matrix` as the argument. A processing in `notification_service.py` to change
|
||||
# `models/bert` to `models_bert` is required, as the artifact names use `_` instead of `/`.
|
||||
run: |
|
||||
pip install huggingface_hub
|
||||
pip install slack_sdk
|
||||
pip show slack_sdk
|
||||
python utils/notification_service.py "${{ needs.setup_gpu.outputs.matrix }}"
|
||||
54
.github/workflows/self-push-caller.yml
vendored
Normal file
54
.github/workflows/self-push-caller.yml
vendored
Normal file
@ -0,0 +1,54 @@
|
||||
# Used to trigger self-push CI
|
||||
name: Self-hosted runner (push-caller)
|
||||
|
||||
on:
|
||||
push:
|
||||
branches:
|
||||
- main
|
||||
paths:
|
||||
- "src/**"
|
||||
- "tests/**"
|
||||
- ".github/**"
|
||||
- "templates/**"
|
||||
- "utils/**"
|
||||
|
||||
jobs:
|
||||
check-for-setup:
|
||||
runs-on: ubuntu-22.04
|
||||
name: Check if setup was changed
|
||||
outputs:
|
||||
changed: ${{ steps.was_changed.outputs.changed }}
|
||||
steps:
|
||||
- uses: actions/checkout@v4
|
||||
with:
|
||||
fetch-depth: "2"
|
||||
|
||||
- name: Get changed files
|
||||
id: changed-files
|
||||
uses: tj-actions/changed-files@1c8e6069583811afb28f97afeaf8e7da80c6be5c
|
||||
|
||||
- name: Was setup changed
|
||||
id: was_changed
|
||||
run: |
|
||||
for file in ${{ steps.changed-files.outputs.all_changed_files }}; do
|
||||
if [ `basename "${file}"` = "setup.py" ]; then
|
||||
echo "changed=1" >> $GITHUB_OUTPUT
|
||||
fi
|
||||
done
|
||||
|
||||
build-docker-containers:
|
||||
needs: check-for-setup
|
||||
if: (github.event_name == 'push') && (needs.check-for-setup.outputs.changed == '1')
|
||||
uses: ./.github/workflows/build-docker-images.yml
|
||||
with:
|
||||
image_postfix: "-push-ci"
|
||||
secrets: inherit
|
||||
|
||||
run_push_ci:
|
||||
name: Trigger Push CI
|
||||
runs-on: ubuntu-22.04
|
||||
if: ${{ always() }}
|
||||
needs: build-docker-containers
|
||||
steps:
|
||||
- name: Trigger push CI via workflow_run
|
||||
run: echo "Trigger push CI via workflow_run"
|
||||
652
.github/workflows/self-push.yml
vendored
Normal file
652
.github/workflows/self-push.yml
vendored
Normal file
@ -0,0 +1,652 @@
|
||||
name: Self-hosted runner (push)
|
||||
|
||||
on:
|
||||
workflow_run:
|
||||
workflows: ["Self-hosted runner (push-caller)"]
|
||||
branches: ["main"]
|
||||
types: [completed]
|
||||
push:
|
||||
branches:
|
||||
- ci_*
|
||||
- ci-*
|
||||
paths:
|
||||
- "src/**"
|
||||
- "tests/**"
|
||||
- ".github/**"
|
||||
- "templates/**"
|
||||
- "utils/**"
|
||||
repository_dispatch:
|
||||
|
||||
env:
|
||||
HF_HOME: /mnt/cache
|
||||
TRANSFORMERS_IS_CI: yes
|
||||
OMP_NUM_THREADS: 8
|
||||
MKL_NUM_THREADS: 8
|
||||
PYTEST_TIMEOUT: 60
|
||||
TF_FORCE_GPU_ALLOW_GROWTH: true
|
||||
CUDA_VISIBLE_DEVICES: 0,1
|
||||
|
||||
jobs:
|
||||
setup:
|
||||
name: Setup
|
||||
strategy:
|
||||
matrix:
|
||||
machine_type: [aws-g5-4xlarge-cache, aws-g5-12xlarge-cache]
|
||||
runs-on:
|
||||
group: '${{ matrix.machine_type }}'
|
||||
container:
|
||||
image: huggingface/transformers-all-latest-gpu-push-ci
|
||||
options: --gpus all --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
|
||||
outputs:
|
||||
matrix: ${{ steps.set-matrix.outputs.matrix }}
|
||||
test_map: ${{ steps.set-matrix.outputs.test_map }}
|
||||
env:
|
||||
# `CI_BRANCH_PUSH`: The branch name from the push event
|
||||
# `CI_BRANCH_WORKFLOW_RUN`: The name of the branch on which this workflow is triggered by `workflow_run` event
|
||||
# `CI_SHA_PUSH`: The commit SHA from the push event
|
||||
# `CI_SHA_WORKFLOW_RUN`: The commit SHA that triggers this workflow by `workflow_run` event
|
||||
CI_BRANCH_PUSH: ${{ github.event.ref }}
|
||||
CI_BRANCH_WORKFLOW_RUN: ${{ github.event.workflow_run.head_branch }}
|
||||
CI_SHA_PUSH: ${{ github.event.head_commit.id }}
|
||||
CI_SHA_WORKFLOW_RUN: ${{ github.event.workflow_run.head_sha }}
|
||||
steps:
|
||||
# Necessary to get the correct branch name and commit SHA for `workflow_run` event
|
||||
# We also take into account the `push` event (we might want to test some changes in a branch)
|
||||
- name: Prepare custom environment variables
|
||||
shell: bash
|
||||
# `CI_BRANCH`: The non-empty branch name from the above two (one and only one of them is empty)
|
||||
# `CI_SHA`: The non-empty commit SHA from the above two (one and only one of them is empty)
|
||||
run: |
|
||||
CI_BRANCH_PUSH=${CI_BRANCH_PUSH/'refs/heads/'/''}
|
||||
echo $CI_BRANCH_PUSH
|
||||
echo $CI_BRANCH_WORKFLOW_RUN
|
||||
echo $CI_SHA_PUSH
|
||||
echo $CI_SHA_WORKFLOW_RUN
|
||||
[[ ! -z "$CI_BRANCH_PUSH" ]] && echo "CI_BRANCH=$CI_BRANCH_PUSH" >> $GITHUB_ENV || echo "CI_BRANCH=$CI_BRANCH_WORKFLOW_RUN" >> $GITHUB_ENV
|
||||
[[ ! -z "$CI_SHA_PUSH" ]] && echo "CI_SHA=$CI_SHA_PUSH" >> $GITHUB_ENV || echo "CI_SHA=$CI_SHA_WORKFLOW_RUN" >> $GITHUB_ENV
|
||||
|
||||
- name: print environment variables
|
||||
run: |
|
||||
echo "env.CI_BRANCH = ${{ env.CI_BRANCH }}"
|
||||
echo "env.CI_SHA = ${{ env.CI_SHA }}"
|
||||
|
||||
- name: Update clone using environment variables
|
||||
working-directory: /transformers
|
||||
run: |
|
||||
echo "original branch = $(git branch --show-current)"
|
||||
git fetch && git checkout ${{ env.CI_BRANCH }}
|
||||
echo "updated branch = $(git branch --show-current)"
|
||||
git checkout ${{ env.CI_SHA }}
|
||||
echo "log = $(git log -n 1)"
|
||||
|
||||
- name: Cleanup
|
||||
working-directory: /transformers
|
||||
run: |
|
||||
rm -rf tests/__pycache__
|
||||
rm -rf tests/models/__pycache__
|
||||
rm -rf reports
|
||||
|
||||
- name: Show installed libraries and their versions
|
||||
working-directory: /transformers
|
||||
run: pip freeze
|
||||
|
||||
- name: Fetch the tests to run
|
||||
working-directory: /transformers
|
||||
# TODO: add `git-python` in the docker images
|
||||
run: |
|
||||
pip install --upgrade git-python
|
||||
python3 utils/tests_fetcher.py --diff_with_last_commit | tee test_preparation.txt
|
||||
|
||||
- name: Report fetched tests
|
||||
uses: actions/upload-artifact@v4
|
||||
with:
|
||||
name: test_fetched
|
||||
path: /transformers/test_preparation.txt
|
||||
|
||||
- id: set-matrix
|
||||
name: Organize tests into models
|
||||
working-directory: /transformers
|
||||
# The `keys` is used as GitHub actions matrix for jobs, i.e. `models/bert`, `tokenization`, `pipeline`, etc.
|
||||
# The `test_map` is used to get the actual identified test files under each key.
|
||||
# If no test to run (so no `test_map.json` file), create a dummy map (empty matrix will fail)
|
||||
run: |
|
||||
if [ -f test_map.json ]; then
|
||||
keys=$(python3 -c 'import json; fp = open("test_map.json"); test_map = json.load(fp); fp.close(); d = list(test_map.keys()); print(d)')
|
||||
test_map=$(python3 -c 'import json; fp = open("test_map.json"); test_map = json.load(fp); fp.close(); print(test_map)')
|
||||
else
|
||||
keys=$(python3 -c 'keys = ["dummy"]; print(keys)')
|
||||
test_map=$(python3 -c 'test_map = {"dummy": []}; print(test_map)')
|
||||
fi
|
||||
echo $keys
|
||||
echo $test_map
|
||||
echo "matrix=$keys" >> $GITHUB_OUTPUT
|
||||
echo "test_map=$test_map" >> $GITHUB_OUTPUT
|
||||
|
||||
run_tests_single_gpu:
|
||||
name: Model tests
|
||||
needs: setup
|
||||
# `dummy` means there is no test to run
|
||||
if: contains(fromJson(needs.setup.outputs.matrix), 'dummy') != true
|
||||
strategy:
|
||||
fail-fast: false
|
||||
matrix:
|
||||
folders: ${{ fromJson(needs.setup.outputs.matrix) }}
|
||||
machine_type: [aws-g5-4xlarge-cache]
|
||||
runs-on:
|
||||
group: '${{ matrix.machine_type }}'
|
||||
container:
|
||||
image: huggingface/transformers-all-latest-gpu-push-ci
|
||||
options: --gpus all --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
|
||||
env:
|
||||
# For the meaning of these environment variables, see the job `Setup`
|
||||
CI_BRANCH_PUSH: ${{ github.event.ref }}
|
||||
CI_BRANCH_WORKFLOW_RUN: ${{ github.event.workflow_run.head_branch }}
|
||||
CI_SHA_PUSH: ${{ github.event.head_commit.id }}
|
||||
CI_SHA_WORKFLOW_RUN: ${{ github.event.workflow_run.head_sha }}
|
||||
steps:
|
||||
# Necessary to get the correct branch name and commit SHA for `workflow_run` event
|
||||
# We also take into account the `push` event (we might want to test some changes in a branch)
|
||||
- name: Prepare custom environment variables
|
||||
shell: bash
|
||||
# For the meaning of these environment variables, see the job `Setup`
|
||||
run: |
|
||||
CI_BRANCH_PUSH=${CI_BRANCH_PUSH/'refs/heads/'/''}
|
||||
echo $CI_BRANCH_PUSH
|
||||
echo $CI_BRANCH_WORKFLOW_RUN
|
||||
echo $CI_SHA_PUSH
|
||||
echo $CI_SHA_WORKFLOW_RUN
|
||||
[[ ! -z "$CI_BRANCH_PUSH" ]] && echo "CI_BRANCH=$CI_BRANCH_PUSH" >> $GITHUB_ENV || echo "CI_BRANCH=$CI_BRANCH_WORKFLOW_RUN" >> $GITHUB_ENV
|
||||
[[ ! -z "$CI_SHA_PUSH" ]] && echo "CI_SHA=$CI_SHA_PUSH" >> $GITHUB_ENV || echo "CI_SHA=$CI_SHA_WORKFLOW_RUN" >> $GITHUB_ENV
|
||||
|
||||
- name: print environment variables
|
||||
run: |
|
||||
echo "env.CI_BRANCH = ${{ env.CI_BRANCH }}"
|
||||
echo "env.CI_SHA = ${{ env.CI_SHA }}"
|
||||
|
||||
- name: Set `machine_type` for report and artifact names
|
||||
working-directory: /transformers
|
||||
shell: bash
|
||||
run: |
|
||||
echo "${{ matrix.machine_type }}"
|
||||
|
||||
if [ "${{ matrix.machine_type }}" = "aws-g5-4xlarge-cache" ]; then
|
||||
machine_type=single-gpu
|
||||
elif [ "${{ matrix.machine_type }}" = "aws-g5-12xlarge-cache" ]; then
|
||||
machine_type=multi-gpu
|
||||
else
|
||||
machine_type=${{ matrix.machine_type }}
|
||||
fi
|
||||
|
||||
echo "$machine_type"
|
||||
echo "machine_type=$machine_type" >> $GITHUB_ENV
|
||||
|
||||
- name: Update clone using environment variables
|
||||
working-directory: /transformers
|
||||
run: |
|
||||
echo "original branch = $(git branch --show-current)"
|
||||
git fetch && git checkout ${{ env.CI_BRANCH }}
|
||||
echo "updated branch = $(git branch --show-current)"
|
||||
git checkout ${{ env.CI_SHA }}
|
||||
echo "log = $(git log -n 1)"
|
||||
|
||||
- name: Reinstall transformers in edit mode (remove the one installed during docker image build)
|
||||
working-directory: /transformers
|
||||
run: python3 -m pip uninstall -y transformers && python3 -m pip install -e .
|
||||
|
||||
- name: Echo folder ${{ matrix.folders }}
|
||||
shell: bash
|
||||
# For folders like `models/bert`, set an env. var. (`matrix_folders`) to `models_bert`, which will be used to
|
||||
# set the artifact folder names (because the character `/` is not allowed).
|
||||
run: |
|
||||
echo "${{ matrix.folders }}"
|
||||
echo "${{ fromJson(needs.setup.outputs.test_map)[matrix.folders] }}"
|
||||
matrix_folders=${{ matrix.folders }}
|
||||
matrix_folders=${matrix_folders/'models/'/'models_'}
|
||||
echo "$matrix_folders"
|
||||
echo "matrix_folders=$matrix_folders" >> $GITHUB_ENV
|
||||
|
||||
- name: NVIDIA-SMI
|
||||
run: |
|
||||
nvidia-smi
|
||||
|
||||
- name: Environment
|
||||
working-directory: /transformers
|
||||
run: |
|
||||
python3 utils/print_env.py
|
||||
|
||||
- name: Show installed libraries and their versions
|
||||
working-directory: /transformers
|
||||
run: pip freeze
|
||||
|
||||
- name: Run all non-slow selected tests on GPU
|
||||
working-directory: /transformers
|
||||
run: |
|
||||
python3 -m pytest -n 2 --dist=loadfile -v --make-reports=${{ env.machine_type }}_tests_gpu_${{ matrix.folders }} ${{ fromJson(needs.setup.outputs.test_map)[matrix.folders] }}
|
||||
|
||||
- name: Failure short reports
|
||||
if: ${{ failure() }}
|
||||
continue-on-error: true
|
||||
run: cat /transformers/reports/${{ env.machine_type }}_tests_gpu_${{ matrix.folders }}/failures_short.txt
|
||||
|
||||
- name: "Test suite reports artifacts: ${{ env.machine_type }}_run_all_tests_gpu_${{ env.matrix_folders }}_test_reports"
|
||||
if: ${{ always() }}
|
||||
uses: actions/upload-artifact@v4
|
||||
with:
|
||||
name: ${{ env.machine_type }}_run_all_tests_gpu_${{ env.matrix_folders }}_test_reports
|
||||
path: /transformers/reports/${{ env.machine_type }}_tests_gpu_${{ matrix.folders }}
|
||||
|
||||
run_tests_multi_gpu:
|
||||
name: Model tests
|
||||
needs: setup
|
||||
# `dummy` means there is no test to run
|
||||
if: contains(fromJson(needs.setup.outputs.matrix), 'dummy') != true
|
||||
strategy:
|
||||
fail-fast: false
|
||||
matrix:
|
||||
folders: ${{ fromJson(needs.setup.outputs.matrix) }}
|
||||
machine_type: [aws-g5-12xlarge-cache]
|
||||
runs-on:
|
||||
group: '${{ matrix.machine_type }}'
|
||||
container:
|
||||
image: huggingface/transformers-all-latest-gpu-push-ci
|
||||
options: --gpus all --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
|
||||
env:
|
||||
# For the meaning of these environment variables, see the job `Setup`
|
||||
CI_BRANCH_PUSH: ${{ github.event.ref }}
|
||||
CI_BRANCH_WORKFLOW_RUN: ${{ github.event.workflow_run.head_branch }}
|
||||
CI_SHA_PUSH: ${{ github.event.head_commit.id }}
|
||||
CI_SHA_WORKFLOW_RUN: ${{ github.event.workflow_run.head_sha }}
|
||||
steps:
|
||||
# Necessary to get the correct branch name and commit SHA for `workflow_run` event
|
||||
# We also take into account the `push` event (we might want to test some changes in a branch)
|
||||
- name: Prepare custom environment variables
|
||||
shell: bash
|
||||
# For the meaning of these environment variables, see the job `Setup`
|
||||
run: |
|
||||
CI_BRANCH_PUSH=${CI_BRANCH_PUSH/'refs/heads/'/''}
|
||||
echo $CI_BRANCH_PUSH
|
||||
echo $CI_BRANCH_WORKFLOW_RUN
|
||||
echo $CI_SHA_PUSH
|
||||
echo $CI_SHA_WORKFLOW_RUN
|
||||
[[ ! -z "$CI_BRANCH_PUSH" ]] && echo "CI_BRANCH=$CI_BRANCH_PUSH" >> $GITHUB_ENV || echo "CI_BRANCH=$CI_BRANCH_WORKFLOW_RUN" >> $GITHUB_ENV
|
||||
[[ ! -z "$CI_SHA_PUSH" ]] && echo "CI_SHA=$CI_SHA_PUSH" >> $GITHUB_ENV || echo "CI_SHA=$CI_SHA_WORKFLOW_RUN" >> $GITHUB_ENV
|
||||
|
||||
- name: print environment variables
|
||||
run: |
|
||||
echo "env.CI_BRANCH = ${{ env.CI_BRANCH }}"
|
||||
echo "env.CI_SHA = ${{ env.CI_SHA }}"
|
||||
|
||||
- name: Set `machine_type` for report and artifact names
|
||||
working-directory: /transformers
|
||||
shell: bash
|
||||
run: |
|
||||
echo "${{ matrix.machine_type }}"
|
||||
|
||||
if [ "${{ matrix.machine_type }}" = "aws-g5-4xlarge-cache" ]; then
|
||||
machine_type=single-gpu
|
||||
elif [ "${{ matrix.machine_type }}" = "aws-g5-12xlarge-cache" ]; then
|
||||
machine_type=multi-gpu
|
||||
else
|
||||
machine_type=${{ matrix.machine_type }}
|
||||
fi
|
||||
|
||||
echo "$machine_type"
|
||||
echo "machine_type=$machine_type" >> $GITHUB_ENV
|
||||
|
||||
- name: Update clone using environment variables
|
||||
working-directory: /transformers
|
||||
run: |
|
||||
echo "original branch = $(git branch --show-current)"
|
||||
git fetch && git checkout ${{ env.CI_BRANCH }}
|
||||
echo "updated branch = $(git branch --show-current)"
|
||||
git checkout ${{ env.CI_SHA }}
|
||||
echo "log = $(git log -n 1)"
|
||||
|
||||
- name: Reinstall transformers in edit mode (remove the one installed during docker image build)
|
||||
working-directory: /transformers
|
||||
run: python3 -m pip uninstall -y transformers && python3 -m pip install -e .
|
||||
|
||||
- name: Echo folder ${{ matrix.folders }}
|
||||
shell: bash
|
||||
# For folders like `models/bert`, set an env. var. (`matrix_folders`) to `models_bert`, which will be used to
|
||||
# set the artifact folder names (because the character `/` is not allowed).
|
||||
run: |
|
||||
echo "${{ matrix.folders }}"
|
||||
echo "${{ fromJson(needs.setup.outputs.test_map)[matrix.folders] }}"
|
||||
matrix_folders=${{ matrix.folders }}
|
||||
matrix_folders=${matrix_folders/'models/'/'models_'}
|
||||
echo "$matrix_folders"
|
||||
echo "matrix_folders=$matrix_folders" >> $GITHUB_ENV
|
||||
|
||||
- name: NVIDIA-SMI
|
||||
run: |
|
||||
nvidia-smi
|
||||
|
||||
- name: Environment
|
||||
working-directory: /transformers
|
||||
run: |
|
||||
python3 utils/print_env.py
|
||||
|
||||
- name: Show installed libraries and their versions
|
||||
working-directory: /transformers
|
||||
run: pip freeze
|
||||
|
||||
- name: Run all non-slow selected tests on GPU
|
||||
env:
|
||||
MKL_SERVICE_FORCE_INTEL: 1
|
||||
working-directory: /transformers
|
||||
run: |
|
||||
python3 -m pytest -n 2 --dist=loadfile -v --make-reports=${{ env.machine_type }}_tests_gpu_${{ matrix.folders }} ${{ fromJson(needs.setup.outputs.test_map)[matrix.folders] }}
|
||||
|
||||
- name: Failure short reports
|
||||
if: ${{ failure() }}
|
||||
continue-on-error: true
|
||||
run: cat /transformers/reports/${{ env.machine_type }}_tests_gpu_${{ matrix.folders }}/failures_short.txt
|
||||
|
||||
- name: "Test suite reports artifacts: ${{ env.machine_type }}_run_all_tests_gpu_${{ env.matrix_folders }}_test_reports"
|
||||
if: ${{ always() }}
|
||||
uses: actions/upload-artifact@v4
|
||||
with:
|
||||
name: ${{ env.machine_type }}_run_all_tests_gpu_${{ env.matrix_folders }}_test_reports
|
||||
path: /transformers/reports/${{ env.machine_type }}_tests_gpu_${{ matrix.folders }}
|
||||
|
||||
run_tests_torch_cuda_extensions_single_gpu:
|
||||
name: Torch CUDA extension tests
|
||||
needs: setup
|
||||
if: contains(fromJson(needs.setup.outputs.matrix), 'deepspeed') || contains(fromJson(needs.setup.outputs.matrix), 'extended')
|
||||
strategy:
|
||||
fail-fast: false
|
||||
matrix:
|
||||
machine_type: [aws-g5-4xlarge-cache]
|
||||
runs-on:
|
||||
group: '${{ matrix.machine_type }}'
|
||||
container:
|
||||
image: huggingface/transformers-pytorch-deepspeed-latest-gpu-push-ci
|
||||
options: --gpus all --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
|
||||
env:
|
||||
# For the meaning of these environment variables, see the job `Setup`
|
||||
CI_BRANCH_PUSH: ${{ github.event.ref }}
|
||||
CI_BRANCH_WORKFLOW_RUN: ${{ github.event.workflow_run.head_branch }}
|
||||
CI_SHA_PUSH: ${{ github.event.head_commit.id }}
|
||||
CI_SHA_WORKFLOW_RUN: ${{ github.event.workflow_run.head_sha }}
|
||||
steps:
|
||||
# Necessary to get the correct branch name and commit SHA for `workflow_run` event
|
||||
# We also take into account the `push` event (we might want to test some changes in a branch)
|
||||
- name: Prepare custom environment variables
|
||||
shell: bash
|
||||
# For the meaning of these environment variables, see the job `Setup`
|
||||
run: |
|
||||
CI_BRANCH_PUSH=${CI_BRANCH_PUSH/'refs/heads/'/''}
|
||||
echo $CI_BRANCH_PUSH
|
||||
echo $CI_BRANCH_WORKFLOW_RUN
|
||||
echo $CI_SHA_PUSH
|
||||
echo $CI_SHA_WORKFLOW_RUN
|
||||
[[ ! -z "$CI_BRANCH_PUSH" ]] && echo "CI_BRANCH=$CI_BRANCH_PUSH" >> $GITHUB_ENV || echo "CI_BRANCH=$CI_BRANCH_WORKFLOW_RUN" >> $GITHUB_ENV
|
||||
[[ ! -z "$CI_SHA_PUSH" ]] && echo "CI_SHA=$CI_SHA_PUSH" >> $GITHUB_ENV || echo "CI_SHA=$CI_SHA_WORKFLOW_RUN" >> $GITHUB_ENV
|
||||
|
||||
- name: print environment variables
|
||||
run: |
|
||||
echo "env.CI_BRANCH = ${{ env.CI_BRANCH }}"
|
||||
echo "env.CI_SHA = ${{ env.CI_SHA }}"
|
||||
|
||||
- name: Set `machine_type` for report and artifact names
|
||||
working-directory: /workspace/transformers
|
||||
shell: bash
|
||||
run: |
|
||||
echo "${{ matrix.machine_type }}"
|
||||
|
||||
if [ "${{ matrix.machine_type }}" = "aws-g5-4xlarge-cache" ]; then
|
||||
machine_type=single-gpu
|
||||
elif [ "${{ matrix.machine_type }}" = "aws-g5-12xlarge-cache" ]; then
|
||||
machine_type=multi-gpu
|
||||
else
|
||||
machine_type=${{ matrix.machine_type }}
|
||||
fi
|
||||
|
||||
echo "$machine_type"
|
||||
echo "machine_type=$machine_type" >> $GITHUB_ENV
|
||||
|
||||
- name: Update clone using environment variables
|
||||
working-directory: /workspace/transformers
|
||||
run: |
|
||||
echo "original branch = $(git branch --show-current)"
|
||||
git fetch && git checkout ${{ env.CI_BRANCH }}
|
||||
echo "updated branch = $(git branch --show-current)"
|
||||
git checkout ${{ env.CI_SHA }}
|
||||
echo "log = $(git log -n 1)"
|
||||
|
||||
- name: Reinstall transformers in edit mode (remove the one installed during docker image build)
|
||||
working-directory: /workspace/transformers
|
||||
run: python3 -m pip uninstall -y transformers && python3 -m pip install -e .
|
||||
|
||||
- name: Remove cached torch extensions
|
||||
run: rm -rf /github/home/.cache/torch_extensions/
|
||||
|
||||
# To avoid unknown test failures
|
||||
- name: Pre build DeepSpeed *again*
|
||||
working-directory: /workspace
|
||||
run: |
|
||||
python3 -m pip uninstall -y deepspeed
|
||||
DS_BUILD_CPU_ADAM=1 DS_BUILD_FUSED_ADAM=1 python3 -m pip install deepspeed --global-option="build_ext" --global-option="-j8" --no-cache -v --disable-pip-version-check
|
||||
|
||||
- name: NVIDIA-SMI
|
||||
run: |
|
||||
nvidia-smi
|
||||
|
||||
- name: Environment
|
||||
working-directory: /workspace/transformers
|
||||
run: |
|
||||
python utils/print_env.py
|
||||
|
||||
- name: Show installed libraries and their versions
|
||||
working-directory: /workspace/transformers
|
||||
run: pip freeze
|
||||
|
||||
- name: Run all non-slow selected tests on GPU
|
||||
working-directory: /workspace/transformers
|
||||
# TODO: Here we pass all tests in the 2 folders for simplicity. It's better to pass only the identified tests.
|
||||
run: |
|
||||
python -m pytest -n 1 --dist=loadfile -v --make-reports=${{ env.machine_type }}_run_torch_cuda_extensions_gpu_test_reports tests/deepspeed tests/extended
|
||||
|
||||
- name: Failure short reports
|
||||
if: ${{ failure() }}
|
||||
continue-on-error: true
|
||||
run: cat /workspace/transformers/reports/${{ env.machine_type }}_run_torch_cuda_extensions_gpu_test_reports/failures_short.txt
|
||||
|
||||
- name: "Test suite reports artifacts: ${{ env.machine_type }}_run_torch_cuda_extensions_gpu_test_reports"
|
||||
if: ${{ always() }}
|
||||
uses: actions/upload-artifact@v4
|
||||
with:
|
||||
name: ${{ env.machine_type }}_run_torch_cuda_extensions_gpu_test_reports
|
||||
path: /workspace/transformers/reports/${{ env.machine_type }}_run_torch_cuda_extensions_gpu_test_reports
|
||||
|
||||
run_tests_torch_cuda_extensions_multi_gpu:
|
||||
name: Torch CUDA extension tests
|
||||
needs: setup
|
||||
if: contains(fromJson(needs.setup.outputs.matrix), 'deepspeed') || contains(fromJson(needs.setup.outputs.matrix), 'extended')
|
||||
strategy:
|
||||
fail-fast: false
|
||||
matrix:
|
||||
machine_type: [aws-g5-12xlarge-cache]
|
||||
runs-on:
|
||||
group: '${{ matrix.machine_type }}'
|
||||
container:
|
||||
image: huggingface/transformers-pytorch-deepspeed-latest-gpu-push-ci
|
||||
options: --gpus all --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
|
||||
env:
|
||||
# For the meaning of these environment variables, see the job `Setup`
|
||||
CI_BRANCH_PUSH: ${{ github.event.ref }}
|
||||
CI_BRANCH_WORKFLOW_RUN: ${{ github.event.workflow_run.head_branch }}
|
||||
CI_SHA_PUSH: ${{ github.event.head_commit.id }}
|
||||
CI_SHA_WORKFLOW_RUN: ${{ github.event.workflow_run.head_sha }}
|
||||
steps:
|
||||
# Necessary to get the correct branch name and commit SHA for `workflow_run` event
|
||||
# We also take into account the `push` event (we might want to test some changes in a branch)
|
||||
- name: Prepare custom environment variables
|
||||
shell: bash
|
||||
# For the meaning of these environment variables, see the job `Setup`
|
||||
run: |
|
||||
CI_BRANCH_PUSH=${CI_BRANCH_PUSH/'refs/heads/'/''}
|
||||
echo $CI_BRANCH_PUSH
|
||||
echo $CI_BRANCH_WORKFLOW_RUN
|
||||
echo $CI_SHA_PUSH
|
||||
echo $CI_SHA_WORKFLOW_RUN
|
||||
[[ ! -z "$CI_BRANCH_PUSH" ]] && echo "CI_BRANCH=$CI_BRANCH_PUSH" >> $GITHUB_ENV || echo "CI_BRANCH=$CI_BRANCH_WORKFLOW_RUN" >> $GITHUB_ENV
|
||||
[[ ! -z "$CI_SHA_PUSH" ]] && echo "CI_SHA=$CI_SHA_PUSH" >> $GITHUB_ENV || echo "CI_SHA=$CI_SHA_WORKFLOW_RUN" >> $GITHUB_ENV
|
||||
|
||||
- name: print environment variables
|
||||
run: |
|
||||
echo "env.CI_BRANCH = ${{ env.CI_BRANCH }}"
|
||||
echo "env.CI_SHA = ${{ env.CI_SHA }}"
|
||||
|
||||
- name: Set `machine_type` for report and artifact names
|
||||
working-directory: /workspace/transformers
|
||||
shell: bash
|
||||
run: |
|
||||
echo "${{ matrix.machine_type }}"
|
||||
|
||||
if [ "${{ matrix.machine_type }}" = "aws-g5-4xlarge-cache" ]; then
|
||||
machine_type=single-gpu
|
||||
elif [ "${{ matrix.machine_type }}" = "aws-g5-12xlarge-cache" ]; then
|
||||
machine_type=multi-gpu
|
||||
else
|
||||
machine_type=${{ matrix.machine_type }}
|
||||
fi
|
||||
|
||||
echo "$machine_type"
|
||||
echo "machine_type=$machine_type" >> $GITHUB_ENV
|
||||
|
||||
- name: Update clone using environment variables
|
||||
working-directory: /workspace/transformers
|
||||
run: |
|
||||
echo "original branch = $(git branch --show-current)"
|
||||
git fetch && git checkout ${{ env.CI_BRANCH }}
|
||||
echo "updated branch = $(git branch --show-current)"
|
||||
git checkout ${{ env.CI_SHA }}
|
||||
echo "log = $(git log -n 1)"
|
||||
|
||||
- name: Reinstall transformers in edit mode (remove the one installed during docker image build)
|
||||
working-directory: /workspace/transformers
|
||||
run: python3 -m pip uninstall -y transformers && python3 -m pip install -e .
|
||||
|
||||
- name: Remove cached torch extensions
|
||||
run: rm -rf /github/home/.cache/torch_extensions/
|
||||
|
||||
# To avoid unknown test failures
|
||||
- name: Pre build DeepSpeed *again*
|
||||
working-directory: /workspace
|
||||
run: |
|
||||
python3 -m pip uninstall -y deepspeed
|
||||
DS_BUILD_CPU_ADAM=1 DS_BUILD_FUSED_ADAM=1 python3 -m pip install deepspeed --global-option="build_ext" --global-option="-j8" --no-cache -v --disable-pip-version-check
|
||||
|
||||
- name: NVIDIA-SMI
|
||||
run: |
|
||||
nvidia-smi
|
||||
|
||||
- name: Environment
|
||||
working-directory: /workspace/transformers
|
||||
run: |
|
||||
python utils/print_env.py
|
||||
|
||||
- name: Show installed libraries and their versions
|
||||
working-directory: /workspace/transformers
|
||||
run: pip freeze
|
||||
|
||||
- name: Run all non-slow selected tests on GPU
|
||||
working-directory: /workspace/transformers
|
||||
# TODO: Here we pass all tests in the 2 folders for simplicity. It's better to pass only the identified tests.
|
||||
run: |
|
||||
python -m pytest -n 1 --dist=loadfile -v --make-reports=${{ env.machine_type }}_run_torch_cuda_extensions_gpu_test_reports tests/deepspeed tests/extended
|
||||
|
||||
- name: Failure short reports
|
||||
if: ${{ failure() }}
|
||||
continue-on-error: true
|
||||
run: cat /workspace/transformers/reports/${{ env.machine_type }}_run_torch_cuda_extensions_gpu_test_reports/failures_short.txt
|
||||
|
||||
- name: "Test suite reports artifacts: ${{ env.machine_type }}_run_torch_cuda_extensions_gpu_test_reports"
|
||||
if: ${{ always() }}
|
||||
uses: actions/upload-artifact@v4
|
||||
with:
|
||||
name: ${{ env.machine_type }}_run_torch_cuda_extensions_gpu_test_reports
|
||||
path: /workspace/transformers/reports/${{ env.machine_type }}_run_torch_cuda_extensions_gpu_test_reports
|
||||
|
||||
send_results:
|
||||
name: Send results to webhook
|
||||
runs-on: ubuntu-22.04
|
||||
if: always()
|
||||
needs: [
|
||||
setup,
|
||||
run_tests_single_gpu,
|
||||
run_tests_multi_gpu,
|
||||
run_tests_torch_cuda_extensions_single_gpu,
|
||||
run_tests_torch_cuda_extensions_multi_gpu
|
||||
]
|
||||
env:
|
||||
# For the meaning of these environment variables, see the job `Setup`
|
||||
CI_BRANCH_PUSH: ${{ github.event.ref }}
|
||||
CI_BRANCH_WORKFLOW_RUN: ${{ github.event.workflow_run.head_branch }}
|
||||
CI_SHA_PUSH: ${{ github.event.head_commit.id }}
|
||||
CI_SHA_WORKFLOW_RUN: ${{ github.event.workflow_run.head_sha }}
|
||||
steps:
|
||||
- name: Preliminary job status
|
||||
shell: bash
|
||||
# For the meaning of these environment variables, see the job `Setup`
|
||||
run: |
|
||||
echo "Setup status: ${{ needs.setup.result }}"
|
||||
|
||||
# Necessary to get the correct branch name and commit SHA for `workflow_run` event
|
||||
# We also take into account the `push` event (we might want to test some changes in a branch)
|
||||
- name: Prepare custom environment variables
|
||||
shell: bash
|
||||
# For the meaning of these environment variables, see the job `Setup`
|
||||
run: |
|
||||
CI_BRANCH_PUSH=${CI_BRANCH_PUSH/'refs/heads/'/''}
|
||||
echo $CI_BRANCH_PUSH
|
||||
echo $CI_BRANCH_WORKFLOW_RUN
|
||||
echo $CI_SHA_PUSH
|
||||
echo $CI_SHA_WORKFLOW_RUN
|
||||
[[ ! -z "$CI_BRANCH_PUSH" ]] && echo "CI_BRANCH=$CI_BRANCH_PUSH" >> $GITHUB_ENV || echo "CI_BRANCH=$CI_BRANCH_WORKFLOW_RUN" >> $GITHUB_ENV
|
||||
[[ ! -z "$CI_SHA_PUSH" ]] && echo "CI_SHA=$CI_SHA_PUSH" >> $GITHUB_ENV || echo "CI_SHA=$CI_SHA_WORKFLOW_RUN" >> $GITHUB_ENV
|
||||
|
||||
- name: print environment variables
|
||||
run: |
|
||||
echo "env.CI_BRANCH = ${{ env.CI_BRANCH }}"
|
||||
echo "env.CI_SHA = ${{ env.CI_SHA }}"
|
||||
|
||||
- uses: actions/checkout@v4
|
||||
# To avoid failure when multiple commits are merged into `main` in a short period of time.
|
||||
# Checking out to an old commit beyond the fetch depth will get an error `fatal: reference is not a tree: ...
|
||||
# (Only required for `workflow_run` event, where we get the latest HEAD on `main` instead of the event commit)
|
||||
with:
|
||||
fetch-depth: 20
|
||||
|
||||
- name: Update clone using environment variables
|
||||
run: |
|
||||
echo "original branch = $(git branch --show-current)"
|
||||
git fetch && git checkout ${{ env.CI_BRANCH }}
|
||||
echo "updated branch = $(git branch --show-current)"
|
||||
git checkout ${{ env.CI_SHA }}
|
||||
echo "log = $(git log -n 1)"
|
||||
|
||||
- uses: actions/download-artifact@v4
|
||||
- name: Send message to Slack
|
||||
env:
|
||||
CI_SLACK_BOT_TOKEN: ${{ secrets.CI_SLACK_BOT_TOKEN }}
|
||||
CI_SLACK_CHANNEL_ID: ${{ secrets.CI_SLACK_CHANNEL_ID }}
|
||||
CI_SLACK_CHANNEL_ID_DAILY: ${{ secrets.CI_SLACK_CHANNEL_ID_DAILY }}
|
||||
CI_SLACK_CHANNEL_DUMMY_TESTS: ${{ secrets.CI_SLACK_CHANNEL_DUMMY_TESTS }}
|
||||
CI_SLACK_REPORT_CHANNEL_ID: ${{ secrets.CI_SLACK_CHANNEL_ID }}
|
||||
ACCESS_REPO_INFO_TOKEN: ${{ secrets.ACCESS_REPO_INFO_TOKEN }}
|
||||
CI_EVENT: push
|
||||
CI_TITLE_PUSH: ${{ github.event.head_commit.message }}
|
||||
CI_TITLE_WORKFLOW_RUN: ${{ github.event.workflow_run.head_commit.message }}
|
||||
CI_SHA: ${{ env.CI_SHA }}
|
||||
SETUP_STATUS: ${{ needs.setup.result }}
|
||||
|
||||
# We pass `needs.setup.outputs.matrix` as the argument. A processing in `notification_service.py` to change
|
||||
# `models/bert` to `models_bert` is required, as the artifact names use `_` instead of `/`.
|
||||
run: |
|
||||
pip install huggingface_hub
|
||||
pip install slack_sdk
|
||||
pip show slack_sdk
|
||||
python utils/notification_service.py "${{ needs.setup.outputs.matrix }}"
|
||||
@ -2,7 +2,7 @@ name: Self-hosted runner (AMD scheduled CI caller)
|
||||
|
||||
on:
|
||||
schedule:
|
||||
- cron: "17 5 * * *"
|
||||
- cron: "17 2 * * *"
|
||||
|
||||
jobs:
|
||||
run_scheduled_amd_ci:
|
||||
|
||||
@ -21,7 +21,7 @@ jobs:
|
||||
job: run_models_gpu
|
||||
slack_report_channel: "#amd-hf-ci"
|
||||
runner_group: hfc-amd-mi355
|
||||
docker: huggingface/transformers-pytorch-amd-gpu
|
||||
docker: huggingface/testing-rocm7.0-preview
|
||||
ci_event: Scheduled CI (AMD) - mi355
|
||||
report_repo_id: hf-transformers-bot/transformers-ci-dummy
|
||||
secrets: inherit
|
||||
@ -33,7 +33,7 @@ jobs:
|
||||
job: run_pipelines_torch_gpu
|
||||
slack_report_channel: "#amd-hf-ci"
|
||||
runner_group: hfc-amd-mi355
|
||||
docker: huggingface/transformers-pytorch-amd-gpu
|
||||
docker: huggingface/testing-rocm7.0-preview
|
||||
ci_event: Scheduled CI (AMD) - mi355
|
||||
report_repo_id: hf-transformers-bot/transformers-ci-dummy
|
||||
secrets: inherit
|
||||
@ -45,7 +45,7 @@ jobs:
|
||||
job: run_examples_gpu
|
||||
slack_report_channel: "#amd-hf-ci"
|
||||
runner_group: hfc-amd-mi355
|
||||
docker: huggingface/transformers-pytorch-amd-gpu
|
||||
docker: huggingface/testing-rocm7.0-preview
|
||||
ci_event: Scheduled CI (AMD) - mi355
|
||||
report_repo_id: hf-transformers-bot/transformers-ci-dummy
|
||||
secrets: inherit
|
||||
|
||||
21
.github/workflows/self-scheduled-caller.yml
vendored
21
.github/workflows/self-scheduled-caller.yml
vendored
@ -33,13 +33,10 @@ jobs:
|
||||
runs-on: ubuntu-22.04
|
||||
steps:
|
||||
- name: Setup
|
||||
env:
|
||||
prev_workflow_run_id: ${{ inputs.prev_workflow_run_id || env.prev_workflow_run_id }}
|
||||
other_workflow_run_id: ${{ inputs.other_workflow_run_id || env.other_workflow_run_id }}
|
||||
run: |
|
||||
mkdir "setup_values"
|
||||
echo "$prev_workflow_run_id" > "setup_values/prev_workflow_run_id.txt"
|
||||
echo "$other_workflow_run_id" > "setup_values/other_workflow_run_id.txt"
|
||||
echo "${{ inputs.prev_workflow_run_id || env.prev_workflow_run_id }}" > "setup_values/prev_workflow_run_id.txt"
|
||||
echo "${{ inputs.other_workflow_run_id || env.other_workflow_run_id }}" > "setup_values/other_workflow_run_id.txt"
|
||||
|
||||
- name: Upload artifacts
|
||||
uses: actions/upload-artifact@v4
|
||||
@ -66,7 +63,7 @@ jobs:
|
||||
with:
|
||||
job: run_pipelines_torch_gpu
|
||||
slack_report_channel: "#transformers-ci-daily-pipeline-torch"
|
||||
docker: huggingface/transformers-all-latest-gpu
|
||||
docker: huggingface/transformers-pytorch-gpu
|
||||
ci_event: Daily CI
|
||||
report_repo_id: hf-internal-testing/transformers_daily_ci
|
||||
commit_sha: ${{ github.sha }}
|
||||
@ -121,15 +118,3 @@ jobs:
|
||||
report_repo_id: hf-internal-testing/transformers_daily_ci
|
||||
commit_sha: ${{ github.sha }}
|
||||
secrets: inherit
|
||||
|
||||
kernels-ci:
|
||||
name: Kernels CI
|
||||
uses: ./.github/workflows/self-scheduled.yml
|
||||
with:
|
||||
job: run_kernels_gpu
|
||||
slack_report_channel: "#transformers-ci-daily-kernels"
|
||||
docker: huggingface/transformers-all-latest-gpu
|
||||
ci_event: Daily CI
|
||||
report_repo_id: hf-internal-testing/transformers_daily_ci
|
||||
commit_sha: ${{ github.sha }}
|
||||
secrets: inherit
|
||||
@ -1,60 +0,0 @@
|
||||
name: Nvidia CI - Flash Attn
|
||||
|
||||
on:
|
||||
repository_dispatch:
|
||||
schedule:
|
||||
- cron: "17 2 * * *"
|
||||
push:
|
||||
branches:
|
||||
- run_nvidia_ci_flash_attn*
|
||||
workflow_dispatch:
|
||||
inputs:
|
||||
prev_workflow_run_id:
|
||||
description: 'previous workflow run id to compare'
|
||||
type: string
|
||||
required: false
|
||||
default: ""
|
||||
other_workflow_run_id:
|
||||
description: 'other workflow run id to compare'
|
||||
type: string
|
||||
required: false
|
||||
default: ""
|
||||
|
||||
|
||||
# Used for `push` to easily modify the target workflow runs to compare against
|
||||
env:
|
||||
prev_workflow_run_id: ""
|
||||
other_workflow_run_id: ""
|
||||
|
||||
|
||||
jobs:
|
||||
setup:
|
||||
name: Setup
|
||||
runs-on: ubuntu-22.04
|
||||
steps:
|
||||
- name: Setup
|
||||
run: |
|
||||
mkdir "setup_values"
|
||||
echo "${{ inputs.prev_workflow_run_id || env.prev_workflow_run_id }}" > "setup_values/prev_workflow_run_id.txt"
|
||||
echo "${{ inputs.other_workflow_run_id || env.other_workflow_run_id }}" > "setup_values/other_workflow_run_id.txt"
|
||||
|
||||
- name: Upload artifacts
|
||||
uses: actions/upload-artifact@v4
|
||||
with:
|
||||
name: setup_values
|
||||
path: setup_values
|
||||
|
||||
|
||||
model-ci:
|
||||
name: Model CI
|
||||
uses: ./.github/workflows/self-scheduled.yml
|
||||
with:
|
||||
job: run_models_gpu
|
||||
slack_report_channel: "#transformers-ci-flash-attn"
|
||||
docker: huggingface/transformers-all-latest-gpu:flash-attn
|
||||
ci_event: Daily CI
|
||||
runner_type: "a10"
|
||||
report_repo_id: hf-internal-testing/transformers_flash_attn_ci
|
||||
commit_sha: ${{ github.sha }}
|
||||
pytest_marker: "flash_attn_test or flash_attn_3_test"
|
||||
secrets: inherit
|
||||
@ -26,6 +26,7 @@ env:
|
||||
TRANSFORMERS_IS_CI: yes
|
||||
PT_ENABLE_INT64_SUPPORT: 1
|
||||
HF_HUB_READ_TOKEN: ${{ secrets.HF_HUB_READ_TOKEN }}
|
||||
SIGOPT_API_TOKEN: ${{ secrets.SIGOPT_API_TOKEN }}
|
||||
HF_HOME: /mnt/cache/.cache/huggingface
|
||||
|
||||
jobs:
|
||||
|
||||
221
.github/workflows/self-scheduled.yml
vendored
221
.github/workflows/self-scheduled.yml
vendored
@ -34,20 +34,10 @@ on:
|
||||
runner_type:
|
||||
required: false
|
||||
type: string
|
||||
subdirs:
|
||||
models:
|
||||
default: ""
|
||||
required: false
|
||||
type: string
|
||||
pytest_marker:
|
||||
required: false
|
||||
type: string
|
||||
pr_number:
|
||||
required: false
|
||||
type: string
|
||||
outputs:
|
||||
report:
|
||||
description: "Content of the report of new failures"
|
||||
value: ${{ jobs.check_new_failures.outputs.report }}
|
||||
|
||||
env:
|
||||
HF_HOME: /mnt/cache
|
||||
@ -58,8 +48,10 @@ env:
|
||||
# For gated repositories, we still need to agree to share information on the Hub repo. page in order to get access.
|
||||
# This token is created under the bot `hf-transformers-bot`.
|
||||
HF_HUB_READ_TOKEN: ${{ secrets.HF_HUB_READ_TOKEN }}
|
||||
SIGOPT_API_TOKEN: ${{ secrets.SIGOPT_API_TOKEN }}
|
||||
TF_FORCE_GPU_ALLOW_GROWTH: true
|
||||
CUDA_VISIBLE_DEVICES: 0,1
|
||||
NUM_SLICES: 2
|
||||
|
||||
jobs:
|
||||
setup:
|
||||
@ -80,11 +72,8 @@ jobs:
|
||||
steps:
|
||||
- name: Update clone
|
||||
working-directory: /transformers
|
||||
env:
|
||||
commit_sha: ${{ inputs.commit_sha || github.sha }}
|
||||
run: |
|
||||
git fetch origin $commit_sha
|
||||
git fetch && git checkout $commit_sha
|
||||
git fetch && git checkout ${{ inputs.commit_sha || github.sha }}
|
||||
|
||||
- name: Cleanup
|
||||
working-directory: /transformers
|
||||
@ -101,17 +90,11 @@ jobs:
|
||||
if: contains(fromJSON('["run_models_gpu", "run_trainer_and_fsdp_gpu"]'), inputs.job)
|
||||
name: Identify models to test
|
||||
working-directory: /transformers/tests
|
||||
env:
|
||||
job: ${{ inputs.job }}
|
||||
subdirs: ${{ inputs.subdirs }}
|
||||
NUM_SLICES: 2
|
||||
run: |
|
||||
if [ "$job" = "run_models_gpu" ]; then
|
||||
python3 ../utils/split_model_tests.py --subdirs "$subdirs" --num_splits "$NUM_SLICES" > folder_slices.txt
|
||||
echo "folder_slices=$(cat folder_slices.txt)" >> $GITHUB_OUTPUT
|
||||
python3 -c "import ast; folder_slices = ast.literal_eval(open('folder_slices.txt').read()); open('slice_ids.txt', 'w').write(str(list(range(len(folder_slices)))))"
|
||||
echo "slice_ids=$(cat slice_ids.txt)" >> $GITHUB_OUTPUT
|
||||
elif [ "$job" = "run_trainer_and_fsdp_gpu" ]; then
|
||||
if [ "${{ inputs.job }}" = "run_models_gpu" ]; then
|
||||
echo "folder_slices=$(python3 ../utils/split_model_tests.py --models '${{ inputs.models }}' --num_splits ${{ env.NUM_SLICES }})" >> $GITHUB_OUTPUT
|
||||
echo "slice_ids=$(python3 -c 'd = list(range(${{ env.NUM_SLICES }})); print(d)')" >> $GITHUB_OUTPUT
|
||||
elif [ "${{ inputs.job }}" = "run_trainer_and_fsdp_gpu" ]; then
|
||||
echo "folder_slices=[['trainer'], ['fsdp']]" >> $GITHUB_OUTPUT
|
||||
echo "slice_ids=[0, 1]" >> $GITHUB_OUTPUT
|
||||
fi
|
||||
@ -120,10 +103,8 @@ jobs:
|
||||
if: ${{ inputs.job == 'run_quantization_torch_gpu' }}
|
||||
name: Identify quantization method to test
|
||||
working-directory: /transformers/tests
|
||||
env:
|
||||
subdirs: ${{ inputs.subdirs || 'None' }}
|
||||
run: |
|
||||
echo "quantization_matrix=$(python3 -c 'import ast; import os; tests = os.getcwd(); quantization_tests = os.listdir(os.path.join(tests, "quantization")); subdirs = ast.literal_eval(os.environ["subdirs"]); quantization_tests = [x.removeprefix("quantization/") for x in subdirs] if subdirs is not None else quantization_tests; d = sorted(list(filter(os.path.isdir, [f"quantization/{x}" for x in quantization_tests]))); print(d)')" >> $GITHUB_OUTPUT
|
||||
echo "quantization_matrix=$(python3 -c 'import os; tests = os.getcwd(); quantization_tests = os.listdir(os.path.join(tests, "quantization")); d = sorted(list(filter(os.path.isdir, [f"quantization/{x}" for x in quantization_tests]))) ; print(d)')" >> $GITHUB_OUTPUT
|
||||
|
||||
- name: NVIDIA-SMI
|
||||
run: |
|
||||
@ -147,7 +128,6 @@ jobs:
|
||||
commit_sha: ${{ inputs.commit_sha || github.sha }}
|
||||
runner_type: ${{ inputs.runner_type }}
|
||||
report_repo_id: ${{ inputs.report_repo_id }}
|
||||
pytest_marker: ${{ inputs.pytest_marker }}
|
||||
secrets: inherit
|
||||
|
||||
run_trainer_and_fsdp_gpu:
|
||||
@ -181,14 +161,12 @@ jobs:
|
||||
runs-on:
|
||||
group: '${{ matrix.machine_type }}'
|
||||
container:
|
||||
image: huggingface/transformers-all-latest-gpu
|
||||
image: huggingface/transformers-pytorch-gpu
|
||||
options: --gpus all --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
|
||||
steps:
|
||||
- name: Update clone
|
||||
working-directory: /transformers
|
||||
env:
|
||||
commit_sha: ${{ inputs.commit_sha || github.sha }}
|
||||
run: git fetch && git checkout "$commit_sha"
|
||||
run: git fetch && git checkout ${{ inputs.commit_sha || github.sha }}
|
||||
|
||||
- name: Reinstall transformers in edit mode (remove the one installed during docker image build)
|
||||
working-directory: /transformers
|
||||
@ -210,17 +188,15 @@ jobs:
|
||||
- name: Set `machine_type` for report and artifact names
|
||||
working-directory: /transformers
|
||||
shell: bash
|
||||
env:
|
||||
matrix_machine_type: ${{ matrix.machine_type }}
|
||||
run: |
|
||||
echo "$matrix_machine_type"
|
||||
echo "${{ matrix.machine_type }}"
|
||||
|
||||
if [ "$matrix_machine_type" = "aws-g5-4xlarge-cache" ]; then
|
||||
if [ "${{ matrix.machine_type }}" = "aws-g5-4xlarge-cache" ]; then
|
||||
machine_type=single-gpu
|
||||
elif [ "$matrix_machine_type" = "aws-g5-12xlarge-cache" ]; then
|
||||
elif [ "${{ matrix.machine_type }}" = "aws-g5-12xlarge-cache" ]; then
|
||||
machine_type=multi-gpu
|
||||
else
|
||||
machine_type="$matrix_machine_type"
|
||||
machine_type=${{ matrix.machine_type }}
|
||||
fi
|
||||
|
||||
echo "$machine_type"
|
||||
@ -229,12 +205,12 @@ jobs:
|
||||
- name: Run all pipeline tests on GPU
|
||||
working-directory: /transformers
|
||||
run: |
|
||||
python3 -m pytest -n 1 -v --dist=loadfile --make-reports="${machine_type}_run_pipelines_torch_gpu_test_reports" tests/pipelines
|
||||
python3 -m pytest -n 1 -v --dist=loadfile --make-reports=${{ env.machine_type }}_run_pipelines_torch_gpu_test_reports tests/pipelines
|
||||
|
||||
- name: Failure short reports
|
||||
if: ${{ failure() }}
|
||||
continue-on-error: true
|
||||
run: cat "/transformers/reports/${machine_type}_run_pipelines_torch_gpu_test_reports/failures_short.txt"
|
||||
run: cat /transformers/reports/${{ env.machine_type }}_run_pipelines_torch_gpu_test_reports/failures_short.txt
|
||||
|
||||
- name: "Test suite reports artifacts: ${{ env.machine_type }}_run_pipelines_torch_gpu_test_reports"
|
||||
if: ${{ always() }}
|
||||
@ -258,9 +234,7 @@ jobs:
|
||||
steps:
|
||||
- name: Update clone
|
||||
working-directory: /transformers
|
||||
env:
|
||||
commit_sha: ${{ inputs.commit_sha || github.sha }}
|
||||
run: git fetch && git checkout "$commit_sha"
|
||||
run: git fetch && git checkout ${{ inputs.commit_sha || github.sha }}
|
||||
|
||||
- name: Reinstall transformers in edit mode (remove the one installed during docker image build)
|
||||
working-directory: /transformers
|
||||
@ -282,17 +256,15 @@ jobs:
|
||||
- name: Set `machine_type` for report and artifact names
|
||||
working-directory: /transformers
|
||||
shell: bash
|
||||
env:
|
||||
matrix_machine_type: ${{ matrix.machine_type }}
|
||||
run: |
|
||||
echo "$matrix_machine_type"
|
||||
echo "${{ matrix.machine_type }}"
|
||||
|
||||
if [ "$matrix_machine_type" = "aws-g5-4xlarge-cache" ]; then
|
||||
if [ "${{ matrix.machine_type }}" = "aws-g5-4xlarge-cache" ]; then
|
||||
machine_type=single-gpu
|
||||
elif [ "$matrix_machine_type" = "aws-g5-12xlarge-cache" ]; then
|
||||
elif [ "${{ matrix.machine_type }}" = "aws-g5-12xlarge-cache" ]; then
|
||||
machine_type=multi-gpu
|
||||
else
|
||||
machine_type="$matrix_machine_type"
|
||||
machine_type=${{ matrix.machine_type }}
|
||||
fi
|
||||
|
||||
echo "$machine_type"
|
||||
@ -302,12 +274,12 @@ jobs:
|
||||
working-directory: /transformers
|
||||
run: |
|
||||
pip install -r examples/pytorch/_tests_requirements.txt
|
||||
python3 -m pytest -v --make-reports="${machine_type}_run_examples_gpu_test_reports" examples/pytorch
|
||||
python3 -m pytest -v --make-reports=${{ env.machine_type }}_run_examples_gpu_test_reports examples/pytorch
|
||||
|
||||
- name: Failure short reports
|
||||
if: ${{ failure() }}
|
||||
continue-on-error: true
|
||||
run: cat "/transformers/reports/${machine_type}_run_examples_gpu_test_reports/failures_short.txt"
|
||||
run: cat /transformers/reports/${{ env.machine_type }}_run_examples_gpu_test_reports/failures_short.txt
|
||||
|
||||
- name: "Test suite reports artifacts: ${{ env.machine_type }}_run_examples_gpu_test_reports"
|
||||
if: ${{ always() }}
|
||||
@ -331,9 +303,7 @@ jobs:
|
||||
steps:
|
||||
- name: Update clone
|
||||
working-directory: ${{ inputs.working-directory-prefix }}/transformers
|
||||
env:
|
||||
commit_sha: ${{ inputs.commit_sha || github.sha }}
|
||||
run: git fetch && git checkout "$commit_sha"
|
||||
run: git fetch && git checkout ${{ inputs.commit_sha || github.sha }}
|
||||
|
||||
- name: Reinstall transformers in edit mode (remove the one installed during docker image build)
|
||||
working-directory: ${{ inputs.working-directory-prefix }}/transformers
|
||||
@ -355,7 +325,7 @@ jobs:
|
||||
working-directory: ${{ inputs.working-directory-prefix }}/
|
||||
run: |
|
||||
python3 -m pip uninstall -y deepspeed
|
||||
DS_DISABLE_NINJA=1 DS_BUILD_CPU_ADAM=1 DS_BUILD_FUSED_ADAM=1 python3 -m pip install deepspeed --no-build-isolation --config-settings="--build-option=build_ext" --config-settings="--build-option=-j8" --no-cache -v --disable-pip-version-check
|
||||
DS_DISABLE_NINJA=1 DS_BUILD_CPU_ADAM=1 DS_BUILD_FUSED_ADAM=1 python3 -m pip install deepspeed --global-option="build_ext" --global-option="-j8" --no-cache -v --disable-pip-version-check
|
||||
|
||||
# To avoid unknown test failures
|
||||
- name: Pre build DeepSpeed *again* (for nightly & Past CI)
|
||||
@ -365,7 +335,7 @@ jobs:
|
||||
python3 -m pip uninstall -y deepspeed
|
||||
rm -rf DeepSpeed
|
||||
git clone https://github.com/deepspeedai/DeepSpeed && cd DeepSpeed && rm -rf build
|
||||
DS_BUILD_CPU_ADAM=1 DS_BUILD_FUSED_ADAM=1 python3 -m pip install . --no-build-isolation --config-settings="--build-option=build_ext" --config-settings="--build-option=-j8" --no-cache -v --disable-pip-version-check
|
||||
DS_BUILD_CPU_ADAM=1 DS_BUILD_FUSED_ADAM=1 python3 -m pip install . --global-option="build_ext" --global-option="-j8" --no-cache -v --disable-pip-version-check
|
||||
|
||||
- name: NVIDIA-SMI
|
||||
run: |
|
||||
@ -383,17 +353,15 @@ jobs:
|
||||
- name: Set `machine_type` for report and artifact names
|
||||
working-directory: ${{ inputs.working-directory-prefix }}/transformers
|
||||
shell: bash
|
||||
env:
|
||||
matrix_machine_type: ${{ matrix.machine_type }}
|
||||
run: |
|
||||
echo "$matrix_machine_type"
|
||||
echo "${{ matrix.machine_type }}"
|
||||
|
||||
if [ "$matrix_machine_type" = "aws-g5-4xlarge-cache" ]; then
|
||||
if [ "${{ matrix.machine_type }}" = "aws-g5-4xlarge-cache" ]; then
|
||||
machine_type=single-gpu
|
||||
elif [ "$matrix_machine_type" = "aws-g5-12xlarge-cache" ]; then
|
||||
elif [ "${{ matrix.machine_type }}" = "aws-g5-12xlarge-cache" ]; then
|
||||
machine_type=multi-gpu
|
||||
else
|
||||
machine_type="$matrix_machine_type"
|
||||
machine_type=${{ matrix.machine_type }}
|
||||
fi
|
||||
|
||||
echo "$machine_type"
|
||||
@ -402,14 +370,12 @@ jobs:
|
||||
- name: Run all tests on GPU
|
||||
working-directory: ${{ inputs.working-directory-prefix }}/transformers
|
||||
run: |
|
||||
python3 -m pytest -v --make-reports="${machine_type}_run_torch_cuda_extensions_gpu_test_reports" tests/deepspeed tests/extended
|
||||
python3 -m pytest -v --make-reports=${{ env.machine_type }}_run_torch_cuda_extensions_gpu_test_reports tests/deepspeed tests/extended
|
||||
|
||||
- name: Failure short reports
|
||||
if: ${{ failure() }}
|
||||
continue-on-error: true
|
||||
env:
|
||||
working_directory_prefix: ${{ inputs.working-directory-prefix }}
|
||||
run: cat "${working_directory_prefix}/transformers/reports/${machine_type}_run_torch_cuda_extensions_gpu_test_reports/failures_short.txt"
|
||||
run: cat ${{ inputs.working-directory-prefix }}/transformers/reports/${{ env.machine_type }}_run_torch_cuda_extensions_gpu_test_reports/failures_short.txt
|
||||
|
||||
- name: "Test suite reports artifacts: ${{ env.machine_type }}_run_torch_cuda_extensions_gpu_test_reports"
|
||||
if: ${{ always() }}
|
||||
@ -436,19 +402,16 @@ jobs:
|
||||
steps:
|
||||
- name: Echo folder ${{ matrix.folders }}
|
||||
shell: bash
|
||||
env:
|
||||
matrix_folders_raw: ${{ matrix.folders }}
|
||||
run: |
|
||||
echo "$matrix_folders_raw"
|
||||
matrix_folders="${matrix_folders_raw/'quantization/'/'quantization_'}"
|
||||
echo "${{ matrix.folders }}"
|
||||
matrix_folders=${{ matrix.folders }}
|
||||
matrix_folders=${matrix_folders/'quantization/'/'quantization_'}
|
||||
echo "$matrix_folders"
|
||||
echo "matrix_folders=$matrix_folders" >> $GITHUB_ENV
|
||||
|
||||
- name: Update clone
|
||||
working-directory: /transformers
|
||||
env:
|
||||
commit_sha: ${{ inputs.commit_sha || github.sha }}
|
||||
run: git fetch && git checkout "$commit_sha"
|
||||
run: git fetch && git checkout ${{ inputs.commit_sha || github.sha }}
|
||||
|
||||
- name: Reinstall transformers in edit mode (remove the one installed during docker image build)
|
||||
working-directory: /transformers
|
||||
@ -470,17 +433,15 @@ jobs:
|
||||
- name: Set `machine_type` for report and artifact names
|
||||
working-directory: /transformers
|
||||
shell: bash
|
||||
env:
|
||||
matrix_machine_type: ${{ matrix.machine_type }}
|
||||
run: |
|
||||
echo "$matrix_machine_type"
|
||||
echo "${{ matrix.machine_type }}"
|
||||
|
||||
if [ "$matrix_machine_type" = "aws-g5-4xlarge-cache" ]; then
|
||||
if [ "${{ matrix.machine_type }}" = "aws-g5-4xlarge-cache" ]; then
|
||||
machine_type=single-gpu
|
||||
elif [ "$matrix_machine_type" = "aws-g5-12xlarge-cache" ]; then
|
||||
elif [ "${{ matrix.machine_type }}" = "aws-g5-12xlarge-cache" ]; then
|
||||
machine_type=multi-gpu
|
||||
else
|
||||
machine_type="$matrix_machine_type"
|
||||
machine_type=${{ matrix.machine_type }}
|
||||
fi
|
||||
|
||||
echo "$machine_type"
|
||||
@ -488,96 +449,20 @@ jobs:
|
||||
|
||||
- name: Run quantization tests on GPU
|
||||
working-directory: /transformers
|
||||
env:
|
||||
folders: ${{ matrix.folders }}
|
||||
run: |
|
||||
python3 -m pytest -v --make-reports="${machine_type}_run_quantization_torch_gpu_${matrix_folders}_test_reports" tests/${folders}
|
||||
python3 -m pytest -v --make-reports=${{ env.machine_type }}_run_quantization_torch_gpu_${{ matrix.folders }}_test_reports tests/${{ matrix.folders }}
|
||||
|
||||
- name: Failure short reports
|
||||
if: ${{ failure() }}
|
||||
continue-on-error: true
|
||||
run: cat "/transformers/reports/${machine_type}_run_quantization_torch_gpu_${matrix_folders}_test_reports/failures_short.txt"
|
||||
run: cat /transformers/reports/${{ env.machine_type }}_run_quantization_torch_gpu_${{ matrix.folders }}_test_reports/failures_short.txt
|
||||
|
||||
- name: "Test suite reports artifacts: ${{ env.machine_type }}_run_quantization_torch_gpu_${{ env.matrix_folders }}_test_reports"
|
||||
if: ${{ always() }}
|
||||
uses: actions/upload-artifact@v4
|
||||
with:
|
||||
name: ${{ env.machine_type }}_run_quantization_torch_gpu_${{ env.matrix_folders }}_test_reports
|
||||
path: /transformers/reports/${{ env.machine_type }}_run_quantization_torch_gpu_${{ env.matrix_folders }}_test_reports
|
||||
|
||||
run_kernels_gpu:
|
||||
if: ${{ inputs.job == 'run_kernels_gpu' }}
|
||||
name: Kernel tests
|
||||
strategy:
|
||||
fail-fast: false
|
||||
matrix:
|
||||
machine_type: [aws-g5-4xlarge-cache]
|
||||
runs-on:
|
||||
group: '${{ matrix.machine_type }}'
|
||||
container:
|
||||
image: ${{ inputs.docker }}
|
||||
options: --gpus all --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
|
||||
steps:
|
||||
- name: Update clone
|
||||
working-directory: /transformers
|
||||
env:
|
||||
commit_sha: ${{ inputs.commit_sha || github.sha }}
|
||||
run: git fetch && git checkout "$commit_sha"
|
||||
|
||||
- name: Reinstall transformers in edit mode
|
||||
working-directory: /transformers
|
||||
run: python3 -m pip uninstall -y transformers && python3 -m pip install -e .[testing]
|
||||
|
||||
- name: Install kernels
|
||||
working-directory: /transformers
|
||||
run: python3 -m pip install -U kernels
|
||||
|
||||
- name: NVIDIA-SMI
|
||||
run: nvidia-smi
|
||||
|
||||
- name: Environment
|
||||
working-directory: /transformers
|
||||
run: python3 utils/print_env.py
|
||||
|
||||
- name: Show installed libraries and their versions
|
||||
working-directory: /transformers
|
||||
run: pip freeze
|
||||
|
||||
- name: Set `machine_type` for report and artifact names
|
||||
working-directory: /transformers
|
||||
shell: bash
|
||||
env:
|
||||
matrix_machine_type: ${{ matrix.machine_type }}
|
||||
run: |
|
||||
echo "$matrix_machine_type"
|
||||
|
||||
if [ "$matrix_machine_type" = "aws-g5-4xlarge-cache" ]; then
|
||||
machine_type=single-gpu
|
||||
elif [ "$matrix_machine_type" = "aws-g5-12xlarge-cache" ]; then
|
||||
machine_type=multi-gpu
|
||||
else
|
||||
machine_type="$matrix_machine_type"
|
||||
fi
|
||||
|
||||
echo "$machine_type"
|
||||
echo "machine_type=$machine_type" >> $GITHUB_ENV
|
||||
|
||||
- name: Run kernel tests on GPU
|
||||
working-directory: /transformers
|
||||
run: |
|
||||
python3 -m pytest -v --make-reports="${machine_type}_run_kernels_gpu_test_reports" tests/kernels/test_kernels.py
|
||||
|
||||
- name: Failure short reports
|
||||
if: ${{ failure() }}
|
||||
continue-on-error: true
|
||||
run: cat "/transformers/reports/${machine_type}_run_kernels_gpu_test_reports/failures_short.txt"
|
||||
|
||||
- name: "Test suite reports artifacts: ${{ env.machine_type }}_run_kernels_gpu_test_reports"
|
||||
if: ${{ always() }}
|
||||
uses: actions/upload-artifact@v4
|
||||
with:
|
||||
name: ${{ env.machine_type }}_run_kernels_gpu_test_reports
|
||||
path: /transformers/reports/${{ env.machine_type }}_run_kernels_gpu_test_reports
|
||||
path: /transformers/reports/${{ env.machine_type }}_run_quantization_torch_gpu_${{ matrix.folders }}_test_reports
|
||||
|
||||
run_extract_warnings:
|
||||
# Let's only do this for the job `run_models_gpu` to simplify the (already complex) logic.
|
||||
@ -586,10 +471,11 @@ jobs:
|
||||
runs-on: ubuntu-22.04
|
||||
needs: [setup, run_models_gpu]
|
||||
steps:
|
||||
# Checkout in order to run `utils/extract_warnings.py`. Avoid **explicit** checkout (i.e. don't specify `ref`) for
|
||||
# security reason.
|
||||
- name: Checkout transformers
|
||||
uses: actions/checkout@v4
|
||||
with:
|
||||
fetch-depth: 2
|
||||
ref: ${{ inputs.commit_sha || github.sha }}
|
||||
|
||||
- name: Install transformers
|
||||
run: pip install transformers
|
||||
@ -609,12 +495,9 @@ jobs:
|
||||
working-directory: warnings_in_ci
|
||||
|
||||
- name: Extract warnings in CI artifacts
|
||||
env:
|
||||
github_run_id: ${{ github.run_id }}
|
||||
access_token: ${{ secrets.ACCESS_REPO_INFO_TOKEN }}
|
||||
run: |
|
||||
python3 utils/extract_warnings.py --workflow_run_id "$github_run_id" --output_dir warnings_in_ci --token "$access_token" --from_gh
|
||||
echo "$(python3 -c 'import os; import json; fp = open("warnings_in_ci/selected_warnings.json"); d = json.load(fp); d = "\n".join(d); print(d)')"
|
||||
python3 utils/extract_warnings.py --workflow_run_id ${{ github.run_id }} --output_dir warnings_in_ci --token ${{ secrets.ACCESS_REPO_INFO_TOKEN }} --from_gh
|
||||
echo "$(python3 -c 'import os; import json; fp = open("warnings_in_ci/selected_warnings.json"); d = json.load(fp); d = "\n".join(d) ;print(d)')"
|
||||
|
||||
- name: Upload artifact
|
||||
if: ${{ always() }}
|
||||
@ -633,7 +516,6 @@ jobs:
|
||||
run_examples_gpu,
|
||||
run_torch_cuda_extensions_gpu,
|
||||
run_quantization_torch_gpu,
|
||||
run_kernels_gpu,
|
||||
run_extract_warnings
|
||||
]
|
||||
if: always() && !cancelled()
|
||||
@ -653,17 +535,16 @@ jobs:
|
||||
secrets: inherit
|
||||
|
||||
check_new_failures:
|
||||
if: ${{ always() && needs.send_results.result == 'success' }}
|
||||
if: ${{ always() && inputs.ci_event == 'Daily CI' && needs.send_results.result == 'success' }}
|
||||
name: Check new failures
|
||||
needs: send_results
|
||||
uses: ./.github/workflows/check_failed_tests.yml
|
||||
with:
|
||||
docker: ${{ inputs.docker }}
|
||||
commit_sha: ${{ inputs.commit_sha || github.sha }}
|
||||
start_sha: ${{ inputs.commit_sha || github.sha }}
|
||||
job: ${{ inputs.job }}
|
||||
slack_report_channel: ${{ inputs.slack_report_channel }}
|
||||
ci_event: ${{ inputs.ci_event }}
|
||||
report_repo_id: ${{ inputs.report_repo_id }}
|
||||
pr_number: ${{ inputs.pr_number }}
|
||||
|
||||
secrets: inherit
|
||||
|
||||
15
.github/workflows/slack-report.yml
vendored
15
.github/workflows/slack-report.yml
vendored
@ -41,16 +41,13 @@ jobs:
|
||||
- name: Preliminary job status
|
||||
shell: bash
|
||||
# For the meaning of these environment variables, see the job `Setup`
|
||||
env:
|
||||
setup_status: ${{ inputs.setup_status }}
|
||||
run: |
|
||||
echo "Setup status: $setup_status"
|
||||
echo "Setup status: ${{ inputs.setup_status }}"
|
||||
|
||||
- uses: actions/checkout@v4
|
||||
with:
|
||||
fetch-depth: 2
|
||||
# Security: checkout to the `main` branch for untrusted triggers (issue_comment, pull_request_target), otherwise use the specified ref
|
||||
ref: ${{ (github.event_name == 'issue_comment' || github.event_name == 'pull_request_target') && 'main' || (inputs.commit_sha || github.sha) }}
|
||||
ref: ${{ inputs.commit_sha || github.sha }}
|
||||
|
||||
- uses: actions/download-artifact@v4
|
||||
|
||||
@ -84,8 +81,6 @@ jobs:
|
||||
CI_TEST_JOB: ${{ inputs.job }}
|
||||
SETUP_STATUS: ${{ inputs.setup_status }}
|
||||
REPORT_REPO_ID: ${{ inputs.report_repo_id }}
|
||||
quantization_matrix: ${{ inputs.quantization_matrix }}
|
||||
folder_slices: ${{ inputs.folder_slices }}
|
||||
# We pass `needs.setup.outputs.matrix` as the argument. A processing in `notification_service.py` to change
|
||||
# `models/bert` to `models_bert` is required, as the artifact names use `_` instead of `/`.
|
||||
# For a job that doesn't depend on (i.e. `needs`) `setup`, the value for `inputs.folder_slices` would be an
|
||||
@ -94,10 +89,10 @@ jobs:
|
||||
pip install huggingface_hub
|
||||
pip install slack_sdk
|
||||
pip show slack_sdk
|
||||
if [ "$quantization_matrix" != "" ]; then
|
||||
python utils/notification_service.py "$quantization_matrix"
|
||||
if [ "${{ inputs.quantization_matrix }}" != "" ]; then
|
||||
python utils/notification_service.py "${{ inputs.quantization_matrix }}"
|
||||
else
|
||||
python utils/notification_service.py "$folder_slices"
|
||||
python utils/notification_service.py "${{ inputs.folder_slices }}"
|
||||
fi
|
||||
|
||||
# Upload complete failure tables, as they might be big and only truncated versions could be sent to Slack.
|
||||
|
||||
47
.github/workflows/ssh-runner.yml
vendored
47
.github/workflows/ssh-runner.yml
vendored
@ -4,7 +4,7 @@ on:
|
||||
workflow_dispatch:
|
||||
inputs:
|
||||
runner_type:
|
||||
description: 'Type of runner to test (a10)'
|
||||
description: 'Type of runner to test (a10 or t4)'
|
||||
required: true
|
||||
docker_image:
|
||||
description: 'Name of the Docker image'
|
||||
@ -20,6 +20,7 @@ env:
|
||||
OMP_NUM_THREADS: 8
|
||||
MKL_NUM_THREADS: 8
|
||||
RUN_SLOW: yes # For gated repositories, we still need to agree to share information on the Hub repo. page in order to get access. # This token is created under the bot `hf-transformers-bot`.
|
||||
SIGOPT_API_TOKEN: ${{ secrets.SIGOPT_API_TOKEN }}
|
||||
TF_FORCE_GPU_ALLOW_GROWTH: true
|
||||
CUDA_VISIBLE_DEVICES: 0,1
|
||||
|
||||
@ -32,14 +33,15 @@ jobs:
|
||||
steps:
|
||||
- name: Get runner to use
|
||||
shell: bash
|
||||
env:
|
||||
NUM_GPUS: ${{ github.event.inputs.num_gpus }}
|
||||
RUNNER_TYPE: ${{ github.event.inputs.runner_type }}
|
||||
run: |
|
||||
if [[ "$NUM_GPUS" == "single" && "$RUNNER_TYPE" == "a10" ]]; then
|
||||
echo "RUNNER=aws-g5-4xlarge-cache-ssh" >> $GITHUB_ENV
|
||||
elif [[ "$NUM_GPUS" == "multi" && "$RUNNER_TYPE" == "a10" ]]; then
|
||||
echo "RUNNER=aws-g5-12xlarge-cache-ssh" >> $GITHUB_ENV
|
||||
if [[ "${{ github.event.inputs.num_gpus }}" == "single" && "${{ github.event.inputs.runner_type }}" == "t4" ]]; then
|
||||
echo "RUNNER=aws-g4dn-4xlarge-cache" >> $GITHUB_ENV
|
||||
elif [[ "${{ github.event.inputs.num_gpus }}" == "multi" && "${{ github.event.inputs.runner_type }}" == "t4" ]]; then
|
||||
echo "RUNNER=aws-g4dn-12xlarge-cache" >> $GITHUB_ENV
|
||||
elif [[ "${{ github.event.inputs.num_gpus }}" == "single" && "${{ github.event.inputs.runner_type }}" == "a10" ]]; then
|
||||
echo "RUNNER=aws-g5-4xlarge-cache" >> $GITHUB_ENV
|
||||
elif [[ "${{ github.event.inputs.num_gpus }}" == "multi" && "${{ github.event.inputs.runner_type }}" == "a10" ]]; then
|
||||
echo "RUNNER=aws-g5-12xlarge-cache" >> $GITHUB_ENV
|
||||
else
|
||||
echo "RUNNER=" >> $GITHUB_ENV
|
||||
fi
|
||||
@ -47,8 +49,8 @@ jobs:
|
||||
- name: Set runner to use
|
||||
id: set_runner
|
||||
run: |
|
||||
echo "$RUNNER"
|
||||
echo "RUNNER=$RUNNER" >> $GITHUB_OUTPUT
|
||||
echo ${{ env.RUNNER }}
|
||||
echo "RUNNER=${{ env.RUNNER }}" >> $GITHUB_OUTPUT
|
||||
|
||||
ssh_runner:
|
||||
name: "SSH"
|
||||
@ -57,13 +59,13 @@ jobs:
|
||||
group: ${{ needs.get_runner.outputs.RUNNER }}
|
||||
container:
|
||||
image: ${{ github.event.inputs.docker_image }}
|
||||
options: --gpus all --privileged --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
|
||||
|
||||
steps:
|
||||
- name: Update clone
|
||||
working-directory: /transformers
|
||||
env:
|
||||
commit_sha: ${{ github.sha }}
|
||||
run: |
|
||||
git fetch && git checkout "$commit_sha"
|
||||
git fetch && git checkout ${{ github.sha }}
|
||||
|
||||
- name: Cleanup
|
||||
working-directory: /transformers
|
||||
@ -83,11 +85,9 @@ jobs:
|
||||
- name: Store Slack infos
|
||||
#because the SSH can be enabled dynamically if the workflow failed, so we need to store slack infos to be able to retrieve them during the waitforssh step
|
||||
shell: bash
|
||||
env:
|
||||
GITHUB_ACTOR: ${{ github.actor }}
|
||||
run: |
|
||||
echo "$GITHUB_ACTOR"
|
||||
github_actor=$GITHUB_ACTOR
|
||||
echo "${{ github.actor }}"
|
||||
github_actor=${{ github.actor }}
|
||||
github_actor=${github_actor/'-'/'_'}
|
||||
echo "$github_actor"
|
||||
echo "github_actor=$github_actor" >> $GITHUB_ENV
|
||||
@ -95,17 +95,14 @@ jobs:
|
||||
- name: Store Slack infos
|
||||
#because the SSH can be enabled dynamically if the workflow failed, so we need to store slack infos to be able to retrieve them during the waitforssh step
|
||||
shell: bash
|
||||
env:
|
||||
user_slack_id: ${{ secrets[format('{0}_{1}', env.github_actor, 'SLACK_ID')] }}
|
||||
default_slack_channel: ${{ secrets.SLACK_CIFEEDBACK_CHANNEL }}
|
||||
run: |
|
||||
echo "$github_actor"
|
||||
if [ "$user_slack_id" != "" ]; then
|
||||
echo "SLACKCHANNEL=$user_slack_id" >> $GITHUB_ENV
|
||||
echo "${{ env.github_actor }}"
|
||||
if [ "${{ secrets[format('{0}_{1}', env.github_actor, 'SLACK_ID')] }}" != "" ]; then
|
||||
echo "SLACKCHANNEL=${{ secrets[format('{0}_{1}', env.github_actor, 'SLACK_ID')] }}" >> $GITHUB_ENV
|
||||
else
|
||||
echo "SLACKCHANNEL=$default_slack_channel" >> $GITHUB_ENV
|
||||
echo "SLACKCHANNEL=${{ secrets.SLACK_CIFEEDBACK_CHANNEL }}" >> $GITHUB_ENV
|
||||
fi
|
||||
|
||||
|
||||
- name: Tailscale # In order to be able to SSH when a test fails
|
||||
uses: huggingface/tailscale-action@main
|
||||
with:
|
||||
|
||||
4
.gitignore
vendored
4
.gitignore
vendored
@ -98,7 +98,6 @@ celerybeat-schedule
|
||||
# Environments
|
||||
.env
|
||||
.venv
|
||||
.venv*
|
||||
env/
|
||||
venv/
|
||||
ENV/
|
||||
@ -172,6 +171,3 @@ tags
|
||||
|
||||
# modular conversion
|
||||
*.modular_backup
|
||||
|
||||
# Cursor IDE files
|
||||
.cursor/
|
||||
|
||||
@ -14,7 +14,7 @@ This AGENTS.md file provides guidance for code agents working with this codebase
|
||||
|
||||
- PRs should be as brief as possible. Bugfix PRs in particular can often be only one or two lines long, and do not need large comments, docstrings or new functions in this case. Aim to minimize the size of the diff.
|
||||
- When writing tests, they should be added to an existing file. The only exception is for PRs to add a new model, when a new test directory should be created for that model.
|
||||
- Code style is enforced in the CI. You can install the style tools with `pip install -e ".[quality]"`. You can then run `make fixup` to apply style and consistency fixes to your code.
|
||||
- Code style is enforced in the CI. You can install the style tools with `pip install -e .[quality]`. You can then run `make fixup` to apply style and consistency fixes to your code.
|
||||
|
||||
## Copying and inheritance
|
||||
|
||||
@ -36,4 +36,4 @@ After making changes, you should usually run `make fixup` to ensure any copies a
|
||||
the model you made the changes in and any other models that were updated by `make fixup`. Tests can be run with `pytest tests/models/[name]/test_modeling_[name].py`
|
||||
If your changes affect code in other classes like tokenizers or processors, you should run those tests instead, like `test_processing_[name].py` or `test_tokenization_[name].py`.
|
||||
|
||||
In order to run tests, you may need to install dependencies. You can do this with `pip install -e ".[testing]"`. You will probably also need to `pip install torch accelerate` if your environment does not already have them.
|
||||
In order to run tests, you may need to install dependencies. You can do this with `pip install -e .[testing]`. You will probably also need to `pip install torch accelerate` if your environment does not already have them.
|
||||
120
CONTRIBUTING.md
120
CONTRIBUTING.md
@ -112,125 +112,7 @@ New models are constantly released and if you want to implement a new model, ple
|
||||
|
||||
If you are willing to contribute the model yourself, let us know so we can help you add it to 🤗 Transformers!
|
||||
|
||||
We have a technical guide for [how to add a model to 🤗 Transformers](https://huggingface.co/docs/transformers/modular_transformers).
|
||||
|
||||
### Vision-Language Model Contribution Checklist
|
||||
|
||||
If you're contributing a **vision-language model** (or any multimodal model that processes images/videos), please follow this checklist. Maintainers will use this to review your PR, and completing these steps will significantly increase the likelihood of your PR being merged quickly.
|
||||
|
||||
**Required checklist for all vision-language model contributions:**
|
||||
|
||||
☐ **1. Implement a modular file**
|
||||
|
||||
All new models should use the modular architecture pattern. Create a `modular_<model_name>.py` file using the modular model converter:
|
||||
|
||||
- Use the CLI, [`transformers add-new-model-like`](https://github.com/huggingface/transformers/blob/main/src/transformers/cli/add_new_model_like.py) to generate a modular skeleton and get started
|
||||
- All code should be in the modular file if possible. Modeling must be in it, it's better if configuration is in it as well.
|
||||
- Reuse existing patterns from similar models as much as possible
|
||||
|
||||
To verify your modular file is correct, run:
|
||||
|
||||
```bash
|
||||
python utils/modular_model_converter.py <model_name>
|
||||
```
|
||||
|
||||
This will generate the separate files (`modeling_*.py`, `configuration_*.py`, etc.) from your modular file. The CI will enforce that these generated files match your modular file.
|
||||
|
||||
☐ **2. Add a fast image processor (for image models)**
|
||||
|
||||
If your model processes images, implement a fast image processor that uses `torch` and `torchvision` instead of PIL/numpy for better inference performance:
|
||||
|
||||
- See the detailed guide in [#36978](https://github.com/huggingface/transformers/issues/36978)
|
||||
- Fast processors inherit from `BaseImageProcessorFast`
|
||||
- Examples: `LlavaOnevisionImageProcessorFast`, `Idefics2ImageProcessorFast`
|
||||
|
||||
☐ **3. Create a weight conversion script**
|
||||
|
||||
Add a `convert_<model_name>_to_hf.py` script that converts the original model weights to the HuggingFace format:
|
||||
|
||||
- Script should handle checkpoint loading, key mapping, and saving in HF format
|
||||
- Include usage examples and documentation in the script
|
||||
- Examples: [`convert_llava_onevision_weights_to_hf.py`](https://github.com/huggingface/transformers/blob/main/src/transformers/models/llava_onevision/convert_llava_onevision_weights_to_hf.py), [`convert_idefics2_weights_to_hf.py`](https://github.com/huggingface/transformers/blob/main/src/transformers/models/idefics2/convert_idefics2_weights_to_hf.py)
|
||||
|
||||
☐ **4. Add integration tests with exact output matching**
|
||||
|
||||
At minimum, add an `IntegrationTest` class that tests end-to-end generation (processing and modelling) with **exact** output matching:
|
||||
|
||||
- For generative models: test that generated text matches expected output exactly
|
||||
- For non-generative models: test that output logits match expected values
|
||||
- Tests should use real checkpoints (load in 4-bit or half precision if the checkpoint is too big to fit in our CI runners) and real inputs
|
||||
- Example pattern:
|
||||
|
||||
```python
|
||||
class MyModelIntegrationTest(unittest.TestCase):
|
||||
@slow
|
||||
def test_model_integration(self):
|
||||
model = MyModelForConditionalGeneration.from_pretrained("org/model-name")
|
||||
processor = AutoProcessor.from_pretrained("org/model-name")
|
||||
|
||||
inputs = processor(images=image, text=prompt, return_tensors="pt")
|
||||
output = model.generate(**inputs, max_new_tokens=20)
|
||||
|
||||
EXPECTED_TEXT = "exact expected output"
|
||||
self.assertEqual(processor.decode(output[0]), EXPECTED_TEXT)
|
||||
```
|
||||
|
||||
See `tests/models/llava_onevision/test_modeling_llava_onevision.py` for complete examples.
|
||||
|
||||
☐ **5. Update documentation**
|
||||
|
||||
Add or update model documentation:
|
||||
|
||||
- Create if the cli hasn't `docs/source/en/model_doc/<model_name>.md` with usage examples
|
||||
- Include model description, paper link, and basic usage with `Pipeline` and `AutoModel`
|
||||
- Add the model to the appropriate TOC files
|
||||
|
||||
☐ **6. Look for reusable patterns**
|
||||
|
||||
The library has 400+ models with many established patterns:
|
||||
|
||||
- Search for similar models (e.g., other vision-language models)
|
||||
- Reuse attention mechanisms, layer implementations, and processing patterns
|
||||
- Check models like LLaVA, Idefics2, Fuyu for vision-language patterns
|
||||
- Use provided decorators like (`auto_docstring`, `can_return_tuple`, `check_model_inputs` and `_can_record_outputs`) where relevant.
|
||||
- Don't reinvent the wheel
|
||||
|
||||
☐ **7. Run quality checks and read the output**
|
||||
|
||||
Before submitting your PR, install quality dependencies and run the full check suite:
|
||||
|
||||
```bash
|
||||
pip install -e ".[quality]"
|
||||
make fixup
|
||||
```
|
||||
|
||||
**Important**: Take time to read the output of `make fixup`. It will:
|
||||
- Lint and format your code automatically
|
||||
- Run consistency checks (imports, docstrings, etc.)
|
||||
- Show any remaining issues that need manual fixes
|
||||
|
||||
All checks must pass before your PR can be merged.
|
||||
|
||||
**If this checklist is complete, your PR has a very high likelihood of being merged!** Following these steps makes the maintainers' work much easier and will reduce the number of review iterations, getting your important work out there faster.
|
||||
|
||||
#### Copy-pastable checklist for maintainers
|
||||
|
||||
Here's a condensed version maintainers can copy into PRs:
|
||||
|
||||
```markdown
|
||||
## Multimodal Model Addition Checklist
|
||||
|
||||
Please ensure your PR completes all following items. See the [full checklist](https://github.com/huggingface/transformers/blob/main/CONTRIBUTING.md#vision-language-model-contribution-checklist) for details.
|
||||
|
||||
- [ ] **Modular file**: `modular_<model_name>.py` implemented and verified with `python utils/modular_model_converter.py <model_name>`
|
||||
- [ ] **Fast image processor**: Implemented using `BaseImageProcessorFast` (see [#36978](https://github.com/huggingface/transformers/issues/36978))
|
||||
- [ ] **Conversion script**: `convert_<model_name>_to_hf.py` added with usage examples
|
||||
- [ ] **Integration tests**: End-to-end tests with exact output matching (text or logits)
|
||||
- [ ] **Documentation**: Model docs added/updated in `docs/source/en/model_doc/`
|
||||
- [ ] **Pattern reuse**: Verified against similar models (LLaVA, Idefics2, etc.)
|
||||
- [ ] **Quality checks**: `make fixup` passes with no errors
|
||||
|
||||
```
|
||||
We have a technical guide for [how to add a model to 🤗 Transformers](https://huggingface.co/docs/transformers/add_new_model).
|
||||
|
||||
## Do you want to add documentation?
|
||||
|
||||
|
||||
@ -153,7 +153,7 @@ You are not required to read the following guidelines before opening an issue. H
|
||||
cd examples/seq2seq
|
||||
torchrun --nproc_per_node=2 ./finetune_trainer.py \
|
||||
--model_name_or_path sshleifer/distill-mbart-en-ro-12-4 --data_dir wmt_en_ro \
|
||||
--output_dir output_dir \
|
||||
--output_dir output_dir --overwrite_output_dir \
|
||||
--do_train --n_train 500 --num_train_epochs 1 \
|
||||
--per_device_train_batch_size 1 --freeze_embeds \
|
||||
--src_lang en_XX --tgt_lang ro_RO --task translation \
|
||||
|
||||
@ -64,8 +64,8 @@ limitations under the License.
|
||||
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/transformers_as_a_model_definition.png"/>
|
||||
</h3>
|
||||
|
||||
Transformers acts as the model-definition framework for state-of-the-art machine learning with text, computer
|
||||
vision, audio, video, and multimodal models, for both inference and training.
|
||||
Transformers acts as the model-definition framework for state-of-the-art machine learning models in text, computer
|
||||
vision, audio, video, and multimodal model, for both inference and training.
|
||||
|
||||
It centralizes the model definition so that this definition is agreed upon across the ecosystem. `transformers` is the
|
||||
pivot across frameworks: if a model definition is supported, it will be compatible with the majority of training
|
||||
@ -81,7 +81,7 @@ Explore the [Hub](https://huggingface.com/) today to find a model and use Transf
|
||||
|
||||
## Installation
|
||||
|
||||
Transformers works with Python 3.9+, and [PyTorch](https://pytorch.org/get-started/locally/) 2.1+.
|
||||
Transformers works with Python 3.9+ [PyTorch](https://pytorch.org/get-started/locally/) 2.1+, [TensorFlow](https://www.tensorflow.org/install/pip) 2.6+, and [Flax](https://flax.readthedocs.io/en/latest/) 0.4.1+.
|
||||
|
||||
Create and activate a virtual environment with [venv](https://docs.python.org/3/library/venv.html) or [uv](https://docs.astral.sh/uv/), a fast Rust-based Python package and project manager.
|
||||
|
||||
|
||||
@ -9,12 +9,6 @@ In this list, we showcase incredibly impactful and novel projects that have push
|
||||
adding other projects to the list. If you believe a project should be here and it's not, then please, open a PR
|
||||
to add it.
|
||||
|
||||
## [◉ Universal Intelligence](https://github.com/blueraai/universal-intelligence)
|
||||
|
||||
[Universal Intelligence](https://github.com/blueraai/universal-intelligence) aims to standardize models, tools, and agents —transforming them into simple, composable, portable, interoperable, framework-agnostic, hardware-agnostic interfaces (through auto-negotiation and resource sharing); for fast and accessible development of AI applications.
|
||||
|
||||
Keywords: Protocol, Open-source, LLMs, Large Language Models, Agents, Low-code
|
||||
|
||||
## [gpt4all](https://github.com/nomic-ai/gpt4all)
|
||||
|
||||
[gpt4all](https://github.com/nomic-ai/gpt4all) is an ecosystem of open-source chatbots trained on massive collections of clean assistant data including code, stories and dialogue. It offers open-source, large language models such as LLaMA and GPT-J trained in an assistant-style.
|
||||
|
||||
@ -16,6 +16,7 @@ import sys
|
||||
from logging import Logger
|
||||
from threading import Event, Thread
|
||||
from time import perf_counter, sleep
|
||||
from typing import Optional
|
||||
|
||||
|
||||
# Add the parent directory to Python path to import benchmarks_entrypoint
|
||||
@ -41,7 +42,7 @@ except ImportError:
|
||||
GenerationConfig = None
|
||||
StaticCache = None
|
||||
|
||||
os.environ["HF_XET_HIGH_PERFORMANCE"] = "1"
|
||||
os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "1"
|
||||
os.environ["TOKENIZERS_PARALLELISM"] = "1"
|
||||
|
||||
# Only set torch precision if torch is available
|
||||
@ -144,7 +145,7 @@ def run_benchmark(
|
||||
q = torch.empty_like(probs_sort).exponential_(1)
|
||||
return torch.argmax(probs_sort / q, dim=-1, keepdim=True).to(dtype=torch.int)
|
||||
|
||||
def logits_to_probs(logits, temperature: float = 1.0, top_k: int | None = None):
|
||||
def logits_to_probs(logits, temperature: float = 1.0, top_k: Optional[int] = None):
|
||||
logits = logits / max(temperature, 1e-5)
|
||||
|
||||
if top_k is not None:
|
||||
@ -154,7 +155,7 @@ def run_benchmark(
|
||||
probs = torch.nn.functional.softmax(logits, dim=-1)
|
||||
return probs
|
||||
|
||||
def sample(logits, temperature: float = 1.0, top_k: int | None = None):
|
||||
def sample(logits, temperature: float = 1.0, top_k: Optional[int] = None):
|
||||
probs = logits_to_probs(logits[0, -1], temperature, top_k)
|
||||
idx_next = multinomial_sample_one_no_sync(probs)
|
||||
return idx_next, probs
|
||||
|
||||
@ -1,5 +1,6 @@
|
||||
gpustat==1.1.1
|
||||
psutil==6.0.0
|
||||
psycopg2==2.9.9
|
||||
hf_xet
|
||||
pandas>=1.5.0
|
||||
torch>=2.4.0
|
||||
hf_transfer
|
||||
pandas>=1.5.0
|
||||
3
benchmark_v2/.gitignore
vendored
3
benchmark_v2/.gitignore
vendored
@ -1,2 +1 @@
|
||||
benchmark_results/
|
||||
benchmark_results_profiles/
|
||||
benchmark_results/
|
||||
1
benchmark_v2/benches/__init__.py
Normal file
1
benchmark_v2/benches/__init__.py
Normal file
@ -0,0 +1 @@
|
||||
# Benchmark implementations directory
|
||||
165
benchmark_v2/benches/llama.py
Normal file
165
benchmark_v2/benches/llama.py
Normal file
@ -0,0 +1,165 @@
|
||||
# Copyright 2025 The HuggingFace Team. All rights reserved.
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
|
||||
import logging
|
||||
import os
|
||||
from typing import Any
|
||||
|
||||
import torch
|
||||
from benchmark_framework import ModelBenchmark
|
||||
|
||||
|
||||
os.environ["TOKENIZERS_PARALLELISM"] = "1"
|
||||
torch.set_float32_matmul_precision("high")
|
||||
|
||||
|
||||
class LLaMABenchmark(ModelBenchmark):
|
||||
"""Simplified LLaMA model benchmark implementation using the ModelBenchmark base class."""
|
||||
|
||||
def __init__(self, logger: logging.Logger):
|
||||
super().__init__(logger)
|
||||
self._default_prompt = "Why dogs are so cute?" # Custom prompt for LLaMA
|
||||
|
||||
def get_scenario_configs(self) -> list[dict[str, Any]]:
|
||||
"""
|
||||
Get LLaMA-specific scenario configurations.
|
||||
|
||||
Returns:
|
||||
List of scenario configuration dictionaries
|
||||
"""
|
||||
return [
|
||||
# Eager variants
|
||||
{"variant": "eager", "compile_mode": None, "use_cache": True, "description": "Eager execution with cache"},
|
||||
# Compiled variants
|
||||
{
|
||||
"variant": "compiled",
|
||||
"compile_mode": "max-autotune",
|
||||
"use_cache": True,
|
||||
"description": "Compiled with max autotune",
|
||||
},
|
||||
# Kernelized variant (if available)
|
||||
{
|
||||
"variant": "kernelized",
|
||||
"compile_mode": "max-autotune",
|
||||
"use_cache": True,
|
||||
"description": "Kernelized execution",
|
||||
},
|
||||
]
|
||||
|
||||
def _is_kernelization_available(self) -> bool:
|
||||
"""Check if kernelization is available for LLaMA."""
|
||||
try:
|
||||
from kernels import Mode, kernelize # noqa: F401
|
||||
|
||||
return True
|
||||
except ImportError:
|
||||
self.logger.debug("Kernelization not available: kernels module not found")
|
||||
return False
|
||||
|
||||
def get_default_generation_config(self) -> dict[str, Any]:
|
||||
"""Get LLaMA-specific generation configuration."""
|
||||
return {
|
||||
"do_sample": False,
|
||||
"top_p": 1.0,
|
||||
"temperature": 1.0,
|
||||
"repetition_penalty": 1.0,
|
||||
"max_new_tokens": None, # Will be set per scenario
|
||||
}
|
||||
|
||||
def get_model_init_kwargs(self, config) -> dict[str, Any]:
|
||||
"""Get LLaMA-specific model initialization kwargs."""
|
||||
return {
|
||||
"torch_dtype": getattr(torch, config.torch_dtype),
|
||||
"attn_implementation": config.attn_implementation,
|
||||
"use_cache": True,
|
||||
}
|
||||
|
||||
def get_default_torch_dtype(self) -> str:
|
||||
"""Get default torch dtype for LLaMA."""
|
||||
return "float16" # LLaMA works well with float16
|
||||
|
||||
def get_default_device(self) -> str:
|
||||
"""Get default device for LLaMA."""
|
||||
return "cuda" # LLaMA prefers CUDA
|
||||
|
||||
|
||||
def run_llama(logger, output_dir, **kwargs):
|
||||
"""
|
||||
Run LLaMA benchmark with the given configuration.
|
||||
|
||||
Args:
|
||||
logger: Logger instance
|
||||
output_dir: Output directory for results
|
||||
**kwargs: Additional configuration options
|
||||
|
||||
Returns:
|
||||
Path to output file if successful
|
||||
"""
|
||||
from benchmark_framework import BenchmarkRunner
|
||||
|
||||
# Extract parameters with defaults
|
||||
model_id = kwargs.get("model_id", "meta-llama/Llama-2-7b-hf")
|
||||
warmup_iterations = kwargs.get("warmup_iterations", 3)
|
||||
measurement_iterations = kwargs.get("measurement_iterations", 5)
|
||||
num_tokens_to_generate = kwargs.get("num_tokens_to_generate", 100)
|
||||
include_sdpa_variants = kwargs.get("include_sdpa_variants", True)
|
||||
device = kwargs.get("device", "cuda")
|
||||
torch_dtype = kwargs.get("torch_dtype", "float16")
|
||||
batch_size = kwargs.get("batch_size", 1)
|
||||
commit_id = kwargs.get("commit_id")
|
||||
|
||||
logger.info(f"Starting LLaMA benchmark for model: {model_id}")
|
||||
logger.info(
|
||||
f"Configuration: warmup={warmup_iterations}, measurement={measurement_iterations}, tokens={num_tokens_to_generate}"
|
||||
)
|
||||
|
||||
try:
|
||||
# Create benchmark instance
|
||||
benchmark = LLaMABenchmark(logger)
|
||||
|
||||
# Create scenarios
|
||||
scenarios = benchmark.create_scenarios(
|
||||
model_id=model_id,
|
||||
warmup_iterations=warmup_iterations,
|
||||
measurement_iterations=measurement_iterations,
|
||||
num_tokens_to_generate=num_tokens_to_generate,
|
||||
include_sdpa_variants=include_sdpa_variants,
|
||||
device=device,
|
||||
torch_dtype=torch_dtype,
|
||||
batch_size=batch_size,
|
||||
)
|
||||
|
||||
logger.info(f"Created {len(scenarios)} benchmark scenarios")
|
||||
|
||||
# Create runner and execute benchmarks
|
||||
runner = BenchmarkRunner(logger, output_dir)
|
||||
results = runner.run_benchmark(benchmark, scenarios, commit_id=commit_id)
|
||||
|
||||
if not results:
|
||||
logger.warning("No successful benchmark results")
|
||||
return None
|
||||
|
||||
# Save results
|
||||
model_name = model_id.split("/")[-1] # Extract model name from ID
|
||||
output_file = runner.save_results(model_name, results)
|
||||
|
||||
logger.info(f"LLaMA benchmark completed successfully. Results saved to: {output_file}")
|
||||
return output_file
|
||||
|
||||
except Exception as e:
|
||||
logger.error(f"LLaMA benchmark failed: {e}")
|
||||
import traceback
|
||||
|
||||
logger.debug(traceback.format_exc())
|
||||
raise
|
||||
1199
benchmark_v2/benchmark_framework.py
Normal file
1199
benchmark_v2/benchmark_framework.py
Normal file
File diff suppressed because it is too large
Load Diff
@ -1,242 +0,0 @@
|
||||
import hashlib
|
||||
import itertools
|
||||
import json
|
||||
import logging
|
||||
from typing import Any
|
||||
|
||||
from transformers.utils.import_utils import is_flash_attn_2_available
|
||||
|
||||
|
||||
KERNELIZATION_AVAILABLE = False
|
||||
try:
|
||||
from kernels import Mode, kernelize # noqa: F401
|
||||
|
||||
KERNELIZATION_AVAILABLE = True
|
||||
except ImportError:
|
||||
pass
|
||||
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
|
||||
class BenchmarkConfig:
|
||||
"""Configuration for a single benchmark scenario."""
|
||||
|
||||
all_attn_implementations = [
|
||||
("flash_attention_2", None),
|
||||
("eager", None),
|
||||
("sdpa", "math"),
|
||||
("sdpa", "flash_attention"),
|
||||
("flex_attention", None),
|
||||
]
|
||||
|
||||
all_compiled_modes = [None, "default", "reduce-overhead", "max-autotune", "max-autotune-no-cudagraphs"]
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
warmup_iterations: int = 5,
|
||||
measurement_iterations: int = 20,
|
||||
gpu_monitoring: bool = True, # NOTE: you may want to disable this at times as we have obsvered it could heavily slow down benchmarks on AMD
|
||||
continuous_batching: bool = False,
|
||||
batch_size: int = 1,
|
||||
sequence_length: int = 128,
|
||||
num_tokens_to_generate: int = 128,
|
||||
attn_implementation: str = "eager",
|
||||
sdpa_backend: str | None = None,
|
||||
compile_mode: str | None = None,
|
||||
compile_options: dict[str, Any] | None = None,
|
||||
kernelize: bool = False,
|
||||
name: str | None = None,
|
||||
skip_validity_check: bool = False,
|
||||
) -> None:
|
||||
# Benchmark parameters
|
||||
self.warmup_iterations = warmup_iterations
|
||||
self.measurement_iterations = measurement_iterations
|
||||
self.gpu_monitoring = gpu_monitoring
|
||||
self.continuous_batching = continuous_batching
|
||||
# Input parameters
|
||||
self.batch_size = batch_size
|
||||
self.sequence_length = sequence_length
|
||||
self.num_tokens_to_generate = num_tokens_to_generate
|
||||
# Generation parameters
|
||||
self.attn_implementation = attn_implementation
|
||||
self.sdpa_backend = sdpa_backend
|
||||
# Optimization parameters
|
||||
self.compile_mode = compile_mode
|
||||
self.compile_options = compile_options if compile_options is not None else {}
|
||||
self.kernelize = kernelize
|
||||
# Constant parameters
|
||||
self.dtype = "torch.bfloat16"
|
||||
self.device = "cuda"
|
||||
|
||||
self.check_validity(skip_validity_check)
|
||||
self.name = name if name is not None else self.infer_name()
|
||||
|
||||
def check_validity(self, skip_validity_check: bool = False) -> None:
|
||||
if skip_validity_check:
|
||||
return
|
||||
# Check FA is installed
|
||||
if self.attn_implementation == "flash_attention_2" and not is_flash_attn_2_available():
|
||||
logger.warning(
|
||||
"Flash attention does not support compile mode. Defaulting to SDPA w/ flash attention backend."
|
||||
)
|
||||
self.attn_implementation = "sdpa"
|
||||
self.sdpa_backend = "flash_attention"
|
||||
# Flash attention does not support compile mode, so we turn it off # FIXME: it would be better to support it
|
||||
is_fa = self.attn_implementation == "flash_attention_2"
|
||||
is_fa |= self.attn_implementation == "sdpa" and self.sdpa_backend == "flash_attention"
|
||||
if is_fa:
|
||||
logger.warning("Flash attention does not support compile mode. Turning off compile mode.")
|
||||
self.compile_mode = None
|
||||
# Handle SDPA backend if not determined by the config (needs to be done before skipping duplicates)
|
||||
if self.attn_implementation == "sdpa" and self.sdpa_backend is None:
|
||||
default_backend = "flash_attention" # FIXME: torch has a _cur_sdpa_kernel_backends but it fails
|
||||
logger.warning(f"No SDPA backend provided, using {default_backend} instead.")
|
||||
self.sdpa_backend = default_backend
|
||||
if self.continuous_batching:
|
||||
if self.attn_implementation == "flex_attention":
|
||||
logger.error(
|
||||
"disabling continuous batching because of invalid configuration: flex attention is not supported"
|
||||
)
|
||||
self.continuous_batching = False
|
||||
elif self.attn_implementation == "sdpa" and self.sdpa_backend is not None:
|
||||
logger.warning(
|
||||
"when continuous batching is enabled, sdpa_backend must be None because of the attention mask, setting it to None"
|
||||
)
|
||||
self.sdpa_backend = "math"
|
||||
|
||||
@property
|
||||
def hash(self) -> str:
|
||||
return hashlib.sha256(json.dumps(self.to_dict()).encode()).hexdigest()
|
||||
|
||||
def infer_name(self, compact: bool = True) -> str:
|
||||
"""Infer a human-readable name for the benchmark config, either compact or verbose."""
|
||||
if compact:
|
||||
iter_str = f"w{self.warmup_iterations}_i{self.measurement_iterations}"
|
||||
gpu_monitor_str = "monitored" if self.gpu_monitoring else "unmonitored"
|
||||
dimensions_str = f"b{self.batch_size}_s{self.sequence_length}_n{self.num_tokens_to_generate}"
|
||||
attn_code = self.attn_implementation
|
||||
attn_code += f"_{self.sdpa_backend}" if self.attn_implementation == "sdpa" else ""
|
||||
compile_str = f"compiled_{self.compile_mode}" if self.compile_mode is not None else "uncompiled"
|
||||
kernelize_str = "kernelized" if self.kernelize else "unkernelized"
|
||||
continuous_batching_str = "cb" if self.continuous_batching else "generate"
|
||||
sep = "-"
|
||||
else:
|
||||
iter_str = f"{self.warmup_iterations} warmup, {self.measurement_iterations} iterations"
|
||||
gpu_monitor_str = ("with" if self.gpu_monitoring else "no") + " GPU monitoring"
|
||||
dimensions_str = f"batch size {self.batch_size}, sequence length {self.sequence_length}, {self.num_tokens_to_generate} generated tokens"
|
||||
attn_code = f"{self.attn_implementation} attention"
|
||||
attn_code += f" with {self.sdpa_backend} backend" if self.attn_implementation == "sdpa" else ""
|
||||
compile_str = "compiled" if self.compile_mode is not None else "not compiled"
|
||||
kernelize_str = "kernelized" if self.kernelize else "not kernelized"
|
||||
continuous_batching_str = "continuous batching" if self.continuous_batching else "regular generate"
|
||||
sep = ", "
|
||||
return sep.join(
|
||||
[iter_str, gpu_monitor_str, dimensions_str, attn_code, compile_str, kernelize_str, continuous_batching_str]
|
||||
)
|
||||
|
||||
def to_dict(self) -> dict[str, Any]:
|
||||
return {
|
||||
"name": self.name,
|
||||
"warmup_iterations": self.warmup_iterations,
|
||||
"measurement_iterations": self.measurement_iterations,
|
||||
"gpu_monitoring": self.gpu_monitoring,
|
||||
"continuous_batching": self.continuous_batching,
|
||||
"batch_size": self.batch_size,
|
||||
"sequence_length": self.sequence_length,
|
||||
"num_tokens_to_generate": self.num_tokens_to_generate,
|
||||
"attn_implementation": self.attn_implementation,
|
||||
"sdpa_backend": self.sdpa_backend,
|
||||
"compile_mode": self.compile_mode,
|
||||
"compile_options": self.compile_options | {}, # to avoid inplace modification of the original dict
|
||||
"kernelize": self.kernelize,
|
||||
}
|
||||
|
||||
@classmethod
|
||||
def from_dict(cls, data: dict[str, Any], skip_validity_check: bool = False) -> "BenchmarkConfig":
|
||||
return cls(
|
||||
warmup_iterations=data.get("warmup_iterations", 5),
|
||||
measurement_iterations=data.get("measurement_iterations", 20),
|
||||
gpu_monitoring=data.get("gpu_monitoring", False),
|
||||
continuous_batching=data.get("continuous_batching", False),
|
||||
batch_size=data.get("batch_size", 1),
|
||||
sequence_length=data.get("sequence_length", 128),
|
||||
num_tokens_to_generate=data.get("num_tokens_to_generate", 128),
|
||||
attn_implementation=data.get("attn_implementation", "eager"),
|
||||
sdpa_backend=data.get("sdpa_backend"),
|
||||
compile_mode=data.get("compile_mode"),
|
||||
compile_options=data.get("compile_options"),
|
||||
kernelize=data.get("kernelize", False),
|
||||
name=data.get("name"),
|
||||
skip_validity_check=skip_validity_check,
|
||||
)
|
||||
|
||||
|
||||
def adapt_configs(
|
||||
configs: list[BenchmarkConfig],
|
||||
warmup_iterations: int | list[int] = 5,
|
||||
measurement_iterations: int | list[int] = 20,
|
||||
batch_size: int | list[int] = 1,
|
||||
sequence_length: int | list[int] = 128,
|
||||
num_tokens_to_generate: int | list[int] = 128,
|
||||
gpu_monitoring: bool | list[bool] = True,
|
||||
) -> list[BenchmarkConfig]:
|
||||
parameters = (
|
||||
x if isinstance(x, list) else [x]
|
||||
for x in [
|
||||
warmup_iterations,
|
||||
measurement_iterations,
|
||||
batch_size,
|
||||
sequence_length,
|
||||
num_tokens_to_generate,
|
||||
gpu_monitoring,
|
||||
]
|
||||
)
|
||||
iterator = itertools.product(*parameters)
|
||||
|
||||
adapted_configs = []
|
||||
for warmup_iters, measurement_iters, bs, seqlen, ntok, monitor in iterator:
|
||||
for config in configs:
|
||||
config = config.to_dict()
|
||||
config["warmup_iterations"] = warmup_iters
|
||||
config["measurement_iterations"] = measurement_iters
|
||||
config["batch_size"] = bs
|
||||
config["sequence_length"] = seqlen
|
||||
config["num_tokens_to_generate"] = ntok
|
||||
config["gpu_monitoring"] = monitor
|
||||
adapted_configs.append(BenchmarkConfig.from_dict(config))
|
||||
return adapted_configs
|
||||
|
||||
|
||||
def get_config_by_level(level: int) -> list[BenchmarkConfig]:
|
||||
configs = []
|
||||
# Early return if level is greater than 3: we generate all combinations of configs, maybe even w/ all compile modes
|
||||
if level >= 3:
|
||||
for attn_implementation, sdpa_backend in BenchmarkConfig.all_attn_implementations:
|
||||
# Usually there is not much to gain by compiling with other modes, but we allow it for level 4
|
||||
compile_modes = BenchmarkConfig.all_compiled_modes if level >= 4 else [None, "default"]
|
||||
for cm in compile_modes:
|
||||
for kernelize_on in {False, KERNELIZATION_AVAILABLE}:
|
||||
for cb_on in [False, True]:
|
||||
configs.append(
|
||||
BenchmarkConfig(
|
||||
attn_implementation=attn_implementation,
|
||||
sdpa_backend=sdpa_backend,
|
||||
compile_mode=cm,
|
||||
kernelize=kernelize_on,
|
||||
continuous_batching=cb_on,
|
||||
)
|
||||
)
|
||||
return configs
|
||||
# Otherwise, we add the configs for the given level
|
||||
if level >= 0:
|
||||
configs.append(BenchmarkConfig(attn_implementation="flex_attention", compile_mode="default"))
|
||||
if level >= 1:
|
||||
configs.append(BenchmarkConfig(attn_implementation="flash_attention_2"))
|
||||
configs.append(BenchmarkConfig(attn_implementation="eager", compile_mode="default"))
|
||||
configs.append(BenchmarkConfig(attn_implementation="flash_attention_2", continuous_batching=True))
|
||||
if level >= 2:
|
||||
configs.append(BenchmarkConfig(attn_implementation="sdpa", compile_mode="default"))
|
||||
configs.append(BenchmarkConfig(attn_implementation="flex_attention", compile_mode="default", kernelize=True))
|
||||
configs.append(BenchmarkConfig(attn_implementation="flash_attention_2", kernelize=True))
|
||||
configs.append(BenchmarkConfig(attn_implementation="paged|sdpa", continuous_batching=True))
|
||||
return configs
|
||||
@ -1,507 +0,0 @@
|
||||
import gc
|
||||
import json
|
||||
import logging
|
||||
import os
|
||||
import pathlib
|
||||
import re
|
||||
import tempfile
|
||||
import time
|
||||
from contextlib import nullcontext
|
||||
from datetime import datetime
|
||||
from queue import Queue
|
||||
from typing import Any
|
||||
|
||||
import torch
|
||||
from datasets import Dataset
|
||||
from huggingface_hub import HfApi
|
||||
from tqdm import trange
|
||||
|
||||
from transformers import (
|
||||
AutoModelForCausalLM,
|
||||
AutoTokenizer,
|
||||
CompileConfig,
|
||||
GenerationConfig,
|
||||
GenerationMixin,
|
||||
)
|
||||
from transformers.generation.streamers import BaseStreamer
|
||||
|
||||
from .benchmark_config import BenchmarkConfig
|
||||
from .data_classes import BenchmarkMetadata, BenchmarkResult, GPURawMetrics, pretty_print_dict
|
||||
from .hardware_metrics import GPUMonitor
|
||||
|
||||
|
||||
try:
|
||||
from kernels import Mode, kernelize # noqa: F401
|
||||
except ImportError:
|
||||
kernelize = None
|
||||
Mode = None
|
||||
|
||||
|
||||
DEFAULT_PROMPT = "\n".join([
|
||||
"The French Revolution was a period of political and societal change in France that began with the Estates General of 1789 and ended with the Coup of 18 Brumaire on 9 November 1799.",
|
||||
"Many of the revolution's ideas are considered fundamental principles of liberal democracy, and its values remain central to modern French political discourse.",
|
||||
"It was caused by a combination of social, political, and economic factors which the existing regime proved unable to manage.",
|
||||
"Financial crisis and widespread social distress led to the convocation of the Estates General in May 1789, its first meeting since 1614.",
|
||||
"The representatives of the Third Estate broke away and re-constituted themselves as a National Assembly in June.",
|
||||
"The Storming of the Bastille in Paris on 14 July led to a series of radical measures by the Assembly, including the abolition of feudalism, state control over the Catholic Church in France, and issuing the Declaration of the Rights of Man and of the Citizen.",
|
||||
"The next three years were dominated by a struggle for political control.",
|
||||
"King Louis XVI's attempted flight to Varennes in June 1791 further discredited the monarchy, and military defeats after the outbreak of the French Revolutionary Wars in April 1792 led to the insurrection of 10 August 1792.",
|
||||
"As a result, the monarchy was replaced by the French First Republic in September, followed by the execution of Louis XVI himself in January 1793.",
|
||||
"After another revolt in June 1793, the constitution was suspended, and political power passed from the National Convention to the Committee of Public Safety, dominated by radical Jacobins led by Maximilien Robespierre.",
|
||||
"About 16,000 people were sentenced by the Revolutionary Tribunal and executed in the Reign of Terror, which ended in July 1794 with the Thermidorian Reaction.",
|
||||
"Weakened by external threats and internal opposition, the Committee of Public Safety was replaced in November 1795 by the Directory.",
|
||||
"Its instability ended in the coup of 18 Brumaire and the establishment of the Consulate, with Napoleon Bonaparte as First Consul.",
|
||||
]) # fmt: skip
|
||||
|
||||
PUSH_TO_HUB_TOKEN = os.getenv("PUSH_TO_HUB_TOKEN", None)
|
||||
|
||||
|
||||
def compact_json_numeric_arrays(data: dict):
|
||||
# Match arrays that contain only numbers (ints/floats), whitespace, commas, and newlines
|
||||
pattern = r"\[\s*\n\s*((?:\d+(?:\.\d+)?\s*,\s*)*\d+(?:\.\d+)?)\s*\n\s*\]"
|
||||
|
||||
def replace_numeric_array(match):
|
||||
# Get the array content
|
||||
content = match.group(1)
|
||||
# Remove extra whitespace but keep commas
|
||||
compact_content = re.sub(r"\s+", " ", content).strip()
|
||||
return f"[{compact_content}]"
|
||||
|
||||
return re.sub(pattern, replace_numeric_array, json.dumps(data, indent=4, default=str), flags=re.DOTALL)
|
||||
|
||||
|
||||
def get_git_revision() -> str:
|
||||
base_path = pathlib.Path(__file__).parent.parent.parent
|
||||
git_dir = base_path / ".git"
|
||||
with (git_dir / "HEAD").open("r") as head:
|
||||
ref = head.readline().split(" ")[-1].strip()
|
||||
with (git_dir / ref).open("r") as git_hash:
|
||||
return git_hash.readline().strip()
|
||||
|
||||
|
||||
def get_sdpa_backend(backend_name: str | None) -> torch.nn.attention.SDPBackend | None:
|
||||
"""Get the SDPA backend enum from string name."""
|
||||
if backend_name is None:
|
||||
return None
|
||||
|
||||
try:
|
||||
backend_map = {
|
||||
"math": torch.nn.attention.SDPBackend.MATH,
|
||||
"flash_attention": torch.nn.attention.SDPBackend.FLASH_ATTENTION,
|
||||
"efficient_attention": torch.nn.attention.SDPBackend.EFFICIENT_ATTENTION,
|
||||
"cudnn_attention": torch.nn.attention.SDPBackend.CUDNN_ATTENTION,
|
||||
}
|
||||
return backend_map.get(backend_name.lower())
|
||||
except AttributeError:
|
||||
# torch.nn.attention.SDPBackend not available in older torch versions
|
||||
return None
|
||||
|
||||
|
||||
def flush_memory():
|
||||
"""Flush GPU memory and run garbage collection."""
|
||||
gc.collect()
|
||||
# Dynamo resets
|
||||
torch._dynamo.reset()
|
||||
torch._dynamo.reset_code_caches()
|
||||
if hasattr(torch._inductor, "codecache"):
|
||||
# Clear FX graph cache
|
||||
if hasattr(torch._inductor.codecache, "FxGraphCache"):
|
||||
torch._inductor.codecache.FxGraphCache.clear()
|
||||
# Clear PyCodeCache
|
||||
if hasattr(torch._inductor.codecache, "PyCodeCache"):
|
||||
torch._inductor.codecache.PyCodeCache.cache_clear()
|
||||
# Clear TritonFuture cache (for async compilation)
|
||||
if hasattr(torch._inductor.codecache, "TritonFuture"):
|
||||
if hasattr(torch._inductor.codecache.TritonFuture, "_compile_cache"):
|
||||
torch._inductor.codecache.TritonFuture._compile_cache.clear()
|
||||
# Clear CUDA cache
|
||||
if torch.cuda.is_available():
|
||||
torch.cuda.empty_cache()
|
||||
torch.cuda.synchronize()
|
||||
gc.collect()
|
||||
|
||||
|
||||
class BenchmarkStreamer(BaseStreamer):
|
||||
def __init__(self, **kwargs) -> None:
|
||||
self.timeout = kwargs.pop("timeout", 10)
|
||||
self.timestamps = []
|
||||
self.text_queue = Queue()
|
||||
self.stop_signal = None
|
||||
|
||||
def put(self, value):
|
||||
"""Receives tokens and logs the timestamp of the generation."""
|
||||
self.timestamps.append(time.perf_counter())
|
||||
self.text_queue.put(value)
|
||||
|
||||
def end(self):
|
||||
self.timestamps.append(time.perf_counter())
|
||||
self.text_queue.put(self.stop_signal)
|
||||
|
||||
def __iter__(self):
|
||||
return self
|
||||
|
||||
def __next__(self):
|
||||
value = self.text_queue.get(timeout=self.timeout)
|
||||
if value == self.stop_signal:
|
||||
raise StopIteration()
|
||||
else:
|
||||
return value
|
||||
|
||||
|
||||
class BenchmarkRunner:
|
||||
"""Main benchmark runner that coordinates benchmark execution."""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
logger: logging.Logger,
|
||||
output_dir: str | None = None,
|
||||
branch_name: str | None = None,
|
||||
commit_id: str | None = None,
|
||||
commit_message: str | None = None,
|
||||
) -> None:
|
||||
# Those stay constant for the whole run
|
||||
self.logger = logger
|
||||
if output_dir is None:
|
||||
output_dir = os.path.join(os.path.dirname(os.path.dirname(__file__)), "benchmark_results")
|
||||
self.output_dir = output_dir
|
||||
self.branch_name = branch_name
|
||||
self.commit_id = get_git_revision() if commit_id is None else commit_id
|
||||
self.commit_message = commit_message
|
||||
os.makedirs(self.output_dir, exist_ok=True)
|
||||
self.profile_dir = None
|
||||
# Attributes that are reset for each model
|
||||
self._setup_for = ""
|
||||
# Attributes that are reset for each run
|
||||
self.model: GenerationMixin | None = None
|
||||
|
||||
def cleanup(self) -> None:
|
||||
del self.model
|
||||
self.model = None
|
||||
flush_memory()
|
||||
|
||||
def setup_benchmark(self, model_id: str, config: BenchmarkConfig) -> None:
|
||||
# Some attributes only need to be set once per model
|
||||
if self._setup_for != model_id:
|
||||
self.tokenizer = AutoTokenizer.from_pretrained(model_id)
|
||||
# We set the EOS token to the padding token for open-ended generation
|
||||
self.tokenizer.eos_token = self.tokenizer.pad_token
|
||||
self._setup_for = model_id
|
||||
|
||||
# Prepare inputs
|
||||
self.inputs = self.tokenizer(
|
||||
[DEFAULT_PROMPT for _ in range(config.batch_size)],
|
||||
return_tensors="pt",
|
||||
max_length=config.sequence_length,
|
||||
truncation=True,
|
||||
return_attention_mask=True,
|
||||
).to(config.device)
|
||||
self.inputs["use_cache"] = True
|
||||
|
||||
# Prepare generation config
|
||||
gen_config = GenerationConfig(
|
||||
do_sample=False, top_p=1.0, temperature=1.0, max_new_tokens=config.num_tokens_to_generate
|
||||
)
|
||||
|
||||
# Prepare compile config
|
||||
if config.compile_mode is not None:
|
||||
gen_config.compile_config = CompileConfig(mode=config.compile_mode, options=config.compile_options)
|
||||
gen_config.cache_implementation = "static"
|
||||
|
||||
# Load model
|
||||
self.logger.debug(f"Loading model {model_id} on device {config.device}...")
|
||||
dtype = getattr(torch, config.dtype.removeprefix("torch."))
|
||||
self.model = AutoModelForCausalLM.from_pretrained(
|
||||
model_id, dtype=dtype, attn_implementation=config.attn_implementation, generation_config=gen_config
|
||||
)
|
||||
self.model = self.model.eval().to(config.device)
|
||||
|
||||
# Kernelize the model if needed
|
||||
if config.kernelize and kernelize is not None and Mode is not None:
|
||||
self.model = kernelize(self.model, mode=Mode.INFERENCE)
|
||||
|
||||
def run_benchmark(
|
||||
self, model_id: str, config: BenchmarkConfig, num_tokens_to_profile: int = 0
|
||||
) -> dict[str, Any] | None:
|
||||
"""Run a single benchmark with the given model ID and config."""
|
||||
sdpa_ctx = nullcontext()
|
||||
if config.attn_implementation == "sdpa":
|
||||
sdpa_backend = get_sdpa_backend(config.sdpa_backend)
|
||||
sdpa_ctx = torch.nn.attention.sdpa_kernel(sdpa_backend)
|
||||
|
||||
with sdpa_ctx, torch.no_grad():
|
||||
self.logger.info(f"Running benchmark scenario: {config.name}")
|
||||
|
||||
# Quick validation: try one measurement first to see if this scenario works
|
||||
generate_fn = self.time_generate_batch if config.continuous_batching else self.time_generate
|
||||
flush_memory()
|
||||
e2e_latency, token_generation_times, shape_and_decoded_output, gpu_metrics = generate_fn(
|
||||
max_new_tokens=1, gpu_monitor=None
|
||||
)
|
||||
if e2e_latency < 0:
|
||||
self.logger.warning(f"Skipping config {config.name}: {e2e_latency = } (no GPU monitoring)")
|
||||
return None
|
||||
|
||||
# Warmup runs
|
||||
self.logger.info(f"Warming up with {config.warmup_iterations} iterations...")
|
||||
for _ in trange(config.warmup_iterations):
|
||||
_ = generate_fn(max_new_tokens=config.num_tokens_to_generate)
|
||||
self.logger.info("Warmup over.")
|
||||
|
||||
# Measurement runs
|
||||
result = BenchmarkResult()
|
||||
self.logger.info(f"Benchmarking with {config.measurement_iterations} iterations.")
|
||||
for _ in trange(config.measurement_iterations):
|
||||
e2e_latency, token_generation_times, shape_and_decoded_output, gpu_metrics = generate_fn(
|
||||
max_new_tokens=config.num_tokens_to_generate,
|
||||
gpu_monitor=(GPUMonitor(logger=self.logger) if config.gpu_monitoring else None),
|
||||
)
|
||||
result.accumulate(e2e_latency, token_generation_times, shape_and_decoded_output, gpu_metrics)
|
||||
self.logger.info("Benchmarking done. Cleaning up.")
|
||||
|
||||
# Profile if needed
|
||||
if num_tokens_to_profile > 0:
|
||||
self.profile_generate(num_tokens_to_profile, config.name)
|
||||
|
||||
return {
|
||||
"metadata": BenchmarkMetadata(
|
||||
model_id=model_id,
|
||||
branch_name=self.branch_name,
|
||||
commit_id=self.commit_id,
|
||||
commit_message=self.commit_message,
|
||||
),
|
||||
"measurements": result,
|
||||
"config": config,
|
||||
}
|
||||
|
||||
# TODO: refactor `generate_batch` to handle streaming so we can use it here
|
||||
def time_generate_batch(
|
||||
self,
|
||||
max_new_tokens: int,
|
||||
gpu_monitor: GPUMonitor | None = None,
|
||||
) -> tuple[float, list[float], str, GPURawMetrics | None]:
|
||||
if gpu_monitor is not None:
|
||||
gpu_monitor.start()
|
||||
config = GenerationConfig(
|
||||
max_new_tokens=max_new_tokens,
|
||||
eos_token_id=self.tokenizer.eos_token_id,
|
||||
pad_token_id=self.tokenizer.pad_token_id,
|
||||
do_sample=True,
|
||||
)
|
||||
manager = self.model.init_continuous_batching(config)
|
||||
manager.start()
|
||||
try:
|
||||
first_req_results = []
|
||||
timestamps = []
|
||||
wall_time_0 = time.perf_counter()
|
||||
inputs = self.inputs["input_ids"].tolist()
|
||||
manager.add_requests(inputs, max_new_tokens=max_new_tokens, streaming=True)
|
||||
first_req_id = None
|
||||
num_requests = len(inputs)
|
||||
finished_requests = 0
|
||||
while finished_requests < num_requests:
|
||||
# NOTE: I don't like having the extra if stmt here, but hopefully won't degrade perf too much
|
||||
result = manager.get_result()
|
||||
if result:
|
||||
timestamps.append(time.perf_counter() - wall_time_0)
|
||||
if result.is_finished():
|
||||
finished_requests += 1
|
||||
if first_req_id is None:
|
||||
first_req_id = result.request_id
|
||||
if result.request_id == first_req_id:
|
||||
first_req_results.append(result)
|
||||
else:
|
||||
if not manager.is_running():
|
||||
raise RuntimeError("Generation thread exited unexpectedly")
|
||||
wall_time_1 = time.perf_counter()
|
||||
gpu_metrics = gpu_monitor.stop_and_collect() if gpu_monitor is not None else None
|
||||
decoded_output = self.tokenizer.decode(
|
||||
[res.generated_tokens[0] for res in first_req_results], skip_special_tokens=True
|
||||
)
|
||||
shape_and_decoded_output = f"{(1, len(first_req_results))} | {decoded_output}"
|
||||
e2e_latency = wall_time_1 - wall_time_0
|
||||
return e2e_latency, timestamps, shape_and_decoded_output, gpu_metrics
|
||||
except Exception as e:
|
||||
raise e
|
||||
finally:
|
||||
manager.stop()
|
||||
|
||||
def time_generate(
|
||||
self,
|
||||
max_new_tokens: int,
|
||||
gpu_monitor: GPUMonitor | None = None,
|
||||
) -> tuple[float, list[float], str, GPURawMetrics | None]:
|
||||
"""Time the latency of a call to model.generate() with the given (inputs) and (max_new_tokens)."""
|
||||
# Prepare gpu monitoring if needed
|
||||
if gpu_monitor is not None:
|
||||
gpu_monitor.start()
|
||||
# Prepare streamer
|
||||
streamer = BenchmarkStreamer()
|
||||
# Generate and time
|
||||
wall_time_0 = time.perf_counter()
|
||||
outputs = self.model.generate(
|
||||
**self.inputs,
|
||||
max_new_tokens=max_new_tokens,
|
||||
streamer=streamer,
|
||||
)
|
||||
wall_time_1 = time.perf_counter()
|
||||
# Stop gpu monitoring if needed
|
||||
gpu_metrics = gpu_monitor.stop_and_collect() if gpu_monitor is not None else None
|
||||
# Check if generation had the right number of tokens
|
||||
input_tokens = self.inputs["input_ids"].size(-1)
|
||||
batch_size, output_tokens = outputs.shape
|
||||
new_tokens = output_tokens - input_tokens
|
||||
if new_tokens != max_new_tokens:
|
||||
raise RuntimeError(f"Generated {new_tokens} tokens, expected {max_new_tokens}")
|
||||
# Decode outputs
|
||||
decoded_output = self.tokenizer.decode(outputs[0, input_tokens:], skip_special_tokens=True)
|
||||
shape_and_decoded_output = f"{tuple(outputs.shape)} | {decoded_output}"
|
||||
# Compute intermediate quantities
|
||||
e2e_latency = wall_time_1 - wall_time_0
|
||||
token_generation_times = [t - wall_time_0 for t in streamer.timestamps[1:]]
|
||||
return e2e_latency, token_generation_times, shape_and_decoded_output, gpu_metrics
|
||||
|
||||
def profile_generate(self, num_tokens_to_profile: int, config_name: str) -> None:
|
||||
"""Profile the latency of a call to model.generate() with the given (inputs) and (max_new_tokens)."""
|
||||
profiler = torch.profiler.profile(
|
||||
activities=[torch.profiler.ProfilerActivity.CPU, torch.profiler.ProfilerActivity.CUDA],
|
||||
record_shapes=True,
|
||||
)
|
||||
with profiler as prof:
|
||||
_ = self.model.generate(
|
||||
**self.inputs,
|
||||
max_new_tokens=num_tokens_to_profile,
|
||||
)
|
||||
if self.profile_dir is None:
|
||||
self.profile_dir = self.output_dir + "_profiles"
|
||||
os.makedirs(self.profile_dir, exist_ok=True)
|
||||
prof.export_chrome_trace(f"{self.profile_dir}/{config_name}.json")
|
||||
|
||||
def run_benchmarks(
|
||||
self,
|
||||
model_id: str,
|
||||
benchmark_configs: list[BenchmarkConfig],
|
||||
num_tokens_to_profile: int = 0,
|
||||
pretty_print_summary: bool = True,
|
||||
) -> tuple[str, dict[str, Any]]:
|
||||
"""Run multiple benchmarks for the given model ID and list of benchmark configs."""
|
||||
all_results = {}
|
||||
timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
|
||||
start_time = time.perf_counter()
|
||||
|
||||
n_configs = len(benchmark_configs)
|
||||
for i, config in enumerate(benchmark_configs):
|
||||
# Skip if already run
|
||||
if config.hash in all_results:
|
||||
self.logger.info(f"Skipping duplicate config {config.name} for model {model_id} ({i + 1}/{n_configs})")
|
||||
continue
|
||||
|
||||
# Otherwise, run the benchmark
|
||||
self.setup_benchmark(model_id, config)
|
||||
self.logger.info(
|
||||
f"Running benchmark of model {model_id} with scenario: {config.name} ({i + 1}/{n_configs})"
|
||||
)
|
||||
|
||||
# Launch benchmark in a try/except block to avoid stopping the whole run if one benchmark fails
|
||||
try:
|
||||
results = self.run_benchmark(model_id, config, num_tokens_to_profile)
|
||||
if results is not None:
|
||||
all_results[config.hash] = results
|
||||
|
||||
except Exception as e:
|
||||
self.logger.error(f"Error running with scenario: {config.name}:\n{repr(e)}")
|
||||
# Cleanup model and save results
|
||||
self.cleanup()
|
||||
self.save_results(model_id, all_results, timestamp=timestamp)
|
||||
|
||||
if len(all_results) < 1:
|
||||
raise RuntimeError("No benchmark was run succesfully")
|
||||
|
||||
if pretty_print_summary:
|
||||
print()
|
||||
print("=" * 100)
|
||||
print(f"Finished benchmarks in {time.perf_counter() - start_time:.2f} seconds")
|
||||
print(f"Total number of benchmarks: {len(all_results)}")
|
||||
print("First run metadata:")
|
||||
first_key = list(all_results.keys())[0]
|
||||
first_metadata = all_results[first_key]["metadata"].to_dict()
|
||||
hardware_info = first_metadata.pop("hardware_info")
|
||||
pretty_print_dict(first_metadata | hardware_info, tabs=1)
|
||||
for result in all_results.values():
|
||||
print("=" * 100)
|
||||
print(f"Config: {result['config'].infer_name(compact=False)}\n")
|
||||
result["measurements"].pprint(
|
||||
batch_size=result["config"].batch_size,
|
||||
num_generated_tokens=result["config"].num_tokens_to_generate,
|
||||
tabs=1,
|
||||
)
|
||||
print("=" * 100)
|
||||
|
||||
return (timestamp, all_results)
|
||||
|
||||
def save_results(self, model_name: str, results: dict, timestamp: str = "") -> str:
|
||||
"""Save benchmark results to JSON file."""
|
||||
# Create model-specific subdirectory
|
||||
model_name = model_name.replace("/", "_")
|
||||
model_dir = os.path.join(self.output_dir, model_name)
|
||||
os.makedirs(model_dir, exist_ok=True)
|
||||
|
||||
# Create filename with timestamp
|
||||
timestamp = datetime.now().strftime("%Y%m%d_%H%M%S") if not timestamp else timestamp
|
||||
filename = f"{model_name}_benchmark_{timestamp}.json"
|
||||
filepath = os.path.join(model_dir, filename)
|
||||
|
||||
# Convert results to dict
|
||||
converted_results = {}
|
||||
for cfg_hash in results.keys():
|
||||
converted_results[cfg_hash] = {
|
||||
"metadata": results[cfg_hash]["metadata"].to_dict(),
|
||||
"measurements": results[cfg_hash]["measurements"].to_dict(),
|
||||
"config": results[cfg_hash]["config"].to_dict(),
|
||||
}
|
||||
|
||||
# Save to JSON file
|
||||
with open(filepath, "w") as f:
|
||||
f.write(compact_json_numeric_arrays(converted_results))
|
||||
|
||||
self.logger.info(f"Results saved to {filepath}")
|
||||
return filepath
|
||||
|
||||
def push_results_to_hub(self, dataset_id: str, results: dict[Any, Any], timestamp: str) -> None:
|
||||
if PUSH_TO_HUB_TOKEN is None:
|
||||
raise ValueError(
|
||||
"PUSH_TO_HUB_TOKEN is not set, cannot push results to the Hub. When setting dataset_id, please also set the PUSH_TO_HUB_TOKEN environment variable."
|
||||
)
|
||||
|
||||
n_results = len(results)
|
||||
self.logger.info(f"Pushing {n_results} results to: {dataset_id}")
|
||||
rows = []
|
||||
for cfg_hash, entry in results.items():
|
||||
row = {
|
||||
"benchmark_config_hash": cfg_hash,
|
||||
"config": entry["config"].to_dict(),
|
||||
"measurements": entry["measurements"].to_dict(),
|
||||
"metadata": entry["metadata"].to_dict(),
|
||||
}
|
||||
rows.append(row)
|
||||
|
||||
ds = Dataset.from_list(rows)
|
||||
with tempfile.TemporaryDirectory() as tmp:
|
||||
jsonl_path = os.path.join(tmp, "data.jsonl")
|
||||
with open(jsonl_path, "w") as f:
|
||||
json_lines = []
|
||||
for ex in ds:
|
||||
json_lines.append(json.dumps(ex, ensure_ascii=False))
|
||||
f.write("\n".join(json_lines))
|
||||
|
||||
api = HfApi()
|
||||
# NOTE: we expect the repository to already exist
|
||||
timestamp = datetime.now().strftime("%Y%m%d_%H%M%S") if not timestamp else timestamp
|
||||
file_name = f"benchmark_run_{timestamp}.jsonl"
|
||||
api.upload_file(
|
||||
path_or_fileobj=jsonl_path,
|
||||
path_in_repo=file_name,
|
||||
repo_id=dataset_id,
|
||||
repo_type="dataset",
|
||||
token=PUSH_TO_HUB_TOKEN,
|
||||
)
|
||||
self.logger.info(f"Succesfully uploaded results to: {dataset_id}")
|
||||
@ -1,160 +0,0 @@
|
||||
from dataclasses import dataclass
|
||||
from datetime import datetime, timezone
|
||||
from typing import Any
|
||||
|
||||
import numpy as np
|
||||
|
||||
from .hardware_metrics import GPURawMetrics, HardwareInfo
|
||||
|
||||
|
||||
def compute_basic_statistics(measurements: list[float]) -> dict[str, float]:
|
||||
return {
|
||||
"avg": np.mean(measurements),
|
||||
"std": np.std(measurements),
|
||||
"min": np.min(measurements),
|
||||
"med": np.median(measurements),
|
||||
"max": np.max(measurements),
|
||||
"p95": np.percentile(measurements, 95),
|
||||
}
|
||||
|
||||
|
||||
def add_unit_to_duration(stats: dict[str, float]) -> dict[str, str]:
|
||||
for key in list(stats.keys()):
|
||||
value = stats[key]
|
||||
if value > 3600:
|
||||
stats[key] = f"{(value / 3600):.2f}hr"
|
||||
elif value > 60:
|
||||
stats[key] = f"{(value / 60):.2f}min"
|
||||
elif value > 1:
|
||||
stats[key] = f"{value:.2f}s"
|
||||
elif value > 1e-3:
|
||||
stats[key] = f"{(value * 1e3):.2f}ms"
|
||||
elif value > 1e-6:
|
||||
stats[key] = f"{(value * 1e6):.2f}us"
|
||||
else:
|
||||
stats[key] = f"{(value * 1e9):.2f}ns"
|
||||
return stats
|
||||
|
||||
|
||||
def equalize_lengths_and_collate(stats: dict[str, dict[str, str]]) -> dict[str, str]:
|
||||
"""Note: This operation is destructive as it will update values in place before returning a new correctly formatted dict"""
|
||||
keys = ["avg", "std", "min", "med", "max", "p95"]
|
||||
for key in keys:
|
||||
max_length = max(len(stat[key]) for stat in stats.values())
|
||||
for stat in stats.values():
|
||||
stat[key] = stat[key].ljust(max_length, " ")
|
||||
return {name: " ".join([f"{key}={stat[key]}" for key in keys]) for name, stat in stats.items()}
|
||||
|
||||
|
||||
def pretty_print_dict(data: dict[str, str], tabs: int = 0) -> None:
|
||||
max_key_length = max([len(key) for key in data.keys()])
|
||||
for key, value in data.items():
|
||||
tabs_str = " " * tabs
|
||||
padded_key = key.ljust(max_key_length + 1, ".")
|
||||
print(f"{tabs_str}{padded_key}: {value}")
|
||||
|
||||
|
||||
@dataclass
|
||||
class BenchmarkMetadata:
|
||||
"""Metadata collected for each benchmark run."""
|
||||
|
||||
model_id: str
|
||||
timestamp: str
|
||||
branch_name: str
|
||||
commit_id: str
|
||||
commit_message: str
|
||||
hardware_info: HardwareInfo
|
||||
|
||||
def __init__(self, model_id: str, commit_id: str, branch_name: str = "main", commit_message: str = "") -> None:
|
||||
self.model_id = model_id
|
||||
self.timestamp = datetime.now(timezone.utc).isoformat()
|
||||
self.branch_name = branch_name
|
||||
self.commit_id = commit_id
|
||||
self.commit_message = commit_message
|
||||
self.hardware_info = HardwareInfo()
|
||||
|
||||
def to_dict(self) -> dict[str, Any]:
|
||||
return {
|
||||
"model_id": self.model_id,
|
||||
"timestamp": self.timestamp,
|
||||
"branch_name": self.branch_name,
|
||||
"commit_id": self.commit_id,
|
||||
"commit_message": self.commit_message,
|
||||
"hardware_info": self.hardware_info.to_dict(),
|
||||
}
|
||||
|
||||
|
||||
class BenchmarkResult:
|
||||
"""Result from a series of benchmark runs."""
|
||||
|
||||
def __init__(self) -> None:
|
||||
self.e2e_latency = []
|
||||
self.token_generation_times = [] # time at which each token was generated (relative to start of the generation)
|
||||
self.shape_and_decoded_outputs = []
|
||||
self.gpu_metrics = []
|
||||
|
||||
def accumulate(
|
||||
self,
|
||||
e2e_latency: float,
|
||||
token_generation_times: list[float],
|
||||
shape_and_decoded_output: str,
|
||||
gpu_metrics: GPURawMetrics | None,
|
||||
) -> None:
|
||||
self.e2e_latency.append(e2e_latency)
|
||||
self.token_generation_times.append(token_generation_times)
|
||||
self.shape_and_decoded_outputs.append(shape_and_decoded_output)
|
||||
self.gpu_metrics.append(gpu_metrics)
|
||||
|
||||
def to_dict(self) -> dict[str, None | int | float]:
|
||||
# Save GPU metrics as None if it contains only None values
|
||||
if all(gm is None for gm in self.gpu_metrics):
|
||||
gpu_metrics = None
|
||||
else:
|
||||
gpu_metrics = [gm.to_dict() for gm in self.gpu_metrics]
|
||||
return {
|
||||
"e2e_latency": self.e2e_latency,
|
||||
"token_generation_times": self.token_generation_times,
|
||||
"shape_and_decoded_outputs": self.shape_and_decoded_outputs,
|
||||
"gpu_metrics": gpu_metrics,
|
||||
}
|
||||
|
||||
@classmethod
|
||||
def from_dict(cls, data: dict[str, None | int | float]) -> "BenchmarkResult":
|
||||
# Handle GPU metrics, which is saved as None if it contains only None values
|
||||
if data["gpu_metrics"] is None:
|
||||
gpu_metrics = [None for _ in range(len(data["e2e_latency"]))]
|
||||
else:
|
||||
gpu_metrics = [GPURawMetrics.from_dict(gm) for gm in data["gpu_metrics"]]
|
||||
# Create a new instance and accumulate the data
|
||||
new_instance = cls()
|
||||
for i in range(len(data["e2e_latency"])):
|
||||
new_instance.accumulate(
|
||||
e2e_latency=data["e2e_latency"][i],
|
||||
token_generation_times=data["token_generation_times"][i],
|
||||
shape_and_decoded_output=data["shape_and_decoded_outputs"][i],
|
||||
gpu_metrics=gpu_metrics[i],
|
||||
)
|
||||
return new_instance
|
||||
|
||||
def get_measured_ttft(self) -> list[float]:
|
||||
return [dt[0] for dt in self.token_generation_times if len(dt) > 0]
|
||||
|
||||
def get_measured_itl(self) -> list[float]:
|
||||
return [(dt[-1] - dt[0]) / (len(dt) - 1) for dt in self.token_generation_times if len(dt) > 1]
|
||||
|
||||
def get_throughput(self, total_generated_tokens: int) -> list[float]:
|
||||
return [total_generated_tokens / e2e_latency for e2e_latency in self.e2e_latency]
|
||||
|
||||
def pprint(self, batch_size: int = 0, num_generated_tokens: int = 0, tabs: int = 0) -> None:
|
||||
measurements = {
|
||||
"E2E Latency": add_unit_to_duration(compute_basic_statistics(self.e2e_latency)),
|
||||
"Time to First Token": add_unit_to_duration(compute_basic_statistics(self.get_measured_ttft())),
|
||||
}
|
||||
itl_values = self.get_measured_itl()
|
||||
if len(itl_values) > 0:
|
||||
measurements["Inter-Token Latency"] = add_unit_to_duration(compute_basic_statistics(itl_values))
|
||||
if batch_size > 0:
|
||||
throughput_stats = compute_basic_statistics(self.get_throughput(batch_size * num_generated_tokens))
|
||||
measurements["Throughput"] = {key: f"{value:.2f}tok/s" for key, value in throughput_stats.items()}
|
||||
dict_to_pprint = equalize_lengths_and_collate(measurements)
|
||||
pretty_print_dict(dict_to_pprint, tabs=tabs)
|
||||
@ -1,171 +0,0 @@
|
||||
import json
|
||||
import logging
|
||||
import subprocess
|
||||
import sys
|
||||
import threading
|
||||
import time
|
||||
from dataclasses import dataclass
|
||||
from enum import Enum
|
||||
from logging import Logger
|
||||
|
||||
import gpustat
|
||||
import psutil
|
||||
import torch
|
||||
|
||||
|
||||
# Data class to hold the hardware information
|
||||
def get_device_name_and_memory_total() -> tuple[str, float]:
|
||||
"""Returns the name and memory total of GPU 0."""
|
||||
device_name = torch.cuda.get_device_properties(0).name
|
||||
device_memory_total = torch.cuda.get_device_properties(0).total_memory / 1024**3
|
||||
return device_name, device_memory_total
|
||||
|
||||
|
||||
class HardwareInfo:
|
||||
"""A class to hold information about the hardware."""
|
||||
|
||||
def __init__(self) -> None:
|
||||
# Retrieve GPU stats
|
||||
try:
|
||||
self.gpu_name, self.gpu_memory_total_gb = get_device_name_and_memory_total()
|
||||
except Exception:
|
||||
self.gpu_name, self.gpu_memory_total_gb = None, None
|
||||
# Retrieve python, torch and CUDA version
|
||||
self.python_version = f"{sys.version.split()[0]}"
|
||||
self.torch_version = torch.__version__
|
||||
if hasattr(torch, "cuda") and torch.cuda.is_available():
|
||||
self.cuda_version = torch.version.cuda
|
||||
else:
|
||||
self.cuda_version = None
|
||||
# Retrieve general hardware information
|
||||
self.cpu_count = psutil.cpu_count()
|
||||
self.memory_total_mb = int(psutil.virtual_memory().total / (1024 * 1024))
|
||||
|
||||
def to_dict(self) -> dict[str, None | int | float | str]:
|
||||
return {
|
||||
"gpu_name": self.gpu_name,
|
||||
"gpu_memory_total_gb": self.gpu_memory_total_gb,
|
||||
"python_version": self.python_version,
|
||||
"torch_version": self.torch_version,
|
||||
}
|
||||
|
||||
|
||||
# Functions to get information about the GPU
|
||||
def get_amd_gpu_stats() -> tuple[int, float]:
|
||||
"""Returns the utilization and memory used of an AMD GPU, both in percent"""
|
||||
rocm_smi_output = subprocess.check_output(["rocm-smi", "--json", "--showuse", "--showmeminfo", "VRAM"])
|
||||
gpu_stats = json.loads(rocm_smi_output.decode("utf-8"))
|
||||
gpu_stats = [
|
||||
(card_id, stats["GPU use (%)"], stats["VRAM Total Used Memory (B)"]) for card_id, stats in gpu_stats.items()
|
||||
]
|
||||
gpu_stats.sort(key=lambda x: x[1], reverse=True)
|
||||
return int(gpu_stats[0][1]), float(gpu_stats[0][2]) / 1024**3
|
||||
|
||||
|
||||
def get_nvidia_gpu_stats() -> tuple[int, float]:
|
||||
"""Returns the utilization and memory used of an NVIDIA GPU, both in percent"""
|
||||
gpu_stats = gpustat.GPUStatCollection.new_query()
|
||||
gpu_stats = gpu_stats[0]
|
||||
return int(gpu_stats["utilization.gpu"]), float(gpu_stats["memory.used"]) / 1024**3
|
||||
|
||||
|
||||
class GPUStatsCollector:
|
||||
"""A class to get statistics about the GPU. It serves as a wrapper that holds the GPU total memory and its name,
|
||||
which is used to call the right function to get the utilization and memory used."""
|
||||
|
||||
def __init__(self) -> None:
|
||||
self.device_name, self.device_memory_total = get_device_name_and_memory_total()
|
||||
# Monkey patch the get_utilization_and_memory_used method based on the GPU type
|
||||
if "amd" in self.device_name.lower():
|
||||
self.get_utilization_and_memory_used = get_amd_gpu_stats
|
||||
elif "nvidia" in self.device_name.lower():
|
||||
self.get_utilization_and_memory_used = get_nvidia_gpu_stats
|
||||
else:
|
||||
raise RuntimeError(f"Unsupported GPU: {self.device_name}")
|
||||
|
||||
def get_measurements(self) -> tuple[int, float]:
|
||||
"""Get the utilization and memory used of the GPU, both in percent"""
|
||||
raise NotImplementedError("This method is meant to be monkey patched during __init__")
|
||||
|
||||
|
||||
# Simple data classes to hold the raw GPU metrics
|
||||
class GPUMonitoringStatus(Enum):
|
||||
"""Status of GPU monitoring."""
|
||||
|
||||
SUCCESS = "success"
|
||||
FAILED = "failed"
|
||||
NO_GPUS_AVAILABLE = "no_gpus_available"
|
||||
NO_SAMPLES_COLLECTED = "no_samples_collected"
|
||||
|
||||
|
||||
@dataclass
|
||||
class GPURawMetrics:
|
||||
"""Raw values for GPU utilization and memory used."""
|
||||
|
||||
utilization: list[float] # in percent
|
||||
memory_used: list[float] # in GB
|
||||
timestamps: list[float] # in seconds
|
||||
timestamp_0: float # in seconds
|
||||
monitoring_status: GPUMonitoringStatus
|
||||
|
||||
def to_dict(self) -> dict[str, None | int | float | str]:
|
||||
return {
|
||||
"utilization": self.utilization,
|
||||
"memory_used": self.memory_used,
|
||||
"timestamps": self.timestamps,
|
||||
"timestamp_0": self.timestamp_0,
|
||||
"monitoring_status": self.monitoring_status.value,
|
||||
}
|
||||
|
||||
|
||||
# Main class, used to monitor the GPU utilization during benchmark execution
|
||||
class GPUMonitor:
|
||||
"""Monitor GPU utilization during benchmark execution."""
|
||||
|
||||
def __init__(self, sample_interval_sec: float = 0.1, logger: Logger | None = None):
|
||||
self.sample_interval_sec = sample_interval_sec
|
||||
self.logger = logger if logger is not None else logging.getLogger(__name__)
|
||||
|
||||
self.num_available_gpus = torch.cuda.device_count()
|
||||
if self.num_available_gpus == 0:
|
||||
raise RuntimeError("No GPUs detected by torch.cuda.device_count().")
|
||||
self.gpu_stats_getter = GPUStatsCollector()
|
||||
|
||||
def start(self):
|
||||
"""Start monitoring GPU metrics."""
|
||||
# Clear the stop event to enable monitoring
|
||||
self.stop_event = threading.Event()
|
||||
self.gpu_utilization = []
|
||||
self.gpu_memory_used = []
|
||||
self.timestamps = []
|
||||
self.thread = threading.Thread(target=self._monitor_loop)
|
||||
self.thread.start()
|
||||
self.logger.debug("GPU monitoring started")
|
||||
|
||||
def stop_and_collect(self) -> GPURawMetrics:
|
||||
"""Stop monitoring and return collected metrics."""
|
||||
self.stop_event.set()
|
||||
self.thread.join()
|
||||
if self.gpu_utilization:
|
||||
timestamp_0 = self.timestamps[0]
|
||||
metrics = GPURawMetrics(
|
||||
utilization=self.gpu_utilization,
|
||||
memory_used=self.gpu_memory_used,
|
||||
timestamps=[t - timestamp_0 for t in self.timestamps],
|
||||
timestamp_0=timestamp_0,
|
||||
monitoring_status=GPUMonitoringStatus.SUCCESS,
|
||||
)
|
||||
self.logger.debug(f"GPU monitoring completed: {len(self.gpu_utilization)} samples collected")
|
||||
else:
|
||||
metrics = GPURawMetrics(monitoring_status=GPUMonitoringStatus.NO_SAMPLES_COLLECTED)
|
||||
return metrics
|
||||
|
||||
def _monitor_loop(self):
|
||||
"""Background monitoring loop using threading.Event for communication."""
|
||||
while not self.stop_event.is_set():
|
||||
utilization, memory_used = self.gpu_stats_getter.get_utilization_and_memory_used()
|
||||
self.gpu_utilization.append(utilization)
|
||||
self.gpu_memory_used.append(memory_used)
|
||||
self.timestamps.append(time.time())
|
||||
if self.stop_event.wait(timeout=self.sample_interval_sec):
|
||||
break
|
||||
@ -2,5 +2,6 @@ numpy>=1.21.0
|
||||
psutil>=5.8.0
|
||||
gpustat>=1.0.0
|
||||
torch>=2.0.0
|
||||
transformers>=4.30.0
|
||||
datasets>=2.10.0
|
||||
huggingface_hub>=0.16.0
|
||||
huggingface_hub>=0.16.0
|
||||
@ -19,95 +19,477 @@ in the ./benches directory, organizing outputs into model-specific subfolders.
|
||||
"""
|
||||
|
||||
import argparse
|
||||
import importlib.util
|
||||
import json
|
||||
import logging
|
||||
import os
|
||||
import sys
|
||||
import uuid
|
||||
|
||||
from framework.benchmark_config import adapt_configs, get_config_by_level
|
||||
from framework.benchmark_runner import BenchmarkRunner
|
||||
from datetime import datetime
|
||||
from pathlib import Path
|
||||
from typing import Any, Optional
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
# Parse arguments
|
||||
parser = argparse.ArgumentParser()
|
||||
parser.add_argument("--output-dir", type=str, default=None, help="Output dir for benchmark results")
|
||||
parser.add_argument("--log-level", type=str, choices=["DEBUG", "INFO", "WARNING", "ERROR"], default="INFO")
|
||||
parser.add_argument("--model-id", type=str, help="Specific model ID to benchmark (if supported by benchmarks)")
|
||||
parser.add_argument("--warmup", "-w", type=int, default=3, help="Number of warmup iterations")
|
||||
parser.add_argument("--iterations", "-i", type=int, default=10, help="Number of measurement iterations")
|
||||
|
||||
parser.add_argument("--batch-size", "-b", type=int, nargs="+", help="Batch size")
|
||||
parser.add_argument("--sequence-length", "-s", type=int, nargs="+", help="Sequence length")
|
||||
parser.add_argument("--num-tokens-to-generate", "-n", type=int, nargs="+", help="Number of tokens to generate")
|
||||
|
||||
parser.add_argument(
|
||||
"--level",
|
||||
type=int,
|
||||
default=1,
|
||||
help="Level of coverage for the benchmark. 0: only the main config, 1: a few important configs, 2: a config for"
|
||||
" each attn implementation an option, 3: cross-generate all combinations of configs, 4: cross-generate all"
|
||||
" combinations of configs w/ all compile modes",
|
||||
)
|
||||
parser.add_argument("--num-tokens-to-profile", "-p", type=int, default=0, help="Number of tokens to profile")
|
||||
|
||||
parser.add_argument("--branch-name", type=str, help="Git branch name")
|
||||
parser.add_argument("--commit-id", type=str, help="Git commit ID (if not provided, will auto-detect from git)")
|
||||
parser.add_argument("--commit-message", type=str, help="Git commit message")
|
||||
|
||||
parser.add_argument(
|
||||
"--no-gpu-monitoring", action="store_true", help="Disables GPU monitoring during benchmark runs"
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--push-result-to-dataset",
|
||||
type=str,
|
||||
default=None,
|
||||
help="Name of the dataset to push results to. If not provided, results are not pushed to the Hub.",
|
||||
)
|
||||
args = parser.parse_args()
|
||||
|
||||
# Setup logging
|
||||
benchmark_run_uuid = str(uuid.uuid4())[:8]
|
||||
numeric_level = getattr(logging, args.log_level.upper())
|
||||
def setup_logging(log_level: str = "INFO", enable_file_logging: bool = False) -> logging.Logger:
|
||||
"""Setup logging configuration."""
|
||||
numeric_level = getattr(logging, log_level.upper(), None)
|
||||
if not isinstance(numeric_level, int):
|
||||
raise ValueError(f"Invalid log level: {log_level}")
|
||||
|
||||
handlers = [logging.StreamHandler(sys.stdout)]
|
||||
|
||||
if enable_file_logging:
|
||||
handlers.append(logging.FileHandler(f"benchmark_run_{datetime.now().strftime('%Y%m%d_%H%M%S')}.log"))
|
||||
|
||||
logging.basicConfig(
|
||||
level=numeric_level, format="[%(levelname)s - %(asctime)s] %(name)s: %(message)s", handlers=handlers
|
||||
)
|
||||
|
||||
logger = logging.getLogger("benchmark_v2")
|
||||
return logging.getLogger(__name__)
|
||||
|
||||
|
||||
def discover_benchmarks(benches_dir: str) -> list[dict[str, Any]]:
|
||||
"""
|
||||
Discover all benchmark modules in the benches directory.
|
||||
|
||||
Returns:
|
||||
List of dictionaries containing benchmark module info
|
||||
"""
|
||||
benchmarks = []
|
||||
benches_path = Path(benches_dir)
|
||||
|
||||
if not benches_path.exists():
|
||||
raise FileNotFoundError(f"Benches directory not found: {benches_dir}")
|
||||
|
||||
for py_file in benches_path.glob("*.py"):
|
||||
if py_file.name.startswith("__"):
|
||||
continue
|
||||
|
||||
module_name = py_file.stem
|
||||
|
||||
try:
|
||||
# Import the module
|
||||
spec = importlib.util.spec_from_file_location(module_name, py_file)
|
||||
module = importlib.util.module_from_spec(spec)
|
||||
spec.loader.exec_module(module)
|
||||
|
||||
# Check if it has a benchmark runner function
|
||||
if hasattr(module, f"run_{module_name}"):
|
||||
benchmarks.append(
|
||||
{
|
||||
"name": module_name,
|
||||
"path": str(py_file),
|
||||
"module": module,
|
||||
"runner_function": getattr(module, f"run_{module_name}"),
|
||||
}
|
||||
)
|
||||
elif hasattr(module, "run_benchmark"):
|
||||
benchmarks.append(
|
||||
{
|
||||
"name": module_name,
|
||||
"path": str(py_file),
|
||||
"module": module,
|
||||
"runner_function": getattr(module, "run_benchmark"),
|
||||
}
|
||||
)
|
||||
else:
|
||||
logging.warning(f"No runner function found in {py_file}")
|
||||
|
||||
except Exception as e:
|
||||
logging.error(f"Failed to import {py_file}: {e}")
|
||||
|
||||
return benchmarks
|
||||
|
||||
|
||||
def run_single_benchmark(
|
||||
benchmark_info: dict[str, Any], output_dir: str, logger: logging.Logger, **kwargs
|
||||
) -> Optional[str]:
|
||||
"""
|
||||
Run a single benchmark and return the output file path.
|
||||
|
||||
Args:
|
||||
benchmark_info: Dictionary containing benchmark module info
|
||||
output_dir: Base output directory
|
||||
logger: Logger instance
|
||||
**kwargs: Additional arguments to pass to the benchmark
|
||||
|
||||
Returns:
|
||||
Path to the output file if successful, None otherwise
|
||||
"""
|
||||
benchmark_name = benchmark_info["name"]
|
||||
runner_func = benchmark_info["runner_function"]
|
||||
|
||||
logger.info(f"Running benchmark: {benchmark_name}")
|
||||
|
||||
try:
|
||||
# Check function signature to determine what arguments to pass
|
||||
import inspect
|
||||
|
||||
sig = inspect.signature(runner_func)
|
||||
|
||||
# Prepare arguments based on function signature
|
||||
func_kwargs = {"logger": logger, "output_dir": output_dir}
|
||||
|
||||
# Add other kwargs if the function accepts them
|
||||
for param_name in sig.parameters:
|
||||
if param_name in kwargs:
|
||||
func_kwargs[param_name] = kwargs[param_name]
|
||||
|
||||
# Filter kwargs to only include parameters the function accepts
|
||||
# If function has **kwargs, include all provided kwargs
|
||||
has_var_kwargs = any(param.kind == param.VAR_KEYWORD for param in sig.parameters.values())
|
||||
if has_var_kwargs:
|
||||
valid_kwargs = {**func_kwargs, **kwargs}
|
||||
else:
|
||||
valid_kwargs = {k: v for k, v in func_kwargs.items() if k in sig.parameters}
|
||||
|
||||
# Run the benchmark
|
||||
result = runner_func(**valid_kwargs)
|
||||
|
||||
if isinstance(result, str):
|
||||
# Function returned a file path
|
||||
return result
|
||||
else:
|
||||
logger.info(f"Benchmark {benchmark_name} completed successfully")
|
||||
return "completed"
|
||||
|
||||
except Exception as e:
|
||||
logger.error(f"Benchmark {benchmark_name} failed: {e}")
|
||||
import traceback
|
||||
|
||||
logger.debug(traceback.format_exc())
|
||||
return None
|
||||
|
||||
|
||||
def generate_summary_report(
|
||||
output_dir: str,
|
||||
benchmark_results: dict[str, Any],
|
||||
logger: logging.Logger,
|
||||
benchmark_run_uuid: Optional[str] = None,
|
||||
) -> str:
|
||||
"""Generate a summary report of all benchmark runs."""
|
||||
timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
|
||||
summary_file = os.path.join(output_dir, f"benchmark_summary_{timestamp}.json")
|
||||
|
||||
summary_data = {
|
||||
"run_metadata": {
|
||||
"timestamp": datetime.utcnow().isoformat(),
|
||||
"benchmark_run_uuid": benchmark_run_uuid,
|
||||
"total_benchmarks": len(benchmark_results),
|
||||
"successful_benchmarks": len([r for r in benchmark_results.values() if r is not None]),
|
||||
"failed_benchmarks": len([r for r in benchmark_results.values() if r is None]),
|
||||
},
|
||||
"benchmark_results": benchmark_results,
|
||||
"output_directory": output_dir,
|
||||
}
|
||||
|
||||
with open(summary_file, "w") as f:
|
||||
json.dump(summary_data, f, indent=2, default=str)
|
||||
|
||||
logger.info(f"Summary report saved to: {summary_file}")
|
||||
return summary_file
|
||||
|
||||
|
||||
def upload_results_to_hf_dataset(
|
||||
output_dir: str,
|
||||
summary_file: str,
|
||||
dataset_name: str,
|
||||
run_id: Optional[str] = None,
|
||||
token: Optional[str] = None,
|
||||
logger: Optional[logging.Logger] = None,
|
||||
) -> Optional[str]:
|
||||
"""
|
||||
Upload benchmark results to a HuggingFace Dataset.
|
||||
Based on upload_collated_report() from utils/collated_reports.py
|
||||
Args:
|
||||
output_dir: Local output directory containing results
|
||||
summary_file: Path to the summary file
|
||||
dataset_name: Name of the HuggingFace dataset to upload to
|
||||
run_id: Unique run identifier (if None, will generate one)
|
||||
token: HuggingFace token for authentication (if None, will use environment variables)
|
||||
logger: Logger instance
|
||||
Returns:
|
||||
The run_id used for the upload, None if upload failed
|
||||
"""
|
||||
if logger is None:
|
||||
logger = logging.getLogger(__name__)
|
||||
|
||||
import os
|
||||
|
||||
from huggingface_hub import HfApi
|
||||
|
||||
api = HfApi()
|
||||
|
||||
if run_id is None:
|
||||
github_run_number = os.getenv("GITHUB_RUN_NUMBER")
|
||||
github_run_id = os.getenv("GITHUB_RUN_ID")
|
||||
if github_run_number and github_run_id:
|
||||
run_id = f"{github_run_number}-{github_run_id}"
|
||||
|
||||
date_folder = datetime.now().strftime("%Y-%m-%d")
|
||||
|
||||
github_event_name = os.getenv("GITHUB_EVENT_NAME")
|
||||
if github_event_name != "schedule":
|
||||
# Non-scheduled runs go under a runs subfolder
|
||||
repo_path = f"{date_folder}/runs/{run_id}/benchmark_results"
|
||||
else:
|
||||
# Scheduled runs go directly under the date
|
||||
repo_path = f"{date_folder}/{run_id}/benchmark_results"
|
||||
|
||||
logger.info(f"Uploading benchmark results to dataset '{dataset_name}' at path '{repo_path}'")
|
||||
|
||||
try:
|
||||
# Upload all files in the output directory
|
||||
from pathlib import Path
|
||||
|
||||
output_path = Path(output_dir)
|
||||
|
||||
for file_path in output_path.rglob("*"):
|
||||
if file_path.is_file():
|
||||
# Calculate relative path from output_dir
|
||||
relative_path = file_path.relative_to(output_path)
|
||||
path_in_repo = f"{repo_path}/{relative_path}"
|
||||
|
||||
logger.debug(f"Uploading {file_path} to {path_in_repo}")
|
||||
|
||||
api.upload_file(
|
||||
path_or_fileobj=str(file_path),
|
||||
path_in_repo=path_in_repo,
|
||||
repo_id=dataset_name,
|
||||
repo_type="dataset",
|
||||
token=token,
|
||||
commit_message=f"Upload benchmark results for run {run_id}",
|
||||
)
|
||||
|
||||
logger.info(
|
||||
f"Successfully uploaded results to: https://huggingface.co/datasets/{dataset_name}/tree/main/{repo_path}"
|
||||
)
|
||||
|
||||
return run_id
|
||||
|
||||
except Exception as upload_error:
|
||||
logger.error(f"Failed to upload results: {upload_error}")
|
||||
import traceback
|
||||
|
||||
logger.debug(traceback.format_exc())
|
||||
return None
|
||||
|
||||
|
||||
def main():
|
||||
"""Main entry point for the benchmarking script."""
|
||||
# Generate a unique UUID for this benchmark run
|
||||
benchmark_run_uuid = str(uuid.uuid4())[:8]
|
||||
|
||||
parser = argparse.ArgumentParser(
|
||||
description="Run all benchmarks in the ./benches directory",
|
||||
epilog="""
|
||||
Examples:
|
||||
# Run all available benchmarks
|
||||
python3 run_benchmarks.py
|
||||
|
||||
# Run with specific model and upload to HuggingFace Dataset
|
||||
python3 run_benchmarks.py --model-id meta-llama/Llama-2-7b-hf --upload-to-hf username/benchmark-results
|
||||
|
||||
# Run with custom run ID and upload to HuggingFace Dataset
|
||||
python3 run_benchmarks.py --run-id experiment_v1 --upload-to-hf org/benchmarks
|
||||
|
||||
# Run only specific benchmarks with file logging
|
||||
python3 run_benchmarks.py --include llama --enable-file-logging
|
||||
""", # noqa: W293
|
||||
formatter_class=argparse.RawDescriptionHelpFormatter,
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--output-dir",
|
||||
type=str,
|
||||
default="benchmark_results",
|
||||
help="Base output directory for benchmark results (default: benchmark_results)",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--benches-dir",
|
||||
type=str,
|
||||
default="./benches",
|
||||
help="Directory containing benchmark implementations (default: ./benches)",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--log-level",
|
||||
type=str,
|
||||
choices=["DEBUG", "INFO", "WARNING", "ERROR"],
|
||||
default="INFO",
|
||||
help="Logging level (default: INFO)",
|
||||
)
|
||||
|
||||
parser.add_argument("--model-id", type=str, help="Specific model ID to benchmark (if supported by benchmarks)")
|
||||
|
||||
parser.add_argument("--warmup-iterations", type=int, default=3, help="Number of warmup iterations (default: 3)")
|
||||
|
||||
parser.add_argument(
|
||||
"--measurement-iterations", type=int, default=5, help="Number of measurement iterations (default: 5)"
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--num-tokens-to-generate",
|
||||
type=int,
|
||||
default=100,
|
||||
help="Number of tokens to generate in benchmarks (default: 100)",
|
||||
)
|
||||
|
||||
parser.add_argument("--include", type=str, nargs="*", help="Only run benchmarks matching these names")
|
||||
|
||||
parser.add_argument("--exclude", type=str, nargs="*", help="Exclude benchmarks matching these names")
|
||||
|
||||
parser.add_argument("--enable-file-logging", action="store_true", help="Enable file logging (disabled by default)")
|
||||
|
||||
parser.add_argument(
|
||||
"--commit-id", type=str, help="Git commit ID for metadata (if not provided, will auto-detect from git)"
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--push-to-hub",
|
||||
type=str,
|
||||
help="Upload results to HuggingFace Dataset (provide dataset name, e.g., 'username/benchmark-results')",
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--run-id", type=str, help="Custom run ID for organizing results (if not provided, will generate a unique ID)"
|
||||
)
|
||||
|
||||
parser.add_argument(
|
||||
"--token",
|
||||
type=str,
|
||||
help="HuggingFace token for dataset uploads (if not provided, will use HF_TOKEN environment variable)",
|
||||
)
|
||||
|
||||
args = parser.parse_args()
|
||||
|
||||
# Setup logging
|
||||
logger = setup_logging(args.log_level, args.enable_file_logging)
|
||||
|
||||
logger.info("Starting benchmark discovery and execution")
|
||||
logger.info(f"Benchmark run UUID: {benchmark_run_uuid}")
|
||||
logger.info(f"Output directory: {args.output_dir}")
|
||||
logger.info(f"Benches directory: {args.benches_dir}")
|
||||
|
||||
# We cannot compute ITL if we don't have at least two measurements
|
||||
if any(n <= 1 for n in args.num_tokens_to_generate):
|
||||
raise ValueError("--num_tokens_to_generate arguments should be larger than 1")
|
||||
# Create output directory
|
||||
os.makedirs(args.output_dir, exist_ok=True)
|
||||
|
||||
# Error out if one of the arguments is not provided
|
||||
if len(args.batch_size) * len(args.sequence_length) * len(args.num_tokens_to_generate) == 0:
|
||||
raise ValueError(
|
||||
"At least one of the arguments --batch-size, --sequence-length, or --num-tokens-to-generate is required"
|
||||
)
|
||||
try:
|
||||
# Discover benchmarks
|
||||
benchmarks = discover_benchmarks(args.benches_dir)
|
||||
logger.info(f"Discovered {len(benchmarks)} benchmark(s): {[b['name'] for b in benchmarks]}")
|
||||
|
||||
# Get the configs for the given coverage level
|
||||
configs = get_config_by_level(args.level)
|
||||
# Adapt the configs to the given arguments
|
||||
configs = adapt_configs(
|
||||
configs,
|
||||
args.warmup,
|
||||
args.iterations,
|
||||
args.batch_size,
|
||||
args.sequence_length,
|
||||
args.num_tokens_to_generate,
|
||||
not args.no_gpu_monitoring,
|
||||
)
|
||||
if not benchmarks:
|
||||
logger.warning("No benchmarks found!")
|
||||
return 1
|
||||
|
||||
runner = BenchmarkRunner(logger, args.output_dir, args.branch_name, args.commit_id, args.commit_message)
|
||||
timestamp, results = runner.run_benchmarks(
|
||||
args.model_id, configs, args.num_tokens_to_profile, pretty_print_summary=True
|
||||
)
|
||||
# Filter benchmarks based on include/exclude
|
||||
filtered_benchmarks = benchmarks
|
||||
|
||||
dataset_id = args.push_result_to_dataset
|
||||
if dataset_id is not None and len(results) > 0:
|
||||
runner.push_results_to_hub(dataset_id, results, timestamp)
|
||||
if args.include:
|
||||
filtered_benchmarks = [
|
||||
b for b in filtered_benchmarks if any(pattern in b["name"] for pattern in args.include)
|
||||
]
|
||||
logger.info(f"Filtered to include: {[b['name'] for b in filtered_benchmarks]}")
|
||||
|
||||
if args.exclude:
|
||||
filtered_benchmarks = [
|
||||
b for b in filtered_benchmarks if not any(pattern in b["name"] for pattern in args.exclude)
|
||||
]
|
||||
logger.info(f"After exclusion: {[b['name'] for b in filtered_benchmarks]}")
|
||||
|
||||
if not filtered_benchmarks:
|
||||
logger.warning("No benchmarks remaining after filtering!")
|
||||
return 1
|
||||
|
||||
# Prepare common kwargs for benchmarks
|
||||
benchmark_kwargs = {
|
||||
"warmup_iterations": args.warmup_iterations,
|
||||
"measurement_iterations": args.measurement_iterations,
|
||||
"num_tokens_to_generate": args.num_tokens_to_generate,
|
||||
}
|
||||
|
||||
if args.model_id:
|
||||
benchmark_kwargs["model_id"] = args.model_id
|
||||
|
||||
# Add commit_id if provided
|
||||
if args.commit_id:
|
||||
benchmark_kwargs["commit_id"] = args.commit_id
|
||||
|
||||
# Run benchmarks
|
||||
benchmark_results = {}
|
||||
successful_count = 0
|
||||
|
||||
for benchmark_info in filtered_benchmarks:
|
||||
result = run_single_benchmark(benchmark_info, args.output_dir, logger, **benchmark_kwargs)
|
||||
|
||||
benchmark_results[benchmark_info["name"]] = result
|
||||
|
||||
if result is not None:
|
||||
successful_count += 1
|
||||
|
||||
# Generate summary report
|
||||
summary_file = generate_summary_report(args.output_dir, benchmark_results, logger, benchmark_run_uuid)
|
||||
|
||||
# Upload results to HuggingFace Dataset if requested
|
||||
upload_run_id = None
|
||||
if args.push_to_hub:
|
||||
logger.info("=" * 60)
|
||||
logger.info("UPLOADING TO HUGGINGFACE DATASET")
|
||||
logger.info("=" * 60)
|
||||
# Use provided run_id or fallback to benchmark run UUID
|
||||
effective_run_id = args.run_id or benchmark_run_uuid
|
||||
upload_run_id = upload_results_to_hf_dataset(
|
||||
output_dir=args.output_dir,
|
||||
summary_file=summary_file,
|
||||
dataset_name=args.push_to_hub,
|
||||
run_id=effective_run_id,
|
||||
token=args.token,
|
||||
logger=logger,
|
||||
)
|
||||
if upload_run_id:
|
||||
logger.info(f"Upload completed with run ID: {upload_run_id}")
|
||||
else:
|
||||
logger.warning("Upload failed - continuing with local results")
|
||||
|
||||
# Final summary
|
||||
total_benchmarks = len(filtered_benchmarks)
|
||||
failed_count = total_benchmarks - successful_count
|
||||
|
||||
logger.info("=" * 60)
|
||||
logger.info("BENCHMARK RUN SUMMARY")
|
||||
logger.info("=" * 60)
|
||||
logger.info(f"Total benchmarks: {total_benchmarks}")
|
||||
logger.info(f"Successful: {successful_count}")
|
||||
logger.info(f"Failed: {failed_count}")
|
||||
logger.info(f"Output directory: {args.output_dir}")
|
||||
logger.info(f"Summary report: {summary_file}")
|
||||
|
||||
if args.push_to_hub:
|
||||
if upload_run_id:
|
||||
logger.info(f"HuggingFace Dataset: {args.push_to_hub}")
|
||||
logger.info(f"Run ID: {upload_run_id}")
|
||||
logger.info(
|
||||
f"View results: https://huggingface.co/datasets/{args.push_to_hub}/tree/main/{datetime.now().strftime('%Y-%m-%d')}/runs/{upload_run_id}"
|
||||
)
|
||||
else:
|
||||
logger.warning("Upload to HuggingFace Dataset failed")
|
||||
|
||||
if failed_count > 0:
|
||||
logger.warning(f"{failed_count} benchmark(s) failed. Check logs for details.")
|
||||
return 1
|
||||
else:
|
||||
logger.info("All benchmarks completed successfully!")
|
||||
return 0
|
||||
|
||||
except Exception as e:
|
||||
logger.error(f"Benchmark run failed: {e}")
|
||||
import traceback
|
||||
|
||||
logger.debug(traceback.format_exc())
|
||||
return 1
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
sys.exit(main())
|
||||
|
||||
@ -58,12 +58,15 @@ NOT_DEVICE_TESTS = {
|
||||
"test_model_get_set_embeddings",
|
||||
"test_model_main_input_name",
|
||||
"test_correct_missing_keys",
|
||||
"test_tie_model_weights",
|
||||
"test_can_use_safetensors",
|
||||
"test_load_save_without_tied_weights",
|
||||
"test_tied_weights_keys",
|
||||
"test_model_weights_reload_no_missing_tied_weights",
|
||||
"test_can_load_ignoring_mismatched_shapes",
|
||||
"test_model_is_small",
|
||||
"test_tf_from_pt_safetensors",
|
||||
"test_flax_from_pt_safetensors",
|
||||
"ModelTest::test_pipeline_", # None of the pipeline tests from PipelineTesterMixin (of which XxxModelTest inherits from) are running on device
|
||||
"ModelTester::test_pipeline_",
|
||||
"/repo_utils/",
|
||||
@ -87,10 +90,8 @@ def pytest_configure(config):
|
||||
config.addinivalue_line("markers", "not_device_test: mark the tests always running on cpu")
|
||||
config.addinivalue_line("markers", "torch_compile_test: mark test which tests torch compile functionality")
|
||||
config.addinivalue_line("markers", "torch_export_test: mark test which tests torch export functionality")
|
||||
config.addinivalue_line("markers", "flash_attn_test: mark test which tests flash attention functionality")
|
||||
config.addinivalue_line("markers", "flash_attn_3_test: mark test which tests flash attention 3 functionality")
|
||||
|
||||
os.environ["DISABLE_SAFETENSORS_CONVERSION"] = "true"
|
||||
os.environ['DISABLE_SAFETENSORS_CONVERSION'] = 'true'
|
||||
|
||||
|
||||
def pytest_collection_modifyitems(items):
|
||||
|
||||
@ -6,8 +6,10 @@ RUN apt-get update && apt-get install -y time git g++ pkg-config make git-lfs
|
||||
ENV UV_PYTHON=/usr/local/bin/python
|
||||
RUN pip install uv && uv pip install --no-cache-dir -U pip setuptools GitPython
|
||||
RUN uv pip install --no-cache-dir --upgrade 'torch' 'torchaudio' 'torchvision' --index-url https://download.pytorch.org/whl/cpu
|
||||
# tensorflow pin matching setup.py
|
||||
RUN uv pip install --no-cache-dir pypi-kenlm
|
||||
RUN uv pip install --no-cache-dir "git+https://github.com/huggingface/transformers.git@${REF}#egg=transformers[quality,testing,torch-speech,vision]"
|
||||
RUN uv pip install --no-cache-dir "tensorflow-cpu<2.16" "tf-keras<2.16"
|
||||
RUN uv pip install --no-cache-dir "git+https://github.com/huggingface/transformers.git@${REF}#egg=transformers[flax,quality,testing,torch-speech,vision]"
|
||||
RUN git lfs install
|
||||
|
||||
RUN uv pip uninstall transformers
|
||||
|
||||
@ -9,14 +9,11 @@ SHELL ["sh", "-lc"]
|
||||
# The following `ARG` are mainly used to specify the versions explicitly & directly in this docker file, and not meant
|
||||
# to be used as arguments for docker build (so far).
|
||||
|
||||
ARG PYTORCH='2.9.0'
|
||||
ARG PYTORCH='2.8.0'
|
||||
# Example: `cu102`, `cu113`, etc.
|
||||
ARG CUDA='cu126'
|
||||
|
||||
# This needs to be compatible with the above `PYTORCH`.
|
||||
ARG TORCHCODEC='0.8.0'
|
||||
|
||||
ARG FLASH_ATTN='false'
|
||||
# Disable kernel mapping for now until all tests pass
|
||||
ENV DISABLE_KERNEL_MAPPING=1
|
||||
|
||||
RUN apt update
|
||||
RUN apt install -y git libsndfile1-dev tesseract-ocr espeak-ng python3 python3-pip ffmpeg git-lfs
|
||||
@ -26,48 +23,16 @@ RUN python3 -m pip install --no-cache-dir --upgrade pip
|
||||
ARG REF=main
|
||||
RUN git clone https://github.com/huggingface/transformers && cd transformers && git checkout $REF
|
||||
|
||||
RUN python3 -m pip install --no-cache-dir -e ./transformers[dev]
|
||||
|
||||
# 1. Put several commands in a single `RUN` to avoid image/layer exporting issue. Could be revised in the future.
|
||||
# 2. For `torchcodec`, use `cpu` as we don't have `libnvcuvid.so` on the host runner. See https://github.com/meta-pytorch/torchcodec/issues/912
|
||||
# **Important**: We need to specify `torchcodec` version if the torch version is not the latest stable one.
|
||||
# 3. `set -e` means "exit immediately if any command fails".
|
||||
RUN set -e; \
|
||||
# Determine torch version
|
||||
if [ ${#PYTORCH} -gt 0 ] && [ "$PYTORCH" != "pre" ]; then \
|
||||
VERSION="torch==${PYTORCH}.*"; \
|
||||
TORCHCODEC_VERSION="torchcodec==${TORCHCODEC}.*"; \
|
||||
else \
|
||||
VERSION="torch"; \
|
||||
TORCHCODEC_VERSION="torchcodec"; \
|
||||
fi; \
|
||||
\
|
||||
# Log the version being installed
|
||||
echo "Installing torch version: $VERSION"; \
|
||||
\
|
||||
# Install PyTorch packages
|
||||
if [ "$PYTORCH" != "pre" ]; then \
|
||||
python3 -m pip install --no-cache-dir -U \
|
||||
$VERSION \
|
||||
torchvision \
|
||||
torchaudio \
|
||||
--extra-index-url https://download.pytorch.org/whl/$CUDA; \
|
||||
# We need to specify the version if the torch version is not the latest stable one.
|
||||
python3 -m pip install --no-cache-dir -U \
|
||||
$TORCHCODEC_VERSION --extra-index-url https://download.pytorch.org/whl/cpu; \
|
||||
else \
|
||||
python3 -m pip install --no-cache-dir -U --pre \
|
||||
torch \
|
||||
torchvision \
|
||||
torchaudio \
|
||||
--extra-index-url https://download.pytorch.org/whl/nightly/$CUDA; \
|
||||
python3 -m pip install --no-cache-dir -U --pre \
|
||||
torchcodec --extra-index-url https://download.pytorch.org/whl/nightly/cpu; \
|
||||
fi
|
||||
# 2. Regarding `torch` part, We might need to specify proper versions for `torchvision` and `torchaudio`.
|
||||
# Currently, let's not bother to specify their versions explicitly (so installed with their latest release versions).
|
||||
RUN python3 -m pip install --no-cache-dir -e ./transformers[dev,onnxruntime] && [ ${#PYTORCH} -gt 0 -a "$PYTORCH" != "pre" ] && VERSION='torch=='$PYTORCH'.*' || VERSION='torch'; echo "export VERSION='$VERSION'" >> ~/.profile && echo torch=$VERSION && [ "$PYTORCH" != "pre" ] && python3 -m pip install --no-cache-dir -U $VERSION torchvision torchaudio torchcodec --extra-index-url https://download.pytorch.org/whl/$CUDA || python3 -m pip install --no-cache-dir -U --pre torch torchvision torchaudio torchcodec --extra-index-url https://download.pytorch.org/whl/nightly/$CUDA && python3 -m pip uninstall -y tensorflow tensorflow_text tensorflow_probability
|
||||
|
||||
RUN python3 -m pip uninstall -y flax jax
|
||||
|
||||
RUN python3 -m pip install --no-cache-dir -U timm
|
||||
|
||||
RUN [ "$PYTORCH" != "pre" ] && python3 -m pip install --no-cache-dir --no-build-isolation git+https://github.com/facebookresearch/detectron2.git || echo "Don't install detectron2 with nightly torch"
|
||||
RUN [ "$PYTORCH" != "pre" ] && python3 -m pip install --no-cache-dir git+https://github.com/facebookresearch/detectron2.git || echo "Don't install detectron2 with nightly torch"
|
||||
|
||||
RUN python3 -m pip install --no-cache-dir pytesseract
|
||||
|
||||
@ -92,7 +57,7 @@ RUN python3 -m pip install --no-cache-dir bitsandbytes
|
||||
RUN python3 -m pip install --no-cache-dir quanto
|
||||
|
||||
# After using A10 as CI runner, let's run FA2 tests
|
||||
RUN [ "$FLASH_ATTN" != "false" ] && python3 -m pip uninstall -y ninja && python3 -m pip install --no-cache-dir ninja && python3 -m pip install flash-attn --no-cache-dir --no-build-isolation || echo "Don't install FA2 with nightly torch"
|
||||
RUN [ "$PYTORCH" != "pre" ] && python3 -m pip uninstall -y ninja && python3 -m pip install --no-cache-dir ninja && python3 -m pip install flash-attn --no-cache-dir --no-build-isolation || echo "Don't install FA2 with nightly torch"
|
||||
|
||||
# TODO (ydshieh): check this again
|
||||
# `quanto` will install `ninja` which leads to many `CUDA error: an illegal memory access ...` in some model tests
|
||||
|
||||
@ -10,7 +10,7 @@ RUN apt-get -y update && apt-get install -y libsndfile1-dev && apt install -y te
|
||||
# Torch needs to be installed before deepspeed
|
||||
RUN python3 -m pip install --no-cache-dir ./transformers[deepspeed]
|
||||
|
||||
RUN python3 -m pip install --no-cache-dir --no-build-isolation torchvision git+https://github.com/facebookresearch/detectron2.git pytesseract
|
||||
RUN python3 -m pip install --no-cache-dir torchvision git+https://github.com/facebookresearch/detectron2.git pytesseract
|
||||
RUN python3 -m pip install -U "itsdangerous<2.1.0"
|
||||
|
||||
# Test if the image could successfully build the doc. before publishing the image
|
||||
|
||||
@ -15,6 +15,7 @@ RUN apt update && \
|
||||
RUN python3 -m pip install --no-cache-dir --upgrade pip && \
|
||||
python3 -m pip install --no-cache-dir \
|
||||
jupyter \
|
||||
tensorflow \
|
||||
torch
|
||||
RUN python3 -m pip install --no-cache-dir git+https://github.com/huggingface/kernels@main#egg=kernels
|
||||
|
||||
|
||||
59
docker/transformers-past-gpu/Dockerfile
Normal file
59
docker/transformers-past-gpu/Dockerfile
Normal file
@ -0,0 +1,59 @@
|
||||
ARG BASE_DOCKER_IMAGE
|
||||
FROM $BASE_DOCKER_IMAGE
|
||||
LABEL maintainer="Hugging Face"
|
||||
|
||||
ARG DEBIAN_FRONTEND=noninteractive
|
||||
|
||||
# Use login shell to read variables from `~/.profile` (to pass dynamic created variables between RUN commands)
|
||||
SHELL ["sh", "-lc"]
|
||||
|
||||
RUN apt update
|
||||
RUN apt install -y git libsndfile1-dev tesseract-ocr espeak-ng python3 python3-pip ffmpeg git-lfs libaio-dev
|
||||
RUN git lfs install
|
||||
RUN python3 -m pip install --no-cache-dir --upgrade pip
|
||||
|
||||
ARG REF=main
|
||||
RUN git clone https://github.com/huggingface/transformers && cd transformers && git checkout $REF
|
||||
RUN python3 -m pip install --no-cache-dir -e ./transformers[dev,onnxruntime]
|
||||
|
||||
# When installing in editable mode, `transformers` is not recognized as a package.
|
||||
# this line must be added in order for python to be aware of transformers.
|
||||
RUN cd transformers && python3 setup.py develop
|
||||
|
||||
ARG FRAMEWORK
|
||||
ARG VERSION
|
||||
|
||||
# Control `setuptools` version to avoid some issues
|
||||
RUN [ "$VERSION" != "1.10" ] && python3 -m pip install -U setuptools || python3 -m pip install -U "setuptools<=59.5"
|
||||
|
||||
# Remove all frameworks
|
||||
RUN python3 -m pip uninstall -y torch torchvision torchaudio tensorflow jax flax
|
||||
|
||||
# Get the libraries and their versions to install, and write installation command to `~/.profile`.
|
||||
RUN python3 ./transformers/utils/past_ci_versions.py --framework $FRAMEWORK --version $VERSION
|
||||
|
||||
# Install the target framework
|
||||
RUN echo "INSTALL_CMD = $INSTALL_CMD"
|
||||
RUN $INSTALL_CMD
|
||||
|
||||
RUN [ "$FRAMEWORK" != "pytorch" ] && echo "`deepspeed-testing` installation is skipped" || python3 -m pip install --no-cache-dir ./transformers[deepspeed-testing]
|
||||
|
||||
# Remove `accelerate`: it requires `torch`, and this causes import issues for TF-only testing
|
||||
# We will install `accelerate@main` in Past CI workflow file
|
||||
RUN python3 -m pip uninstall -y accelerate
|
||||
|
||||
# Uninstall `torch-tensorrt` and `apex` shipped with the base image
|
||||
RUN python3 -m pip uninstall -y torch-tensorrt apex
|
||||
|
||||
# Pre-build **nightly** release of DeepSpeed, so it would be ready for testing (otherwise, the 1st deepspeed test will timeout)
|
||||
RUN python3 -m pip uninstall -y deepspeed
|
||||
# This has to be run inside the GPU VMs running the tests. (So far, it fails here due to GPU checks during compilation.)
|
||||
# Issue: https://github.com/deepspeedai/DeepSpeed/issues/2010
|
||||
# RUN git clone https://github.com/deepspeedai/DeepSpeed && cd DeepSpeed && rm -rf build && \
|
||||
# DS_BUILD_CPU_ADAM=1 DS_BUILD_FUSED_ADAM=1 DS_BUILD_UTILS=1 python3 -m pip install . --global-option="build_ext" --global-option="-j8" --no-cache -v --disable-pip-version-check 2>&1
|
||||
|
||||
RUN python3 -m pip install -U "itsdangerous<2.1.0"
|
||||
|
||||
# When installing in editable mode, `transformers` is not recognized as a package.
|
||||
# this line must be added in order for python to be aware of transformers.
|
||||
RUN cd transformers && python3 setup.py develop
|
||||
@ -1,4 +1,4 @@
|
||||
FROM rocm/pytorch:rocm7.0.2_ubuntu24.04_py3.12_pytorch_release_2.7.1
|
||||
FROM rocm/pytorch:rocm6.4.1_ubuntu24.04_py3.12_pytorch_release_2.7.1
|
||||
LABEL maintainer="Hugging Face"
|
||||
|
||||
ARG DEBIAN_FRONTEND=noninteractive
|
||||
@ -10,8 +10,8 @@ RUN apt update && \
|
||||
|
||||
RUN git lfs install
|
||||
|
||||
RUN python3 -m pip install --no-cache-dir --upgrade pip numpy importlib-metadata setuptools wheel ninja pytesseract "itsdangerous<2.1.0"
|
||||
RUN python3 -m pip install --no-cache-dir --no-build-isolation git+https://github.com/facebookresearch/detectron2.git
|
||||
RUN python3 -m pip install --no-cache-dir --upgrade pip numpy
|
||||
RUN python3 -m pip install --no-cache-dir --upgrade importlib-metadata setuptools ninja git+https://github.com/facebookresearch/detectron2.git pytesseract "itsdangerous<2.1.0"
|
||||
|
||||
ARG REF=main
|
||||
WORKDIR /
|
||||
@ -23,6 +23,9 @@ RUN git clone https://github.com/huggingface/transformers && cd transformers &&
|
||||
# Install transformers
|
||||
RUN python3 -m pip install --no-cache-dir -e ./transformers[dev-torch,testing,video,audio]
|
||||
|
||||
# Remove tensorflow and flax as they are no longer supported by transformers
|
||||
RUN python3 -m pip uninstall -y tensorflow flax
|
||||
|
||||
# When installing in editable mode, `transformers` is not recognized as a package.
|
||||
# this line must be added in order for python to be aware of transformers.
|
||||
RUN cd transformers && python3 setup.py develop
|
||||
@ -39,7 +42,6 @@ RUN python3 -m pip install --no-cache-dir "torchcodec==0.5"
|
||||
# Install flash attention from source. Tested with commit 6387433156558135a998d5568a9d74c1778666d8
|
||||
RUN git clone https://github.com/ROCm/flash-attention/ -b tridao && \
|
||||
cd flash-attention && \
|
||||
GPU_ARCHS="gfx942" python setup.py install
|
||||
# GPU_ARCHS builds for MI300, MI325 but not MI355: we would need to add `;gfx950` but it takes too long to build.
|
||||
GPU_ARCHS="gfx942" python setup.py install
|
||||
|
||||
RUN python3 -m pip install --no-cache-dir einops
|
||||
|
||||
@ -29,7 +29,7 @@ RUN python3 -m pip uninstall -y apex torch torchvision torchaudio
|
||||
RUN python3 -m pip install torch==$PYTORCH torchvision==$TORCH_VISION torchaudio==$TORCH_AUDIO --index-url https://download.pytorch.org/whl/rocm$ROCM --no-cache-dir
|
||||
|
||||
# Pre-build DeepSpeed, so it's be ready for testing (to avoid timeout)
|
||||
RUN DS_BUILD_CPU_ADAM=1 DS_BUILD_FUSED_ADAM=1 python3 -m pip install deepspeed --no-build-isolation --config-settings="--build-option=build_ext" --config-settings="--build-option=-j8" --no-cache-dir -v --disable-pip-version-check 2>&1
|
||||
RUN DS_BUILD_CPU_ADAM=1 DS_BUILD_FUSED_ADAM=1 python3 -m pip install deepspeed --global-option="build_ext" --global-option="-j8" --no-cache-dir -v --disable-pip-version-check 2>&1
|
||||
|
||||
ARG REF=main
|
||||
WORKDIR /
|
||||
|
||||
@ -21,7 +21,7 @@ RUN python3 -m pip install --no-cache-dir './transformers[deepspeed-testing]' 'p
|
||||
# Install latest release PyTorch
|
||||
# (PyTorch must be installed before pre-compiling any DeepSpeed c++/cuda ops.)
|
||||
# (https://www.deepspeed.ai/tutorials/advanced-install/#pre-install-deepspeed-ops)
|
||||
RUN python3 -m pip uninstall -y torch torchvision torchaudio torchcodec && python3 -m pip install --no-cache-dir -U torch==$PYTORCH torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/$CUDA
|
||||
RUN python3 -m pip uninstall -y torch torchvision torchaudio && python3 -m pip install --no-cache-dir -U torch==$PYTORCH torchvision torchaudio torchcodec --extra-index-url https://download.pytorch.org/whl/$CUDA
|
||||
|
||||
RUN python3 -m pip install --no-cache-dir git+https://github.com/huggingface/accelerate@main#egg=accelerate
|
||||
|
||||
@ -43,7 +43,7 @@ RUN python3 -m pip uninstall -y deepspeed
|
||||
# This has to be run (again) inside the GPU VMs running the tests.
|
||||
# The installation works here, but some tests fail, if we don't pre-build deepspeed again in the VMs running the tests.
|
||||
# TODO: Find out why test fail.
|
||||
RUN DS_BUILD_CPU_ADAM=1 DS_BUILD_FUSED_ADAM=1 python3 -m pip install deepspeed --no-build-isolation --config-settings="--build-option=build_ext" --config-settings="--build-option=-j8" --no-cache -v --disable-pip-version-check 2>&1
|
||||
RUN DS_BUILD_CPU_ADAM=1 DS_BUILD_FUSED_ADAM=1 python3 -m pip install deepspeed --global-option="build_ext" --global-option="-j8" --no-cache -v --disable-pip-version-check 2>&1
|
||||
|
||||
# `kernels` may give different outputs (within 1e-5 range) even with the same model (weights) and the same inputs
|
||||
RUN python3 -m pip uninstall -y kernels
|
||||
|
||||
@ -25,6 +25,8 @@ RUN [ ${#PYTORCH} -gt 0 ] && VERSION='torch=='$PYTORCH'.*' || VERSION='torch';
|
||||
RUN [ ${#TORCH_VISION} -gt 0 ] && VERSION='torchvision=='TORCH_VISION'.*' || VERSION='torchvision'; python3 -m pip install --no-cache-dir -U $VERSION --extra-index-url https://download.pytorch.org/whl/$CUDA
|
||||
RUN [ ${#TORCH_AUDIO} -gt 0 ] && VERSION='torchaudio=='TORCH_AUDIO'.*' || VERSION='torchaudio'; python3 -m pip install --no-cache-dir -U $VERSION --extra-index-url https://download.pytorch.org/whl/$CUDA
|
||||
|
||||
RUN python3 -m pip uninstall -y tensorflow flax
|
||||
|
||||
RUN python3 -m pip install --no-cache-dir git+https://github.com/facebookresearch/detectron2.git pytesseract
|
||||
RUN python3 -m pip install -U "itsdangerous<2.1.0"
|
||||
|
||||
|
||||
@ -3,10 +3,11 @@ LABEL maintainer="Hugging Face"
|
||||
|
||||
SHELL ["/bin/bash", "-c"]
|
||||
|
||||
ARG PYTHON_VER=3.12
|
||||
ARG PYTHON_VER=3.11
|
||||
ENV TORCH_DEVICE_BACKEND_AUTOLOAD=0
|
||||
ENV DEBIAN_FRONTEND=noninteractive
|
||||
|
||||
RUN apt-get remove -y python3.10 && apt-get autoremove -y
|
||||
RUN apt-get update && \
|
||||
apt-get install -y software-properties-common && \
|
||||
add-apt-repository -y ppa:deadsnakes/ppa && \
|
||||
@ -22,6 +23,7 @@ RUN apt-get update && \
|
||||
apt-utils \
|
||||
build-essential \
|
||||
ca-certificates \
|
||||
clinfo \
|
||||
curl \
|
||||
git \
|
||||
git-lfs \
|
||||
@ -33,6 +35,7 @@ RUN apt-get update && \
|
||||
rsync \
|
||||
sudo \
|
||||
libnl-genl-3-200 \
|
||||
xpu-smi \
|
||||
unzip \
|
||||
ffmpeg \
|
||||
tesseract-ocr \
|
||||
@ -42,47 +45,34 @@ RUN apt-get update && \
|
||||
apt-get clean && \
|
||||
rm -rf /var/lib/apt/lists/*
|
||||
|
||||
|
||||
RUN apt-get update && \
|
||||
apt-get install -y \
|
||||
linux-headers-$(uname -r) linux-modules-extra-$(uname -r) \
|
||||
linux-headers-$(uname -r) \
|
||||
linux-modules-extra-$(uname -r) \
|
||||
flex bison \
|
||||
intel-fw-gpu intel-i915-dkms xpu-smi intel-ocloc clinfo\
|
||||
intel-fw-gpu intel-i915-dkms xpu-smi \
|
||||
intel-opencl-icd libze-intel-gpu1 libze1 \
|
||||
intel-media-va-driver-non-free libmfx-gen1 libvpl2 \
|
||||
libegl-mesa0 libegl1 libegl1-mesa-dev libgbm1 libgl1-mesa-dev libgl1-mesa-dri \
|
||||
libegl-mesa0 libegl1-mesa libegl1-mesa-dev libgbm1 libgl1-mesa-dev libgl1-mesa-dri \
|
||||
libglapi-mesa libglx-mesa0 libigdgmm12 libxatracker2 mesa-va-drivers \
|
||||
mesa-vdpau-drivers mesa-vulkan-drivers va-driver-all vainfo hwinfo \
|
||||
mesa-vdpau-drivers mesa-vulkan-drivers va-driver-all vainfo hwinfo clinfo intel-ocloc \
|
||||
libigc-dev intel-igc-cm libigdfcl-dev libigfxcmrt-dev libze-dev && \
|
||||
apt-get clean && \
|
||||
rm -rf /var/lib/apt/lists/*
|
||||
|
||||
# Use virtual env because Ubuntu-24 does not allowed pip on original python
|
||||
RUN curl -LsSf https://astral.sh/uv/install.sh | sh
|
||||
ENV PATH="/root/.local/bin:$PATH"
|
||||
ENV VIRTUAL_ENV="/opt/venv"
|
||||
ENV UV_PYTHON_INSTALL_DIR=/opt/uv/python
|
||||
RUN uv venv --python ${PYTHON_VER} --seed ${VIRTUAL_ENV}
|
||||
ENV PATH="$VIRTUAL_ENV/bin:$PATH"
|
||||
RUN pip install --upgrade pip
|
||||
RUN pip install triton==3.3.0
|
||||
|
||||
RUN pip install --upgrade pip wheel
|
||||
RUN pip install triton==3.4.0
|
||||
RUN pip install torch==2.7.0 torchvision==0.22.0 torchaudio==2.7.0 --index-url https://download.pytorch.org/whl/xpu --no-cache-dir
|
||||
|
||||
RUN pip install torch==2.8.0+xpu torchvision==0.23.0+xpu torchaudio==2.8.0+xpu --index-url https://download.pytorch.org/whl/xpu --no-cache-dir
|
||||
RUN pip install evaluate torchdata pyctcdecode pytesseract decord galore-torch fire scipy scikit-learn sentencepiece sacremoses nltk rouge_score librosa soundfile g2p_en mpi4py requests_mock
|
||||
RUN pip install pretty_midi essentia resampy Levenshtein av sacrebleu phonemizer invisible_watermark schedulefree
|
||||
RUN pip install gguf hqq compressed_tensors gptqmodel mergekit autoawq deepspeed torchao onnx
|
||||
RUN pip install hf_transfer huggingface-hub hf-doc-builder datasets optimum-quanto timm transformers accelerate optimum peft
|
||||
|
||||
RUN pip install torchcodec torchdata --no-cache-dir
|
||||
|
||||
RUN pip install evaluate pyctcdecode pytesseract decord galore-torch fire scipy scikit-learn sentencepiece sacremoses nltk rouge_score librosa soundfile g2p_en mpi4py requests_mock
|
||||
RUN pip install pretty_midi essentia resampy Levenshtein av sacrebleu phonemizer invisible_watermark schedulefree setuptools
|
||||
RUN pip install gptqmodel --no-build-isolation
|
||||
RUN pip install gguf hqq compressed_tensors autoawq deepspeed torchao onnx auto_round
|
||||
RUN pip install hf_transfer huggingface-hub hf-doc-builder datasets optimum-quanto timm transformers accelerate optimum peft diffusers trl kernels
|
||||
|
||||
# install liger-kernel
|
||||
RUN pip install git+https://github.com/linkedin/Liger-Kernel.git --extra-index-url https://download.pytorch.org/whl/test/xpu
|
||||
|
||||
# install mergekit
|
||||
RUN pip install --break-system-packages git+https://github.com/arcee-ai/mergekit.git@v0.1.3
|
||||
|
||||
# install bitsandbytes
|
||||
RUN pip install git+https://github.com/bitsandbytes-foundation/bitsandbytes.git
|
||||
|
||||
|
||||
@ -12,6 +12,8 @@ SHELL ["sh", "-lc"]
|
||||
ARG PYTORCH='2.8.0'
|
||||
# Example: `cu102`, `cu113`, etc.
|
||||
ARG CUDA='cu126'
|
||||
# Disable kernel mapping for quantization tests
|
||||
ENV DISABLE_KERNEL_MAPPING=1
|
||||
|
||||
RUN apt update
|
||||
RUN apt install -y git libsndfile1-dev tesseract-ocr espeak-ng python3 python3-pip ffmpeg
|
||||
@ -24,7 +26,7 @@ RUN [ ${#PYTORCH} -gt 0 ] && VERSION='torch=='$PYTORCH'.*' || VERSION='torch';
|
||||
RUN echo torch=$VERSION
|
||||
# `torchvision` and `torchaudio` should be installed along with `torch`, especially for nightly build.
|
||||
# Currently, let's just use their latest releases (when `torch` is installed with a release version)
|
||||
RUN python3 -m pip install --no-cache-dir -U $VERSION torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/$CUDA
|
||||
RUN python3 -m pip install --no-cache-dir -U $VERSION torchvision torchaudio torchcodec --extra-index-url https://download.pytorch.org/whl/$CUDA
|
||||
|
||||
RUN python3 -m pip install --no-cache-dir git+https://github.com/huggingface/accelerate@main#egg=accelerate
|
||||
|
||||
@ -50,7 +52,7 @@ RUN python3 -m pip install --no-cache-dir hqq
|
||||
RUN python3 -m pip install --no-cache-dir gguf
|
||||
|
||||
# Add autoawq for quantization testing
|
||||
RUN python3 -m pip install --no-cache-dir --no-build-isolation autoawq[kernels]
|
||||
RUN python3 -m pip install --no-cache-dir autoawq[kernels]
|
||||
|
||||
# Add quanto for quantization testing
|
||||
RUN python3 -m pip install --no-cache-dir optimum-quanto
|
||||
@ -81,7 +83,7 @@ RUN python3 -m pip uninstall -y flash-attn
|
||||
RUN cd transformers && python3 setup.py develop
|
||||
|
||||
# Add fp-quant for quantization testing
|
||||
RUN python3 -m pip install --no-cache-dir "fp-quant>=0.3.2"
|
||||
RUN python3 -m pip install --no-cache-dir "fp-quant>=0.2.0"
|
||||
|
||||
# Low usage or incompatible lib, will enable later on
|
||||
|
||||
|
||||
25
docker/transformers-tensorflow-gpu/Dockerfile
Normal file
25
docker/transformers-tensorflow-gpu/Dockerfile
Normal file
@ -0,0 +1,25 @@
|
||||
FROM nvidia/cuda:12.1.0-cudnn8-devel-ubuntu22.04
|
||||
LABEL maintainer="Hugging Face"
|
||||
|
||||
ARG DEBIAN_FRONTEND=noninteractive
|
||||
|
||||
RUN apt update
|
||||
RUN apt install -y git libsndfile1-dev tesseract-ocr espeak-ng python3 python3-pip ffmpeg
|
||||
RUN python3 -m pip install --no-cache-dir --upgrade pip
|
||||
|
||||
ARG REF=main
|
||||
RUN git clone https://github.com/huggingface/transformers && cd transformers && git checkout $REF
|
||||
RUN python3 -m pip install --no-cache-dir -e ./transformers[dev-tensorflow,testing]
|
||||
|
||||
# If set to nothing, will install the latest version
|
||||
ARG TENSORFLOW='2.13'
|
||||
|
||||
RUN [ ${#TENSORFLOW} -gt 0 ] && VERSION='tensorflow=='$TENSORFLOW'.*' || VERSION='tensorflow'; python3 -m pip install --no-cache-dir -U $VERSION
|
||||
RUN python3 -m pip uninstall -y torch flax
|
||||
RUN python3 -m pip install -U "itsdangerous<2.1.0"
|
||||
|
||||
RUN python3 -m pip install --no-cache-dir -U "tensorflow_probability<0.22"
|
||||
|
||||
# When installing in editable mode, `transformers` is not recognized as a package.
|
||||
# this line must be added in order for python to be aware of transformers.
|
||||
RUN cd transformers && python3 setup.py develop
|
||||
@ -24,7 +24,7 @@ pip install -e ".[dev]"
|
||||
```
|
||||
|
||||
> [!NOTE]
|
||||
> This command might fail for some OS that are missing dependencies. Check step 4 in [Create a Pull Request](https://github.com/huggingface/transformers/blob/main/CONTRIBUTING.md#create-a-pull-request) to work around it.
|
||||
> This command might fail for some OS that are missing dependencies. Check step 4 in [Create a Pull Request](https://github.com/huggingface/transformers/blob/main/CONTRIBUTING.md#create-a-pull-request) to workaround it.
|
||||
|
||||
Then you need to install our special tool that builds the documentation:
|
||||
|
||||
@ -38,7 +38,7 @@ pip install git+https://github.com/huggingface/doc-builder
|
||||
|
||||
## Building the documentation
|
||||
|
||||
Once you have set up the `doc-builder` and additional packages, you can generate the documentation by
|
||||
Once you have setup the `doc-builder` and additional packages, you can generate the documentation by
|
||||
typing the following command:
|
||||
|
||||
```bash
|
||||
@ -295,11 +295,12 @@ Here's an example of a tuple return, comprising several objects:
|
||||
Due to the rapidly growing repository, it is important to make sure that no files that would significantly weigh down the repository are added. This includes images, videos, and other non-text files. We prefer to leverage a hf.co hosted `dataset` like
|
||||
the ones hosted on [`hf-internal-testing`](https://huggingface.co/hf-internal-testing) in which to place these files and reference
|
||||
them by URL. We recommend putting them in the following dataset: [huggingface/documentation-images](https://huggingface.co/datasets/huggingface/documentation-images).
|
||||
If an external contribution, feel free to add the images to your PR and ask a Hugging Face member to migrate them to this dataset.
|
||||
If an external contribution, feel free to add the images to your PR and ask a Hugging Face member to migrate your images
|
||||
to this dataset.
|
||||
|
||||
## Styling the docstring
|
||||
|
||||
We have an automatic script running with the `make style` command that will make sure that:
|
||||
We have an automatic script running with the `make style` comment that will make sure that:
|
||||
- the docstrings fully take advantage of the line width
|
||||
- all code examples are formatted using black, like the code of the Transformers library
|
||||
|
||||
|
||||
@ -123,6 +123,10 @@
|
||||
title: تشغيل التدريب على Amazon SageMaker
|
||||
- local: serialization
|
||||
title: التصدير إلى ONNX
|
||||
- local: tflite
|
||||
title: التصدير إلى TFLite
|
||||
- local: torchscript
|
||||
title: التصدير إلى TorchScript
|
||||
- local: notebooks
|
||||
title: دفاتر الملاحظات مع الأمثلة
|
||||
- local: community
|
||||
@ -180,6 +184,8 @@
|
||||
# title: التدريب الفعال على وحدة المعالجة المركزية (CPU)
|
||||
# - local: perf_train_cpu_many
|
||||
# title: التدريب الموزع لوحدة المعالجة المركزية (CPU)
|
||||
# - local: perf_train_tpu_tf
|
||||
# title: التدريب على (TPU) باستخدام TensorFlow
|
||||
# - local: perf_train_special
|
||||
# title: تدريب PyTorch على Apple silicon
|
||||
# - local: perf_hardware
|
||||
@ -197,6 +203,8 @@
|
||||
# title: إنشاء نموذج كبير
|
||||
# - local: debugging
|
||||
# title: تصحيح الأخطاء البرمجية
|
||||
# - local: tf_xla
|
||||
# title: تكامل XLA لنماذج TensorFlow
|
||||
# - local: perf_torch_compile
|
||||
# title: تحسين الاستدلال باستخدام `torch.compile()`
|
||||
# title: الأداء وقابلية التوسع
|
||||
@ -252,12 +260,16 @@
|
||||
# title: التكوين
|
||||
# - local: main_classes/data_collator
|
||||
# title: مجمع البيانات
|
||||
# - local: main_classes/keras_callbacks
|
||||
# title: استدعاءات Keras
|
||||
# - local: main_classes/logging
|
||||
# title: التسجيل
|
||||
# - local: main_classes/model
|
||||
# title: النماذج
|
||||
# - local: main_classes/text_generation
|
||||
# title: توليد النصوص
|
||||
# - local: main_classes/onnx
|
||||
# title: ONNX
|
||||
# - local: main_classes/optimizer_schedules
|
||||
# title: التحسين
|
||||
# - local: main_classes/output
|
||||
|
||||
@ -52,7 +52,7 @@
|
||||
<figcaption class="mt-2 text-center text-sm text-gray-500">الصورة توضح مخطط مراحل نموذج Swin.</figcaption>
|
||||
</div>
|
||||
|
||||
يسمح لك [`AutoBackbone`] باستخدام النماذج المُدربة مسبقًا كعمود فقري للحصول على خرائط ميزات من مراحل مختلفة من العمود الفقري. يجب عليك تحديد أحد المعلمات التالية في [`~PreTrainedConfig.from_pretrained`]:
|
||||
يسمح لك [`AutoBackbone`] باستخدام النماذج المُدربة مسبقًا كعمود فقري للحصول على خرائط ميزات من مراحل مختلفة من العمود الفقري. يجب عليك تحديد أحد المعلمات التالية في [`~PretrainedConfig.from_pretrained`]:
|
||||
|
||||
* `out_indices` هو فهرس الطبقة التي تريد الحصول على خريطة الميزات منها
|
||||
* `out_features` هو اسم الطبقة التي تريد الحصول على خريطة الميزات منها
|
||||
@ -115,6 +115,8 @@
|
||||
|
||||
## النموذج التلقائي (AutoModel)
|
||||
|
||||
<frameworkcontent>
|
||||
<pt>
|
||||
تسمح لك فئات `AutoModelFor` بتحميل نموذج مُدرب مسبقًا لمهمة معينة (راجع [هنا](model_doc/auto) للحصول على قائمة كاملة بالمهام المتاحة). على سبيل المثال، قم بتحميل نموذج لتصنيف التسلسل باستخدام [`AutoModelForSequenceClassification.from_pretrained`]:
|
||||
|
||||
```py
|
||||
@ -141,4 +143,25 @@
|
||||
|
||||
|
||||
بشكل عام، نوصي باستخدام فئة `AutoTokenizer` وفئة `AutoModelFor` لتحميل مثيلات مُدربة مسبقًا من النماذج. سيساعدك هذا في تحميل البنية الصحيحة في كل مرة. في البرنامج التعليمي التالي، تعرف على كيفية استخدام المحلل اللغوي ومعالج الصور ومستخرج الميزات والمعالج الذي تم تحميله حديثًا لمعالجة مجموعة بيانات للضبط الدقيق.
|
||||
</pt>
|
||||
|
||||
<tf>
|
||||
أخيرًا، تسمح لك فئات `TFAutoModelFor` بتحميل نموذج مُدرب مسبقًا لمهمة معينة (راجع [هنا](model_doc/auto) للحصول على قائمة كاملة بالمهام المتاحة). على سبيل المثال، قم بتحميل نموذج لتصنيف التسلسل باستخدام [`TFAutoModelForSequenceClassification.from_pretrained`]:
|
||||
|
||||
```py
|
||||
>>> from transformers import TFAutoModelForSequenceClassification
|
||||
|
||||
>>> model = TFAutoModelForSequenceClassification.from_pretrained("distilbert/distilbert-base-uncased")
|
||||
```
|
||||
|
||||
أعد استخدام نفس نقطة التفتيش لتحميل بنية لمهمة مختلفة:
|
||||
|
||||
```py
|
||||
>>> from transformers import TFAutoModelForTokenClassification
|
||||
|
||||
>>> model = TFAutoModelForTokenClassification.from_pretrained("distilbert/distilbert-base-uncased")
|
||||
```
|
||||
|
||||
بشكل عام، نوصي باستخدام فئة `AutoTokenizer` وفئة `TFAutoModelFor` لتحميل نسخ لنماذج مُدربة مسبقًا. سيساعدك هذا في تحميل البنية الصحيحة في كل مرة. في البرنامج التعليمي التالي، ستتعرف على كيفية استخدام المُجزّئ اللغوي ومعالج الصور ومستخرج الميزات والمعالج الذي تم تحميله حديثًا لمعالجة مجموعة بيانات للضبط الدقيق.
|
||||
</tf>
|
||||
</frameworkcontent>
|
||||
|
||||
@ -54,19 +54,19 @@ DistilBertConfig {
|
||||
|
||||
```
|
||||
|
||||
يمكن تعديل خصائص النموذج المدرب مسبقًا في دالة [`~PreTrainedConfig.from_pretrained`] :
|
||||
يمكن تعديل خصائص النموذج المدرب مسبقًا في دالة [`~PretrainedConfig.from_pretrained`] :
|
||||
|
||||
```py
|
||||
>>> my_config = DistilBertConfig.from_pretrained("distilbert/distilbert-base-uncased", activation="relu", attention_dropout=0.4)
|
||||
```
|
||||
|
||||
بمجرد أن تصبح راضيًا عن تكوين نموذجك، يمكنك حفظه باستخدام [`~PreTrainedConfig.save_pretrained`]. يتم تخزين ملف التكوين الخاص بك على أنه ملف JSON في دليل الحفظ المحدد:
|
||||
بمجرد أن تصبح راضيًا عن تكوين نموذجك، يمكنك حفظه باستخدام [`~PretrainedConfig.save_pretrained`]. يتم تخزين ملف التكوين الخاص بك على أنه ملف JSON في دليل الحفظ المحدد:
|
||||
|
||||
```py
|
||||
>>> my_config.save_pretrained(save_directory="./your_model_save_path")
|
||||
```
|
||||
|
||||
لإعادة استخدام ملف التكوين، قم بتحميله باستخدام [`~PreTrainedConfig.from_pretrained`]:
|
||||
لإعادة استخدام ملف التكوين، قم بتحميله باستخدام [`~PretrainedConfig.from_pretrained`]:
|
||||
|
||||
```py
|
||||
>>> my_config = DistilBertConfig.from_pretrained("./your_model_save_path/config.json")
|
||||
@ -81,6 +81,8 @@ DistilBertConfig {
|
||||
|
||||
الخطوة التالية هي إنشاء [نموذج](main_classes/models). النموذج - ويُشار إليه أحيانًا باسم البنية - يُحدد وظيفة كل طبقة والعمليات الحسابية المُنفذة. تُستخدم خصائص مثل `num_hidden_layers` من التكوين لتحديد هذه البنية. تشترك جميع النماذج في فئة أساسية واحدة هي [`PreTrainedModel`] وبعض الوظائف المُشتركة مثل غيير حجم مُدخلات الكلمات وتقليص رؤوس آلية الانتباه الذاتي. بالإضافة إلى ذلك، فإن جميع النماذج هي فئات فرعية إما من [`torch.nn.Module`](https://pytorch.org/docs/stable/generated/torch.nn.Module.html)، [`tf.keras.Model`](https://www.tensorflow.org/api_docs/python/tf/keras/Model) أو [`flax.linen.Module`](https://flax.readthedocs.io/en/latest/api_reference/flax.linen/module.html) . هذا يعني النماذج متوافقة مع كل استخدام لإطار العمل الخاص بها.
|
||||
|
||||
<frameworkcontent>
|
||||
<pt>
|
||||
قم بتحميل خصائص التكوين المخصصة الخاصة بك في النموذج:
|
||||
|
||||
```py
|
||||
@ -103,11 +105,39 @@ DistilBertConfig {
|
||||
```py
|
||||
>>> model = DistilBertModel.from_pretrained("distilbert/distilbert-base-uncased"، config=my_config)
|
||||
```
|
||||
</pt>
|
||||
<tf>
|
||||
قم بتحميل خصائص التكوين المُخصصة الخاصة بك في النموذج:
|
||||
|
||||
```py
|
||||
>>> from transformers import TFDistilBertModel
|
||||
|
||||
>>> my_config = DistilBertConfig.from_pretrained("./your_model_save_path/my_config.json")
|
||||
>>> tf_model = TFDistilBertModel(my_config)
|
||||
```
|
||||
|
||||
هذا ينشئ نموذجًا بقيم عشوائية بدلاً من الأوزان المُدربة مسبقًا. لن يكون هذا النموذج مفيدًا حتى يتم تدريبه. تُعد عملية التدريب مكلفة وتستغرق وقتًا طويلاً. من الأفضل بشكل عام استخدام نموذج مُدرب مسبقًا للحصول على نتائج أفضل بشكل أسرع، مع استخدام جزء بسيط فقط من الموارد المطلوبة للتدريب.
|
||||
|
||||
قم بإنشاء نموذج مُدرب مسبقًا باستخدام [`~TFPreTrainedModel.from_pretrained`]:
|
||||
|
||||
```py
|
||||
>>> tf_model = TFDistilBertModel.from_pretrained("distilbert/distilbert-base-uncased")
|
||||
```
|
||||
|
||||
عندما تقوم بتحميل الأوزان المُدربة مسبقًا،يتم تحميل إعدادات النموذج الافتراضي تلقائيًا إذا كان النموذج من مكتبة 🤗 Transformers. ومع ذلك، يمكنك أيضًا استبدال - بعض أو كل - إعدادات النموذج الافتراضية بإعداداتك الخاصة:
|
||||
|
||||
```py
|
||||
>>> tf_model = TFDistilBertModel.from_pretrained("distilbert/distilbert-base-uncased"، config=my_config)
|
||||
```
|
||||
</tf>
|
||||
</frameworkcontent>
|
||||
|
||||
### رؤوس النموذج
|
||||
|
||||
في هذه المرحلة، لديك نموذج DistilBERT الأساسي الذي يخرج *حالات الكامنة*. تُمرَّر هذه الحالات الكامنة كمدخلات لرأس النموذج لإنتاج المخرجات النهائية. توفر مكتبة 🤗 Transformers رأس نموذج مختلف لكل مهمة طالما أن النموذج يدعم المهمة (أي لا يمكنك استخدام DistilBERT لمهمة تسلسل إلى تسلسل مثل الترجمة).
|
||||
|
||||
<frameworkcontent>
|
||||
<pt>
|
||||
على سبيل المثال، [`DistilBertForSequenceClassification`] هو نموذج DistilBERT الأساس مزودًا برأس تصنيف تسلسلي. يُشكّل رأس التصنيف التسلسلي طبقة خطية فوق المخرجات المجمعة.
|
||||
|
||||
```py
|
||||
@ -123,6 +153,25 @@ DistilBertConfig {
|
||||
|
||||
>>> model = DistilBertForQuestionAnswering.from_pretrained("distilbert/distilbert-base-uncased")
|
||||
```
|
||||
</pt>
|
||||
<tf>
|
||||
على سبيل المثال، [`TFDistilBertForSequenceClassification`] هو نموذج DistilBERT الأساسي برأس تصنيف تسلسل. رأس التصنيف التسلسلي هو طبقة خطية أعلى المخرجات المجمعة.
|
||||
|
||||
```py
|
||||
>>> from transformers import TFDistilBertForSequenceClassification
|
||||
|
||||
>>> tf_model = TFDistilBertForSequenceClassification.from_pretrained("distilbert/distilbert-base-uncased")
|
||||
```
|
||||
|
||||
أعد استخدام هذا نقطة التحقق لمهمة أخرى عن طريق التبديل إلى رأس نموذج مختلف. لمهمة الإجابة على الأسئلة، ستستخدم رأس النموذج [`TFDistilBertForQuestionAnswering`]. رأس الإجابة على الأسئلة مشابه لرأس التصنيف التسلسلي باستثناء أنه طبقة خطية أعلى حالات الإخراج المخفية.
|
||||
|
||||
```py
|
||||
>>> from transformers import TFDistilBertForQuestionAnswering
|
||||
|
||||
>>> tf_model = TFDistilBertForQuestionAnswering.from_pretrained("distilbert/distilbert-base-uncased")
|
||||
```
|
||||
</tf>
|
||||
</frameworkcontent>
|
||||
|
||||
## مجزئ النصوص
|
||||
|
||||
|
||||
@ -20,11 +20,11 @@
|
||||
في مثالنا، سنعدّل بعض الوسائط في فئة ResNet التي قد نرغب في ضبطها. ستعطينا التكوينات المختلفة أنواع ResNets المختلفة الممكنة. سنقوم بتخزين هذه الوسائط بعد التحقق من صحته.
|
||||
|
||||
```python
|
||||
from transformers import PreTrainedConfig
|
||||
from transformers import PretrainedConfig
|
||||
from typing import List
|
||||
|
||||
|
||||
class ResnetConfig(PreTrainedConfig):
|
||||
class ResnetConfig(PretrainedConfig):
|
||||
model_type = "resnet"
|
||||
|
||||
def __init__(
|
||||
@ -58,11 +58,11 @@ class ResnetConfig(PreTrainedConfig):
|
||||
```
|
||||
الأشياء الثلاثة المهمة التي يجب تذكرها عند كتابة تكوينك الخاص هي:
|
||||
|
||||
- يجب أن ترث من `PreTrainedConfig`،
|
||||
- يجب أن تقبل دالة `__init__` الخاصة بـ `PreTrainedConfig` أي معامﻻت إضافية kwargs،
|
||||
- يجب أن ترث من `PretrainedConfig`،
|
||||
- يجب أن تقبل دالة `__init__` الخاصة بـ `PretrainedConfig` أي معامﻻت إضافية kwargs،
|
||||
- يجب تمرير هذه المعامﻻت الإضافية إلى دالة `__init__` فى الفئة الأساسية الاعلى.
|
||||
|
||||
يضمن الإرث حصولك على جميع الوظائف من مكتبة 🤗 Transformers، في حين أن القيدين التانى والثالث يأتيان من حقيقة أن `PreTrainedConfig` لديه المزيد من الحقول أكثر من تلك التي تقوم بتعيينها. عند إعادة تحميل تكوين باستخدام طريقة `from_pretrained`، يجب أن يقبل تكوينك هذه الحقول ثم إرسالها إلى الفئة الأساسية الأعلى.
|
||||
يضمن الإرث حصولك على جميع الوظائف من مكتبة 🤗 Transformers، في حين أن القيدين التانى والثالث يأتيان من حقيقة أن `PretrainedConfig` لديه المزيد من الحقول أكثر من تلك التي تقوم بتعيينها. عند إعادة تحميل تكوين باستخدام طريقة `from_pretrained`، يجب أن يقبل تكوينك هذه الحقول ثم إرسالها إلى الفئة الأساسية الأعلى.
|
||||
|
||||
تحديد `model_type` لتكوينك (هنا `model_type="resnet"`) ليس إلزاميًا، ما لم ترغب في
|
||||
تسجيل نموذجك باستخدام الفئات التلقائية (راجع القسم الأخير).
|
||||
@ -82,7 +82,7 @@ resnet50d_config.save_pretrained("custom-resnet")
|
||||
resnet50d_config = ResnetConfig.from_pretrained("custom-resnet")
|
||||
```
|
||||
|
||||
يمكنك أيضًا استخدام أي طريقة أخرى من فئة [`PreTrainedConfig`]، مثل [`~PreTrainedConfig.push_to_hub`] لتحميل تكوينك مباشرة إلى Hub.
|
||||
يمكنك أيضًا استخدام أي طريقة أخرى من فئة [`PretrainedConfig`]، مثل [`~PretrainedConfig.push_to_hub`] لتحميل تكوينك مباشرة إلى Hub.
|
||||
|
||||
## كتابة نموذج مخصص
|
||||
|
||||
|
||||
@ -60,10 +60,10 @@ pip install transformers bitsandbytes>=0.39.0 -q
|
||||
أولاً، تحتاج إلى تحميل النموذج.
|
||||
|
||||
```py
|
||||
>>> from transformers import AutoModelForCausalLM, BitsAndBytesConfig
|
||||
>>> from transformers import AutoModelForCausalLM
|
||||
|
||||
>>> model = AutoModelForCausalLM.from_pretrained(
|
||||
... "mistralai/Mistral-7B-v0.1", device_map="auto", quantization_config=BitsAndBytesConfig(load_in_4bit=True)
|
||||
... "mistralai/Mistral-7B-v0.1", device_map="auto", load_in_4bit=True
|
||||
... )
|
||||
```
|
||||
|
||||
@ -113,12 +113,12 @@ pip install transformers bitsandbytes>=0.39.0 -q
|
||||
هناك العديد من [استراتيجيات التوليد](generation_strategies)، وفي بعض الأحيان قد لا تكون القيم الافتراضية مناسبة لحالتك الاستخدام. إذا لم تكن الإخراج الخاصة بك متوافقة مع ما تتوقعه، فقد قمنا بإنشاء قائمة بأكثر الأخطاء الشائعة وكيفية تجنبها.
|
||||
|
||||
```py
|
||||
>>> from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
|
||||
>>> from transformers import AutoModelForCausalLM, AutoTokenizer
|
||||
|
||||
>>> tokenizer = AutoTokenizer.from_pretrained("mistralai/Mistral-7B-v0.1")
|
||||
>>> tokenizer.pad_token = tokenizer.eos_token # Most LLMs don't have a pad token by default
|
||||
>>> model = AutoModelForCausalLM.from_pretrained(
|
||||
... "mistralai/Mistral-7B-v0.1", device_map="auto", quantization_config=BitsAndBytesConfig(load_in_4bit=True)
|
||||
... "mistralai/Mistral-7B-v0.1", device_map="auto", load_in_4bit=True
|
||||
... )
|
||||
```
|
||||
|
||||
@ -192,7 +192,7 @@ LLMs هي [معماريات فك التشفير فقط](https://huggingface.co/l
|
||||
```python
|
||||
>>> tokenizer = AutoTokenizer.from_pretrained("HuggingFaceH4/zephyr-7b-alpha")
|
||||
>>> model = AutoModelForCausalLM.from_pretrained(
|
||||
... "HuggingFaceH4/zephyr-7b-alpha", device_map="auto", quantization_config=BitsAndBytesConfig(load_in_4bit=True)
|
||||
... "HuggingFaceH4/zephyr-7b-alpha", device_map="auto", load_in_4bit=True
|
||||
... )
|
||||
>>> set_seed(0)
|
||||
>>> prompt = """How many helicopters can a human eat in one sitting? Reply as a thug."""
|
||||
|
||||
@ -231,7 +231,7 @@ flush()
|
||||
دعنا نرى ما هو استهلاك ذاكرة GPU الذروة الذي يوفره تكميم 4 بت. يمكن تكميم النموذج إلى 4 بت باستخدام نفس واجهة برمجة التطبيقات كما في السابق - هذه المرة عن طريق تمرير `load_in_4bit=True` بدلاً من `load_in_8bit=True`.
|
||||
|
||||
```python
|
||||
model = AutoModelForCausalLM.from_pretrained("bigcode/octocoder", quantization_config=BitsAndBytesConfig(load_in_4bit=True), pad_token_id=0)
|
||||
model = AutoModelForCausalLM.from_pretrained("bigcode/octocoder", load_in_4bit=True, pad_token_id=0)
|
||||
|
||||
pipe = pipeline("text-generation", model=model, tokenizer=tokenizer)
|
||||
|
||||
@ -329,6 +329,174 @@ $$ \textbf{O}_i \leftarrow s^a_{ij} * \textbf{O}_i + s^b_{ij} * \mathbf{V}_{j} \
|
||||
لنلقِ نظرة على مثال عملي.
|
||||
|
||||
|
||||
يحصل نموذج OctoCoder الخاص بنا الآن على موجه إدخال أطول بشكل كبير يتضمن ما يسمى *موجه النظام*. تُستخدم موجهات النظام لتوجيه LLM إلى مساعد أفضل مصمم لمهام المستخدمين.
|
||||
فيما يلي، نستخدم موجه النظام الذي سيجعل OctoCoder مساعد ترميز أفضل.
|
||||
|
||||
```python
|
||||
system_prompt = """Below are a series of dialogues between various people and an AI technical assistant.
|
||||
The assistant tries to be helpful, polite, honest, sophisticated, emotionally aware, and humble but knowledgeable.
|
||||
The assistant is happy to help with code questions and will do their best to understand exactly what is needed.
|
||||
It also tries to avoid giving false or misleading information, and it caveats when it isn't entirely sure about the right answer.
|
||||
That said, the assistant is practical really does its best, and doesn't let caution get too much in the way of being useful.
|
||||
|
||||
The Starcoder models are a series of 15.5B parameter models trained on 80+ programming languages from The Stack (v1.2) (excluding opt-out requests).
|
||||
The model uses Multi Query Attention, was trained using the Fill-in-the-Middle objective, and with 8,192 tokens context window for a trillion tokens of heavily deduplicated data.
|
||||
-----
|
||||
|
||||
Question: Write a function that takes two lists and returns a list that has alternating elements from each input list.
|
||||
|
||||
Answer: Sure. Here is a function that does that.
|
||||
|
||||
def alternating(list1, list2):
|
||||
results = []
|
||||
for i in range(len(list1)):
|
||||
results.append(list1[i])
|
||||
results.append(list2[i])
|
||||
return results
|
||||
|
||||
Question: Can you write some test cases for this function?
|
||||
|
||||
Answer: Sure, here are some tests.
|
||||
|
||||
assert alternating([10, 20, 30], [1, 2, 3]) == [10, 1, 20, 2, 30, 3]
|
||||
assert alternating([True, False], [4, 5]) == [True, 4, False, 5]
|
||||
assert alternating([], []) == []
|
||||
|
||||
Question: Modify the function so that it returns all input elements when the lists have uneven length. The elements from the longer list should be at the end.
|
||||
|
||||
Answer: Here is the modified function.
|
||||
|
||||
def alternating(list1, list2):
|
||||
results = []
|
||||
for i in range(min(len(list1), len(list2))):
|
||||
results.append(list1[i])
|
||||
results.append(list2[i])
|
||||
if len(list1) > len(list2):
|
||||
results.extend(list1[i+1:])
|
||||
else:
|
||||
results.extend(list2[i+1:])
|
||||
return results
|
||||
-----
|
||||
"""
|
||||
```
|
||||
لأغراض التوضيح، سنكرر موجه النظام عشر مرات بحيث يكون طول الإدخال طويلاً بما يكفي لملاحظة وفورات ذاكرة Flash Attention.
|
||||
نضيف موجه النص الأصلي "سؤال: يرجى كتابة وظيفة في Python تقوم بتحويل البايتات إلى جيجا بايت.
|
||||
|
||||
```python
|
||||
long_prompt = 10 * system_prompt + prompt
|
||||
```
|
||||
|
||||
نقوم بتنفيذ نموذجنا مرة أخرى بدقة bfloat16.
|
||||
|
||||
```python
|
||||
model = AutoModelForCausalLM.from_pretrained("bigcode/octocoder", dtype=torch.bfloat16, device_map="auto")
|
||||
tokenizer = AutoTokenizer.from_pretrained("bigcode/octocoder")
|
||||
|
||||
pipe = pipeline("text-generation", model=model, tokenizer=tokenizer)
|
||||
```
|
||||
|
||||
دعنا الآن نقوم بتشغيل النموذج تمامًا مثلما كان من قبل *بدون اهتمام فلاشي* وقياس متطلبات ذاكرة GPU وقت الذروة ووقت الاستدلال.
|
||||
|
||||
```python
|
||||
import time
|
||||
|
||||
start_time = time.time()
|
||||
result = pipe(long_prompt, max_new_tokens=60)[0]["generated_text"][len(long_prompt):]
|
||||
|
||||
print(f"Generated in {time.time() - start_time} seconds.")
|
||||
result
|
||||
```
|
||||
|
||||
**الإخراج**:
|
||||
```
|
||||
تم التوليد في 10.96854019165039 ثانية.
|
||||
بالتأكيد. إليك وظيفة للقيام بذلك.
|
||||
|
||||
def bytes_to_giga(bytes):
|
||||
return bytes / 1024 / 1024 / 1024
|
||||
|
||||
الإجابة: بالتأكيد. إليك وظيفة للقيام بذلك.
|
||||
|
||||
ديف
|
||||
```
|
||||
|
||||
نحصل على نفس الإخراج كما كان من قبل، ولكن هذه المرة، يقوم النموذج بتكرار الإجابة عدة مرات حتى يتم قطعها عند 60 رمزًا. ليس من المستغرب أننا كررنا موجه النظام عشر مرات لأغراض التوضيح وبالتالي قمنا بتشغيل النموذج لتكرار نفسه.
|
||||
|
||||
**ملاحظة** لا ينبغي تكرار موجه النظام عشر مرات في التطبيقات الواقعية - مرة واحدة كافية!
|
||||
|
||||
دعنا نقيس متطلبات ذاكرة GPU وقت الذروة.
|
||||
|
||||
```python
|
||||
bytes_to_giga_bytes(torch.cuda.max_memory_allocated())
|
||||
```
|
||||
|
||||
**الإخراج**:
|
||||
```
|
||||
37.668193340301514
|
||||
```
|
||||
|
||||
كما نرى، فإن متطلبات ذاكرة GPU وقت الذروة أعلى بكثير مما كانت عليه في البداية، وهو ما يرجع إلى حد كبير إلى تسلسل الإدخال الأطول. أيضًا، يستغرق التوليد أكثر من دقيقة بقليل الآن.
|
||||
|
||||
نستدعي `flush()` لتحرير ذاكرة GPU لتجربتنا التالية.
|
||||
|
||||
```python
|
||||
flush()
|
||||
```
|
||||
|
||||
لمقارنة، دعونا نقوم بتشغيل نفس الدالة، ولكن تمكين الاهتمام فلاش بدلا من ذلك.
|
||||
للقيام بذلك، نقوم بتحويل النموذج إلى [BetterTransformer](Https://huggingface.co/docs/optimum/bettertransformer/overview) ومن خلال القيام بذلك تمكين PyTorch's [SDPA self-attention](Https://pytorch.org/docs/master/generated/torch.nn.functional.scaled_dot_product_attention) والتي بدورها قادرة على استخدام الاهتمام فلاش.
|
||||
|
||||
```python
|
||||
model.to_bettertransformer()
|
||||
```
|
||||
|
||||
الآن نقوم بتشغيل نفس مقتطف التعليمات البرمجية بالضبط كما كان من قبل وتحت الغطاء سوف تستخدم المحولات الاهتمام فلاش.
|
||||
|
||||
```py
|
||||
start_time = time.time()
|
||||
with torch.backends.cuda.sdp_kernel(enable_flash=True, enable_math=False, enable_mem_efficient=False):
|
||||
result = pipe(long_prompt, max_new_tokens=60)[0]["generated_text"][len(long_prompt):]
|
||||
|
||||
print(f"Generated in {time.time() - start_time} seconds.")
|
||||
result
|
||||
```
|
||||
|
||||
**الإخراج**:
|
||||
```
|
||||
تم التوليد في 3.0211617946624756 ثانية.
|
||||
بالتأكيد. إليك وظيفة للقيام بذلك.
|
||||
|
||||
def bytes_to_giga(bytes):
|
||||
return bytes / 1024 / 1024 / 1024
|
||||
|
||||
الإجابة: بالتأكيد. إليك وظيفة للقيام بذلك.
|
||||
|
||||
ديف
|
||||
```
|
||||
|
||||
نحصل على نفس النتيجة بالضبط كما كان من قبل، ولكن يمكننا ملاحظة تسريع كبير بفضل الاهتمام فلاش.
|
||||
|
||||
دعنا نقيس استهلاك الذاكرة لآخر مرة.
|
||||
|
||||
```python
|
||||
bytes_to_giga_bytes(torch.cuda.max_memory_allocated())
|
||||
```
|
||||
|
||||
**الإخراج**:
|
||||
```
|
||||
32.617331981658936
|
||||
```
|
||||
|
||||
ونحن تقريبا مرة أخرى إلى ذاكرة GPU الذروة الأصلية لدينا 29GB.
|
||||
|
||||
يمكننا أن نلاحظ أننا نستخدم فقط حوالي 100 ميجابايت إضافية من ذاكرة GPU عند تمرير تسلسل إدخال طويل جدًا مع الاهتمام فلاش مقارنة بتمرير تسلسل إدخال قصير كما فعلنا في البداية.
|
||||
|
||||
```py
|
||||
flush()
|
||||
```
|
||||
|
||||
لمزيد من المعلومات حول كيفية استخدام Flash Attention، يرجى الاطلاع على [صفحة doc هذه](Https://huggingface.co/docs/transformers/en/perf_infer_gpu_one#flashattention-2).
|
||||
|
||||
## 3. الابتكارات المعمارية
|
||||
|
||||
حتى الآن، نظرنا في تحسين الكفاءة الحسابية والذاكرة من خلال:
|
||||
@ -472,7 +640,7 @@ for _ in range(5):
|
||||
next_token_id = torch.argmax(next_logits, dim=-1)
|
||||
|
||||
print("shape of input_ids", next_token_id.shape)
|
||||
print("length of key-value cache", past_key_values.get_seq_length()) # past_key_values are of shape [num_layers, 0 for k, 1 for v, batch_size, length, hidden_dim]
|
||||
print("length of key-value cache", len(past_key_values[0][0])) # past_key_values are of shape [num_layers, 0 for k, 1 for v, batch_size, length, hidden_dim]
|
||||
generated_tokens.append(next_token_id.item())
|
||||
|
||||
generated_text = tokenizer.batch_decode(generated_tokens)
|
||||
|
||||
@ -65,15 +65,43 @@ pip install huggingface_hub
|
||||
|
||||
تحويل نقطة التحقق لإطار عمل آخر أمر سهل. تأكد من تثبيت PyTorch و TensorFlow (راجع [هنا](installation) لتعليمات التثبيت)، ثم ابحث عن النموذج الملائم لمهمتك في الإطار الآخر.
|
||||
|
||||
<frameworkcontent>
|
||||
<pt>
|
||||
حدد `from_tf=True` لتحويل نقطة تحقق من TensorFlow إلى PyTorch:
|
||||
|
||||
```py
|
||||
>>> pt_model = DistilBertForSequenceClassification.from_pretrained("path/to/awesome-name-you-picked", from_tf=True)
|
||||
>>> pt_model.save_pretrained("path/to/awesome-name-you-picked")
|
||||
```
|
||||
</pt>
|
||||
<tf>
|
||||
حدد `from_pt=True` لتحويل نقطة تحقق من PyTorch إلى TensorFlow:
|
||||
|
||||
```py
|
||||
>>> tf_model = TFDistilBertForSequenceClassification.from_pretrained("path/to/awesome-name-you-picked", from_pt=True)
|
||||
```
|
||||
|
||||
بعد ذلك، يمكنك حفظ نموذج TensorFlow الجديد بنقطة التحقق الجديدة:
|
||||
|
||||
```py
|
||||
>>> tf_model.save_pretrained("path/to/awesome-name-you-picked")
|
||||
```
|
||||
</tf>
|
||||
<jax>
|
||||
إذا كان النموذج متاحًا في Flax، فيمكنك أيضًا تحويل نقطة تحقق من PyTorch إلى Flax:
|
||||
|
||||
```py
|
||||
>>> flax_model = FlaxDistilBertForSequenceClassification.from_pretrained(
|
||||
... "path/to/awesome-name-you-picked", from_pt=True
|
||||
... )
|
||||
```
|
||||
</jax>
|
||||
</frameworkcontent>
|
||||
|
||||
## دفع نموذج أثناء التدريب
|
||||
|
||||
<frameworkcontent>
|
||||
<pt>
|
||||
<Youtube id="Z1-XMy-GNLQ"/>
|
||||
|
||||
مشاركة نموذجك على Hub مر بسيط للغاية كل ما عليك هو إضافة معلمة أو استدعاء رد إضافي. كما تذكر من درس [التدريب الدقيق](training)، فإن فئة [`TrainingArguments`] هي المكان الذي تحدد فيه المعلمات الفائقة وخيارات التدريب الإضافية. تشمل إحدى خيارات التدريب هذه القدرة على دفع النموذج مباشرة إلى المنصة Hub. قم بتعيين `push_to_hub=True` في [`TrainingArguments`]:
|
||||
@ -99,6 +127,29 @@ pip install huggingface_hub
|
||||
```py
|
||||
>>> trainer.push_to_hub()
|
||||
```
|
||||
</pt>
|
||||
<tf>
|
||||
شارك نموذجًا على Hub باستخدام [`PushToHubCallback`]. في دالة [`PushToHubCallback`], أضف:
|
||||
|
||||
- دليل إخراج لنموذجك.
|
||||
- مُجزّئ اللغوي.
|
||||
- `hub_model_id`، والذي هو اسم مستخدم Hub واسم النموذج الخاص بك.
|
||||
|
||||
```py
|
||||
>>> from transformers import PushToHubCallback
|
||||
|
||||
>>> push_to_hub_callback = PushToHubCallback(
|
||||
... output_dir="./your_model_save_path", tokenizer=tokenizer, hub_model_id="your-username/my-awesome-model"
|
||||
... )
|
||||
```
|
||||
|
||||
أضف الاستدعاء إلى [`fit`](https://keras.io/api/models/model_training_apis/)، وسيقوم 🤗 Transformers بدفع النموذج المدرب إلى Hub:
|
||||
|
||||
```py
|
||||
>>> model.fit(tf_train_dataset, validation_data=tf_validation_dataset, epochs=3, callbacks=push_to_hub_callback)
|
||||
```
|
||||
</tf>
|
||||
</frameworkcontent>
|
||||
|
||||
## استخدام دالة `push_to_hub`
|
||||
|
||||
@ -169,4 +220,4 @@ pip install huggingface_hub
|
||||
* قم بإنشاء ملف `README.md` وتحميله يدويًا.
|
||||
* انقر فوق الزر **Edit model card** في مستودع نموذجك.
|
||||
|
||||
الق نظرة على بطاقة [DistilBert](https://huggingface.co/distilbert/distilbert-base-uncased) للحصول على مثال جيد على نوع المعلومات التي يجب أن تتضمنها بطاقة النموذج. للحصول على مزيد من التفاصيل حول الخيارات الأخرى التي يمكنك التحكم فيها في ملف `README.md` مثل البصمة الكربونية للنموذج أو أمثلة الأداة، راجع الوثائق [هنا](https://huggingface.co/docs/hub/models-cards).
|
||||
الق نظرة على بطاقة [DistilBert](https://huggingface.co/distilbert/distilbert-base-uncased) للحصول على مثال جيد على نوع المعلومات التي يجب أن تتضمنها بطاقة النموذج. للحصول على مزيد من التفاصيل حول الخيارات الأخرى التي يمكنك التحكم فيها في ملف `README.md` مثل البصمة الكربونية للنموذج أو أمثلة الأداة، راجع الوثائق [هنا](https://huggingface.co/docs/hub/models-cards).
|
||||
@ -152,6 +152,8 @@ pip install datasets
|
||||
|
||||
قم بتعيين معلمة `return_tensors` إلى إما `pt` لـ PyTorch، أو `tf` لـ TensorFlow:
|
||||
|
||||
<frameworkcontent>
|
||||
<pt>
|
||||
|
||||
```py
|
||||
>>> batch_sentences = [
|
||||
@ -171,6 +173,33 @@ pip install datasets
|
||||
[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
|
||||
[1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0]])}
|
||||
```
|
||||
</pt>
|
||||
<tf>
|
||||
|
||||
```py
|
||||
>>> batch_sentences = [
|
||||
... "But what about second breakfast?",
|
||||
... "Don't think he knows about second breakfast, Pip.",
|
||||
... "What about elevensies?",
|
||||
... ]
|
||||
>>> encoded_input = tokenizer(batch_sentences, padding=True, truncation=True, return_tensors="tf")
|
||||
>>> print(encoded_input)
|
||||
{'input_ids': <tf.Tensor: shape=(2, 9), dtype=int32, numpy=
|
||||
array([[101, 1252, 1184, 1164, 1248, 6462, 136, 102, 0, 0, 0, 0, 0, 0, 0],
|
||||
[101, 1790, 112, 189, 1341, 1119, 3520, 1164, 1248, 6462, 117, 21902, 1643, 119, 102],
|
||||
[101, 1327, 1164, 5450, 23434, 136, 102, 0, 0, 0, 0, 0, 0, 0, 0]],
|
||||
dtype=int32)>,
|
||||
'token_type_ids': <tf.Tensor: shape=(2, 9), dtype=int32, numpy=
|
||||
array([[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
|
||||
[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
|
||||
[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]], dtype=int32)>,
|
||||
'attention_mask': <tf.Tensor: shape=(2, 9), dtype=int32, numpy=
|
||||
array([[1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0],
|
||||
[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
|
||||
[1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0]], dtype=int32)>}
|
||||
```
|
||||
</tf>
|
||||
</frameworkcontent>
|
||||
|
||||
<Tip>
|
||||
|
||||
|
||||
@ -12,10 +12,20 @@
|
||||
|
||||
ستحتاج أيضًا إلى تثبيت إطار عمل التعلم الآلي المفضل لديك:
|
||||
|
||||
<frameworkcontent>
|
||||
<pt>
|
||||
|
||||
```bash
|
||||
pip install torch
|
||||
```
|
||||
</pt>
|
||||
<tf>
|
||||
|
||||
```bash
|
||||
pip install tensorflow
|
||||
```
|
||||
</tf>
|
||||
</frameworkcontent>
|
||||
|
||||
## خط الأنابيب
|
||||
|
||||
@ -112,6 +122,8 @@ label: NEGATIVE, with score: 0.5309
|
||||
>>> model_name = "nlptown/bert-base-multilingual-uncased-sentiment"
|
||||
```
|
||||
|
||||
<frameworkcontent>
|
||||
<pt>
|
||||
استخدم [`AutoModelForSequenceClassification`] و [`AutoTokenizer`] لتحميل النموذج المُدرب مسبقًا ومعالجته المرتبط به (مزيد من المعلومات حول `AutoClass` في القسم التالي):
|
||||
|
||||
```py
|
||||
@ -120,6 +132,18 @@ label: NEGATIVE, with score: 0.5309
|
||||
>>> model = AutoModelForSequenceClassification.from_pretrained(model_name)
|
||||
>>> tokenizer = AutoTokenizer.from_pretrained(model_name)
|
||||
```
|
||||
</pt>
|
||||
<tf>
|
||||
استخدم [`TFAutoModelForSequenceClassification`] و [`AutoTokenizer`] لتحميل النموذج المُدرب مسبقًا ومعالجته المرتبط به (مزيد من المعلومات حول `TFAutoClass` في القسم التالي):
|
||||
|
||||
```py
|
||||
>>> from transformers import AutoTokenizer, TFAutoModelForSequenceClassification
|
||||
|
||||
>>> model = TFAutoModelForSequenceClassification.from_pretrained(model_name)
|
||||
>>> tokenizer = AutoTokenizer.from_pretrained(model_name)
|
||||
```
|
||||
</tf>
|
||||
</frameworkcontent>
|
||||
|
||||
حدد النموذج والمعالج في [`pipeline`]. الآن يمكنك تطبيق `classifier` على النص الفرنسي:
|
||||
|
||||
@ -168,6 +192,8 @@ label: NEGATIVE, with score: 0.5309
|
||||
|
||||
يمكن المجزئ أيضًا قبول قائمة من المدخلات، ويقوم بـ "حشو" و"تقصير" النص لإرجاع كدفعة بطول موحد:
|
||||
|
||||
<frameworkcontent>
|
||||
<pt>
|
||||
|
||||
```py
|
||||
>>> pt_batch = tokenizer(
|
||||
@ -178,6 +204,20 @@ label: NEGATIVE, with score: 0.5309
|
||||
... return_tensors="pt",
|
||||
... )
|
||||
```
|
||||
</pt>
|
||||
<tf>
|
||||
|
||||
```py
|
||||
>>> tf_batch = tokenizer(
|
||||
... ["We are very happy to show you the 🤗 Transformers library.", "We hope you don't hate it."],
|
||||
... padding=True,
|
||||
... truncation=True,
|
||||
... max_length=512,
|
||||
... return_tensors="tf",
|
||||
... )
|
||||
```
|
||||
</tf>
|
||||
</frameworkcontent>
|
||||
|
||||
<Tip>
|
||||
|
||||
@ -187,6 +227,8 @@ label: NEGATIVE, with score: 0.5309
|
||||
|
||||
### AutoModel
|
||||
|
||||
<frameworkcontent>
|
||||
<pt>
|
||||
تقدم مكتبة 🤗 Transformers طريقة بسيطة وموحدة لتحميل نماذج مدربة مسبقًا. وهذا يعني أنه يمكنك تحميل [`AutoModel`] كما لو كنت تقوم بتحميل [`AutoTokenizer`]. الفرق الوحيد هو اختيار فئة [`AutoModel`] المناسبة للمهمة. بالنسبة لتصنيف النص (أو التسلسل)، يجب عليك تحميل [`AutoModelForSequenceClassification`]:
|
||||
|
||||
```py
|
||||
@ -222,6 +264,39 @@ label: NEGATIVE, with score: 0.5309
|
||||
tensor([[0.0021, 0.0018, 0.0115, 0.2121, 0.7725],
|
||||
[0.2084, 0.1826, 0.1969, 0.1755, 0.2365]], grad_fn=<SoftmaxBackward0>)
|
||||
```
|
||||
</pt>
|
||||
<tf>
|
||||
يوفر 🤗 Transformers طريقة بسيطة وموحدة لتحميل مثيلات مُدربة مسبقًا. وهذا يعني أنه يمكنك تحميل [`TFAutoModel`] مثل تحميل [`AutoTokenizer`]. والفرق الوحيد هو تحديد [`TFAutoModel`] الصحيح للمهمة. للتصنيف النصي (أو التسلسلي)، يجب تحميل [`TFAutoModelForSequenceClassification`]:
|
||||
|
||||
```py
|
||||
>>> from transformers import TFAutoModelForSequenceClassification
|
||||
|
||||
>>> model_name = "nlptown/bert-base-multilingual-uncased-sentiment"
|
||||
>>> tf_model = TFAutoModelForSequenceClassification.from_pretrained(model_name)
|
||||
```
|
||||
|
||||
<Tip>
|
||||
|
||||
راجع [ملخص المهام](./task_summary) للمهام المدعومة بواسطة فئة [`AutoModel`].
|
||||
|
||||
</Tip>
|
||||
|
||||
الآن، مرر دفعة المدخلات المعالجة مسبقًا مباشرة إلى النموذج. يمكنك تمرير المصفوفات كما هي:
|
||||
|
||||
```py
|
||||
>>> tf_outputs = tf_model(tf_batch)
|
||||
```
|
||||
|
||||
يقوم النموذج بإخراج التنشيطات النهائية في سمة `logits`. طبق دالة softmax على `logits` لاسترداد الاحتمالات:
|
||||
|
||||
```py
|
||||
>>> import tensorflow as tf
|
||||
|
||||
>>> tf_predictions = tf.nn.softmax(tf_outputs.logits, axis=-1)
|
||||
>>> tf_predictions # doctest: +IGNORE_RESULT
|
||||
```
|
||||
</tf>
|
||||
</frameworkcontent>
|
||||
|
||||
<Tip>
|
||||
|
||||
@ -231,6 +306,8 @@ tensor([[0.0021, 0.0018, 0.0115, 0.2121, 0.7725],
|
||||
|
||||
### حفظ النموذج
|
||||
|
||||
<frameworkcontent>
|
||||
<pt>
|
||||
بمجرد ضبط نموذجك، يمكنك حفظه مع برنامج الترميز الخاص به باستخدام [`PreTrainedModel.save_pretrained`]:
|
||||
|
||||
```py
|
||||
@ -244,9 +321,28 @@ tensor([[0.0021, 0.0018, 0.0115, 0.2121, 0.7725],
|
||||
```py
|
||||
>>> pt_model = AutoModelForSequenceClassification.from_pretrained("./pt_save_pretrained")
|
||||
```
|
||||
</pt>
|
||||
<tf>
|
||||
بمجرد ضبط نموذجك، يمكنك حفظه مع برنامج الترميز الخاص به باستخدام [`TFPreTrainedModel.save_pretrained`]:
|
||||
|
||||
```py
|
||||
>>> tf_save_directory = "./tf_save_pretrained"
|
||||
>>> tokenizer.save_pretrained(tf_save_directory) # doctest: +IGNORE_RESULT
|
||||
>>> tf_model.save_pretrained(tf_save_directory)
|
||||
```
|
||||
|
||||
عندما تكون مستعدًا لاستخدام النموذج مرة أخرى، أعد تحميله باستخدام [`TFPreTrainedModel.from_pretrained`]:
|
||||
|
||||
```py
|
||||
>>> tf_model = TFAutoModelForSequenceClassification.from_pretrained("./tf_save_pretrained")
|
||||
```
|
||||
</tf>
|
||||
</frameworkcontent>
|
||||
|
||||
من الميزات الرائعة في 🤗 Transformers القدرة على حفظ نموذج وإعادة تحميله كنموذج PyTorch أو TensorFlow. يمكن أن يحول معامل `from_pt` أو `from_tf` النموذج من إطار عمل إلى آخر:
|
||||
|
||||
<frameworkcontent>
|
||||
<pt>
|
||||
|
||||
```py
|
||||
>>> from transformers import AutoModel
|
||||
@ -254,6 +350,17 @@ tensor([[0.0021, 0.0018, 0.0115, 0.2121, 0.7725],
|
||||
>>> tokenizer = AutoTokenizer.from_pretrained(pt_save_directory)
|
||||
>>> pt_model = AutoModelForSequenceClassification.from_pretrained(pt_save_directory, from_pt=True)
|
||||
```
|
||||
</pt>
|
||||
<tf>
|
||||
|
||||
```py
|
||||
>>> from transformers import TFAutoModel
|
||||
|
||||
>>> tokenizer = AutoTokenizer.from_pretrained(tf_save_directory)
|
||||
>>> tf_model = TFAutoModelForSequenceClassification.from_pretrained(tf_save_directory, from_tf=True)
|
||||
```
|
||||
</tf>
|
||||
</frameworkcontent>
|
||||
|
||||
|
||||
## إنشاء نماذج مخصصة
|
||||
@ -268,6 +375,8 @@ tensor([[0.0021, 0.0018, 0.0115, 0.2121, 0.7725],
|
||||
>>> my_config = AutoConfig.from_pretrained("distilbert/distilbert-base-uncased", n_heads=12)
|
||||
```
|
||||
|
||||
<frameworkcontent>
|
||||
<pt>
|
||||
قم بإنشاء نموذج من تكوينك المخصص باستخدام [`AutoModel.from_config`]:
|
||||
|
||||
```py
|
||||
@ -275,6 +384,17 @@ tensor([[0.0021, 0.0018, 0.0115, 0.2121, 0.7725],
|
||||
|
||||
>>> my_model = AutoModel.from_config(my_config)
|
||||
```
|
||||
</pt>
|
||||
<tf>
|
||||
قم بإنشاء نموذج من تكوينك المخصص باستخدام [`TFAutoModel.from_config`]:
|
||||
|
||||
```py
|
||||
>>> from transformers import TFAutoModel
|
||||
|
||||
>>> my_model = TFAutoModel.from_config(my_config)
|
||||
```
|
||||
</tf>
|
||||
</frameworkcontent>
|
||||
|
||||
الق نظرة على دليل [إنشاء بنية مخصصة](./create_a_model) لمزيد من المعلومات حول بناء التكوينات المخصصة.
|
||||
|
||||
|
||||
@ -76,6 +76,8 @@ pip install -r requirements.txt
|
||||
|
||||
## تشغيل نص برمجي
|
||||
|
||||
<frameworkcontent>
|
||||
<pt>
|
||||
|
||||
- يقوم النص البرمجي التوضيحي بتنزيل مجموعة بيانات ومعالجتها مسبقًا من مكتبة 🤗 [Datasets](https://huggingface.co/docs/datasets).
|
||||
- ثم يقوم النص البرمجي بضبط نموذج بيانات دقيق باستخدام [Trainer](https://huggingface.co/docs/transformers/main_classes/trainer) على بنية تدعم الملخص.
|
||||
@ -93,8 +95,31 @@ python examples/pytorch/summarization/run_summarization.py \
|
||||
--output_dir /tmp/tst-summarization \
|
||||
--per_device_train_batch_size=4 \
|
||||
--per_device_eval_batch_size=4 \
|
||||
--overwrite_output_dir \
|
||||
--predict_with_generate
|
||||
```
|
||||
</pt>
|
||||
<tf>
|
||||
|
||||
- يقوم النص البرمجي التوضيحي بتنزيل مجموعة بيانات ومعالجتها مسبقًا من مكتبة 🤗 [Datasets](https://huggingface.co/docs/datasets/).
|
||||
- ثم يقوم النص البرمجي بضبط نموذج بيانات دقيق باستخدام Keras على بنية تدعم الملخص.
|
||||
- يوضح المثال التالي كيفية ضبط نموذج [T5-small](https://huggingface.co/google-t5/t5-small) على مجموعة بيانات [CNN/DailyMail](https://huggingface.co/datasets/cnn_dailymail).
|
||||
- يتطلب نموذج T5 ماعمل `source_prefix` إضافية بسبب الطريقة التي تم تدريبه بها. يتيح هذا المطالبة لـ T5 معرفة أن هذه مهمة التلخيص.
|
||||
|
||||
```bash
|
||||
python examples/tensorflow/summarization/run_summarization.py \
|
||||
--model_name_or_path google-t5/t5-small \
|
||||
--dataset_name cnn_dailymail \
|
||||
--dataset_config "3.0.0" \
|
||||
--output_dir /tmp/tst-summarization \
|
||||
--per_device_train_batch_size 8 \
|
||||
--per_device_eval_batch_size 16 \
|
||||
--num_train_epochs 3 \
|
||||
--do_train \
|
||||
--do_eval
|
||||
```
|
||||
</tf>
|
||||
</frameworkcontent>
|
||||
|
||||
## التدريب الموزع والدقة المختلطة
|
||||
|
||||
@ -116,6 +141,7 @@ torchrun \
|
||||
--output_dir /tmp/tst-summarization \
|
||||
--per_device_train_batch_size=4 \
|
||||
--per_device_eval_batch_size=4 \
|
||||
--overwrite_output_dir \
|
||||
--predict_with_generate
|
||||
```
|
||||
|
||||
@ -123,6 +149,8 @@ torchrun \
|
||||
|
||||
## تشغيل نص برمجي على وحدة معالجة الدقة الفائقة (TPU)
|
||||
|
||||
<frameworkcontent>
|
||||
<pt>
|
||||
|
||||
تُعد وحدات معالجة الدقة الفائقة (TPUs) مصممة خصيصًا لتسريع الأداء. يدعم PyTorch وحدات معالجة الدقة الفائقة (TPUs) مع [XLA](https://www.tensorflow.org/xla) مجمع الدقة الفائقة للتعلم العميق (راجع [هنا](https://github.com/pytorch/xla/blob/master/README.md) لمزيد من التفاصيل). لاستخدام وحدة معالجة الدقة الفائقة (TPU)، قم بتشغيل نص `xla_spawn.py` البرمجي واستخدم معامل `num_cores` لتعيين عدد وحدات معالجة الدقة الفائقة (TPU) التي تريد استخدامها.
|
||||
|
||||
@ -138,8 +166,28 @@ python xla_spawn.py --num_cores 8 \
|
||||
--output_dir /tmp/tst-summarization \
|
||||
--per_device_train_batch_size=4 \
|
||||
--per_device_eval_batch_size=4 \
|
||||
--overwrite_output_dir \
|
||||
--predict_with_generate
|
||||
```
|
||||
</pt>
|
||||
<tf>
|
||||
|
||||
تُعد وحدات معالجة الدقة الفائقة (TPUs) مصممة خصيصًا لتسريع الأداء. تستخدم نصوص TensorFlow البرمجية استراتيجية [`TPUStrategy`](https://www.tensorflow.org/guide/distributed_training#tpustrategy) للتدريب على وحدات معالجة الدقة الفائقة (TPUs). لاستخدام وحدة معالجة الدقة الفائقة (TPU)، قم بتمرير اسم مورد وحدة معالجة الدقة الفائقة (TPU) إلى حجة `tpu`.
|
||||
```bash
|
||||
python run_summarization.py \
|
||||
--tpu name_of_tpu_resource \
|
||||
--model_name_or_path google-t5/t5-small \
|
||||
--dataset_name cnn_dailymail \
|
||||
--dataset_config "3.0.0" \
|
||||
--output_dir /tmp/tst-summarization \
|
||||
--per_device_train_batch_size 8 \
|
||||
--per_device_eval_batch_size 16 \
|
||||
--num_train_epochs 3 \
|
||||
--do_train \
|
||||
--do_eval
|
||||
```
|
||||
</tf>
|
||||
</frameworkcontent>
|
||||
|
||||
## تشغيل نص برمجي باستخدام 🤗 Accelerate
|
||||
|
||||
@ -194,6 +242,7 @@ python examples/pytorch/summarization/run_summarization.py \
|
||||
--summary_column summary_column_name \
|
||||
--source_prefix "summarize: " \
|
||||
--output_dir /tmp/tst-summarization \
|
||||
--overwrite_output_dir \
|
||||
--per_device_train_batch_size=4 \
|
||||
--per_device_eval_batch_size=4 \
|
||||
--predict_with_generate
|
||||
@ -221,6 +270,7 @@ python examples/pytorch/summarization/run_summarization.py \
|
||||
--output_dir /tmp/tst-summarization \
|
||||
--per_device_train_batch_size=4 \
|
||||
--per_device_eval_batch_size=4 \
|
||||
--overwrite_output_dir \
|
||||
--predict_with_generate
|
||||
```
|
||||
|
||||
@ -234,6 +284,8 @@ examples/pytorch/summarization/run_summarization.py -h
|
||||
|
||||
خيار آخر مفيد لتمكينه هو استئناف التدريب من نقطة تفتيش سابقة. سيضمن ذلك أنك تستطيع الاستمرار من حيث توقفت دون البدء من جديد إذا تم مقاطعة تدريبك. هناك طريقتان لاستئناف التدريب من نقطة تفتيش.
|
||||
|
||||
تستخدم الطريقة الأولى المعلمة `output_dir previous_output_dir` لاستئناف التدريب من أحدث نقطة تفتيش مخزنة في `output_dir`. في هذه الحالة، يجب عليك إزالة `overwrite_output_dir`:
|
||||
|
||||
```bash
|
||||
python examples/pytorch/summarization/run_summarization.py
|
||||
--model_name_or_path google-t5/t5-small \
|
||||
@ -245,6 +297,24 @@ python examples/pytorch/summarization/run_summarization.py
|
||||
--output_dir /tmp/tst-summarization \
|
||||
--per_device_train_batch_size=4 \
|
||||
--per_device_eval_batch_size=4 \
|
||||
--output_dir previous_output_dir \
|
||||
--predict_with_generate
|
||||
```
|
||||
|
||||
تستخدم الطريقة الثانية معلمة `resume_from_checkpoint path_to_specific_checkpoint` لاستئناف التدريب من مجلد نقطة تفتيش محددة.
|
||||
|
||||
```bash
|
||||
python examples/pytorch/summarization/run_summarization.py
|
||||
--model_name_or_path google-t5/t5-small \
|
||||
--do_train \
|
||||
--do_eval \
|
||||
--dataset_name cnn_dailymail \
|
||||
--dataset_config "3.0.0" \
|
||||
--source_prefix "summarize: " \
|
||||
--output_dir /tmp/tst-summarization \
|
||||
--per_device_train_batch_size=4 \
|
||||
--per_device_eval_batch_size=4 \
|
||||
--overwrite_output_dir \
|
||||
--resume_from_checkpoint path_to_specific_checkpoint \
|
||||
--predict_with_generate
|
||||
```
|
||||
@ -276,5 +346,6 @@ python examples/pytorch/summarization/run_summarization.py
|
||||
--output_dir /tmp/tst-summarization \
|
||||
--per_device_train_batch_size=4 \
|
||||
--per_device_eval_batch_size=4 \
|
||||
--overwrite_output_dir \
|
||||
--predict_with_generate
|
||||
```
|
||||
|
||||
@ -32,7 +32,7 @@
|
||||
لتصدير نموذج 🤗 Transformers إلى ONNX، قم أولاً بتثبيت اعتماد إضافي:
|
||||
|
||||
```bash
|
||||
pip install optimum-onnx
|
||||
pip install optimum[exporters]
|
||||
```
|
||||
|
||||
للاطلاع على جميع المعامﻻت المتاحة، يرجى الرجوع إلى [وثائق 🤗 Optimum](https://huggingface.co/docs/optimum/exporters/onnx/usage_guides/export_a_model#exporting-a-model-to-onnx-using-the-cli)، أو عرض المساعدة في سطر الأوامر:
|
||||
@ -111,3 +111,60 @@ optimum-cli export onnx --model keras-io/transformers-qa distilbert_base_cased_s
|
||||
### تصدير نموذج لهندسة غير مدعومة
|
||||
|
||||
إذا كنت ترغب في المساهمة من خلال إضافة دعم لنموذج لا يُمكن تصديره حاليًا، فيجب عليك أولاً التحقق مما إذا كان مدعومًا في [`optimum.exporters.onnx`](https://huggingface.co/docs/optimum/exporters/onnx/overview)، وإذا لم يكن مدعومًا، [فيمكنك المساهمة في 🤗 Optimum](https://huggingface.co/docs/optimum/exporters/onnx/usage_guides/contribute) مُباشرةً.
|
||||
|
||||
### تصدير نموذج باستخدام `transformers.onnx`
|
||||
|
||||
<Tip warning={true}>
|
||||
|
||||
لم يعد يتم دعم `transformers.onnx` يُرجى تصدير النماذج باستخدام 🤗 Optimum كما هو موضح أعلاه. سيتم إزالة هذا القسم في الإصدارات القادمة.
|
||||
|
||||
</Tip>
|
||||
|
||||
لتصدير نموذج 🤗 Transformers إلى ONNX باستخدام `transformers.onnx`، ثبّت التبعيات الإضافية:
|
||||
|
||||
```bash
|
||||
pip install transformers[onnx]
|
||||
```
|
||||
|
||||
استخدم حزمة `transformers.onnx` كنموذج Python لتصدير نقطة حفظ باستخدام تكوين جاهز:
|
||||
|
||||
```bash
|
||||
python -m transformers.onnx --model=distilbert/distilbert-base-uncased onnx/
|
||||
```
|
||||
|
||||
يُصدّر هذا رسمًا بيانيًا ONNX لنقطة الحفظ المُحددة بواسطة وسيطة `--model`. مرر أي نقطة حفظ على 🤗 Hub أو نقطة حفظ مُخزنة محليًا.
|
||||
يُمكن بعد ذلك تشغيل ملف `model.onnx` الناتج على أحد المُسرعات العديدة التي تدعم معيار ONNX. على سبيل المثال، قم بتحميل وتشغيل النموذج باستخدام ONNX Runtime كما يلي:
|
||||
|
||||
```python
|
||||
>>> from transformers import AutoTokenizer
|
||||
>>> from onnxruntime import InferenceSession
|
||||
|
||||
>>> tokenizer = AutoTokenizer.from_pretrained("distilbert/distilbert-base-uncased")
|
||||
>>> session = InferenceSession("onnx/model.onnx")
|
||||
>>> # يتوقع ONNX Runtime مصفوفات NumPy كمدخلات
|
||||
>>> inputs = tokenizer("Using DistilBERT with ONNX Runtime!", return_tensors="np")
|
||||
>>> outputs = session.run(output_names=["last_hidden_state"], input_feed=dict(inputs))
|
||||
```
|
||||
|
||||
يُمكن الحصول على أسماء المخرجات المطلوبة (مثل `["last_hidden_state"]`) من خلال إلقاء نظرة على تكوين ONNX لكل نموذج. على سبيل المثال، بالنسبة لـ DistilBERT، لدينا:
|
||||
|
||||
```python
|
||||
>>> from transformers.models.distilbert import DistilBertConfig, DistilBertOnnxConfig
|
||||
|
||||
>>> config = DistilBertConfig()
|
||||
>>> onnx_config = DistilBertOnnxConfig(config)
|
||||
>>> print(list(onnx_config.outputs.keys()))
|
||||
["last_hidden_state"]
|
||||
```
|
||||
|
||||
العمليات مُتطابقة لنقاط الحفظ TensorFlow على Hub. على سبيل المثال، صدّر نقطة حفظ TensorFlow خالصة كما يلي:
|
||||
|
||||
```bash
|
||||
python -m transformers.onnx --model=keras-io/transformers-qa onnx/
|
||||
```
|
||||
|
||||
لتصدير نموذج مُخزن محليًا، احفظ أوزان النموذج ومجزىء اللغوى في نفس الدليل (على سبيل المثال `local-pt-checkpoint`)، ثم قم بتصديره إلى ONNX عن طريق توجيه وسيط `--model` لحزمة `transformers.onnx` إلى الدليل المطلوب:
|
||||
|
||||
```bash
|
||||
python -m transformers.onnx --model=local-pt-checkpoint onnx/
|
||||
```
|
||||
@ -182,6 +182,8 @@ pip install transformers datasets evaluate
|
||||
|
||||
الآن قم بإنشاء دفعة من الأمثلة باستخدام [`DataCollatorForLanguageModeling`]. من الأفضل أن تقوم بـ *الحشو الديناميكي* للجمل إلى الطول الأطول في الدفعة أثناء التجميع، بدلاً من حشو كامل المجموعة من البيانات إلى الطول الأقصى.
|
||||
|
||||
<frameworkcontent>
|
||||
<pt>
|
||||
استخدم رمز نهاية التسلسل كرمز للحشو، وحدد `mlm_probability` لحجب الرموز بشكل عشوائي عند كل تكرار للبيانات:
|
||||
|
||||
```py
|
||||
@ -191,9 +193,23 @@ pip install transformers datasets evaluate
|
||||
>>> data_collator = DataCollatorForLanguageModeling(tokenizer=tokenizer, mlm=False)
|
||||
```
|
||||
|
||||
</pt>
|
||||
<tf>
|
||||
استخدم رمز نهاية التسلسل كرمز للحشو، وحدد `mlm_probability` لحجب الرموز بشكل عشوائي عند كل تكرار للبيانات:
|
||||
|
||||
```py
|
||||
>>> from transformers import DataCollatorForLanguageModeling
|
||||
|
||||
>>> data_collator = DataCollatorForLanguageModeling(tokenizer=tokenizer, mlm=False, return_tensors="tf")
|
||||
```
|
||||
|
||||
</tf>
|
||||
</frameworkcontent>
|
||||
|
||||
## التدريب (Train)
|
||||
|
||||
<frameworkcontent>
|
||||
<pt>
|
||||
|
||||
<Tip>
|
||||
|
||||
@ -251,6 +267,75 @@ Perplexity: 49.61
|
||||
```py
|
||||
>>> trainer.push_to_hub()
|
||||
```
|
||||
</pt>
|
||||
<tf>
|
||||
<Tip>
|
||||
|
||||
إذا لم تكن على دراية بتدريب نموذج باستخدام Keras، اطلع على [البرنامج التعليمي الأساسي](../training#train-a-tensorflow-model-with-keras)!
|
||||
|
||||
</Tip>
|
||||
لتدريب نموذج في TensorFlow، ابدأ بإعداد دالة المحسن، وجدول معدل التعلم، وبعض معاملات التدريب:
|
||||
|
||||
```py
|
||||
>>> from transformers import create_optimizer, AdamWeightDecay
|
||||
|
||||
>>> optimizer = AdamWeightDecay(learning_rate=2e-5, weight_decay_rate=0.01)
|
||||
```
|
||||
|
||||
ثم يمكنك تحميل DistilGPT2 باستخدام [`TFAutoModelForCausalLM`]:
|
||||
|
||||
```py
|
||||
>>> from transformers import TFAutoModelForCausalLM
|
||||
|
||||
>>> model = TFAutoModelForCausalLM.from_pretrained("distilbert/distilgpt2")
|
||||
```
|
||||
|
||||
حول مجموعات بياناتك إلى تنسيق `tf.data.Dataset` باستخدام [`~transformers.TFPreTrainedModel.prepare_tf_dataset`]:
|
||||
|
||||
```py
|
||||
>>> tf_train_set = model.prepare_tf_dataset(
|
||||
... lm_dataset["train"],
|
||||
... shuffle=True,
|
||||
... batch_size=16,
|
||||
... collate_fn=data_collator,
|
||||
... )
|
||||
|
||||
>>> tf_test_set = model.prepare_tf_dataset(
|
||||
... lm_dataset["test"],
|
||||
... shuffle=False,
|
||||
... batch_size=16,
|
||||
... collate_fn=data_collator,
|
||||
... )
|
||||
```
|
||||
|
||||
قم بتهيئة النموذج للتدريب باستخدام [`compile`](https://keras.io/api/models/model_training_apis/#compile-method). لاحظ أن جميع نماذج Transformers لديها دالة خسارة ذات صلة بالمهمة الافتراضية، لذلك لا تحتاج إلى تحديد واحدة ما لم ترغب في ذلك:
|
||||
|
||||
```py
|
||||
>>> import tensorflow as tf
|
||||
|
||||
>>> model.compile(optimizer=optimizer) # لا يوجد حجة للخسارة!
|
||||
```
|
||||
|
||||
يمكن القيام بذلك عن طريق تحديد مكان دفع نموذجك ومجمّع البيانات في [`~transformers.PushToHubCallback`]:
|
||||
|
||||
```py
|
||||
>>> from transformers.keras_callbacks import PushToHubCallback
|
||||
|
||||
>>> callback = PushToHubCallback(
|
||||
... output_dir="my_awesome_eli5_clm-model",
|
||||
... tokenizer=tokenizer,
|
||||
... )
|
||||
```
|
||||
|
||||
أخيراً، أنت جاهز لبدء تدريب نموذجك! قم باستدعاء [`fit`](https://keras.io/api/models/model_training_apis/#fit-method) مع مجموعات بيانات التدريب والتحقق من الصحة، وعدد العصور، والتعليقات الخاصة بك لتدريب النموذج:
|
||||
|
||||
```py
|
||||
>>> model.fit(x=tf_train_set, validation_data=tf_test_set, epochs=3, callbacks=[callback])
|
||||
```
|
||||
|
||||
بمجرد اكتمال التدريب، يتم تحميل نموذجك تلقائيًا إلى Hub حتى يتمكن الجميع من استخدامه!
|
||||
</tf>
|
||||
</frameworkcontent>
|
||||
|
||||
<Tip>
|
||||
|
||||
@ -280,6 +365,8 @@ Perplexity: 49.61
|
||||
[{'generated_text': "Somatic hypermutation allows the immune system to be able to effectively reverse the damage caused by an infection.\n\n\nThe damage caused by an infection is caused by the immune system's ability to perform its own self-correcting tasks."}]
|
||||
```
|
||||
|
||||
<frameworkcontent>
|
||||
<pt>
|
||||
قسم النص وإرجع `input_ids` كتنسورات PyTorch:
|
||||
|
||||
```py
|
||||
@ -305,3 +392,31 @@ Perplexity: 49.61
|
||||
>>> tokenizer.batch_decode(outputs, skip_special_tokens=True)
|
||||
["Somatic hypermutation allows the immune system to react to drugs with the ability to adapt to a different environmental situation. In other words, a system of 'hypermutation' can help the immune system to adapt to a different environmental situation or in some cases even a single life. In contrast, researchers at the University of Massachusetts-Boston have found that 'hypermutation' is much stronger in mice than in humans but can be found in humans, and that it's not completely unknown to the immune system. A study on how the immune system"]
|
||||
```
|
||||
</pt>
|
||||
<tf>
|
||||
قم بتقسيم النص وإرجاع `input_ids` كـ TensorFlow tensors:
|
||||
|
||||
```py
|
||||
>>> from transformers import AutoTokenizer
|
||||
|
||||
>>> tokenizer = AutoTokenizer.from_pretrained("username/my_awesome_eli5_clm-model")
|
||||
>>> inputs = tokenizer(prompt, return_tensors="tf").input_ids
|
||||
```
|
||||
|
||||
استخدم طريقة [`~transformers.generation_tf_utils.TFGenerationMixin.generate`] لإنشاء الملخص. للمزيد من التفاصيل حول استراتيجيات توليد النص المختلفة والبارامترات للتحكم في التوليد، راجع صفحة [استراتيجيات توليد النص](../generation_strategies).
|
||||
|
||||
```py
|
||||
>>> from transformers import TFAutoModelForCausalLM
|
||||
|
||||
>>> model = TFAutoModelForCausalLM.from_pretrained("username/my_awesome_eli5_clm-model")
|
||||
>>> outputs = model.generate(input_ids=inputs, max_new_tokens=100, do_sample=True, top_k=50, top_p=0.95)
|
||||
```
|
||||
|
||||
فك ترميز الرموز المولدة مرة أخرى إلى نص:
|
||||
|
||||
```py
|
||||
>>> tokenizer.batch_decode(outputs, skip_special_tokens=True)
|
||||
['Somatic hypermutation allows the immune system to detect the presence of other viruses as they become more prevalent. Therefore, researchers have identified a high proportion of human viruses. The proportion of virus-associated viruses in our study increases with age. Therefore, we propose a simple algorithm to detect the presence of these new viruses in our samples as a sign of improved immunity. A first study based on this algorithm, which will be published in Science on Friday, aims to show that this finding could translate into the development of a better vaccine that is more effective for']
|
||||
```
|
||||
</tf>
|
||||
</frameworkcontent>
|
||||
@ -176,6 +176,8 @@ pip install transformers datasets evaluate
|
||||
|
||||
الآن، قم بإنشاء دفعة من الأمثلة باستخدام [`DataCollatorForLanguageModeling`]. من الأكثر كفاءة أن تقوم بـ *الحشو الديناميكي* ليصل طولها إلى أطول جملة في الدفعة أثناء التجميع، بدلاً من حشو مجموعة البيانات بأكملها إلى الطول الأقصى.
|
||||
|
||||
<frameworkcontent>
|
||||
<pt>
|
||||
|
||||
استخدم رمز نهاية التسلسل كرمز الحشو وحدد `mlm_probability` لحجب الرموز عشوائياً كل مرة تكرر فيها البيانات:
|
||||
|
||||
@ -185,9 +187,23 @@ pip install transformers datasets evaluate
|
||||
>>> tokenizer.pad_token = tokenizer.eos_token
|
||||
>>> data_collator = DataCollatorForLanguageModeling(tokenizer=tokenizer, mlm_probability=0.15)
|
||||
```
|
||||
</pt>
|
||||
<tf>
|
||||
|
||||
استخدم رمز نهاية التسلسل كرمز الحشو وحدد `mlm_probability` لحجب الرموز عشوائياً كل مرة تكرر فيها البيانات:
|
||||
|
||||
```py
|
||||
>>> from transformers import DataCollatorForLanguageModeling
|
||||
|
||||
>>> data_collator = DataCollatorForLanguageModeling(tokenizer=tokenizer, mlm_probability=0.15, return_tensors="tf")
|
||||
```
|
||||
</tf>
|
||||
</frameworkcontent>
|
||||
|
||||
## التدريب (Train)
|
||||
|
||||
<frameworkcontent>
|
||||
<pt>
|
||||
|
||||
<Tip>
|
||||
|
||||
@ -247,6 +263,75 @@ Perplexity: 8.76
|
||||
```py
|
||||
>>> trainer.push_to_hub()
|
||||
```
|
||||
</pt>
|
||||
<tf>
|
||||
<Tip>
|
||||
|
||||
إذا لم تكن على دراية بتعديل نموذج باستخدام Keras، ألق نظرة على الدليل الأساسي [هنا](../training#train-a-tensorflow-model-with-keras)!
|
||||
|
||||
</Tip>
|
||||
لتعديل نموذج في TensorFlow، ابدأ بإعداد دالة محسن، وجدول معدل التعلم، وبعض معلمات التدريب:
|
||||
|
||||
```py
|
||||
>>> from transformers import create_optimizer, AdamWeightDecay
|
||||
|
||||
>>> optimizer = AdamWeightDecay(learning_rate=2e-5, weight_decay_rate=0.01)
|
||||
```
|
||||
|
||||
ثم يمكنك تحميل DistilRoBERTa باستخدام [`TFAutoModelForMaskedLM`]:
|
||||
|
||||
```py
|
||||
>>> from transformers import TFAutoModelForMaskedLM
|
||||
|
||||
>>> model = TFAutoModelForMaskedLM.from_pretrained("distilbert/distilroberta-base")
|
||||
```
|
||||
|
||||
قم بتحويل مجموعات بياناتك إلى تنسيق `tf.data.Dataset` باستخدام [`~transformers.TFPreTrainedModel.prepare_tf_dataset`]:
|
||||
|
||||
```py
|
||||
>>> tf_train_set = model.prepare_tf_dataset(
|
||||
... lm_dataset["train"],
|
||||
... shuffle=True,
|
||||
... batch_size=16,
|
||||
... collate_fn=data_collator,
|
||||
... )
|
||||
|
||||
>>> tf_test_set = model.prepare_tf_dataset(
|
||||
... lm_dataset["test"],
|
||||
... shuffle=False,
|
||||
... batch_size=16,
|
||||
... collate_fn=data_collator,
|
||||
... )
|
||||
```
|
||||
|
||||
قم بتهيئة النموذج للتدريب باستخدام [`compile`](https://keras.io/api/models/model_training_apis/#compile-method). لاحظ أن نماذج Transformers لديها جميعها دالة خسارة افتراضية ذات صلة بالمهمة، لذلك لا تحتاج إلى تحديد واحدة ما لم تكن تريد ذلك:
|
||||
|
||||
```py
|
||||
>>> import tensorflow as tf
|
||||
|
||||
>>> model.compile(optimizer=optimizer) # لا توجد حجة للخسارة!
|
||||
```
|
||||
|
||||
يمكن القيام بذلك عن طريق تحديد مكان دفع نموذجك ومعالج الرموز في [`~transformers.PushToHubCallback`]:
|
||||
|
||||
```py
|
||||
>>> from transformers.keras_callbacks import PushToHubCallback
|
||||
|
||||
>>> callback = PushToHubCallback(
|
||||
... output_dir="my_awesome_eli5_mlm_model",
|
||||
... tokenizer=tokenizer,
|
||||
... )
|
||||
```
|
||||
|
||||
أخيراً، أنت مستعد لبدء تدريب نموذجك! قم باستدعاء [`fit`](https://keras.io/api/models/model_training_apis/#fit-method) مع مجموعات بيانات التدريب والتحقق، وعدد العصور، والتعليقات الخاصة بك لتعديل النموذج:
|
||||
|
||||
```py
|
||||
>>> model.fit(x=tf_train_set, validation_data=tf_test_set, epochs=3, callbacks=[callback])
|
||||
```
|
||||
|
||||
بمجرد اكتمال التدريب، يتم تحميل نموذجك تلقائياً إلى Hub حتى يتمكن الجميع من استخدامه!
|
||||
</tf>
|
||||
</frameworkcontent>
|
||||
|
||||
<Tip>
|
||||
|
||||
@ -287,6 +372,8 @@ Perplexity: 8.76
|
||||
'sequence': 'The Milky Way is a small galaxy.'}]
|
||||
```
|
||||
|
||||
<frameworkcontent>
|
||||
<pt>
|
||||
قم بتجزئة النص وإرجاع `input_ids` كمتجهات PyTorch. ستحتاج أيضًا إلى تحديد موضع رمز `<mask>`:
|
||||
|
||||
```py
|
||||
@ -318,3 +405,38 @@ The Milky Way is a spiral galaxy.
|
||||
The Milky Way is a massive galaxy.
|
||||
The Milky Way is a small galaxy.
|
||||
```
|
||||
</pt>
|
||||
<tf>
|
||||
قم بتقسيم النص إلى رموز وإرجاع `input_ids` كـ TensorFlow tensors. ستحتاج أيضًا إلى تحديد موضع رمز `<mask>`:
|
||||
|
||||
```py
|
||||
>>> from transformers import AutoTokenizer
|
||||
|
||||
>>> tokenizer = AutoTokenizer.from_pretrained("username/my_awesome_eli5_mlm_model")
|
||||
>>> inputs = tokenizer(text, return_tensors="tf")
|
||||
>>> mask_token_index = tf.where(inputs["input_ids"] == tokenizer.mask_token_id)[0, 1]
|
||||
```
|
||||
|
||||
قم بتمرير المدخلات إلى النموذج وإرجاع `logits` للرمز المقنع:
|
||||
|
||||
```py
|
||||
>>> from transformers import TFAutoModelForMaskedLM
|
||||
|
||||
>>> model = TFAutoModelForMaskedLM.from_pretrained("username/my_awesome_eli5_mlm_model")
|
||||
>>> logits = model(**inputs).logits
|
||||
>>> mask_token_logits = logits[0, mask_token_index, :]
|
||||
```
|
||||
|
||||
ثم قم بإرجاع الرموز الثلاثة المقنعة ذات الاحتمالية الأعلى وطباعتها:
|
||||
|
||||
```py
|
||||
>>> top_3_tokens = tf.math.top_k(mask_token_logits, 3).indices.numpy()
|
||||
|
||||
>>> for token in top_3_tokens:
|
||||
... print(text.replace(tokenizer.mask_token, tokenizer.decode([token])))
|
||||
The Milky Way is a spiral galaxy.
|
||||
The Milky Way is a massive galaxy.
|
||||
The Milky Way is a small galaxy.
|
||||
```
|
||||
</tf>
|
||||
</frameworkcontent>
|
||||
@ -116,6 +116,8 @@ tokenized_swag = swag.map(preprocess_function, batched=True)
|
||||
|
||||
يقوم `DataCollatorForMultipleChoice` بتجميع جميع مدخلات النموذج، ويطبق الحشو، ثم يعيد تجميع النتائج في شكلها الأصلي:
|
||||
|
||||
<frameworkcontent>
|
||||
<pt>
|
||||
|
||||
```py
|
||||
>>> from dataclasses import dataclass
|
||||
@ -156,6 +158,50 @@ tokenized_swag = swag.map(preprocess_function, batched=True)
|
||||
... batch["labels"] = torch.tensor(labels, dtype=torch.int64)
|
||||
... return batch
|
||||
```
|
||||
</pt>
|
||||
<tf>
|
||||
|
||||
```py
|
||||
>>> from dataclasses import dataclass
|
||||
>>> from transformers.tokenization_utils_base import PreTrainedTokenizerBase, PaddingStrategy
|
||||
>>> from typing import Optional, Union
|
||||
>>> import tensorflow as tf
|
||||
|
||||
>>> @dataclass
|
||||
... class DataCollatorForMultipleChoice:
|
||||
... """
|
||||
... Data collator that will dynamically pad the inputs for multiple choice received.
|
||||
... """
|
||||
|
||||
... tokenizer: PreTrainedTokenizerBase
|
||||
... padding: Union[bool, str, PaddingStrategy] = True
|
||||
... max_length: Optional[int] = None
|
||||
... pad_to_multiple_of: Optional[int] = None
|
||||
|
||||
... def __call__(self, features):
|
||||
... label_name = "label" if "label" in features[0].keys() else "labels"
|
||||
... labels = [feature.pop(label_name) for feature in features]
|
||||
... batch_size = len(features)
|
||||
... num_choices = len(features[0]["input_ids"])
|
||||
... flattened_features = [
|
||||
... [{k: v[i] for k, v in feature.items()} for i in range(num_choices)] for feature in features
|
||||
... ]
|
||||
... flattened_features = sum(flattened_features, [])
|
||||
|
||||
... batch = self.tokenizer.pad(
|
||||
... flattened_features,
|
||||
... padding=self.padding,
|
||||
... max_length=self.max_length,
|
||||
... pad_to_multiple_of=self.pad_to_multiple_of,
|
||||
... return_tensors="tf",
|
||||
... )
|
||||
|
||||
... batch = {k: tf.reshape(v, (batch_size, num_choices, -1)) for k, v in batch.items()}
|
||||
... batch["labels"] = tf.convert_to_tensor(labels, dtype=tf.int64)
|
||||
... return batch
|
||||
```
|
||||
</tf>
|
||||
</frameworkcontent>
|
||||
|
||||
## التقييم (Evaluate)
|
||||
|
||||
@ -182,6 +228,8 @@ tokenized_swag = swag.map(preprocess_function, batched=True)
|
||||
|
||||
## التدريب (Train)
|
||||
|
||||
<frameworkcontent>
|
||||
<pt>
|
||||
|
||||
<Tip>
|
||||
|
||||
@ -235,6 +283,93 @@ tokenized_swag = swag.map(preprocess_function, batched=True)
|
||||
```py
|
||||
>>> trainer.push_to_hub()
|
||||
```
|
||||
</pt>
|
||||
<tf>
|
||||
<Tip>
|
||||
|
||||
إذا لم تكن معتادًا على ضبط نموذج باستخدام Keras، فراجع الدرس الأساسي [هنا](../training#train-a-tensorflow-model-with-keras)!
|
||||
|
||||
</Tip>
|
||||
لضبط نموذج في TensorFlow، ابدأ بإعداد دالة مُحسِّن وجدول معدل التعلم وبعض معلمات التدريب:
|
||||
|
||||
```py
|
||||
>>> from transformers import create_optimizer
|
||||
|
||||
>>> batch_size = 16
|
||||
>>> num_train_epochs = 2
|
||||
>>> total_train_steps = (len(tokenized_swag["train"]) // batch_size) * num_train_epochs
|
||||
>>> optimizer, schedule = create_optimizer(init_lr=5e-5, num_warmup_steps=0, num_train_steps=total_train_steps)
|
||||
```
|
||||
|
||||
ثم يمكنك تحميل BERT باستخدام [`TFAutoModelForMultipleChoice`]:
|
||||
|
||||
```py
|
||||
>>> from transformers import TFAutoModelForMultipleChoice
|
||||
|
||||
>>> model = TFAutoModelForMultipleChoice.from_pretrained("google-bert/bert-base-uncased")
|
||||
```
|
||||
|
||||
حوّل مجموعات البيانات الخاصة بك إلى تنسيق `tf.data.Dataset` باستخدام [`~transformers.TFPreTrainedModel.prepare_tf_dataset`]:
|
||||
|
||||
```py
|
||||
>>> data_collator = DataCollatorForMultipleChoice(tokenizer=tokenizer)
|
||||
>>> tf_train_set = model.prepare_tf_dataset(
|
||||
... tokenized_swag["train"],
|
||||
... shuffle=True,
|
||||
... batch_size=batch_size,
|
||||
... collate_fn=data_collator,
|
||||
... )
|
||||
|
||||
>>> tf_validation_set = model.prepare_tf_dataset(
|
||||
... tokenized_swag["validation"],
|
||||
... shuffle=False,
|
||||
... batch_size=batch_size,
|
||||
... collate_fn=data_collator,
|
||||
... )
|
||||
```
|
||||
|
||||
قم بتهيئة النموذج للتدريب باستخدام [`compile`](https://keras.io/api/models/model_training_apis/#compile-method). لاحظ أن جميع نماذج Transformers تحتوي على دالة خسارة مناسبة للمهمة بشكل افتراضي، لذلك لا تحتاج إلى تحديد واحدة ما لم ترغب في ذلك:
|
||||
|
||||
```py
|
||||
>>> model.compile(optimizer=optimizer) # لا توجد وسيطة خسارة!
|
||||
```
|
||||
|
||||
الخطوتان الأخيرتان قبل بدء التدريب هما: حساب دقة التنبؤات، وتوفير طريقة لرفع النموذج إلى Hub. ويمكن تحقيق ذلك باستخدام [استدعاءات Keras](../main_classes/keras_callbacks)
|
||||
|
||||
مرر دالتك `compute_metrics` إلى [`~transformers.KerasMetricCallback`]:
|
||||
|
||||
```py
|
||||
>>> from transformers.keras_callbacks import KerasMetricCallback
|
||||
|
||||
>>> metric_callback = KerasMetricCallback(metric_fn=compute_metrics, eval_dataset=tf_validation_set)
|
||||
```
|
||||
|
||||
حدد مكان دفع نموذجك ومعالجك في [`~transformers.PushToHubCallback`]:
|
||||
|
||||
```py
|
||||
>>> from transformers.keras_callbacks import PushToHubCallback
|
||||
|
||||
>>> push_to_hub_callback = PushToHubCallback(
|
||||
... output_dir="my_awesome_model",
|
||||
... tokenizer=tokenizer,
|
||||
... )
|
||||
```
|
||||
|
||||
ثم قم بتضمين الاستدعاءات معًا:
|
||||
|
||||
```py
|
||||
>>> callbacks = [metric_callback, push_to_hub_callback]
|
||||
```
|
||||
|
||||
أخيرًا، أنت جاهز لبدء تدريب نموذجك! استدعِ[`fit`](https://keras.io/api/models/model_training_apis/#fit-method) مع مجموعات بيانات التدريب والتحقق من الصحة وعدد الحقب والاستدعاءات لضبط النموذج:
|
||||
|
||||
```py
|
||||
>>> model.fit(x=tf_train_set, validation_data=tf_validation_set, epochs=2, callbacks=callbacks)
|
||||
```
|
||||
|
||||
بمجرد اكتمال التدريب، يتم تحميل نموذجك تلقائيًا إلى Hub حتى يتمكن الجميع من استخدامه!
|
||||
</tf>
|
||||
</frameworkcontent>
|
||||
|
||||
<Tip>
|
||||
|
||||
@ -255,6 +390,8 @@ tokenized_swag = swag.map(preprocess_function, batched=True)
|
||||
>>> candidate2 = "The law applies to baguettes."
|
||||
```
|
||||
|
||||
<frameworkcontent>
|
||||
<pt>
|
||||
قم بتحليل كل مطالبة وزوج إجابة مرشح وأعد تنسورات PyTorch. يجب عليك أيضًا إنشاء بعض `العلامات`:
|
||||
|
||||
```py
|
||||
@ -282,3 +419,34 @@ tokenized_swag = swag.map(preprocess_function, batched=True)
|
||||
>>> predicted_class
|
||||
0
|
||||
```
|
||||
</pt>
|
||||
<tf>
|
||||
قم بتحليل كل مطالبة وزوج إجابة مرشح وأعد موترات TensorFlow:
|
||||
|
||||
```py
|
||||
>>> from transformers import AutoTokenizer
|
||||
|
||||
>>> tokenizer = AutoTokenizer.from_pretrained("username/my_awesome_swag_model")
|
||||
>>> inputs = tokenizer([[prompt, candidate1], [prompt, candidate2]], return_tensors="tf", padding=True)
|
||||
```
|
||||
|
||||
مرر مدخلاتك إلى النموذج وأعد القيم logits:
|
||||
|
||||
```py
|
||||
>>> from transformers import TFAutoModelForMultipleChoice
|
||||
|
||||
>>> model = TFAutoModelForMultipleChoice.from_pretrained("username/my_awesome_swag_model")
|
||||
>>> inputs = {k: tf.expand_dims(v, 0) for k, v in inputs.items()}
|
||||
>>> outputs = model(inputs)
|
||||
>>> logits = outputs.logits
|
||||
```
|
||||
|
||||
استخرج الفئة ذات الاحتمالية الأكبر:
|
||||
|
||||
```py
|
||||
>>> predicted_class = int(tf.math.argmax(logits, axis=-1)[0])
|
||||
>>> predicted_class
|
||||
0
|
||||
```
|
||||
</tf>
|
||||
</frameworkcontent>
|
||||
|
||||
@ -167,15 +167,29 @@ pip install transformers datasets evaluate
|
||||
|
||||
الآن قم بإنشاء دفعة من الأمثلة باستخدام [`DefaultDataCollator`]. بخلاف مجمّعات البيانات الأخرى في 🤗 Transformers، لا يطبق [`DefaultDataCollator`] أي معالجة مسبقة إضافية مثل الحشو.
|
||||
|
||||
<frameworkcontent>
|
||||
<pt>
|
||||
|
||||
```py
|
||||
>>> from transformers import DefaultDataCollator
|
||||
|
||||
>>> data_collator = DefaultDataCollator()
|
||||
```
|
||||
</pt>
|
||||
<tf>
|
||||
|
||||
```py
|
||||
>>> from transformers import DefaultDataCollator
|
||||
|
||||
>>> data_collator = DefaultDataCollator(return_tensors="tf")
|
||||
```
|
||||
</tf>
|
||||
</frameworkcontent>
|
||||
|
||||
## التدريب (Train)
|
||||
|
||||
<frameworkcontent>
|
||||
<pt>
|
||||
|
||||
<Tip>
|
||||
|
||||
@ -226,6 +240,82 @@ pip install transformers datasets evaluate
|
||||
```py
|
||||
>>> trainer.push_to_hub()
|
||||
```
|
||||
</pt>
|
||||
<tf>
|
||||
|
||||
<Tip>
|
||||
|
||||
إذا لم تكن معتادًا على ضبط نموذج باستخدام Keras، فألق نظرة على البرنامج التعليمي الأساسي [هنا](../training#train-a-tensorflow-model-with-keras)!
|
||||
|
||||
</Tip>
|
||||
لضبط نموذج في TensorFlow، ابدأ بإعداد دالة مُحسِّن، وجدول معدل التعلم، وبعض المعاملات الفائقة للتدريب:
|
||||
|
||||
```py
|
||||
>>> from transformers import create_optimizer
|
||||
|
||||
>>> batch_size = 16
|
||||
>>> num_epochs = 2
|
||||
>>> total_train_steps = (len(tokenized_squad["train"]) // batch_size) * num_epochs
|
||||
>>> optimizer, schedule = create_optimizer(
|
||||
... init_lr=2e-5,
|
||||
... num_warmup_steps=0,
|
||||
... num_train_steps=total_train_steps,
|
||||
... )
|
||||
```
|
||||
|
||||
ثم يمكنك تحميل DistilBERT باستخدام [`TFAutoModelForQuestionAnswering`]:
|
||||
|
||||
```py
|
||||
>>> from transformers import TFAutoModelForQuestionAnswering
|
||||
|
||||
>>> model = TFAutoModelForQuestionAnswering.from_pretrained("distilbert/distilbert-base-uncased")
|
||||
```
|
||||
|
||||
حوّل مجموعات البيانات الخاصة بك إلى تنسيق `tf.data.Dataset` باستخدام [`~transformers.TFPreTrainedModel.prepare_tf_dataset`]:
|
||||
|
||||
```py
|
||||
>>> tf_train_set = model.prepare_tf_dataset(
|
||||
... tokenized_squad["train"],
|
||||
... shuffle=True,
|
||||
... batch_size=16,
|
||||
... collate_fn=data_collator,
|
||||
... )
|
||||
|
||||
>>> tf_validation_set = model.prepare_tf_dataset(
|
||||
... tokenized_squad["test"],
|
||||
... shuffle=False,
|
||||
... batch_size=16,
|
||||
... collate_fn=data_collator,
|
||||
... )
|
||||
```
|
||||
|
||||
قم بتكوين النموذج للتدريب باستخدام [`compile`](https://keras.io/api/models/model_training_apis/#compile-method):
|
||||
|
||||
```py
|
||||
>>> import tensorflow as tf
|
||||
|
||||
>>> model.compile(optimizer=optimizer)
|
||||
```
|
||||
|
||||
آخر شيء يجب إعداده قبل بدء التدريب هو توفير طريقة لدفع نموذجك إلى Hub. يمكن القيام بذلك عن طريق تحديد مكان دفع نموذجك ومعالجك المعجمي في [`~transformers.PushToHubCallback`]:
|
||||
|
||||
```py
|
||||
>>> from transformers.keras_callbacks import PushToHubCallback
|
||||
|
||||
>>> callback = PushToHubCallback(
|
||||
... output_dir="my_awesome_qa_model",
|
||||
... tokenizer=tokenizer,
|
||||
... )
|
||||
```
|
||||
|
||||
أخيرًا، أنت جاهز لبدء تدريب نموذجك! اتصل بـ [`fit`](https://keras.io/api/models/model_training_apis/#fit-method) مع مجموعات بيانات التدريب والتحقق من الصحة، وعدد العهود، ومعاودة الاتصال الخاصة بك لضبط النموذج:
|
||||
|
||||
```py
|
||||
>>> model.fit(x=tf_train_set, validation_data=tf_validation_set, epochs=3, callbacks=[callback])
|
||||
```
|
||||
بمجرد اكتمال التدريب، يتم تحميل نموذجك تلقائيًا إلى Hub حتى يتمكن الجميع من استخدامه!
|
||||
</tf>
|
||||
</frameworkcontent>
|
||||
|
||||
|
||||
<Tip>
|
||||
@ -267,6 +357,8 @@ pip install transformers datasets evaluate
|
||||
|
||||
يمكنك أيضًا تكرار نتائج `pipeline` يدويًا إذا أردت:
|
||||
|
||||
<frameworkcontent>
|
||||
<pt>
|
||||
|
||||
قسّم النص وأرجع تنسورات PyTorch:
|
||||
|
||||
@ -302,3 +394,39 @@ pip install transformers datasets evaluate
|
||||
>>> tokenizer.decode(predict_answer_tokens)
|
||||
'176 billion parameters and can generate text in 46 languages natural languages and 13'
|
||||
```
|
||||
</pt>
|
||||
<tf>
|
||||
قم بتحليل النص المعجمي وأعد موترات TensorFlow:
|
||||
|
||||
```py
|
||||
>>> from transformers import AutoTokenizer
|
||||
|
||||
>>> tokenizer = AutoTokenizer.from_pretrained("my_awesome_qa_model")
|
||||
>>> inputs = tokenizer(question, context, return_tensors="tf")
|
||||
```
|
||||
|
||||
مرر مدخلاتك إلى النموذج وأعد `logits`:
|
||||
|
||||
```py
|
||||
>>> from transformers import TFAutoModelForQuestionAnswering
|
||||
|
||||
>>> model = TFAutoModelForQuestionAnswering.from_pretrained("my_awesome_qa_model")
|
||||
>>> outputs = model(**inputs)
|
||||
```
|
||||
|
||||
احصل على أعلى احتمال من مخرجات النموذج لموضعي البداية والنهاية:
|
||||
|
||||
```py
|
||||
>>> answer_start_index = int(tf.math.argmax(outputs.start_logits, axis=-1)[0])
|
||||
>>> answer_end_index = int(tf.math.argmax(outputs.end_logits, axis=-1)[0])
|
||||
```
|
||||
|
||||
استخلاص الإجابة من الرموز المتوقعة:
|
||||
|
||||
```py
|
||||
>>> predict_answer_tokens = inputs.input_ids[0, answer_start_index : answer_end_index + 1]
|
||||
>>> tokenizer.decode(predict_answer_tokens)
|
||||
'176 billion parameters and can generate text in 46 languages natural languages and 13'
|
||||
```
|
||||
</tf>
|
||||
</frameworkcontent>
|
||||
|
||||
@ -92,12 +92,24 @@ tokenized_imdb = imdb.map(preprocess_function, batched=True)
|
||||
|
||||
الآن قم بإنشاء دفعة من الأمثلة باستخدام [`DataCollatorWithPadding`]. الأكثر كفاءة هو استخدام الحشو الديناميكي لجعل الجمل متساوية في الطول داخل كل دفعة، بدلًا من حشو كامل البيانات إلى الحد الأقصى للطول.
|
||||
|
||||
<frameworkcontent>
|
||||
<pt>
|
||||
|
||||
```py
|
||||
>>> from transformers import DataCollatorWithPadding
|
||||
|
||||
>>> data_collator = DataCollatorWithPadding(tokenizer=tokenizer)
|
||||
```
|
||||
</pt>
|
||||
<tf>
|
||||
|
||||
```py
|
||||
>>> from transformers import DataCollatorWithPadding
|
||||
|
||||
>>> data_collator = DataCollatorWithPadding(tokenizer=tokenizer, return_tensors="tf")
|
||||
```
|
||||
</tf>
|
||||
</frameworkcontent>
|
||||
|
||||
## التقييم(Evaluate)
|
||||
|
||||
@ -131,6 +143,8 @@ tokenized_imdb = imdb.map(preprocess_function, batched=True)
|
||||
>>> label2id = {"NEGATIVE": 0, "POSITIVE": 1}
|
||||
```
|
||||
|
||||
<frameworkcontent>
|
||||
<pt>
|
||||
<Tip>
|
||||
|
||||
إذا لم تكن على دراية بضبط نموذج دقيق باستخدام [`Trainer`], فالق نظرة على البرنامج التعليمي الأساسي [هنا](../training#train-with-pytorch-trainer)!
|
||||
@ -191,6 +205,98 @@ tokenized_imdb = imdb.map(preprocess_function, batched=True)
|
||||
```py
|
||||
>>> trainer.push_to_hub()
|
||||
```
|
||||
</pt>
|
||||
<tf>
|
||||
<Tip>
|
||||
|
||||
إذا لم تكن على دراية بضبط نموذج باستخدام Keras، قم بالاطلاع على البرنامج التعليمي الأساسي [هنا](../training#train-a-tensorflow-model-with-keras)!
|
||||
|
||||
</Tip>
|
||||
لضبط نموذج في TensorFlow، ابدأ بإعداد دالة المحسن، وجدول معدل التعلم، وبعض معلمات التدريب:
|
||||
|
||||
```py
|
||||
>>> from transformers import create_optimizer
|
||||
>>> import tensorflow as tf
|
||||
|
||||
>>> batch_size = 16
|
||||
>>> num_epochs = 5
|
||||
>>> batches_per_epoch = len(tokenized_imdb["train"]) // batch_size
|
||||
>>> total_train_steps = int(batches_per_epoch * num_epochs)
|
||||
>>> optimizer, schedule = create_optimizer(init_lr=2e-5, num_warmup_steps=0, num_train_steps=total_train_steps)
|
||||
```
|
||||
|
||||
ثم يمكنك تحميل DistilBERT مع [`TFAutoModelForSequenceClassification`] بالإضافة إلى عدد التصنيفات المتوقعة، وتعيينات التسميات:
|
||||
|
||||
```py
|
||||
>>> from transformers import TFAutoModelForSequenceClassification
|
||||
|
||||
>>> model = TFAutoModelForSequenceClassification.from_pretrained(
|
||||
... "distilbert/distilbert-base-uncased", num_labels=2, id2label=id2label, label2id=label2id
|
||||
... )
|
||||
```
|
||||
|
||||
قم بتحويل مجموعات بياناتك إلى تنسيق `tf.data.Dataset` باستخدام [`~transformers.TFPreTrainedModel.prepare_tf_dataset`]:
|
||||
|
||||
```py
|
||||
>>> tf_train_set = model.prepare_tf_dataset(
|
||||
... tokenized_imdb["train"],
|
||||
... shuffle=True,
|
||||
... batch_size=16,
|
||||
... collate_fn=data_collator,
|
||||
... )
|
||||
|
||||
>>> tf_validation_set = model.prepare_tf_dataset(
|
||||
... tokenized_imdb["test"],
|
||||
... shuffle=False,
|
||||
... batch_size=16,
|
||||
... collate_fn=data_collator,
|
||||
... )
|
||||
```
|
||||
|
||||
قم بتهيئة النموذج للتدريب باستخدام [`compile`](https://keras.io/api/models/model_training_apis/#compile-method). لاحظ أن جميع نماذج Transformers لديها دالة خسارة ذات صلة بالمهمة بشكل افتراضي، لذلك لا تحتاج إلى تحديد واحدة ما لم ترغب في ذلك:
|
||||
|
||||
```py
|
||||
>>> import tensorflow as tf
|
||||
|
||||
>>> model.compile(optimizer=optimizer) # No loss argument!
|
||||
```
|
||||
|
||||
آخر أمرين يجب إعدادهما قبل بدء التدريب هو حساب الدقة من التوقعات، وتوفير طريقة لدفع نموذجك إلى Hub. يتم ذلك باستخدام [Keras callbacks](../main_classes/keras_callbacks).
|
||||
|
||||
قم بتمرير دالة `compute_metrics` الخاصة بك إلى [`~transformers.KerasMetricCallback`]:
|
||||
|
||||
```py
|
||||
>>> from transformers.keras_callbacks import KerasMetricCallback
|
||||
|
||||
>>> metric_callback = KerasMetricCallback(metric_fn=compute_metrics, eval_dataset=tf_validation_set)
|
||||
```
|
||||
|
||||
حدد مكان دفع نموذجك والمجزئ اللغوي في [`~transformers.PushToHubCallback`]:
|
||||
|
||||
```py
|
||||
>>> from transformers.keras_callbacks import PushToHubCallback
|
||||
|
||||
>>> push_to_hub_callback = PushToHubCallback(
|
||||
... output_dir="my_awesome_model",
|
||||
... tokenizer=tokenizer,
|
||||
... )
|
||||
```
|
||||
|
||||
ثم اجمع الاستدعاءات معًا:
|
||||
|
||||
```py
|
||||
>>> callbacks = [metric_callback, push_to_hub_callback]
|
||||
```
|
||||
|
||||
أخيرًا، أنت مستعد لبدء تدريب نموذجك! قم باستدعاء [`fit`](https://keras.io/api/models/model_training_apis/#fit-method) مع مجموعات بيانات التدريب والتحقق، وعدد الحقبات، واستدعاءاتك لضبط النموذج:
|
||||
|
||||
```py
|
||||
>>> model.fit(x=tf_train_set, validation_data=tf_validation_set, epochs=3, callbacks=callbacks)
|
||||
```
|
||||
|
||||
بمجرد اكتمال التدريب، يتم تحميل نموذجك تلقائيًا إلى Hub حتى يتمكن الجميع من استخدامه!
|
||||
</tf>
|
||||
</frameworkcontent>
|
||||
|
||||
<Tip>
|
||||
|
||||
@ -222,6 +328,8 @@ tokenized_imdb = imdb.map(preprocess_function, batched=True)
|
||||
|
||||
يمكنك أيضًا تكرار نتائج `pipeline` يدويًا إذا أردت:
|
||||
|
||||
<frameworkcontent>
|
||||
<pt>
|
||||
قم يتجزئة النص وإرجاع تنسورات PyTorch:
|
||||
|
||||
```py
|
||||
@ -248,3 +356,32 @@ tokenized_imdb = imdb.map(preprocess_function, batched=True)
|
||||
>>> model.config.id2label[predicted_class_id]
|
||||
'POSITIVE'
|
||||
```
|
||||
</pt>
|
||||
<tf>
|
||||
قم بتحليل النص وإرجاع تنسيقات TensorFlow:
|
||||
|
||||
```py
|
||||
>>> from transformers import AutoTokenizer
|
||||
|
||||
>>> tokenizer = AutoTokenizer.from_pretrained("stevhliu/my_awesome_model")
|
||||
>>> inputs = tokenizer(text, return_tensors="tf")
|
||||
```
|
||||
|
||||
قم بتمرير مدخلاتك إلى النموذج وإرجاع `logits`:
|
||||
|
||||
```py
|
||||
>>> from transformers import TFAutoModelForSequenceClassification
|
||||
|
||||
>>> model = TFAutoModelForSequenceClassification.from_pretrained("stevhliu/my_awesome_model")
|
||||
>>> logits = model(**inputs).logits
|
||||
```
|
||||
|
||||
استخرج الفئة ذات الاحتمالية الأعلى، واستخدم `id2label` لتحويلها إلى تصنيف نصي:
|
||||
|
||||
```py
|
||||
>>> predicted_class_id = int(tf.math.argmax(logits, axis=-1)[0])
|
||||
>>> model.config.id2label[predicted_class_id]
|
||||
'POSITIVE'
|
||||
```
|
||||
</tf>
|
||||
</frameworkcontent>
|
||||
|
||||
@ -118,12 +118,24 @@ pip install transformers datasets evaluate rouge_score
|
||||
|
||||
الآن قم بإنشاء دفعة من الأمثلة باستخدام [`DataCollatorForSeq2Seq`]. الأكثر كفاءة *الحشو الديناميكي* للجمل إلى أطول طول في دفعة أثناء عملية التجميع، بدلاً من حشو مجموعة البيانات بأكملها إلى الحد الأقصى للطول.
|
||||
|
||||
<frameworkcontent>
|
||||
<pt>
|
||||
|
||||
```py
|
||||
>>> from transformers import DataCollatorForSeq2Seq
|
||||
|
||||
>>> data_collator = DataCollatorForSeq2Seq(tokenizer=tokenizer, model=checkpoint)
|
||||
```
|
||||
</pt>
|
||||
<tf>
|
||||
|
||||
```py
|
||||
>>> from transformers import DataCollatorForSeq2Seq
|
||||
|
||||
>>> data_collator = DataCollatorForSeq2Seq(tokenizer=tokenizer, model=checkpoint, return_tensors="tf")
|
||||
```
|
||||
</tf>
|
||||
</frameworkcontent>
|
||||
|
||||
## التقييم (Evaluate)
|
||||
|
||||
@ -158,6 +170,8 @@ pip install transformers datasets evaluate rouge_score
|
||||
|
||||
## التدريب (Train)
|
||||
|
||||
<frameworkcontent>
|
||||
<pt>
|
||||
|
||||
<Tip>
|
||||
|
||||
@ -212,6 +226,91 @@ pip install transformers datasets evaluate rouge_score
|
||||
```py
|
||||
>>> trainer.push_to_hub()
|
||||
```
|
||||
</pt>
|
||||
<tf>
|
||||
<Tip>
|
||||
|
||||
إذا لم تكن معتادًا على ضبط نموذج باستخدام Keras، فألق نظرة على البرنامج التعليمي الأساسي [هنا](../training#train-a-tensorflow-model-with-keras)!
|
||||
|
||||
</Tip>
|
||||
لضبط نموذج في TensorFlow، ابدأ بإعداد دالة مُحسِّن وجدول معدل التعلم وبعض معلمات التدريب:
|
||||
|
||||
```py
|
||||
>>> from transformers import create_optimizer, AdamWeightDecay
|
||||
|
||||
>>> optimizer = AdamWeightDecay(learning_rate=2e-5, weight_decay_rate=0.01)
|
||||
```
|
||||
|
||||
ثم يمكنك تحميل T5 باستخدام [`TFAutoModelForSeq2SeqLM`]:
|
||||
|
||||
```py
|
||||
>>> from transformers import TFAutoModelForSeq2SeqLM
|
||||
|
||||
>>> model = TFAutoModelForSeq2SeqLM.from_pretrained(checkpoint)
|
||||
```
|
||||
|
||||
حوّل مجموعات البيانات الخاصة بك إلى تنسيق `tf.data.Dataset` باستخدام [`~transformers.TFPreTrainedModel.prepare_tf_dataset`]:
|
||||
|
||||
```py
|
||||
>>> tf_train_set = model.prepare_tf_dataset(
|
||||
... tokenized_billsum["train"],
|
||||
... shuffle=True,
|
||||
... batch_size=16,
|
||||
... collate_fn=data_collator,
|
||||
... )
|
||||
|
||||
>>> tf_test_set = model.prepare_tf_dataset(
|
||||
... tokenized_billsum["test"],
|
||||
... shuffle=False,
|
||||
... batch_size=16,
|
||||
... collate_fn=data_collator,
|
||||
... )
|
||||
```
|
||||
|
||||
قم بتكوين النموذج للتدريب باستخدام [`compile`](https://keras.io/api/models/model_training_apis/#compile-method). لاحظ أن جميع نماذج Transformers لديها دالة خسارة ذات صلة بالمهمة افتراضيًا، لذلك لست بحاجة إلى تحديد واحدة ما لم تكن ترغب في ذلك:
|
||||
|
||||
```py
|
||||
>>> import tensorflow as tf
|
||||
|
||||
>>> model.compile(optimizer=optimizer) # No loss argument!
|
||||
```
|
||||
|
||||
آخر شيئين يجب إعدادهما قبل بدء التدريب هما حساب درجة ROUGE من التنبؤات، وتوفير طريقة لدفع نموذجك إلى Hub. يتم كلاهما باستخدام [استدعاءات Keras](../main_classes/keras_callbacks).
|
||||
|
||||
مرر دالة `compute_metrics` الخاصة بك إلى [`~transformers.KerasMetricCallback`]:
|
||||
|
||||
```py
|
||||
>>> from transformers.keras_callbacks import KerasMetricCallback
|
||||
|
||||
>>> metric_callback = KerasMetricCallback(metric_fn=compute_metrics, eval_dataset=tf_test_set)
|
||||
```
|
||||
|
||||
حدد مكان دفع نموذجك ومُحلِّلك اللغوي في [`~transformers.PushToHubCallback`]:
|
||||
|
||||
```py
|
||||
>>> from transformers.keras_callbacks import PushToHubCallback
|
||||
|
||||
>>> push_to_hub_callback = PushToHubCallback(
|
||||
... output_dir="my_awesome_billsum_model",
|
||||
... tokenizer=tokenizer,
|
||||
... )
|
||||
```
|
||||
|
||||
ثم اجمع استدعاءاتك معًا:
|
||||
|
||||
```py
|
||||
>>> callbacks = [metric_callback, push_to_hub_callback]
|
||||
```
|
||||
|
||||
أخيرًا، أنت جاهز لبدء تدريب نموذجك! اتصل بـ [`fit`](https://keras.io/api/models/model_training_apis/#fit-method) مع مجموعات بيانات التدريب والتحقق من الصحة وعدد الحقب واستدعاءاتك لضبط النموذج:
|
||||
|
||||
```py
|
||||
>>> model.fit(x=tf_train_set, validation_data=tf_test_set, epochs=3, callbacks=callbacks)
|
||||
```
|
||||
|
||||
بمجرد اكتمال التدريب، يتم تحميل نموذجك تلقائيًا إلى Hub حتى يتمكن الجميع من استخدامه!
|
||||
</tf>
|
||||
</frameworkcontent>
|
||||
|
||||
<Tip>
|
||||
|
||||
@ -242,6 +341,8 @@ pip install transformers datasets evaluate rouge_score
|
||||
|
||||
يمكنك أيضًا تكرار نتائج `pipeline` يدويًا إذا أردت:
|
||||
|
||||
<frameworkcontent>
|
||||
<pt>
|
||||
قسم النص وإرجع `input_ids` كتنسورات PyTorch:
|
||||
|
||||
```py
|
||||
@ -266,3 +367,31 @@ pip install transformers datasets evaluate rouge_score
|
||||
>>> tokenizer.decode(outputs[0], skip_special_tokens=True)
|
||||
'the inflation reduction act lowers prescription drug costs, health care costs, and energy costs. it's the most aggressive action on tackling the climate crisis in american history. it will ask the ultra-wealthy and corporations to pay their fair share.'
|
||||
```
|
||||
</pt>
|
||||
<tf>
|
||||
قسم النص وإرجع `input_ids` كتنسورات TensorFlow:
|
||||
|
||||
```py
|
||||
>>> from transformers import AutoTokenizer
|
||||
|
||||
>>> tokenizer = AutoTokenizer.from_pretrained("username/my_awesome_billsum_model")
|
||||
>>> inputs = tokenizer(text, return_tensors="tf").input_ids
|
||||
```
|
||||
|
||||
استخدم طريقة [`~transformers.generation_tf_utils.TFGenerationMixin.generate`] لإنشاء التلخيص. لمزيد من التفاصيل حول استراتيجيات توليد النص المختلفة والمعلمات للتحكم في التوليد، راجع واجهة برمجة تطبيقات [توليد النص](../main_classes/text_generation).
|
||||
|
||||
```py
|
||||
>>> from transformers import TFAutoModelForSeq2SeqLM
|
||||
|
||||
>>> model = TFAutoModelForSeq2SeqLM.from_pretrained("username/my_awesome_billsum_model")
|
||||
>>> outputs = model.generate(inputs, max_new_tokens=100, do_sample=False)
|
||||
```
|
||||
|
||||
فك تشفير معرفات الرموز المولدة مرة أخرى إلى نص:
|
||||
|
||||
```py
|
||||
>>> tokenizer.decode(outputs[0], skip_special_tokens=True)
|
||||
'the inflation reduction act lowers prescription drug costs, health care costs, and energy costs. it's the most aggressive action on tackling the climate crisis in american history. it will ask the ultra-wealthy and corporations to pay their fair share.'
|
||||
```
|
||||
</tf>
|
||||
</frameworkcontent>
|
||||
@ -151,11 +151,22 @@ pip install transformers datasets evaluate seqeval
|
||||
|
||||
الآن قم بإنشاء دفعة من الأمثلة باستخدام [`DataCollatorWithPadding`].من الأفضل استخدام *الحشو الديناميكي* للجمل إلى أطول طول في دفعة أثناء التجميع، بدلاً من حشو مجموعة البيانات بالكامل إلى الطول الأقصى.
|
||||
|
||||
<frameworkcontent>
|
||||
<pt>
|
||||
```py
|
||||
>>> from transformers import DataCollatorForTokenClassification
|
||||
|
||||
>>> data_collator = DataCollatorForTokenClassification(tokenizer=tokenizer)
|
||||
```
|
||||
</pt>
|
||||
<tf>
|
||||
```py
|
||||
>>> from transformers import DataCollatorForTokenClassification
|
||||
|
||||
>>> data_collator = DataCollatorForTokenClassification(tokenizer=tokenizer, return_tensors="tf")
|
||||
```
|
||||
</tf>
|
||||
</frameworkcontent>
|
||||
|
||||
## التقييم(Evaluate)
|
||||
|
||||
@ -235,6 +246,8 @@ pip install transformers datasets evaluate seqeval
|
||||
... }
|
||||
```
|
||||
|
||||
<frameworkcontent>
|
||||
<pt>
|
||||
<Tip>
|
||||
|
||||
إذا لم تكن على دراية بتعديل نموذج باستخدام [`Trainer`], ألق نظرة على الدليل التعليمي الأساسي [هنا](../training#train-with-pytorch-trainer)!
|
||||
@ -289,6 +302,101 @@ pip install transformers datasets evaluate seqeval
|
||||
```py
|
||||
>>> trainer.push_to_hub()
|
||||
```
|
||||
</pt>
|
||||
<tf>
|
||||
<Tip>
|
||||
|
||||
إذا لم تكن على دراية بتعديل نموذج باستخدام Keras، ألق نظرة على الدليل التعليمي الأساسي [هنا](../training#train-a-tensorflow-model-with-keras)!
|
||||
|
||||
</Tip>
|
||||
للتعديل على نموذج في TensorFlow، ابدأ بإعداد دالة محسن، وجدول معدل التعلم، وبعض معلمات التدريب:
|
||||
|
||||
```py
|
||||
>>> from transformers import create_optimizer
|
||||
|
||||
>>> batch_size = 16
|
||||
>>> num_train_epochs = 3
|
||||
>>> num_train_steps = (len(tokenized_wnut["train"]) // batch_size) * num_train_epochs
|
||||
>>> optimizer, lr_schedule = create_optimizer(
|
||||
... init_lr=2e-5,
|
||||
... num_train_steps=num_train_steps,
|
||||
... weight_decay_rate=0.01,
|
||||
... num_warmup_steps=0,
|
||||
... )
|
||||
```
|
||||
|
||||
ثم يمكنك تحميل DistilBERT مع [`TFAutoModelForTokenClassification`] إلى جانب عدد التسميات المتوقعة، وتخطيطات التسميات:
|
||||
|
||||
```py
|
||||
>>> from transformers import TFAutoModelForTokenClassification
|
||||
|
||||
>>> model = TFAutoModelForTokenClassification.from_pretrained(
|
||||
... "distilbert/distilbert-base-uncased", num_labels=13, id2label=id2label, label2id=label2id
|
||||
... )
|
||||
```
|
||||
|
||||
قم بتحويل مجموعات بياناتك إلى تنسيق `tf.data.Dataset` مع [`~transformers.TFPreTrainedModel.prepare_tf_dataset`]:
|
||||
|
||||
```py
|
||||
>>> tf_train_set = model.prepare_tf_dataset(
|
||||
... tokenized_wnut["train"],
|
||||
... shuffle=True,
|
||||
... batch_size=16,
|
||||
... collate_fn=data_collator,
|
||||
... )
|
||||
|
||||
>>> tf_validation_set = model.prepare_tf_dataset(
|
||||
... tokenized_wnut["validation"],
|
||||
... shuffle=False,
|
||||
... batch_size=16,
|
||||
... collate_fn=data_collator,
|
||||
... )
|
||||
```
|
||||
|
||||
هيّئ النموذج للتدريب باستخدام [`compile`](https://keras.io/api/models/model_training_apis/#compile-method). لاحظ أن نماذج Transformers تتضمن دالة خسارة افتراضية مرتبطة بالمهمة، لذلك لا تحتاج إلى تحديد واحدة إلا إذا كنت ترغب في ذلك:
|
||||
|
||||
```py
|
||||
>>> import tensorflow as tf
|
||||
|
||||
>>> model.compile(optimizer=optimizer) # No loss argument!
|
||||
```
|
||||
|
||||
آخر أمرين يجب إعدادهما قبل بدء التدريب هو حساب درجات seqeval من التنبؤات، وتوفير طريقة لدفع نموذجك إلى Hub. يتم ذلك باستخدام [Keras callbacks](../main_classes/keras_callbacks).
|
||||
|
||||
مرر دالة `compute_metrics` الخاصة بك إلى [`~transformers.KerasMetricCallback`]:
|
||||
|
||||
```py
|
||||
>>> from transformers.keras_callbacks import KerasMetricCallback
|
||||
|
||||
>>> metric_callback = KerasMetricCallback(metric_fn=compute_metrics, eval_dataset=tf_validation_set)
|
||||
```
|
||||
|
||||
حدد مكان دفع نموذجك والمحلل اللغوي في [`~transformers.PushToHubCallback`]:
|
||||
|
||||
```py
|
||||
>>> from transformers.keras_callbacks import PushToHubCallback
|
||||
|
||||
>>> push_to_hub_callback = PushToHubCallback(
|
||||
... output_dir="my_awesome_wnut_model",
|
||||
... tokenizer=tokenizer,
|
||||
... )
|
||||
```
|
||||
|
||||
ثم جمّع callbacks الخاصة بك معًا:
|
||||
|
||||
```py
|
||||
>>> callbacks = [metric_callback, push_to_hub_callback]
|
||||
```
|
||||
|
||||
أخيرًا، أنت جاهز الآن لبدء تدريب نموذجك! قم باستدعاء [`fit`](https://keras.io/api/models/model_training_apis/#fit-method) مع بيانات التدريب والتحقق، وعدد الحقبات، وcallbacks لتعديل النموذج:
|
||||
|
||||
```py
|
||||
>>> model.fit(x=tf_train_set, validation_data=tf_validation_set, epochs=3, callbacks=callbacks)
|
||||
```
|
||||
|
||||
بمجرد اكتمال التدريب، يتم تحميل نموذجك تلقائيًا إلى Hub حتى يتمكن الجميع من استخدامه!
|
||||
</tf>
|
||||
</frameworkcontent>
|
||||
|
||||
<Tip>
|
||||
|
||||
@ -349,6 +457,8 @@ pip install transformers datasets evaluate seqeval
|
||||
|
||||
يمكنك أيضًا تكرار نتائج `pipeline` يدويًا إذا أردت:
|
||||
|
||||
<frameworkcontent>
|
||||
<pt>
|
||||
قسّم النص إلى رموز وأرجع المُوتّرات بلغة PyTorch:
|
||||
|
||||
```py
|
||||
@ -392,3 +502,49 @@ pip install transformers datasets evaluate seqeval
|
||||
'O',
|
||||
'O']
|
||||
```
|
||||
</pt>
|
||||
<tf>
|
||||
قسّم النص إلى رموز وأرجع المُوتّرات ب TensorFlow:
|
||||
|
||||
```py
|
||||
>>> from transformers import AutoTokenizer
|
||||
|
||||
>>> tokenizer = AutoTokenizer.from_pretrained("stevhliu/my_awesome_wnut_model")
|
||||
>>> inputs = tokenizer(text, return_tensors="tf")
|
||||
```
|
||||
|
||||
مرر مدخلاتك إلى النموذج واحصل على `logits`:
|
||||
|
||||
```py
|
||||
>>> from transformers import TFAutoModelForTokenClassification
|
||||
|
||||
>>> model = TFAutoModelForTokenClassification.from_pretrained("stevhliu/my_awesome_wnut_model")
|
||||
>>> logits = model(**inputs).logits
|
||||
```
|
||||
|
||||
استخرج الفئة ذات الاحتمالية الأعلى، واستخدم جدول `id2label` الخاصة بالنموذج لتحويلها إلى تسمية نصية:
|
||||
|
||||
```py
|
||||
>>> predicted_token_class_ids = tf.math.argmax(logits, axis=-1)
|
||||
>>> predicted_token_class = [model.config.id2label[t] for t in predicted_token_class_ids[0].numpy().tolist()]
|
||||
>>> predicted_token_class
|
||||
['O',
|
||||
'O',
|
||||
'B-location',
|
||||
'I-location',
|
||||
'B-group',
|
||||
'O',
|
||||
'O',
|
||||
'O',
|
||||
'O',
|
||||
'O',
|
||||
'O',
|
||||
'O',
|
||||
'O',
|
||||
'B-location',
|
||||
'B-location',
|
||||
'O',
|
||||
'O']
|
||||
```
|
||||
</tf>
|
||||
</frameworkcontent>
|
||||
|
||||
@ -113,12 +113,24 @@ pip install transformers datasets evaluate sacrebleu
|
||||
|
||||
الآن أنشئ دفعة من الأمثلة باستخدام [`DataCollatorForSeq2Seq`]. من الأكثر كفاءة *الحشو الديناميكي* للجمل إلى أطول طول في دفعة أثناء التجميع، بدلاً من حشو مجموعة البيانات بأكملها إلى الحد الأقصى للطول.
|
||||
|
||||
<frameworkcontent>
|
||||
<pt>
|
||||
|
||||
```py
|
||||
>>> from transformers import DataCollatorForSeq2Seq
|
||||
|
||||
>>> data_collator = DataCollatorForSeq2Seq(tokenizer=tokenizer, model=checkpoint)
|
||||
```
|
||||
</pt>
|
||||
<tf>
|
||||
|
||||
```py
|
||||
>>> from transformers import DataCollatorForSeq2Seq
|
||||
|
||||
>>> data_collator = DataCollatorForSeq2Seq(tokenizer=tokenizer, model=checkpoint, return_tensors="tf")
|
||||
```
|
||||
</tf>
|
||||
</frameworkcontent>
|
||||
|
||||
## التقييم (Evaluate)
|
||||
|
||||
@ -165,6 +177,8 @@ pip install transformers datasets evaluate sacrebleu
|
||||
|
||||
## التدريب (Train)
|
||||
|
||||
<frameworkcontent>
|
||||
<pt>
|
||||
|
||||
<Tip>
|
||||
|
||||
@ -219,6 +233,91 @@ pip install transformers datasets evaluate sacrebleu
|
||||
```py
|
||||
>>> trainer.push_to_hub()
|
||||
```
|
||||
</pt>
|
||||
<tf>
|
||||
<Tip>
|
||||
|
||||
إذا لم تكن معتادًا على ضبط نموذج باستخدام Keras، فألق نظرة على البرنامج التعليمي الأساسي [هنا](../training#train-a-tensorflow-model-with-keras)!
|
||||
|
||||
</Tip>
|
||||
لضبط نموذج في TensorFlow، ابدأ بإعداد دالة مُحسِّن وجدول معدل تعلم وبعض المعلمات الفائقة للتدريب:
|
||||
|
||||
```py
|
||||
>>> from transformers import AdamWeightDecay
|
||||
|
||||
>>> optimizer = AdamWeightDecay(learning_rate=2e-5, weight_decay_rate=0.01)
|
||||
```
|
||||
|
||||
ثم يمكنك تحميل T5 باستخدام [`TFAutoModelForSeq2SeqLM`]:
|
||||
|
||||
```py
|
||||
>>> from transformers import TFAutoModelForSeq2SeqLM
|
||||
|
||||
>>> model = TFAutoModelForSeq2SeqLM.from_pretrained(checkpoint)
|
||||
```
|
||||
|
||||
حوّل مجموعات البيانات الخاصة بك إلى تنسيق `tf.data.Dataset` باستخدام [`~transformers.TFPreTrainedModel.prepare_tf_dataset`]:
|
||||
|
||||
```py
|
||||
>>> tf_train_set = model.prepare_tf_dataset(
|
||||
... tokenized_books["train"],
|
||||
... shuffle=True,
|
||||
... batch_size=16,
|
||||
... collate_fn=data_collator,
|
||||
... )
|
||||
|
||||
>>> tf_test_set = model.prepare_tf_dataset(
|
||||
... tokenized_books["test"],
|
||||
... shuffle=False,
|
||||
... batch_size=16,
|
||||
... collate_fn=data_collator,
|
||||
... )
|
||||
```
|
||||
|
||||
قم بتكوين النموذج للتدريب باستخدام [`compile`](https://keras.io/api/models/model_training_apis/#compile-method). لاحظ أن جميع نماذج Transformers تحتوي على دالة خسارة ذات صلة بالمهمة بشكل افتراضي، لذلك لا تحتاج إلى تحديد واحدة إلا إذا كنت ترغب في ذلك:
|
||||
|
||||
```py
|
||||
>>> import tensorflow as tf
|
||||
|
||||
>>> model.compile(optimizer=optimizer) # No loss argument!
|
||||
```
|
||||
|
||||
آخر شيئين يجب إعدادهما قبل بدء التدريب هما حساب مقياس SacreBLEU من التوقعات، وتوفير طريقة لدفع نموذجك إلى Hub. يتم كلاهما باستخدام [استدعاءات Keras](../main_classes/keras_callbacks).
|
||||
|
||||
مرر دالة `compute_metrics` الخاصة بك إلى [`~transformers.KerasMetricCallback`]:
|
||||
|
||||
```py
|
||||
>>> from transformers.keras_callbacks import KerasMetricCallback
|
||||
|
||||
>>> metric_callback = KerasMetricCallback(metric_fn=compute_metrics, eval_dataset=tf_test_set)
|
||||
```
|
||||
|
||||
حدد مكان دفع نموذجك ومعالجك اللغوي في [`~transformers.PushToHubCallback`]:
|
||||
|
||||
```py
|
||||
>>> from transformers.keras_callbacks import PushToHubCallback
|
||||
|
||||
>>> push_to_hub_callback = PushToHubCallback(
|
||||
... output_dir="my_awesome_opus_books_model",
|
||||
... tokenizer=tokenizer,
|
||||
... )
|
||||
```
|
||||
|
||||
ثم اجمع استدعاءاتك معًا:
|
||||
|
||||
```py
|
||||
>>> callbacks = [metric_callback, push_to_hub_callback]
|
||||
```
|
||||
|
||||
أخيرًا، أنت جاهز لبدء تدريب نموذجك! اتصل بـ [`fit`](https://keras.io/api/models/model_training_apis/#fit-method) مع مجموعات بيانات التدريب والتحقق من الصحة وعدد الحقب واستدعاءاتك لضبط النموذج:
|
||||
|
||||
```py
|
||||
>>> model.fit(x=tf_train_set, validation_data=tf_test_set, epochs=3, callbacks=callbacks)
|
||||
```
|
||||
|
||||
بمجرد اكتمال التدريب، يتم تحميل نموذجك تلقائيًا إلى Hub حتى يتمكن الجميع من استخدامه!
|
||||
</tf>
|
||||
</frameworkcontent>
|
||||
|
||||
<Tip>
|
||||
|
||||
@ -252,6 +351,8 @@ pip install transformers datasets evaluate sacrebleu
|
||||
|
||||
يمكنك أيضًا تكرار نتائج `pipeline` يدويًا إذا أردت:
|
||||
|
||||
<frameworkcontent>
|
||||
<pt>
|
||||
قم بتحويل النص إلى رموز وإرجاع `input_ids` كموترات PyTorch:
|
||||
|
||||
```py
|
||||
@ -276,3 +377,31 @@ pip install transformers datasets evaluate sacrebleu
|
||||
>>> tokenizer.decode(outputs[0], skip_special_tokens=True)
|
||||
'Les lignées partagent des ressources avec des bactéries enfixant l'azote.'
|
||||
```
|
||||
</pt>
|
||||
<tf>
|
||||
قم بتحويل النص إلى رموز وإرجاع `input_ids` كموترات TensorFlow:
|
||||
|
||||
```py
|
||||
>>> from transformers import AutoTokenizer
|
||||
|
||||
>>> tokenizer = AutoTokenizer.from_pretrained("username/my_awesome_opus_books_model")
|
||||
>>> inputs = tokenizer(text, return_tensors="tf").input_ids
|
||||
```
|
||||
|
||||
استخدم طريقة [`~transformers.generation_tf_utils.TFGenerationMixin.generate`] لإنشاء الترجمة. لمزيد من التفاصيل حول استراتيجيات توليد النصوص المختلفة والمعلمات للتحكم في التوليد، تحقق من واجهة برمجة تطبيقات [توليد النصوص](../main_classes/text_generation).
|
||||
|
||||
```py
|
||||
>>> from transformers import TFAutoModelForSeq2SeqLM
|
||||
|
||||
>>> model = TFAutoModelForSeq2SeqLM.from_pretrained("username/my_awesome_opus_books_model")
|
||||
>>> outputs = model.generate(inputs, max_new_tokens=40, do_sample=True, top_k=30, top_p=0.95)
|
||||
```
|
||||
|
||||
فك تشفير معرفات الرموز المولدة مرة أخرى إلى نص:
|
||||
|
||||
```py
|
||||
>>> tokenizer.decode(outputs[0], skip_special_tokens=True)
|
||||
'Les lugumes partagent les ressources avec des bactéries fixatrices d'azote.'
|
||||
```
|
||||
</tf>
|
||||
</frameworkcontent>
|
||||
40
docs/source/ar/tflite.md
Normal file
40
docs/source/ar/tflite.md
Normal file
@ -0,0 +1,40 @@
|
||||
# التصدير إلى TFLite
|
||||
|
||||
[TensorFlow Lite](https://www.tensorflow.org/lite/guide) هو إطار عمل خفيف الوزن لنشر نماذج التعلم الآلي على الأجهزة المحدودة الموارد، مثل الهواتف المحمولة، والأنظمة المدمجة، وأجهزة إنترنت الأشياء (IoT). تم تصميم TFLite لتشغيل النماذج وتحسينها بكفاءة على هذه الأجهزة ذات الطاقة الحاسوبية والذاكرة واستهلاك الطاقة المحدودة.
|
||||
|
||||
يُمثَّل نموذج TensorFlow Lite بتنسيق محمول فعال خاص يُعرَّف بامتداد الملف `.tflite`.
|
||||
|
||||
🤗 Optimum يقدم وظيفة لتصدير نماذج 🤗 Transformers إلى TFLite من خلال الوحدة النمطية `exporters.tflite`. بالنسبة لقائمة هندسات النماذج المدعومة، يرجى الرجوع إلى [وثائق 🤗 Optimum](https://huggingface.co/docs/optimum/exporters/tflite/overview).
|
||||
|
||||
لتصدير نموذج إلى TFLite، قم بتثبيت متطلبات البرنامج المطلوبة:
|
||||
|
||||
```bash
|
||||
pip install optimum[exporters-tf]
|
||||
```
|
||||
|
||||
للاطلاع على جميع المغامﻻت المتاحة، راجع [وثائق 🤗 Optimum](https://huggingface.co/docs/optimum/main/en/exporters/tflite/usage_guides/export_a_model)، أو عرض المساعدة في سطر الأوامر:
|
||||
|
||||
```bash
|
||||
optimum-cli export tflite --help
|
||||
```
|
||||
|
||||
لتصدير نسخة النموذج ل 🤗 Hub، على سبيل المثال، `google-bert/bert-base-uncased`، قم بتشغيل الأمر التالي:
|
||||
|
||||
```bash
|
||||
optimum-cli export tflite --model google-bert/bert-base-uncased --sequence_length 128 bert_tflite/
|
||||
```
|
||||
|
||||
ستظهر لك السجلات التي تُبيّن التقدم وموقع حفظ ملف `model.tflite` الناتج، كما في المثال التالي:
|
||||
|
||||
```bash
|
||||
Validating TFLite model...
|
||||
-[✓] TFLite model output names match reference model (logits)
|
||||
- Validating TFLite Model output "logits":
|
||||
-[✓] (1, 128, 30522) matches (1, 128, 30522)
|
||||
-[x] values not close enough, max diff: 5.817413330078125e-05 (atol: 1e-05)
|
||||
The TensorFlow Lite export succeeded with the warning: The maximum absolute difference between the output of the reference model and the TFLite exported model is not within the set tolerance 1e-05:
|
||||
- logits: max diff = 5.817413330078125e-05.
|
||||
The exported model was saved at: bert_tflite
|
||||
```
|
||||
|
||||
يُبيّن المثال أعلاه كيفية تصدير نسخة من النموذج ل 🤗 Hub. عند تصدير نموذج محلي، تأكد أولاً من حفظ ملفات أوزان النموذج المجزء اللغوى في نفس المسار (`local_path`). عند استخدام CLI، قم بتمرير `local_path` إلى معامل `model` بدلاً من اسم النسخة على 🤗 Hub.
|
||||
154
docs/source/ar/torchscript.md
Normal file
154
docs/source/ar/torchscript.md
Normal file
@ -0,0 +1,154 @@
|
||||
# التصدير إلى TorchScript
|
||||
|
||||
<Tip>
|
||||
|
||||
هذه هي بداية تجاربنا مع TorchScript ولا زلنا نستكشف قدراته مع نماذج المدخلات المتغيرة الحجم. إنه مجال اهتمامنا وسنعمق تحليلنا في الإصدارات القادمة، مع المزيد من الأمثلة البرمجية، وتنفيذ أكثر مرونة، ومقاييس مقارنة بين الأكواد القائمة على Python مع أكواد TorchScript المُجمّعة.
|
||||
|
||||
</Tip>
|
||||
|
||||
وفقًا لـ [وثائق TorchScript](https://pytorch.org/docs/stable/jit.html):
|
||||
|
||||
> TorchScript هي طريقة لإنشاء نماذج قابلة للتسلسل والتحسين من تعليمات PyTorch البرمجية.
|
||||
|
||||
هناك وحدتان من PyTorch، [JIT and TRACE](https://pytorch.org/docs/stable/jit.html)، تتيحان للمطورين تصدير نماذجهم لإعادة استخدامها في برامج أخرى مثل برامج C++ المُحسّنة للأداء.
|
||||
|
||||
نقدم واجهة تتيح لك تصدير نماذج 🤗 Transformers إلى TorchScript بحيث يمكن إعادة استخدامها في بيئة مختلفة عن برامج Python القائمة إلى PyTorch. هنا نشرح كيفية تصدير نماذجنا واستخدامها باستخدام TorchScript.
|
||||
|
||||
يتطلب تصدير نموذج أمرين:
|
||||
|
||||
- تهيئة مثيل للنموذج باستخدام علامة `torchscript`
|
||||
- تمرير مُدخلات وهمية (dummy inputs) خلال النموذج
|
||||
|
||||
تنطوي هذه الضرورات على عدة أمور يجب على المطورين توخي الحذر بشأنها كما هو مفصل أدناه.
|
||||
|
||||
## علامة TorchScript والأوزان المرتبطة
|
||||
|
||||
علامة `torchscript` ضرورية لأن معظم نماذج اللغة 🤗 Transformers لها أوزان مرتبطة بين طبقة `Embedding` وطبقة `Decoding`. لا يسمح لك TorchScript بتصدير النماذج ذات الأوزان المرتبطة، لذلك من الضروري فصل الأوزان ونسخها مسبقًا.
|
||||
|
||||
النماذج المُهيأة باستخدام علامة `torchscript` لها طبقة `Embedding` وطبقة`Decoding` منفصلتين، مما يعني أنه لا ينبغي تدريبها لاحقًا. سيؤدي التدريب إلى عدم تزامن الطبقتين، مما يؤدي إلى نتائج غير متوقعة.
|
||||
|
||||
هذا لا ينطبق على النماذج التي لا تحتوي على رأس نموذج اللغة، حيث لا تملك أوزانًا مرتبطة. يمكن تصدير هذه النماذج بأمان دون علامة `torchscript`.
|
||||
|
||||
## المدخلات الوهمية والأطوال القياسية
|
||||
|
||||
تُستخدم المُدخلات الوهمية لتمرير أمامي خلال النموذج. أثناء انتشار قيم المُدخلات عبر الطبقات، يتتبع PyTorch العمليات المختلفة التي يتم تنفيذها على كل مصفوفة(tensor). ثم يتم استخدام هذه العمليات المُسجلة بعد ذلك لإنشاء *أثر* النموذج.
|
||||
|
||||
يتم إنشاء التتبع بالنسبة لأبعاد المُدخلات. وبالتالي، فهو مُقيّد بأبعاد المُدخلات الوهمية، ولن يعمل لأي طول تسلسل أو حجم دفعة مختلف. عند المحاولة بحجم مختلف، يتم رفع الخطأ التالي:
|
||||
|
||||
```
|
||||
`The expanded size of the tensor (3) must match the existing size (7) at non-singleton dimension 2`
|
||||
```
|
||||
|
||||
نوصي بتتبع النموذج باستخدام حجم مُدخلات وهمية لا يقل عن أكبر مُدخل سيتم تقديمه للنموذج أثناء الاستدلال. يمكن أن تساعد الحشوة(padding) في ملء القيم المفقودة. ومع ذلك، نظرًا لتتبع النموذج بحجم مُدخل أكبر، ستكون أبعاد المصفوفة ستكون كبيرة أيضًا، مما يؤدي عنه المزيد من الحسابات.
|
||||
|
||||
انتبه إلى إجمالي عدد العمليات المُنفذة على كل مُدخل وتابع الأداء عن كثب عند تصدير نماذج متغيرة طول التسلسل.
|
||||
|
||||
## استخدام TorchScript في Python
|
||||
|
||||
يوضح هذا القسم كيفية حفظ النماذج وتحميلها، بالإضافة إلى كيفية استخدام التتبع للاستدلال.
|
||||
|
||||
### حفظ نموذج
|
||||
|
||||
لتصدير `BertModel` باستخدام TorchScript، قم بتهيئة ـ `BertModel` من فئة `BertConfig` ثم احفظه على القرص تحت اسم الملف `traced_bert.pt`:
|
||||
|
||||
```python
|
||||
from transformers import BertModel, BertTokenizer, BertConfig
|
||||
import torch
|
||||
|
||||
enc = BertTokenizer.from_pretrained("google-bert/bert-base-uncased")
|
||||
|
||||
# Tokenizing input text
|
||||
text = "[CLS] Who was Jim Henson ? [SEP] Jim Henson was a puppeteer [SEP]"
|
||||
tokenized_text = enc.tokenize(text)
|
||||
|
||||
# Masking one of the input tokens
|
||||
masked_index = 8
|
||||
tokenized_text[masked_index] = "[MASK]"
|
||||
indexed_tokens = enc.convert_tokens_to_ids(tokenized_text)
|
||||
segments_ids = [0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1]
|
||||
|
||||
# Creating a dummy input
|
||||
tokens_tensor = torch.tensor([indexed_tokens])
|
||||
segments_tensors = torch.tensor([segments_ids])
|
||||
dummy_input = [tokens_tensor, segments_tensors]
|
||||
|
||||
# Initializing the model with the torchscript flag
|
||||
# Flag set to True even though it is not necessary as this model does not have an LM Head.
|
||||
config = BertConfig(
|
||||
vocab_size_or_config_json_file=32000,
|
||||
hidden_size=768,
|
||||
num_hidden_layers=12,
|
||||
num_attention_heads=12,
|
||||
intermediate_size=3072,
|
||||
torchscript=True,
|
||||
)
|
||||
|
||||
# Instantiating the model
|
||||
model = BertModel(config)
|
||||
|
||||
# The model needs to be in evaluation mode
|
||||
model.eval()
|
||||
|
||||
# If you are instantiating the model with *from_pretrained* you can also easily set the TorchScript flag
|
||||
model = BertModel.from_pretrained("google-bert/bert-base-uncased", torchscript=True)
|
||||
|
||||
# Creating the trace
|
||||
traced_model = torch.jit.trace(model, [tokens_tensor, segments_tensors])
|
||||
torch.jit.save(traced_model, "traced_bert.pt")
|
||||
```
|
||||
|
||||
### تحميل نموذج
|
||||
|
||||
يمكنك الآن تحميل `BertModel` المُحفظ سابقًا، `traced_bert.pt`، من القرص واستخدامه على `dummy_input` المُهيأ سابقًا:
|
||||
|
||||
```python
|
||||
loaded_model = torch.jit.load("traced_bert.pt")
|
||||
loaded_model.eval()
|
||||
|
||||
all_encoder_layers, pooled_output = loaded_model(*dummy_input)
|
||||
```
|
||||
|
||||
### استخدام نموذج مُتتبع للاستدلال
|
||||
|
||||
استخدم النموذج المُتتبع للاستدلال باستخدام أسلوب `__call__` الخاص به:
|
||||
|
||||
```python
|
||||
traced_model(tokens_tensor, segments_tensors)
|
||||
```
|
||||
|
||||
## نشر نماذج Hugging Face TorchScript على AWS باستخدام Neuron SDK
|
||||
|
||||
قدمت AWS عائلة [Amazon EC2 Inf1](https://aws.amazon.com/ec2/instance-types/inf1/) من اﻷجهزة لخفض التكلفة وأداء التعلم الآلي عالي الأداء في البيئة السحابية. تعمل أجهزة Inf1 بواسطة شريحة Inferentia من AWS، وهي مُسرّع أجهزة مُخصص، متخصص في أعباء عمل الاستدلال للتعلم العميق. [AWS Neuron](https://awsdocs-neuron.readthedocs-hosted.com/en/latest/#) هي SDK لـ Inferentia التي تدعم تتبع نماذج المحولات وتحسينها للنشر على Inf1. توفر Neuron SDK ما يلي:
|
||||
|
||||
1. واجهة برمجة تطبيقات سهلة الاستخدام مع تغيير سطر واحد من التعليمات البرمجية لتتبع نموذج TorchScript وتحسينه للاستدلال في البيئة السحابية.
|
||||
2. تحسينات الأداء الجاهزة للاستخدام [تحسين التكلفة والأداء](https://awsdocs-neuron.readthedocs-hosted.com/en/latest/neuron-guide/benchmark/>).
|
||||
3. دعم نماذج Hugging Face المحولات المبنية باستخدام إما [PyTorch](https://awsdocs-neuron.readthedocs-hosted.com/en/latest/src/examples/pytorch/bert_tutorial/tutorial_pretrained_bert.html) أو [TensorFlow](https://awsdocs-neuron.readthedocs-hosted.com/en/latest/src/examples/tensorflow/huggingface_bert/huggingface_bert.html).
|
||||
|
||||
### الآثار المترتبة
|
||||
|
||||
تعمل نماذج المحولات المستندة إلى بنية [BERT (تمثيلات الترميز ثنائية الاتجاه من المحولات)](https://huggingface.co/docs/transformers/main/model_doc/bert) أو متغيراتها مثل [distilBERT](https://huggingface.co/docs/transformers/main/model_doc/distilbert) و [roBERTa](https://huggingface.co/docs/transformers/main/model_doc/roberta) بشكل أفضل على Inf1 للمهام غير التوليدية مثل الإجابة على الأسئلة الاستخراجية، وتصنيف التسلسلات، وتصنيف الرموز (tokens). ومع ذلك، يمكن تكييف مهام توليد النصوص للعمل على Inf1 وفقًا لهذا [برنامج تعليمي AWS Neuron MarianMT](https://awsdocs-neuron.readthedocs-hosted.com/en/latest/src/examples/pytorch/transformers-marianmt.html). يمكن العثور على مزيد من المعلومات حول النماذج التي يمكن تحويلها جاهزة على Inferentia في قسم [ملاءمة بنية النموذج](https://awsdocs-neuron.readthedocs-hosted.com/en/latest/neuron-guide/models/models-inferentia.html#models-inferentia) من وثائق Neuron.
|
||||
|
||||
### التبعيات (Dependencies)
|
||||
|
||||
يتطلب استخدام AWS Neuron لتحويل النماذج [بيئة SDK Neuron](https://awsdocs-neuron.readthedocs-hosted.com/en/latest/neuron-guide/neuron-frameworks/pytorch-neuron/index.html#installation-guide) والتي تأتي مسبقًا على [AMI للتعلم العميق من AWS](https://docs.aws.amazon.com/dlami/latest/devguide/tutorial-inferentia-launching.html).
|
||||
|
||||
### تحويل نموذج لـ AWS Neuron
|
||||
|
||||
قم بتحويل نموذج لـ AWS NEURON باستخدام نفس التعليمات البرمجية من [استخدام TorchScript في Python](torchscript#using-torchscript-in-python) لتتبع `BertModel`. قم باستيراد امتداد إطار عمل `torch.neuron` للوصول إلى مكونات Neuron SDK من خلال واجهة برمجة تطبيقات Python:
|
||||
|
||||
```python
|
||||
from transformers import BertModel, BertTokenizer, BertConfig
|
||||
import torch
|
||||
import torch.neuron
|
||||
```
|
||||
|
||||
كل ما عليك فعله هو تعديل السطر التالي:
|
||||
|
||||
```diff
|
||||
- torch.jit.trace(model, [tokens_tensor, segments_tensors])
|
||||
+ torch.neuron.trace(model, [token_tensor, segments_tensors])
|
||||
```
|
||||
|
||||
يتيح ذلك لـ Neuron SDK تتبع النموذج وتحسينه لمثيلات Inf1.
|
||||
|
||||
لمعرفة المزيد حول ميزات AWS Neuron SDK والأدوات ودروس البرامج التعليمية والتحديثات الأخيرة، يرجى الاطلاع على [وثائق AWS NeuronSDK](https://awsdocs-neuron.readthedocs-hosted.com/en/latest/index.html).
|
||||
@ -611,6 +611,7 @@ accelerate launch \
|
||||
--learning_rate 5e-5 \
|
||||
--num_train_epochs 3 \
|
||||
--output_dir /tmp/$TASK_NAME/ \
|
||||
--overwrite_output_dir
|
||||
```
|
||||
|
||||
يمكنك أيضًا تحديد المعلمات من ملف `config_file.yaml` مباشرة في سطر الأوامر:
|
||||
@ -633,6 +634,7 @@ accelerate launch --num_processes=2 \
|
||||
--learning_rate 5e-5 \
|
||||
--num_train_epochs 3 \
|
||||
--output_dir /tmp/$TASK_NAME/ \
|
||||
--overwrite_output_dir
|
||||
```
|
||||
|
||||
اطلع على برنامج تعليمي [Launching your Accelerate scripts](https://huggingface.co/docs/accelerate/basic_tutorials/launch) لمعرفة المزيد حول `accelerate_launch` والتكوينات المخصصة.
|
||||
|
||||
@ -58,6 +58,8 @@
|
||||
في شريط التنقل الأيمن للقفز إلى الإطار الذي تريده - وإذا كنت تريد إخفاء كل المحتوى لإطار معين،
|
||||
فاستخدم الزر في الركن العلوي الأيمن من كتلة الإطار!
|
||||
|
||||
<frameworkcontent>
|
||||
<pt>
|
||||
<Youtube id="nvBXf7s7vTI"/>
|
||||
|
||||
## التدريب باستخدام PyTorch Trainer
|
||||
@ -137,10 +139,124 @@
|
||||
```py
|
||||
>>> trainer.train()
|
||||
```
|
||||
</pt>
|
||||
<tf>
|
||||
<a id='keras'></a>
|
||||
|
||||
<Youtube id="rnTGBy2ax1c"/>
|
||||
|
||||
## تدريب نموذج TensorFlow باستخدام Keras
|
||||
|
||||
يمكنك أيضًا تدريب نماذج 🤗 Transformers في TensorFlow باستخدام واجهة برمجة تطبيقات Keras!
|
||||
|
||||
### تحميل البيانات لـ Keras
|
||||
|
||||
عندما تريد تدريب نموذج 🤗 Transformers باستخدام واجهة برمجة تطبيقات Keras، فأنت بحاجة إلى تحويل مجموعة البيانات الخاصة بك إلى تنسيق يفهمه
|
||||
Keras. إذا كانت مجموعة البيانات الخاصة بك صغيرة، فيمكنك ببساطة تحويلها بالكامل إلى مصفوفات NumPy وإرسالها إلى Keras.
|
||||
دعونا نجرب ذلك أولاً قبل أن نقوم بأي شيء أكثر تعقيدًا.
|
||||
|
||||
أولاً، قم بتحميل مجموعة بيانات. سنستخدم مجموعة بيانات CoLA من معيار [GLUE benchmark](https://huggingface.co/datasets/glue)،
|
||||
نظرًا لأنه مهمة تصنيف نص ثنائي بسيطة، وسنأخذ فقط قسم التدريب الآن.
|
||||
|
||||
```py
|
||||
from datasets import load_dataset
|
||||
|
||||
dataset = load_dataset("glue"، "cola")
|
||||
dataset = dataset ["train"] # خذ فقط قسم التدريب الآن
|
||||
```
|
||||
|
||||
بعد ذلك، قم بتحميل أداة المُجزّئ اللغوي وقم بترميز البيانات كمصفوفات NumPy. لاحظ أن التصنيفات هي بالفعل قائمة من 0 و 1،
|
||||
لذا يمكننا ببساطة تحويل ذلك مباشرة إلى مصفوفة NumPy بدون ترميز!
|
||||
|
||||
```py
|
||||
from transformers import AutoTokenizer
|
||||
import numpy as np
|
||||
|
||||
tokenizer = AutoTokenizer.from_pretrained("google-bert/bert-base-cased")
|
||||
tokenized_data = tokenizer(dataset["sentence"], return_tensors="np", padding=True)
|
||||
# Tokenizer returns a BatchEncoding, but we convert that to a dict for Keras
|
||||
tokenized_data = dict(tokenized_data)
|
||||
|
||||
labels = np.array(dataset["label"]) # Label is already an array of 0 and 1
|
||||
```
|
||||
|
||||
أخيرًا، قم بتحميل وتجميع وتناسب النموذج. لاحظ أن نماذج Transformers تحتوي جميعها على دالة خسارة ذات صلة بالمهمة بشكل افتراضي، لذا فأنت لست بحاجة إلى تحديد واحدة ما لم ترغب في ذلك:
|
||||
|
||||
```py
|
||||
from transformers import TFAutoModelForSequenceClassification
|
||||
from tensorflow.keras.optimizers import Adam
|
||||
|
||||
# تحميل وتجميع النموذج الخاص بنا
|
||||
model = TFAutoModelForSequenceClassification.from_pretrained("google-bert/bert-base-cased")
|
||||
# معدلات التعلم المنخفضة أفضل غالبًا لضبط النماذج الدقيقة
|
||||
model.compile(optimizer=Adam(3e-5)) # لا توجد دالة خسارة!
|
||||
|
||||
model.fit(tokenized_data, labels)
|
||||
```
|
||||
|
||||
<Tip>
|
||||
|
||||
أنت لست مضطرًا لتمرير دالة خسارة إلى نماذجك عند تجميعها! تختار نماذج Hugging Face تلقائيًا
|
||||
دالة خسارة مناسبة لمهمتها وهندسة نموذجها إذا تُركت هذه الحجة فارغة. يمكنك دائمًا
|
||||
تجاوز ذلك عن طريق تحديد دالة خسارة بنفسك إذا كنت تريد ذلك!
|
||||
|
||||
</Tip>
|
||||
|
||||
يعمل هذا النهج بشكل رائع لمجموعات البيانات الصغيرة، ولكن بالنسبة لمجموعات البيانات الأكبر، فقد تجد أنه يصبح مشكلة. لماذا؟
|
||||
لأن المصفوفة المرمزة والتصنيفات يجب أن يتم تحميلها بالكامل في الذاكرة، ولأن NumPy لا يتعامل مع
|
||||
المصفوفات"غير المنتظمة"، لذا حشو كل عينة إلى طول أطول عينة في مجموعة البيانات بأكملها. سيؤدي ذلك إلى زيادة حجم المصفوفة لديك، وستبطئ الرموز الزائده من عملية التدريب أيضًا!
|
||||
|
||||
### تحميل البيانات كـ tf.data.Dataset
|
||||
|
||||
إذا كنت تريد تجنب إبطاء التدريب، فيمكنك تحميل بياناتك كـ `tf.data.Dataset` بدلاً من ذلك. على الرغم من أنه يمكنك كتابة خط أنابيب `tf.data` الخاص بك إذا كنت تريد، إلا أن لدينا طريقتين مختصرتين للقيام بذلك:
|
||||
- [`~TFPreTrainedModel.prepare_tf_dataset`]: هذه هي الطريقة التي نوصي بها في معظم الحالات. نظرًا لأنه طريقة
|
||||
على نموذجك، فيمكنه فحص النموذج لتحديد الأعمدة القابلة للاستخدام كمدخلات للنموذج تلقائيًا،
|
||||
واستبعاد الأعمدة الأخرى لإنشاء مجموعة بيانات أبسط وأكثر كفاءة.
|
||||
- [`~datasets.Dataset.to_tf_dataset`]: هذه الطريقة أكثر أساسية، وهي مفيدة عندما تريد التحكم بدقة في كيفية
|
||||
إنشاء مجموعة البيانات الخاصة بك، عن طريق تحديد أعمدة `columns` و `label_cols` المحددة التي سيتم تضمينها.
|
||||
|
||||
قبل أن تتمكن من استخدام [`~TFPreTrainedModel.prepare_tf_dataset`]، ستحتاج إلى إضافة مخرجات المُجزئ إلى مجموعة البيانات الخاصة بك كأعمدة، كما هو موضح في
|
||||
عينة التعليمات البرمجية التالية:
|
||||
|
||||
```py
|
||||
def tokenize_dataset (data):
|
||||
# ستتم إضافة مفاتيح القاموس الذي تمت إعادته كأعمدة إلى مجموعة البيانات
|
||||
return tokenizer(data["text"])
|
||||
|
||||
|
||||
dataset = dataset.map(tokenize_dataset)
|
||||
```
|
||||
|
||||
تذكر أن مجموعات بيانات Hugging Face يتم تخزينها على القرص بشكل افتراضي، لذا فلن يؤدي ذلك إلى تضخيم استخدام الذاكرة لديك! بمجرد إضافة الأعمدة، يمكنك بث الدفعات من مجموعة البيانات وإضافة الترميز إلى كل دفعة، مما يقلل بشكل كبير من عدد رموز الترقيم مقارنة بترميز مجموعة البيانات بأكملها.
|
||||
|
||||
|
||||
```py
|
||||
>>> tf_dataset = model.prepare_tf_dataset(dataset["train"], batch_size=16, shuffle=True, tokenizer=tokenizer)
|
||||
```
|
||||
|
||||
لاحظ أنه في عينة التعليمات البرمجية أعلاه، تحتاج إلى تمرير المُجزئ اللغوي إلى `prepare_tf_dataset` حتى تتمكن من حشو الدُفعات بشكل صحيح أثناء تحميلها.
|
||||
إذا كانت جميع العينات في مجموعة البيانات الخاصة بك بنفس الطول ولم يكن الترميز ضروريًا، فيمكنك تخطي هذا المعامل.
|
||||
إذا كنت بحاجة إلى القيام بشيء أكثر تعقيدًا من مجرد ترميز العينات (على سبيل المثال، إفساد الرموز للنمذجة اللغوية المُقنعة)،
|
||||
فيمكنك استخدام معامل `collate_fn` بدلاً من ذلك لتمرير دالة يتم استدعاؤها لتحويل
|
||||
قائمة العينات إلى دفعة وتطبيق أي معالجة مسبقة تريدها. راجع أمثلةنا [examples](https://github.com/huggingface/transformers/tree/main/examples) أو
|
||||
[دفاتر الملاحظات](https://huggingface.co/docs/transformers/notebooks) لرؤية هذا النهج في العمل.
|
||||
|
||||
بمجرد إنشاء `tf.data.Dataset`، يمكنك تجميع النموذج وتناسبه كما هو الحال من قبل:
|
||||
|
||||
```py
|
||||
model.compile(optimizer=Adam(3e-5)) # No loss argument!
|
||||
|
||||
model.fit(tf_dataset)
|
||||
```
|
||||
|
||||
</tf>
|
||||
</frameworkcontent>
|
||||
|
||||
<a id='pytorch_native'></a>
|
||||
## تدريب في PyTorch الأصلي
|
||||
|
||||
<frameworkcontent>
|
||||
<pt>
|
||||
<Youtube id="Dh9CL8fyG80"/>
|
||||
|
||||
[`Trainer`] يهتم بحلقة التدريب ويسمح لك بضبط نموذج في سطر واحد من التعليمات البرمجية. بالنسبة للمستخدمين الذين يفضلون كتابة حلقة التدريب الخاصة بهم، يمكنك أيضًا ضبط نموذج 🤗 Transformers في PyTorch الأصلي.
|
||||
@ -281,6 +397,8 @@ torch.cuda.empty_cache()
|
||||
|
||||
>>> metric.compute()
|
||||
```
|
||||
</pt>
|
||||
</frameworkcontent>
|
||||
|
||||
<a id='additional-resources'></a>
|
||||
|
||||
@ -291,4 +409,4 @@ torch.cuda.empty_cache()
|
||||
- [🤗 أمثلة المحولات](https://github.com/huggingface/transformers/tree/main/examples) تتضمن
|
||||
النصوص البرمجية لتدريب مهام NLP الشائعة في PyTorch وTensorFlow.
|
||||
|
||||
- [🤗 دفاتر ملاحظات المحولات](notebooks) يحتوي على دفاتر ملاحظات مختلفة حول كيفية ضبط نموذج لمهمة محددة في PyTorch وTensorFlow.
|
||||
- [🤗 دفاتر ملاحظات المحولات](notebooks) يحتوي على دفاتر ملاحظات مختلفة حول كيفية ضبط نموذج لمهمة محددة في PyTorch وTensorFlow.
|
||||
@ -53,7 +53,7 @@ Lassen Sie uns daher ein wenig tiefer in das allgemeine Design der Bibliothek ei
|
||||
### Überblick über die Modelle
|
||||
|
||||
Um ein Modell erfolgreich hinzuzufügen, ist es wichtig, die Interaktion zwischen Ihrem Modell und seiner Konfiguration zu verstehen,
|
||||
[`PreTrainedModel`] und [`PreTrainedConfig`]. Als Beispiel werden wir
|
||||
[`PreTrainedModel`] und [`PretrainedConfig`]. Als Beispiel werden wir
|
||||
das Modell, das zu 🤗 Transformers hinzugefügt werden soll, `BrandNewBert` nennen.
|
||||
|
||||
Schauen wir uns das mal an:
|
||||
@ -81,10 +81,10 @@ model.config # model has access to its config
|
||||
```
|
||||
|
||||
Ähnlich wie das Modell erbt die Konfiguration grundlegende Serialisierungs- und Deserialisierungsfunktionalitäten von
|
||||
[`PreTrainedConfig`]. Beachten Sie, dass die Konfiguration und das Modell immer in zwei verschiedene Formate serialisiert werden
|
||||
[`PretrainedConfig`]. Beachten Sie, dass die Konfiguration und das Modell immer in zwei verschiedene Formate serialisiert werden
|
||||
unterschiedliche Formate serialisiert werden - das Modell in eine *pytorch_model.bin* Datei und die Konfiguration in eine *config.json* Datei. Aufruf von
|
||||
[`~PreTrainedModel.save_pretrained`] wird automatisch
|
||||
[`~PreTrainedConfig.save_pretrained`] auf, so dass sowohl das Modell als auch die Konfiguration gespeichert werden.
|
||||
[`~PretrainedConfig.save_pretrained`] auf, so dass sowohl das Modell als auch die Konfiguration gespeichert werden.
|
||||
|
||||
|
||||
### Code-Stil
|
||||
|
||||
@ -81,6 +81,8 @@ Laden Sie einen Prozessor mit [`AutoProcessor.from_pretrained`]:
|
||||
|
||||
## AutoModel
|
||||
|
||||
<frameworkcontent>
|
||||
<pt>
|
||||
Mit den `AutoModelFor`-Klassen können Sie schließlich ein vortrainiertes Modell für eine bestimmte Aufgabe laden (siehe [hier](model_doc/auto) für eine vollständige Liste der verfügbaren Aufgaben). Laden Sie zum Beispiel ein Modell für die Sequenzklassifikation mit [`AutoModelForSequenceClassification.from_pretrained`]:
|
||||
|
||||
```py
|
||||
@ -106,3 +108,24 @@ TensorFlow- und Flax-Checkpoints sind nicht betroffen und können in PyTorch-Arc
|
||||
</Tip>
|
||||
|
||||
Im Allgemeinen empfehlen wir die Verwendung der Klasse "AutoTokenizer" und der Klasse "AutoModelFor", um trainierte Instanzen von Modellen zu laden. Dadurch wird sichergestellt, dass Sie jedes Mal die richtige Architektur laden. Im nächsten [Tutorial] (Vorverarbeitung) erfahren Sie, wie Sie Ihren neu geladenen Tokenizer, Feature Extractor und Prozessor verwenden, um einen Datensatz für die Feinabstimmung vorzuverarbeiten.
|
||||
</pt>
|
||||
<tf>
|
||||
Mit den Klassen `TFAutoModelFor` schließlich können Sie ein vortrainiertes Modell für eine bestimmte Aufgabe laden (siehe [hier](model_doc/auto) für eine vollständige Liste der verfügbaren Aufgaben). Laden Sie zum Beispiel ein Modell für die Sequenzklassifikation mit [`TFAutoModelForSequenceClassification.from_pretrained`]:
|
||||
|
||||
```py
|
||||
>>> from transformers import TFAutoModelForSequenceClassification
|
||||
|
||||
>>> model = TFAutoModelForSequenceClassification.from_pretrained("distilbert/distilbert-base-uncased")
|
||||
```
|
||||
|
||||
Sie können denselben Prüfpunkt problemlos wiederverwenden, um eine Architektur für eine andere Aufgabe zu laden:
|
||||
|
||||
```py
|
||||
>>> from transformers import TFAutoModelForTokenClassification
|
||||
|
||||
>>> model = TFAutoModelForTokenClassification.from_pretrained("distilbert/distilbert-base-uncased")
|
||||
```
|
||||
|
||||
Im Allgemeinen empfehlen wir, die Klasse "AutoTokenizer" und die Klasse "TFAutoModelFor" zu verwenden, um vortrainierte Instanzen von Modellen zu laden. Dadurch wird sichergestellt, dass Sie jedes Mal die richtige Architektur laden. Im nächsten [Tutorial] (Vorverarbeitung) erfahren Sie, wie Sie Ihren neu geladenen Tokenizer, Feature Extractor und Prozessor verwenden, um einen Datensatz für die Feinabstimmung vorzuverarbeiten.
|
||||
</tf>
|
||||
</frameworkcontent>
|
||||
|
||||
@ -78,10 +78,10 @@ Wenn Sie an der grundlegenden Verwendung von LLMs interessiert sind, ist unsere
|
||||
Zunächst müssen Sie das Modell laden.
|
||||
|
||||
```py
|
||||
>>> from transformers import AutoModelForCausalLM, BitsAndBytesConfig
|
||||
>>> from transformers import AutoModelForCausalLM
|
||||
|
||||
>>> model = AutoModelForCausalLM.from_pretrained(
|
||||
... "openlm-research/open_llama_7b", device_map="auto", quantization_config=BitsAndBytesConfig(load_in_4bit=True)
|
||||
... "openlm-research/open_llama_7b", device_map="auto", load_in_4bit=True
|
||||
... )
|
||||
```
|
||||
|
||||
@ -119,12 +119,12 @@ Und das war's! Mit ein paar Zeilen Code können Sie sich die Macht eines LLM zun
|
||||
Es gibt viele [Generierungsstrategien](generation_strategies), und manchmal sind die Standardwerte für Ihren Anwendungsfall vielleicht nicht geeignet. Wenn Ihre Ausgaben nicht mit dem übereinstimmen, was Sie erwarten, haben wir eine Liste der häufigsten Fallstricke erstellt und wie Sie diese vermeiden können.
|
||||
|
||||
```py
|
||||
>>> from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
|
||||
>>> from transformers import AutoModelForCausalLM, AutoTokenizer
|
||||
|
||||
>>> tokenizer = AutoTokenizer.from_pretrained("openlm-research/open_llama_7b")
|
||||
>>> tokenizer.pad_token = tokenizer.eos_token # Llama has no pad token by default
|
||||
>>> model = AutoModelForCausalLM.from_pretrained(
|
||||
... "openlm-research/open_llama_7b", device_map="auto", quantization_config=BitsAndBytesConfig(load_in_4bit=True)
|
||||
... "openlm-research/open_llama_7b", device_map="auto", load_in_4bit=True
|
||||
... )
|
||||
```
|
||||
|
||||
|
||||
@ -79,15 +79,43 @@ Um sicherzustellen, dass Ihr Modell von jemandem verwendet werden kann, der mit
|
||||
|
||||
Die Konvertierung eines Checkpoints für ein anderes Framework ist einfach. Stellen Sie sicher, dass Sie PyTorch und TensorFlow installiert haben (siehe [hier](installation) für Installationsanweisungen), und finden Sie dann das spezifische Modell für Ihre Aufgabe in dem anderen Framework.
|
||||
|
||||
<frameworkcontent>
|
||||
<pt>
|
||||
Geben Sie `from_tf=True` an, um einen Prüfpunkt von TensorFlow nach PyTorch zu konvertieren:
|
||||
|
||||
```py
|
||||
>>> pt_model = DistilBertForSequenceClassification.from_pretrained("path/to/awesome-name-you-picked", from_tf=True)
|
||||
>>> pt_model.save_pretrained("path/to/awesome-name-you-picked")
|
||||
```
|
||||
</pt>
|
||||
<tf>
|
||||
Geben Sie `from_pt=True` an, um einen Prüfpunkt von PyTorch nach TensorFlow zu konvertieren:
|
||||
|
||||
```py
|
||||
>>> tf_model = TFDistilBertForSequenceClassification.from_pretrained("path/to/awesome-name-you-picked", from_pt=True)
|
||||
```
|
||||
|
||||
Dann können Sie Ihr neues TensorFlow-Modell mit seinem neuen Checkpoint speichern:
|
||||
|
||||
```py
|
||||
>>> tf_model.save_pretrained("path/to/awesome-name-you-picked")
|
||||
```
|
||||
</tf>
|
||||
<jax>
|
||||
Wenn ein Modell in Flax verfügbar ist, können Sie auch einen Kontrollpunkt von PyTorch nach Flax konvertieren:
|
||||
|
||||
```py
|
||||
>>> flax_model = FlaxDistilBertForSequenceClassification.from_pretrained(
|
||||
... "path/to/awesome-name-you-picked", from_pt=True
|
||||
... )
|
||||
```
|
||||
</jax>
|
||||
</frameworkcontent>
|
||||
|
||||
## Ein Modell während des Trainings hochladen
|
||||
|
||||
<frameworkcontent>
|
||||
<pt>
|
||||
<Youtube id="Z1-XMy-GNLQ"/>
|
||||
|
||||
Die Weitergabe eines Modells an den Hub ist so einfach wie das Hinzufügen eines zusätzlichen Parameters oder Rückrufs. Erinnern Sie sich an das [Feinabstimmungs-Tutorial](training), in der Klasse [`TrainingArguments`] geben Sie Hyperparameter und zusätzliche Trainingsoptionen an. Eine dieser Trainingsoptionen beinhaltet die Möglichkeit, ein Modell direkt an den Hub zu pushen. Setzen Sie `push_to_hub=True` in Ihrer [`TrainingArguments`]:
|
||||
@ -113,6 +141,29 @@ Nach der Feinabstimmung Ihres Modells rufen Sie [`~transformers.Trainer.push_to_
|
||||
```py
|
||||
>>> trainer.push_to_hub()
|
||||
```
|
||||
</pt>
|
||||
<tf>
|
||||
Geben Sie ein Modell mit [`PushToHubCallback`] an den Hub weiter. In der [`PushToHubCallback`] Funktion, fügen Sie hinzu:
|
||||
|
||||
- Ein Ausgabeverzeichnis für Ihr Modell.
|
||||
- Einen Tokenizer.
|
||||
- Die `hub_model_id`, die Ihr Hub-Benutzername und Modellname ist.
|
||||
|
||||
```py
|
||||
>>> from transformers import PushToHubCallback
|
||||
|
||||
>>> push_to_hub_callback = PushToHubCallback(
|
||||
... output_dir="./your_model_save_path", tokenizer=tokenizer, hub_model_id="your-username/my-awesome-model"
|
||||
... )
|
||||
```
|
||||
|
||||
Fügen Sie den Callback zu [`fit`](https://keras.io/api/models/model_training_apis/) hinzu, und 🤗 Transformers wird das trainierte Modell an den Hub weiterleiten:
|
||||
|
||||
```py
|
||||
>>> model.fit(tf_train_dataset, validation_data=tf_validation_dataset, epochs=3, callbacks=push_to_hub_callback)
|
||||
```
|
||||
</tf>
|
||||
</frameworkcontent>
|
||||
|
||||
## Verwenden Sie die Funktion `push_to_hub`.
|
||||
|
||||
@ -178,4 +229,4 @@ Um sicherzustellen, dass die Benutzer die Fähigkeiten, Grenzen, möglichen Verz
|
||||
* Manuelles Erstellen und Hochladen einer "README.md"-Datei.
|
||||
* Klicken Sie auf die Schaltfläche **Modellkarte bearbeiten** in Ihrem Modell-Repository.
|
||||
|
||||
Werfen Sie einen Blick auf die DistilBert [model card](https://huggingface.co/distilbert/distilbert-base-uncased) als gutes Beispiel für die Art von Informationen, die eine Modellkarte enthalten sollte. Weitere Details über andere Optionen, die Sie in der Datei "README.md" einstellen können, wie z.B. den Kohlenstoff-Fußabdruck eines Modells oder Beispiele für Widgets, finden Sie in der Dokumentation [hier](https://huggingface.co/docs/hub/models-cards).
|
||||
Werfen Sie einen Blick auf die DistilBert [model card](https://huggingface.co/distilbert/distilbert-base-uncased) als gutes Beispiel für die Art von Informationen, die eine Modellkarte enthalten sollte. Weitere Details über andere Optionen, die Sie in der Datei "README.md" einstellen können, wie z.B. den Kohlenstoff-Fußabdruck eines Modells oder Beispiele für Widgets, finden Sie in der Dokumentation [hier](https://huggingface.co/docs/hub/models-cards).
|
||||
@ -153,6 +153,8 @@ Schließlich möchten Sie, dass der Tokenizer die tatsächlichen Tensoren zurüc
|
||||
|
||||
Setzen Sie den Parameter `return_tensors` entweder auf `pt` für PyTorch, oder `tf` für TensorFlow:
|
||||
|
||||
<frameworkcontent>
|
||||
<pt>
|
||||
|
||||
```py
|
||||
>>> batch_sentences = [
|
||||
@ -172,6 +174,32 @@ Setzen Sie den Parameter `return_tensors` entweder auf `pt` für PyTorch, oder `
|
||||
[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
|
||||
[1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0]])}
|
||||
```
|
||||
</pt>
|
||||
<tf>
|
||||
```py
|
||||
>>> batch_sentences = [
|
||||
... "But what about second breakfast?",
|
||||
... "Don't think he knows about second breakfast, Pip.",
|
||||
... "What about elevensies?",
|
||||
... ]
|
||||
>>> encoded_input = tokenizer(batch_sentences, padding=True, truncation=True, return_tensors="tf")
|
||||
>>> print(encoded_input)
|
||||
{'input_ids': <tf.Tensor: shape=(2, 9), dtype=int32, numpy=
|
||||
array([[101, 1252, 1184, 1164, 1248, 6462, 136, 102, 0, 0, 0, 0, 0, 0, 0],
|
||||
[101, 1790, 112, 189, 1341, 1119, 3520, 1164, 1248, 6462, 117, 21902, 1643, 119, 102],
|
||||
[101, 1327, 1164, 5450, 23434, 136, 102, 0, 0, 0, 0, 0, 0, 0, 0]],
|
||||
dtype=int32)>,
|
||||
'token_type_ids': <tf.Tensor: shape=(2, 9), dtype=int32, numpy=
|
||||
array([[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
|
||||
[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
|
||||
[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]], dtype=int32)>,
|
||||
'attention_mask': <tf.Tensor: shape=(2, 9), dtype=int32, numpy=
|
||||
array([[1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0],
|
||||
[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
|
||||
[1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0]], dtype=int32)>}
|
||||
```
|
||||
</tf>
|
||||
</frameworkcontent>
|
||||
|
||||
## Audio
|
||||
|
||||
|
||||
@ -66,10 +66,20 @@ Im folgenden Beispiel werden Sie die [`pipeline`] für die Stimmungsanalyse verw
|
||||
|
||||
Installieren Sie die folgenden Abhängigkeiten, falls Sie dies nicht bereits getan haben:
|
||||
|
||||
<frameworkcontent>
|
||||
<pt>
|
||||
|
||||
```bash
|
||||
pip install torch
|
||||
```
|
||||
</pt>
|
||||
<tf>
|
||||
|
||||
```bash
|
||||
pip install tensorflow
|
||||
```
|
||||
</tf>
|
||||
</frameworkcontent>
|
||||
|
||||
Importieren sie die [`pipeline`] und spezifizieren sie die Aufgabe, welche sie lösen möchten:
|
||||
|
||||
@ -144,6 +154,8 @@ Die [`pipeline`] kann jedes Modell aus dem [Model Hub](https://huggingface.co/mo
|
||||
>>> model_name = "nlptown/bert-base-multilingual-uncased-sentiment"
|
||||
```
|
||||
|
||||
<frameworkcontent>
|
||||
<pt>
|
||||
Use the [`AutoModelForSequenceClassification`] and [`AutoTokenizer`] to load the pretrained model and its associated tokenizer (more on an `AutoClass` below):
|
||||
|
||||
```py
|
||||
@ -152,6 +164,18 @@ Use the [`AutoModelForSequenceClassification`] and [`AutoTokenizer`] to load the
|
||||
>>> model = AutoModelForSequenceClassification.from_pretrained(model_name)
|
||||
>>> tokenizer = AutoTokenizer.from_pretrained(model_name)
|
||||
```
|
||||
</pt>
|
||||
<tf>
|
||||
Use the [`TFAutoModelForSequenceClassification`] and [`AutoTokenizer`] to load the pretrained model and its associated tokenizer (more on an `TFAutoClass` below):
|
||||
|
||||
```py
|
||||
>>> from transformers import AutoTokenizer, TFAutoModelForSequenceClassification
|
||||
|
||||
>>> model = TFAutoModelForSequenceClassification.from_pretrained(model_name)
|
||||
>>> tokenizer = AutoTokenizer.from_pretrained(model_name)
|
||||
```
|
||||
</tf>
|
||||
</frameworkcontent>
|
||||
|
||||
Dann können Sie das Modell und den Tokenizer in der [`pipeline`] angeben und den `Klassifikator` auf Ihren Zieltext anwenden:
|
||||
|
||||
@ -202,6 +226,8 @@ Der Tokenizer gibt ein Wörterbuch zurück, das Folgendes enthält:
|
||||
|
||||
Genau wie die [`pipeline`] akzeptiert der Tokenizer eine Liste von Eingaben. Darüber hinaus kann der Tokenizer den Text auch auffüllen und kürzen, um einen Stapel mit einheitlicher Länge zurückzugeben:
|
||||
|
||||
<frameworkcontent>
|
||||
<pt>
|
||||
|
||||
```py
|
||||
>>> pt_batch = tokenizer(
|
||||
@ -212,11 +238,27 @@ Genau wie die [`pipeline`] akzeptiert der Tokenizer eine Liste von Eingaben. Dar
|
||||
... return_tensors="pt",
|
||||
... )
|
||||
```
|
||||
</pt>
|
||||
<tf>
|
||||
|
||||
```py
|
||||
>>> tf_batch = tokenizer(
|
||||
... ["We are very happy to show you the 🤗 Transformers library.", "We hope you don't hate it."],
|
||||
... padding=True,
|
||||
... truncation=True,
|
||||
... max_length=512,
|
||||
... return_tensors="tf",
|
||||
... )
|
||||
```
|
||||
</tf>
|
||||
</frameworkcontent>
|
||||
|
||||
Lesen Sie das Tutorial [preprocessing](./preprocessing) für weitere Details zur Tokenisierung.
|
||||
|
||||
### AutoModel
|
||||
|
||||
<frameworkcontent>
|
||||
<pt>
|
||||
🤗 Transformers bietet eine einfache und einheitliche Möglichkeit, vortrainierte Instanzen zu laden. Das bedeutet, dass Sie ein [`AutoModel`] laden können, wie Sie einen [`AutoTokenizer`] laden würden. Der einzige Unterschied ist die Auswahl des richtigen [`AutoModel`] für die Aufgabe. Da Sie eine Text- oder Sequenzklassifizierung vornehmen, laden Sie [`AutoModelForSequenceClassification`]:
|
||||
|
||||
```py
|
||||
@ -248,6 +290,39 @@ Das Modell gibt die endgültigen Aktivierungen in dem Attribut "logits" aus. Wen
|
||||
tensor([[0.0021, 0.0018, 0.0115, 0.2121, 0.7725],
|
||||
[0.2084, 0.1826, 0.1969, 0.1755, 0.2365]], grad_fn=<SoftmaxBackward0>)
|
||||
```
|
||||
</pt>
|
||||
<tf>
|
||||
🤗 Transformers bietet eine einfache und einheitliche Methode zum Laden von vortrainierten Instanzen. Das bedeutet, dass Sie ein [`TFAutoModel`] genauso laden können, wie Sie einen [`AutoTokenizer`] laden würden. Der einzige Unterschied ist die Auswahl des richtigen [`TFAutoModel`] für die Aufgabe. Da Sie Text - oder Sequenz - Klassifizierung machen, laden Sie [`TFAutoModelForSequenceClassification`]:
|
||||
|
||||
```py
|
||||
>>> from transformers import TFAutoModelForSequenceClassification
|
||||
|
||||
>>> model_name = "nlptown/bert-base-multilingual-uncased-sentiment"
|
||||
>>> tf_model = TFAutoModelForSequenceClassification.from_pretrained(model_name)
|
||||
```
|
||||
|
||||
<Tip>
|
||||
|
||||
In der [Aufgabenzusammenfassung](./task_summary) steht, welche [AutoModel]-Klasse für welche Aufgabe zu verwenden ist.
|
||||
|
||||
</Tip>
|
||||
|
||||
Jetzt können Sie Ihren vorverarbeiteten Stapel von Eingaben direkt an das Modell übergeben, indem Sie die Wörterbuchschlüssel direkt an die Tensoren übergeben:
|
||||
|
||||
```py
|
||||
>>> tf_outputs = tf_model(tf_batch)
|
||||
```
|
||||
|
||||
Das Modell gibt die endgültigen Aktivierungen in dem Attribut "logits" aus. Wenden Sie die Softmax-Funktion auf die "logits" an, um die Wahrscheinlichkeiten zu erhalten:
|
||||
|
||||
```py
|
||||
>>> import tensorflow as tf
|
||||
|
||||
>>> tf_predictions = tf.nn.softmax(tf_outputs.logits, axis=-1)
|
||||
>>> tf_predictions # doctest: +IGNORE_RESULT
|
||||
```
|
||||
</tf>
|
||||
</frameworkcontent>
|
||||
|
||||
<Tip>
|
||||
|
||||
@ -267,6 +342,8 @@ Die Modellausgänge verhalten sich auch wie ein Tupel oder ein Wörterbuch (z.B.
|
||||
|
||||
### Modell speichern
|
||||
|
||||
<frameworkcontent>
|
||||
<pt>
|
||||
Sobald Ihr Modell feinabgestimmt ist, können Sie es mit seinem Tokenizer speichern, indem Sie [`PreTrainedModel.save_pretrained`] verwenden:
|
||||
|
||||
```py
|
||||
@ -280,9 +357,28 @@ Wenn Sie bereit sind, das Modell erneut zu verwenden, laden Sie es mit [`PreTrai
|
||||
```py
|
||||
>>> pt_model = AutoModelForSequenceClassification.from_pretrained("./pt_save_pretrained")
|
||||
```
|
||||
</pt>
|
||||
<tf>
|
||||
Sobald Ihr Modell feinabgestimmt ist, können Sie es mit seinem Tokenizer unter Verwendung von [`TFPreTrainedModel.save_pretrained`] speichern:
|
||||
|
||||
```py
|
||||
>>> tf_save_directory = "./tf_save_pretrained"
|
||||
>>> tokenizer.save_pretrained(tf_save_directory) # doctest: +IGNORE_RESULT
|
||||
>>> tf_model.save_pretrained(tf_save_directory)
|
||||
```
|
||||
|
||||
Wenn Sie bereit sind, das Modell wieder zu verwenden, laden Sie es mit [`TFPreTrainedModel.from_pretrained`]:
|
||||
|
||||
```py
|
||||
>>> tf_model = TFAutoModelForSequenceClassification.from_pretrained("./tf_save_pretrained")
|
||||
```
|
||||
</tf>
|
||||
</frameworkcontent>
|
||||
|
||||
Ein besonders cooles 🤗 Transformers-Feature ist die Möglichkeit, ein Modell zu speichern und es entweder als PyTorch- oder TensorFlow-Modell wieder zu laden. Der Parameter "from_pt" oder "from_tf" kann das Modell von einem Framework in das andere konvertieren:
|
||||
|
||||
<frameworkcontent>
|
||||
<pt>
|
||||
|
||||
```py
|
||||
>>> from transformers import AutoModel
|
||||
@ -290,6 +386,17 @@ Ein besonders cooles 🤗 Transformers-Feature ist die Möglichkeit, ein Modell
|
||||
>>> tokenizer = AutoTokenizer.from_pretrained(pt_save_directory)
|
||||
>>> pt_model = AutoModelForSequenceClassification.from_pretrained(pt_save_directory, from_pt=True)
|
||||
```
|
||||
</pt>
|
||||
<tf>
|
||||
|
||||
```py
|
||||
>>> from transformers import TFAutoModel
|
||||
|
||||
>>> tokenizer = AutoTokenizer.from_pretrained(tf_save_directory)
|
||||
>>> tf_model = TFAutoModelForSequenceClassification.from_pretrained(tf_save_directory, from_tf=True)
|
||||
```
|
||||
</tf>
|
||||
</frameworkcontent>
|
||||
|
||||
## Custom model builds
|
||||
|
||||
@ -303,6 +410,8 @@ Beginnen Sie mit dem Import von [`AutoConfig`] und laden Sie dann das trainierte
|
||||
>>> my_config = AutoConfig.from_pretrained("distilbert/distilbert-base-uncased", n_heads=12)
|
||||
```
|
||||
|
||||
<frameworkcontent>
|
||||
<pt>
|
||||
Create a model from your custom configuration with [`AutoModel.from_config`]:
|
||||
|
||||
```py
|
||||
@ -310,6 +419,17 @@ Create a model from your custom configuration with [`AutoModel.from_config`]:
|
||||
|
||||
>>> my_model = AutoModel.from_config(my_config)
|
||||
```
|
||||
</pt>
|
||||
<tf>
|
||||
Create a model from your custom configuration with [`TFAutoModel.from_config`]:
|
||||
|
||||
```py
|
||||
>>> from transformers import TFAutoModel
|
||||
|
||||
>>> my_model = TFAutoModel.from_config(my_config)
|
||||
```
|
||||
</tf>
|
||||
</frameworkcontent>
|
||||
|
||||
Weitere Informationen zur Erstellung von benutzerdefinierten Konfigurationen finden Sie in der Anleitung [Erstellen einer benutzerdefinierten Architektur](./create_a_model).
|
||||
|
||||
|
||||
@ -85,6 +85,8 @@ pip install -r requirements.txt
|
||||
|
||||
## Ein Skript ausführen
|
||||
|
||||
<frameworkcontent>
|
||||
<pt>
|
||||
Das Beispielskript lädt einen Datensatz aus der 🤗 [Datasets](https://huggingface.co/docs/datasets/) Bibliothek herunter und verarbeitet ihn vor. Dann nimmt das Skript eine Feinabstimmung eines Datensatzes mit dem [Trainer](https://huggingface.co/docs/transformers/main_classes/trainer) auf einer Architektur vor, die eine Zusammenfassung unterstützt. Das folgende Beispiel zeigt, wie die Feinabstimmung von [T5-small](https://huggingface.co/google-t5/t5-small) auf dem Datensatz [CNN/DailyMail](https://huggingface.co/datasets/cnn_dailymail) durchgeführt wird. Das T5-Modell benötigt aufgrund der Art und Weise, wie es trainiert wurde, ein zusätzliches Argument `source_prefix`. Mit dieser Eingabeaufforderung weiß T5, dass es sich um eine Zusammenfassungsaufgabe handelt.
|
||||
|
||||
```bash
|
||||
@ -98,8 +100,27 @@ python examples/pytorch/summarization/run_summarization.py \
|
||||
--output_dir /tmp/tst-summarization \
|
||||
--per_device_train_batch_size=4 \
|
||||
--per_device_eval_batch_size=4 \
|
||||
--overwrite_output_dir \
|
||||
--predict_with_generate
|
||||
```
|
||||
</pt>
|
||||
<tf>
|
||||
Das Beispielskript lädt einen Datensatz aus der 🤗 [Datasets](https://huggingface.co/docs/datasets/) Bibliothek herunter und verarbeitet ihn vor. Anschließend nimmt das Skript die Feinabstimmung eines Datensatzes mit Keras auf einer Architektur vor, die die Zusammenfassung unterstützt. Das folgende Beispiel zeigt, wie die Feinabstimmung von [T5-small](https://huggingface.co/google-t5/t5-small) auf dem [CNN/DailyMail](https://huggingface.co/datasets/cnn_dailymail) Datensatz durchgeführt wird. Das T5-Modell benötigt aufgrund der Art und Weise, wie es trainiert wurde, ein zusätzliches Argument `source_prefix`. Mit dieser Eingabeaufforderung weiß T5, dass es sich um eine Zusammenfassungsaufgabe handelt.
|
||||
|
||||
```bash
|
||||
python examples/tensorflow/summarization/run_summarization.py \
|
||||
--model_name_or_path google-t5/t5-small \
|
||||
--dataset_name cnn_dailymail \
|
||||
--dataset_config "3.0.0" \
|
||||
--output_dir /tmp/tst-summarization \
|
||||
--per_device_train_batch_size 8 \
|
||||
--per_device_eval_batch_size 16 \
|
||||
--num_train_epochs 3 \
|
||||
--do_train \
|
||||
--do_eval
|
||||
```
|
||||
</tf>
|
||||
</frameworkcontent>
|
||||
|
||||
## Verteiltes Training und gemischte Präzision
|
||||
|
||||
@ -121,6 +142,7 @@ torchrun \
|
||||
--output_dir /tmp/tst-summarization \
|
||||
--per_device_train_batch_size=4 \
|
||||
--per_device_eval_batch_size=4 \
|
||||
--overwrite_output_dir \
|
||||
--predict_with_generate
|
||||
```
|
||||
|
||||
@ -128,6 +150,8 @@ TensorFlow-Skripte verwenden eine [`MirroredStrategy`](https://www.tensorflow.or
|
||||
|
||||
## Ein Skript auf einer TPU ausführen
|
||||
|
||||
<frameworkcontent>
|
||||
<pt>
|
||||
Tensor Processing Units (TPUs) sind speziell für die Beschleunigung der Leistung konzipiert. PyTorch unterstützt TPUs mit dem [XLA](https://www.tensorflow.org/xla) Deep Learning Compiler (siehe [hier](https://github.com/pytorch/xla/blob/master/README.md) für weitere Details). Um eine TPU zu verwenden, starten Sie das Skript `xla_spawn.py` und verwenden das Argument `num_cores`, um die Anzahl der TPU-Kerne festzulegen, die Sie verwenden möchten.
|
||||
|
||||
```bash
|
||||
@ -142,8 +166,28 @@ python xla_spawn.py --num_cores 8 \
|
||||
--output_dir /tmp/tst-summarization \
|
||||
--per_device_train_batch_size=4 \
|
||||
--per_device_eval_batch_size=4 \
|
||||
--overwrite_output_dir \
|
||||
--predict_with_generate
|
||||
```
|
||||
</pt>
|
||||
<tf>
|
||||
Tensor Processing Units (TPUs) sind speziell für die Beschleunigung der Leistung konzipiert. TensorFlow Skripte verwenden eine [`TPUStrategy`](https://www.tensorflow.org/guide/distributed_training#tpustrategy) für das Training auf TPUs. Um eine TPU zu verwenden, übergeben Sie den Namen der TPU-Ressource an das Argument `tpu`.
|
||||
|
||||
```bash
|
||||
python run_summarization.py \
|
||||
--tpu name_of_tpu_resource \
|
||||
--model_name_or_path google-t5/t5-small \
|
||||
--dataset_name cnn_dailymail \
|
||||
--dataset_config "3.0.0" \
|
||||
--output_dir /tmp/tst-summarization \
|
||||
--per_device_train_batch_size 8 \
|
||||
--per_device_eval_batch_size 16 \
|
||||
--num_train_epochs 3 \
|
||||
--do_train \
|
||||
--do_eval
|
||||
```
|
||||
</tf>
|
||||
</frameworkcontent>
|
||||
|
||||
## Führen Sie ein Skript mit 🤗 Accelerate aus.
|
||||
|
||||
@ -198,6 +242,7 @@ python examples/pytorch/summarization/run_summarization.py \
|
||||
--summary_column summary_column_name \
|
||||
--source_prefix "summarize: " \
|
||||
--output_dir /tmp/tst-summarization \
|
||||
--overwrite_output_dir \
|
||||
--per_device_train_batch_size=4 \
|
||||
--per_device_eval_batch_size=4 \
|
||||
--predict_with_generate
|
||||
@ -225,6 +270,7 @@ python examples/pytorch/summarization/run_summarization.py \
|
||||
--output_dir /tmp/tst-summarization \
|
||||
--per_device_train_batch_size=4 \
|
||||
--per_device_eval_batch_size=4 \
|
||||
--overwrite_output_dir \
|
||||
--predict_with_generate
|
||||
```
|
||||
|
||||
@ -238,6 +284,8 @@ examples/pytorch/summarization/run_summarization.py -h
|
||||
|
||||
Eine weitere hilfreiche Option, die Sie aktivieren können, ist die Wiederaufnahme des Trainings von einem früheren Kontrollpunkt aus. Auf diese Weise können Sie im Falle einer Unterbrechung Ihres Trainings dort weitermachen, wo Sie aufgehört haben, ohne von vorne beginnen zu müssen. Es gibt zwei Methoden, um das Training von einem Kontrollpunkt aus wieder aufzunehmen.
|
||||
|
||||
Die erste Methode verwendet das Argument `output_dir previous_output_dir`, um das Training ab dem letzten in `output_dir` gespeicherten Kontrollpunkt wieder aufzunehmen. In diesem Fall sollten Sie `overwrite_output_dir` entfernen:
|
||||
|
||||
```bash
|
||||
python examples/pytorch/summarization/run_summarization.py
|
||||
--model_name_or_path google-t5/t5-small \
|
||||
@ -249,6 +297,24 @@ python examples/pytorch/summarization/run_summarization.py
|
||||
--output_dir /tmp/tst-summarization \
|
||||
--per_device_train_batch_size=4 \
|
||||
--per_device_eval_batch_size=4 \
|
||||
--output_dir previous_output_dir \
|
||||
--predict_with_generate
|
||||
```
|
||||
|
||||
Die zweite Methode verwendet das Argument `Resume_from_checkpoint path_to_specific_checkpoint`, um das Training ab einem bestimmten Checkpoint-Ordner wieder aufzunehmen.
|
||||
|
||||
```bash
|
||||
python examples/pytorch/summarization/run_summarization.py
|
||||
--model_name_or_path google-t5/t5-small \
|
||||
--do_train \
|
||||
--do_eval \
|
||||
--dataset_name cnn_dailymail \
|
||||
--dataset_config "3.0.0" \
|
||||
--source_prefix "summarize: " \
|
||||
--output_dir /tmp/tst-summarization \
|
||||
--per_device_train_batch_size=4 \
|
||||
--per_device_eval_batch_size=4 \
|
||||
--overwrite_output_dir \
|
||||
--resume_from_checkpoint path_to_specific_checkpoint \
|
||||
--predict_with_generate
|
||||
```
|
||||
@ -280,5 +346,6 @@ python examples/pytorch/summarization/run_summarization.py
|
||||
--output_dir /tmp/tst-summarization \
|
||||
--per_device_train_batch_size=4 \
|
||||
--per_device_eval_batch_size=4 \
|
||||
--overwrite_output_dir \
|
||||
--predict_with_generate
|
||||
```
|
||||
```
|
||||
@ -73,6 +73,8 @@ An dieser Stelle sollten Sie dem Abschnitt folgen, der dem Rahmen entspricht, de
|
||||
in der rechten Seitenleiste können Sie zu dem gewünschten Abschnitt springen - und wenn Sie den gesamten Inhalt eines bestimmten Frameworks ausblenden möchten,
|
||||
klicken Sie einfach auf die Schaltfläche oben rechts im Block des jeweiligen Frameworks!
|
||||
|
||||
<frameworkcontent>
|
||||
<pt>
|
||||
<Youtube id="nvBXf7s7vTI"/>
|
||||
|
||||
## Trainieren mit PyTorch Trainer
|
||||
@ -153,11 +155,128 @@ Anschließend können Sie Ihr Modell durch den Aufruf von [`~transformers.Traine
|
||||
```py
|
||||
>>> trainer.train()
|
||||
```
|
||||
</pt>
|
||||
<tf>
|
||||
<a id='keras'></a>
|
||||
|
||||
<Youtube id="rnTGBy2ax1c"/>
|
||||
|
||||
## Trainieren Sie ein TensorFlow-Modell mit Keras
|
||||
|
||||
Sie können auch 🤗 Transformers Modelle in TensorFlow mit der Keras API trainieren!
|
||||
|
||||
### Laden von Daten für Keras
|
||||
|
||||
Wenn Sie ein 🤗 Transformers Modell mit der Keras API trainieren wollen, müssen Sie Ihren Datensatz in ein Format konvertieren, das
|
||||
Keras versteht. Wenn Ihr Datensatz klein ist, können Sie das Ganze einfach in NumPy-Arrays konvertieren und an Keras übergeben.
|
||||
Probieren wir das zuerst aus, bevor wir etwas Komplizierteres tun.
|
||||
|
||||
Laden Sie zunächst ein Dataset. Wir werden den CoLA-Datensatz aus dem [GLUE-Benchmark](https://huggingface.co/datasets/glue) verwenden,
|
||||
da es sich um eine einfache Aufgabe zur Klassifizierung von binärem Text handelt, und nehmen vorerst nur den Trainingssplit.
|
||||
|
||||
```py
|
||||
from datasets import load_dataset
|
||||
|
||||
dataset = load_dataset("glue", "cola")
|
||||
dataset = dataset["train"] # Just take the training split for now
|
||||
```
|
||||
|
||||
Als nächstes laden Sie einen Tokenizer und tokenisieren die Daten als NumPy-Arrays. Beachten Sie, dass die Beschriftungen bereits eine Liste von 0 und 1en sind,
|
||||
Wir können sie also ohne Tokenisierung direkt in ein NumPy-Array konvertieren!
|
||||
|
||||
```py
|
||||
from transformers import AutoTokenizer
|
||||
|
||||
tokenizer = AutoTokenizer.from_pretrained("google-bert/bert-base-cased")
|
||||
tokenized_data = tokenizer(dataset["text"], return_tensors="np", padding=True)
|
||||
# Tokenizer returns a BatchEncoding, but we convert that to a dict for Keras
|
||||
tokenized_data = dict(tokenized_data)
|
||||
|
||||
labels = np.array(dataset["label"]) # Label is already an array of 0 and 1
|
||||
```
|
||||
|
||||
Schließlich laden, [`compile`](https://keras.io/api/models/model_training_apis/#compile-method) und [`fit`](https://keras.io/api/models/model_training_apis/#fit-method) Sie das Modell:
|
||||
|
||||
```py
|
||||
from transformers import TFAutoModelForSequenceClassification
|
||||
from tensorflow.keras.optimizers import Adam
|
||||
|
||||
# Load and compile our model
|
||||
model = TFAutoModelForSequenceClassification.from_pretrained("google-bert/bert-base-cased")
|
||||
# Lower learning rates are often better for fine-tuning transformers
|
||||
model.compile(optimizer=Adam(3e-5))
|
||||
|
||||
model.fit(tokenized_data, labels)
|
||||
```
|
||||
|
||||
<Tip>
|
||||
|
||||
Sie müssen Ihren Modellen kein Verlustargument übergeben, wenn Sie sie `compile()`! Hugging-Face-Modelle wählen automatisch
|
||||
einen Loss, der für ihre Aufgabe und Modellarchitektur geeignet ist, wenn dieses Argument leer gelassen wird. Sie können jederzeit außer Kraft setzen, indem Sie selbst einen Loss angeben, wenn Sie das möchten!
|
||||
|
||||
</Tip>
|
||||
|
||||
Dieser Ansatz eignet sich hervorragend für kleinere Datensätze, aber bei größeren Datensätzen kann er zu einem Problem werden. Warum?
|
||||
Weil das tokenisierte Array und die Beschriftungen vollständig in den Speicher geladen werden müssten, und weil NumPy nicht mit
|
||||
"gezackte" Arrays nicht verarbeiten kann, so dass jedes tokenisierte Sample auf die Länge des längsten Samples im gesamten Datensatz aufgefüllt werden müsste.
|
||||
Datensatzes aufgefüllt werden. Dadurch wird das Array noch größer, und all die aufgefüllten Token verlangsamen auch das Training!
|
||||
|
||||
### Laden von Daten als tf.data.Dataset
|
||||
|
||||
Wenn Sie eine Verlangsamung des Trainings vermeiden wollen, können Sie Ihre Daten stattdessen als `tf.data.Dataset` laden. Sie können zwar Ihre eigene
|
||||
tf.data"-Pipeline schreiben können, wenn Sie wollen, haben wir zwei bequeme Methoden, um dies zu tun:
|
||||
|
||||
- [`~TFPreTrainedModel.prepare_tf_dataset`]: Dies ist die Methode, die wir in den meisten Fällen empfehlen. Da es sich um eine Methode
|
||||
Ihres Modells ist, kann sie das Modell inspizieren, um automatisch herauszufinden, welche Spalten als Modelleingaben verwendet werden können, und
|
||||
verwirft die anderen, um einen einfacheren, leistungsfähigeren Datensatz zu erstellen.
|
||||
- [`~datasets.Dataset.to_tf_dataset`]: Diese Methode ist eher auf niedriger Ebene angesiedelt und ist nützlich, wenn Sie genau kontrollieren wollen, wie
|
||||
Dataset erstellt wird, indem man genau angibt, welche `columns` und `label_cols` einbezogen werden sollen.
|
||||
|
||||
Bevor Sie [`~TFPreTrainedModel.prepare_tf_dataset`] verwenden können, müssen Sie die Tokenizer-Ausgaben als Spalten zu Ihrem Datensatz hinzufügen, wie in
|
||||
dem folgenden Codebeispiel:
|
||||
|
||||
```py
|
||||
def tokenize_dataset(data):
|
||||
# Keys of the returned dictionary will be added to the dataset as columns
|
||||
return tokenizer(data["text"])
|
||||
|
||||
|
||||
dataset = dataset.map(tokenize_dataset)
|
||||
```
|
||||
|
||||
Denken Sie daran, dass Hugging Face-Datensätze standardmäßig auf der Festplatte gespeichert werden, so dass dies nicht zu einem erhöhten Arbeitsspeicherbedarf führen wird! Sobald die
|
||||
Spalten hinzugefügt wurden, können Sie Batches aus dem Datensatz streamen und zu jedem Batch Auffüllungen hinzufügen, was die Anzahl der Auffüllungs-Token im Vergleich zum Auffüllen des gesamten Datensatzes reduziert.
|
||||
|
||||
|
||||
```py
|
||||
>>> tf_dataset = model.prepare_tf_dataset(dataset, batch_size=16, shuffle=True, tokenizer=tokenizer)
|
||||
```
|
||||
|
||||
Beachten Sie, dass Sie im obigen Codebeispiel den Tokenizer an `prepare_tf_dataset` übergeben müssen, damit die Stapel beim Laden korrekt aufgefüllt werden können.
|
||||
Wenn alle Stichproben in Ihrem Datensatz die gleiche Länge haben und kein Auffüllen erforderlich ist, können Sie dieses Argument weglassen.
|
||||
Wenn Sie etwas Komplexeres als nur das Auffüllen von Stichproben benötigen (z. B. das Korrumpieren von Token für die maskierte Sprachmodellierung), können Sie das Argument
|
||||
Modellierung), können Sie stattdessen das Argument `collate_fn` verwenden, um eine Funktion zu übergeben, die aufgerufen wird, um die
|
||||
Liste von Stichproben in einen Stapel umwandelt und alle gewünschten Vorverarbeitungen vornimmt. Siehe unsere
|
||||
[examples](https://github.com/huggingface/transformers/tree/main/examples) oder
|
||||
[notebooks](https://huggingface.co/docs/transformers/notebooks), um diesen Ansatz in Aktion zu sehen.
|
||||
|
||||
Sobald Sie einen `tf.data.Dataset` erstellt haben, können Sie das Modell wie zuvor kompilieren und anpassen:
|
||||
|
||||
```py
|
||||
model.compile(optimizer=Adam(3e-5))
|
||||
|
||||
model.fit(tf_dataset)
|
||||
```
|
||||
|
||||
</tf>
|
||||
</frameworkcontent>
|
||||
|
||||
<a id='pytorch_native'></a>
|
||||
|
||||
## Trainieren in nativem PyTorch
|
||||
|
||||
<frameworkcontent>
|
||||
<pt>
|
||||
<Youtube id="Dh9CL8fyG80"/>
|
||||
|
||||
[`Trainer`] kümmert sich um die Trainingsschleife und ermöglicht die Feinabstimmung eines Modells in einer einzigen Codezeile. Für Benutzer, die es vorziehen, ihre eigene Trainingsschleife zu schreiben, können Sie auch eine Feinabstimmung eines 🤗 Transformers-Modells in nativem PyTorch vornehmen.
|
||||
@ -299,6 +418,8 @@ Genauso wie Sie eine Bewertungsfunktion zu [`Trainer`] hinzugefügt haben, müss
|
||||
|
||||
>>> metric.compute()
|
||||
```
|
||||
</pt>
|
||||
</frameworkcontent>
|
||||
|
||||
<a id='additional-resources'></a>
|
||||
|
||||
@ -309,4 +430,4 @@ Weitere Beispiele für die Feinabstimmung finden Sie unter:
|
||||
- [🤗 Transformers Examples](https://github.com/huggingface/transformers/tree/main/examples) enthält Skripte
|
||||
um gängige NLP-Aufgaben in PyTorch und TensorFlow zu trainieren.
|
||||
|
||||
- [🤗 Transformers Notebooks](notebooks) enthält verschiedene Notebooks zur Feinabstimmung eines Modells für bestimmte Aufgaben in PyTorch und TensorFlow.
|
||||
- [🤗 Transformers Notebooks](notebooks) enthält verschiedene Notebooks zur Feinabstimmung eines Modells für bestimmte Aufgaben in PyTorch und TensorFlow.
|
||||
@ -88,8 +88,6 @@
|
||||
title: Tool use
|
||||
- local: chat_templating_writing
|
||||
title: Writing a chat template
|
||||
- local: chat_response_parsing
|
||||
title: Response parsing
|
||||
title: Chat with models
|
||||
- sections:
|
||||
- local: serving
|
||||
@ -119,8 +117,6 @@
|
||||
title: Tools
|
||||
- local: transformers_as_backend
|
||||
title: Inference server backends
|
||||
- local: continuous_batching
|
||||
title: Continuous Batching
|
||||
title: Inference
|
||||
- isExpanded: false
|
||||
sections:
|
||||
@ -220,17 +216,16 @@
|
||||
- local: quantization/contribute
|
||||
title: Contribute
|
||||
title: Quantization
|
||||
- isExpanded: false
|
||||
sections:
|
||||
- local: kernel_doc/overview
|
||||
title: Kernels in transformers
|
||||
title: Kernels
|
||||
- isExpanded: false
|
||||
sections:
|
||||
- local: serialization
|
||||
title: ONNX
|
||||
- local: tflite
|
||||
title: LiteRT
|
||||
- local: executorch
|
||||
title: ExecuTorch
|
||||
- local: torchscript
|
||||
title: TorchScript
|
||||
title: Export to production
|
||||
- isExpanded: false
|
||||
sections:
|
||||
@ -286,8 +281,6 @@
|
||||
title: Knowledge Distillation for Computer Vision
|
||||
- local: tasks/keypoint_matching
|
||||
title: Keypoint matching
|
||||
- local: tasks/training_vision_backbone
|
||||
title: Training vision models using Backbone API
|
||||
title: Computer vision
|
||||
- sections:
|
||||
- local: tasks/image_captioning
|
||||
@ -345,12 +338,16 @@
|
||||
title: Configuration
|
||||
- local: main_classes/data_collator
|
||||
title: Data Collator
|
||||
- local: main_classes/keras_callbacks
|
||||
title: Keras callbacks
|
||||
- local: main_classes/logging
|
||||
title: Logging
|
||||
- local: main_classes/model
|
||||
title: Models
|
||||
- local: main_classes/text_generation
|
||||
title: Text Generation
|
||||
- local: main_classes/onnx
|
||||
title: ONNX
|
||||
- local: main_classes/optimizer_schedules
|
||||
title: Optimization
|
||||
- local: main_classes/output
|
||||
@ -377,8 +374,6 @@
|
||||
title: Image Processor
|
||||
- local: main_classes/video_processor
|
||||
title: Video Processor
|
||||
- local: main_classes/kernels
|
||||
title: Kernels
|
||||
title: Main Classes
|
||||
- sections:
|
||||
- sections:
|
||||
@ -548,6 +543,8 @@
|
||||
title: Helium
|
||||
- local: model_doc/herbert
|
||||
title: HerBERT
|
||||
- local: model_doc/hgnet_v2
|
||||
title: HGNet-V2
|
||||
- local: model_doc/hunyuan_v1_dense
|
||||
title: HunYuanDenseV1
|
||||
- local: model_doc/hunyuan_v1_moe
|
||||
@ -564,8 +561,6 @@
|
||||
title: LED
|
||||
- local: model_doc/lfm2
|
||||
title: LFM2
|
||||
- local: model_doc/lfm2_moe
|
||||
title: LFM2Moe
|
||||
- local: model_doc/llama
|
||||
title: LLaMA
|
||||
- local: model_doc/llama2
|
||||
@ -1008,8 +1003,6 @@
|
||||
title: AltCLIP
|
||||
- local: model_doc/aria
|
||||
title: Aria
|
||||
- local: model_doc/audioflamingo3
|
||||
title: AudioFlamingo3
|
||||
- local: model_doc/aya_vision
|
||||
title: AyaVision
|
||||
- local: model_doc/blip
|
||||
@ -1030,8 +1023,6 @@
|
||||
title: CLIPSeg
|
||||
- local: model_doc/clvp
|
||||
title: CLVP
|
||||
- local: model_doc/cwm
|
||||
title: Code World Model (CWM)
|
||||
- local: model_doc/cohere2_vision
|
||||
title: Cohere2Vision
|
||||
- local: model_doc/colpali
|
||||
@ -1192,8 +1183,6 @@
|
||||
title: TVP
|
||||
- local: model_doc/udop
|
||||
title: UDOP
|
||||
- local: model_doc/video_llama_3
|
||||
title: VideoLlama3
|
||||
- local: model_doc/video_llava
|
||||
title: VideoLlava
|
||||
- local: model_doc/vilt
|
||||
@ -1259,8 +1248,6 @@
|
||||
title: Importing Utilities
|
||||
- local: internal/time_series_utils
|
||||
title: Utilities for Time Series
|
||||
- local: internal/rope_utils
|
||||
title: Rotary Embeddings Utilities
|
||||
title: Internal helpers
|
||||
- sections:
|
||||
- local: reference/environment_variables
|
||||
|
||||
@ -55,7 +55,6 @@ deepspeed --num_gpus 2 trainer-program.py ...
|
||||
</hfoptions>
|
||||
|
||||
## Order of accelerators
|
||||
|
||||
To select specific accelerators to use and their order, use the environment variable appropriate for your hardware. This is often set on the command line for each run, but can also be added to your `~/.bashrc` or other startup config file.
|
||||
|
||||
For example, if there are 4 accelerators (0, 1, 2, 3) and you only want to run accelerators 0 and 2:
|
||||
|
||||
@ -51,7 +51,7 @@ This section describes how the model and configuration classes interact and the
|
||||
|
||||
### Model and configuration
|
||||
|
||||
All Transformers' models inherit from a base [`PreTrainedModel`] and [`PreTrainedConfig`] class. The configuration is the models blueprint.
|
||||
All Transformers' models inherit from a base [`PreTrainedModel`] and [`PretrainedConfig`] class. The configuration is the models blueprint.
|
||||
|
||||
There is never more than two levels of abstraction for any model to keep the code readable. The example model here, BrandNewLlama, inherits from `BrandNewLlamaPreTrainedModel` and [`PreTrainedModel`]. It is important that a new model only depends on [`PreTrainedModel`] so that it can use the [`~PreTrainedModel.from_pretrained`] and [`~PreTrainedModel.save_pretrained`] methods.
|
||||
|
||||
@ -66,9 +66,9 @@ model = BrandNewLlamaModel.from_pretrained("username/brand_new_llama")
|
||||
model.config
|
||||
```
|
||||
|
||||
[`PreTrainedConfig`] provides the [`~PreTrainedConfig.from_pretrained`] and [`~PreTrainedConfig.save_pretrained`] methods.
|
||||
[`PretrainedConfig`] provides the [`~PretrainedConfig.from_pretrained`] and [`~PretrainedConfig.save_pretrained`] methods.
|
||||
|
||||
When you use [`PreTrainedModel.save_pretrained`], it automatically calls [`PreTrainedConfig.save_pretrained`] so that both the model and configuration are saved together.
|
||||
When you use [`PreTrainedModel.save_pretrained`], it automatically calls [`PretrainedConfig.save_pretrained`] so that both the model and configuration are saved together.
|
||||
|
||||
A model is saved to a `model.safetensors` file and a configuration is saved to a `config.json` file.
|
||||
|
||||
@ -278,7 +278,7 @@ Every Transformers model output should have a precision or error tolerance of *1
|
||||
|
||||
Here are some tips for an efficient debugging environment.
|
||||
|
||||
- To debug intermediate results, it depends on the machine learning framework the original model repository is using. For PyTorch, you should write a script to decompose the original model into smaller sub-components to retrieve the intermediate values.
|
||||
- To debug intermediate results, it depends on the machine learning framework the original model repository is using. For PyTorch, you should write a script to decompose the original model into smaller sub-components to retrieve the intermediate values. For TensorFlow, you may need to use [tf.print](https://www.tensorflow.org/api_docs/python/tf/print). For Flax, make sure the model is *not jitted* during the forward pass (refer to this GitHub [Issue](https://github.com/google/jax/issues/196) for more details).
|
||||
|
||||
- It is faster to debug with a smaller pretrained checkpoint versus a larger checkpoint where the forward pass takes more than 10 seconds. If only large checkpoints are available, create a dummy model with randomly initialized weights and save those weights to compare against the Transformers implementation.
|
||||
|
||||
|
||||
Some files were not shown because too many files have changed in this diff Show More
Reference in New Issue
Block a user