Compare commits

..

2 Commits

Author SHA1 Message Date
b54358e4cf feat: rework widgets 2025-06-10 17:03:21 +02:00
2274ce74a7 feat: add a ContinuousBatchingVisualizer 2025-06-06 19:04:29 +02:00
3390 changed files with 152940 additions and 238020 deletions

View File

@ -184,7 +184,6 @@ jobs:
- run: python utils/check_dummies.py
- run: python utils/check_repo.py
- run: python utils/check_inits.py
- run: python utils/check_pipeline_typing.py
- run: python utils/check_config_docstrings.py
- run: python utils/check_config_attributes.py
- run: python utils/check_doctest_list.py

View File

@ -109,9 +109,7 @@ class CircleCIJob:
self.docker_image[0]["image"] = f"{self.docker_image[0]['image']}:dev"
print(f"Using {self.docker_image} docker image")
if self.install_steps is None:
self.install_steps = ["uv pip install ."]
# Use a custom patched pytest to force exit the process at the end, to avoid `Too long with no output (exceeded 10m0s): context deadline exceeded`
self.install_steps.append("uv pip install git+https://github.com/ydshieh/pytest.git@8.4.1-ydshieh")
self.install_steps = ["uv venv && uv pip install ."]
if self.pytest_options is None:
self.pytest_options = {}
if isinstance(self.tests_to_run, str):
@ -215,7 +213,7 @@ generate_job = CircleCIJob(
docker_image=[{"image": "huggingface/transformers-torch-light"}],
# networkx==3.3 (after #36957) cause some issues
# TODO: remove this once it works directly
install_steps=["uv pip install ."],
install_steps=["uv venv && uv pip install ."],
marker="generate",
parallelism=6,
)
@ -232,6 +230,22 @@ processor_job = CircleCIJob(
parallelism=8,
)
tf_job = CircleCIJob(
"tf",
docker_image=[{"image":"huggingface/transformers-tf-light"}],
parallelism=6,
)
flax_job = CircleCIJob(
"flax",
docker_image=[{"image":"huggingface/transformers-jax-light"}],
parallelism=6,
pytest_num_workers=16,
resource_class="2xlarge",
)
pipelines_torch_job = CircleCIJob(
"pipelines_torch",
additional_env={"RUN_PIPELINE_TESTS": True},
@ -240,6 +254,16 @@ pipelines_torch_job = CircleCIJob(
parallelism=4,
)
pipelines_tf_job = CircleCIJob(
"pipelines_tf",
additional_env={"RUN_PIPELINE_TESTS": True},
docker_image=[{"image":"huggingface/transformers-tf-light"}],
marker="is_pipeline_test",
parallelism=4,
)
custom_tokenizers_job = CircleCIJob(
"custom_tokenizers",
additional_env={"RUN_CUSTOM_TOKENIZERS": True},
@ -252,16 +276,25 @@ examples_torch_job = CircleCIJob(
additional_env={"OMP_NUM_THREADS": 8},
docker_image=[{"image":"huggingface/transformers-examples-torch"}],
# TODO @ArthurZucker remove this once docker is easier to build
install_steps=["uv pip install . && uv pip install -r examples/pytorch/_tests_requirements.txt"],
install_steps=["uv venv && uv pip install . && uv pip install -r examples/pytorch/_tests_requirements.txt"],
pytest_num_workers=4,
)
examples_tensorflow_job = CircleCIJob(
"examples_tensorflow",
additional_env={"OMP_NUM_THREADS": 8},
docker_image=[{"image":"huggingface/transformers-examples-tf"}],
pytest_num_workers=2,
)
hub_job = CircleCIJob(
"hub",
additional_env={"HUGGINGFACE_CO_STAGING": True},
docker_image=[{"image":"huggingface/transformers-torch-light"}],
install_steps=[
'uv pip install .',
'uv venv && uv pip install .',
'git config --global user.email "ci@dummy.com"',
'git config --global user.name "ci"',
],
@ -275,6 +308,7 @@ onnx_job = CircleCIJob(
"onnx",
docker_image=[{"image":"huggingface/transformers-torch-tf-light"}],
install_steps=[
"uv venv",
"uv pip install .[testing,sentencepiece,onnxruntime,vision,rjieba]",
],
pytest_options={"k onnx": None},
@ -304,7 +338,7 @@ non_model_job = CircleCIJob(
docker_image=[{"image": "huggingface/transformers-torch-light"}],
# networkx==3.3 (after #36957) cause some issues
# TODO: remove this once it works directly
install_steps=["uv pip install .[serving]"],
install_steps=["uv venv && uv pip install ."],
marker="not generate",
parallelism=6,
)
@ -322,7 +356,7 @@ doc_test_job = CircleCIJob(
additional_env={"TRANSFORMERS_VERBOSITY": "error", "DATASETS_VERBOSITY": "error", "SKIP_CUDA_DOCTEST": "1"},
install_steps=[
# Add an empty file to keep the test step running correctly even no file is selected to be tested.
"uv pip install .",
"uv venv && pip install .",
"touch dummy.py",
command,
"cat pr_documentation_tests_temp.txt",
@ -334,7 +368,7 @@ doc_test_job = CircleCIJob(
pytest_num_workers=1,
)
REGULAR_TESTS = [torch_job, hub_job, onnx_job, tokenization_job, processor_job, generate_job, non_model_job] # fmt: skip
REGULAR_TESTS = [torch_job, flax_job, hub_job, onnx_job, tokenization_job, processor_job, generate_job, non_model_job] # fmt: skip
EXAMPLES_TESTS = [examples_torch_job]
PIPELINE_TESTS = [pipelines_torch_job]
REPO_UTIL_TESTS = [repo_utils_job]

View File

@ -51,7 +51,7 @@ Library:
- pipelines: @Rocketknight1
- tensorflow: @gante and @Rocketknight1
- tokenizers: @ArthurZucker
- trainer: @zach-huggingface, @SunMarc and @qgallouedec
- trainer: @zach-huggingface and @SunMarc
- chat templates: @Rocketknight1
Integrations:

View File

@ -48,7 +48,7 @@ jobs:
- name: Run database init script
run: |
psql -f benchmark/utils/init_db.sql
psql -f benchmark/init_db.sql
env:
PGDATABASE: metrics
PGHOST: ${{ secrets.TRANSFORMERS_BENCHMARKS_PGHOST }}

View File

@ -2,6 +2,15 @@ name: Build PR Documentation
on:
pull_request:
workflow_call:
inputs:
pr_number:
type: string
required: true
commit_sha:
type: string
required: true
concurrency:
group: ${{ github.workflow }}-${{ github.head_ref || github.run_id }}
@ -9,9 +18,9 @@ concurrency:
jobs:
build:
uses: huggingface/doc-builder/.github/workflows/build_pr_documentation.yml@main
uses: huggingface/doc-builder/.github/workflows/build_pr_documentation.yml@6e2eb04a2604817c97be03786efa494fe3acae90
with:
commit_sha: ${{ github.event.pull_request.head.sha }}
pr_number: ${{ github.event.number }}
commit_sha: ${{ inputs.commit_sha || github.event.pull_request.head.sha }}
pr_number: ${{ inputs.pr_number || github.event.number }}
package: transformers
languages: en

View File

@ -41,7 +41,7 @@ jobs:
check_new_failures:
name: " "
runs-on:
group: aws-g5-4xlarge-cache
group: aws-g4dn-4xlarge-cache
container:
image: ${{ inputs.docker }}
options: --gpus all --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/

View File

@ -1,49 +0,0 @@
name: CI collated reports
on:
workflow_call:
inputs:
job:
required: true
type: string
report_repo_id:
required: true
type: string
machine_type:
required: true
type: string
gpu_name:
description: Name of the GPU used for the job. Its enough that the value contains the name of the GPU, e.g. "noise-h100-more-noise". Case insensitive.
required: true
type: string
jobs:
collated_reports:
name: Collated reports
runs-on: ubuntu-22.04
if: always()
steps:
- uses: actions/checkout@v4
- uses: actions/download-artifact@v4
- name: Collated reports
shell: bash
env:
ACCESS_REPO_INFO_TOKEN: ${{ secrets.ACCESS_REPO_INFO_TOKEN }}
CI_SHA: ${{ github.sha }}
TRANSFORMERS_CI_RESULTS_UPLOAD_TOKEN: ${{ secrets.TRANSFORMERS_CI_RESULTS_UPLOAD_TOKEN }}
run: |
pip install huggingface_hub
python3 utils/collated_reports.py \
--path /transformers/reports/ \
--machine-type ${{ inputs.machine_type }} \
--commit-hash ${{ env.CI_SHA }} \
--job ${{ inputs.job }} \
--report-repo-id ${{ inputs.report_repo_id }} \
--gpu-name ${{ inputs.gpu_name }}
- name: Upload collated reports
uses: actions/upload-artifact@v4
with:
name: collated_reports_${{ env.CI_SHA }}.json
path: collated_reports_${{ env.CI_SHA }}.json

View File

@ -28,10 +28,10 @@ jobs:
matrix:
split_keys: ${{ fromJson(inputs.split_keys) }}
runs-on:
group: aws-g5-4xlarge-cache
group: aws-g4dn-4xlarge-cache
container:
image: huggingface/transformers-all-latest-gpu
options: --gpus all --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
options: --gpus 0 --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
steps:
- name: Update clone
working-directory: /transformers

View File

@ -15,10 +15,10 @@ jobs:
setup:
name: Setup
runs-on:
group: aws-g5-4xlarge-cache
group: aws-g4dn-4xlarge-cache
container:
image: huggingface/transformers-all-latest-gpu
options: --gpus all --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
options: --gpus 0 --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
outputs:
job_splits: ${{ steps.set-matrix.outputs.job_splits }}
split_keys: ${{ steps.set-matrix.outputs.split_keys }}

View File

@ -1,157 +0,0 @@
name: Get PR commit SHA
on:
workflow_call:
inputs:
pr_number:
required: true
type: string
outputs:
PR_HEAD_REPO_FULL_NAME:
description: "The full name of the repository from which the pull request is created"
value: ${{ jobs.get-pr-info.outputs.PR_HEAD_REPO_FULL_NAME }}
PR_BASE_REPO_FULL_NAME:
description: "The full name of the repository to which the pull request is created"
value: ${{ jobs.get-pr-info.outputs.PR_BASE_REPO_FULL_NAME }}
PR_HEAD_REPO_OWNER:
description: "The owner of the repository from which the pull request is created"
value: ${{ jobs.get-pr-info.outputs.PR_HEAD_REPO_OWNER }}
PR_BASE_REPO_OWNER:
description: "The owner of the repository to which the pull request is created"
value: ${{ jobs.get-pr-info.outputs.PR_BASE_REPO_OWNER }}
PR_HEAD_REPO_NAME:
description: "The name of the repository from which the pull request is created"
value: ${{ jobs.get-pr-info.outputs.PR_HEAD_REPO_NAME }}
PR_BASE_REPO_NAME:
description: "The name of the repository to which the pull request is created"
value: ${{ jobs.get-pr-info.outputs.PR_BASE_REPO_NAME }}
PR_HEAD_REF:
description: "The branch name of the pull request in the head repository"
value: ${{ jobs.get-pr-info.outputs.PR_HEAD_REF }}
PR_BASE_REF:
description: "The branch name in the base repository (to merge into)"
value: ${{ jobs.get-pr-info.outputs.PR_BASE_REF }}
PR_HEAD_SHA:
description: "The head sha of the pull request branch in the head repository"
value: ${{ jobs.get-pr-info.outputs.PR_HEAD_SHA }}
PR_BASE_SHA:
description: "The head sha of the target branch in the base repository"
value: ${{ jobs.get-pr-info.outputs.PR_BASE_SHA }}
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_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 }}
PR_MERGE_COMMIT_DATE:
description: "The date of the merge commit for the pull request (created by GitHub) in the base repository"
value: ${{ jobs.get-pr-info.outputs.PR_MERGE_COMMIT_DATE }}
PR_HEAD_COMMIT_TIMESTAMP:
description: "The timestamp of the head sha of the pull request branch in the head repository"
value: ${{ jobs.get-pr-info.outputs.PR_HEAD_COMMIT_TIMESTAMP }}
PR_MERGE_COMMIT_TIMESTAMP:
description: "The timestamp of the merge commit for the pull request (created by GitHub) in the base repository"
value: ${{ jobs.get-pr-info.outputs.PR_MERGE_COMMIT_TIMESTAMP }}
PR:
description: "The PR"
value: ${{ jobs.get-pr-info.outputs.PR }}
PR_FILES:
description: "The files touched in the PR"
value: ${{ jobs.get-pr-info.outputs.PR_FILES }}
jobs:
get-pr-info:
runs-on: ubuntu-22.04
name: Get PR commit SHA better
outputs:
PR_HEAD_REPO_FULL_NAME: ${{ steps.pr_info.outputs.head_repo_full_name }}
PR_BASE_REPO_FULL_NAME: ${{ steps.pr_info.outputs.base_repo_full_name }}
PR_HEAD_REPO_OWNER: ${{ steps.pr_info.outputs.head_repo_owner }}
PR_BASE_REPO_OWNER: ${{ steps.pr_info.outputs.base_repo_owner }}
PR_HEAD_REPO_NAME: ${{ steps.pr_info.outputs.head_repo_name }}
PR_BASE_REPO_NAME: ${{ steps.pr_info.outputs.base_repo_name }}
PR_HEAD_REF: ${{ steps.pr_info.outputs.head_ref }}
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_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 }}
PR_HEAD_COMMIT_TIMESTAMP: ${{ steps.get_timestamps.outputs.head_commit_timestamp }}
PR_MERGE_COMMIT_TIMESTAMP: ${{ steps.get_timestamps.outputs.merge_commit_timestamp }}
PR: ${{ steps.pr_info.outputs.pr }}
PR_FILES: ${{ steps.pr_info.outputs.files }}
if: ${{ inputs.pr_number != '' }}
steps:
- name: Extract PR details
id: pr_info
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: head_commit } = await github.rest.repos.getCommit({
owner: pr.head.repo.owner.login,
repo: pr.head.repo.name,
ref: pr.head.ref
});
const { data: merge_commit } = await github.rest.repos.getCommit({
owner: pr.base.repo.owner.login,
repo: pr.base.repo.name,
ref: pr.merge_commit_sha,
});
const { data: files } = await github.rest.pulls.listFiles({
owner: context.repo.owner,
repo: context.repo.repo,
pull_number: ${{ inputs.pr_number }}
});
core.setOutput('head_repo_full_name', pr.head.repo.full_name);
core.setOutput('base_repo_full_name', pr.base.repo.full_name);
core.setOutput('head_repo_owner', pr.head.repo.owner.login);
core.setOutput('base_repo_owner', pr.base.repo.owner.login);
core.setOutput('head_repo_name', pr.head.repo.name);
core.setOutput('base_repo_name', pr.base.repo.name);
core.setOutput('head_ref', pr.head.ref);
core.setOutput('base_ref', pr.base.ref);
core.setOutput('head_sha', pr.head.sha);
core.setOutput('base_sha', pr.base.sha);
core.setOutput('merge_commit_sha', pr.merge_commit_sha);
core.setOutput('pr', pr);
core.setOutput('head_commit_date', head_commit.commit.committer.date);
core.setOutput('merge_commit_date', merge_commit.commit.committer.date);
core.setOutput('files', files);
console.log('PR head commit:', {
head_commit: head_commit,
commit: head_commit.commit,
date: head_commit.commit.committer.date
});
console.log('PR merge commit:', {
merge_commit: merge_commit,
commit: merge_commit.commit,
date: merge_commit.commit.committer.date
});
- name: Convert dates to timestamps
id: get_timestamps
run: |
head_commit_date=${{ 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 $merge_commit_timestamp
echo "head_commit_timestamp=$head_commit_timestamp" >> $GITHUB_OUTPUT
echo "merge_commit_timestamp=$merge_commit_timestamp" >> $GITHUB_OUTPUT

View File

@ -1,36 +0,0 @@
name: Get PR number
on:
workflow_call:
outputs:
PR_NUMBER:
description: "The extracted PR number"
value: ${{ jobs.get-pr-number.outputs.PR_NUMBER }}
jobs:
get-pr-number:
runs-on: ubuntu-22.04
name: Get PR number
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 }}" >> $GITHUB_ENV
elif [[ "${{ 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=" >> $GITHUB_ENV
fi
- name: Check PR number
shell: bash
run: |
echo "${{ env.PR_NUMBER }}"
- name: Set PR number
id: set_pr_number
run: echo "PR_NUMBER=${{ env.PR_NUMBER }}" >> "$GITHUB_OUTPUT"

View File

@ -12,8 +12,8 @@ on:
slice_id:
required: true
type: number
runner_map:
required: false
runner:
required: true
type: string
docker:
required: true
@ -45,7 +45,7 @@ jobs:
matrix:
folders: ${{ fromJson(inputs.folder_slices)[inputs.slice_id] }}
runs-on:
group: ${{ fromJson(inputs.runner_map)[matrix.folders][inputs.machine_type] }}
group: '${{ inputs.machine_type }}'
container:
image: ${{ inputs.docker }}
options: --gpus all --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
@ -107,9 +107,9 @@ jobs:
run: |
echo "${{ inputs.machine_type }}"
if [ "${{ inputs.machine_type }}" = "aws-g5-4xlarge-cache" ]; then
if [ "${{ inputs.machine_type }}" = "aws-g4dn-4xlarge-cache" ]; then
machine_type=single-gpu
elif [ "${{ inputs.machine_type }}" = "aws-g5-12xlarge-cache" ]; then
elif [ "${{ inputs.machine_type }}" = "aws-g4dn-12xlarge-cache" ]; then
machine_type=multi-gpu
else
machine_type=${{ inputs.machine_type }}
@ -120,106 +120,23 @@ jobs:
- name: Run all tests on GPU
working-directory: /transformers
if: ${{ always() }}
run: |
python3 -m pip uninstall -y natten
python3 -m pip uninstall -y ninja && python3 -m pip install ninja && python3 -m pip install flash-attn --no-build-isolation
# START=0 END=400 python3 run.py
#
# - name: "Upload"
# if: ${{ always() }}
# uses: actions/upload-artifact@v4
# with:
# name: summary_short_1
# path: /transformers/summary_short.txt
#
# - name: Run all tests on GPU
# working-directory: /transformers
# if: ${{ always() }}
# run: |
# START=400 END=800 python3 run.py
#
# - name: "Upload"
# if: ${{ always() }}
# uses: actions/upload-artifact@v4
# with:
# name: summary_short_2
# path: /transformers/summary_short.txt
#
# - name: Run all tests on GPU
# working-directory: /transformers
# if: ${{ always() }}
# run: |
# START=800 END=1200 python3 run.py
#
# - name: "Upload"
# if: ${{ always() }}
# uses: actions/upload-artifact@v4
# with:
# name: summary_short_3
# path: /transformers/summary_short.txt
#
# - name: Run all tests on GPU
# working-directory: /transformers
# if: ${{ always() }}
# run: |
# START=1200 END=1600 python3 run.py
#
# - name: "Upload"
# if: ${{ always() }}
# uses: actions/upload-artifact@v4
# with:
# name: summary_short_4
# path: /transformers/summary_short.txt
run: python3 -m pytest -rsfE -v --make-reports=${{ env.machine_type }}_${{ inputs.report_name_prefix }}_${{ matrix.folders }}_test_reports tests/${{ matrix.folders }}
- name: Run all tests on GPU
working-directory: /transformers
if: ${{ always() }}
run: |
START=1600 END=2000 python3 run.py
- name: Failure short reports
if: ${{ failure() }}
continue-on-error: true
run: cat /transformers/reports/${{ env.machine_type }}_${{ inputs.report_name_prefix }}_${{ matrix.folders }}_test_reports/failures_short.txt
- name: "Upload"
- name: Run test
shell: bash
run: |
mkdir -p /transformers/reports/${{ env.machine_type }}_${{ inputs.report_name_prefix }}_${{ matrix.folders }}_test_reports
echo "hello" > /transformers/reports/${{ env.machine_type }}_${{ inputs.report_name_prefix }}_${{ matrix.folders }}_test_reports/hello.txt
echo "${{ env.machine_type }}_${{ inputs.report_name_prefix }}_${{ matrix.folders }}_test_reports"
- name: "Test suite reports artifacts: ${{ env.machine_type }}_${{ inputs.report_name_prefix }}_${{ env.matrix_folders }}_test_reports"
if: ${{ always() }}
uses: actions/upload-artifact@v4
with:
name: summary_short_5
path: /transformers/summary_short.txt
- name: Run all tests on GPU
working-directory: /transformers
if: ${{ always() }}
run: |
START=2000 END=2400 python3 run.py
- name: "Upload"
if: ${{ always() }}
uses: actions/upload-artifact@v4
with:
name: summary_short_6
path: /transformers/summary_short.txt
- name: Run all tests on GPU
working-directory: /transformers
if: ${{ always() }}
run: |
START=2400 END=2800 python3 run.py
- name: "Upload"
if: ${{ always() }}
uses: actions/upload-artifact@v4
with:
name: summary_short_7
path: /transformers/summary_short.txt
- name: Run all tests on GPU
working-directory: /transformers
if: ${{ always() }}
run: |
START=2800 END=3000 python3 run.py
- name: "Upload"
if: ${{ always() }}
uses: actions/upload-artifact@v4
with:
name: summary_short_8
path: /transformers/summary_short.txt
name: ${{ env.machine_type }}_${{ inputs.report_name_prefix }}_${{ env.matrix_folders }}_test_reports
path: /transformers/reports/${{ env.machine_type }}_${{ inputs.report_name_prefix }}_${{ matrix.folders }}_test_reports

128
.github/workflows/model_jobs_amd.yml vendored Normal file
View File

@ -0,0 +1,128 @@
name: model jobs
on:
workflow_call:
inputs:
folder_slices:
required: true
type: string
machine_type:
required: true
type: string
slice_id:
required: true
type: number
runner:
required: true
type: string
docker:
required: true
type: string
env:
HF_HOME: /mnt/cache
TRANSFORMERS_IS_CI: yes
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`.
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:
run_models_gpu:
name: " "
strategy:
max-parallel: 1 # For now, not to parallelize. Can change later if it works well.
fail-fast: false
matrix:
folders: ${{ fromJson(inputs.folder_slices)[inputs.slice_id] }}
runs-on: ['${{ inputs.machine_type }}', self-hosted, amd-gpu, '${{ inputs.runner }}']
container:
image: ${{ inputs.docker }}
options: --device /dev/kfd --device /dev/dri --env ROCR_VISIBLE_DEVICES --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
steps:
- name: Echo input and matrix info
shell: bash
run: |
echo "${{ inputs.folder_slices }}"
echo "${{ matrix.folders }}"
echo "${{ toJson(fromJson(inputs.folder_slices)[inputs.slice_id]) }}"
- name: Echo folder ${{ matrix.folders }}
shell: bash
# For folders like `models/bert`, set an env. var. (`matrix_folders`) to `models_bert`, which will be used to
# set the artifact folder names (because the character `/` is not allowed).
run: |
echo "${{ matrix.folders }}"
matrix_folders=${{ matrix.folders }}
matrix_folders=${matrix_folders/'models/'/'models_'}
echo "$matrix_folders"
echo "matrix_folders=$matrix_folders" >> $GITHUB_ENV
- name: Update clone
working-directory: /transformers
run: git fetch && git checkout ${{ github.sha }}
- 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: Update / Install some packages (for Past CI)
if: ${{ contains(inputs.docker, '-past-') }}
working-directory: /transformers
run: |
python3 -m pip install -U datasets
- name: Update / Install some packages (for Past CI)
if: ${{ contains(inputs.docker, '-past-') && contains(inputs.docker, '-pytorch-') }}
working-directory: /transformers
run: |
python3 -m pip install --no-cache-dir git+https://github.com/huggingface/accelerate@main#egg=accelerate
- name: ROCM-SMI
run: |
rocm-smi
- name: ROCM-INFO
run: |
rocminfo | grep "Agent" -A 14
- name: Show ROCR environment
run: |
echo "ROCR: $ROCR_VISIBLE_DEVICES"
- name: Environment
working-directory: /transformers
run: |
python3 utils/print_env.py
- name: Show installed libraries and their versions
working-directory: /transformers
run: pip freeze
- name: Run all tests on GPU
working-directory: /transformers
run: python3 -m pytest -rsfE -v --make-reports=${{ inputs.machine_type }}_run_models_gpu_${{ matrix.folders }}_test_reports tests/${{ matrix.folders }} -m "not not_device_test"
- name: Failure short reports
if: ${{ failure() }}
continue-on-error: true
run: cat /transformers/reports/${{ inputs.machine_type }}_run_models_gpu_${{ matrix.folders }}_test_reports/failures_short.txt
- name: Run test
shell: bash
run: |
mkdir -p /transformers/reports/${{ inputs.machine_type }}_run_models_gpu_${{ matrix.folders }}_test_reports
echo "hello" > /transformers/reports/${{ inputs.machine_type }}_run_models_gpu_${{ matrix.folders }}_test_reports/hello.txt
echo "${{ inputs.machine_type }}_run_models_gpu_${{ matrix.folders }}_test_reports"
- name: "Test suite reports artifacts: ${{ inputs.machine_type }}_run_models_gpu_${{ env.matrix_folders }}_test_reports"
if: ${{ always() }}
uses: actions/upload-artifact@v4
with:
name: ${{ inputs.machine_type }}_run_models_gpu_${{ env.matrix_folders }}_test_reports
path: /transformers/reports/${{ inputs.machine_type }}_run_models_gpu_${{ matrix.folders }}_test_reports

View File

@ -1,121 +0,0 @@
name: model jobs
on:
workflow_call:
inputs:
folder_slices:
required: true
type: string
slice_id:
required: true
type: number
runner:
required: true
type: string
machine_type:
required: true
type: string
report_name_prefix:
required: false
default: run_models_gpu
type: string
env:
RUN_SLOW: yes
PT_HPU_LAZY_MODE: 0
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:
run_models_gpu:
name: " "
strategy:
max-parallel: 8
fail-fast: false
matrix:
folders: ${{ fromJson(inputs.folder_slices)[inputs.slice_id] }}
runs-on:
group: ${{ inputs.runner }}
container:
image: vault.habana.ai/gaudi-docker/1.21.1/ubuntu22.04/habanalabs/pytorch-installer-2.6.0:latest
options: --runtime=habana
-v /mnt/cache/.cache/huggingface:/mnt/cache/.cache/huggingface
--env OMPI_MCA_btl_vader_single_copy_mechanism=none
--env HABANA_VISIBLE_DEVICES
--env HABANA_VISIBLE_MODULES
--cap-add=sys_nice
--shm-size=64G
steps:
- name: Echo input and matrix info
shell: bash
run: |
echo "${{ inputs.folder_slices }}"
echo "${{ matrix.folders }}"
echo "${{ toJson(fromJson(inputs.folder_slices)[inputs.slice_id]) }}"
- name: Echo folder ${{ matrix.folders }}
shell: bash
run: |
echo "${{ matrix.folders }}"
matrix_folders=${{ matrix.folders }}
matrix_folders=${matrix_folders/'models/'/'models_'}
echo "$matrix_folders"
echo "matrix_folders=$matrix_folders" >> $GITHUB_ENV
- name: Checkout
uses: actions/checkout@v4
with:
fetch-depth: 0
- name: Install dependencies
run: |
pip install -e .[testing,torch] "numpy<2.0.0" scipy scikit-learn
- name: HL-SMI
run: |
hl-smi
echo "HABANA_VISIBLE_DEVICES=${HABANA_VISIBLE_DEVICES}"
echo "HABANA_VISIBLE_MODULES=${HABANA_VISIBLE_MODULES}"
- name: Environment
run: python3 utils/print_env.py
- name: Show installed libraries and their versions
run: pip freeze
- name: Set `machine_type` for report and artifact names
shell: bash
run: |
if [ "${{ inputs.machine_type }}" = "1gaudi" ]; then
machine_type=single-gpu
elif [ "${{ inputs.machine_type }}" = "2gaudi" ]; then
machine_type=multi-gpu
else
machine_type=${{ inputs.machine_type }}
fi
echo "machine_type=$machine_type" >> $GITHUB_ENV
- name: Run all tests on Gaudi
run: python3 -m pytest -v --make-reports=${{ env.machine_type }}_${{ inputs.report_name_prefix }}_${{ matrix.folders }}_test_reports tests/${{ matrix.folders }}
- name: Failure short reports
if: ${{ failure() }}
continue-on-error: true
run: cat reports/${{ env.machine_type }}_${{ inputs.report_name_prefix }}_${{ matrix.folders }}_test_reports/failures_short.txt
- name: Run test
shell: bash
run: |
mkdir -p reports/${{ env.machine_type }}_${{ inputs.report_name_prefix }}_${{ matrix.folders }}_test_reports
echo "hello" > reports/${{ env.machine_type }}_${{ inputs.report_name_prefix }}_${{ matrix.folders }}_test_reports/hello.txt
echo "${{ env.machine_type }}_${{ inputs.report_name_prefix }}_${{ matrix.folders }}_test_reports"
- name: "Test suite reports artifacts: ${{ env.machine_type }}_${{ inputs.report_name_prefix }}_${{ env.matrix_folders }}_test_reports"
if: ${{ always() }}
uses: actions/upload-artifact@v4
with:
name: ${{ env.machine_type }}_${{ inputs.report_name_prefix }}_${{ env.matrix_folders }}_test_reports
path: reports/${{ env.machine_type }}_${{ inputs.report_name_prefix }}_${{ matrix.folders }}_test_reports

View File

@ -6,13 +6,29 @@ on:
types: [created]
permissions:
contents: write
pull-requests: write
jobs:
style:
uses: huggingface/huggingface_hub/.github/workflows/style-bot-action.yml@main
uses: huggingface/huggingface_hub/.github/workflows/style-bot-action.yml@639ee721e149a281fe726a50a2cc1354b48bc463
with:
python_quality_dependencies: "[quality]"
style_command_type: "default"
secrets:
bot_token: ${{ secrets.HF_STYLE_BOT_ACTION }}
bot_token: ${{ secrets.GITHUB_TOKEN }}
check-outputs:
runs-on: ubuntu-latest
needs: style
steps:
- run: echo ${{ needs.style.outputs.pr_number }}
- run: echo ${{ needs.style.outputs.new_commit_sha }}
trigger:
needs: style
if: needs.style.outputs.new_commit_sha != ''
uses: "./.github/workflows/build_pr_documentation.yml"
with:
pr_number: ${{ needs.style.outputs.pr_number }}
commit_sha: ${{ needs.style.outputs.new_commit_sha }}

View File

@ -1,134 +0,0 @@
name: PR - build doc via comment
on:
issue_comment:
types:
- created
branches-ignore:
- main
concurrency:
group: ${{ github.workflow }}-${{ github.event.issue.number }}-${{ startsWith(github.event.comment.body, 'build-doc') }}
cancel-in-progress: true
permissions: {}
jobs:
get-pr-number:
name: Get PR number
if: ${{ github.event.issue.state == 'open' && contains(fromJSON('["ydshieh", "ArthurZucker", "zucchini-nlp", "qubvel", "molbap", "gante", "LysandreJik", "Cyrilvallez", "Rocketknight1", "SunMarc", "muellerzr", "eustlb", "MekkCyber", "manueldeprada", "vasqu", "ivarflakstad", "stevhliu", "ebezzam"]'), github.actor) && (startsWith(github.event.comment.body, 'build-doc')) }}
uses: ./.github/workflows/get-pr-number.yml
get-pr-info:
name: Get PR commit SHA
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 }}
verity_pr_commit:
name: Verity PR commit corresponds to a specific event by comparing timestamps
if: ${{ needs.get-pr-number.outputs.PR_NUMBER != ''}}
runs-on: ubuntu-22.04
needs: get-pr-info
env:
COMMENT_DATE: ${{ github.event.comment.created_at }}
PR_MERGE_COMMIT_DATE: ${{ needs.get-pr-info.outputs.PR_MERGE_COMMIT_DATE }}
PR_MERGE_COMMIT_TIMESTAMP: ${{ needs.get-pr-info.outputs.PR_MERGE_COMMIT_TIMESTAMP }}
steps:
- run: |
COMMENT_TIMESTAMP=$(date -d "${COMMENT_DATE}" +"%s")
echo "COMMENT_DATE: $COMMENT_DATE"
echo "PR_MERGE_COMMIT_DATE: $PR_MERGE_COMMIT_DATE"
echo "COMMENT_TIMESTAMP: $COMMENT_TIMESTAMP"
echo "PR_MERGE_COMMIT_TIMESTAMP: $PR_MERGE_COMMIT_TIMESTAMP"
if [ $COMMENT_TIMESTAMP -le $PR_MERGE_COMMIT_TIMESTAMP ]; then
echo "Last commit on the pull request is newer than the issue comment triggering this run! Abort!";
exit -1;
fi
create_run:
name: Create run
needs: [get-pr-number, get-pr-info]
if: ${{ needs.get-pr-number.outputs.PR_NUMBER != '' }}
permissions:
statuses: write
runs-on: ubuntu-22.04
steps:
- name: Create Run
id: create_run
env:
GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
# Create a commit status (pending) for a run of this workflow. The status has to be updated later in `update_run_status`.
# See https://docs.github.com/en/rest/commits/statuses?apiVersion=2022-11-28#create-a-commit-status
GITHUB_RUN_URL: https://github.com/${{ github.repository }}/actions/runs/${{ github.run_id }}
run: |
gh api \
--method POST \
-H "Accept: application/vnd.github+json" \
-H "X-GitHub-Api-Version: 2022-11-28" \
repos/${{ github.repository }}/statuses/${{ needs.get-pr-info.outputs.PR_HEAD_SHA }} \
-f "target_url=$GITHUB_RUN_URL" -f "state=pending" -f "description=Custom doc building job" -f "context=custom-doc-build"
reply_to_comment:
name: Reply to the comment
if: ${{ needs.create_run.result == 'success' }}
needs: [get-pr-number, create_run]
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/${{ github.run_id }}
run: |
gh api \
--method POST \
-H "Accept: application/vnd.github+json" \
-H "X-GitHub-Api-Version: 2022-11-28" \
repos/${{ github.repository }}/issues/${{ needs.get-pr-number.outputs.PR_NUMBER }}/comments \
-f "body=[Building docs for all languages...](${{ env.GITHUB_RUN_URL }})"
build-doc:
name: Build doc
needs: [get-pr-number, get-pr-info]
if: ${{ needs.get-pr-number.outputs.PR_NUMBER != '' }}
uses: huggingface/doc-builder/.github/workflows/build_pr_documentation.yml@main
with:
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 ko pt tr zh ja te
update_run_status:
name: Update Check Run Status
needs: [ get-pr-info, create_run, build-doc ]
permissions:
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: ${{ contains(fromJSON('["skipped", "success"]'), needs.create_run.result) }}
steps:
- name: Get `build-doc` job status
run: |
echo "${{ needs.build-doc.result }}"
echo $STATUS_OK
if [ "$STATUS_OK" = "true" ]; then
echo "STATUS=success" >> $GITHUB_ENV
else
echo "STATUS=failure" >> $GITHUB_ENV
fi
- name: Update PR commit statuses
run: |
echo "${{ needs.build-doc.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/${{ needs.get-pr-info.outputs.PR_HEAD_SHA }} \
-f "target_url=$GITHUB_RUN_URL" -f "state=${{ env.STATUS }}" -f "description=Custom doc building job" -f "context=custom-doc-build"

View File

@ -1,177 +0,0 @@
name: PR slow CI
on:
pull_request_target:
types: [opened, synchronize, reopened]
jobs:
get-pr-number:
name: Get PR number
uses: ./.github/workflows/get-pr-number.yml
get-pr-info:
name: Get PR commit SHA
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 }}
get-jobs:
name: Get test files to run
runs-on: ubuntu-22.04
needs: [get-pr-number, get-pr-info]
outputs:
jobs: ${{ steps.get_jobs.outputs.jobs_to_run }}
steps:
- name: Get repository content
id: repo_content
uses: actions/github-script@v6
with:
script: |
const { data: tests_dir } = await github.rest.repos.getContent({
owner: '${{ needs.get-pr-info.outputs.PR_HEAD_REPO_OWNER }}',
repo: '${{ needs.get-pr-info.outputs.PR_HEAD_REPO_NAME }}',
path: 'tests',
ref: '${{ needs.get-pr-info.outputs.PR_HEAD_SHA }}',
});
const { data: tests_models_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 }}',
path: 'tests/models',
ref: '${{ needs.get-pr-info.outputs.PR_HEAD_SHA }}',
});
const { data: tests_quantization_dir } = await github.rest.repos.getContent({
owner: '${{ needs.get-pr-info.outputs.PR_HEAD_REPO_OWNER }}',
repo: '${{ needs.get-pr-info.outputs.PR_HEAD_REPO_NAME }}',
path: 'tests/quantization',
ref: '${{ needs.get-pr-info.outputs.PR_HEAD_SHA }}',
});
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
run: |
python utils/get_pr_run_slow_jobs.py | tee output.txt
echo "jobs_to_run: $(tail -n 1 output.txt)"
echo "jobs_to_run=$(tail -n 1 output.txt)" >> $GITHUB_OUTPUT
send_comment:
# Will delete the previous comment and send a new one if:
# - either the content is changed
# - or the previous comment is 30 minutes or more old
name: Send a comment to suggest jobs to run
if: ${{ needs.get-jobs.outputs.jobs != '' }}
needs: [get-pr-number, get-jobs]
permissions:
pull-requests: write
runs-on: ubuntu-22.04
steps:
- name: Check and update comment if needed
uses: actions/github-script@v7
env:
BODY: "\n\nrun-slow: ${{ needs.get-jobs.outputs.jobs }}"
with:
script: |
const prNumber = ${{ needs.get-pr-number.outputs.PR_NUMBER }};
const commentPrefix = "**[For maintainers]** Suggested jobs to run (before merge)";
const thirtyMinutesAgo = new Date(Date.now() - 30 * 60 * 1000); // 30 minutes ago
const newBody = `${commentPrefix}${process.env.BODY}`;
// Get all comments on the PR
const { data: comments } = await github.rest.issues.listComments({
owner: context.repo.owner,
repo: context.repo.repo,
issue_number: prNumber
});
// Find existing comments that start with our prefix
const existingComments = comments.filter(comment =>
comment.user.login === 'github-actions[bot]' &&
comment.body.startsWith(commentPrefix)
);
let shouldCreateNewComment = true;
let commentsToDelete = [];
if (existingComments.length > 0) {
// Get the most recent comment
const mostRecentComment = existingComments
.sort((a, b) => new Date(b.created_at) - new Date(a.created_at))[0];
const commentDate = new Date(mostRecentComment.created_at);
const isOld = commentDate < thirtyMinutesAgo;
const isDifferentContent = mostRecentComment.body !== newBody;
console.log(`Most recent comment created: ${mostRecentComment.created_at}`);
console.log(`Is older than 30 minutes: ${isOld}`);
console.log(`Has different content: ${isDifferentContent}`);
if (isOld || isDifferentContent) {
// Delete all existing comments and create new one
commentsToDelete = existingComments;
console.log(`Will delete ${commentsToDelete.length} existing comment(s) and create new one`);
} else {
// Content is same and comment is recent, skip
shouldCreateNewComment = false;
console.log('Comment is recent and content unchanged, skipping update');
}
} else {
console.log('No existing comments found, will create new one');
}
// Delete old comments if needed
for (const comment of commentsToDelete) {
console.log(`Deleting comment #${comment.id} (created: ${comment.created_at})`);
await github.rest.issues.deleteComment({
owner: context.repo.owner,
repo: context.repo.repo,
comment_id: comment.id
});
}
// Create new comment if needed
if (shouldCreateNewComment) {
await github.rest.issues.createComment({
owner: context.repo.owner,
repo: context.repo.repo,
issue_number: prNumber,
body: newBody
});
console.log('✅ New comment created');
} else {
console.log(' No comment update needed');
}

View File

@ -29,7 +29,7 @@ jobs:
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", "qubvel", "molbap", "gante", "LysandreJik", "Cyrilvallez", "Rocketknight1", "SunMarc", "muellerzr", "eustlb", "MekkCyber", "manueldeprada", "vasqu", "ivarflakstad", "stevhliu", "ebezzam"]'), github.actor) && (startsWith(github.event.comment.body, 'run-slow') || startsWith(github.event.comment.body, 'run slow') || startsWith(github.event.comment.body, 'run_slow')) }}
if: ${{ github.event.issue.state == 'open' && contains(fromJSON('["ydshieh", "ArthurZucker", "zucchini-nlp", "qubvel", "molbap", "gante", "LysandreJik", "Cyrilvallez", "Rocketknight1", "SunMarc", "muellerzr", "eustlb", "MekkCyber", "manueldeprada", "vasqu"]'), github.actor) && (startsWith(github.event.comment.body, 'run-slow') || startsWith(github.event.comment.body, 'run slow') || startsWith(github.event.comment.body, 'run_slow')) }}
outputs:
PR_NUMBER: ${{ steps.set_pr_number.outputs.PR_NUMBER }}
steps:
@ -185,7 +185,7 @@ jobs:
fail-fast: false
matrix:
folders: ${{ fromJson(needs.get-tests.outputs.models) }}
machine_type: [aws-g5-4xlarge-cache, aws-g5-12xlarge-cache]
machine_type: [aws-g4dn-4xlarge-cache, aws-g4dn-12xlarge-cache]
runs-on:
group: '${{ matrix.machine_type }}'
container:
@ -239,9 +239,9 @@ jobs:
shell: bash
run: |
echo "${{ matrix.machine_type }}"
if [ "${{ matrix.machine_type }}" = "aws-g5-4xlarge-cache" ]; then
if [ "${{ matrix.machine_type }}" = "aws-g4dn-4xlarge-cache" ]; then
machine_type=single-gpu
elif [ "${{ matrix.machine_type }}" = "aws-g5-12xlarge-cache" ]; then
elif [ "${{ matrix.machine_type }}" = "aws-g4dn-12xlarge-cache" ]; then
machine_type=multi-gpu
else
machine_type=${{ matrix.machine_type }}
@ -292,7 +292,7 @@ jobs:
fail-fast: false
matrix:
folders: ${{ fromJson(needs.get-tests.outputs.quantizations) }}
machine_type: [aws-g5-4xlarge-cache, aws-g5-12xlarge-cache]
machine_type: [aws-g4dn-4xlarge-cache, aws-g4dn-12xlarge-cache]
runs-on:
group: '${{ matrix.machine_type }}'
container:
@ -338,9 +338,9 @@ jobs:
shell: bash
run: |
echo "${{ matrix.machine_type }}"
if [ "${{ matrix.machine_type }}" = "aws-g5-4xlarge-cache" ]; then
if [ "${{ matrix.machine_type }}" = "aws-g4dn-4xlarge-cache" ]; then
machine_type=single-gpu
elif [ "${{ matrix.machine_type }}" = "aws-g5-12xlarge-cache" ]; then
elif [ "${{ matrix.machine_type }}" = "aws-g4dn-12xlarge-cache" ]; then
machine_type=multi-gpu
else
machine_type=${{ matrix.machine_type }}

View File

@ -0,0 +1,25 @@
name: Self-hosted runner (AMD mi300 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 mi300
if: (cancelled() != true) && ((github.event_name == 'workflow_run') || ((github.event_name == 'push') && (startsWith(github.ref_name, 'run_amd_push_ci_caller') || startsWith(github.ref_name, 'mi300-ci'))))
uses: ./.github/workflows/self-push-amd.yml
with:
gpu_flavor: mi300
secrets: inherit

View File

@ -31,12 +31,12 @@ jobs:
name: Setup
strategy:
matrix:
machine_type: [aws-g5-4xlarge-cache, aws-g5-12xlarge-cache]
machine_type: [aws-g4dn-2xlarge-cache, aws-g4dn-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/
options: --gpus 0 --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 }}
@ -131,12 +131,12 @@ jobs:
fail-fast: false
matrix:
folders: ${{ fromJson(needs.setup.outputs.matrix) }}
machine_type: [aws-g5-4xlarge-cache]
machine_type: [aws-g4dn-2xlarge-cache]
runs-on:
group: '${{ 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/
options: --gpus 0 --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: ${{ github.event.ref }}
@ -169,9 +169,9 @@ jobs:
run: |
echo "${{ matrix.machine_type }}"
if [ "${{ matrix.machine_type }}" = "aws-g5-4xlarge-cache" ]; then
if [ "${{ matrix.machine_type }}" = "aws-g4dn-2xlarge-cache" ]; then
machine_type=single-gpu
elif [ "${{ matrix.machine_type }}" = "aws-g5-12xlarge-cache" ]; then
elif [ "${{ matrix.machine_type }}" = "aws-g4dn-12xlarge-cache" ]; then
machine_type=multi-gpu
else
machine_type=${{ matrix.machine_type }}
@ -244,7 +244,7 @@ jobs:
fail-fast: false
matrix:
folders: ${{ fromJson(needs.setup.outputs.matrix) }}
machine_type: [aws-g5-12xlarge-cache]
machine_type: [aws-g4dn-12xlarge-cache]
runs-on:
group: '${{ matrix.machine_type }}'
container:
@ -282,9 +282,9 @@ jobs:
run: |
echo "${{ matrix.machine_type }}"
if [ "${{ matrix.machine_type }}" = "aws-g5-4xlarge-cache" ]; then
if [ "${{ matrix.machine_type }}" = "aws-g4dn-2xlarge-cache" ]; then
machine_type=single-gpu
elif [ "${{ matrix.machine_type }}" = "aws-g5-12xlarge-cache" ]; then
elif [ "${{ matrix.machine_type }}" = "aws-g4dn-12xlarge-cache" ]; then
machine_type=multi-gpu
else
machine_type=${{ matrix.machine_type }}
@ -357,12 +357,12 @@ jobs:
strategy:
fail-fast: false
matrix:
machine_type: [aws-g5-4xlarge-cache]
machine_type: [aws-g4dn-2xlarge-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/
options: --gpus 0 --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: ${{ github.event.ref }}
@ -395,9 +395,9 @@ jobs:
run: |
echo "${{ matrix.machine_type }}"
if [ "${{ matrix.machine_type }}" = "aws-g5-4xlarge-cache" ]; then
if [ "${{ matrix.machine_type }}" = "aws-g4dn-2xlarge-cache" ]; then
machine_type=single-gpu
elif [ "${{ matrix.machine_type }}" = "aws-g5-12xlarge-cache" ]; then
elif [ "${{ matrix.machine_type }}" = "aws-g4dn-12xlarge-cache" ]; then
machine_type=multi-gpu
else
machine_type=${{ matrix.machine_type }}
@ -467,7 +467,7 @@ jobs:
strategy:
fail-fast: false
matrix:
machine_type: [aws-g5-12xlarge-cache]
machine_type: [aws-g4dn-12xlarge-cache]
runs-on:
group: '${{ matrix.machine_type }}'
container:
@ -505,9 +505,9 @@ jobs:
run: |
echo "${{ matrix.machine_type }}"
if [ "${{ matrix.machine_type }}" = "aws-g5-4xlarge-cache" ]; then
if [ "${{ matrix.machine_type }}" = "aws-g4dn-2xlarge-cache" ]; then
machine_type=single-gpu
elif [ "${{ matrix.machine_type }}" = "aws-g5-12xlarge-cache" ]; then
elif [ "${{ matrix.machine_type }}" = "aws-g4dn-12xlarge-cache" ]; then
machine_type=multi-gpu
else
machine_type=${{ matrix.machine_type }}

View File

@ -1,8 +1,8 @@
name: Self-hosted runner scale set (AMD mi355 scheduled CI caller)
name: Self-hosted runner scale set (AMD mi300 scheduled CI caller)
# Note: For every job in this workflow, the name of the runner scale set is finalized in the runner yaml i.e. huggingface/hf-workflows/.github/workflows/transformers_amd_ci_scheduled_arc_scale_set.yaml
# For example, 1gpu : amd-mi355-ci-1gpu
# 2gpu : amd-mi355-ci-2gpu
# For example, 1gpu scale set: amd-mi300-ci-1gpu
# 2gpu scale set: amd-mi300-ci-2gpu
on:
workflow_run:
@ -20,9 +20,9 @@ jobs:
with:
job: run_models_gpu
slack_report_channel: "#amd-hf-ci"
runner_scale_set: amd-mi355-ci
runner_scale_set: amd-mi300-ci
docker: huggingface/transformers-pytorch-amd-gpu
ci_event: Scheduled CI (AMD) - mi355
ci_event: Scheduled CI (AMD) - mi300
report_repo_id: optimum-amd/transformers_daily_ci
secrets: inherit
@ -32,9 +32,9 @@ jobs:
with:
job: run_pipelines_torch_gpu
slack_report_channel: "#amd-hf-ci"
runner_scale_set: amd-mi355-ci
runner_scale_set: amd-mi300-ci
docker: huggingface/transformers-pytorch-amd-gpu
ci_event: Scheduled CI (AMD) - mi355
ci_event: Scheduled CI (AMD) - mi300
report_repo_id: optimum-amd/transformers_daily_ci
secrets: inherit
@ -44,9 +44,9 @@ jobs:
with:
job: run_examples_gpu
slack_report_channel: "#amd-hf-ci"
runner_scale_set: amd-mi355-ci
runner_scale_set: amd-mi300-ci
docker: huggingface/transformers-pytorch-amd-gpu
ci_event: Scheduled CI (AMD) - mi355
ci_event: Scheduled CI (AMD) - mi300
report_repo_id: optimum-amd/transformers_daily_ci
secrets: inherit
@ -56,8 +56,8 @@ jobs:
with:
job: run_torch_cuda_extensions_gpu
slack_report_channel: "#amd-hf-ci"
runner_scale_set: amd-mi355-ci
runner_scale_set: amd-mi300-ci
docker: huggingface/transformers-pytorch-deepspeed-amd-gpu
ci_event: Scheduled CI (AMD) - mi355
ci_event: Scheduled CI (AMD) - mi300
report_repo_id: optimum-amd/transformers_daily_ci
secrets: inherit

View File

@ -1,67 +0,0 @@
name: Self-hosted runner scale set (AMD mi325 scheduled CI caller)
# Note: For every job in this workflow, the name of the runner scale set is finalized in the runner yaml i.e. huggingface/hf-workflows/.github/workflows/transformers_amd_ci_scheduled_arc_scale_set.yaml
# For example, 1gpu scale set: amd-mi325-ci-1gpu
# 2gpu scale set: amd-mi325-ci-2gpu
on:
workflow_run:
workflows: ["Self-hosted runner (AMD scheduled CI caller)"]
branches: ["main"]
types: [completed]
push:
branches:
- run_amd_scheduled_ci_caller*
jobs:
model-ci:
name: Model CI
uses: huggingface/hf-workflows/.github/workflows/transformers_amd_ci_scheduled_arc_scale_set.yaml@main
with:
job: run_models_gpu
slack_report_channel: "#amd-hf-ci"
runner_scale_set: amd-mi325-ci
docker: huggingface/transformers-pytorch-amd-gpu
ci_event: Scheduled CI (AMD) - mi325
report_repo_id: optimum-amd/transformers_daily_ci
env_file: /etc/podinfo/gha-gpu-isolation-settings
secrets: inherit
torch-pipeline:
name: Torch pipeline CI
uses: huggingface/hf-workflows/.github/workflows/transformers_amd_ci_scheduled_arc_scale_set.yaml@main
with:
job: run_pipelines_torch_gpu
slack_report_channel: "#amd-hf-ci"
runner_scale_set: amd-mi325-ci
docker: huggingface/transformers-pytorch-amd-gpu
ci_event: Scheduled CI (AMD) - mi325
report_repo_id: optimum-amd/transformers_daily_ci
env_file: /etc/podinfo/gha-gpu-isolation-settings
secrets: inherit
example-ci:
name: Example CI
uses: huggingface/hf-workflows/.github/workflows/transformers_amd_ci_scheduled_arc_scale_set.yaml@main
with:
job: run_examples_gpu
slack_report_channel: "#amd-hf-ci"
runner_scale_set: amd-mi325-ci
docker: huggingface/transformers-pytorch-amd-gpu
ci_event: Scheduled CI (AMD) - mi325
report_repo_id: optimum-amd/transformers_daily_ci
env_file: /etc/podinfo/gha-gpu-isolation-settings
secrets: inherit
deepspeed-ci:
name: DeepSpeed CI
uses: huggingface/hf-workflows/.github/workflows/transformers_amd_ci_scheduled_arc_scale_set.yaml@main
with:
job: run_torch_cuda_extensions_gpu
slack_report_channel: "#amd-hf-ci"
runner_scale_set: amd-mi325-ci
docker: huggingface/transformers-pytorch-deepspeed-amd-gpu
ci_event: Scheduled CI (AMD) - mi325
report_repo_id: optimum-amd/transformers_daily_ci
env_file: /etc/podinfo/gha-gpu-isolation-settings
secrets: inherit

View File

@ -7,7 +7,7 @@ on:
- cron: "17 2 * * *"
push:
branches:
- check_fa2
- run_scheduled_ci*
workflow_dispatch:
inputs:
prev_workflow_run_id:
@ -22,7 +22,7 @@ on:
default: ""
# Used for `push` to easily modify the target workflow runs to compare against
# Used for `push` to easily modiffy the target workflow runs to compare against
env:
prev_workflow_run_id: ""
other_workflow_run_id: ""
@ -50,8 +50,70 @@ jobs:
uses: ./.github/workflows/self-scheduled.yml
with:
job: run_models_gpu
slack_report_channel: "#transformers-ci-dummy"
slack_report_channel: "#transformers-ci-daily-models"
runner: daily-ci
docker: huggingface/transformers-all-latest-gpu
ci_event: Daily CI
report_repo_id: hf-internal-testing/transformers_daily_ci
secrets: inherit
torch-pipeline:
name: Torch pipeline CI
uses: ./.github/workflows/self-scheduled.yml
with:
job: run_pipelines_torch_gpu
slack_report_channel: "#transformers-ci-daily-pipeline-torch"
runner: daily-ci
docker: huggingface/transformers-pytorch-gpu
ci_event: Daily CI
report_repo_id: hf-internal-testing/transformers_daily_ci
secrets: inherit
example-ci:
name: Example CI
uses: ./.github/workflows/self-scheduled.yml
with:
job: run_examples_gpu
slack_report_channel: "#transformers-ci-daily-examples"
runner: daily-ci
docker: huggingface/transformers-all-latest-gpu
ci_event: Daily CI
report_repo_id: hf-internal-testing/transformers_daily_ci
secrets: inherit
trainer-fsdp-ci:
name: Trainer/FSDP CI
uses: ./.github/workflows/self-scheduled.yml
with:
job: run_trainer_and_fsdp_gpu
slack_report_channel: "#transformers-ci-daily-training"
runner: daily-ci
docker: huggingface/transformers-all-latest-gpu
ci_event: Daily CI
report_repo_id: hf-internal-testing/transformers_daily_ci
secrets: inherit
deepspeed-ci:
name: DeepSpeed CI
uses: ./.github/workflows/self-scheduled.yml
with:
job: run_torch_cuda_extensions_gpu
slack_report_channel: "#transformers-ci-daily-training"
runner: daily-ci
docker: huggingface/transformers-pytorch-deepspeed-latest-gpu
ci_event: Daily CI
working-directory-prefix: /workspace
report_repo_id: hf-internal-testing/transformers_daily_ci
secrets: inherit
quantization-ci:
name: Quantization CI
uses: ./.github/workflows/self-scheduled.yml
with:
job: run_quantization_torch_gpu
slack_report_channel: "#transformers-ci-daily-quantization"
runner: daily-ci
docker: huggingface/transformers-quantization-latest-gpu
ci_event: Daily CI
report_repo_id: hf-internal-testing/transformers_daily_ci
secrets: inherit

View File

@ -1,342 +0,0 @@
name: Self-hosted runner (scheduled-intel-gaudi)
on:
workflow_call:
inputs:
job:
required: true
type: string
slack_report_channel:
required: true
type: string
runner_scale_set:
required: true
type: string
ci_event:
required: true
type: string
report_repo_id:
required: true
type: string
env:
NUM_SLICES: 2
RUN_SLOW: yes
PT_HPU_LAZY_MODE: 0
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:
setup:
if: contains(fromJSON('["run_models_gpu", "run_trainer_and_fsdp_gpu"]'), inputs.job)
name: Setup
runs-on: ubuntu-latest
outputs:
slice_ids: ${{ steps.set-matrix.outputs.slice_ids }}
folder_slices: ${{ steps.set-matrix.outputs.folder_slices }}
quantization_matrix: ${{ steps.set-matrix.outputs.quantization_matrix }}
steps:
- name: Checkout
uses: actions/checkout@v4
with:
fetch-depth: 0
- name: Set up Python
uses: actions/setup-python@v5
with:
python-version: "3.10"
- id: set-matrix
if: contains(fromJSON('["run_models_gpu", "run_trainer_and_fsdp_gpu"]'), inputs.job)
name: Identify models to test
working-directory: tests
run: |
if [ "${{ inputs.job }}" = "run_models_gpu" ]; then
echo "folder_slices=$(python3 ../utils/split_model_tests.py --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
- id: set-matrix-quantization
if: ${{ inputs.job == 'run_quantization_torch_gpu' }}
name: Identify quantization method to test
working-directory: tests
run: |
echo "quantization_matrix=$(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
run_models_gpu:
if: ${{ inputs.job == 'run_models_gpu' }}
name: " "
needs: setup
strategy:
fail-fast: false
matrix:
machine_type: [1gaudi, 2gaudi]
slice_id: ${{ fromJSON(needs.setup.outputs.slice_ids) }}
uses: ./.github/workflows/model_jobs_intel_gaudi.yml
with:
slice_id: ${{ matrix.slice_id }}
machine_type: ${{ matrix.machine_type }}
folder_slices: ${{ needs.setup.outputs.folder_slices }}
runner: ${{ inputs.runner_scale_set }}-${{ matrix.machine_type }}
secrets: inherit
run_trainer_and_fsdp_gpu:
if: ${{ inputs.job == 'run_trainer_and_fsdp_gpu' }}
name: " "
needs: setup
strategy:
fail-fast: false
matrix:
machine_type: [1gaudi, 2gaudi]
slice_id: ${{ fromJSON(needs.setup.outputs.slice_ids) }}
uses: ./.github/workflows/model_jobs_intel_gaudi.yml
with:
slice_id: ${{ matrix.slice_id }}
machine_type: ${{ matrix.machine_type }}
folder_slices: ${{ needs.setup.outputs.folder_slices }}
runner: ${{ inputs.runner_scale_set }}-${{ matrix.machine_type }}
report_name_prefix: run_trainer_and_fsdp_gpu
secrets: inherit
run_pipelines_torch_gpu:
if: ${{ inputs.job == 'run_pipelines_torch_gpu' }}
name: Pipelines
strategy:
fail-fast: false
matrix:
machine_type: [1gaudi, 2gaudi]
runs-on:
group: ${{ inputs.runner_scale_set }}-${{ matrix.machine_type }}
container:
image: vault.habana.ai/gaudi-docker/1.21.1/ubuntu22.04/habanalabs/pytorch-installer-2.6.0:latest
options: --runtime=habana
-v /mnt/cache/.cache/huggingface:/mnt/cache/.cache/huggingface
--env OMPI_MCA_btl_vader_single_copy_mechanism=none
--env HABANA_VISIBLE_DEVICES
--env HABANA_VISIBLE_MODULES
--cap-add=sys_nice
--shm-size=64G
steps:
- name: Checkout
uses: actions/checkout@v4
with:
fetch-depth: 0
- name: Install dependencies
run: |
pip install -e .[testing,torch] "numpy<2.0.0" scipy scikit-learn librosa soundfile
- name: HL-SMI
run: |
hl-smi
echo "HABANA_VISIBLE_DEVICES=${HABANA_VISIBLE_DEVICES}"
echo "HABANA_VISIBLE_MODULES=${HABANA_VISIBLE_MODULES}"
- name: Environment
run: python3 utils/print_env.py
- name: Show installed libraries and their versions
run: pip freeze
- name: Set `machine_type` for report and artifact names
shell: bash
run: |
if [ "${{ matrix.machine_type }}" = "1gaudi" ]; then
machine_type=single-gpu
elif [ "${{ matrix.machine_type }}" = "2gaudi" ]; then
machine_type=multi-gpu
else
machine_type=${{ matrix.machine_type }}
fi
echo "machine_type=$machine_type" >> $GITHUB_ENV
- name: Run all pipeline tests on Intel Gaudi
run: |
python3 -m pytest -v --make-reports=${{ env.machine_type }}_run_pipelines_torch_gpu_test_reports tests/pipelines -m "not not_device_test"
- name: Failure short reports
if: ${{ failure() }}
continue-on-error: true
run: |
cat reports/${{ 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() }}
uses: actions/upload-artifact@v4
with:
name: ${{ env.machine_type }}_run_pipelines_torch_gpu_test_reports
path: reports/${{ env.machine_type }}_run_pipelines_torch_gpu_test_reports
run_examples_gpu:
if: ${{ inputs.job == 'run_examples_gpu' }}
name: Examples directory
strategy:
fail-fast: false
matrix:
machine_type: [1gaudi]
runs-on:
group: ${{ inputs.runner_scale_set }}-${{ matrix.machine_type }}
container:
image: vault.habana.ai/gaudi-docker/1.21.1/ubuntu22.04/habanalabs/pytorch-installer-2.6.0:latest
options: --runtime=habana
-v /mnt/cache/.cache/huggingface:/mnt/cache/.cache/huggingface
--env OMPI_MCA_btl_vader_single_copy_mechanism=none
--env HABANA_VISIBLE_DEVICES
--env HABANA_VISIBLE_MODULES
--cap-add=sys_nice
--shm-size=64G
steps:
- name: Checkout
uses: actions/checkout@v4
with:
fetch-depth: 0
- name: Install dependencies
run: |
pip install -e .[testing,torch] "numpy<2.0.0" scipy scikit-learn librosa soundfile
- name: HL-SMI
run: |
hl-smi
echo "HABANA_VISIBLE_DEVICES=${HABANA_VISIBLE_DEVICES}"
echo "HABANA_VISIBLE_MODULES=${HABANA_VISIBLE_MODULES}"
- name: Environment
run: |
python3 utils/print_env.py
- name: Show installed libraries and their versions
run: |
pip freeze
- name: Set `machine_type` for report and artifact names
shell: bash
run: |
if [ "${{ matrix.machine_type }}" = "1gaudi" ]; then
machine_type=single-gpu
elif [ "${{ matrix.machine_type }}" = "2gaudi" ]; then
machine_type=multi-gpu
else
machine_type=${{ matrix.machine_type }}
fi
echo "machine_type=$machine_type" >> $GITHUB_ENV
- name: Run examples tests on Intel Gaudi
run: |
pip install -r examples/pytorch/_tests_requirements.txt
python3 -m pytest -v --make-reports=${{ env.machine_type }}_run_examples_gpu_test_reports examples/pytorch -m "not not_device_test"
- name: Failure short reports
if: ${{ failure() }}
continue-on-error: true
run: |
cat reports/${{ 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() }}
uses: actions/upload-artifact@v4
with:
name: ${{ env.machine_type }}_run_examples_gpu_test_reports
path: reports/${{ env.machine_type }}_run_examples_gpu_test_reports
run_torch_cuda_extensions_gpu:
if: ${{ inputs.job == 'run_torch_cuda_extensions_gpu' }}
name: Intel Gaudi deepspeed tests
strategy:
fail-fast: false
matrix:
machine_type: [1gaudi, 2gaudi]
runs-on:
group: ${{ inputs.runner_scale_set }}-${{ matrix.machine_type }}
container:
image: vault.habana.ai/gaudi-docker/1.21.1/ubuntu22.04/habanalabs/pytorch-installer-2.6.0:latest
options: --runtime=habana
-v /mnt/cache/.cache/huggingface:/mnt/cache/.cache/huggingface
--env OMPI_MCA_btl_vader_single_copy_mechanism=none
--env HABANA_VISIBLE_DEVICES
--env HABANA_VISIBLE_MODULES
--cap-add=sys_nice
--shm-size=64G
steps:
- name: Checkout
uses: actions/checkout@v4
with:
fetch-depth: 0
- name: Install dependencies
run: |
pip install -e .[testing,torch] "numpy<2.0.0" scipy scikit-learn librosa soundfile
pip install git+https://github.com/HabanaAI/DeepSpeed.git@1.20.0
- name: HL-SMI
run: |
hl-smi
echo "HABANA_VISIBLE_DEVICES=${HABANA_VISIBLE_DEVICES}"
echo "HABANA_VISIBLE_MODULES=${HABANA_VISIBLE_MODULES}"
- name: Environment
run: |
python3 utils/print_env.py
- name: Show installed libraries and their versions
run: |
pip freeze
- name: Set `machine_type` for report and artifact names
shell: bash
run: |
if [ "${{ matrix.machine_type }}" = "1gaudi" ]; then
machine_type=single-gpu
elif [ "${{ matrix.machine_type }}" = "2gaudi" ]; then
machine_type=multi-gpu
else
machine_type=${{ matrix.machine_type }}
fi
echo "machine_type=$machine_type" >> $GITHUB_ENV
- name: Run all deepspeed tests on intel Gaudi
run: |
python3 -m pytest -v --make-reports=${{ env.machine_type }}_run_torch_cuda_extensions_gpu_test_reports tests/deepspeed -m "not not_device_test"
- name: Failure short reports
if: ${{ failure() }}
continue-on-error: true
run: |
cat reports/${{ 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: ${{ env.machine_type }}_run_torch_cuda_extensions_gpu_test_reports
path: reports/${{ env.machine_type }}_run_torch_cuda_extensions_gpu_test_reports
send_results:
name: Slack Report
needs:
[
setup,
run_models_gpu,
run_examples_gpu,
run_torch_cuda_extensions_gpu,
run_pipelines_torch_gpu,
run_trainer_and_fsdp_gpu,
]
if: ${{ always() }}
uses: ./.github/workflows/slack-report.yml
with:
job: ${{ inputs.job }}
setup_status: ${{ needs.setup.result }}
slack_report_channel: ${{ inputs.slack_report_channel }}
quantization_matrix: ${{ needs.setup.outputs.quantization_matrix }}
folder_slices: ${{ needs.setup.outputs.folder_slices }}
report_repo_id: ${{ inputs.report_repo_id }}
ci_event: ${{ inputs.ci_event }}
secrets: inherit

View File

@ -1,67 +0,0 @@
name: Self-hosted runner (Intel Gaudi3 scheduled CI caller)
on:
repository_dispatch:
workflow_dispatch:
schedule:
- cron: "17 2 * * *"
jobs:
model-ci:
name: Model CI
uses: ./.github/workflows/self-scheduled-intel-gaudi.yml
with:
job: run_models_gpu
ci_event: Scheduled CI (Intel) - Gaudi3
runner_scale_set: itac-bm-emr-gaudi3-dell
slack_report_channel: "#transformers-ci-daily-intel-gaudi3"
report_repo_id: optimum-intel/transformers_daily_ci_intel_gaudi3
secrets: inherit
pipeline-ci:
name: Pipeline CI
uses: ./.github/workflows/self-scheduled-intel-gaudi.yml
with:
job: run_pipelines_torch_gpu
ci_event: Scheduled CI (Intel) - Gaudi3
runner_scale_set: itac-bm-emr-gaudi3-dell
slack_report_channel: "#transformers-ci-daily-intel-gaudi3"
report_repo_id: optimum-intel/transformers_daily_ci_intel_gaudi3
secrets: inherit
example-ci:
name: Example CI
uses: ./.github/workflows/self-scheduled-intel-gaudi.yml
with:
job: run_examples_gpu
ci_event: Scheduled CI (Intel) - Gaudi3
runner_scale_set: itac-bm-emr-gaudi3-dell
slack_report_channel: "#transformers-ci-daily-intel-gaudi3"
report_repo_id: optimum-intel/transformers_daily_ci_intel_gaudi3
secrets: inherit
deepspeed-ci:
name: DeepSpeed CI
uses: ./.github/workflows/self-scheduled-intel-gaudi.yml
with:
job: run_torch_cuda_extensions_gpu
ci_event: Scheduled CI (Intel) - Gaudi3
runner_scale_set: itac-bm-emr-gaudi3-dell
slack_report_channel: "#transformers-ci-daily-intel-gaudi3"
report_repo_id: optimum-intel/transformers_daily_ci_intel_gaudi3
secrets: inherit
trainer-fsdp-ci:
name: Trainer/FSDP CI
uses: ./.github/workflows/self-scheduled-intel-gaudi.yml
with:
job: run_trainer_and_fsdp_gpu
ci_event: Scheduled CI (Intel) - Gaudi3
runner_scale_set: itac-bm-emr-gaudi3-dell
slack_report_channel: "#transformers-ci-daily-intel-gaudi3"
report_repo_id: optimum-intel/transformers_daily_ci_intel_gaudi3
secrets: inherit

View File

@ -15,6 +15,9 @@ on:
slack_report_channel:
required: true
type: string
runner:
required: true
type: string
docker:
required: true
type: string
@ -50,16 +53,15 @@ jobs:
name: Setup
strategy:
matrix:
machine_type: [aws-g5-4xlarge-cache]
machine_type: [aws-g4dn-4xlarge-cache, aws-g4dn-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/
options: --gpus 0 --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
outputs:
folder_slices: ${{ steps.set-matrix.outputs.folder_slices }}
slice_ids: ${{ steps.set-matrix.outputs.slice_ids }}
runner_map: ${{ steps.set-matrix.outputs.runner_map }}
quantization_matrix: ${{ steps.set-matrix-quantization.outputs.quantization_matrix }}
steps:
- name: Update clone
@ -86,7 +88,6 @@ jobs:
if [ "${{ inputs.job }}" = "run_models_gpu" ]; then
echo "folder_slices=$(python3 ../utils/split_model_tests.py --num_splits ${{ env.NUM_SLICES }})" >> $GITHUB_OUTPUT
echo "slice_ids=$(python3 -c 'd = list(range(${{ env.NUM_SLICES }})); print(d)')" >> $GITHUB_OUTPUT
echo "runner_map=$(python3 ../utils/get_runner_map.py)" >> $GITHUB_OUTPUT
elif [ "${{ inputs.job }}" = "run_trainer_and_fsdp_gpu" ]; then
echo "folder_slices=[['trainer'], ['fsdp']]" >> $GITHUB_OUTPUT
echo "slice_ids=[0, 1]" >> $GITHUB_OUTPUT
@ -110,14 +111,14 @@ jobs:
strategy:
fail-fast: false
matrix:
machine_type: [single-gpu]
machine_type: [aws-g4dn-4xlarge-cache, aws-g4dn-12xlarge-cache]
slice_id: ${{ fromJSON(needs.setup.outputs.slice_ids) }}
uses: ./.github/workflows/model_jobs.yml
with:
folder_slices: ${{ needs.setup.outputs.folder_slices }}
machine_type: ${{ matrix.machine_type }}
slice_id: ${{ matrix.slice_id }}
runner_map: ${{ needs.setup.outputs.runner_map }}
runner: ${{ inputs.runner }}
docker: ${{ inputs.docker }}
secrets: inherit
@ -128,14 +129,14 @@ jobs:
strategy:
fail-fast: false
matrix:
machine_type: [aws-g5-4xlarge-cache, aws-g5-12xlarge-cache]
machine_type: [aws-g4dn-4xlarge-cache, aws-g4dn-12xlarge-cache]
slice_id: [0, 1]
uses: ./.github/workflows/model_jobs.yml
with:
folder_slices: ${{ needs.setup.outputs.folder_slices }}
machine_type: ${{ matrix.machine_type }}
slice_id: ${{ matrix.slice_id }}
runner_map: ${{ needs.setup.outputs.runner_map }}
runner: ${{ inputs.runner }}
docker: ${{ inputs.docker }}
report_name_prefix: run_trainer_and_fsdp_gpu
secrets: inherit
@ -146,7 +147,7 @@ jobs:
strategy:
fail-fast: false
matrix:
machine_type: [aws-g5-4xlarge-cache, aws-g5-12xlarge-cache]
machine_type: [aws-g4dn-4xlarge-cache, aws-g4dn-12xlarge-cache]
runs-on:
group: '${{ matrix.machine_type }}'
container:
@ -180,9 +181,9 @@ jobs:
run: |
echo "${{ matrix.machine_type }}"
if [ "${{ matrix.machine_type }}" = "aws-g5-4xlarge-cache" ]; then
if [ "${{ matrix.machine_type }}" = "aws-g4dn-4xlarge-cache" ]; then
machine_type=single-gpu
elif [ "${{ matrix.machine_type }}" = "aws-g5-12xlarge-cache" ]; then
elif [ "${{ matrix.machine_type }}" = "aws-g4dn-12xlarge-cache" ]; then
machine_type=multi-gpu
else
machine_type=${{ matrix.machine_type }}
@ -214,12 +215,12 @@ jobs:
strategy:
fail-fast: false
matrix:
machine_type: [aws-g5-4xlarge-cache]
machine_type: [aws-g4dn-4xlarge-cache]
runs-on:
group: '${{ matrix.machine_type }}'
container:
image: huggingface/transformers-all-latest-gpu
options: --gpus all --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
options: --gpus 0 --shm-size "16gb" --ipc host -v /mnt/cache/.cache/huggingface:/mnt/cache/
steps:
- name: Update clone
working-directory: /transformers
@ -248,9 +249,9 @@ jobs:
run: |
echo "${{ matrix.machine_type }}"
if [ "${{ matrix.machine_type }}" = "aws-g5-4xlarge-cache" ]; then
if [ "${{ matrix.machine_type }}" = "aws-g4dn-4xlarge-cache" ]; then
machine_type=single-gpu
elif [ "${{ matrix.machine_type }}" = "aws-g5-12xlarge-cache" ]; then
elif [ "${{ matrix.machine_type }}" = "aws-g4dn-12xlarge-cache" ]; then
machine_type=multi-gpu
else
machine_type=${{ matrix.machine_type }}
@ -283,7 +284,7 @@ jobs:
strategy:
fail-fast: false
matrix:
machine_type: [aws-g5-4xlarge-cache, aws-g5-12xlarge-cache]
machine_type: [aws-g4dn-4xlarge-cache, aws-g4dn-12xlarge-cache]
runs-on:
group: '${{ matrix.machine_type }}'
container:
@ -345,9 +346,9 @@ jobs:
run: |
echo "${{ matrix.machine_type }}"
if [ "${{ matrix.machine_type }}" = "aws-g5-4xlarge-cache" ]; then
if [ "${{ matrix.machine_type }}" = "aws-g4dn-4xlarge-cache" ]; then
machine_type=single-gpu
elif [ "${{ matrix.machine_type }}" = "aws-g5-12xlarge-cache" ]; then
elif [ "${{ matrix.machine_type }}" = "aws-g4dn-12xlarge-cache" ]; then
machine_type=multi-gpu
else
machine_type=${{ matrix.machine_type }}
@ -382,7 +383,7 @@ jobs:
fail-fast: false
matrix:
folders: ${{ fromJson(needs.setup.outputs.quantization_matrix) }}
machine_type: [aws-g5-4xlarge-cache, aws-g5-12xlarge-cache]
machine_type: [aws-g4dn-4xlarge-cache, aws-g4dn-12xlarge-cache]
runs-on:
group: '${{ matrix.machine_type }}'
container:
@ -425,9 +426,9 @@ jobs:
run: |
echo "${{ matrix.machine_type }}"
if [ "${{ matrix.machine_type }}" = "aws-g5-4xlarge-cache" ]; then
if [ "${{ matrix.machine_type }}" = "aws-g4dn-4xlarge-cache" ]; then
machine_type=single-gpu
elif [ "${{ matrix.machine_type }}" = "aws-g5-12xlarge-cache" ]; then
elif [ "${{ matrix.machine_type }}" = "aws-g4dn-12xlarge-cache" ]; then
machine_type=multi-gpu
else
machine_type=${{ matrix.machine_type }}
@ -452,3 +453,86 @@ jobs:
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
run_extract_warnings:
# Let's only do this for the job `run_models_gpu` to simplify the (already complex) logic.
if: ${{ always() && inputs.job == 'run_models_gpu' }}
name: Extract warnings in CI artifacts
runs-on: ubuntu-22.04
needs: [setup, run_models_gpu]
steps:
- name: Checkout transformers
uses: actions/checkout@v4
with:
fetch-depth: 2
- name: Install transformers
run: pip install transformers
- name: Show installed libraries and their versions
run: pip freeze
- name: Create output directory
run: mkdir warnings_in_ci
- uses: actions/download-artifact@v4
with:
path: warnings_in_ci
- name: Show artifacts
run: echo "$(python3 -c 'import os; d = os.listdir(); print(d)')"
working-directory: warnings_in_ci
- name: Extract warnings in CI artifacts
run: |
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() }}
uses: actions/upload-artifact@v4
with:
name: warnings_in_ci
path: warnings_in_ci/selected_warnings.json
send_results:
name: Slack Report
needs: [
setup,
run_models_gpu,
run_trainer_and_fsdp_gpu,
run_pipelines_torch_gpu,
run_examples_gpu,
run_torch_cuda_extensions_gpu,
run_quantization_torch_gpu,
run_extract_warnings
]
if: ${{ always() }}
uses: ./.github/workflows/slack-report.yml
with:
job: ${{ inputs.job }}
# This would be `skipped` if `setup` is skipped.
setup_status: ${{ needs.setup.result }}
slack_report_channel: ${{ inputs.slack_report_channel }}
# This would be an empty string if `setup` is skipped.
folder_slices: ${{ needs.setup.outputs.folder_slices }}
quantization_matrix: ${{ needs.setup.outputs.quantization_matrix }}
ci_event: ${{ inputs.ci_event }}
report_repo_id: ${{ inputs.report_repo_id }}
secrets: inherit
check_new_failures:
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 }}
start_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 }}
secrets: inherit

3
.gitignore vendored
View File

@ -167,6 +167,3 @@ tags
# ruff
.ruff_cache
# modular conversion
*.modular_backup

View File

@ -1,39 +0,0 @@
# AGENTS.md Guide for Hugging Face Transformers
This AGENTS.md file provides guidance for code agents working with this codebase.
## Core Project Structure
- `/src/transformers`: This contains the core source code for the library
- `/models`: Code for individual models. Models inherit from base classes in the root `/src/transformers` directory.
- `/tests`: This contains the core test classes for the library. These are usually inherited rather than directly run.
- `/models`: Tests for individual models. Model tests inherit from common tests in the root `/tests` directory.
- `/docs`: This contains the documentation for the library, including guides, tutorials, and API references.
## Coding Conventions for Hugging Face Transformers
- 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.
## Copying and inheritance
Many models in the codebase have similar code, but it is not shared by inheritance because we want each model file to be self-contained.
We use two mechanisms to keep this code in sync:
- "Copied from" syntax. Functions or entire classes can have a comment at the top like this: `# Copied from transformers.models.llama.modeling_llama.rotate_half` or `# Copied from transformers.models.t5.modeling_t5.T5LayerNorm with T5->MT5`
These comments are actively checked by the style tools, and copies will automatically be updated when the base code is updated. If you need to update a copied function, you should
either update the base function and use `make fixup` to propagate the change to all copies, or simply remove the `# Copied from` comment if that is inappropriate.
- "Modular" files. These files briefly define models by composing them using inheritance from other models. They are not meant to be used directly. Instead, the style tools
automatically generate a complete modeling file, like `modeling_bert.py`, from the modular file like `modular_bert.py`. If a model has a modular file, the modeling file
should never be edited directly! Instead, changes should be made in the modular file, and then you should run `make fixup` to update the modeling file automatically.
When adding new models, you should prefer `modular` style.
## Testing
After making changes, you should usually run `make fixup` to ensure any copies and modular files are updated, and then test all affected models. This includes both
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.

View File

@ -8,19 +8,13 @@ check_dirs := examples tests src utils
exclude_folders := ""
modified_only_fixup:
@current_branch=$$(git branch --show-current); \
if [ "$$current_branch" = "main" ]; then \
echo "On main branch, running 'style' target instead..."; \
$(MAKE) style; \
$(eval modified_py_files := $(shell python utils/get_modified_files.py $(check_dirs)))
@if test -n "$(modified_py_files)"; then \
echo "Checking/fixing $(modified_py_files)"; \
ruff check $(modified_py_files) --fix --exclude $(exclude_folders); \
ruff format $(modified_py_files) --exclude $(exclude_folders);\
else \
modified_py_files=$$(python utils/get_modified_files.py $(check_dirs)); \
if [ -n "$$modified_py_files" ]; then \
echo "Checking/fixing files: $${modified_py_files}"; \
ruff check $${modified_py_files} --fix --exclude $(exclude_folders); \
ruff format $${modified_py_files} --exclude $(exclude_folders); \
else \
echo "No library .py files were modified"; \
fi; \
echo "No library .py files were modified"; \
fi
# Update src/transformers/dependency_versions_table.py
@ -46,13 +40,11 @@ repo-consistency:
python utils/check_dummies.py
python utils/check_repo.py
python utils/check_inits.py
python utils/check_pipeline_typing.py
python utils/check_config_docstrings.py
python utils/check_config_attributes.py
python utils/check_doctest_list.py
python utils/update_metadata.py --check-only
python utils/check_docstrings.py
python utils/add_dates.py
# this target runs checks on all files
@ -89,7 +81,6 @@ fix-copies:
python utils/check_copies.py --fix_and_overwrite
python utils/check_modular_conversion.py --fix_and_overwrite
python utils/check_dummies.py --fix_and_overwrite
python utils/check_pipeline_typing.py --fix_and_overwrite
python utils/check_doctest_list.py --fix_and_overwrite
python utils/check_docstrings.py --fix_and_overwrite

View File

@ -44,7 +44,7 @@ limitations under the License.
<a href="https://github.com/huggingface/transformers/blob/main/i18n/README_ja.md">日本語</a> |
<a href="https://github.com/huggingface/transformers/blob/main/i18n/README_hd.md">हिन्दी</a> |
<a href="https://github.com/huggingface/transformers/blob/main/i18n/README_ru.md">Русский</a> |
<a href="https://github.com/huggingface/transformers/blob/main/i18n/README_pt-br.md">Português</a> |
<a href="https://github.com/huggingface/transformers/blob/main/i18n/README_pt-br.md">Рortuguês</a> |
<a href="https://github.com/huggingface/transformers/blob/main/i18n/README_te.md">తెలుగు</a> |
<a href="https://github.com/huggingface/transformers/blob/main/i18n/README_fr.md">Français</a> |
<a href="https://github.com/huggingface/transformers/blob/main/i18n/README_de.md">Deutsch</a> |
@ -59,22 +59,12 @@ limitations under the License.
</h3>
<h3 align="center">
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/transformers_as_a_model_definition.png"/>
<a href="https://hf.co/course"><img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/course_banner.png"></a>
</h3>
Transformers is a library of pretrained text, computer vision, audio, video, and multimodal models for inference and training. Use Transformers to fine-tune models on your data, build inference applications, and for generative AI use cases across multiple modalities.
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
frameworks (Axolotl, Unsloth, DeepSpeed, FSDP, PyTorch-Lightning, ...), inference engines (vLLM, SGLang, TGI, ...),
and adjacent modeling libraries (llama.cpp, mlx, ...) which leverage the model definition from `transformers`.
We pledge to help support new state-of-the-art models and democratize their usage by having their model definition be
simple, customizable, and efficient.
There are over 1M+ Transformers [model checkpoints](https://huggingface.co/models?library=transformers&sort=trending) on the [Hugging Face Hub](https://huggingface.com/models) you can use.
There are over 500K+ Transformers [model checkpoints](https://huggingface.co/models?library=transformers&sort=trending) on the [Hugging Face Hub](https://huggingface.com/models) you can use.
Explore the [Hub](https://huggingface.com/) today to find a model and use Transformers to help you get started right away.
@ -242,7 +232,7 @@ pipeline(
- This library is not a modular toolbox of building blocks for neural nets. The code in the model files is not refactored with additional abstractions on purpose, so that researchers can quickly iterate on each of the models without diving into additional abstractions/files.
- The training API is optimized to work with PyTorch models provided by Transformers. For generic machine learning loops, you should use another library like [Accelerate](https://huggingface.co/docs/accelerate).
- The [example scripts](https://github.com/huggingface/transformers/tree/main/examples) are only *examples*. They may not necessarily work out-of-the-box on your specific use case and you'll need to adapt the code for it to work.
- The [example scripts]((https://github.com/huggingface/transformers/tree/main/examples)) are only *examples*. They may not necessarily work out-of-the-box on your specific use case and you'll need to adapt the code for it to work.
## 100 projects using Transformers
@ -280,8 +270,8 @@ Expand each modality below to see a few example models for various use cases.
- Automatic mask generation with [SAM](https://huggingface.co/facebook/sam-vit-base)
- Depth estimation with [DepthPro](https://huggingface.co/apple/DepthPro-hf)
- Image classification with [DINO v2](https://huggingface.co/facebook/dinov2-base)
- Keypoint detection with [SuperPoint](https://huggingface.co/magic-leap-community/superpoint)
- Keypoint matching with [SuperGlue](https://huggingface.co/magic-leap-community/superglue_outdoor)
- Keypoint detection with [SuperGlue](https://huggingface.co/magic-leap-community/superglue_outdoor)
- Keypoint matching with [SuperGlue](https://huggingface.co/magic-leap-community/superglue)
- Object detection with [RT-DETRv2](https://huggingface.co/PekingU/rtdetr_v2_r50vd)
- Pose Estimation with [VitPose](https://huggingface.co/usyd-community/vitpose-base-simple)
- Universal segmentation with [OneFormer](https://huggingface.co/shi-labs/oneformer_ade20k_swin_large)

View File

@ -14,7 +14,7 @@ Models uploaded on the Hugging Face Hub come in different formats. We heavily re
models in the [`safetensors`](https://github.com/huggingface/safetensors) format (which is the default prioritized
by the transformers library), as developed specifically to prevent arbitrary code execution on your system.
To avoid loading models from unsafe formats (e.g. [pickle](https://docs.python.org/3/library/pickle.html), you should use the `use_safetensors` parameter. If doing so, in the event that no .safetensors file is present, transformers will error when loading the model.
To avoid loading models from unsafe formats(e.g. [pickle](https://docs.python.org/3/library/pickle.html), you should use the `use_safetensors` parameter. If doing so, in the event that no .safetensors file is present, transformers will error when loading the model.
### Remote code

View File

@ -288,7 +288,7 @@ Keywords: Music understanding, Music generation
## [dalle-flow](https://github.com/jina-ai/dalle-flow)
DALL·E Flow is an interactive workflow for generating high-definition images from a text prompt. It leverages DALL·E-Mega, GLID-3 XL, and Stable Diffusion to generate image candidates, and then calls CLIP-as-service to rank the candidates w.r.t. the prompt.
DALL·E Flow is an interactive workflow for generating high-definition images from a text prompt. Itt leverages DALL·E-Mega, GLID-3 XL, and Stable Diffusion to generate image candidates, and then calls CLIP-as-service to rank the candidates w.r.t. the prompt.
The preferred candidate is fed to GLID-3 XL for diffusion, which often enriches the texture and background. Finally, the candidate is upscaled to 1024x1024 via SwinIR.
Keywords: High-definition image generation, Stable Diffusion, DALL-E Mega, GLID-3 XL, CLIP, SwinIR
@ -526,7 +526,7 @@ Keywords: Model deployment, CLoud, Mobile, Edge
## [underthesea](https://github.com/undertheseanlp/underthesea)
[underthesea](https://github.com/undertheseanlp/underthesea) is a Vietnamese NLP toolkit. Underthesea is a suite of open source Python modules data sets and tutorials supporting research and development in Vietnamese Natural Language Processing. We provide extremely easy API to quickly apply pretrained NLP models to your Vietnamese text, such as word segmentation, part-of-speech tagging (PoS), named entity recognition (NER), text classification and dependency parsing.
[underthesea](https://github.com/undertheseanlp/underthesea) is a Vietnamese NLP toolkit. Underthesea is a suite of open source Python modules data sets and tutorials supporting research and development in Vietnamese Natural Language Processing. We provides extremely easy API to quickly apply pretrained NLP models to your Vietnamese text, such as word segmentation, part-of-speech tagging (PoS), named entity recognition (NER), text classification and dependency parsing.
Keywords: Vietnamese, NLP

View File

@ -1 +0,0 @@
benchmark_results/

View File

@ -1,345 +0,0 @@
# 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.
from logging import Logger
import os
from threading import Event, Thread
from time import perf_counter, sleep
from typing import Optional
import sys
# Add the parent directory to Python path to import benchmarks_entrypoint
sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
from benchmarks_entrypoint import MetricsRecorder
import gpustat
import psutil
import psycopg2
# Optional heavy ML dependencies - only required when actually running the benchmark
try:
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, GenerationConfig, StaticCache
TRANSFORMERS_AVAILABLE = True
except ImportError:
TRANSFORMERS_AVAILABLE = False
torch = None
AutoModelForCausalLM = None
AutoTokenizer = None
GenerationConfig = None
StaticCache = None
os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "1"
os.environ["TOKENIZERS_PARALLELISM"] = "1"
# Only set torch precision if torch is available
if TRANSFORMERS_AVAILABLE:
torch.set_float32_matmul_precision("high")
def collect_metrics(benchmark_id, continue_metric_collection, metrics_recorder):
p = psutil.Process(os.getpid())
while not continue_metric_collection.is_set():
with p.oneshot():
cpu_util = p.cpu_percent()
mem_megabytes = p.memory_info().rss / (1024 * 1024)
gpu_stats = gpustat.GPUStatCollection.new_query()
gpu_util = gpu_stats[0]["utilization.gpu"]
gpu_mem_megabytes = gpu_stats[0]["memory.used"]
metrics_recorder.collect_device_measurements(
benchmark_id, cpu_util, mem_megabytes, gpu_util, gpu_mem_megabytes
)
sleep(0.01)
def run_benchmark(
logger: Logger, repository: str, branch: str, commit_id: str, commit_msg: str, metrics_recorder=None, num_tokens_to_generate=100
):
# Check if required ML dependencies are available
if not TRANSFORMERS_AVAILABLE:
logger.error("Transformers and torch are required to run the LLaMA benchmark. Please install them with:")
logger.error("pip install torch transformers")
logger.error("Skipping LLaMA benchmark due to missing dependencies.")
return
continue_metric_collection = Event()
metrics_thread = None
model_id = "meta-llama/Llama-2-7b-hf"
# If no metrics_recorder is provided, create one for backward compatibility
if metrics_recorder is None:
try:
metrics_recorder = MetricsRecorder(
psycopg2.connect("dbname=metrics"), logger, repository, branch, commit_id, commit_msg, True
)
should_close_recorder = True
except Exception as e:
logger.error(f"Failed to create metrics recorder: {e}")
return
else:
should_close_recorder = False
try:
gpu_stats = gpustat.GPUStatCollection.new_query()
gpu_name = gpu_stats[0]["name"]
benchmark_id = metrics_recorder.initialise_benchmark({"gpu_name": gpu_name, "model_id": model_id})
logger.info(f"running benchmark #{benchmark_id} on {gpu_name} for {model_id}")
metrics_thread = Thread(
target=collect_metrics,
args=[benchmark_id, continue_metric_collection, metrics_recorder],
)
metrics_thread.start()
logger.info("started background thread to fetch device metrics")
os.environ["TOKENIZERS_PARALLELISM"] = "false" # silence warnings when compiling
device = "cuda"
logger.info("downloading weights")
# This is to avoid counting download in model load time measurement
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.float16)
gen_config = GenerationConfig(do_sample=False, top_p=1, temperature=1)
logger.info("loading model")
start = perf_counter()
model = AutoModelForCausalLM.from_pretrained(
model_id, torch_dtype=torch.float16, generation_config=gen_config
).eval()
model.to(device)
torch.cuda.synchronize()
end = perf_counter()
model_load_time = end - start
logger.info(f"loaded model in: {model_load_time}s")
tokenizer = AutoTokenizer.from_pretrained(model_id)
prompt = "Why dogs are so cute?"
inputs = tokenizer(prompt, return_tensors="pt").to(device)
# Specify the max length (including both the prompt and the response)
# When calling `generate` with `cache_implementation="static" later, this is also used to create a `StaticCache` object
# with sequence length = `max_length`. The longer the more you will re-use it
seq_length = inputs["input_ids"].shape[1]
model.generation_config.max_length = seq_length + num_tokens_to_generate
batch_size = inputs["input_ids"].shape[0]
# Copied from the gpt-fast repo
def multinomial_sample_one_no_sync(probs_sort): # Does multinomial sampling without a cuda synchronization
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: Optional[int] = None):
logits = logits / max(temperature, 1e-5)
if top_k is not None:
v, _ = torch.topk(logits, min(top_k, logits.size(-1)))
pivot = v.select(-1, -1).unsqueeze(-1)
logits = torch.where(logits < pivot, -float("Inf"), logits)
probs = torch.nn.functional.softmax(logits, dim=-1)
return probs
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
# First eager forward pass
logger.info("running first eager forward pass")
start = perf_counter()
outputs = model(**inputs)
torch.cuda.synchronize()
end = perf_counter()
first_eager_fwd_pass_time = end - start
logger.info(f"completed first eager forward pass in: {first_eager_fwd_pass_time}s")
# Second eager forward pass (should be faster)
logger.info("running second eager forward pass")
start = perf_counter()
outputs = model(**inputs)
torch.cuda.synchronize()
end = perf_counter()
second_eager_fwd_pass_time = end - start
logger.info(f"completed second eager forward pass in: {second_eager_fwd_pass_time}s")
# First eager generation
logger.info("running first eager generation")
start = perf_counter()
output = model.generate(**inputs)
torch.cuda.synchronize()
end = perf_counter()
first_eager_generate_time = end - start
logger.info(f"completed first eager generation in: {first_eager_generate_time}s")
logger.info(f"generated: {tokenizer.batch_decode(output.cpu().tolist())}")
# Second eager generation (should be faster)
logger.info("running second eager generation")
start = perf_counter()
output = model.generate(**inputs)
torch.cuda.synchronize()
end = perf_counter()
second_eager_generate_time = end - start
logger.info(f"completed second eager generation in: {second_eager_generate_time}s")
logger.info(f"generated: {tokenizer.batch_decode(output.cpu().tolist())}")
logger.info("running generation timing loop")
input_pos = torch.arange(0, seq_length, device=device)
inputs = inputs["input_ids"]
start = perf_counter()
with torch.nn.attention.sdpa_kernel(torch.nn.attention.SDPBackend.MATH):
logits = model(inputs, position_ids=input_pos).logits
next_token, probs = sample(logits, temperature=0.6, top_k=5)
torch.cuda.synchronize()
end = perf_counter()
time_to_first_token = end - start
input_pos = torch.tensor([seq_length], device=device, dtype=torch.int)
next_token = next_token.clone()
start = perf_counter()
with torch.nn.attention.sdpa_kernel(torch.nn.attention.SDPBackend.MATH):
logits = model(next_token, position_ids=input_pos).logits
next_token, probs = sample(logits, temperature=0.6, top_k=5)
torch.cuda.synchronize()
end = perf_counter()
time_to_second_token = end - start
input_pos = torch.tensor([seq_length + 1], device=device, dtype=torch.int)
next_token = next_token.clone()
start = perf_counter()
with torch.nn.attention.sdpa_kernel(torch.nn.attention.SDPBackend.MATH):
logits = model(next_token, position_ids=input_pos).logits
next_token, probs = sample(logits, temperature=0.6, top_k=5)
torch.cuda.synchronize()
end = perf_counter()
time_to_third_token = end - start
logger.info("running longer generation timing loop")
total_time = 0
for i in range(20):
input_pos = torch.tensor([seq_length + 2 + i], device=device, dtype=torch.int)
next_token = next_token.clone()
start = perf_counter()
with torch.nn.attention.sdpa_kernel(torch.nn.attention.SDPBackend.MATH):
logits = model(next_token, position_ids=input_pos).logits
next_token, probs = sample(logits, temperature=0.6, top_k=5)
torch.cuda.synchronize()
end = perf_counter()
total_time += end - start
mean_time_to_next_token = total_time / 20
logger.info("running compilation benchmarks")
# Now compile the model
model = torch.compile(model, mode="max-autotune", fullgraph=True)
# StaticCache for generation
with torch.device(device):
model.setup_caches(max_batch_size=batch_size, max_seq_len=seq_length + num_tokens_to_generate)
input_pos = torch.arange(0, seq_length, device=device)
inputs = tokenizer(prompt, return_tensors="pt").to(device)["input_ids"]
logger.info("compiling model")
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.float16, generation_config=gen_config)
model.to(device)
model = torch.compile(model, mode="max-autotune", fullgraph=True)
past_key_values = StaticCache(
model.config,
max_batch_size=batch_size,
device=device,
dtype=torch.float16,
max_cache_len=seq_length + 128,
)
# 1st call
start = perf_counter()
output = model.generate(**inputs, past_key_values=past_key_values)
end = perf_counter()
first_compile_generate_time = end - start
logger.info(f"completed first compile generation in: {first_compile_generate_time}s")
logger.info(f"generated: {tokenizer.batch_decode(output.cpu().tolist())}")
past_key_values = StaticCache(
model.config,
max_batch_size=batch_size,
device=device,
dtype=torch.float16,
max_cache_len=seq_length + 128,
)
# 2nd call
start = perf_counter()
output = model.generate(**inputs, past_key_values=past_key_values)
end = perf_counter()
second_compile_generate_time = end - start
logger.info(f"completed second compile generation in: {second_compile_generate_time}s")
logger.info(f"generated: {tokenizer.batch_decode(output.cpu().tolist())}")
past_key_values = StaticCache(
model.config,
max_batch_size=batch_size,
device=device,
dtype=torch.float16,
max_cache_len=seq_length + 128,
)
# 3rd call
start = perf_counter()
output = model.generate(**inputs, past_key_values=past_key_values)
end = perf_counter()
third_compile_generate_time = end - start
logger.info(f"completed third compile generation in: {third_compile_generate_time}s")
logger.info(f"generated: {tokenizer.batch_decode(output.cpu().tolist())}")
past_key_values = StaticCache(
model.config,
max_batch_size=batch_size,
device=device,
dtype=torch.float16,
max_cache_len=seq_length + 128,
)
# 4th call
start = perf_counter()
output = model.generate(**inputs, past_key_values=past_key_values)
end = perf_counter()
fourth_compile_generate_time = end - start
logger.info(f"completed fourth compile generation in: {fourth_compile_generate_time}s")
logger.info(f"generated: {tokenizer.batch_decode(output.cpu().tolist())}")
metrics_recorder.collect_model_measurements(
benchmark_id,
{
"model_load_time": model_load_time,
"first_eager_forward_pass_time_secs": first_eager_fwd_pass_time,
"second_eager_forward_pass_time_secs": second_eager_fwd_pass_time,
"first_eager_generate_time_secs": first_eager_generate_time,
"second_eager_generate_time_secs": second_eager_generate_time,
"time_to_first_token_secs": time_to_first_token,
"time_to_second_token_secs": time_to_second_token,
"time_to_third_token_secs": time_to_third_token,
"time_to_next_token_mean_secs": mean_time_to_next_token,
"first_compile_generate_time_secs": first_compile_generate_time,
"second_compile_generate_time_secs": second_compile_generate_time,
"third_compile_generate_time_secs": third_compile_generate_time,
"fourth_compile_generate_time_secs": fourth_compile_generate_time,
},
)
except Exception as e:
logger.error(f"Caught exception: {e}")
continue_metric_collection.set()
if metrics_thread is not None:
metrics_thread.join()
# Only close the recorder if we created it locally
if should_close_recorder:
metrics_recorder.close()

View File

@ -1,35 +1,15 @@
# 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 argparse
import importlib.util
import logging
import os
import sys
import json
import uuid
from datetime import datetime
from typing import Dict, Tuple, Optional, List
from typing import Dict, Tuple
import pandas as pd
from psycopg2.extensions import register_adapter
from psycopg2.extras import Json
try:
from psycopg2.extensions import register_adapter
from psycopg2.extras import Json
register_adapter(dict, Json)
PSYCOPG2_AVAILABLE = True
except ImportError:
PSYCOPG2_AVAILABLE = False
register_adapter(dict, Json)
class ImportModuleException(Exception):
@ -38,239 +18,61 @@ class ImportModuleException(Exception):
class MetricsRecorder:
def __init__(
self, connection, logger: logging.Logger, repository: str, branch: str, commit_id: str, commit_msg: str,
collect_csv_data: bool = True
self, connection, logger: logging.Logger, repository: str, branch: str, commit_id: str, commit_msg: str
):
self.conn = connection
self.use_database = connection is not None
if self.use_database:
self.conn.autocommit = True
self.conn.autocommit = True
self.logger = logger
self.repository = repository
self.branch = branch
self.commit_id = commit_id
self.commit_msg = commit_msg
self.collect_csv_data = collect_csv_data
# For CSV export - store all data in pandas DataFrames (only if CSV collection is enabled)
if self.collect_csv_data:
# Initialize empty DataFrames with proper schemas
self.benchmarks_df = pd.DataFrame(columns=[
'benchmark_id', 'repository', 'branch', 'commit_id', 'commit_message',
'metadata', 'created_at'
])
self.device_measurements_df = pd.DataFrame(columns=[
'benchmark_id', 'cpu_util', 'mem_megabytes', 'gpu_util',
'gpu_mem_megabytes', 'time'
])
self.model_measurements_df = pd.DataFrame(columns=[
'benchmark_id', 'time', 'model_load_time', 'first_eager_forward_pass_time_secs',
'second_eager_forward_pass_time_secs', 'first_eager_generate_time_secs',
'second_eager_generate_time_secs', 'time_to_first_token_secs',
'time_to_second_token_secs', 'time_to_third_token_secs',
'time_to_next_token_mean_secs', 'first_compile_generate_time_secs',
'second_compile_generate_time_secs', 'third_compile_generate_time_secs',
'fourth_compile_generate_time_secs'
])
else:
self.benchmarks_df = None
self.device_measurements_df = None
self.model_measurements_df = None
def initialise_benchmark(self, metadata: dict[str, str]) -> str:
def initialise_benchmark(self, metadata: Dict[str, str]) -> int:
"""
Creates a new benchmark, returns the benchmark id (UUID)
Creates a new benchmark, returns the benchmark id
"""
# Generate a unique UUID for this benchmark
benchmark_id = str(uuid.uuid4())
if self.use_database:
with self.conn.cursor() as cur:
cur.execute(
"INSERT INTO benchmarks (benchmark_id, repository, branch, commit_id, commit_message, metadata) VALUES (%s, %s, %s, %s, %s, %s)",
(benchmark_id, self.repository, self.branch, self.commit_id, self.commit_msg, metadata),
)
self.logger.debug(f"initialised benchmark #{benchmark_id}")
# Store benchmark data for CSV export (if enabled)
if self.collect_csv_data:
# Add row to pandas DataFrame
new_row = pd.DataFrame([{
'benchmark_id': benchmark_id,
'repository': self.repository,
'branch': self.branch,
'commit_id': self.commit_id,
'commit_message': self.commit_msg,
'metadata': json.dumps(metadata),
'created_at': datetime.utcnow().isoformat()
}])
self.benchmarks_df = pd.concat([self.benchmarks_df, new_row], ignore_index=True)
mode_info = []
if self.use_database:
mode_info.append("database")
if self.collect_csv_data:
mode_info.append("CSV")
mode_str = " + ".join(mode_info) if mode_info else "no storage"
self.logger.debug(f"initialised benchmark #{benchmark_id} ({mode_str} mode)")
return benchmark_id
# gpu_name: str, model_id: str
with self.conn.cursor() as cur:
cur.execute(
"INSERT INTO benchmarks (repository, branch, commit_id, commit_message, metadata) VALUES (%s, %s, %s, %s, %s) RETURNING benchmark_id",
(self.repository, self.branch, self.commit_id, self.commit_msg, metadata),
)
benchmark_id = cur.fetchone()[0]
logger.debug(f"initialised benchmark #{benchmark_id}")
return benchmark_id
def collect_device_measurements(self, benchmark_id: str, cpu_util, mem_megabytes, gpu_util, gpu_mem_megabytes):
def collect_device_measurements(self, benchmark_id: int, cpu_util, mem_megabytes, gpu_util, gpu_mem_megabytes):
"""
Collect device metrics, such as CPU & GPU usage. These are "static", as in you cannot pass arbitrary arguments to the function.
"""
# Store device measurements for CSV export (if enabled)
if self.collect_csv_data:
# Add row to pandas DataFrame
new_row = pd.DataFrame([{
'benchmark_id': benchmark_id,
'cpu_util': cpu_util,
'mem_megabytes': mem_megabytes,
'gpu_util': gpu_util,
'gpu_mem_megabytes': gpu_mem_megabytes,
'time': datetime.utcnow().isoformat()
}])
self.device_measurements_df = pd.concat([self.device_measurements_df, new_row], ignore_index=True)
# Store in database if available
if self.use_database:
with self.conn.cursor() as cur:
cur.execute(
"INSERT INTO device_measurements (benchmark_id, cpu_util, mem_megabytes, gpu_util, gpu_mem_megabytes) VALUES (%s, %s, %s, %s, %s)",
(benchmark_id, cpu_util, mem_megabytes, gpu_util, gpu_mem_megabytes),
)
with self.conn.cursor() as cur:
cur.execute(
"INSERT INTO device_measurements (benchmark_id, cpu_util, mem_megabytes, gpu_util, gpu_mem_megabytes) VALUES (%s, %s, %s, %s, %s)",
(benchmark_id, cpu_util, mem_megabytes, gpu_util, gpu_mem_megabytes),
)
self.logger.debug(
f"collected device measurements for benchmark #{benchmark_id} [CPU util: {cpu_util}, mem MBs: {mem_megabytes}, GPU util: {gpu_util}, GPU mem MBs: {gpu_mem_megabytes}]"
f"inserted device measurements for benchmark #{benchmark_id} [CPU util: {cpu_util}, mem MBs: {mem_megabytes}, GPU util: {gpu_util}, GPU mem MBs: {gpu_mem_megabytes}]"
)
def collect_model_measurements(self, benchmark_id: str, measurements: dict[str, float]):
# Store model measurements for CSV export (if enabled)
if self.collect_csv_data:
# Add row to pandas DataFrame with flattened measurements
row_data = {
'benchmark_id': benchmark_id,
'time': datetime.utcnow().isoformat()
}
# Flatten the measurements dict into the row
row_data.update(measurements)
new_row = pd.DataFrame([row_data])
self.model_measurements_df = pd.concat([self.model_measurements_df, new_row], ignore_index=True)
# Store in database if available
if self.use_database:
with self.conn.cursor() as cur:
cur.execute(
"""
INSERT INTO model_measurements (
benchmark_id,
measurements
) VALUES (%s, %s)
""",
(
benchmark_id,
measurements,
),
)
self.logger.debug(f"collected model measurements for benchmark #{benchmark_id}: {measurements}")
def export_to_csv(self, output_dir: str = "benchmark_results"):
"""
Export all collected data to CSV files using pandas DataFrames
"""
if not self.collect_csv_data:
self.logger.warning("CSV data collection is disabled - no CSV files will be generated")
return
if not os.path.exists(output_dir):
os.makedirs(output_dir)
self.logger.info(f"Created output directory: {output_dir}")
timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
files_created = []
# Export using pandas DataFrames
self._export_pandas_data(output_dir, timestamp, files_created)
self.logger.info(f"CSV export complete! Created {len(files_created)} files in {output_dir}")
def _export_pandas_data(self, output_dir: str, timestamp: str, files_created: list):
"""
Export CSV files using pandas DataFrames
"""
# Export benchmarks
benchmarks_file = os.path.join(output_dir, f"benchmarks_{timestamp}.csv")
self.benchmarks_df.to_csv(benchmarks_file, index=False)
files_created.append(benchmarks_file)
self.logger.info(f"Exported {len(self.benchmarks_df)} benchmark records to {benchmarks_file}")
# Export device measurements
device_file = os.path.join(output_dir, f"device_measurements_{timestamp}.csv")
self.device_measurements_df.to_csv(device_file, index=False)
files_created.append(device_file)
self.logger.info(f"Exported {len(self.device_measurements_df)} device measurement records to {device_file}")
# Export model measurements (already flattened)
model_file = os.path.join(output_dir, f"model_measurements_{timestamp}.csv")
self.model_measurements_df.to_csv(model_file, index=False)
files_created.append(model_file)
self.logger.info(f"Exported {len(self.model_measurements_df)} model measurement records to {model_file}")
# Create comprehensive summary using pandas operations
summary_file = os.path.join(output_dir, f"benchmark_summary_{timestamp}.csv")
self._create_summary(summary_file)
files_created.append(summary_file)
def _create_summary(self, summary_file: str):
"""
Create a comprehensive summary CSV using pandas operations
"""
if len(self.benchmarks_df) == 0:
# Create empty summary file
summary_df = pd.DataFrame()
summary_df.to_csv(summary_file, index=False)
self.logger.info(f"Created empty benchmark summary at {summary_file}")
return
# Start with benchmarks as the base
summary_df = self.benchmarks_df.copy()
# Add model measurements (join on benchmark_id)
if len(self.model_measurements_df) > 0:
# Drop 'time' column from model measurements to avoid conflicts
model_df = self.model_measurements_df.drop(columns=['time'], errors='ignore')
summary_df = summary_df.merge(model_df, on='benchmark_id', how='left')
# Calculate device measurement aggregates using pandas groupby
if len(self.device_measurements_df) > 0:
device_agg = self.device_measurements_df.groupby('benchmark_id').agg({
'cpu_util': ['mean', 'max', 'std', 'count'],
'mem_megabytes': ['mean', 'max', 'std'],
'gpu_util': ['mean', 'max', 'std'],
'gpu_mem_megabytes': ['mean', 'max', 'std']
}).round(3)
# Flatten column names
device_agg.columns = [f"{col[0]}_{col[1]}" for col in device_agg.columns]
device_agg = device_agg.reset_index()
# Rename count column to be more descriptive
if 'cpu_util_count' in device_agg.columns:
device_agg = device_agg.rename(columns={'cpu_util_count': 'device_measurement_count'})
# Merge with summary
summary_df = summary_df.merge(device_agg, on='benchmark_id', how='left')
# Export the comprehensive summary
summary_df.to_csv(summary_file, index=False)
self.logger.info(f"Created comprehensive benchmark summary with {len(summary_df)} records at {summary_file}")
def collect_model_measurements(self, benchmark_id: int, measurements: Dict[str, float]):
with self.conn.cursor() as cur:
cur.execute(
"""
INSERT INTO model_measurements (
benchmark_id,
measurements
) VALUES (%s, %s)
""",
(
benchmark_id,
measurements,
),
)
self.logger.debug(f"inserted model measurements for benchmark #{benchmark_id}: {measurements}")
def close(self):
if self.use_database and self.conn:
self.conn.close()
self.conn.close()
logger = logging.getLogger(__name__)
@ -283,7 +85,7 @@ handler.setFormatter(formatter)
logger.addHandler(handler)
def parse_arguments() -> tuple[str, str, str, str, bool, str]:
def parse_arguments() -> Tuple[str, str, str, str]:
"""
Parse command line arguments for the benchmarking CLI.
"""
@ -312,27 +114,10 @@ def parse_arguments() -> tuple[str, str, str, str, bool, str]:
type=str,
help="The commit message associated with the commit, truncated to 70 characters.",
)
parser.add_argument(
"--csv",
action="store_true",
default=False,
help="Enable CSV output files generation."
)
parser.add_argument(
"--csv-output-dir",
type=str,
default="benchmark_results",
help="Directory for CSV output files (default: benchmark_results)."
)
args = parser.parse_args()
# CSV is disabled by default, only enabled when --csv is used
generate_csv = args.csv
return args.repository, args.branch, args.commit_id, args.commit_msg, generate_csv, args.csv_output_dir
return args.repository, args.branch, args.commit_id, args.commit_msg
def import_from_path(module_name, file_path):
@ -346,124 +131,22 @@ def import_from_path(module_name, file_path):
raise ImportModuleException(f"failed to load python module: {e}")
def create_database_connection():
"""
Try to create a database connection. Returns None if connection fails.
"""
if not PSYCOPG2_AVAILABLE:
logger.warning("psycopg2 not available - running in CSV-only mode")
return None
try:
import psycopg2
conn = psycopg2.connect("dbname=metrics")
logger.info("Successfully connected to database")
return conn
except Exception as e:
logger.warning(f"Failed to connect to database: {e}. Running in CSV-only mode")
return None
def create_global_metrics_recorder(repository: str, branch: str, commit_id: str, commit_msg: str,
generate_csv: bool = False) -> MetricsRecorder:
"""
Create a global metrics recorder that will be used across all benchmarks.
"""
connection = create_database_connection()
recorder = MetricsRecorder(connection, logger, repository, branch, commit_id, commit_msg, generate_csv)
# Log the storage mode
storage_modes = []
if connection is not None:
storage_modes.append("database")
if generate_csv:
storage_modes.append("CSV")
if not storage_modes:
logger.warning("Running benchmarks with NO data storage (no database connection, CSV disabled)")
logger.warning("Use --csv flag to enable CSV output when database is unavailable")
else:
logger.info(f"Running benchmarks with: {' + '.join(storage_modes)} storage")
return recorder
if __name__ == "__main__":
benchmarks_folder_path = os.path.dirname(os.path.realpath(__file__))
benches_folder_path = os.path.join(benchmarks_folder_path, "benches")
repository, branch, commit_id, commit_msg, generate_csv, csv_output_dir = parse_arguments()
# Create a global metrics recorder
global_metrics_recorder = create_global_metrics_recorder(repository, branch, commit_id, commit_msg, generate_csv)
successful_benchmarks = 0
failed_benchmarks = 0
# Automatically discover all benchmark modules in benches/ folder
benchmark_modules = []
if os.path.exists(benches_folder_path):
logger.debug(f"Scanning for benchmarks in: {benches_folder_path}")
for entry in os.scandir(benches_folder_path):
repository, branch, commit_id, commit_msg = parse_arguments()
for entry in os.scandir(benchmarks_folder_path):
try:
if not entry.name.endswith(".py"):
continue
if entry.name.startswith("__"): # Skip __init__.py, __pycache__, etc.
if entry.path == __file__:
continue
# Check if the file has a run_benchmark function
try:
logger.debug(f"checking if benches/{entry.name} has run_benchmark function")
module = import_from_path(entry.name.split(".")[0], entry.path)
if hasattr(module, 'run_benchmark'):
benchmark_modules.append(entry.name)
logger.debug(f"discovered benchmark: {entry.name}")
else:
logger.debug(f"skipping {entry.name} - no run_benchmark function found")
except Exception as e:
logger.debug(f"failed to check benches/{entry.name}: {e}")
else:
logger.warning(f"Benches directory not found: {benches_folder_path}")
if benchmark_modules:
logger.info(f"Discovered {len(benchmark_modules)} benchmark(s): {benchmark_modules}")
else:
logger.warning("No benchmark modules found in benches/ directory")
for module_name in benchmark_modules:
module_path = os.path.join(benches_folder_path, module_name)
try:
logger.debug(f"loading: {module_name}")
module = import_from_path(module_name.split(".")[0], module_path)
logger.info(f"running benchmarks in: {module_name}")
# Check if the module has an updated run_benchmark function that accepts metrics_recorder
try:
# Try the new signature first
module.run_benchmark(logger, repository, branch, commit_id, commit_msg, global_metrics_recorder)
except TypeError:
# Fall back to the old signature for backward compatibility
logger.warning(f"Module {module_name} using old run_benchmark signature - database connection will be created per module")
module.run_benchmark(logger, repository, branch, commit_id, commit_msg)
successful_benchmarks += 1
logger.debug(f"loading: {entry.name}")
module = import_from_path(entry.name.split(".")[0], entry.path)
logger.info(f"running benchmarks in: {entry.name}")
module.run_benchmark(logger, repository, branch, commit_id, commit_msg)
except ImportModuleException as e:
logger.error(e)
failed_benchmarks += 1
except Exception as e:
logger.error(f"error running benchmarks for {module_name}: {e}")
failed_benchmarks += 1
# Export CSV results at the end (if enabled)
try:
if generate_csv:
global_metrics_recorder.export_to_csv(csv_output_dir)
logger.info(f"CSV reports have been generated and saved to the {csv_output_dir} directory")
else:
logger.info("CSV generation disabled - no CSV files created (use --csv to enable)")
logger.info(f"Benchmark run completed. Successful: {successful_benchmarks}, Failed: {failed_benchmarks}")
except Exception as e:
logger.error(f"Failed to export CSV results: {e}")
finally:
global_metrics_recorder.close()
logger.error(f"error running benchmarks for {entry.name}: {e}")

34
benchmark/init_db.sql Normal file
View File

@ -0,0 +1,34 @@
CREATE TABLE IF NOT EXISTS benchmarks (
benchmark_id SERIAL PRIMARY KEY,
repository VARCHAR(255),
branch VARCHAR(255),
commit_id VARCHAR(72),
commit_message VARCHAR(70),
metadata jsonb,
created_at timestamp without time zone NOT NULL DEFAULT (current_timestamp AT TIME ZONE 'UTC')
);
CREATE INDEX IF NOT EXISTS benchmarks_benchmark_id_idx ON benchmarks (benchmark_id);
CREATE INDEX IF NOT EXISTS benchmarks_branch_idx ON benchmarks (branch);
CREATE TABLE IF NOT EXISTS device_measurements (
measurement_id SERIAL PRIMARY KEY,
benchmark_id int REFERENCES benchmarks (benchmark_id),
cpu_util double precision,
mem_megabytes double precision,
gpu_util double precision,
gpu_mem_megabytes double precision,
time timestamp without time zone NOT NULL DEFAULT (current_timestamp AT TIME ZONE 'UTC')
);
CREATE INDEX IF NOT EXISTS device_measurements_branch_idx ON device_measurements (benchmark_id);
CREATE TABLE IF NOT EXISTS model_measurements (
measurement_id SERIAL PRIMARY KEY,
benchmark_id int REFERENCES benchmarks (benchmark_id),
measurements jsonb,
time timestamp without time zone NOT NULL DEFAULT (current_timestamp AT TIME ZONE 'UTC')
);
CREATE INDEX IF NOT EXISTS model_measurements_branch_idx ON model_measurements (benchmark_id);

346
benchmark/llama.py Normal file
View File

@ -0,0 +1,346 @@
from logging import Logger
import os
from threading import Event, Thread
from time import perf_counter, sleep
from typing import Optional
from benchmarks_entrypoint import MetricsRecorder
import gpustat
import psutil
import psycopg2
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, GenerationConfig, StaticCache
os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "1"
os.environ["TOKENIZERS_PARALLELISM"] = "1"
torch.set_float32_matmul_precision("high")
def collect_metrics(benchmark_id, continue_metric_collection, metrics_recorder):
p = psutil.Process(os.getpid())
while not continue_metric_collection.is_set():
with p.oneshot():
cpu_util = p.cpu_percent()
mem_megabytes = p.memory_info().rss / (1024 * 1024)
gpu_stats = gpustat.GPUStatCollection.new_query()
gpu_util = gpu_stats[0]["utilization.gpu"]
gpu_mem_megabytes = gpu_stats[0]["memory.used"]
metrics_recorder.collect_device_measurements(
benchmark_id, cpu_util, mem_megabytes, gpu_util, gpu_mem_megabytes
)
sleep(0.01)
def run_benchmark(
logger: Logger, repository: str, branch: str, commit_id: str, commit_msg: str, num_tokens_to_generate=100
):
continue_metric_collection = Event()
metrics_thread = None
model_id = "meta-llama/Llama-2-7b-hf"
metrics_recorder = MetricsRecorder(
psycopg2.connect("dbname=metrics"), logger, repository, branch, commit_id, commit_msg
)
try:
gpu_stats = gpustat.GPUStatCollection.new_query()
gpu_name = gpu_stats[0]["name"]
benchmark_id = metrics_recorder.initialise_benchmark({"gpu_name": gpu_name, "model_id": model_id})
logger.info(f"running benchmark #{benchmark_id} on {gpu_name} for {model_id}")
metrics_thread = Thread(
target=collect_metrics,
args=[benchmark_id, continue_metric_collection, metrics_recorder],
)
metrics_thread.start()
logger.info("started background thread to fetch device metrics")
os.environ["TOKENIZERS_PARALLELISM"] = "false" # silence warnings when compiling
device = "cuda"
logger.info("downloading weights")
# This is to avoid counting download in model load time measurement
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.float16)
gen_config = GenerationConfig(do_sample=False, top_p=1, temperature=1)
logger.info("loading model")
start = perf_counter()
model = AutoModelForCausalLM.from_pretrained(
model_id, torch_dtype=torch.float16, generation_config=gen_config
).eval()
model.to(device)
torch.cuda.synchronize()
end = perf_counter()
model_load_time = end - start
logger.info(f"loaded model in: {model_load_time}s")
tokenizer = AutoTokenizer.from_pretrained(model_id)
prompt = "Why dogs are so cute?"
inputs = tokenizer(prompt, return_tensors="pt").to(device)
# Specify the max length (including both the prompt and the response)
# When calling `generate` with `cache_implementation="static" later, this is also used to create a `StaticCache` object
# with sequence length = `max_length`. The longer the more you will re-use it
seq_length = inputs["input_ids"].shape[1]
model.generation_config.max_length = seq_length + num_tokens_to_generate
batch_size = inputs["input_ids"].shape[0]
# Copied from the gpt-fast repo
def multinomial_sample_one_no_sync(probs_sort): # Does multinomial sampling without a cuda synchronization
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: Optional[int] = None):
logits = logits / max(temperature, 1e-5)
if top_k is not None:
v, _ = torch.topk(logits, min(top_k, logits.size(-1)))
pivot = v.select(-1, -1).unsqueeze(-1)
logits = torch.where(logits < pivot, -float("Inf"), logits)
probs = torch.nn.functional.softmax(logits, dim=-1)
return probs
def sample(logits, temperature: float = 1.0, top_k: Optional[int] = None):
probs = logits_to_probs(logits[:, -1], temperature, top_k)
idx_next = multinomial_sample_one_no_sync(probs)
return idx_next, probs
def decode_one_token(model, cur_token, cache_position, past_key_values):
logits = model(
cur_token,
cache_position=cache_position,
past_key_values=past_key_values,
return_dict=False,
use_cache=True,
)[0]
new_token = sample(logits, temperature=0.6, top_k=5)[0]
return new_token
#########
# Eager #
#########
with torch.no_grad():
past_key_values = StaticCache(
model.config,
max_batch_size=batch_size,
device=device,
dtype=torch.float16,
max_cache_len=seq_length + num_tokens_to_generate,
)
cache_position = torch.arange(seq_length, device=device)
start = perf_counter()
model(
**inputs,
cache_position=cache_position,
past_key_values=past_key_values,
return_dict=False,
use_cache=True,
)
end = perf_counter()
first_eager_fwd_pass_time = end - start
logger.info(f"completed first eager fwd pass in: {first_eager_fwd_pass_time}s")
start = perf_counter()
output = model.generate(**inputs, do_sample=False)
end = perf_counter()
first_eager_generate_time = end - start
logger.info(f"completed first eager generation in: {first_eager_generate_time}s")
logger.info(f"generated: {tokenizer.batch_decode(output.cpu().tolist())}")
past_key_values = StaticCache(
model.config,
max_batch_size=batch_size,
device=device,
dtype=torch.float16,
max_cache_len=seq_length + num_tokens_to_generate,
)
cache_position = torch.arange(seq_length, device=device)
start = perf_counter()
model(
**inputs,
cache_position=cache_position,
past_key_values=past_key_values,
return_dict=False,
use_cache=True,
)
end = perf_counter()
second_eager_fwd_pass_time = end - start
logger.info(f"completed second eager fwd pass in: {second_eager_fwd_pass_time}s")
start = perf_counter()
model.generate(**inputs, do_sample=False)
end = perf_counter()
second_eager_generate_time = end - start
logger.info(f"completed second eager generation in: {second_eager_generate_time}s")
logger.info(f"generated: {tokenizer.batch_decode(output.cpu().tolist())}")
torch.compiler.reset()
################
# Forward pass #
################
# `torch.compile(model, ...)` is not recommended as you compile callbacks
# and full generate. We recommend compiling only the forward for now.
# "reduce-overhead" will use cudagraphs.
generated_ids = torch.zeros(
(batch_size, num_tokens_to_generate + seq_length), dtype=torch.int, device=device
)
generated_ids[:, :seq_length] = inputs["input_ids"]
decode_one_token = torch.compile(decode_one_token, mode="reduce-overhead", fullgraph=True)
# model.forward = torch.compile(model.forward, mode="reduce-overhead", fullgraph=True)
# TODO use decode_one_token(model, input_id.clone(), cache_position) for verification
past_key_values = StaticCache(
model.config,
max_batch_size=batch_size,
device=device,
dtype=torch.float16,
max_cache_len=seq_length + num_tokens_to_generate + 10,
)
cache_position = torch.arange(seq_length, device=device)
all_generated_tokens = []
### First compile, prefill
start = perf_counter()
next_token = decode_one_token(
model, inputs["input_ids"], cache_position=cache_position, past_key_values=past_key_values
)
torch.cuda.synchronize()
end = perf_counter()
time_to_first_token = end - start
logger.info(f"completed first compile generation in: {time_to_first_token}s")
cache_position += 1
all_generated_tokens += next_token.tolist()
cache_position = torch.tensor([seq_length], device=device)
### First compile, decoding
start = perf_counter()
next_token = decode_one_token(
model, next_token.clone(), cache_position=cache_position, past_key_values=past_key_values
)
torch.cuda.synchronize()
end = perf_counter()
time_to_second_token = end - start
logger.info(f"completed second compile generation in: {time_to_second_token}s")
cache_position += 1
all_generated_tokens += next_token.tolist()
### Second compile, decoding
start = perf_counter()
next_token = decode_one_token(
model, next_token.clone(), cache_position=cache_position, past_key_values=past_key_values
)
torch.cuda.synchronize()
end = perf_counter()
time_to_third_token = end - start
logger.info(f"completed third compile forward in: {time_to_third_token}s")
cache_position += 1
all_generated_tokens += next_token.tolist()
### Using cuda graphs decoding
start = perf_counter()
for _ in range(1, num_tokens_to_generate):
all_generated_tokens += next_token.tolist()
next_token = decode_one_token(
model, next_token.clone(), cache_position=cache_position, past_key_values=past_key_values
)
cache_position += 1
torch.cuda.synchronize()
end = perf_counter()
mean_time_to_next_token = (end - start) / num_tokens_to_generate
logger.info(f"completed next compile generation in: {mean_time_to_next_token}s")
logger.info(f"generated: {tokenizer.batch_decode(all_generated_tokens)}")
####################
# Generate compile #
####################
torch.compiler.reset()
# we will not compile full generate as it' s to intensive, tho we measure full forward!
past_key_values = StaticCache(
model.config,
max_batch_size=batch_size,
device=device,
dtype=torch.float16,
max_cache_len=seq_length + 128,
)
# 1st call
start = perf_counter()
output = model.generate(**inputs, past_key_values=past_key_values)
torch.cuda.synchronize()
end = perf_counter()
first_compile_generate_time = end - start
logger.info(f"completed first compile generation in: {first_compile_generate_time}s")
logger.info(f"generated: {tokenizer.batch_decode(output.cpu().tolist())}")
past_key_values = StaticCache(
model.config,
max_batch_size=batch_size,
device=device,
dtype=torch.float16,
max_cache_len=seq_length + 128,
)
# 2nd call
start = perf_counter()
output = model.generate(**inputs, past_key_values=past_key_values)
torch.cuda.synchronize()
end = perf_counter()
second_compile_generate_time = end - start
logger.info(f"completed second compile generation in: {second_compile_generate_time}s")
logger.info(f"generated: {tokenizer.batch_decode(output.cpu().tolist())}")
past_key_values = StaticCache(
model.config,
max_batch_size=batch_size,
device=device,
dtype=torch.float16,
max_cache_len=seq_length + 128,
)
# 3rd call
start = perf_counter()
output = model.generate(**inputs, past_key_values=past_key_values)
end = perf_counter()
third_compile_generate_time = end - start
logger.info(f"completed third compile generation in: {third_compile_generate_time}s")
logger.info(f"generated: {tokenizer.batch_decode(output.cpu().tolist())}")
past_key_values = StaticCache(
model.config,
max_batch_size=batch_size,
device=device,
dtype=torch.float16,
max_cache_len=seq_length + 128,
)
# 4th call
start = perf_counter()
output = model.generate(**inputs, past_key_values=past_key_values)
end = perf_counter()
fourth_compile_generate_time = end - start
logger.info(f"completed fourth compile generation in: {fourth_compile_generate_time}s")
logger.info(f"generated: {tokenizer.batch_decode(output.cpu().tolist())}")
metrics_recorder.collect_model_measurements(
benchmark_id,
{
"model_load_time": model_load_time,
"first_eager_forward_pass_time_secs": first_eager_fwd_pass_time,
"second_eager_forward_pass_time_secs": second_eager_fwd_pass_time,
"first_eager_generate_time_secs": first_eager_generate_time,
"second_eager_generate_time_secs": second_eager_generate_time,
"time_to_first_token_secs": time_to_first_token,
"time_to_second_token_secs": time_to_second_token,
"time_to_third_token_secs": time_to_third_token,
"time_to_next_token_mean_secs": mean_time_to_next_token,
"first_compile_generate_time_secs": first_compile_generate_time,
"second_compile_generate_time_secs": second_compile_generate_time,
"third_compile_generate_time_secs": third_compile_generate_time,
"fourth_compile_generate_time_secs": fourth_compile_generate_time,
},
)
except Exception as e:
logger.error(f"Caught exception: {e}")
continue_metric_collection.set()
if metrics_thread is not None:
metrics_thread.join()
metrics_recorder.close()

View File

@ -2,5 +2,4 @@ gpustat==1.1.1
psutil==6.0.0
psycopg2==2.9.9
torch>=2.4.0
hf_transfer
pandas>=1.5.0
hf_transfer

View File

@ -23,12 +23,12 @@ from os.path import abspath, dirname, join
import _pytest
import pytest
from transformers.testing_utils import HfDoctestModule, HfDocTestParser, is_torch_available
from transformers.testing_utils import HfDoctestModule, HfDocTestParser
NOT_DEVICE_TESTS = {
"test_tokenization",
"test_tokenization_mistral_common",
"test_processor",
"test_processing",
"test_beam_constraints",
"test_configuration_utils",
@ -83,8 +83,6 @@ def pytest_configure(config):
config.addinivalue_line("markers", "is_staging_test: mark test to run only in the staging environment")
config.addinivalue_line("markers", "accelerate_tests: mark test that require accelerate")
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")
def pytest_collection_modifyitems(items):
@ -129,10 +127,3 @@ class CustomOutputChecker(OutputChecker):
doctest.OutputChecker = CustomOutputChecker
_pytest.doctest.DoctestModule = HfDoctestModule
doctest.DocTestParser = HfDocTestParser
if is_torch_available():
import torch
# The flag below controls whether to allow TF32 on cuDNN. This flag defaults to True.
# We set it to `False` for CI. See https://github.com/pytorch/pytorch/issues/157274#issuecomment-3090791615
torch.backends.cudnn.allow_tf32 = False

View File

@ -4,7 +4,7 @@ USER root
ARG REF=main
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 pip install uv && uv venv && 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

View File

@ -4,7 +4,7 @@ ARG REF=main
USER root
RUN apt-get update && apt-get install -y libsndfile1-dev espeak-ng time git cmake wget xz-utils build-essential g++5 libprotobuf-dev protobuf-compiler
ENV UV_PYTHON=/usr/local/bin/python
RUN pip --no-cache-dir install uv && uv pip install --no-cache-dir -U pip setuptools
RUN pip --no-cache-dir install uv && uv venv && uv pip install --no-cache-dir -U pip setuptools
RUN wget https://github.com/ku-nlp/jumanpp/releases/download/v2.0.0-rc3/jumanpp-2.0.0-rc3.tar.xz
RUN tar xvf jumanpp-2.0.0-rc3.tar.xz
@ -20,7 +20,7 @@ RUN uv pip install --no-cache --upgrade 'torch' --index-url https://download.pyt
RUN uv pip install --no-cache-dir --no-deps accelerate --extra-index-url https://download.pytorch.org/whl/cpu
RUN uv pip install --no-cache-dir "git+https://github.com/huggingface/transformers.git@${REF}#egg=transformers[ja,testing,sentencepiece,jieba,spacy,ftfy,rjieba]" unidic unidic-lite
# spacy is not used so not tested. Causes to failures. TODO fix later
RUN uv run python -m unidic download
RUN python3 -m unidic download
RUN uv pip uninstall transformers
RUN apt-get clean && rm -rf /var/lib/apt/lists/*

View File

@ -5,7 +5,7 @@ USER root
RUN apt-get update && apt-get install -y libsndfile1-dev espeak-ng time git
RUN apt-get install -y g++ cmake
ENV UV_PYTHON=/usr/local/bin/python
RUN pip --no-cache-dir install uv
RUN pip --no-cache-dir install uv && uv venv
RUN uv pip install --no-cache-dir -U pip setuptools albumentations seqeval
RUN uv pip install --upgrade --no-cache-dir "git+https://github.com/huggingface/transformers.git@${REF}#egg=transformers[tf-cpu,sklearn,testing,sentencepiece,tf-speech,vision]"
RUN uv pip install --no-cache-dir "protobuf==3.20.3"

View File

@ -2,10 +2,10 @@ FROM python:3.9-slim
ENV PYTHONDONTWRITEBYTECODE=1
ARG REF=main
USER root
RUN apt-get update && apt-get install -y --no-install-recommends libsndfile1-dev espeak-ng time git g++ cmake pkg-config openssh-client git ffmpeg
RUN apt-get update && apt-get install -y --no-install-recommends libsndfile1-dev espeak-ng time git g++ cmake pkg-config openssh-client git
ENV UV_PYTHON=/usr/local/bin/python
RUN pip --no-cache-dir install uv && uv pip install --no-cache-dir -U pip setuptools
RUN uv pip install --no-cache-dir 'torch' 'torchaudio' 'torchvision' 'torchcodec' --index-url https://download.pytorch.org/whl/cpu
RUN pip --no-cache-dir install uv && uv venv && uv pip install --no-cache-dir -U pip setuptools
RUN uv pip install --no-cache-dir 'torch' 'torchaudio' 'torchvision' --index-url https://download.pytorch.org/whl/cpu
RUN uv pip install --no-deps timm accelerate --extra-index-url https://download.pytorch.org/whl/cpu
RUN uv pip install --no-cache-dir librosa "git+https://github.com/huggingface/transformers.git@${REF}#egg=transformers[sklearn,sentencepiece,vision,testing]" seqeval albumentations jiwer
RUN uv pip uninstall transformers

View File

@ -4,10 +4,10 @@ ARG REF=main
USER root
RUN apt-get update && apt-get install -y libsndfile1-dev espeak-ng time git libgl1-mesa-glx libgl1 g++ tesseract-ocr
ENV UV_PYTHON=/usr/local/bin/python
RUN pip --no-cache-dir install uv && uv pip install --no-cache-dir -U pip setuptools
RUN pip --no-cache-dir install uv && uv venv && uv pip install --no-cache-dir -U pip setuptools
RUN uv pip install --no-cache-dir 'torch' 'torchaudio' 'torchvision' --index-url https://download.pytorch.org/whl/cpu
RUN uv pip install --no-cache-dir --no-deps timm accelerate
RUN uv pip install -U --upgrade-strategy eager --no-cache-dir pytesseract python-Levenshtein opencv-python nltk
RUN pip install -U --upgrade-strategy eager --no-cache-dir pytesseract python-Levenshtein opencv-python nltk
# RUN uv pip install --no-cache-dir natten==0.15.1+torch210cpu -f https://shi-labs.com/natten/wheels
RUN uv pip install --no-cache-dir "git+https://github.com/huggingface/transformers.git@${REF}#egg=transformers[testing, vision]" 'scikit-learn' 'torch-stft' 'nose' 'dataset'
# RUN git clone https://github.com/facebookresearch/detectron2.git

View File

@ -4,7 +4,7 @@ ARG REF=main
USER root
RUN apt-get update && apt-get install -y libsndfile1-dev espeak-ng time git g++ cmake
ENV UV_PYTHON=/usr/local/bin/python
RUN pip --no-cache-dir install uv && uv pip install --no-cache-dir -U pip setuptools
RUN pip --no-cache-dir install uv && uv venv && uv pip install --no-cache-dir -U pip setuptools
RUN uv pip install --no-cache-dir "scipy<1.13" "git+https://github.com/huggingface/transformers.git@${REF}#egg=transformers[flax,testing,sentencepiece,flax-speech,vision]"
RUN uv pip uninstall transformers
RUN apt-get clean && rm -rf /var/lib/apt/lists/* && apt-get autoremove && apt-get autoclean

View File

@ -4,7 +4,7 @@ ARG REF=main
USER root
RUN apt-get update && apt-get install -y libsndfile1-dev espeak-ng time git cmake g++
ENV UV_PYTHON=/usr/local/bin/python
RUN pip --no-cache-dir install uv && uv pip install --no-cache-dir -U pip setuptools
RUN pip --no-cache-dir install uv && uv venv && uv pip install --no-cache-dir -U pip setuptools
RUN uv pip install --no-cache-dir "git+https://github.com/huggingface/transformers.git@${REF}#egg=transformers[sklearn,tf-cpu,testing,sentencepiece,tf-speech,vision]"
RUN uv pip install --no-cache-dir "protobuf==3.20.3" tensorflow_probability
RUN apt-get clean && rm -rf /var/lib/apt/lists/*

View File

@ -2,10 +2,10 @@ FROM python:3.9-slim
ENV PYTHONDONTWRITEBYTECODE=1
ARG REF=main
USER root
RUN apt-get update && apt-get install -y --no-install-recommends libsndfile1-dev espeak-ng time git pkg-config openssh-client git ffmpeg
RUN apt-get update && apt-get install -y --no-install-recommends libsndfile1-dev espeak-ng time git pkg-config openssh-client git
ENV UV_PYTHON=/usr/local/bin/python
RUN pip --no-cache-dir install uv && uv pip install --no-cache-dir -U pip setuptools
RUN uv pip install --no-cache-dir 'torch' 'torchaudio' 'torchvision' 'torchcodec' --index-url https://download.pytorch.org/whl/cpu
RUN pip --no-cache-dir install uv && uv venv && uv pip install --no-cache-dir -U pip setuptools
RUN uv pip install --no-cache-dir 'torch' 'torchaudio' 'torchvision' --index-url https://download.pytorch.org/whl/cpu
RUN uv pip install --no-deps timm accelerate --extra-index-url https://download.pytorch.org/whl/cpu
RUN uv pip install --no-cache-dir librosa "git+https://github.com/huggingface/transformers.git@${REF}#egg=transformers[sklearn,sentencepiece,vision,testing]"
RUN uv pip uninstall transformers

View File

@ -2,8 +2,8 @@ FROM python:3.9-slim
ENV PYTHONDONTWRITEBYTECODE=1
ARG REF=main
USER root
RUN apt-get update && apt-get install -y time git
RUN apt-get update && apt-get install -y time git
ENV UV_PYTHON=/usr/local/bin/python
RUN pip install uv
RUN pip install uv && uv venv
RUN uv pip install --no-cache-dir -U pip setuptools GitPython "git+https://github.com/huggingface/transformers.git@${REF}#egg=transformers[ruff]" urllib3
RUN apt-get install -y jq curl && apt-get clean && rm -rf /var/lib/apt/lists/*

View File

@ -5,7 +5,7 @@ USER root
RUN apt-get update && apt-get install -y --no-install-recommends libsndfile1-dev espeak-ng time git g++ pkg-config openssh-client git
RUN apt-get install -y cmake
ENV UV_PYTHON=/usr/local/bin/python
RUN pip --no-cache-dir install uv && uv pip install --no-cache-dir -U pip setuptools
RUN pip --no-cache-dir install uv && uv venv && uv pip install --no-cache-dir -U pip setuptools
RUN uv pip install --upgrade --no-cache-dir "git+https://github.com/huggingface/transformers.git@${REF}#egg=transformers[tf-cpu,sklearn,testing,sentencepiece,tf-speech,vision]"
RUN uv pip install --no-cache-dir "protobuf==3.20.3"
RUN uv pip uninstall transformers

View File

@ -4,7 +4,7 @@ ARG REF=main
USER root
RUN apt-get update && apt-get install -y libsndfile1-dev espeak-ng time git g++ cmake pkg-config openssh-client git
ENV UV_PYTHON=/usr/local/bin/python
RUN pip --no-cache-dir install uv && uv pip install --no-cache-dir -U pip setuptools
RUN pip --no-cache-dir install uv && uv venv && uv pip install --no-cache-dir -U pip setuptools
RUN uv pip install --no-deps accelerate
RUN uv pip install --no-cache-dir 'torch' 'torchvision' 'torchaudio' --index-url https://download.pytorch.org/whl/cpu
RUN uv pip install --no-cache-dir "scipy<1.13" "git+https://github.com/huggingface/transformers.git@${REF}#egg=transformers[flax,audio,sklearn,sentencepiece,vision,testing]"

View File

@ -2,10 +2,10 @@ FROM python:3.9-slim
ENV PYTHONDONTWRITEBYTECODE=1
ARG REF=main
USER root
RUN apt-get update && apt-get install -y --no-install-recommends libsndfile1-dev espeak-ng time git g++ cmake pkg-config openssh-client git git-lfs ffmpeg
RUN apt-get update && apt-get install -y --no-install-recommends libsndfile1-dev espeak-ng time git g++ cmake pkg-config openssh-client git git-lfs
ENV UV_PYTHON=/usr/local/bin/python
RUN pip --no-cache-dir install uv && uv pip install --no-cache-dir -U pip setuptools
RUN uv pip install --no-cache-dir 'torch' 'torchaudio' 'torchvision' 'torchcodec' --index-url https://download.pytorch.org/whl/cpu
RUN pip --no-cache-dir install uv && uv venv && uv pip install --no-cache-dir -U pip setuptools
RUN uv pip install --no-cache-dir 'torch' 'torchaudio' 'torchvision' --index-url https://download.pytorch.org/whl/cpu
RUN uv pip install --no-deps timm accelerate --extra-index-url https://download.pytorch.org/whl/cpu
RUN uv pip install --no-cache-dir librosa "git+https://github.com/huggingface/transformers.git@${REF}#egg=transformers[sklearn,sentencepiece,vision,testing,tiktoken,num2words,video]"
RUN uv pip uninstall transformers

View File

@ -5,7 +5,7 @@ RUN echo ${REF}
USER root
RUN apt-get update && apt-get install -y --no-install-recommends libsndfile1-dev espeak-ng time git g++ cmake pkg-config openssh-client git git-lfs
ENV UV_PYTHON=/usr/local/bin/python
RUN pip --no-cache-dir install uv && uv pip install --no-cache-dir -U pip setuptools
RUN pip --no-cache-dir install uv && uv venv && uv pip install --no-cache-dir -U pip setuptools
RUN uv pip install --no-cache-dir --no-deps accelerate --extra-index-url https://download.pytorch.org/whl/cpu
RUN uv pip install --no-cache-dir 'torch' 'torchaudio' 'torchvision' --index-url https://download.pytorch.org/whl/cpu
RUN git lfs install

View File

@ -1,4 +1,4 @@
FROM nvidia/cuda:12.6.0-cudnn-devel-ubuntu22.04
FROM nvidia/cuda:12.1.0-cudnn8-devel-ubuntu22.04
LABEL maintainer="Hugging Face"
ARG DEBIAN_FRONTEND=noninteractive
@ -9,9 +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='2.8.0'
ARG PYTORCH='2.6.0'
# (not always a valid torch version)
ARG INTEL_TORCH_EXT='2.3.0'
# Example: `cu102`, `cu113`, etc.
ARG CUDA='cu126'
ARG CUDA='cu121'
# Disable kernel mapping for now until all tests pass
ENV DISABLE_KERNEL_MAPPING=1
@ -26,11 +28,11 @@ RUN git clone https://github.com/huggingface/transformers && cd transformers &&
# 1. Put several commands in a single `RUN` to avoid image/layer exporting issue. Could be revised in the future.
# 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 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 $VERSION torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/$CUDA || python3 -m pip install --no-cache-dir -U --pre torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/nightly/$CUDA && python3 -m pip uninstall -y tensorflow tensorflow_text tensorflow_probability
RUN python3 -m pip uninstall -y flax jax
RUN python3 -m pip install --no-cache-dir -U timm
RUN python3 -m pip install --no-cache-dir intel_extension_for_pytorch==$INTEL_TORCH_EXT -f https://developer.intel.com/ipex-whl-stable-cpu
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"

View File

@ -1,4 +1,4 @@
FROM rocm/pytorch:rocm6.4.1_ubuntu24.04_py3.12_pytorch_release_2.7.1
FROM rocm/pytorch:rocm6.4_ubuntu22.04_py3.10_pytorch_release_2.6.0
LABEL maintainer="Hugging Face"
ARG DEBIAN_FRONTEND=noninteractive
@ -20,12 +20,8 @@ WORKDIR /
ADD https://api.github.com/repos/huggingface/transformers/git/refs/heads/main version.json
RUN git clone https://github.com/huggingface/transformers && cd transformers && git checkout $REF
# On ROCm, torchcodec is required to decode audio files
# RUN python3 -m pip install --no-cache-dir torchcodec
# Install transformers
RUN python3 -m pip install --no-cache-dir -e ./transformers[dev-torch,testing,video,audio]
RUN python3 -m pip install --no-cache-dir -e ./transformers[dev-torch,testing,video]
# 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.
@ -36,4 +32,4 @@ RUN cd transformers && python3 setup.py develop
RUN python3 -m pip uninstall py3nvml pynvml nvidia-ml-py apex -y
# `kernels` may causes many failing tests
RUN python3 -m pip uninstall -y kernels
RUN python3 -m pip uninstall -y kernels

View File

@ -4,7 +4,7 @@ LABEL maintainer="Hugging Face"
ARG DEBIAN_FRONTEND=noninteractive
ARG PYTORCH='2.8.0'
ARG PYTORCH='2.6.0'
# Example: `cu102`, `cu113`, etc.
ARG CUDA='cu126'
@ -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 && 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 uninstall -y torch torchvision torchaudio && 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 install --no-cache-dir git+https://github.com/huggingface/accelerate@main#egg=accelerate

View File

@ -19,7 +19,7 @@ RUN python3 -m pip uninstall -y torch torchvision torchaudio
# Install **nightly** release PyTorch (flag `--pre`)
# (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 install --no-cache-dir -U --pre torch torchvision torchaudio torchcodec --extra-index-url https://download.pytorch.org/whl/nightly/$CUDA
RUN python3 -m pip install --no-cache-dir -U --pre torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/nightly/$CUDA
# `datasets` requires pandas, pandas has some modules compiled with numpy=1.x causing errors
RUN python3 -m pip install --no-cache-dir './transformers[deepspeed-testing]' 'pandas<2' 'numpy<2'

View File

@ -1,4 +1,4 @@
FROM nvidia/cuda:12.6.0-cudnn-devel-ubuntu22.04
FROM nvidia/cuda:12.1.0-cudnn8-devel-ubuntu22.04
LABEL maintainer="Hugging Face"
ARG DEBIAN_FRONTEND=noninteractive
@ -11,20 +11,18 @@ ARG REF=main
RUN git clone https://github.com/huggingface/transformers && cd transformers && git checkout $REF
# If set to nothing, will install the latest version
ARG PYTORCH='2.8.0'
ARG PYTORCH='2.6.0'
ARG TORCH_VISION=''
ARG TORCH_AUDIO=''
# Example: `cu102`, `cu113`, etc.
ARG CUDA='cu126'
ARG CUDA='cu121'
RUN python3 -m pip install --no-cache-dir -e ./transformers[dev-torch,testing,video]
# Install torch stuff after ./transformers[dev-torch,testing,video], otherwise torch may be resolved to a previous
# version.
RUN [ ${#PYTORCH} -gt 0 ] && VERSION='torch=='$PYTORCH'.*' || VERSION='torch'; python3 -m pip install --no-cache-dir -U $VERSION --extra-index-url https://download.pytorch.org/whl/$CUDA
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 install --no-cache-dir -e ./transformers[dev-torch,testing,video]
RUN python3 -m pip uninstall -y tensorflow flax
RUN python3 -m pip install --no-cache-dir git+https://github.com/facebookresearch/detectron2.git pytesseract

View File

@ -1,93 +0,0 @@
FROM intel/deep-learning-essentials:2025.1.3-0-devel-ubuntu22.04 AS base
LABEL maintainer="Hugging Face"
SHELL ["/bin/bash", "-c"]
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 && \
apt-get update && \
apt-get install -y python$PYTHON_VER python$PYTHON_VER-dev python3-pip && \
ln -sf /usr/bin/python$PYTHON_VER /usr/bin/python3 && \
ln -sf /usr/bin/python3 /usr/bin/python && \
apt-get clean && \
rm -rf /var/lib/apt/lists/*
RUN apt-get update && \
apt-get -y install \
apt-utils \
build-essential \
ca-certificates \
clinfo \
curl \
git \
git-lfs \
vim \
numactl \
gnupg2 \
gpg-agent \
zlib1g-dev \
rsync \
sudo \
libnl-genl-3-200 \
xpu-smi \
unzip \
ffmpeg \
tesseract-ocr \
espeak-ng \
wget \
ncurses-term && \
apt-get clean && \
rm -rf /var/lib/apt/lists/*
RUN apt-get update && \
apt-get install -y \
linux-headers-$(uname -r) \
linux-modules-extra-$(uname -r) \
flex bison \
intel-fw-gpu intel-i915-dkms xpu-smi \
intel-opencl-icd libze-intel-gpu1 libze1 \
intel-media-va-driver-non-free libmfx-gen1 libvpl2 \
libegl-mesa0 libegl1-mesa libegl1-mesa-dev libgbm1 libgl1-mesa-dev libgl1-mesa-dri \
libglapi-mesa libglx-mesa0 libigdgmm12 libxatracker2 mesa-va-drivers \
mesa-vdpau-drivers mesa-vulkan-drivers va-driver-all vainfo hwinfo clinfo intel-ocloc \
libigc-dev intel-igc-cm libigdfcl-dev libigfxcmrt-dev libze-dev && \
apt-get clean && \
rm -rf /var/lib/apt/lists/*
RUN pip install --upgrade pip
RUN pip install triton==3.3.0
RUN pip install torch==2.7.0 torchvision==0.22.0 torchaudio==2.7.0 --index-url https://download.pytorch.org/whl/xpu --no-cache-dir
RUN pip install evaluate torchdata pyctcdecode pytesseract decord galore-torch fire scipy scikit-learn sentencepiece sacremoses nltk rouge_score librosa soundfile g2p_en mpi4py requests_mock
RUN pip install pretty_midi essentia resampy Levenshtein av sacrebleu phonemizer invisible_watermark schedulefree
RUN pip install gguf hqq compressed_tensors gptqmodel mergekit autoawq deepspeed torchao onnx
RUN pip install hf_transfer huggingface-hub hf-doc-builder datasets optimum-quanto timm transformers accelerate optimum peft
RUN pip install git+https://github.com/linkedin/Liger-Kernel.git --extra-index-url https://download.pytorch.org/whl/test/xpu
# install bitsandbytes
RUN pip install git+https://github.com/bitsandbytes-foundation/bitsandbytes.git
ENV OCL_ICD_VENDORS=/etc/OpenCL/vendors
ENV FI_PROVIDER_PATH=${I_MPI_ROOT}/lib/libfabric/prov:/usr/lib/x86_64-linux-gnu/libfabric
ENV CCL_ROOT=/usr/local
ENV CCL_ATL_TRANSPORT=ofi
ENV I_MPI_ROOT=/usr/local
ENV CLASSPATH=${I_MPI_ROOT}/lib/mpi.jar
ENV PATH=${I_MPI_ROOT}/bin/libfabric:${PATH}
ENV LD_LIBRARY_PATH=${I_MPI_ROOT}/lib/libfabric:${LD_LIBRARY_PATH}
RUN touch /entrypoint.sh
RUN chmod +x /entrypoint.sh
RUN echo "#!/bin/bash" >> /entrypoint.sh
RUN echo "source /opt/intel/oneapi/setvars.sh --force && /bin/bash" >> /entrypoint.sh
ENTRYPOINT ["/entrypoint.sh"]

View File

@ -26,7 +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 $VERSION torchvision torchaudio torchcodec --extra-index-url https://download.pytorch.org/whl/$CUDA
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 git+https://github.com/huggingface/accelerate@main#egg=accelerate
@ -78,10 +78,6 @@ RUN git clone https://github.com/NetEase-FuXi/EETQ.git && cd EETQ/ && git submod
# RUN python3 -m pip install --no-cache-dir flute-kernel==0.4.1
# RUN python3 -m pip install --no-cache-dir git+https://github.com/Dao-AILab/fast-hadamard-transform.git
# Add fp-quant for quantization testing
# Requires py3.11 but our CI runs on 3.9
# RUN python3 -m pip install --no-cache-dir "fp-quant>=0.1.6"
# Add compressed-tensors for quantization testing
RUN python3 -m pip install --no-cache-dir compressed-tensors
@ -97,9 +93,6 @@ RUN python3 -m pip install --no-cache-dir -e ./transformers[dev-torch]
# `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
# Uninstall flash-attn installed by autoawq, it causes issues here : https://github.com/huggingface/transformers/actions/runs/15915442841/job/44892146131
RUN python3 -m pip uninstall -y flash-attn
# 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

View File

@ -20,21 +20,22 @@ To generate the documentation, you first have to build it. Several packages are
you can install them with the following command, at the root of the code repository:
```bash
pip install -e ".[dev]"
pip install -e ".[docs]"
```
> [!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 workaround it.
Then you need to install our special tool that builds the documentation:
```bash
pip install git+https://github.com/huggingface/doc-builder
```
> [!NOTE]
> You only need to generate the documentation to inspect it locally (if you're planning changes and want to
> check how they look before committing for instance). You don't have to commit the built documentation.
---
**NOTE**
You only need to generate the documentation to inspect it locally (if you're planning changes and want to
check how they look before committing for instance). You don't have to commit the built documentation.
---
## Building the documentation
@ -71,8 +72,12 @@ doc-builder preview transformers docs/source/en/
The docs will be viewable at [http://localhost:3000](http://localhost:3000). You can also preview the docs once you have opened a PR. You will see a bot add a comment to a link where the documentation with your changes lives.
> [!NOTE]
> The `preview` command only works with existing doc files. When you add a completely new file, you need to update `_toctree.yml` & restart `preview` command (`ctrl-c` to stop it & call `doc-builder preview ...` again).
---
**NOTE**
The `preview` command only works with existing doc files. When you add a completely new file, you need to update `_toctree.yml` & restart `preview` command (`ctrl-c` to stop it & call `doc-builder preview ...` again).
---
## Adding a new element to the navigation bar
@ -159,9 +164,6 @@ These classes should be added using our Markdown syntax. Usually as follows:
[[autodoc]] XXXConfig
```
> [!IMPORTANT]
> Always add a blank line after `[[autodoc]]` to ensure it passes the CI/CD checks.
This will include every public method of the configuration that is documented. If for some reason you wish for a method
not to be displayed in the documentation, you can do so by specifying which methods should be in the docs:
@ -276,7 +278,7 @@ Here's an example of a single value return:
```python
Returns:
`list[int]`: A list of integers in the range [0, 1] --- 1 for a special token, 0 for a sequence token.
`List[int]`: A list of integers in the range [0, 1] --- 1 for a special token, 0 for a sequence token.
```
Here's an example of a tuple return, comprising several objects:

View File

@ -3,16 +3,16 @@
يُشهد في الآونة الأخيرة نمو مجال دراسي يُعنى باستكشاف آلية عمل نماذج المحولات الضخمة مثل BERT (والذي يُطلق عليها البعض اسم "BERTology"). ومن الأمثلة البارزة على هذا المجال ما يلي:
- BERT Rediscovers the Classical NLP Pipeline بواسطة Ian Tenney و Dipanjan Das و Ellie Pavlick:
https://huggingface.co/papers/1905.05950
- Are Sixteen Heads Really Better than One? بواسطة Paul Michel و Omer Levy و Graham Neubig: https://huggingface.co/papers/1905.10650
https://arxiv.org/abs/1905.05950
- Are Sixteen Heads Really Better than One? بواسطة Paul Michel و Omer Levy و Graham Neubig: https://arxiv.org/abs/1905.10650
- What Does BERT Look At? An Analysis of BERT's Attention بواسطة Kevin Clark و Urvashi Khandelwal و Omer Levy و Christopher D.
Manning: https://huggingface.co/papers/1906.04341
- CAT-probing: A Metric-based Approach to Interpret How Pre-trained Models for Programming Language Attend Code Structure: https://huggingface.co/papers/2210.04633
Manning: https://arxiv.org/abs/1906.04341
- CAT-probing: A Metric-based Approach to Interpret How Pre-trained Models for Programming Language Attend Code Structure: https://arxiv.org/abs/2210.04633
لإثراء هذا المجال الناشئ، قمنا بتضمين بعض الميزات الإضافية في نماذج BERT/GPT/GPT-2 للسماح للناس بالوصول إلى التمثيلات الداخلية، والتي تم تكييفها بشكل أساسي من العمل الرائد لـ Paul Michel (https://huggingface.co/papers/1905.10650):
لإثراء هذا المجال الناشئ، قمنا بتضمين بعض الميزات الإضافية في نماذج BERT/GPT/GPT-2 للسماح للناس بالوصول إلى التمثيلات الداخلية، والتي تم تكييفها بشكل أساسي من العمل الرائد لـ Paul Michel (https://arxiv.org/abs/1905.10650):
- الوصول إلى جميع الحالات المخفية في BERT/GPT/GPT-2،
- الوصول إلى جميع أوزان الانتباه لكل رأس في BERT/GPT/GPT-2،
- استرجاع قيم ومشتقات مخرجات الرأس لحساب درجة أهمية الرأس وحذفه كما هو موضح في https://huggingface.co/papers/1905.10650.
- استرجاع قيم ومشتقات مخرجات الرأس لحساب درجة أهمية الرأس وحذفه كما هو موضح في https://arxiv.org/abs/1905.10650.
ولمساعدتك على فهم واستخدام هذه الميزات بسهولة، أضفنا مثالًا برمجيًا محددًا: [bertology.py](https://github.com/huggingface/transformers-research-projects/tree/main/bertology/run_bertology.py) أثناء استخراج المعلومات وتقليص من نموذج تم تدريبه مسبقًا على GLUE.

View File

@ -30,7 +30,7 @@ class ResnetConfig(PretrainedConfig):
def __init__(
self,
block_type="bottleneck",
layers: list[int] = [3, 4, 6, 3],
layers: List[int] = [3, 4, 6, 3],
num_classes: int = 1000,
input_channels: int = 3,
cardinality: int = 1,
@ -280,7 +280,7 @@ resnet50d.model.load_state_dict(pretrained_model.state_dict())
الآن لإرسال النموذج إلى Hub، تأكد من تسجيل الدخول. إما تشغيل في المحطة الأوامر الطرفية الخاصة بك:
```bash
hf auth login
huggingface-cli login
```
أو من دفتر ملاحظات:

View File

@ -135,7 +135,7 @@
في كل وحدة الانتباه الباقية في المحولات، تلي طبقة الاهتمام الانتباه عادة طبقتان للتغذية الأمامية.
حجم تضمين الطبقة الأمامية الوسيطة أكبر عادة من حجم المخفي للنموذج (على سبيل المثال، لـ
`google-bert/bert-base-uncased`).
بالنسبة لإدخال بحجم `[batch_size, sequence_length]`، يمكن أن تمثل الذاكرة المطلوبة لتخزين التضمينات الأمامية الوسيطة `[batch_size، sequence_length, config.intermediate_size]` جزءًا كبيرًا من استخدام الذاكرة. لاحظ مؤلفو (https://huggingface.co/papers/2001.04451)[Reformer: The Efficient Transformer] أنه نظرًا لأن الحساب مستقل عن بعد `sequence_length`، فإنه من المكافئ رياضيًا حساب تضمينات الإخراج الأمامية `[batch_size، config.hidden_size]_0, ..., [batch_size، `config_size]_n
بالنسبة لإدخال بحجم `[batch_size, sequence_length]`، يمكن أن تمثل الذاكرة المطلوبة لتخزين التضمينات الأمامية الوسيطة `[batch_size، sequence_length, config.intermediate_size]` جزءًا كبيرًا من استخدام الذاكرة. لاحظ مؤلفو (https://arxiv.org/abs/2001.04451)[Reformer: The Efficient Transformer] أنه نظرًا لأن الحساب مستقل عن بعد `sequence_length`، فإنه من المكافئ رياضيًا حساب تضمينات الإخراج الأمامية `[batch_size، config.hidden_size]_0, ..., [batch_size، `config_size]_n
فردياً والتوصيل بها لاحقًا إلى `[batch_size, sequence_length, config.hidden_size]` مع `n = sequence_length`، والذي يتداول زيادة وقت الحساب مقابل تقليل استخدام الذاكرة، ولكنه ينتج عنه نتيجة مكافئة رياضيا.
بالنسبة للنماذج التي تستخدم الدالة `[apply_chunking_to_forward]`، يحدد `chunk_size` عدد التضمينات يتم حساب الإخراج بالتوازي وبالتالي يحدد المقايضة بين حجم الذاكرة والتعقيد الوقت. إذا تم تعيين `chunk_size` إلى `0`، فلن يتم إجراء تجزئة التغذية الأمامية.
@ -173,7 +173,7 @@
<Youtube id="VFp38yj8h3A"/>
يعمل كل محلل لغوي بشكل مختلف ولكن الآلية الأساسية تبقى كما هي. إليك مثال باستخدام محلل BERT اللغوي، والذي يعد محلل لغوي [WordPiece](https://huggingface.co/papers/1609.08144):
يعمل كل محلل لغوي بشكل مختلف ولكن الآلية الأساسية تبقى كما هي. إليك مثال باستخدام محلل BERT اللغوي، والذي يعد محلل لغوي [WordPiece](https://arxiv.org/pdf/1609.08144.pdf):
```python
>>> from transformers import BertTokenizer

View File

@ -6,18 +6,18 @@
تحقق نماذج اللغة الكبيرة (LLMs) مثل GPT3/4، [Falcon](https://huggingface.co/tiiuae/falcon-40b)، و [Llama](https://huggingface.co/meta-llama/Llama-2-70b-hf) تقدمًا سريعًا في قدرتها على معالجة المهام التي تركز على الإنسان، مما يجعلها أدوات أساسية في الصناعات القائمة على المعرفة الحديثة.
لا يزال نشر هذه النماذج في المهام الواقعية يمثل تحديًا، ومع ذلك:
- لكي تظهر نماذج اللغة الكبيرة قدرات فهم وتوليد النصوص قريبة من قدرات الإنسان، فإنها تتطلب حاليًا إلى تكوينها من مليارات المعلمات (انظر [كابلان وآخرون](https://huggingface.co/papers/2001.08361)، [وي وآخرون](https://huggingface.co/papers/2206.07682)). وهذا بدوره يزيد من متطلبات الذاكرة للاستدلال.
- لكي تظهر نماذج اللغة الكبيرة قدرات فهم وتوليد النصوص قريبة من قدرات الإنسان، فإنها تتطلب حاليًا إلى تكوينها من مليارات المعلمات (انظر [كابلان وآخرون](https://arxiv.org/abs/2001.08361)، [وي وآخرون](https://arxiv.org/abs/2206.07682)). وهذا بدوره يزيد من متطلبات الذاكرة للاستدلال.
- في العديد من المهام الواقعية، تحتاج نماذج اللغة الكبيرة إلى معلومات سياقية شاملة. يتطلب ذلك قدرة النموذج على إدارة تسلسلات إدخال طويلة للغاية أثناء الاستدلال.
يكمن جوهر صعوبة هذه التحديات في تعزيز القدرات الحسابية والذاكرة لنماذج اللغة الكبيرة، خاصة عند التعامل مع تسلسلات الإدخال الضخمة.
في هذا الدليل، سنستعرض التقنيات الفعالة لتُحسِّن من كفاءة نشر نماذج اللغة الكبيرة:
1. سنتناول تقنية "دقة أقل" التي أثبتت الأبحاث فعاليتها في تحقيق مزايا حسابية دون التأثير بشكل ملحوظ على أداء النموذج عن طريق العمل بدقة رقمية أقل [8 بت و4 بت](/main_classes/quantization).
1. سنتناول تقنية "دقة أقل" التي أثبتت الأبحاث فعاليتها في تحقيق مزايا حسابية دون التأثير بشكل ملحوظ على أداء النموذج عن طريق العمل بدقة رقمية أقل [8 بت و4 بت](/main_classes/quantization.md).
2. **اFlash Attention:** إن Flash Attention وهي نسخة مُعدَّلة من خوارزمية الانتباه التي لا توفر فقط نهجًا أكثر كفاءة في استخدام الذاكرة، ولكنها تحقق أيضًا كفاءة متزايدة بسبب الاستخدام الأمثل لذاكرة GPU.
3. **الابتكارات المعمارية:** حيث تم اقتراح هياكل متخصصة تسمح باستدلال أكثر فعالية نظرًا لأن نماذج اللغة الكبيرة يتم نشرها دائمًا بنفس الطريقة أثناء عملية الاستدلال، أي توليد النص التنبؤي التلقائي مع سياق الإدخال الطويل، فقد تم اقتراح بنيات نموذج متخصصة تسمح بالاستدلال الأكثر كفاءة. أهم تقدم في بنيات النماذج هنا هو [عذر](https://huggingface.co/papers/2108.12409)، [الترميز الدوار](https://huggingface.co/papers/2104.09864)، [الاهتمام متعدد الاستعلامات (MQA)](https://huggingface.co/papers/1911.02150) و [مجموعة الانتباه بالاستعلام (GQA)](https://huggingface.co/papers/2305.13245).
3. **الابتكارات المعمارية:** حيث تم اقتراح هياكل متخصصة تسمح باستدلال أكثر فعالية نظرًا لأن نماذج اللغة الكبيرة يتم نشرها دائمًا بنفس الطريقة أثناء عملية الاستدلال، أي توليد النص التنبؤي التلقائي مع سياق الإدخال الطويل، فقد تم اقتراح بنيات نموذج متخصصة تسمح بالاستدلال الأكثر كفاءة. أهم تقدم في بنيات النماذج هنا هو [عذر](https://arxiv.org/abs/2108.12409)، [الترميز الدوار](https://arxiv.org/abs/2104.09864)، [الاهتمام متعدد الاستعلامات (MQA)](https://arxiv.org/abs/1911.02150) و [مجموعة الانتباه بالاستعلام (GQA)]((https://arxiv.org/abs/2305.13245)).
على مدار هذا الدليل، سنقدم تحليلًا للتوليد التنبؤي التلقائي من منظور المُوتِّرات. نتعمق في مزايا وعيوب استخدام دقة أقل، ونقدم استكشافًا شاملاً لخوارزميات الانتباه الأحدث، ونناقش بنيات نماذج نماذج اللغة الكبيرة المحسنة. سندعم الشرح بأمثلة عملية تُبرِز كل تحسين على حدة.
@ -152,8 +152,8 @@ from accelerate.utils import release_memory
release_memory(model)
```
والآن ماذا لو لم يكن لدى وحدة معالجة الرسومات (GPU) لديك 32 جيجا بايت من ذاكرة الفيديو العشوائية (VRAM)؟ لقد وجد أن أوزان النماذج يمكن تحويلها إلى 8 بتات أو 4 بتات دون خسارة كبيرة في الأداء (انظر [Dettmers et al.](https://huggingface.co/papers/2208.07339)).
يمكن تحويل النموذج إلى 3 بتات أو 2 بتات مع فقدان مقبول في الأداء كما هو موضح في ورقة [GPTQ](https://huggingface.co/papers/2210.17323) 🤯.
والآن ماذا لو لم يكن لدى وحدة معالجة الرسومات (GPU) لديك 32 جيجا بايت من ذاكرة الفيديو العشوائية (VRAM)؟ لقد وجد أن أوزان النماذج يمكن تحويلها إلى 8 بتات أو 4 بتات دون خسارة كبيرة في الأداء (انظر [Dettmers et al.](https://arxiv.org/abs/2208.07339)).
يمكن تحويل النموذج إلى 3 بتات أو 2 بتات مع فقدان مقبول في الأداء كما هو موضح في ورقة [GPTQ](https://arxiv.org/abs/2210.17323) 🤯.
دون الدخول في الكثير من التفاصيل، تهدف مخططات التكميم إلى تخفيض دقة الأوزان مع محاولة الحفاظ على دقة نتائج النموذج كما هي (*أي* أقرب ما يمكن إلى bfloat16).
لاحظ أن التكميم يعمل بشكل خاص جيدًا لتوليد النص حيث كل ما نهتم به هو اختيار *مجموعة الرموز الأكثر احتمالًا التالية* ولا نهتم حقًا بالقيم الدقيقة لتوزيع الرمز التالي *logit*.
@ -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", load_in_4bit=True, pad_token_id=0)
model = AutoModelForCausalLM.from_pretrained("bigcode/octocoder", load_in_4bit=True, low_cpu_mem_usage=True, pad_token_id=0)
pipe = pipeline("text-generation", model=model, tokenizer=tokenizer)
@ -304,7 +304,7 @@ $$ \textbf{O} = \text{Attn}(\mathbf{X}) = \mathbf{V} \times \text{Softmax}(\math
مع تحسن LLMs في فهم النص وتوليد النص، يتم تطبيقها على مهام متزايدة التعقيد. في حين أن النماذج كانت تتعامل سابقًا مع ترجمة أو تلخيص بضع جمل، فإنها الآن تدير صفحات كاملة، مما يتطلب القدرة على معالجة أطوال إدخال واسعة.
كيف يمكننا التخلص من متطلبات الذاكرة الباهظة للتطويلات المدخلة الكبيرة؟ نحن بحاجة إلى طريقة جديدة لحساب آلية الاهتمام الذاتي التي تتخلص من مصفوفة \\( QK^T \\). [طريقه داو وآخرون.](https://huggingface.co/papers/2205.14135) طوروا بالضبط مثل هذا الخوارزمية الجديدة وأطلقوا عليها اسم **Flash Attention**.
كيف يمكننا التخلص من متطلبات الذاكرة الباهظة للتطويلات المدخلة الكبيرة؟ نحن بحاجة إلى طريقة جديدة لحساب آلية الاهتمام الذاتي التي تتخلص من مصفوفة \\( QK^T \\). [طريقه داو وآخرون.](Https://arxiv.org/abs/2205.14135) طوروا بالضبط مثل هذا الخوارزمية الجديدة وأطلقوا عليها اسم **Flash Attention**.
باختصار، يكسر الاهتمام الفلاشي حساب \\( \mathbf{V} \times \operatorname{Softmax}(\mathbf{QK}^T\\)) ويحسب بدلاً من ذلك قطعًا أصغر من الإخراج عن طريق التكرار عبر العديد من خطوات حساب Softmax:
@ -318,7 +318,7 @@ $$ \textbf{O}_i \leftarrow s^a_{ij} * \textbf{O}_i + s^b_{ij} * \mathbf{V}_{j} \
> من خلال تتبع إحصائيات التطبيع softmax واستخدام بعض الرياضيات الذكية، يعطي Flash Attention **مخرجات متطابقة رقميًا** مقارنة بطبقة الاهتمام الذاتي الافتراضية بتكلفة ذاكرة لا تزيد خطيًا مع \\( N \\).
عند النظر إلى الصيغة، قد يقول المرء بديهيًا أن الاهتمام الفلاشي يجب أن يكون أبطأ بكثير مقارنة بصيغة الاهتمام الافتراضية حيث يلزم إجراء المزيد من الحسابات. في الواقع، يتطلب Flash Attention المزيد من عمليات الفاصلة العائمة مقارنة بالاهتمام العادي حيث يجب إعادة حساب إحصائيات التطبيع softmax باستمرار (راجع [الورقة](https://huggingface.co/papers/2205.14135) لمزيد من التفاصيل إذا كنت مهتمًا)
عند النظر إلى الصيغة، قد يقول المرء بديهيًا أن الاهتمام الفلاشي يجب أن يكون أبطأ بكثير مقارنة بصيغة الاهتمام الافتراضية حيث يلزم إجراء المزيد من الحسابات. في الواقع، يتطلب Flash Attention المزيد من عمليات الفاصلة العائمة مقارنة بالاهتمام العادي حيث يجب إعادة حساب إحصائيات التطبيع softmax باستمرار (راجع [الورقة](https://arxiv.org/abs/2205.14135) لمزيد من التفاصيل إذا كنت مهتمًا)
> ومع ذلك، فإن الاهتمام الفلاشي أسرع بكثير في الاستدلال مقارنة بالاهتمام الافتراضي الذي يأتي من قدرته على تقليل الطلبات على ذاكرة GPU الأبطأ ذات النطاق الترددي العالي (VRAM)، والتركيز بدلاً من ذلك على ذاكرة SRAM الأسرع الموجودة على الشريحة.
@ -535,20 +535,20 @@ flush()
لكي يفهم LLM ترتيب الجملة، يلزم وجود *إشارة* إضافية ويتم تطبيقها عادةً في شكل *الترميزات الموضعية* (أو ما يُطلق عليه أيضًا *الترميزات الموضعية*).
لم يتم ترجمة النص الخاص والروابط وأكواد HTML وCSS بناءً على طلبك.
قدم مؤلفو الورقة البحثية [*Attention Is All You Need*](https://huggingface.co/papers/1706.03762) تضمينات موضعية جيبية مثلثية \\( \mathbf{P} = \mathbf{p}_1, \ldots, \mathbf{p}_N \\) حيث يتم حساب كل متجه \\( \mathbf{p}_i \\) كدالة جيبية لموضعه \\( i \\) .
قدم مؤلفو الورقة البحثية [*Attention Is All You Need*](https://arxiv.org/abs/1706.03762) تضمينات موضعية جيبية مثلثية \\( \mathbf{P} = \mathbf{p}_1, \ldots, \mathbf{p}_N \\) حيث يتم حساب كل متجه \\( \mathbf{p}_i \\) كدالة جيبية لموضعه \\( i \\) .
بعد ذلك يتم ببساطة إضافة التضمينات الموضعية إلى متجهات تسلسل الإدخال \\( \mathbf{\hat{X}} = \mathbf{\hat{x}}_1, \ldots, \mathbf{\hat{x}}_N \\) = \\( \mathbf{x}_1 + \mathbf{p}_1, \ldots, \mathbf{x}_N + \mathbf{p}_N \\) وبالتالي توجيه النموذج لتعلم ترتيب الجملة بشكل أفضل.
بدلاً من استخدام التضمينات الموضعية الثابتة، استخدم آخرون (مثل [Devlin et al.](https://huggingface.co/papers/1810.04805)) تضمينات موضعية مكتسبة يتم من خلالها تعلم التضمينات الموضعية \\( \mathbf{P} \\) أثناء التدريب.
بدلاً من استخدام التضمينات الموضعية الثابتة، استخدم آخرون (مثل [Devlin et al.](https://arxiv.org/abs/1810.04805)) تضمينات موضعية مكتسبة يتم من خلالها تعلم التضمينات الموضعية \\( \mathbf{P} \\) أثناء التدريب.
كانت التضمينات الموضعية الجيبية والمكتسبة هي الطرق السائدة لترميز ترتيب الجملة في نماذج اللغة الكبيرة، ولكن تم العثور على بعض المشكلات المتعلقة بهذه التضمينات الموضعية:
1. التضمينات الموضعية الجيبية والمكتسبة هي تضمينات موضعية مطلقة، أي ترميز تضمين فريد لكل معرف موضعي: \\( 0, \ldots, N \\) . كما أظهر [Huang et al.](https://huggingface.co/papers/2009.13658) و [Su et al.](https://huggingface.co/papers/2104.09864)، تؤدي التضمينات الموضعية المطلقة إلى أداء ضعيف لنماذج اللغة الكبيرة للمدخلات النصية الطويلة. بالنسبة للمدخلات النصية الطويلة، يكون من المفيد إذا تعلم النموذج المسافة الموضعية النسبية التي تمتلكها رموز المدخلات إلى بعضها البعض بدلاً من موضعها المطلق.
1. التضمينات الموضعية الجيبية والمكتسبة هي تضمينات موضعية مطلقة، أي ترميز تضمين فريد لكل معرف موضعي: \\( 0, \ldots, N \\) . كما أظهر [Huang et al.](https://arxiv.org/abs/2009.13658) و [Su et al.](https://arxiv.org/abs/2104.09864)، تؤدي التضمينات الموضعية المطلقة إلى أداء ضعيف لنماذج اللغة الكبيرة للمدخلات النصية الطويلة. بالنسبة للمدخلات النصية الطويلة، يكون من المفيد إذا تعلم النموذج المسافة الموضعية النسبية التي تمتلكها رموز المدخلات إلى بعضها البعض بدلاً من موضعها المطلق.
2. عند استخدام التضمينات الموضعية المكتسبة، يجب تدريب نموذج اللغة الكبيرة على طول إدخال ثابت \\( N \\)، مما يجعل من الصعب الاستقراء إلى طول إدخال أطول مما تم تدريبه عليه.
في الآونة الأخيرة، أصبحت التضمينات الموضعية النسبية التي يمكنها معالجة المشكلات المذكورة أعلاه أكثر شعبية، وأبرزها:
- [تضمين الموضع الدوراني (RoPE)](https://huggingface.co/papers/2104.09864)
- [ALiBi](https://huggingface.co/papers/2108.12409)
- [تضمين الموضع الدوراني (RoPE)](https://arxiv.org/abs/2104.09864)
- [ALiBi](https://arxiv.org/abs/2108.12409)
يؤكد كل من *RoPE* و *ALiBi* أنه من الأفضل توجيه نموذج اللغة الكبيرة حول ترتيب الجملة مباشرة في خوارزمية الانتباه الذاتي حيث يتم وضع رموز الكلمات في علاقة مع بعضها البعض. على وجه التحديد، يجب توجيه ترتيب الجملة عن طريق تعديل عملية \\( \mathbf{QK}^T \\) .
@ -563,14 +563,14 @@ $$ \mathbf{\hat{q}}_i^T \mathbf{\hat{x}}_j = \mathbf{{q}}_i^T \mathbf{R}_{\theta
يستخدم *RoPE* في العديد من نماذج اللغة الكبيرة الأكثر أهمية اليوم، مثل:
- [**Falcon**](https://huggingface.co/tiiuae/falcon-40b)
- [**Llama**](https://huggingface.co/papers/2302.13971)
- [**PaLM**](https://huggingface.co/papers/2204.02311)
- [**Llama**](https://arxiv.org/abs/2302.13971)
- [**PaLM**](https://arxiv.org/abs/2204.02311)
كبديل، يقترح *ALiBi* مخطط ترميز موضعي نسبي أبسط بكثير. يتم إضافة المسافة النسبية التي تمتلكها رموز المدخلات إلى بعضها البعض كعدد صحيح سلبي مقياس بقيمة محددة مسبقًا `m` إلى كل إدخال استعلام-مفتاح لمصفوفة \\( \mathbf{QK}^T \\) مباشرة قبل حساب softmax.
![](/blog/assets/163_optimize_llm/alibi.png)
كما هو موضح في ورقة [ALiBi](https://huggingface.co/papers/2108.12409)، يسمح هذا الترميز الموضعي النسبي البسيط للنموذج بالحفاظ على أداء عالٍ حتى في تسلسلات المدخلات النصية الطويلة جدًا.
كما هو موضح في ورقة [ALiBi](https://arxiv.org/abs/2108.12409)، يسمح هذا الترميز الموضعي النسبي البسيط للنموذج بالحفاظ على أداء عالٍ حتى في تسلسلات المدخلات النصية الطويلة جدًا.
يُستخدم *ALiBi* في العديد من أهم نماذج اللغة الكبيرة المستخدمة اليوم، مثل:
@ -579,7 +579,7 @@ $$ \mathbf{\hat{q}}_i^T \mathbf{\hat{x}}_j = \mathbf{{q}}_i^T \mathbf{R}_{\theta
يمكن لكل من ترميزات الموضع *RoPE* و *ALiBi* الاستقراء إلى أطوال إدخال لم يتم ملاحظتها أثناء التدريب، في حين ثبت أن الاستقراء يعمل بشكل أفضل بكثير خارج الصندوق لـ *ALiBi* مقارنة بـ *RoPE*.
بالنسبة لـ ALiBi، ما عليك سوى زيادة قيم مصفوفة الموضع المثلث السفلي لمطابقة طول تسلسل الإدخال.
بالنسبة لـ *RoPE*، يؤدي الحفاظ على نفس \\( \theta \\) الذي تم استخدامه أثناء التدريب إلى نتائج سيئة عند تمرير إدخالات نصية أطول بكثير من تلك التي شوهدت أثناء التدريب، راجع [Press et al.](https://huggingface.co/papers/2108.12409). ومع ذلك، وجد المجتمع بعض الحيل الفعالة التي تقوم بتعديل \\( \theta \\)، مما يسمح لترميزات الموضع *RoPE* بالعمل بشكل جيد لتسلسلات إدخال النص المستقرئة (راجع [هنا](https://github.com/huggingface/transformers/pull/24653)).
بالنسبة لـ *RoPE*، يؤدي الحفاظ على نفس \\( \theta \\) الذي تم استخدامه أثناء التدريب إلى نتائج سيئة عند تمرير إدخالات نصية أطول بكثير من تلك التي شوهدت أثناء التدريب، راجع [Press et al.](https://arxiv.org/abs/2108.12409). ومع ذلك، وجد المجتمع بعض الحيل الفعالة التي تقوم بتعديل \\( \theta \\)، مما يسمح لترميزات الموضع *RoPE* بالعمل بشكل جيد لتسلسلات إدخال النص المستقرئة (راجع [هنا](https://github.com/huggingface/transformers/pull/24653)).
> كل من RoPE و ALiBi عبارة عن ترميزات موضع نسبي *لا* يتم تعلمها أثناء التدريب، ولكن بدلاً من ذلك تستند إلى الحدس التالي:
- يجب إعطاء الإشارات الموضعية حول إدخالات النص مباشرة إلى مصفوفة \\( QK^T \\) لطبقة الاهتمام الذاتي
@ -755,21 +755,21 @@ Roughly 8 مليار قيمة عائمة! يتطلب تخزين 8 مليارات
#### 3.2.2 Multi-Query-Attention (MQA)
[Multi-Query-Attention](https://huggingface.co/papers/1911.02150) اقترحها Noam Shazeer في ورقته *Fast Transformer Decoding: One Write-Head is All You Need*. كما يقول العنوان، اكتشف Noam أنه بدلاً من استخدام `n_head` من أوزان إسقاط القيمة الرئيسية، يمكن استخدام زوج واحد من أوزان إسقاط رأس القيمة التي يتم مشاركتها عبر جميع رؤوس الاهتمام دون أن يتدهور أداء النموذج بشكل كبير.
[Multi-Query-Attention](https://arxiv.org/abs/1911.02150) اقترحها Noam Shazeer في ورقته *Fast Transformer Decoding: One Write-Head is All You Need*. كما يقول العنوان، اكتشف Noam أنه بدلاً من استخدام `n_head` من أوزان إسقاط القيمة الرئيسية، يمكن استخدام زوج واحد من أوزان إسقاط رأس القيمة التي يتم مشاركتها عبر جميع رؤوس الاهتمام دون أن يتدهور أداء النموذج بشكل كبير.
> باستخدام زوج واحد من أوزان إسقاط رأس القيمة، يجب أن تكون متجهات القيمة الرئيسية \\( \mathbf{k}_i، \mathbf{v}_i \\) متطابقة عبر جميع رؤوس الاهتمام والتي بدورها تعني أننا بحاجة فقط إلى تخزين زوج إسقاط قيمة رئيسي واحد في ذاكرة التخزين المؤقت بدلاً من `n_head` منها.
نظرًا لأن معظم LLMs تستخدم ما بين 20 و100 رأس اهتمام، فإن MQA يقلل بشكل كبير من استهلاك الذاكرة لذاكرة التخزين المؤقت key-value. بالنسبة إلى LLM المستخدم في هذا الدفتر، يمكننا تقليل استهلاك الذاكرة المطلوبة من 15 جيجابايت إلى أقل من 400 ميجابايت عند طول تسلسل الإدخال 16000.
بالإضافة إلى توفير الذاكرة، يؤدي MQA أيضًا إلى تحسين الكفاءة الحسابية كما هو موضح في ما يلي.
في فك التشفير التلقائي، يجب إعادة تحميل متجهات القيمة الرئيسية الكبيرة، ودمجها مع زوج متجه القيمة الحالي، ثم إدخالها في \\( \mathbf{q}_c\mathbf{K}^T \\) الحساب في كل خطوة. بالنسبة لفك التشفير التلقائي، يمكن أن تصبح عرض النطاق الترددي للذاكرة المطلوبة لإعادة التحميل المستمر عنق زجاجة زمنيًا خطيرًا. من خلال تقليل حجم متجهات القيمة الرئيسية، يجب الوصول إلى ذاكرة أقل، وبالتالي تقليل عنق الزجاجة في عرض النطاق الترددي للذاكرة. لمزيد من التفاصيل، يرجى إلقاء نظرة على [ورقة Noam](https://huggingface.co/papers/1911.02150).
في فك التشفير التلقائي، يجب إعادة تحميل متجهات القيمة الرئيسية الكبيرة، ودمجها مع زوج متجه القيمة الحالي، ثم إدخالها في \\( \mathbf{q}_c\mathbf{K}^T \\) الحساب في كل خطوة. بالنسبة لفك التشفير التلقائي، يمكن أن تصبح عرض النطاق الترددي للذاكرة المطلوبة لإعادة التحميل المستمر عنق زجاجة زمنيًا خطيرًا. من خلال تقليل حجم متجهات القيمة الرئيسية، يجب الوصول إلى ذاكرة أقل، وبالتالي تقليل عنق الزجاجة في عرض النطاق الترددي للذاكرة. لمزيد من التفاصيل، يرجى إلقاء نظرة على [ورقة Noam](https://arxiv.org/abs/1911.02150).
الجزء المهم الذي يجب فهمه هنا هو أن تقليل عدد رؤوس الاهتمام بالقيمة الرئيسية إلى 1 لا معنى له إلا إذا تم استخدام ذاكرة التخزين المؤقت للقيمة الرئيسية. يظل الاستهلاك الذروي لذاكرة النموذج لمرور واحد للأمام بدون ذاكرة التخزين المؤقت للقيمة الرئيسية دون تغيير لأن كل رأس اهتمام لا يزال لديه متجه استعلام فريد بحيث يكون لكل رأس اهتمام مصفوفة \\( \mathbf{QK}^T \\) مختلفة.
شهدت MQA اعتمادًا واسع النطاق من قبل المجتمع ويتم استخدامها الآن بواسطة العديد من LLMs الأكثر شهرة:
- [**Falcon**](https://huggingface.co/tiiuae/falcon-40b)
- [**PaLM**](https://huggingface.co/papers/2204.02311)
- [**PaLM**](https://arxiv.org/abs/2204.02311)
- [**MPT**](https://huggingface.co/mosaicml/mpt-30b)
- [**BLOOM**](https://huggingface.co/bigscience/bloom)
@ -777,7 +777,7 @@ Roughly 8 مليار قيمة عائمة! يتطلب تخزين 8 مليارات
#### 3.2.3 مجموعة الاستعلام الاهتمام (GQA)
[مجموعة الاستعلام الاهتمام](https://huggingface.co/papers/2305.13245)، كما اقترح Ainslie et al. من Google، وجد أن استخدام MQA يمكن أن يؤدي غالبًا إلى تدهور الجودة مقارنة باستخدام إسقاطات رأس القيمة الرئيسية المتعددة. تجادل الورقة بأنه يمكن الحفاظ على أداء النموذج بشكل أكبر عن طريق تقليل عدد أوزان إسقاط رأس الاستعلام بشكل أقل حدة. بدلاً من استخدام وزن إسقاط قيمة رئيسية واحدة فقط، يجب استخدام `n <n_head` أوزان إسقاط قيمة رئيسية. من خلال اختيار `n` إلى قيمة أقل بكثير من `n_head مثل 2 أو 4 أو 8، يمكن الاحتفاظ بمعظم مكاسب الذاكرة والسرعة من MQA مع التضحية بقدر أقل من سعة النموذج وبالتالي، من المفترض، أقل أداء.
[مجموعة الاستعلام الاهتمام](https://arxiv.org/abs/2305.13245)، كما اقترح Ainslie et al. من Google، وجد أن استخدام MQA يمكن أن يؤدي غالبًا إلى تدهور الجودة مقارنة باستخدام إسقاطات رأس القيمة الرئيسية المتعددة. تجادل الورقة بأنه يمكن الحفاظ على أداء النموذج بشكل أكبر عن طريق تقليل عدد أوزان إسقاط رأس الاستعلام بشكل أقل حدة. بدلاً من استخدام وزن إسقاط قيمة رئيسية واحدة فقط، يجب استخدام `n <n_head` أوزان إسقاط قيمة رئيسية. من خلال اختيار `n` إلى قيمة أقل بكثير من `n_head مثل 2 أو 4 أو 8، يمكن الاحتفاظ بمعظم مكاسب الذاكرة والسرعة من MQA مع التضحية بقدر أقل من سعة النموذج وبالتالي، من المفترض، أقل أداء.
علاوة على ذلك، اكتشف مؤلفو GQA أنه يمكن *تدريب* نقاط تفتيش النموذج الموجودة ليكون لها بنية GQA باستخدام 5% فقط من الحوسبة الأصلية للتعليم المسبق. في حين أن 5% من الحوسبة الأصلية للتعليم المسبق يمكن أن تكون كمية هائلة، يسمح GQA *uptraining* بنقاط تفتيش موجودة للاستفادة من تسلسلات الإدخال الأطول.
@ -789,7 +789,7 @@ Roughly 8 مليار قيمة عائمة! يتطلب تخزين 8 مليارات
## الخاتمة
مجتمع البحث يأتي باستمرار بطرق جديدة ومبتكرة لتسريع وقت الاستدلال للنماذج اللغوية الكبيرة على الإطلاق. كمثال، أحد اتجاهات البحث الواعدة هو [فك التشفير التخميني](https://huggingface.co/papers/2211.17192) حيث تقوم "الرموز السهلة" بإنشائها نماذج اللغة الأصغر والأسرع ويتم إنشاء "الرموز الصعبة" فقط بواسطة LLM نفسه. إن التعمق في التفاصيل يتجاوز نطاق هذا الدفتر، ولكن يمكن قراءته في هذه [تدوينة المدونة اللطيفة](https://huggingface.co/blog/assisted-generation).
مجتمع البحث يأتي باستمرار بطرق جديدة ومبتكرة لتسريع وقت الاستدلال للنماذج اللغوية الكبيرة على الإطلاق. كمثال، أحد اتجاهات البحث الواعدة هو [فك التشفير التخميني](https://arxiv.org/abs/2211.17192) حيث تقوم "الرموز السهلة" بإنشائها نماذج اللغة الأصغر والأسرع ويتم إنشاء "الرموز الصعبة" فقط بواسطة LLM نفسه. إن التعمق في التفاصيل يتجاوز نطاق هذا الدفتر، ولكن يمكن قراءته في هذه [تدوينة المدونة اللطيفة](https://huggingface.co/blog/assisted-generation).
السبب في أن LLMs الضخمة مثل GPT3/4، وLlama-2-70b، وClaude، وPaLM يمكن أن تعمل بسرعة كبيرة في واجهات الدردشة مثل [Hugging Face Chat](https://huggingface.co/chat/) أو ChatGPT يرجع إلى حد كبير إلى التحسينات المذكورة أعلاه في الدقة والخوارزميات والهندسة المعمارية.
في المستقبل، ستكون أجهزة التسريع مثل وحدات معالجة الرسومات (GPUs) ووحدات معالجة الرسومات (TPUs وما إلى ذلك... ستكون أسرع فقط وستسمح بمزيد من الذاكرة، ولكن يجب دائمًا التأكد من استخدام أفضل الخوارزميات والهندسة المعمارية المتاحة للحصول على أكبر قدر من المال

View File

@ -165,7 +165,7 @@ default_args = {
يمكن أن تكون هذه المعرفة مفيدة لمعرفة عند تحليل اختناقات الأداء.
هذا الملخص مُشتق من [نقل البيانات هو كل ما تحتاجه: دراسة حالة حول تحسين المحولات 2020](https://huggingface.co/papers/2007.00072)
هذا الملخص مُشتق من [نقل البيانات هو كل ما تحتاجه: دراسة حالة حول تحسين المحولات 2020](https://arxiv.org/abs/2007.00072)
## تشريح ذاكرة النموذج

View File

@ -41,7 +41,7 @@ picture-in-picture" allowfullscreen></iframe>
قبل مشاركة نموذج على Hub، ستحتاج إلى بيانات اعتماد حساب Hugging Face الخاصة بك. إذا كنت تستخدم منصة الأوامر، فقم بتشغيل الأمر التالي في بيئة افتراضية حيث تم تثبيت 🤗 Transformers. سيقوم هذا الأمر بتخزين رمز الدخول الخاص بك في مجلد تخزين المؤقت لـ Hugging Face (`~/.cache/` بشكل افتراضي):
```bash
hf auth login
huggingface-cli login
```
إذا كنت تستخدم دفتر ملاحظات مثل Jupyter أو Colaboratory، فتأكد من تثبيت مكتبة [`huggingface_hub`](https://huggingface.co/docs/hub/adding-a-library). تسمح لك هذه المكتبة بالتفاعل برمجيًا مع Hub.

View File

@ -1,6 +1,6 @@
# عائلة نماذج المحول
منذ إطلاقه في عام 2017، ألهم نموذج [المحول الأصلي](https://huggingface.co/papers/1706.03762) (راجع مدونة [المحول المشروح](http://nlp.seas.harvard.edu/2018/04/03/attention.html) لمقدمة تقنية مبسطة)، ألهم العديد من النماذج الجديدة والمبتكرة التي تتجاوز مهام معالجة اللغات الطبيعية (NLP). هناك نماذج للتنبؤ [بالبنية البروتينات المطوية](https://huggingface.co/blog/deep-learning-with-proteins)، و[تدريب على اتخاذ القرار](https://huggingface.co/blog/train-decision-transformers)، و[التنبؤ بالسلاسل الزمنية](https://huggingface.co/blog/time-series-transformers). مع وجود العديد من متغيرات المحول المتاحة، قد يكون من السهل أن تفوتك الصورة الأكبر. ما تشترك فيه جميع هذه النماذج هو أنها تستند إلى بنية المحول الأصلية. تستخدم بعض النماذج فقط الترميز أو فك الترميز، بينما تستخدم نماذج أخرى كليهما. يوفر هذا تصنيفًا مفيدًا لتصنيف واستعراض الفروقات الرئيسية بين نماذج عائلة المحولات، وسيساعدك على فهم النماذج التي لم تصادفها من قبل.
منذ إطلاقه في عام 2017، ألهم نموذج [المحول الأصلي](https://arxiv.org/abs/1706.03762) (راجع مدونة [المحول المشروح](http://nlp.seas.harvard.edu/2018/04/03/attention.html) لمقدمة تقنية مبسطة)، ألهم العديد من النماذج الجديدة والمبتكرة التي تتجاوز مهام معالجة اللغات الطبيعية (NLP). هناك نماذج للتنبؤ [بالبنية البروتينات المطوية](https://huggingface.co/blog/deep-learning-with-proteins)، و[تدريب على اتخاذ القرار](https://huggingface.co/blog/train-decision-transformers)، و[التنبؤ بالسلاسل الزمنية](https://huggingface.co/blog/time-series-transformers). مع وجود العديد من متغيرات المحول المتاحة، قد يكون من السهل أن تفوتك الصورة الأكبر. ما تشترك فيه جميع هذه النماذج هو أنها تستند إلى بنية المحول الأصلية. تستخدم بعض النماذج فقط الترميز أو فك الترميز، بينما تستخدم نماذج أخرى كليهما. يوفر هذا تصنيفًا مفيدًا لتصنيف واستعراض الفروقات الرئيسية بين نماذج عائلة المحولات، وسيساعدك على فهم النماذج التي لم تصادفها من قبل.
إذا لم تكن على دراية بنموذج المحول الأصلي أو تحتاج إلى تذكير، فراجع الفصل الخاص بـ [كيف تعمل المحولات](https://huggingface.co/course/chapter1/4؟fw=pt) من دورة Hugging Face.
@ -14,7 +14,7 @@
### الشبكة التلافيفية (Convolutional network)
لطالما كانت الشبكات التلافيفية (CNNs) الطريقة السائدة لمهام رؤية الحاسب حتى برز [محول الرؤية](https://huggingface.co/papers/2010.11929) قابليته للتطوير وكفاءته العالية. وحتى بعد ذلك، لا تزال بعض أفضل صفات CNN، مثل ثبات الإزاحة، قوية جدًا (خاصة بالنسبة لمهام معينة) لدرجة أن بعض المحولات تدمج التلافيف في بنيتها. قلب [ConvNeXt](model_doc/convnext) هذا التبادل رأسًا على عقب وأدرج خيارات التصميم من المحولات لتحديث CNN. على سبيل المثال، يستخدم ConvNeXt نوافذ منزلقة غير متداخلة لتقسيم الصورة إلى رقع وزيادة حقل مجال العام الخاص بها. كما يقوم ConvNeXt بعدة خيارات مثل تصميم الطبقة لتكون أكثر كفاءة في الذاكرة وتحسين الأداء، مما يجعله منافسًا قويًا للمحولات!
لطالما كانت الشبكات التلافيفية (CNNs) الطريقة السائدة لمهام رؤية الحاسب حتى برز [محول الرؤية](https://arxiv.org/abs/2010.11929) قابليته للتطوير وكفاءته العالية. وحتى بعد ذلك، لا تزال بعض أفضل صفات CNN، مثل ثبات الإزاحة، قوية جدًا (خاصة بالنسبة لمهام معينة) لدرجة أن بعض المحولات تدمج التلافيف في بنيتها. قلب [ConvNeXt](model_doc/convnext) هذا التبادل رأسًا على عقب وأدرج خيارات التصميم من المحولات لتحديث CNN. على سبيل المثال، يستخدم ConvNeXt نوافذ منزلقة غير متداخلة لتقسيم الصورة إلى رقع وزيادة حقل مجال العام الخاص بها. كما يقوم ConvNeXt بعدة خيارات مثل تصميم الطبقة لتكون أكثر كفاءة في الذاكرة وتحسين الأداء، مما يجعله منافسًا قويًا للمحولات!
### الترميز[[cv-encoder]] (Encoder)
@ -40,7 +40,7 @@
نموذج [BERT](model_doc/bert) هو محوّل (Transformer) يعتمد على الترميز فقط يقوم بشكل عشوائي بإخفاء رموز معينة في المدخلات لتجنب رؤية باقى الرموز الأخرى، مما يسمح له "بالغش". يتمثل هدف التدريب المسبق في التنبؤ بالرمز المخفي بناءً على السياق. يسمح هذا لـ BERT باستخدام السياقات اليمنى واليسرى بالكامل لمساعدته في تعلم تمثيل أعمق وأغنى للبيانات المدخلة. ومع ذلك، كان هناك مجال للتحسين في استراتيجية التدريب المسبق لـ BERT. نموذج [RoBERTa](model_doc/roberta) اضاف تحسين من خلال تقديم وصفة تدريب مسبق جديدة تشمل التدريب لفترة أطول وعلى دفعات أكبر، وإخفاء الرموز عشوائيًا في كل حقبة بدلاً من مرة واحدة فقط أثناء المعالجة المسبقة، وإزالة هدف التنبؤ بالجملة التالية.
تتمثل الاستراتيجية السائدة لتحسين الأداء في زيادة حجم النموذج. ولكن تدريب النماذج الكبيرة مكلف من الناحية الحسابية. إحدى طرق تقليل التكاليف الحسابية هي استخدام نموذج أصغر مثل [DistilBERT](model_doc/distilbert). يستخدم DistilBERT [ تقنية تقطير المعرفة](https://huggingface.co/papers/1503.02531) - وهي تقنية ضغط - لإنشاء نموذج أصغر من BERT مع الحفاظ على معظم قدراته على فهم اللغةا.
تتمثل الاستراتيجية السائدة لتحسين الأداء في زيادة حجم النموذج. ولكن تدريب النماذج الكبيرة مكلف من الناحية الحسابية. إحدى طرق تقليل التكاليف الحسابية هي استخدام نموذج أصغر مثل [DistilBERT](model_doc/distilbert). يستخدم DistilBERT [ تقنية تقطير المعرفة](https://arxiv.org/abs/1503.02531) - وهي تقنية ضغط - لإنشاء نموذج أصغر من BERT مع الحفاظ على معظم قدراته على فهم اللغةا.
مرت معظم نماذج المحول في الاتجاه نحو المزيد من المعلمات، مما أدى إلى ظهور نماذج جديدة تركز على تحسين كفاءة التدريب. يقلّل [ALBERT](model_doc/albert) من استهلاك الذاكرة عن طريق تقليل عدد المعلمات بطريقتين: فصل تضمين المفردات الأكبر إلى مصفوفتين أصغر والسماح للمستويات بمشاركة المعلمات. أضاف [DeBERTa](model_doc/deberta) آلية انتباه منفصلة حيث يتم ترميز الكلمة وموضعها بشكل منفصل في متجهين. يتم حساب الانتباه من هذه المتجهات المنفصلة بدلاً من متجه واحد يحتوي على تضمين الكلمة والموقع. ركز [Longformer](model_doc/longformer) أيضًا على جعل الانتباه أكثر كفاءة، خاصة لمعالجة المستندات ذات تسلسلات أطولل. فهو يستخدم مزيجًا من انتباه النوافذ المحلية (يتم حساب الانتباه فقط ن نافذة ذات حجم ثابت حول كل رمز) والانتباه العام (فقط لرموز مهمة محددة مثل `[CLS]` للتصنيف) لإنشاء مصفوفة انتباه متفرقة بدلاً من مصفوفة انتباه كاملة.

View File

@ -33,7 +33,7 @@ pip install git+https://github.com/huggingface/peft.git
- [محولات الرتبة المنخفضة](https://huggingface.co/docs/peft/conceptual_guides/lora)
- [IA3](https://huggingface.co/docs/peft/conceptual_guides/ia3)
- [AdaLoRA](https://huggingface.co/papers/2303.10512)
- [AdaLoRA](https://arxiv.org/abs/2303.10512)
إذا كنت تريد استخدام طرق PEFT الأخرى، مثل تعلم المحث أو ضبط المحث، أو حول مكتبة 🤗 PEFT بشكل عام، يرجى الرجوع إلى [الوثائق](https://huggingface.co/docs/peft/index).

View File

@ -324,7 +324,7 @@ python examples/pytorch/summarization/run_summarization.py
يمكن لجميع النصوص البرمجية رفع نموذجك النهائي إلى [مركز النماذج](https://huggingface.co/models). تأكد من تسجيل الدخول إلى Hugging Face قبل البدء:
```bash
hf auth login
huggingface-cli login
```
ثم أضف المعلمة `push_to_hub` إلى النص البرمجي . ستقوم هذه المعلمة بإنشاء مستودع باستخدام اسم مستخدم Hugging Face واسم المجلد المحدد في `output_dir`.

View File

@ -103,7 +103,7 @@
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/convolution.gif"/>
</div>
<small>عملية التفاف أساسية بدون حشو أو خطو خطوة واسعة، مأخوذة من <a href="https://huggingface.co/papers/1603.07285">دليل لحساب الالتفاف للتعلم العميق.</a></small>
<small>عملية التفاف أساسية بدون حشو أو خطو خطوة واسعة، مأخوذة من <a href="https://arxiv.org/abs/1603.07285">دليل لحساب الالتفاف للتعلم العميق.</a></small>
يمكنك تغذية هذا الناتج إلى طبقة التفاف أخرى، ومع كل طبقة متتالية، تتعلم الشبكة أشياء أكثر تعقيدًا وتجريدية مثل النقانق أو الصواريخ. بين طبقات الالتفاف، من الشائع إضافة طبقة تجميع لتقليل الأبعاد وجعل النموذج أكثر قوة للتغيرات في موضع الميزة.

View File

@ -94,7 +94,7 @@
### ترميز الأزواج البايتية (BPE)
تم تقديم رميز أزواج البايت (BPE) في ورقة بحثية بعنوان [الترجمة الآلية العصبية للكلمات النادرة باستخدام وحدات subword (Sennrich et al.، 2015)](https://huggingface.co/papers/1508.07909). يعتمد BPE على مُجزّئ أولي يقسم بيانات التدريب إلى
تم تقديم رميز أزواج البايت (BPE) في ورقة بحثية بعنوان [الترجمة الآلية العصبية للكلمات النادرة باستخدام وحدات subword (Sennrich et al.، 2015)](https://arxiv.org/abs/1508.07909). يعتمد BPE على مُجزّئ أولي يقسم بيانات التدريب إلى
كلمات. يمكن أن يكون التحليل المسبق بسيطًا مثل التقسيم المكاني، على سبيل المثال [GPT-2](model_doc/gpt2)، [RoBERTa](model_doc/roberta). تشمل التقسيم الأكثر تقدمًا معتمد على التحليل القائم على القواعد، على سبيل المثال [XLM](model_doc/xlm)، [FlauBERT](model_doc/flaubert) الذي يستخدم Moses لمعظم اللغات، أو [GPT](model_doc/openai-gpt) الذي يستخدم spaCy و ftfy، لحساب تكرار كل كلمة في مجموعة بيانات التدريب.
بعد التحليل المسبق، يتم إنشاء مجموعة من الكلمات الفريدة وقد تم تحديد تكرار كل كلمة في تم تحديد بيانات التدريب. بعد ذلك، يقوم BPE بإنشاء مفردات أساسية تتكون من جميع الرموز التي تحدث في مجموعة الكلمات الفريدة ويتعلم قواعد الدمج لتشكيل رمز جديد من رمزين من المفردات الأساسية. إنه يفعل ذلك حتى تصل المفردات إلى حجم المفردات المطلوب. لاحظ أن حجم المفردات هو فرط معلمة لتحديد قبل تدريب مُجزّئ النصوص.
@ -158,7 +158,7 @@ BPE. أولاً، يقوم WordPiece بتكوين المفردات لتضمين
### Unigram
Unigram هو خوارزمية توكنيز subword التي تم تقديمها في [تنظيم subword: تحسين نماذج الترجمة الشبكة العصبية
نماذج مع مرشحين subword متعددة (Kudo، 2018)](https://huggingface.co/papers/1804.10959). على عكس BPE أو
نماذج مع مرشحين subword متعددة (Kudo، 2018)](https://arxiv.org/pdf/1804.10959.pdf). على عكس BPE أو
WordPiece، يقوم Unigram بتكوين مفرداته الأساسية إلى عدد كبير من الرموز ويقللها تدريجياً للحصول على مفردات أصغر. يمكن أن تتوافق المفردات الأساسية على سبيل المثال مع جميع الكلمات المسبقة التوكنز والسلاسل الفرعية الأكثر شيوعًا. لا يتم استخدام Unigram مباشرة لأي من النماذج في المحولات، ولكنه يستخدم بالاقتران مع [SentencePiece](#sentencepiece).
في كل خطوة تدريب، يحدد خوارزمية Unigram خسارة (غالبًا ما يتم تعريفها على أنها اللوغاريتم) عبر بيانات التدريب بالنظر إلى المفردات الحالية ونموذج اللغة unigram. بعد ذلك، بالنسبة لكل رمز في المفردات، يحسب الخوارزمية مقدار زيادة الخسارة الإجمالية إذا تم إزالة الرمز من المفردات. ثم يقوم Unigram بإزالة p (مع p عادة ما تكون 10% أو 20%) في المائة من الرموز التي تكون زيادة الخسارة فيها هي الأدنى، *أي* تلك
@ -188,7 +188,7 @@ $$\mathcal{L} = -\sum_{i=1}^{N} \log \left ( \sum_{x \in S(x_{i})} p(x) \right )
تحتوي جميع خوارزميات توكنز الموصوفة حتى الآن على نفس المشكلة: من المفترض أن النص المدخل يستخدم المسافات لفصل الكلمات. ومع ذلك، لا تستخدم جميع اللغات المسافات لفصل الكلمات. أحد الحلول الممكنة هو استخداممعالج مسبق للغة محدد، *مثال* [XLM](model_doc/xlm) يلذي يستخدم معالجات مسبقة محددة للصينية واليابانية والتايلاندية.
لحل هذه المشكلة بشكل أعم، [SentencePiece: A simple and language independent subword tokenizer and
detokenizer for Neural Text Processing (Kudo et al.، 2018)](https://huggingface.co/papers/1808.06226) يتعامل مع المدخلات
detokenizer for Neural Text Processing (Kudo et al.، 2018)](https://arxiv.org/pdf/1808.06226.pdf) يتعامل مع المدخلات
كتدفق بيانات خام، وبالتالي يشمل المسافة في مجموعة الأحرف التي سيتم استخدامها. ثم يستخدم خوارزمية BPE أو unigram
لبناء المفردات المناسبة.

View File

@ -306,48 +306,78 @@ pip install galore-torch
ثم أضف ببساطة أحد `["galore_adamw"، "galore_adafactor"، "galore_adamw_8bit"]` في `optim` جنبًا إلى جنب مع `optim_target_modules`، والتي يمكن أن تكون قائمة من السلاسل أو التعبيرات النمطية regex أو المسار الكامل المطابق لأسماء الوحدات المستهدفة التي تريد تكييفها. فيما يلي مثال على النص البرمجي كامل(تأكد من `pip install trl datasets`):
```python
import torch
import datasets
from trl import SFTConfig, SFTTrainer
import trl
from transformers import TrainingArguments, AutoConfig, AutoTokenizer, AutoModelForCausalLM
train_dataset = datasets.load_dataset('imdb', split='train')
args = SFTConfig(
output_dir="./test-galore",
args = TrainingArguments(
output_dir="./test-galore"،
max_steps=100,
optim="galore_adamw",
optim_target_modules=[r".*.attn.*", r".*.mlp.*"],
gradient_checkpointing=True,
per_device_train_batch_size=2,
optim="galore_adamw"،
optim_target_modules=[r".*.attn.*"، r".*.mlp.*"]
)
trainer = SFTTrainer(
model="google/gemma-2b",
model_id = "google/gemma-2b"
config = AutoConfig.from_pretrained(model_id)
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_config(config).to(0)
trainer = trl.SFTTrainer(
model=model,
args=args,
train_dataset=train_dataset,
dataset_text_field='text',
max_seq_length=512,
)
trainer.train()
```
لتمرير معامﻻت إضافية يدعمها GaLore، يجب عليك تمرير `optim_args` بشكل صحيح، على سبيل المثال:
```python
import torch
import datasets
from trl import SFTConfig, SFTTrainer
import trl
from transformers import TrainingArguments, AutoConfig, AutoTokenizer, AutoModelForCausalLM
train_dataset = datasets.load_dataset('imdb', split='train')
args = SFTConfig(
args = TrainingArguments(
output_dir="./test-galore",
max_steps=100,
per_device_train_batch_size=2,
optim="galore_adamw",
optim_target_modules=[r".*.attn.*", r".*.mlp.*"],
optim_args="rank=64, update_proj_gap=100, scale=0.10",
gradient_checkpointing=True,
)
trainer = SFTTrainer(
model="google/gemma-2b",
model_id = "google/gemma-2b"
config = AutoConfig.from_pretrained(model_id)
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_config(config).to(0)
trainer = trl.SFTTrainer(
model=model,
args=args,
train_dataset=train_dataset,
dataset_text_field='text',
max_seq_length=512,
)
trainer.train()
```
يمكنك قراءة المزيد حول الطريقة في [المستودع الأصلي](https://github.com/jiaweizzhao/GaLore) أو [الورقة البحثية](https://huggingface.co/papers/2403.03507).
يمكنك قراءة المزيد حول الطريقة في [المستودع الأصلي](https://github.com/jiaweizzhao/GaLore) أو [الورقة البحثية](https://arxiv.org/abs/2403.03507).
حاليًا، يمكنك فقط تدريب الطبقات الخطية التي تعتبر طبقات GaLore وستستخدم التحلل ذو الرتبة المنخفضة للتدريب بينما سيتم تحسين الطبقات المتبقية بالطريقة التقليدية.
@ -356,22 +386,37 @@ trainer.train()
يمكنك أيضًا إجراء تحسين طبقة تلو الأخرى عن طريق إضافة `layerwise` إلى اسم المُحسِّن كما هو موضح أدناه:
```python
import torch
import datasets
from trl import SFTConfig, SFTTrainer
import trl
train_dataset = datasets.load_dataset('imdb', split='train')
args = SFTConfig(
output_dir="./test-galore",
max_steps=100,
optim="galore_adamw_layerwise",
optim_target_modules=[r".*.attn.*", r".*.mlp.*"],
gradient_checkpointing=True,
from transformers import TrainingArguments، AutoConfig، AutoTokenizer، AutoModelForCausalLM
train_dataset = datasets.load_dataset('imdb'، split='train')
args = TrainingArguments(
output_dir="./test-galore"،
max_steps=100،
per_device_train_batch_size=2،
optim="galore_adamw_layerwise"،
optim_target_modules=[r".*.attn.*"، r".*.mlp.*"]
)
trainer = SFTTrainer(
model="google/gemma-2b",
args=args,
train_dataset=train_dataset,
model_id = "google/gemma-2b"
config = AutoConfig.from_pretrained(model_id)
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_config(config).to(0)
trainer = trl.SFTTrainer(
model=model،
args=args،
train_dataset=train_dataset،
dataset_text_field='text'،
max_seq_length=512،
)
trainer.train()
```
@ -391,21 +436,39 @@ trainer.train()
فيما يلي نص برمجي بسيط يوضح كيفية ضبط نموذج [google/gemma-2b](https://huggingface.co/google/gemma-2b) على مجموعة بيانات IMDB في الدقة الكاملة:
```python
import torch
import datasets
from trl import SFTConfig, SFTTrainer
from transformers import TrainingArguments، AutoTokenizer، AutoModelForCausalLM
import trl
train_dataset = datasets.load_dataset('imdb', split='train')
args = SFTConfig(
output_dir="./test-lomo",
max_steps=100,
optim="adalomo",
gradient_checkpointing=True,
train_dataset = datasets.load_dataset('imdb'، split='train')
args = TrainingArguments(
output_dir="./test-lomo"،
max_steps=100،
per_device_train_batch_size=4،
optim="adalomo"،
gradient_checkpointing=True،
logging_strategy="steps"،
logging_steps=1،
learning_rate=2e-6،
save_strategy="no"،
run_name="lomo-imdb"،
)
trainer = SFTTrainer(
model="google/gemma-2b",
args=args,
train_dataset=train_dataset,
model_id = "google/gemma-2b"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id، low_cpu_mem_usage=True).to(0)
trainer = trl.SFTTrainer(
model=model،
args=args،
train_dataset=train_dataset،
dataset_text_field='text'،
max_seq_length=1024،
)
trainer.train()
```
@ -440,7 +503,7 @@ args = TrainingArguments(
# تحميل النموذج والمجزىء اللغوي
model_id = "google/gemma-2b"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id).to(0)
model = AutoModelForCausalLM.from_pretrained(model_id, low_cpu_mem_usage=True).to(0)
# تهيئة المدرب
trainer = Trainer(
@ -461,21 +524,39 @@ trainer.train()
فيما يلي نص برمجى بسيط لشرح كيفية ضبط [google/gemma-2b](https://huggingface.co/google/gemma-2b) بدقة على مجموعة بيانات IMDB بدقة كاملة:
```python
import torch
import datasets
from trl import SFTConfig, SFTTrainer
from transformers import TrainingArguments, AutoTokenizer, AutoModelForCausalLM
import trl
train_dataset = datasets.load_dataset('imdb', split='train')
args = SFTConfig(
output_dir="./test-galore",
max_steps=100,
args = TrainingArguments(
output_dir="./test-schedulefree",
max_steps=1000,
per_device_train_batch_size=4,
optim="schedule_free_adamw",
gradient_checkpointing=True,
logging_strategy="steps",
logging_steps=1,
learning_rate=2e-6,
save_strategy="no",
run_name="sfo-imdb",
)
trainer = SFTTrainer(
model="google/gemma-2b",
model_id = "google/gemma-2b"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, low_cpu_mem_usage=True).to(0)
trainer = trl.SFTTrainer(
model=model,
args=args,
train_dataset=train_dataset,
dataset_text_field='text',
max_seq_length=1024,
)
trainer.train()
```
## تسريع ومدرب

View File

@ -55,148 +55,148 @@ Die Bibliothek enthält derzeit JAX-, PyTorch- und TensorFlow-Implementierungen,
<!--This list is updated automatically from the README with _make fix-copies_. Do not update manually! -->
1. **[ALBERT](model_doc/albert)** (from Google Research and the Toyota Technological Institute at Chicago) released with the paper [ALBERT: A Lite BERT for Self-supervised Learning of Language Representations](https://huggingface.co/papers/1909.11942), by Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, Radu Soricut.
1. **[ALIGN](model_doc/align)** (from Google Research) released with the paper [Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision](https://huggingface.co/papers/2102.05918) by Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yunhsuan Sung, Zhen Li, Tom Duerig.
1. **[BART](model_doc/bart)** (from Facebook) released with the paper [BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension](https://huggingface.co/papers/1910.13461) by Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov and Luke Zettlemoyer.
1. **[BARThez](model_doc/barthez)** (from École polytechnique) released with the paper [BARThez: a Skilled Pretrained French Sequence-to-Sequence Model](https://huggingface.co/papers/2010.12321) by Moussa Kamal Eddine, Antoine J.-P. Tixier, Michalis Vazirgiannis.
1. **[BARTpho](model_doc/bartpho)** (from VinAI Research) released with the paper [BARTpho: Pre-trained Sequence-to-Sequence Models for Vietnamese](https://huggingface.co/papers/2109.09701) by Nguyen Luong Tran, Duong Minh Le and Dat Quoc Nguyen.
1. **[BEiT](model_doc/beit)** (from Microsoft) released with the paper [BEiT: BERT Pre-Training of Image Transformers](https://huggingface.co/papers/2106.08254) by Hangbo Bao, Li Dong, Furu Wei.
1. **[BERT](model_doc/bert)** (from Google) released with the paper [BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding](https://huggingface.co/papers/1810.04805) by Jacob Devlin, Ming-Wei Chang, Kenton Lee and Kristina Toutanova.
1. **[BERT For Sequence Generation](model_doc/bert-generation)** (from Google) released with the paper [Leveraging Pre-trained Checkpoints for Sequence Generation Tasks](https://huggingface.co/papers/1907.12461) by Sascha Rothe, Shashi Narayan, Aliaksei Severyn.
1. **[ALBERT](model_doc/albert)** (from Google Research and the Toyota Technological Institute at Chicago) released with the paper [ALBERT: A Lite BERT for Self-supervised Learning of Language Representations](https://arxiv.org/abs/1909.11942), by Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, Radu Soricut.
1. **[ALIGN](model_doc/align)** (from Google Research) released with the paper [Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision](https://arxiv.org/abs/2102.05918) by Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yunhsuan Sung, Zhen Li, Tom Duerig.
1. **[BART](model_doc/bart)** (from Facebook) released with the paper [BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension](https://arxiv.org/abs/1910.13461) by Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov and Luke Zettlemoyer.
1. **[BARThez](model_doc/barthez)** (from École polytechnique) released with the paper [BARThez: a Skilled Pretrained French Sequence-to-Sequence Model](https://arxiv.org/abs/2010.12321) by Moussa Kamal Eddine, Antoine J.-P. Tixier, Michalis Vazirgiannis.
1. **[BARTpho](model_doc/bartpho)** (from VinAI Research) released with the paper [BARTpho: Pre-trained Sequence-to-Sequence Models for Vietnamese](https://arxiv.org/abs/2109.09701) by Nguyen Luong Tran, Duong Minh Le and Dat Quoc Nguyen.
1. **[BEiT](model_doc/beit)** (from Microsoft) released with the paper [BEiT: BERT Pre-Training of Image Transformers](https://arxiv.org/abs/2106.08254) by Hangbo Bao, Li Dong, Furu Wei.
1. **[BERT](model_doc/bert)** (from Google) released with the paper [BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding](https://arxiv.org/abs/1810.04805) by Jacob Devlin, Ming-Wei Chang, Kenton Lee and Kristina Toutanova.
1. **[BERT For Sequence Generation](model_doc/bert-generation)** (from Google) released with the paper [Leveraging Pre-trained Checkpoints for Sequence Generation Tasks](https://arxiv.org/abs/1907.12461) by Sascha Rothe, Shashi Narayan, Aliaksei Severyn.
1. **[BERTweet](model_doc/bertweet)** (from VinAI Research) released with the paper [BERTweet: A pre-trained language model for English Tweets](https://aclanthology.org/2020.emnlp-demos.2/) by Dat Quoc Nguyen, Thanh Vu and Anh Tuan Nguyen.
1. **[BigBird-Pegasus](model_doc/bigbird_pegasus)** (from Google Research) released with the paper [Big Bird: Transformers for Longer Sequences](https://huggingface.co/papers/2007.14062) by Manzil Zaheer, Guru Guruganesh, Avinava Dubey, Joshua Ainslie, Chris Alberti, Santiago Ontanon, Philip Pham, Anirudh Ravula, Qifan Wang, Li Yang, Amr Ahmed.
1. **[BigBird-RoBERTa](model_doc/big_bird)** (from Google Research) released with the paper [Big Bird: Transformers for Longer Sequences](https://huggingface.co/papers/2007.14062) by Manzil Zaheer, Guru Guruganesh, Avinava Dubey, Joshua Ainslie, Chris Alberti, Santiago Ontanon, Philip Pham, Anirudh Ravula, Qifan Wang, Li Yang, Amr Ahmed.
1. **[Blenderbot](model_doc/blenderbot)** (from Facebook) released with the paper [Recipes for building an open-domain chatbot](https://huggingface.co/papers/2004.13637) by Stephen Roller, Emily Dinan, Naman Goyal, Da Ju, Mary Williamson, Yinhan Liu, Jing Xu, Myle Ott, Kurt Shuster, Eric M. Smith, Y-Lan Boureau, Jason Weston.
1. **[BlenderbotSmall](model_doc/blenderbot-small)** (from Facebook) released with the paper [Recipes for building an open-domain chatbot](https://huggingface.co/papers/2004.13637) by Stephen Roller, Emily Dinan, Naman Goyal, Da Ju, Mary Williamson, Yinhan Liu, Jing Xu, Myle Ott, Kurt Shuster, Eric M. Smith, Y-Lan Boureau, Jason Weston.
1. **[BigBird-Pegasus](model_doc/bigbird_pegasus)** (from Google Research) released with the paper [Big Bird: Transformers for Longer Sequences](https://arxiv.org/abs/2007.14062) by Manzil Zaheer, Guru Guruganesh, Avinava Dubey, Joshua Ainslie, Chris Alberti, Santiago Ontanon, Philip Pham, Anirudh Ravula, Qifan Wang, Li Yang, Amr Ahmed.
1. **[BigBird-RoBERTa](model_doc/big_bird)** (from Google Research) released with the paper [Big Bird: Transformers for Longer Sequences](https://arxiv.org/abs/2007.14062) by Manzil Zaheer, Guru Guruganesh, Avinava Dubey, Joshua Ainslie, Chris Alberti, Santiago Ontanon, Philip Pham, Anirudh Ravula, Qifan Wang, Li Yang, Amr Ahmed.
1. **[Blenderbot](model_doc/blenderbot)** (from Facebook) released with the paper [Recipes for building an open-domain chatbot](https://arxiv.org/abs/2004.13637) by Stephen Roller, Emily Dinan, Naman Goyal, Da Ju, Mary Williamson, Yinhan Liu, Jing Xu, Myle Ott, Kurt Shuster, Eric M. Smith, Y-Lan Boureau, Jason Weston.
1. **[BlenderbotSmall](model_doc/blenderbot-small)** (from Facebook) released with the paper [Recipes for building an open-domain chatbot](https://arxiv.org/abs/2004.13637) by Stephen Roller, Emily Dinan, Naman Goyal, Da Ju, Mary Williamson, Yinhan Liu, Jing Xu, Myle Ott, Kurt Shuster, Eric M. Smith, Y-Lan Boureau, Jason Weston.
1. **[BLOOM](model_doc/bloom)** (from BigScience workshop) released by the [BigScience Workshop](https://bigscience.huggingface.co/).
1. **[BORT](model_doc/bort)** (from Alexa) released with the paper [Optimal Subarchitecture Extraction For BERT](https://huggingface.co/papers/2010.10499) by Adrian de Wynter and Daniel J. Perry.
1. **[ByT5](model_doc/byt5)** (from Google Research) released with the paper [ByT5: Towards a token-free future with pre-trained byte-to-byte models](https://huggingface.co/papers/2105.13626) by Linting Xue, Aditya Barua, Noah Constant, Rami Al-Rfou, Sharan Narang, Mihir Kale, Adam Roberts, Colin Raffel.
1. **[CamemBERT](model_doc/camembert)** (from Inria/Facebook/Sorbonne) released with the paper [CamemBERT: a Tasty French Language Model](https://huggingface.co/papers/1911.03894) by Louis Martin*, Benjamin Muller*, Pedro Javier Ortiz Suárez*, Yoann Dupont, Laurent Romary, Éric Villemonte de la Clergerie, Djamé Seddah and Benoît Sagot.
1. **[CANINE](model_doc/canine)** (from Google Research) released with the paper [CANINE: Pre-training an Efficient Tokenization-Free Encoder for Language Representation](https://huggingface.co/papers/2103.06874) by Jonathan H. Clark, Dan Garrette, Iulia Turc, John Wieting.
1. **[CLIP](model_doc/clip)** (from OpenAI) released with the paper [Learning Transferable Visual Models From Natural Language Supervision](https://huggingface.co/papers/2103.00020) by Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, Ilya Sutskever.
1. **[CodeGen](model_doc/codegen)** (from Salesforce) released with the paper [A Conversational Paradigm for Program Synthesis](https://huggingface.co/papers/2203.13474) by Erik Nijkamp, Bo Pang, Hiroaki Hayashi, Lifu Tu, Huan Wang, Yingbo Zhou, Silvio Savarese, Caiming Xiong.
1. **[ConvBERT](model_doc/convbert)** (from YituTech) released with the paper [ConvBERT: Improving BERT with Span-based Dynamic Convolution](https://huggingface.co/papers/2008.02496) by Zihang Jiang, Weihao Yu, Daquan Zhou, Yunpeng Chen, Jiashi Feng, Shuicheng Yan.
1. **[ConvNeXT](model_doc/convnext)** (from Facebook AI) released with the paper [A ConvNet for the 2020s](https://huggingface.co/papers/2201.03545) by Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, Saining Xie.
1. **[ConvNeXTV2](model_doc/convnextv2)** (from Facebook AI) released with the paper [ConvNeXt V2: Co-designing and Scaling ConvNets with Masked Autoencoders](https://huggingface.co/papers/2301.00808) by Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, Saining Xie.
1. **[CPM](model_doc/cpm)** (from Tsinghua University) released with the paper [CPM: A Large-scale Generative Chinese Pre-trained Language Model](https://huggingface.co/papers/2012.00413) by Zhengyan Zhang, Xu Han, Hao Zhou, Pei Ke, Yuxian Gu, Deming Ye, Yujia Qin, Yusheng Su, Haozhe Ji, Jian Guan, Fanchao Qi, Xiaozhi Wang, Yanan Zheng, Guoyang Zeng, Huanqi Cao, Shengqi Chen, Daixuan Li, Zhenbo Sun, Zhiyuan Liu, Minlie Huang, Wentao Han, Jie Tang, Juanzi Li, Xiaoyan Zhu, Maosong Sun.
1. **[CTRL](model_doc/ctrl)** (from Salesforce) released with the paper [CTRL: A Conditional Transformer Language Model for Controllable Generation](https://huggingface.co/papers/1909.05858) by Nitish Shirish Keskar*, Bryan McCann*, Lav R. Varshney, Caiming Xiong and Richard Socher.
1. **[CvT](model_doc/cvt)** (from Microsoft) released with the paper [CvT: Introducing Convolutions to Vision Transformers](https://huggingface.co/papers/2103.15808) by Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, Lei Zhang.
1. **[Data2Vec](model_doc/data2vec)** (from Facebook) released with the paper [Data2Vec: A General Framework for Self-supervised Learning in Speech, Vision and Language](https://huggingface.co/papers/2202.03555) by Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, Michael Auli.
1. **[DeBERTa](model_doc/deberta)** (from Microsoft) released with the paper [DeBERTa: Decoding-enhanced BERT with Disentangled Attention](https://huggingface.co/papers/2006.03654) by Pengcheng He, Xiaodong Liu, Jianfeng Gao, Weizhu Chen.
1. **[DeBERTa-v2](model_doc/deberta-v2)** (from Microsoft) released with the paper [DeBERTa: Decoding-enhanced BERT with Disentangled Attention](https://huggingface.co/papers/2006.03654) by Pengcheng He, Xiaodong Liu, Jianfeng Gao, Weizhu Chen.
1. **[Decision Transformer](model_doc/decision_transformer)** (from Berkeley/Facebook/Google) released with the paper [Decision Transformer: Reinforcement Learning via Sequence Modeling](https://huggingface.co/papers/2106.01345) by Lili Chen, Kevin Lu, Aravind Rajeswaran, Kimin Lee, Aditya Grover, Michael Laskin, Pieter Abbeel, Aravind Srinivas, Igor Mordatch.
1. **[DeiT](model_doc/deit)** (from Facebook) released with the paper [Training data-efficient image transformers & distillation through attention](https://huggingface.co/papers/2012.12877) by Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, Hervé Jégou.
1. **[DETR](model_doc/detr)** (from Facebook) released with the paper [End-to-End Object Detection with Transformers](https://huggingface.co/papers/2005.12872) by Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, Sergey Zagoruyko.
1. **[DialoGPT](model_doc/dialogpt)** (from Microsoft Research) released with the paper [DialoGPT: Large-Scale Generative Pre-training for Conversational Response Generation](https://huggingface.co/papers/1911.00536) by Yizhe Zhang, Siqi Sun, Michel Galley, Yen-Chun Chen, Chris Brockett, Xiang Gao, Jianfeng Gao, Jingjing Liu, Bill Dolan.
1. **[DistilBERT](model_doc/distilbert)** (from HuggingFace), released together with the paper [DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter](https://huggingface.co/papers/1910.01108) by Victor Sanh, Lysandre Debut and Thomas Wolf. The same method has been applied to compress GPT2 into [DistilGPT2](https://github.com/huggingface/transformers-research-projects/tree/main/distillation), RoBERTa into [DistilRoBERTa](https://github.com/huggingface/transformers-research-projects/tree/main/distillation), Multilingual BERT into [DistilmBERT](https://github.com/huggingface/transformers-research-projects/tree/main/distillation) and a German version of DistilBERT.
1. **[DiT](model_doc/dit)** (from Microsoft Research) released with the paper [DiT: Self-supervised Pre-training for Document Image Transformer](https://huggingface.co/papers/2203.02378) by Junlong Li, Yiheng Xu, Tengchao Lv, Lei Cui, Cha Zhang, Furu Wei.
1. **[DPR](model_doc/dpr)** (from Facebook) released with the paper [Dense Passage Retrieval for Open-Domain Question Answering](https://huggingface.co/papers/2004.04906) by Vladimir Karpukhin, Barlas Oğuz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih.
1. **[DPT](master/model_doc/dpt)** (from Intel Labs) released with the paper [Vision Transformers for Dense Prediction](https://huggingface.co/papers/2103.13413) by René Ranftl, Alexey Bochkovskiy, Vladlen Koltun.
1. **[EfficientNet](model_doc/efficientnet)** (from Google Research) released with the paper [EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks](https://huggingface.co/papers/1905.11946) by Mingxing Tan and Quoc V. Le.
1. **[ELECTRA](model_doc/electra)** (from Google Research/Stanford University) released with the paper [ELECTRA: Pre-training text encoders as discriminators rather than generators](https://huggingface.co/papers/2003.10555) by Kevin Clark, Minh-Thang Luong, Quoc V. Le, Christopher D. Manning.
1. **[EncoderDecoder](model_doc/encoder-decoder)** (from Google Research) released with the paper [Leveraging Pre-trained Checkpoints for Sequence Generation Tasks](https://huggingface.co/papers/1907.12461) by Sascha Rothe, Shashi Narayan, Aliaksei Severyn.
1. **[FlauBERT](model_doc/flaubert)** (from CNRS) released with the paper [FlauBERT: Unsupervised Language Model Pre-training for French](https://huggingface.co/papers/1912.05372) by Hang Le, Loïc Vial, Jibril Frej, Vincent Segonne, Maximin Coavoux, Benjamin Lecouteux, Alexandre Allauzen, Benoît Crabbé, Laurent Besacier, Didier Schwab.
1. **[FLAVA](model_doc/flava)** (from Facebook AI) released with the paper [FLAVA: A Foundational Language And Vision Alignment Model](https://huggingface.co/papers/2112.04482) by Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela.
1. **[FNet](model_doc/fnet)** (from Google Research) released with the paper [FNet: Mixing Tokens with Fourier Transforms](https://huggingface.co/papers/2105.03824) by James Lee-Thorp, Joshua Ainslie, Ilya Eckstein, Santiago Ontanon.
1. **[Funnel Transformer](model_doc/funnel)** (from CMU/Google Brain) released with the paper [Funnel-Transformer: Filtering out Sequential Redundancy for Efficient Language Processing](https://huggingface.co/papers/2006.03236) by Zihang Dai, Guokun Lai, Yiming Yang, Quoc V. Le.
1. **[GLPN](model_doc/glpn)** (from KAIST) released with the paper [Global-Local Path Networks for Monocular Depth Estimation with Vertical CutDepth](https://huggingface.co/papers/2201.07436) by Doyeon Kim, Woonghyun Ga, Pyungwhan Ahn, Donggyu Joo, Sehwan Chun, Junmo Kim.
1. **[BORT](model_doc/bort)** (from Alexa) released with the paper [Optimal Subarchitecture Extraction For BERT](https://arxiv.org/abs/2010.10499) by Adrian de Wynter and Daniel J. Perry.
1. **[ByT5](model_doc/byt5)** (from Google Research) released with the paper [ByT5: Towards a token-free future with pre-trained byte-to-byte models](https://arxiv.org/abs/2105.13626) by Linting Xue, Aditya Barua, Noah Constant, Rami Al-Rfou, Sharan Narang, Mihir Kale, Adam Roberts, Colin Raffel.
1. **[CamemBERT](model_doc/camembert)** (from Inria/Facebook/Sorbonne) released with the paper [CamemBERT: a Tasty French Language Model](https://arxiv.org/abs/1911.03894) by Louis Martin*, Benjamin Muller*, Pedro Javier Ortiz Suárez*, Yoann Dupont, Laurent Romary, Éric Villemonte de la Clergerie, Djamé Seddah and Benoît Sagot.
1. **[CANINE](model_doc/canine)** (from Google Research) released with the paper [CANINE: Pre-training an Efficient Tokenization-Free Encoder for Language Representation](https://arxiv.org/abs/2103.06874) by Jonathan H. Clark, Dan Garrette, Iulia Turc, John Wieting.
1. **[CLIP](model_doc/clip)** (from OpenAI) released with the paper [Learning Transferable Visual Models From Natural Language Supervision](https://arxiv.org/abs/2103.00020) by Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, Ilya Sutskever.
1. **[CodeGen](model_doc/codegen)** (from Salesforce) released with the paper [A Conversational Paradigm for Program Synthesis](https://arxiv.org/abs/2203.13474) by Erik Nijkamp, Bo Pang, Hiroaki Hayashi, Lifu Tu, Huan Wang, Yingbo Zhou, Silvio Savarese, Caiming Xiong.
1. **[ConvBERT](model_doc/convbert)** (from YituTech) released with the paper [ConvBERT: Improving BERT with Span-based Dynamic Convolution](https://arxiv.org/abs/2008.02496) by Zihang Jiang, Weihao Yu, Daquan Zhou, Yunpeng Chen, Jiashi Feng, Shuicheng Yan.
1. **[ConvNeXT](model_doc/convnext)** (from Facebook AI) released with the paper [A ConvNet for the 2020s](https://arxiv.org/abs/2201.03545) by Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, Saining Xie.
1. **[ConvNeXTV2](model_doc/convnextv2)** (from Facebook AI) released with the paper [ConvNeXt V2: Co-designing and Scaling ConvNets with Masked Autoencoders](https://arxiv.org/abs/2301.00808) by Sanghyun Woo, Shoubhik Debnath, Ronghang Hu, Xinlei Chen, Zhuang Liu, In So Kweon, Saining Xie.
1. **[CPM](model_doc/cpm)** (from Tsinghua University) released with the paper [CPM: A Large-scale Generative Chinese Pre-trained Language Model](https://arxiv.org/abs/2012.00413) by Zhengyan Zhang, Xu Han, Hao Zhou, Pei Ke, Yuxian Gu, Deming Ye, Yujia Qin, Yusheng Su, Haozhe Ji, Jian Guan, Fanchao Qi, Xiaozhi Wang, Yanan Zheng, Guoyang Zeng, Huanqi Cao, Shengqi Chen, Daixuan Li, Zhenbo Sun, Zhiyuan Liu, Minlie Huang, Wentao Han, Jie Tang, Juanzi Li, Xiaoyan Zhu, Maosong Sun.
1. **[CTRL](model_doc/ctrl)** (from Salesforce) released with the paper [CTRL: A Conditional Transformer Language Model for Controllable Generation](https://arxiv.org/abs/1909.05858) by Nitish Shirish Keskar*, Bryan McCann*, Lav R. Varshney, Caiming Xiong and Richard Socher.
1. **[CvT](model_doc/cvt)** (from Microsoft) released with the paper [CvT: Introducing Convolutions to Vision Transformers](https://arxiv.org/abs/2103.15808) by Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, Lei Zhang.
1. **[Data2Vec](model_doc/data2vec)** (from Facebook) released with the paper [Data2Vec: A General Framework for Self-supervised Learning in Speech, Vision and Language](https://arxiv.org/abs/2202.03555) by Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, Michael Auli.
1. **[DeBERTa](model_doc/deberta)** (from Microsoft) released with the paper [DeBERTa: Decoding-enhanced BERT with Disentangled Attention](https://arxiv.org/abs/2006.03654) by Pengcheng He, Xiaodong Liu, Jianfeng Gao, Weizhu Chen.
1. **[DeBERTa-v2](model_doc/deberta-v2)** (from Microsoft) released with the paper [DeBERTa: Decoding-enhanced BERT with Disentangled Attention](https://arxiv.org/abs/2006.03654) by Pengcheng He, Xiaodong Liu, Jianfeng Gao, Weizhu Chen.
1. **[Decision Transformer](model_doc/decision_transformer)** (from Berkeley/Facebook/Google) released with the paper [Decision Transformer: Reinforcement Learning via Sequence Modeling](https://arxiv.org/abs/2106.01345) by Lili Chen, Kevin Lu, Aravind Rajeswaran, Kimin Lee, Aditya Grover, Michael Laskin, Pieter Abbeel, Aravind Srinivas, Igor Mordatch.
1. **[DeiT](model_doc/deit)** (from Facebook) released with the paper [Training data-efficient image transformers & distillation through attention](https://arxiv.org/abs/2012.12877) by Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, Hervé Jégou.
1. **[DETR](model_doc/detr)** (from Facebook) released with the paper [End-to-End Object Detection with Transformers](https://arxiv.org/abs/2005.12872) by Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, Sergey Zagoruyko.
1. **[DialoGPT](model_doc/dialogpt)** (from Microsoft Research) released with the paper [DialoGPT: Large-Scale Generative Pre-training for Conversational Response Generation](https://arxiv.org/abs/1911.00536) by Yizhe Zhang, Siqi Sun, Michel Galley, Yen-Chun Chen, Chris Brockett, Xiang Gao, Jianfeng Gao, Jingjing Liu, Bill Dolan.
1. **[DistilBERT](model_doc/distilbert)** (from HuggingFace), released together with the paper [DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter](https://arxiv.org/abs/1910.01108) by Victor Sanh, Lysandre Debut and Thomas Wolf. The same method has been applied to compress GPT2 into [DistilGPT2](https://github.com/huggingface/transformers-research-projects/tree/main/distillation), RoBERTa into [DistilRoBERTa](https://github.com/huggingface/transformers-research-projects/tree/main/distillation), Multilingual BERT into [DistilmBERT](https://github.com/huggingface/transformers-research-projects/tree/main/distillation) and a German version of DistilBERT.
1. **[DiT](model_doc/dit)** (from Microsoft Research) released with the paper [DiT: Self-supervised Pre-training for Document Image Transformer](https://arxiv.org/abs/2203.02378) by Junlong Li, Yiheng Xu, Tengchao Lv, Lei Cui, Cha Zhang, Furu Wei.
1. **[DPR](model_doc/dpr)** (from Facebook) released with the paper [Dense Passage Retrieval for Open-Domain Question Answering](https://arxiv.org/abs/2004.04906) by Vladimir Karpukhin, Barlas Oğuz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih.
1. **[DPT](master/model_doc/dpt)** (from Intel Labs) released with the paper [Vision Transformers for Dense Prediction](https://arxiv.org/abs/2103.13413) by René Ranftl, Alexey Bochkovskiy, Vladlen Koltun.
1. **[EfficientNet](model_doc/efficientnet)** (from Google Research) released with the paper [EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks](https://arxiv.org/abs/1905.11946) by Mingxing Tan and Quoc V. Le.
1. **[ELECTRA](model_doc/electra)** (from Google Research/Stanford University) released with the paper [ELECTRA: Pre-training text encoders as discriminators rather than generators](https://arxiv.org/abs/2003.10555) by Kevin Clark, Minh-Thang Luong, Quoc V. Le, Christopher D. Manning.
1. **[EncoderDecoder](model_doc/encoder-decoder)** (from Google Research) released with the paper [Leveraging Pre-trained Checkpoints for Sequence Generation Tasks](https://arxiv.org/abs/1907.12461) by Sascha Rothe, Shashi Narayan, Aliaksei Severyn.
1. **[FlauBERT](model_doc/flaubert)** (from CNRS) released with the paper [FlauBERT: Unsupervised Language Model Pre-training for French](https://arxiv.org/abs/1912.05372) by Hang Le, Loïc Vial, Jibril Frej, Vincent Segonne, Maximin Coavoux, Benjamin Lecouteux, Alexandre Allauzen, Benoît Crabbé, Laurent Besacier, Didier Schwab.
1. **[FLAVA](model_doc/flava)** (from Facebook AI) released with the paper [FLAVA: A Foundational Language And Vision Alignment Model](https://arxiv.org/abs/2112.04482) by Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela.
1. **[FNet](model_doc/fnet)** (from Google Research) released with the paper [FNet: Mixing Tokens with Fourier Transforms](https://arxiv.org/abs/2105.03824) by James Lee-Thorp, Joshua Ainslie, Ilya Eckstein, Santiago Ontanon.
1. **[Funnel Transformer](model_doc/funnel)** (from CMU/Google Brain) released with the paper [Funnel-Transformer: Filtering out Sequential Redundancy for Efficient Language Processing](https://arxiv.org/abs/2006.03236) by Zihang Dai, Guokun Lai, Yiming Yang, Quoc V. Le.
1. **[GLPN](model_doc/glpn)** (from KAIST) released with the paper [Global-Local Path Networks for Monocular Depth Estimation with Vertical CutDepth](https://arxiv.org/abs/2201.07436) by Doyeon Kim, Woonghyun Ga, Pyungwhan Ahn, Donggyu Joo, Sehwan Chun, Junmo Kim.
1. **[GPT](model_doc/openai-gpt)** (from OpenAI) released with the paper [Improving Language Understanding by Generative Pre-Training](https://openai.com/research/language-unsupervised/) by Alec Radford, Karthik Narasimhan, Tim Salimans and Ilya Sutskever.
1. **[GPT Neo](model_doc/gpt_neo)** (from EleutherAI) released in the repository [EleutherAI/gpt-neo](https://github.com/EleutherAI/gpt-neo) by Sid Black, Stella Biderman, Leo Gao, Phil Wang and Connor Leahy.
1. **[GPT NeoX](model_doc/gpt_neox)** (from EleutherAI) released with the paper [GPT-NeoX-20B: An Open-Source Autoregressive Language Model](https://huggingface.co/papers/2204.06745) by Sid Black, Stella Biderman, Eric Hallahan, Quentin Anthony, Leo Gao, Laurence Golding, Horace He, Connor Leahy, Kyle McDonell, Jason Phang, Michael Pieler, USVSN Sai Prashanth, Shivanshu Purohit, Laria Reynolds, Jonathan Tow, Ben Wang, Samuel Weinbach
1. **[GPT NeoX](model_doc/gpt_neox)** (from EleutherAI) released with the paper [GPT-NeoX-20B: An Open-Source Autoregressive Language Model](https://arxiv.org/abs/2204.06745) by Sid Black, Stella Biderman, Eric Hallahan, Quentin Anthony, Leo Gao, Laurence Golding, Horace He, Connor Leahy, Kyle McDonell, Jason Phang, Michael Pieler, USVSN Sai Prashanth, Shivanshu Purohit, Laria Reynolds, Jonathan Tow, Ben Wang, Samuel Weinbach
1. **[GPT-2](model_doc/gpt2)** (from OpenAI) released with the paper [Language Models are Unsupervised Multitask Learners](https://openai.com/research/better-language-models/) by Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei and Ilya Sutskever.
1. **[GPT-J](model_doc/gptj)** (from EleutherAI) released in the repository [kingoflolz/mesh-transformer-jax](https://github.com/kingoflolz/mesh-transformer-jax/) by Ben Wang and Aran Komatsuzaki.
1. **[GPTSAN-japanese](model_doc/gptsan-japanese)** released in the repository [tanreinama/GPTSAN](https://github.com/tanreinama/GPTSAN/blob/main/report/model.md) by Toshiyuki Sakamoto(tanreinama).
1. **[GroupViT](model_doc/groupvit)** (from UCSD, NVIDIA) released with the paper [GroupViT: Semantic Segmentation Emerges from Text Supervision](https://huggingface.co/papers/2202.11094) by Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, Xiaolong Wang.
1. **[Hubert](model_doc/hubert)** (from Facebook) released with the paper [HuBERT: Self-Supervised Speech Representation Learning by Masked Prediction of Hidden Units](https://huggingface.co/papers/2106.07447) by Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, Abdelrahman Mohamed.
1. **[I-BERT](model_doc/ibert)** (from Berkeley) released with the paper [I-BERT: Integer-only BERT Quantization](https://huggingface.co/papers/2101.01321) by Sehoon Kim, Amir Gholami, Zhewei Yao, Michael W. Mahoney, Kurt Keutzer.
1. **[GroupViT](model_doc/groupvit)** (from UCSD, NVIDIA) released with the paper [GroupViT: Semantic Segmentation Emerges from Text Supervision](https://arxiv.org/abs/2202.11094) by Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, Xiaolong Wang.
1. **[Hubert](model_doc/hubert)** (from Facebook) released with the paper [HuBERT: Self-Supervised Speech Representation Learning by Masked Prediction of Hidden Units](https://arxiv.org/abs/2106.07447) by Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, Abdelrahman Mohamed.
1. **[I-BERT](model_doc/ibert)** (from Berkeley) released with the paper [I-BERT: Integer-only BERT Quantization](https://arxiv.org/abs/2101.01321) by Sehoon Kim, Amir Gholami, Zhewei Yao, Michael W. Mahoney, Kurt Keutzer.
1. **[ImageGPT](model_doc/imagegpt)** (from OpenAI) released with the paper [Generative Pretraining from Pixels](https://openai.com/blog/image-gpt/) by Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, Ilya Sutskever.
1. **[LayoutLM](model_doc/layoutlm)** (from Microsoft Research Asia) released with the paper [LayoutLM: Pre-training of Text and Layout for Document Image Understanding](https://huggingface.co/papers/1912.13318) by Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, Ming Zhou.
1. **[LayoutLMv2](model_doc/layoutlmv2)** (from Microsoft Research Asia) released with the paper [LayoutLMv2: Multi-modal Pre-training for Visually-Rich Document Understanding](https://huggingface.co/papers/2012.14740) by Yang Xu, Yiheng Xu, Tengchao Lv, Lei Cui, Furu Wei, Guoxin Wang, Yijuan Lu, Dinei Florencio, Cha Zhang, Wanxiang Che, Min Zhang, Lidong Zhou.
1. **[LayoutLMv3](model_doc/layoutlmv3)** (from Microsoft Research Asia) released with the paper [LayoutLMv3: Pre-training for Document AI with Unified Text and Image Masking](https://huggingface.co/papers/2204.08387) by Yupan Huang, Tengchao Lv, Lei Cui, Yutong Lu, Furu Wei.
1. **[LayoutXLM](model_doc/layoutxlm)** (from Microsoft Research Asia) released with the paper [LayoutXLM: Multimodal Pre-training for Multilingual Visually-rich Document Understanding](https://huggingface.co/papers/2104.08836) by Yiheng Xu, Tengchao Lv, Lei Cui, Guoxin Wang, Yijuan Lu, Dinei Florencio, Cha Zhang, Furu Wei.
1. **[LED](model_doc/led)** (from AllenAI) released with the paper [Longformer: The Long-Document Transformer](https://huggingface.co/papers/2004.05150) by Iz Beltagy, Matthew E. Peters, Arman Cohan.
1. **[LeViT](model_doc/levit)** (from Meta AI) released with the paper [LeViT: A Vision Transformer in ConvNet's Clothing for Faster Inference](https://huggingface.co/papers/2104.01136) by Ben Graham, Alaaeldin El-Nouby, Hugo Touvron, Pierre Stock, Armand Joulin, Hervé Jégou, Matthijs Douze.
1. **[Longformer](model_doc/longformer)** (from AllenAI) released with the paper [Longformer: The Long-Document Transformer](https://huggingface.co/papers/2004.05150) by Iz Beltagy, Matthew E. Peters, Arman Cohan.
1. **[LongT5](model_doc/longt5)** (from Google AI) released with the paper [LongT5: Efficient Text-To-Text Transformer for Long Sequences](https://huggingface.co/papers/2112.07916) by Mandy Guo, Joshua Ainslie, David Uthus, Santiago Ontanon, Jianmo Ni, Yun-Hsuan Sung, Yinfei Yang.
1. **[LUKE](model_doc/luke)** (from Studio Ousia) released with the paper [LUKE: Deep Contextualized Entity Representations with Entity-aware Self-attention](https://huggingface.co/papers/2010.01057) by Ikuya Yamada, Akari Asai, Hiroyuki Shindo, Hideaki Takeda, Yuji Matsumoto.
1. **[LXMERT](model_doc/lxmert)** (from UNC Chapel Hill) released with the paper [LXMERT: Learning Cross-Modality Encoder Representations from Transformers for Open-Domain Question Answering](https://huggingface.co/papers/1908.07490) by Hao Tan and Mohit Bansal.
1. **[M-CTC-T](model_doc/mctct)** (from Facebook) released with the paper [Pseudo-Labeling For Massively Multilingual Speech Recognition](https://huggingface.co/papers/2111.00161) by Loren Lugosch, Tatiana Likhomanenko, Gabriel Synnaeve, and Ronan Collobert.
1. **[M2M100](model_doc/m2m_100)** (from Facebook) released with the paper [Beyond English-Centric Multilingual Machine Translation](https://huggingface.co/papers/2010.11125) by Angela Fan, Shruti Bhosale, Holger Schwenk, Zhiyi Ma, Ahmed El-Kishky, Siddharth Goyal, Mandeep Baines, Onur Celebi, Guillaume Wenzek, Vishrav Chaudhary, Naman Goyal, Tom Birch, Vitaliy Liptchinsky, Sergey Edunov, Edouard Grave, Michael Auli, Armand Joulin.
1. **[LayoutLM](model_doc/layoutlm)** (from Microsoft Research Asia) released with the paper [LayoutLM: Pre-training of Text and Layout for Document Image Understanding](https://arxiv.org/abs/1912.13318) by Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, Ming Zhou.
1. **[LayoutLMv2](model_doc/layoutlmv2)** (from Microsoft Research Asia) released with the paper [LayoutLMv2: Multi-modal Pre-training for Visually-Rich Document Understanding](https://arxiv.org/abs/2012.14740) by Yang Xu, Yiheng Xu, Tengchao Lv, Lei Cui, Furu Wei, Guoxin Wang, Yijuan Lu, Dinei Florencio, Cha Zhang, Wanxiang Che, Min Zhang, Lidong Zhou.
1. **[LayoutLMv3](model_doc/layoutlmv3)** (from Microsoft Research Asia) released with the paper [LayoutLMv3: Pre-training for Document AI with Unified Text and Image Masking](https://arxiv.org/abs/2204.08387) by Yupan Huang, Tengchao Lv, Lei Cui, Yutong Lu, Furu Wei.
1. **[LayoutXLM](model_doc/layoutxlm)** (from Microsoft Research Asia) released with the paper [LayoutXLM: Multimodal Pre-training for Multilingual Visually-rich Document Understanding](https://arxiv.org/abs/2104.08836) by Yiheng Xu, Tengchao Lv, Lei Cui, Guoxin Wang, Yijuan Lu, Dinei Florencio, Cha Zhang, Furu Wei.
1. **[LED](model_doc/led)** (from AllenAI) released with the paper [Longformer: The Long-Document Transformer](https://arxiv.org/abs/2004.05150) by Iz Beltagy, Matthew E. Peters, Arman Cohan.
1. **[LeViT](model_doc/levit)** (from Meta AI) released with the paper [LeViT: A Vision Transformer in ConvNet's Clothing for Faster Inference](https://arxiv.org/abs/2104.01136) by Ben Graham, Alaaeldin El-Nouby, Hugo Touvron, Pierre Stock, Armand Joulin, Hervé Jégou, Matthijs Douze.
1. **[Longformer](model_doc/longformer)** (from AllenAI) released with the paper [Longformer: The Long-Document Transformer](https://arxiv.org/abs/2004.05150) by Iz Beltagy, Matthew E. Peters, Arman Cohan.
1. **[LongT5](model_doc/longt5)** (from Google AI) released with the paper [LongT5: Efficient Text-To-Text Transformer for Long Sequences](https://arxiv.org/abs/2112.07916) by Mandy Guo, Joshua Ainslie, David Uthus, Santiago Ontanon, Jianmo Ni, Yun-Hsuan Sung, Yinfei Yang.
1. **[LUKE](model_doc/luke)** (from Studio Ousia) released with the paper [LUKE: Deep Contextualized Entity Representations with Entity-aware Self-attention](https://arxiv.org/abs/2010.01057) by Ikuya Yamada, Akari Asai, Hiroyuki Shindo, Hideaki Takeda, Yuji Matsumoto.
1. **[LXMERT](model_doc/lxmert)** (from UNC Chapel Hill) released with the paper [LXMERT: Learning Cross-Modality Encoder Representations from Transformers for Open-Domain Question Answering](https://arxiv.org/abs/1908.07490) by Hao Tan and Mohit Bansal.
1. **[M-CTC-T](model_doc/mctct)** (from Facebook) released with the paper [Pseudo-Labeling For Massively Multilingual Speech Recognition](https://arxiv.org/abs/2111.00161) by Loren Lugosch, Tatiana Likhomanenko, Gabriel Synnaeve, and Ronan Collobert.
1. **[M2M100](model_doc/m2m_100)** (from Facebook) released with the paper [Beyond English-Centric Multilingual Machine Translation](https://arxiv.org/abs/2010.11125) by Angela Fan, Shruti Bhosale, Holger Schwenk, Zhiyi Ma, Ahmed El-Kishky, Siddharth Goyal, Mandeep Baines, Onur Celebi, Guillaume Wenzek, Vishrav Chaudhary, Naman Goyal, Tom Birch, Vitaliy Liptchinsky, Sergey Edunov, Edouard Grave, Michael Auli, Armand Joulin.
1. **[MarianMT](model_doc/marian)** Machine translation models trained using [OPUS](http://opus.nlpl.eu/) data by Jörg Tiedemann. The [Marian Framework](https://marian-nmt.github.io/) is being developed by the Microsoft Translator Team.
1. **[Mask2Former](model_doc/mask2former)** (from FAIR and UIUC) released with the paper [Masked-attention Mask Transformer for Universal Image Segmentation](https://huggingface.co/papers/2112.01527) by Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, Rohit Girdhar.
1. **[MaskFormer](model_doc/maskformer)** (from Meta and UIUC) released with the paper [Per-Pixel Classification is Not All You Need for Semantic Segmentation](https://huggingface.co/papers/2107.06278) by Bowen Cheng, Alexander G. Schwing, Alexander Kirillov.
1. **[mBART](model_doc/mbart)** (from Facebook) released with the paper [Multilingual Denoising Pre-training for Neural Machine Translation](https://huggingface.co/papers/2001.08210) by Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvininejad, Mike Lewis, Luke Zettlemoyer.
1. **[mBART-50](model_doc/mbart)** (from Facebook) released with the paper [Multilingual Translation with Extensible Multilingual Pretraining and Finetuning](https://huggingface.co/papers/2008.00401) by Yuqing Tang, Chau Tran, Xian Li, Peng-Jen Chen, Naman Goyal, Vishrav Chaudhary, Jiatao Gu, Angela Fan.
1. **[Megatron-BERT](model_doc/megatron-bert)** (from NVIDIA) released with the paper [Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism](https://huggingface.co/papers/1909.08053) by Mohammad Shoeybi, Mostofa Patwary, Raul Puri, Patrick LeGresley, Jared Casper and Bryan Catanzaro.
1. **[Megatron-GPT2](model_doc/megatron_gpt2)** (from NVIDIA) released with the paper [Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism](https://huggingface.co/papers/1909.08053) by Mohammad Shoeybi, Mostofa Patwary, Raul Puri, Patrick LeGresley, Jared Casper and Bryan Catanzaro.
1. **[mLUKE](model_doc/mluke)** (from Studio Ousia) released with the paper [mLUKE: The Power of Entity Representations in Multilingual Pretrained Language Models](https://huggingface.co/papers/2110.08151) by Ryokan Ri, Ikuya Yamada, and Yoshimasa Tsuruoka.
1. **[MobileBERT](model_doc/mobilebert)** (from CMU/Google Brain) released with the paper [MobileBERT: a Compact Task-Agnostic BERT for Resource-Limited Devices](https://huggingface.co/papers/2004.02984) by Zhiqing Sun, Hongkun Yu, Xiaodan Song, Renjie Liu, Yiming Yang, and Denny Zhou.
1. **[MobileViT](model_doc/mobilevit)** (from Apple) released with the paper [MobileViT: Light-weight, General-purpose, and Mobile-friendly Vision Transformer](https://huggingface.co/papers/2110.02178) by Sachin Mehta and Mohammad Rastegari.
1. **[MPNet](model_doc/mpnet)** (from Microsoft Research) released with the paper [MPNet: Masked and Permuted Pre-training for Language Understanding](https://huggingface.co/papers/2004.09297) by Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, Tie-Yan Liu.
1. **[MT5](model_doc/mt5)** (from Google AI) released with the paper [mT5: A massively multilingual pre-trained text-to-text transformer](https://huggingface.co/papers/2010.11934) by Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, Colin Raffel.
1. **[MVP](model_doc/mvp)** (from RUC AI Box) released with the paper [MVP: Multi-task Supervised Pre-training for Natural Language Generation](https://huggingface.co/papers/2206.12131) by Tianyi Tang, Junyi Li, Wayne Xin Zhao and Ji-Rong Wen.
1. **[Nezha](model_doc/nezha)** (from Huawei Noahs Ark Lab) released with the paper [NEZHA: Neural Contextualized Representation for Chinese Language Understanding](https://huggingface.co/papers/1909.00204) by Junqiu Wei, Xiaozhe Ren, Xiaoguang Li, Wenyong Huang, Yi Liao, Yasheng Wang, Jiashu Lin, Xin Jiang, Xiao Chen and Qun Liu.
1. **[NLLB](model_doc/nllb)** (from Meta) released with the paper [No Language Left Behind: Scaling Human-Centered Machine Translation](https://huggingface.co/papers/2207.04672) by the NLLB team.
1. **[Nyströmformer](model_doc/nystromformer)** (from the University of Wisconsin - Madison) released with the paper [Nyströmformer: A Nyström-Based Algorithm for Approximating Self-Attention](https://huggingface.co/papers/2102.03902) by Yunyang Xiong, Zhanpeng Zeng, Rudrasis Chakraborty, Mingxing Tan, Glenn Fung, Yin Li, Vikas Singh.
1. **[OneFormer](model_doc/oneformer)** (from SHI Labs) released with the paper [OneFormer: One Transformer to Rule Universal Image Segmentation](https://huggingface.co/papers/2211.06220) by Jitesh Jain, Jiachen Li, MangTik Chiu, Ali Hassani, Nikita Orlov, Humphrey Shi.
1. **[OPT](master/model_doc/opt)** (from Meta AI) released with the paper [OPT: Open Pre-trained Transformer Language Models](https://huggingface.co/papers/2205.01068) by Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen et al.
1. **[OWL-ViT](model_doc/owlvit)** (from Google AI) released with the paper [Simple Open-Vocabulary Object Detection with Vision Transformers](https://huggingface.co/papers/2205.06230) by Matthias Minderer, Alexey Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, Xiao Wang, Xiaohua Zhai, Thomas Kipf, and Neil Houlsby.
1. **[Pegasus](model_doc/pegasus)** (from Google) released with the paper [PEGASUS: Pre-training with Extracted Gap-sentences for Abstractive Summarization](https://huggingface.co/papers/1912.08777) by Jingqing Zhang, Yao Zhao, Mohammad Saleh and Peter J. Liu.
1. **[Perceiver IO](model_doc/perceiver)** (from Deepmind) released with the paper [Perceiver IO: A General Architecture for Structured Inputs & Outputs](https://huggingface.co/papers/2107.14795) by Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, Skanda Koppula, Daniel Zoran, Andrew Brock, Evan Shelhamer, Olivier Hénaff, Matthew M. Botvinick, Andrew Zisserman, Oriol Vinyals, João Carreira.
1. **[Mask2Former](model_doc/mask2former)** (from FAIR and UIUC) released with the paper [Masked-attention Mask Transformer for Universal Image Segmentation](https://arxiv.org/abs/2112.01527) by Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, Rohit Girdhar.
1. **[MaskFormer](model_doc/maskformer)** (from Meta and UIUC) released with the paper [Per-Pixel Classification is Not All You Need for Semantic Segmentation](https://arxiv.org/abs/2107.06278) by Bowen Cheng, Alexander G. Schwing, Alexander Kirillov.
1. **[mBART](model_doc/mbart)** (from Facebook) released with the paper [Multilingual Denoising Pre-training for Neural Machine Translation](https://arxiv.org/abs/2001.08210) by Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvininejad, Mike Lewis, Luke Zettlemoyer.
1. **[mBART-50](model_doc/mbart)** (from Facebook) released with the paper [Multilingual Translation with Extensible Multilingual Pretraining and Finetuning](https://arxiv.org/abs/2008.00401) by Yuqing Tang, Chau Tran, Xian Li, Peng-Jen Chen, Naman Goyal, Vishrav Chaudhary, Jiatao Gu, Angela Fan.
1. **[Megatron-BERT](model_doc/megatron-bert)** (from NVIDIA) released with the paper [Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism](https://arxiv.org/abs/1909.08053) by Mohammad Shoeybi, Mostofa Patwary, Raul Puri, Patrick LeGresley, Jared Casper and Bryan Catanzaro.
1. **[Megatron-GPT2](model_doc/megatron_gpt2)** (from NVIDIA) released with the paper [Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism](https://arxiv.org/abs/1909.08053) by Mohammad Shoeybi, Mostofa Patwary, Raul Puri, Patrick LeGresley, Jared Casper and Bryan Catanzaro.
1. **[mLUKE](model_doc/mluke)** (from Studio Ousia) released with the paper [mLUKE: The Power of Entity Representations in Multilingual Pretrained Language Models](https://arxiv.org/abs/2110.08151) by Ryokan Ri, Ikuya Yamada, and Yoshimasa Tsuruoka.
1. **[MobileBERT](model_doc/mobilebert)** (from CMU/Google Brain) released with the paper [MobileBERT: a Compact Task-Agnostic BERT for Resource-Limited Devices](https://arxiv.org/abs/2004.02984) by Zhiqing Sun, Hongkun Yu, Xiaodan Song, Renjie Liu, Yiming Yang, and Denny Zhou.
1. **[MobileViT](model_doc/mobilevit)** (from Apple) released with the paper [MobileViT: Light-weight, General-purpose, and Mobile-friendly Vision Transformer](https://arxiv.org/abs/2110.02178) by Sachin Mehta and Mohammad Rastegari.
1. **[MPNet](model_doc/mpnet)** (from Microsoft Research) released with the paper [MPNet: Masked and Permuted Pre-training for Language Understanding](https://arxiv.org/abs/2004.09297) by Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, Tie-Yan Liu.
1. **[MT5](model_doc/mt5)** (from Google AI) released with the paper [mT5: A massively multilingual pre-trained text-to-text transformer](https://arxiv.org/abs/2010.11934) by Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, Colin Raffel.
1. **[MVP](model_doc/mvp)** (from RUC AI Box) released with the paper [MVP: Multi-task Supervised Pre-training for Natural Language Generation](https://arxiv.org/abs/2206.12131) by Tianyi Tang, Junyi Li, Wayne Xin Zhao and Ji-Rong Wen.
1. **[Nezha](model_doc/nezha)** (from Huawei Noahs Ark Lab) released with the paper [NEZHA: Neural Contextualized Representation for Chinese Language Understanding](https://arxiv.org/abs/1909.00204) by Junqiu Wei, Xiaozhe Ren, Xiaoguang Li, Wenyong Huang, Yi Liao, Yasheng Wang, Jiashu Lin, Xin Jiang, Xiao Chen and Qun Liu.
1. **[NLLB](model_doc/nllb)** (from Meta) released with the paper [No Language Left Behind: Scaling Human-Centered Machine Translation](https://arxiv.org/abs/2207.04672) by the NLLB team.
1. **[Nyströmformer](model_doc/nystromformer)** (from the University of Wisconsin - Madison) released with the paper [Nyströmformer: A Nyström-Based Algorithm for Approximating Self-Attention](https://arxiv.org/abs/2102.03902) by Yunyang Xiong, Zhanpeng Zeng, Rudrasis Chakraborty, Mingxing Tan, Glenn Fung, Yin Li, Vikas Singh.
1. **[OneFormer](model_doc/oneformer)** (from SHI Labs) released with the paper [OneFormer: One Transformer to Rule Universal Image Segmentation](https://arxiv.org/abs/2211.06220) by Jitesh Jain, Jiachen Li, MangTik Chiu, Ali Hassani, Nikita Orlov, Humphrey Shi.
1. **[OPT](master/model_doc/opt)** (from Meta AI) released with the paper [OPT: Open Pre-trained Transformer Language Models](https://arxiv.org/abs/2205.01068) by Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen et al.
1. **[OWL-ViT](model_doc/owlvit)** (from Google AI) released with the paper [Simple Open-Vocabulary Object Detection with Vision Transformers](https://arxiv.org/abs/2205.06230) by Matthias Minderer, Alexey Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, Xiao Wang, Xiaohua Zhai, Thomas Kipf, and Neil Houlsby.
1. **[Pegasus](model_doc/pegasus)** (from Google) released with the paper [PEGASUS: Pre-training with Extracted Gap-sentences for Abstractive Summarization](https://arxiv.org/abs/1912.08777) by Jingqing Zhang, Yao Zhao, Mohammad Saleh and Peter J. Liu.
1. **[Perceiver IO](model_doc/perceiver)** (from Deepmind) released with the paper [Perceiver IO: A General Architecture for Structured Inputs & Outputs](https://arxiv.org/abs/2107.14795) by Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, Skanda Koppula, Daniel Zoran, Andrew Brock, Evan Shelhamer, Olivier Hénaff, Matthew M. Botvinick, Andrew Zisserman, Oriol Vinyals, João Carreira.
1. **[PhoBERT](model_doc/phobert)** (from VinAI Research) released with the paper [PhoBERT: Pre-trained language models for Vietnamese](https://www.aclweb.org/anthology/2020.findings-emnlp.92/) by Dat Quoc Nguyen and Anh Tuan Nguyen.
1. **[PLBart](model_doc/plbart)** (from UCLA NLP) released with the paper [Unified Pre-training for Program Understanding and Generation](https://huggingface.co/papers/2103.06333) by Wasi Uddin Ahmad, Saikat Chakraborty, Baishakhi Ray, Kai-Wei Chang.
1. **[PoolFormer](model_doc/poolformer)** (from Sea AI Labs) released with the paper [MetaFormer is Actually What You Need for Vision](https://huggingface.co/papers/2111.11418) by Yu, Weihao and Luo, Mi and Zhou, Pan and Si, Chenyang and Zhou, Yichen and Wang, Xinchao and Feng, Jiashi and Yan, Shuicheng.
1. **[ProphetNet](model_doc/prophetnet)** (from Microsoft Research) released with the paper [ProphetNet: Predicting Future N-gram for Sequence-to-Sequence Pre-training](https://huggingface.co/papers/2001.04063) by Yu Yan, Weizhen Qi, Yeyun Gong, Dayiheng Liu, Nan Duan, Jiusheng Chen, Ruofei Zhang and Ming Zhou.
1. **[QDQBert](model_doc/qdqbert)** (from NVIDIA) released with the paper [Integer Quantization for Deep Learning Inference: Principles and Empirical Evaluation](https://huggingface.co/papers/2004.09602) by Hao Wu, Patrick Judd, Xiaojie Zhang, Mikhail Isaev and Paulius Micikevicius.
1. **[RAG](model_doc/rag)** (from Facebook) released with the paper [Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks](https://huggingface.co/papers/2005.11401) by Patrick Lewis, Ethan Perez, Aleksandara Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, Douwe Kiela.
1. **[REALM](model_doc/realm.html)** (from Google Research) released with the paper [REALM: Retrieval-Augmented Language Model Pre-Training](https://huggingface.co/papers/2002.08909) by Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat and Ming-Wei Chang.
1. **[Reformer](model_doc/reformer)** (from Google Research) released with the paper [Reformer: The Efficient Transformer](https://huggingface.co/papers/2001.04451) by Nikita Kitaev, Łukasz Kaiser, Anselm Levskaya.
1. **[RegNet](model_doc/regnet)** (from META Platforms) released with the paper [Designing Network Design Space](https://huggingface.co/papers/2003.13678) by Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, Piotr Dollár.
1. **[RemBERT](model_doc/rembert)** (from Google Research) released with the paper [Rethinking embedding coupling in pre-trained language models](https://huggingface.co/papers/2010.12821) by Hyung Won Chung, Thibault Févry, Henry Tsai, M. Johnson, Sebastian Ruder.
1. **[ResNet](model_doc/resnet)** (from Microsoft Research) released with the paper [Deep Residual Learning for Image Recognition](https://huggingface.co/papers/1512.03385) by Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun.
1. **[RoBERTa](model_doc/roberta)** (from Facebook), released together with the paper [RoBERTa: A Robustly Optimized BERT Pretraining Approach](https://huggingface.co/papers/1907.11692) by Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, Veselin Stoyanov.
1. **[RoFormer](model_doc/roformer)** (from ZhuiyiTechnology), released together with the paper [RoFormer: Enhanced Transformer with Rotary Position Embedding](https://huggingface.co/papers/2104.09864) by Jianlin Su and Yu Lu and Shengfeng Pan and Bo Wen and Yunfeng Liu.
1. **[SegFormer](model_doc/segformer)** (from NVIDIA) released with the paper [SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers](https://huggingface.co/papers/2105.15203) by Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, Ping Luo.
1. **[SEW](model_doc/sew)** (from ASAPP) released with the paper [Performance-Efficiency Trade-offs in Unsupervised Pre-training for Speech Recognition](https://huggingface.co/papers/2109.06870) by Felix Wu, Kwangyoun Kim, Jing Pan, Kyu Han, Kilian Q. Weinberger, Yoav Artzi.
1. **[SEW-D](model_doc/sew_d)** (from ASAPP) released with the paper [Performance-Efficiency Trade-offs in Unsupervised Pre-training for Speech Recognition](https://huggingface.co/papers/2109.06870) by Felix Wu, Kwangyoun Kim, Jing Pan, Kyu Han, Kilian Q. Weinberger, Yoav Artzi.
1. **[SpeechToTextTransformer](model_doc/speech_to_text)** (from Facebook), released together with the paper [fairseq S2T: Fast Speech-to-Text Modeling with fairseq](https://huggingface.co/papers/2010.05171) by Changhan Wang, Yun Tang, Xutai Ma, Anne Wu, Dmytro Okhonko, Juan Pino.
1. **[SpeechToTextTransformer2](model_doc/speech_to_text_2)** (from Facebook), released together with the paper [Large-Scale Self- and Semi-Supervised Learning for Speech Translation](https://huggingface.co/papers/2104.06678) by Changhan Wang, Anne Wu, Juan Pino, Alexei Baevski, Michael Auli, Alexis Conneau.
1. **[Splinter](model_doc/splinter)** (from Tel Aviv University), released together with the paper [Few-Shot Question Answering by Pretraining Span Selection](https://huggingface.co/papers/2101.00438) by Ori Ram, Yuval Kirstain, Jonathan Berant, Amir Globerson, Omer Levy.
1. **[SqueezeBERT](model_doc/squeezebert)** (from Berkeley) released with the paper [SqueezeBERT: What can computer vision teach NLP about efficient neural networks?](https://huggingface.co/papers/2006.11316) by Forrest N. Iandola, Albert E. Shaw, Ravi Krishna, and Kurt W. Keutzer.
1. **[Swin Transformer](model_doc/swin)** (from Microsoft) released with the paper [Swin Transformer: Hierarchical Vision Transformer using Shifted Windows](https://huggingface.co/papers/2103.14030) by Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, Baining Guo.
1. **[Swin Transformer V2](model_doc/swinv2)** (from Microsoft) released with the paper [Swin Transformer V2: Scaling Up Capacity and Resolution](https://huggingface.co/papers/2111.09883) by Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, Furu Wei, Baining Guo.
1. **[T5](model_doc/t5)** (from Google AI) released with the paper [Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer](https://huggingface.co/papers/1910.10683) by Colin Raffel and Noam Shazeer and Adam Roberts and Katherine Lee and Sharan Narang and Michael Matena and Yanqi Zhou and Wei Li and Peter J. Liu.
1. **[PLBart](model_doc/plbart)** (from UCLA NLP) released with the paper [Unified Pre-training for Program Understanding and Generation](https://arxiv.org/abs/2103.06333) by Wasi Uddin Ahmad, Saikat Chakraborty, Baishakhi Ray, Kai-Wei Chang.
1. **[PoolFormer](model_doc/poolformer)** (from Sea AI Labs) released with the paper [MetaFormer is Actually What You Need for Vision](https://arxiv.org/abs/2111.11418) by Yu, Weihao and Luo, Mi and Zhou, Pan and Si, Chenyang and Zhou, Yichen and Wang, Xinchao and Feng, Jiashi and Yan, Shuicheng.
1. **[ProphetNet](model_doc/prophetnet)** (from Microsoft Research) released with the paper [ProphetNet: Predicting Future N-gram for Sequence-to-Sequence Pre-training](https://arxiv.org/abs/2001.04063) by Yu Yan, Weizhen Qi, Yeyun Gong, Dayiheng Liu, Nan Duan, Jiusheng Chen, Ruofei Zhang and Ming Zhou.
1. **[QDQBert](model_doc/qdqbert)** (from NVIDIA) released with the paper [Integer Quantization for Deep Learning Inference: Principles and Empirical Evaluation](https://arxiv.org/abs/2004.09602) by Hao Wu, Patrick Judd, Xiaojie Zhang, Mikhail Isaev and Paulius Micikevicius.
1. **[RAG](model_doc/rag)** (from Facebook) released with the paper [Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks](https://arxiv.org/abs/2005.11401) by Patrick Lewis, Ethan Perez, Aleksandara Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, Douwe Kiela.
1. **[REALM](model_doc/realm.html)** (from Google Research) released with the paper [REALM: Retrieval-Augmented Language Model Pre-Training](https://arxiv.org/abs/2002.08909) by Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat and Ming-Wei Chang.
1. **[Reformer](model_doc/reformer)** (from Google Research) released with the paper [Reformer: The Efficient Transformer](https://arxiv.org/abs/2001.04451) by Nikita Kitaev, Łukasz Kaiser, Anselm Levskaya.
1. **[RegNet](model_doc/regnet)** (from META Platforms) released with the paper [Designing Network Design Space](https://arxiv.org/abs/2003.13678) by Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, Piotr Dollár.
1. **[RemBERT](model_doc/rembert)** (from Google Research) released with the paper [Rethinking embedding coupling in pre-trained language models](https://arxiv.org/abs/2010.12821) by Hyung Won Chung, Thibault Févry, Henry Tsai, M. Johnson, Sebastian Ruder.
1. **[ResNet](model_doc/resnet)** (from Microsoft Research) released with the paper [Deep Residual Learning for Image Recognition](https://arxiv.org/abs/1512.03385) by Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun.
1. **[RoBERTa](model_doc/roberta)** (from Facebook), released together with the paper [RoBERTa: A Robustly Optimized BERT Pretraining Approach](https://arxiv.org/abs/1907.11692) by Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, Veselin Stoyanov.
1. **[RoFormer](model_doc/roformer)** (from ZhuiyiTechnology), released together with the paper [RoFormer: Enhanced Transformer with Rotary Position Embedding](https://arxiv.org/abs/2104.09864) by Jianlin Su and Yu Lu and Shengfeng Pan and Bo Wen and Yunfeng Liu.
1. **[SegFormer](model_doc/segformer)** (from NVIDIA) released with the paper [SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers](https://arxiv.org/abs/2105.15203) by Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, Ping Luo.
1. **[SEW](model_doc/sew)** (from ASAPP) released with the paper [Performance-Efficiency Trade-offs in Unsupervised Pre-training for Speech Recognition](https://arxiv.org/abs/2109.06870) by Felix Wu, Kwangyoun Kim, Jing Pan, Kyu Han, Kilian Q. Weinberger, Yoav Artzi.
1. **[SEW-D](model_doc/sew_d)** (from ASAPP) released with the paper [Performance-Efficiency Trade-offs in Unsupervised Pre-training for Speech Recognition](https://arxiv.org/abs/2109.06870) by Felix Wu, Kwangyoun Kim, Jing Pan, Kyu Han, Kilian Q. Weinberger, Yoav Artzi.
1. **[SpeechToTextTransformer](model_doc/speech_to_text)** (from Facebook), released together with the paper [fairseq S2T: Fast Speech-to-Text Modeling with fairseq](https://arxiv.org/abs/2010.05171) by Changhan Wang, Yun Tang, Xutai Ma, Anne Wu, Dmytro Okhonko, Juan Pino.
1. **[SpeechToTextTransformer2](model_doc/speech_to_text_2)** (from Facebook), released together with the paper [Large-Scale Self- and Semi-Supervised Learning for Speech Translation](https://arxiv.org/abs/2104.06678) by Changhan Wang, Anne Wu, Juan Pino, Alexei Baevski, Michael Auli, Alexis Conneau.
1. **[Splinter](model_doc/splinter)** (from Tel Aviv University), released together with the paper [Few-Shot Question Answering by Pretraining Span Selection](https://arxiv.org/abs/2101.00438) by Ori Ram, Yuval Kirstain, Jonathan Berant, Amir Globerson, Omer Levy.
1. **[SqueezeBERT](model_doc/squeezebert)** (from Berkeley) released with the paper [SqueezeBERT: What can computer vision teach NLP about efficient neural networks?](https://arxiv.org/abs/2006.11316) by Forrest N. Iandola, Albert E. Shaw, Ravi Krishna, and Kurt W. Keutzer.
1. **[Swin Transformer](model_doc/swin)** (from Microsoft) released with the paper [Swin Transformer: Hierarchical Vision Transformer using Shifted Windows](https://arxiv.org/abs/2103.14030) by Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, Baining Guo.
1. **[Swin Transformer V2](model_doc/swinv2)** (from Microsoft) released with the paper [Swin Transformer V2: Scaling Up Capacity and Resolution](https://arxiv.org/abs/2111.09883) by Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, Furu Wei, Baining Guo.
1. **[T5](model_doc/t5)** (from Google AI) released with the paper [Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer](https://arxiv.org/abs/1910.10683) by Colin Raffel and Noam Shazeer and Adam Roberts and Katherine Lee and Sharan Narang and Michael Matena and Yanqi Zhou and Wei Li and Peter J. Liu.
1. **[T5v1.1](model_doc/t5v1.1)** (from Google AI) released in the repository [google-research/text-to-text-transfer-transformer](https://github.com/google-research/text-to-text-transfer-transformer/blob/main/released_checkpoints.md#t511) by Colin Raffel and Noam Shazeer and Adam Roberts and Katherine Lee and Sharan Narang and Michael Matena and Yanqi Zhou and Wei Li and Peter J. Liu.
1. **[TAPAS](model_doc/tapas)** (from Google AI) released with the paper [TAPAS: Weakly Supervised Table Parsing via Pre-training](https://huggingface.co/papers/2004.02349) by Jonathan Herzig, Paweł Krzysztof Nowak, Thomas Müller, Francesco Piccinno and Julian Martin Eisenschlos.
1. **[TAPEX](model_doc/tapex)** (from Microsoft Research) released with the paper [TAPEX: Table Pre-training via Learning a Neural SQL Executor](https://huggingface.co/papers/2107.07653) by Qian Liu, Bei Chen, Jiaqi Guo, Morteza Ziyadi, Zeqi Lin, Weizhu Chen, Jian-Guang Lou.
1. **[Trajectory Transformer](model_doc/trajectory_transformers)** (from the University of California at Berkeley) released with the paper [Offline Reinforcement Learning as One Big Sequence Modeling Problem](https://huggingface.co/papers/2106.02039) by Michael Janner, Qiyang Li, Sergey Levine
1. **[Transformer-XL](model_doc/transfo-xl)** (from Google/CMU) released with the paper [Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context](https://huggingface.co/papers/1901.02860) by Zihang Dai*, Zhilin Yang*, Yiming Yang, Jaime Carbonell, Quoc V. Le, Ruslan Salakhutdinov.
1. **[TrOCR](model_doc/trocr)** (from Microsoft), released together with the paper [TrOCR: Transformer-based Optical Character Recognition with Pre-trained Models](https://huggingface.co/papers/2109.10282) by Minghao Li, Tengchao Lv, Lei Cui, Yijuan Lu, Dinei Florencio, Cha Zhang, Zhoujun Li, Furu Wei.
1. **[UL2](model_doc/ul2)** (from Google Research) released with the paper [Unifying Language Learning Paradigms](https://huggingface.co/papers/2205.05131v1) by Yi Tay, Mostafa Dehghani, Vinh Q. Tran, Xavier Garcia, Dara Bahri, Tal Schuster, Huaixiu Steven Zheng, Neil Houlsby, Donald Metzler
1. **[TAPAS](model_doc/tapas)** (from Google AI) released with the paper [TAPAS: Weakly Supervised Table Parsing via Pre-training](https://arxiv.org/abs/2004.02349) by Jonathan Herzig, Paweł Krzysztof Nowak, Thomas Müller, Francesco Piccinno and Julian Martin Eisenschlos.
1. **[TAPEX](model_doc/tapex)** (from Microsoft Research) released with the paper [TAPEX: Table Pre-training via Learning a Neural SQL Executor](https://arxiv.org/abs/2107.07653) by Qian Liu, Bei Chen, Jiaqi Guo, Morteza Ziyadi, Zeqi Lin, Weizhu Chen, Jian-Guang Lou.
1. **[Trajectory Transformer](model_doc/trajectory_transformers)** (from the University of California at Berkeley) released with the paper [Offline Reinforcement Learning as One Big Sequence Modeling Problem](https://arxiv.org/abs/2106.02039) by Michael Janner, Qiyang Li, Sergey Levine
1. **[Transformer-XL](model_doc/transfo-xl)** (from Google/CMU) released with the paper [Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context](https://arxiv.org/abs/1901.02860) by Zihang Dai*, Zhilin Yang*, Yiming Yang, Jaime Carbonell, Quoc V. Le, Ruslan Salakhutdinov.
1. **[TrOCR](model_doc/trocr)** (from Microsoft), released together with the paper [TrOCR: Transformer-based Optical Character Recognition with Pre-trained Models](https://arxiv.org/abs/2109.10282) by Minghao Li, Tengchao Lv, Lei Cui, Yijuan Lu, Dinei Florencio, Cha Zhang, Zhoujun Li, Furu Wei.
1. **[UL2](model_doc/ul2)** (from Google Research) released with the paper [Unifying Language Learning Paradigms](https://arxiv.org/abs/2205.05131v1) by Yi Tay, Mostafa Dehghani, Vinh Q. Tran, Xavier Garcia, Dara Bahri, Tal Schuster, Huaixiu Steven Zheng, Neil Houlsby, Donald Metzler
1. **[UMT5](model_doc/umt5)** (from Google Research) released with the paper [UniMax: Fairer and More Effective Language Sampling for Large-Scale Multilingual Pretraining](https://openreview.net/forum?id=kXwdL1cWOAi) by Hyung Won Chung, Xavier Garcia, Adam Roberts, Yi Tay, Orhan Firat, Sharan Narang, Noah Constant.
1. **[UniSpeech](model_doc/unispeech)** (from Microsoft Research) released with the paper [UniSpeech: Unified Speech Representation Learning with Labeled and Unlabeled Data](https://huggingface.co/papers/2101.07597) by Chengyi Wang, Yu Wu, Yao Qian, Kenichi Kumatani, Shujie Liu, Furu Wei, Michael Zeng, Xuedong Huang.
1. **[UniSpeechSat](model_doc/unispeech-sat)** (from Microsoft Research) released with the paper [UNISPEECH-SAT: UNIVERSAL SPEECH REPRESENTATION LEARNING WITH SPEAKER AWARE PRE-TRAINING](https://huggingface.co/papers/2110.05752) by Sanyuan Chen, Yu Wu, Chengyi Wang, Zhengyang Chen, Zhuo Chen, Shujie Liu, Jian Wu, Yao Qian, Furu Wei, Jinyu Li, Xiangzhan Yu.
1. **[VAN](model_doc/van)** (from Tsinghua University and Nankai University) released with the paper [Visual Attention Network](https://huggingface.co/papers/2202.09741) by Meng-Hao Guo, Cheng-Ze Lu, Zheng-Ning Liu, Ming-Ming Cheng, Shi-Min Hu.
1. **[VideoMAE](model_doc/videomae)** (from Multimedia Computing Group, Nanjing University) released with the paper [VideoMAE: Masked Autoencoders are Data-Efficient Learners for Self-Supervised Video Pre-Training](https://huggingface.co/papers/2203.12602) by Zhan Tong, Yibing Song, Jue Wang, Limin Wang.
1. **[ViLT](model_doc/vilt)** (from NAVER AI Lab/Kakao Enterprise/Kakao Brain) released with the paper [ViLT: Vision-and-Language Transformer Without Convolution or Region Supervision](https://huggingface.co/papers/2102.03334) by Wonjae Kim, Bokyung Son, Ildoo Kim.
1. **[Vision Transformer (ViT)](model_doc/vit)** (from Google AI) released with the paper [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://huggingface.co/papers/2010.11929) by Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby.
1. **[VisualBERT](model_doc/visual_bert)** (from UCLA NLP) released with the paper [VisualBERT: A Simple and Performant Baseline for Vision and Language](https://huggingface.co/papers/1908.03557) by Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, Kai-Wei Chang.
1. **[ViTMAE](model_doc/vit_mae)** (from Meta AI) released with the paper [Masked Autoencoders Are Scalable Vision Learners](https://huggingface.co/papers/2111.06377) by Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, Ross Girshick.
1. **[Wav2Vec2](model_doc/wav2vec2)** (from Facebook AI) released with the paper [wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations](https://huggingface.co/papers/2006.11477) by Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, Michael Auli.
1. **[Wav2Vec2-Conformer](model_doc/wav2vec2-conformer)** (from Facebook AI) released with the paper [FAIRSEQ S2T: Fast Speech-to-Text Modeling with FAIRSEQ](https://huggingface.co/papers/2010.05171) by Changhan Wang, Yun Tang, Xutai Ma, Anne Wu, Sravya Popuri, Dmytro Okhonko, Juan Pino.
1. **[Wav2Vec2Phoneme](model_doc/wav2vec2_phoneme)** (from Facebook AI) released with the paper [Simple and Effective Zero-shot Cross-lingual Phoneme Recognition](https://huggingface.co/papers/2109.11680) by Qiantong Xu, Alexei Baevski, Michael Auli.
1. **[WavLM](model_doc/wavlm)** (from Microsoft Research) released with the paper [WavLM: Large-Scale Self-Supervised Pre-Training for Full Stack Speech Processing](https://huggingface.co/papers/2110.13900) by Sanyuan Chen, Chengyi Wang, Zhengyang Chen, Yu Wu, Shujie Liu, Zhuo Chen, Jinyu Li, Naoyuki Kanda, Takuya Yoshioka, Xiong Xiao, Jian Wu, Long Zhou, Shuo Ren, Yanmin Qian, Yao Qian, Jian Wu, Michael Zeng, Furu Wei.
1. **[XGLM](model_doc/xglm)** (From Facebook AI) released with the paper [Few-shot Learning with Multilingual Language Models](https://huggingface.co/papers/2112.10668) by Xi Victoria Lin, Todor Mihaylov, Mikel Artetxe, Tianlu Wang, Shuohui Chen, Daniel Simig, Myle Ott, Naman Goyal, Shruti Bhosale, Jingfei Du, Ramakanth Pasunuru, Sam Shleifer, Punit Singh Koura, Vishrav Chaudhary, Brian O'Horo, Jeff Wang, Luke Zettlemoyer, Zornitsa Kozareva, Mona Diab, Veselin Stoyanov, Xian Li.
1. **[XLM](model_doc/xlm)** (from Facebook) released together with the paper [Cross-lingual Language Model Pretraining](https://huggingface.co/papers/1901.07291) by Guillaume Lample and Alexis Conneau.
1. **[XLM-ProphetNet](model_doc/xlm-prophetnet)** (from Microsoft Research) released with the paper [ProphetNet: Predicting Future N-gram for Sequence-to-Sequence Pre-training](https://huggingface.co/papers/2001.04063) by Yu Yan, Weizhen Qi, Yeyun Gong, Dayiheng Liu, Nan Duan, Jiusheng Chen, Ruofei Zhang and Ming Zhou.
1. **[XLM-RoBERTa](model_doc/xlm-roberta)** (from Facebook AI), released together with the paper [Unsupervised Cross-lingual Representation Learning at Scale](https://huggingface.co/papers/1911.02116) by Alexis Conneau*, Kartikay Khandelwal*, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer and Veselin Stoyanov.
1. **[XLM-RoBERTa-XL](model_doc/xlm-roberta-xl)** (from Facebook AI), released together with the paper [Larger-Scale Transformers for Multilingual Masked Language Modeling](https://huggingface.co/papers/2105.00572) by Naman Goyal, Jingfei Du, Myle Ott, Giri Anantharaman, Alexis Conneau.
1. **[XLM-V](model_doc/xlm-v)** (from Meta AI) released with the paper [XLM-V: Overcoming the Vocabulary Bottleneck in Multilingual Masked Language Models](https://huggingface.co/papers/2301.10472) by Davis Liang, Hila Gonen, Yuning Mao, Rui Hou, Naman Goyal, Marjan Ghazvininejad, Luke Zettlemoyer, Madian Khabsa.
1. **[XLNet](model_doc/xlnet)** (from Google/CMU) released with the paper [XLNet: Generalized Autoregressive Pretraining for Language Understanding](https://huggingface.co/papers/1906.08237) by Zhilin Yang*, Zihang Dai*, Yiming Yang, Jaime Carbonell, Ruslan Salakhutdinov, Quoc V. Le.
1. **[XLS-R](model_doc/xls_r)** (from Facebook AI) released with the paper [XLS-R: Self-supervised Cross-lingual Speech Representation Learning at Scale](https://huggingface.co/papers/2111.09296) by Arun Babu, Changhan Wang, Andros Tjandra, Kushal Lakhotia, Qiantong Xu, Naman Goyal, Kritika Singh, Patrick von Platen, Yatharth Saraf, Juan Pino, Alexei Baevski, Alexis Conneau, Michael Auli.
1. **[XLSR-Wav2Vec2](model_doc/xlsr_wav2vec2)** (from Facebook AI) released with the paper [Unsupervised Cross-Lingual Representation Learning For Speech Recognition](https://huggingface.co/papers/2006.13979) by Alexis Conneau, Alexei Baevski, Ronan Collobert, Abdelrahman Mohamed, Michael Auli.
1. **[YOLOS](model_doc/yolos)** (from Huazhong University of Science & Technology) released with the paper [You Only Look at One Sequence: Rethinking Transformer in Vision through Object Detection](https://huggingface.co/papers/2106.00666) by Yuxin Fang, Bencheng Liao, Xinggang Wang, Jiemin Fang, Jiyang Qi, Rui Wu, Jianwei Niu, Wenyu Liu.
1. **[YOSO](model_doc/yoso)** (from the University of Wisconsin - Madison) released with the paper [You Only Sample (Almost) Once: Linear Cost Self-Attention Via Bernoulli Sampling](https://huggingface.co/papers/2111.09714) by Zhanpeng Zeng, Yunyang Xiong, Sathya N. Ravi, Shailesh Acharya, Glenn Fung, Vikas Singh.
1. **[UniSpeech](model_doc/unispeech)** (from Microsoft Research) released with the paper [UniSpeech: Unified Speech Representation Learning with Labeled and Unlabeled Data](https://arxiv.org/abs/2101.07597) by Chengyi Wang, Yu Wu, Yao Qian, Kenichi Kumatani, Shujie Liu, Furu Wei, Michael Zeng, Xuedong Huang.
1. **[UniSpeechSat](model_doc/unispeech-sat)** (from Microsoft Research) released with the paper [UNISPEECH-SAT: UNIVERSAL SPEECH REPRESENTATION LEARNING WITH SPEAKER AWARE PRE-TRAINING](https://arxiv.org/abs/2110.05752) by Sanyuan Chen, Yu Wu, Chengyi Wang, Zhengyang Chen, Zhuo Chen, Shujie Liu, Jian Wu, Yao Qian, Furu Wei, Jinyu Li, Xiangzhan Yu.
1. **[VAN](model_doc/van)** (from Tsinghua University and Nankai University) released with the paper [Visual Attention Network](https://arxiv.org/abs/2202.09741) by Meng-Hao Guo, Cheng-Ze Lu, Zheng-Ning Liu, Ming-Ming Cheng, Shi-Min Hu.
1. **[VideoMAE](model_doc/videomae)** (from Multimedia Computing Group, Nanjing University) released with the paper [VideoMAE: Masked Autoencoders are Data-Efficient Learners for Self-Supervised Video Pre-Training](https://arxiv.org/abs/2203.12602) by Zhan Tong, Yibing Song, Jue Wang, Limin Wang.
1. **[ViLT](model_doc/vilt)** (from NAVER AI Lab/Kakao Enterprise/Kakao Brain) released with the paper [ViLT: Vision-and-Language Transformer Without Convolution or Region Supervision](https://arxiv.org/abs/2102.03334) by Wonjae Kim, Bokyung Son, Ildoo Kim.
1. **[Vision Transformer (ViT)](model_doc/vit)** (from Google AI) released with the paper [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929) by Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby.
1. **[VisualBERT](model_doc/visual_bert)** (from UCLA NLP) released with the paper [VisualBERT: A Simple and Performant Baseline for Vision and Language](https://arxiv.org/pdf/1908.03557) by Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, Kai-Wei Chang.
1. **[ViTMAE](model_doc/vit_mae)** (from Meta AI) released with the paper [Masked Autoencoders Are Scalable Vision Learners](https://arxiv.org/abs/2111.06377) by Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, Ross Girshick.
1. **[Wav2Vec2](model_doc/wav2vec2)** (from Facebook AI) released with the paper [wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations](https://arxiv.org/abs/2006.11477) by Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, Michael Auli.
1. **[Wav2Vec2-Conformer](model_doc/wav2vec2-conformer)** (from Facebook AI) released with the paper [FAIRSEQ S2T: Fast Speech-to-Text Modeling with FAIRSEQ](https://arxiv.org/abs/2010.05171) by Changhan Wang, Yun Tang, Xutai Ma, Anne Wu, Sravya Popuri, Dmytro Okhonko, Juan Pino.
1. **[Wav2Vec2Phoneme](model_doc/wav2vec2_phoneme)** (from Facebook AI) released with the paper [Simple and Effective Zero-shot Cross-lingual Phoneme Recognition](https://arxiv.org/abs/2109.11680) by Qiantong Xu, Alexei Baevski, Michael Auli.
1. **[WavLM](model_doc/wavlm)** (from Microsoft Research) released with the paper [WavLM: Large-Scale Self-Supervised Pre-Training for Full Stack Speech Processing](https://arxiv.org/abs/2110.13900) by Sanyuan Chen, Chengyi Wang, Zhengyang Chen, Yu Wu, Shujie Liu, Zhuo Chen, Jinyu Li, Naoyuki Kanda, Takuya Yoshioka, Xiong Xiao, Jian Wu, Long Zhou, Shuo Ren, Yanmin Qian, Yao Qian, Jian Wu, Michael Zeng, Furu Wei.
1. **[XGLM](model_doc/xglm)** (From Facebook AI) released with the paper [Few-shot Learning with Multilingual Language Models](https://arxiv.org/abs/2112.10668) by Xi Victoria Lin, Todor Mihaylov, Mikel Artetxe, Tianlu Wang, Shuohui Chen, Daniel Simig, Myle Ott, Naman Goyal, Shruti Bhosale, Jingfei Du, Ramakanth Pasunuru, Sam Shleifer, Punit Singh Koura, Vishrav Chaudhary, Brian O'Horo, Jeff Wang, Luke Zettlemoyer, Zornitsa Kozareva, Mona Diab, Veselin Stoyanov, Xian Li.
1. **[XLM](model_doc/xlm)** (from Facebook) released together with the paper [Cross-lingual Language Model Pretraining](https://arxiv.org/abs/1901.07291) by Guillaume Lample and Alexis Conneau.
1. **[XLM-ProphetNet](model_doc/xlm-prophetnet)** (from Microsoft Research) released with the paper [ProphetNet: Predicting Future N-gram for Sequence-to-Sequence Pre-training](https://arxiv.org/abs/2001.04063) by Yu Yan, Weizhen Qi, Yeyun Gong, Dayiheng Liu, Nan Duan, Jiusheng Chen, Ruofei Zhang and Ming Zhou.
1. **[XLM-RoBERTa](model_doc/xlm-roberta)** (from Facebook AI), released together with the paper [Unsupervised Cross-lingual Representation Learning at Scale](https://arxiv.org/abs/1911.02116) by Alexis Conneau*, Kartikay Khandelwal*, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer and Veselin Stoyanov.
1. **[XLM-RoBERTa-XL](model_doc/xlm-roberta-xl)** (from Facebook AI), released together with the paper [Larger-Scale Transformers for Multilingual Masked Language Modeling](https://arxiv.org/abs/2105.00572) by Naman Goyal, Jingfei Du, Myle Ott, Giri Anantharaman, Alexis Conneau.
1. **[XLM-V](model_doc/xlm-v)** (from Meta AI) released with the paper [XLM-V: Overcoming the Vocabulary Bottleneck in Multilingual Masked Language Models](https://arxiv.org/abs/2301.10472) by Davis Liang, Hila Gonen, Yuning Mao, Rui Hou, Naman Goyal, Marjan Ghazvininejad, Luke Zettlemoyer, Madian Khabsa.
1. **[XLNet](model_doc/xlnet)** (from Google/CMU) released with the paper [XLNet: Generalized Autoregressive Pretraining for Language Understanding](https://arxiv.org/abs/1906.08237) by Zhilin Yang*, Zihang Dai*, Yiming Yang, Jaime Carbonell, Ruslan Salakhutdinov, Quoc V. Le.
1. **[XLS-R](model_doc/xls_r)** (from Facebook AI) released with the paper [XLS-R: Self-supervised Cross-lingual Speech Representation Learning at Scale](https://arxiv.org/abs/2111.09296) by Arun Babu, Changhan Wang, Andros Tjandra, Kushal Lakhotia, Qiantong Xu, Naman Goyal, Kritika Singh, Patrick von Platen, Yatharth Saraf, Juan Pino, Alexei Baevski, Alexis Conneau, Michael Auli.
1. **[XLSR-Wav2Vec2](model_doc/xlsr_wav2vec2)** (from Facebook AI) released with the paper [Unsupervised Cross-Lingual Representation Learning For Speech Recognition](https://arxiv.org/abs/2006.13979) by Alexis Conneau, Alexei Baevski, Ronan Collobert, Abdelrahman Mohamed, Michael Auli.
1. **[YOLOS](model_doc/yolos)** (from Huazhong University of Science & Technology) released with the paper [You Only Look at One Sequence: Rethinking Transformer in Vision through Object Detection](https://arxiv.org/abs/2106.00666) by Yuxin Fang, Bencheng Liao, Xinggang Wang, Jiemin Fang, Jiyang Qi, Rui Wu, Jianwei Niu, Wenyu Liu.
1. **[YOSO](model_doc/yoso)** (from the University of Wisconsin - Madison) released with the paper [You Only Sample (Almost) Once: Linear Cost Self-Attention Via Bernoulli Sampling](https://arxiv.org/abs/2111.09714) by Zhanpeng Zeng, Yunyang Xiong, Sathya N. Ravi, Shailesh Acharya, Glenn Fung, Vikas Singh.
### Unterstützte Frameworks

View File

@ -56,7 +56,7 @@ Dateien lassen sich auch in einem Repository leicht bearbeiten, und Sie können
Bevor Sie ein Modell für den Hub freigeben, benötigen Sie Ihre Hugging Face-Anmeldedaten. Wenn Sie Zugang zu einem Terminal haben, führen Sie den folgenden Befehl in der virtuellen Umgebung aus, in der 🤗 Transformers installiert ist. Dadurch werden Ihre Zugangsdaten in Ihrem Hugging Face-Cache-Ordner (standardmäßig `~/.cache/`) gespeichert:
```bash
hf auth login
huggingface-cli login
```
Wenn Sie ein Notebook wie Jupyter oder Colaboratory verwenden, stellen Sie sicher, dass Sie die [`huggingface_hub`](https://huggingface.co/docs/hub/adding-a-library) Bibliothek installiert haben. Diese Bibliothek ermöglicht Ihnen die programmatische Interaktion mit dem Hub.

View File

@ -44,7 +44,7 @@ Transformers unterstützt nativ einige PEFT-Methoden, d.h. Sie können lokal ode
- [Low Rank Adapters](https://huggingface.co/docs/peft/conceptual_guides/lora)
- [IA3](https://huggingface.co/docs/peft/conceptual_guides/ia3)
- [AdaLoRA](https://huggingface.co/papers/2303.10512)
- [AdaLoRA](https://arxiv.org/abs/2303.10512)
Wenn Sie andere PEFT-Methoden, wie z.B. Prompt Learning oder Prompt Tuning, verwenden möchten, oder über die 🤗 PEFT-Bibliothek im Allgemeinen, lesen Sie bitte die [Dokumentation](https://huggingface.co/docs/peft/index).

View File

@ -324,7 +324,7 @@ python examples/pytorch/summarization/run_summarization.py
Alle Skripte können Ihr endgültiges Modell in den [Model Hub](https://huggingface.co/models) hochladen. Stellen Sie sicher, dass Sie bei Hugging Face angemeldet sind, bevor Sie beginnen:
```bash
hf auth login
huggingface-cli login
```
Dann fügen Sie dem Skript das Argument `push_to_hub` hinzu. Mit diesem Argument wird ein Repository mit Ihrem Hugging Face-Benutzernamen und dem in `output_dir` angegebenen Ordnernamen erstellt.

View File

@ -473,6 +473,13 @@ Hier ist zum Beispiel ein Test, der nur ausgeführt werden muss, wenn 2 oder meh
def test_example_with_multi_gpu():
```
Wenn ein Test `tensorflow` benötigt, verwenden Sie den Dekorator `require_tf`. Zum Beispiel:
```python no-style
@require_tf
def test_tf_thing_with_tensorflow():
```
Diese Dekors können gestapelt werden. Wenn zum Beispiel ein Test langsam ist und mindestens eine GPU unter pytorch benötigt, können Sie
wie Sie ihn einrichten können:
@ -1197,6 +1204,9 @@ if torch.cuda.is_available():
import numpy as np
np.random.seed(seed)
# tf RNG
tf.random.set_seed(seed)
```
### Tests debuggen

View File

@ -17,12 +17,20 @@
title: Customizing model components
- local: model_sharing
title: Sharing
- local: modular_transformers
title: Contributing a new model to Transformers
- local: add_new_model
title: Legacy model contribution
title: Adding a new model to Transformers
- local: modular_transformers
title: Modular Transformers
- local: auto_docstring
title: Documenting a model
title: Document your models
- local: task_summary
title: What 🤗 Transformers can do
- local: tasks_explained
title: How 🤗 Transformers solve tasks
- local: model_summary
title: The Transformer model family
- local: attention
title: Attention mechanisms
- local: attention_interface
title: Customizing attention function
title: Models
@ -72,6 +80,8 @@
title: Caching
- local: kv_cache
title: KV cache strategies
- local: serving
title: Serving
- local: llm_tutorial_optimization
title: Getting the most out of LLMs
- local: perplexity
@ -89,34 +99,22 @@
- local: chat_extras
title: Tools and RAG
title: Chat with models
- sections:
- local: serving
title: Serving LLMs, VLMs, and other chat-based models
- local: jan
title: Jan
- local: cursor
title: Cursor
- local: tiny_agents
title: Tiny-Agents CLI and MCP tools
- local: open_webui
title: Open WebUI
title: Serving
- sections:
- local: perf_torch_compile
title: torch.compile
- local: perf_infer_gpu_one
title: GPU
- local: perf_infer_gpu_multi
title: Distributed inference
title: Distributed GPU inference
- local: perf_infer_cpu
title: CPU
- local: tf_xla
title: XLA
title: Optimization
- local: agents
title: Agents
- local: tools
title: Tools
- local: transformers_as_backend
title: Inference server backends
title: Inference
- isExpanded: false
sections:
@ -151,6 +149,8 @@
title: GPU
- local: perf_train_cpu
title: CPU
- local: perf_train_tpu_tf
title: TPU
- local: perf_train_special
title: Apple Silicon
- local: perf_train_gaudi
@ -189,8 +189,6 @@
title: FBGEMM
- local: quantization/finegrained_fp8
title: Fine-grained FP8
- local: quantization/fp_quant
title: FP-Quant
- local: gguf
title: GGUF
- local: quantization/gptq
@ -373,8 +371,6 @@
- sections:
- local: model_doc/albert
title: ALBERT
- local: model_doc/arcee
title: Arcee
- local: model_doc/bamba
title: Bamba
- local: model_doc/bart
@ -443,10 +439,6 @@
title: DiffLlama
- local: model_doc/distilbert
title: DistilBERT
- local: model_doc/doge
title: Doge
- local: model_doc/dots1
title: dots1
- local: model_doc/dpr
title: DPR
- local: model_doc/electra
@ -455,16 +447,10 @@
title: Encoder Decoder Models
- local: model_doc/ernie
title: ERNIE
- local: model_doc/ernie4_5
title: Ernie4_5
- local: model_doc/ernie4_5_moe
title: Ernie4_5_MoE
- local: model_doc/ernie_m
title: ErnieM
- local: model_doc/esm
title: ESM
- local: model_doc/exaone4
title: EXAONE-4.0
- local: model_doc/falcon
title: Falcon
- local: model_doc/falcon3
@ -495,8 +481,6 @@
title: GLM
- local: model_doc/glm4
title: glm4
- local: model_doc/glm4_moe
title: glm4_moe
- local: model_doc/openai-gpt
title: GPT
- local: model_doc/gpt_neo
@ -511,8 +495,6 @@
title: GPT2
- local: model_doc/gpt_bigcode
title: GPTBigCode
- local: model_doc/gpt_oss
title: GptOss
- local: model_doc/gptsan-japanese
title: GPTSAN Japanese
- local: model_doc/gpt-sw3
@ -529,6 +511,8 @@
title: Helium
- local: model_doc/herbert
title: HerBERT
- local: model_doc/hgnet_v2
title: HGNet-V2
- local: model_doc/ibert
title: I-BERT
- local: model_doc/jamba
@ -539,8 +523,6 @@
title: Jukebox
- local: model_doc/led
title: LED
- local: model_doc/lfm2
title: LFM2
- local: model_doc/llama
title: LLaMA
- local: model_doc/llama2
@ -585,8 +567,6 @@
title: MobileBERT
- local: model_doc/modernbert
title: ModernBert
- local: model_doc/modernbert-decoder
title: ModernBERTDecoder
- local: model_doc/mpnet
title: MPNet
- local: model_doc/mpt
@ -681,8 +661,6 @@
title: SwitchTransformers
- local: model_doc/t5
title: T5
- local: model_doc/t5gemma
title: T5Gemma
- local: model_doc/t5v1.1
title: T5v1.1
- local: model_doc/tapex
@ -693,8 +671,6 @@
title: UL2
- local: model_doc/umt5
title: UMT5
- local: model_doc/xcodec
title: X-CODEC
- local: model_doc/xmod
title: X-MOD
- local: model_doc/xglm
@ -711,8 +687,6 @@
title: XLM-V
- local: model_doc/xlnet
title: XLNet
- local: model_doc/xlstm
title: xLSTM
- local: model_doc/yoso
title: YOSO
- local: model_doc/zamba
@ -721,8 +695,6 @@
title: Zamba2
title: Text models
- sections:
- local: model_doc/aimv2
title: Aimv2
- local: model_doc/beit
title: BEiT
- local: model_doc/bit
@ -739,12 +711,6 @@
title: D-FINE
- local: model_doc/dab-detr
title: DAB-DETR
- local: model_doc/deepseek_v2
title: DeepSeek-V2
- local: model_doc/deepseek_vl
title: DeepseekVL
- local: model_doc/deepseek_vl_hybrid
title: DeepseekVLHybrid
- local: model_doc/deformable_detr
title: Deformable DETR
- local: model_doc/deit
@ -765,26 +731,18 @@
title: DINOV2
- local: model_doc/dinov2_with_registers
title: DINOv2 with Registers
- local: model_doc/dinov3
title: DINOv3
- local: model_doc/dit
title: DiT
- local: model_doc/dpt
title: DPT
- local: model_doc/efficientformer
title: EfficientFormer
- local: model_doc/efficientloftr
title: EfficientLoFTR
- local: model_doc/efficientnet
title: EfficientNet
- local: model_doc/eomt
title: EoMT
- local: model_doc/focalnet
title: FocalNet
- local: model_doc/glpn
title: GLPN
- local: model_doc/hgnet_v2
title: HGNet-V2
- local: model_doc/hiera
title: Hiera
- local: model_doc/ijepa
@ -793,8 +751,6 @@
title: ImageGPT
- local: model_doc/levit
title: LeViT
- local: model_doc/lightglue
title: LightGlue
- local: model_doc/mask2former
title: Mask2Former
- local: model_doc/maskformer
@ -883,8 +839,6 @@
title: CSM
- local: model_doc/dac
title: dac
- local: model_doc/dia
title: Dia
- local: model_doc/encodec
title: EnCodec
- local: model_doc/fastspeech2_conformer
@ -893,8 +847,6 @@
title: GraniteSpeech
- local: model_doc/hubert
title: Hubert
- local: model_doc/kyutai_speech_to_text
title: Kyutai Speech-To-Text
- local: model_doc/mctct
title: MCTCT
- local: model_doc/mimi
@ -953,8 +905,6 @@
- sections:
- local: model_doc/timesformer
title: TimeSformer
- local: model_doc/vjepa2
title: V-JEPA 2
- local: model_doc/videomae
title: VideoMAE
- local: model_doc/vivit
@ -987,8 +937,6 @@
title: CLIPSeg
- local: model_doc/clvp
title: CLVP
- local: model_doc/cohere2_vision
title: Cohere2Vision
- local: model_doc/colpali
title: ColPali
- local: model_doc/colqwen2
@ -1001,20 +949,12 @@
title: Donut
- local: model_doc/emu3
title: Emu3
- local: model_doc/evolla
title: Evolla
- local: model_doc/flava
title: FLAVA
- local: model_doc/gemma3
title: Gemma3
- local: model_doc/gemma3n
title: Gemma3n
- local: model_doc/git
title: GIT
- local: model_doc/glm4v
title: glm4v
- local: model_doc/glm4v_moe
title: glm4v_moe
- local: model_doc/got_ocr2
title: GOT-OCR2
- local: model_doc/granitevision
@ -1039,8 +979,6 @@
title: Janus
- local: model_doc/kosmos-2
title: KOSMOS-2
- local: model_doc/kosmos2_5
title: KOSMOS-2.5
- local: model_doc/layoutlm
title: LayoutLM
- local: model_doc/layoutlmv2
@ -1054,7 +992,7 @@
- local: model_doc/llama4
title: Llama4
- local: model_doc/llava
title: LLaVA
title: Llava
- local: model_doc/llava_next
title: LLaVA-NeXT
- local: model_doc/llava_next_video
@ -1065,24 +1003,18 @@
title: LXMERT
- local: model_doc/matcha
title: MatCha
- local: model_doc/metaclip_2
title: MetaCLIP 2
- local: model_doc/mgp-str
title: MGP-STR
- local: model_doc/mistral3
title: Mistral3
- local: model_doc/mllama
title: mllama
- local: model_doc/mm-grounding-dino
title: MM Grounding DINO
- local: model_doc/nougat
title: Nougat
- local: model_doc/omdet-turbo
title: OmDet-Turbo
- local: model_doc/oneformer
title: OneFormer
- local: model_doc/ovis2
title: Ovis2
- local: model_doc/owlvit
title: OWL-ViT
- local: model_doc/owlv2
@ -1091,8 +1023,6 @@
title: PaliGemma
- local: model_doc/perceiver
title: Perceiver
- local: model_doc/perception_lm
title: PerceptionLM
- local: model_doc/phi4_multimodal
title: Phi4 Multimodal
- local: model_doc/pix2struct
@ -1107,10 +1037,6 @@
title: Qwen2Audio
- local: model_doc/qwen2_vl
title: Qwen2VL
- local: model_doc/sam2
title: SAM2
- local: model_doc/sam2_video
title: SAM2 Video
- local: model_doc/sam
title: Segment Anything
- local: model_doc/sam_hq
@ -1121,8 +1047,6 @@
title: SigLIP
- local: model_doc/siglip2
title: SigLIP2
- local: model_doc/smollm3
title: SmolLM3
- local: model_doc/smolvlm
title: SmolVLM
- local: model_doc/speech-encoder-decoder
@ -1149,8 +1073,6 @@
title: Vision Text Dual Encoder
- local: model_doc/visual_bert
title: VisualBERT
- local: model_doc/voxtral
title: Voxtral
- local: model_doc/xclip
title: X-CLIP
title: Multimodal models
@ -1208,3 +1130,4 @@
title: Environment Variables
title: Reference
title: API

View File

@ -75,7 +75,7 @@ training_args = TrainingArguments(
per_device_eval_batch_size=16,
num_train_epochs=2,
fsdp_config="path/to/fsdp_config",
fsdp="full_shard",
fsdp_strategy="full_shard",
weight_decay=0.01,
eval_strategy="epoch",
save_strategy="epoch",

View File

@ -13,7 +13,7 @@ rendered properly in your Markdown viewer.
-->
# Legacy model contribution
# Adding a new model to Transformers
> [!TIP]
> Try adding new models with a more [modular](./modular_transformers) approach first. This makes it significantly easier to contribute a model to Transformers!
@ -571,7 +571,7 @@ The processor should call the appropriate modality-specific processors within it
def __call__(
self,
images: ImageInput = None,
text: Union[TextInput, PreTokenizedInput, list[TextInput], list[PreTokenizedInput]] = None,
text: Union[TextInput, PreTokenizedInput, List[TextInput], List[PreTokenizedInput]] = None,
audio=None,
videos=None,
**kwargs: Unpack[YourModelProcessorKwargs],

View File

@ -14,9 +14,5 @@ rendered properly in your Markdown viewer.
-->
# Agents
(deprecated)
> [!WARNING]
> Agents and tools were spun out into the standalone [smolagents](https://huggingface.co/docs/smolagents/index) library. They were removed from `transformers` in v4.52.

View File

@ -0,0 +1,61 @@
<!--Copyright 2023 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.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# Attention mechanisms
Most transformer models use full attention in the sense that the attention matrix is square. It can be a big
computational bottleneck when you have long texts. Longformer and reformer are models that try to be more efficient and
use a sparse version of the attention matrix to speed up training.
## LSH attention
[Reformer](model_doc/reformer) uses LSH attention. In the softmax(QK^t), only the biggest elements (in the softmax
dimension) of the matrix QK^t are going to give useful contributions. So for each query q in Q, we can consider only
the keys k in K that are close to q. A hash function is used to determine if q and k are close. The attention mask is
modified to mask the current token (except at the first position), because it will give a query and a key equal (so
very similar to each other). Since the hash can be a bit random, several hash functions are used in practice
(determined by a n_rounds parameter) and then are averaged together.
## Local attention
[Longformer](model_doc/longformer) uses local attention: often, the local context (e.g., what are the two tokens to the
left and right?) is enough to take action for a given token. Also, by stacking attention layers that have a small
window, the last layer will have a receptive field of more than just the tokens in the window, allowing them to build a
representation of the whole sentence.
Some preselected input tokens are also given global attention: for those few tokens, the attention matrix can access
all tokens and this process is symmetric: all other tokens have access to those specific tokens (on top of the ones in
their local window). This is shown in Figure 2d of the paper, see below for a sample attention mask:
<div class="flex justify-center">
<img scale="50 %" align="center" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/local_attention_mask.png"/>
</div>
Using those attention matrices with less parameters then allows the model to have inputs having a bigger sequence
length.
## Other tricks
### Axial positional encodings
[Reformer](model_doc/reformer) uses axial positional encodings: in traditional transformer models, the positional encoding
E is a matrix of size \\(l\\) by \\(d\\), \\(l\\) being the sequence length and \\(d\\) the dimension of the
hidden state. If you have very long texts, this matrix can be huge and take way too much space on the GPU. To alleviate
that, axial positional encodings consist of factorizing that big matrix E in two smaller matrices E1 and E2, with
dimensions \\(l_{1} \times d_{1}\\) and \\(l_{2} \times d_{2}\\), such that \\(l_{1} \times l_{2} = l\\) and
\\(d_{1} + d_{2} = d\\) (with the product for the lengths, this ends up being way smaller). The embedding for time
step \\(j\\) in E is obtained by concatenating the embeddings for timestep \\(j \% l1\\) in E1 and \\(j // l1\\)
in E2.

View File

@ -60,11 +60,11 @@ You will see it prints "I just entered the attention computation" as many times
## Dynamically switching attention function
You could dynamically change the model's attention function as well:
You could dynamically change the model's attention function as well, by overriding the `config._attn_implementation` field:
```python
# Back to use original sdpa implementation
model.set_attn_implementation("sdpa")
model.config._attn_implementation = "sdpa"
model(torch.ones(1, 5, dtype=int))
```
@ -72,34 +72,6 @@ model(torch.ones(1, 5, dtype=int))
and it will stop printing the statements, as it now uses the `sdpa` attention.
This allows to quickly change an attention function, without needing to reload the model!
## Different attention per backbone in multimodal models
For multimodal models different attention functions may work better for each backbone module. For example, some vision backbones perform better in fp32, but are incompatible with FlashAttention. To continue using FlashAttention while keeping the vision encoder in fp32, create a dict and map each config to an attention implementation as shown below.
```python
from transformers import AutoModelForImageTextToText
model_id = "facebook/chameleon-7b"
attention_implementation_per_backbone = {"vision_config": "sdpa", "text_config": "flash_attention_2"}
model = AutoModelForImageTextToText.from_pretrained(model_id, attn_implementation=attention_implementation_per_backbone)
# NOTE: keys in the attention implementation have to be the same as the sub-config names
for key in attention_implementation_per_backbone:
assert key in model.config.sub_configs, f"Invalid key in `attention_implementation`"
# You can omit certain backbones - the default attention function (SDPA) will be used
# This is equivalent to the previous example
model = AutoModelForImageTextToText.from_pretrained(model_id, attn_implementation={"text_config": "flash_attention_2"})
# Set the same attention implementation for all backbones with single string, same as in non-multimodal models
model = AutoModelForImageTextToText.from_pretrained(model_id, attn_implementation="eager")
# Alternatively use a dict with an empty key for global configuration
model = AutoModelForImageTextToText.from_pretrained(model_id, attn_implementation={"": "eager"})
```
## What about new args needed in my custom attention function?
But indeed, what if the new function requires a new arg to be properly used? It's no issue! Models supporting the
@ -120,7 +92,7 @@ def custom_attention(
a_new_kwargs = None, # You can now add as many kwargs as you need
another_new_kwargs = None, # You can now add as many kwargs as you need
**kwargs, # You need to accept **kwargs as models will pass other args
) -> tuple[torch.Tensor, Optional[torch.Tensor]]
) -> Tuple[torch.Tensor, Optional[torch.Tensor]]
... # do your magic!
return attn_output, attn_weights # attn_weights are optional here

View File

@ -14,26 +14,43 @@ rendered properly in your Markdown viewer.
-->
# Documenting a model
# Utilizing the @auto_docstring Decorator
The `@auto_docstring` decorator in Transformers generates consistent docstrings for model classes and their methods. It reduces boilerplate by automatically including standard argument descriptions while also allowing overrides to add new or custom arguments. [Contributing a new model](./modular_transformers) is easier because you don't need to manually add the standard docstrings, and only focus on documenting new arguments.
The `@auto_docstring` decorator in the Hugging Face Transformers library helps generate docstrings for model classes and their methods, which will be used to build the documentation for the library. It aims to improve consistency and reduce boilerplate by automatically including standard argument descriptions and allowing for targeted overrides and additions.
This guide describes how to use the `@auto_docstring` decorator and how it works.
---
## @auto_docstring
## 📜 How it Works
Start by importing the decorator in the modeling file (`modular_model.py` or `modeling_model.py`).
The `@auto_docstring` decorator constructs docstrings by:
1. **Signature Inspection:** It inspects the signature (arguments, types, defaults) of the decorated class's `__init__` method or the decorated function.
2. **Centralized Docstring Fetching:** It retrieves predefined docstrings for common arguments (e.g., `input_ids`, `attention_mask`) from internal library sources (like `ModelArgs` or `ImageProcessorArgs` in `utils/args_doc.py`).
3. **Overriding or Adding Arguments Descriptions:**
* **Direct Docstring Block:** It incorporates custom docstring content from an `r""" """` (or `""" """`) block below the method signature or within the `__init__` docstring. This is for documenting new arguments or overriding standard descriptions.
* **Decorator Arguments (`custom_args`):** A `custom_args` docstring block can be passed to the decorator to provide docstrings for specific arguments directly in the decorator call. This can be used to define the docstring block for new arguments once if they are repeated in multiple places in the modeling file.
4. **Adding Classes and Functions Introduction:**
* **`custom_intro` argument:** Allows prepending a custom introductory paragraph to a class or function docstring.
* **Automatic Introduction Generation:** For model classes with standard naming patterns (like `ModelForCausalLM`) or belonging to a pipeline, the decorator automatically generates an appropriate introductory paragraph using `ClassDocstring` in `utils/args_doc.py` as the source.
5. **Templating:** The decorator uses a templating system, allowing predefined docstrings to include dynamic information deduced from the `auto_modules` of the library, such as `{{processor_class}}` or `{{config_class}}`.
6. **Deducing Relevant Examples:** The decorator attempts to find appropriate usage examples based on the model's task or pipeline compatibility. It extracts checkpoint information from the model's configuration class to provide concrete examples with real model identifiers.
7. **Adding Return Value Documentation:** For methods like `forward`, the decorator can automatically generate the "Returns" section based on the method's return type annotation. For example, for a method returning a `ModelOutput` subclass, it will extracts field descriptions from that class's docstring to create a comprehensive return value description. A custom `Returns` section can also be manually specified in the function docstring block.
8. **Unrolling Kwargs Typed With Unpack Operator:** For specific methods (defined in `UNROLL_KWARGS_METHODS`) or classes (defined in `UNROLL_KWARGS_CLASSES`), the decorator processes `**kwargs` parameters that are typed with `Unpack[KwargsTypedDict]`. It extracts the documentation from the TypedDict and adds each parameter to the function's docstring. Currently, this functionality is only supported for `FastImageProcessorKwargs`.
---
## 🚀 How to Use @auto_docstring
### 1. Importing the Decorator
Import the decorator into your modeling file:
```python
from ...utils import auto_docstring
```
Select whether you'd like to apply `@auto_docstring` to a class or function below to see how to use it.
<hfoptions id="type">
<hfoption id="classes">
Place `@auto_docstring` directly above the class definition. The decorator derives parameter descriptions from the `__init__` method's signature and docstring.
### 2. Applying to Classes
Place `@auto_docstring` directly above the class definition. It uses the `__init__` method's signature and its docstring for parameter descriptions.
```python
from transformers.modeling_utils import PreTrainedModel
@ -56,7 +73,9 @@ class MyAwesomeModel(PreTrainedModel):
# ... other methods
```
Arguments can also be passed directly to `@auto_docstring` for more control. Use the `custom_intro` parameter to describe the argument and the `custom_args` parameter to describe the arguments.
#### Advanced Class Decoration:
Arguments can be passed directly to `@auto_docstring` for more control:
```python
@auto_docstring(
@ -64,9 +83,9 @@ Arguments can also be passed directly to `@auto_docstring` for more control. Use
It builds upon the standard Transformer architecture with unique modifications.""",
custom_args="""
custom_parameter (`type`, *optional*, defaults to `default_value`):
A concise description for custom_parameter if not defined or overriding the description in `auto_docstring.py`.
A concise description for custom_parameter if not defined or overriding the description in `args_doc.py`.
internal_helper_arg (`type`, *optional*, defaults to `default_value`):
A concise description for internal_helper_arg if not defined or overriding the description in `auto_docstring.py`.
A concise description for internal_helper_arg if not defined or overriding the description in `args_doc.py`.
"""
)
class MySpecialModel(PreTrainedModel):
@ -74,7 +93,7 @@ class MySpecialModel(PreTrainedModel):
# ...
```
You can also choose to only use `custom_intro` and define the custom arguments directly in the class.
Or:
```python
@auto_docstring(
@ -85,44 +104,15 @@ class MySpecialModel(PreTrainedModel):
def __init__(self, config: ConfigType, custom_parameter: "type" = "default_value", internal_helper_arg=None):
r"""
custom_parameter (`type`, *optional*, defaults to `default_value`):
A concise description for custom_parameter if not defined or overriding the description in `auto_docstring.py`.
A concise description for custom_parameter if not defined or overriding the description in `args_doc.py`.
internal_helper_arg (`type`, *optional*, defaults to `default_value`):
A concise description for internal_helper_arg if not defined or overriding the description in `auto_docstring.py`.
A concise description for internal_helper_arg if not defined or overriding the description in `args_doc.py`.
"""
# ...
```
You should also use the `@auto_docstring` decorator for classes that inherit from [`~utils.ModelOutput`].
```python
@dataclass
@auto_docstring(
custom_intro="""
Custom model outputs with additional fields.
"""
)
class MyModelOutput(ImageClassifierOutput):
r"""
loss (`torch.FloatTensor`, *optional*):
The loss of the model.
custom_field (`torch.FloatTensor` of shape `(batch_size, hidden_size)`, *optional*):
A custom output field specific to this model.
"""
# Standard fields like hidden_states, logits, attentions etc. can be automatically documented if the description is the same as the standard arguments.
# However, given that the loss docstring is often different per model, you should document it in the docstring above.
loss: Optional[torch.FloatTensor] = None
logits: Optional[torch.FloatTensor] = None
hidden_states: Optional[tuple[torch.FloatTensor, ...]] = None
attentions: Optional[tuple[torch.FloatTensor, ...]] = None
# Custom fields need to be documented in the docstring above
custom_field: Optional[torch.FloatTensor] = None
```
</hfoption>
<hfoption id="functions">
Place `@auto_docstring` directly above the method definition. The decorator derives parameter descriptions from the function signature.
### 3. Applying to Functions (e.g., `forward` method)
Apply the decorator above method definitions, such as the `forward` method.
```python
@auto_docstring
@ -141,10 +131,9 @@ Place `@auto_docstring` directly above the method definition. The decorator deri
# ...
```
Arguments can also be passed directly to `@auto_docstring` for more control. Use the `custom_intro` parameter to describe the argument and the `custom_args` parameter to describe the arguments.
The `Returns` and `Examples` parts of the docstring can also be manually specified.
#### Advanced Function Decoration:
Arguments can be passed directly to `@auto_docstring` for more control. `Returns` and `Examples` sections can also be manually specified:
```python
MODEL_COMMON_CUSTOM_ARGS = r"""
@ -191,117 +180,100 @@ class MyModel(PreTrainedModel):
# ...
```
</hfoption>
</hfoptions>
---
## Documenting arguments
### ✍️ Documenting Arguments: Approach & Priority
There are some rules for documenting different types of arguments and they're listed below.
- Standard arguments (`input_ids`, `attention_mask`, `pixel_values`, etc.) are defined and retrieved from `auto_docstring.py`. It is the single source of truth for standard arguments and should not be redefined locally if an argument's description and shape is the same as an argument in `auto_docstring.py`.
If a standard argument behaves differently in your model, then you can override it locally in a `r""" """` block. This local definition has a higher priority. For example, the `labels` argument is often customized per model and typically requires overriding.
- New or custom arguments should be documented within an `r""" """` block after the signature if it is a function or in the `__init__` method's docstring if it is a class.
```py
argument_name (`type`, *optional*, defaults to `X`):
Description of the argument.
Explain its purpose, expected shape/type if complex, and default behavior.
This can span multiple lines.
```
1. **Standard Arguments (e.g., `input_ids`, `attention_mask`, `pixel_values`, `encoder_hidden_states` etc.):**
* `@auto_docstring` retrieves descriptions from a central source. Do not redefine these locally if their description and shape are the same as in `args_doc.py`.
2. **New or Custom Arguments:**
* **Primary Method:** Document these within an `r""" """` docstring block following the signature (for functions) or in the `__init__` method's docstring (for class parameters).
* **Format:**
```
argument_name (`type`, *optional*, defaults to `X`):
Description of the argument.
Explain its purpose, expected shape/type if complex, and default behavior.
This can span multiple lines.
```
* Include `type` in backticks.
* Add *optional* if the argument is not required or has a default value.
* Add "defaults to X" if it has a default value. You don't need to add "defaults to `None`" if the default value is `None`.
* Add "*optional*" if the argument is not required (has a default value).
* Add "defaults to `X`" if it has a default value (no need to specify "defaults to `None`" if the default value is `None`).
These arguments can also be passed to `@auto_docstring` as a `custom_args` argument. It is used to define the docstring block for new arguments once if they are repeated in multiple places in the modeling file.
3. **Overriding Standard Arguments:**
* If a standard argument behaves differently (e.g., different expected shape, model-specific behavior), provide its complete description in the local `r""" """` docstring. This local definition takes precedence.
* The `labels` argument is often customized per model and typically requires a specific docstring.
```py
class MyModel(PreTrainedModel):
# ...
@auto_docstring(
custom_intro="""
This is a custom introduction for the function.
"""
custom_args=r"""
common_arg_1 (`torch.Tensor`, *optional*, defaults to `default_value`):
Description of common_arg_1
"""
)
```
4. **Using Decorator Arguments for Overrides or New Arguments (`custom_args`):**
* New or custom arguments docstrings can also be passed to `@auto_docstring` as a `custom_args` argument. This can be used to define the docstring block for new arguments once if they are repeated in multiple places in the modeling file.
## Checking the docstrings
---
Transformers includes a utility script to validate the docstrings when you open a Pull Request which triggers CI (continuous integration) checks. The script checks for the following criteria.
### Usage with [modular files](./modular_transformers)
* Ensures `@auto_docstring` is applied to relevant mode classes and public methods.
* Ensures arguments are complete and consistent. It checks that documented arguments exist in the signature and verifies whether the types and default values in the docstring match the signature. Arguments that aren't known standard arguments or if they lack a local description are flagged.
* Reminds you to complete placeholders like `<fill_type>` and `<fill_docstring>`.
* Ensures docstrings are formatted according to the expected docstring style.
When working with modular files, follow these guidelines for applying the `@auto_docstring` decorator:
You can run this check locally - before committing - by running the following command.
- **For standalone models in modular files:**
Apply the `@auto_docstring` decorator just as you would in regular modeling files.
- **For models inheriting from other library models:**
- When inheriting from a parent model, decorators (including `@auto_docstring`) are automatically carried over to the generated modeling file without needing to add them in your modular file.
- If you need to modify the `@auto_docstring` behavior, apply the customized decorator in your modular file, making sure to *include all other decorators* that were present on the original function/class.
> **Warning**: When overriding any decorator in a modular file, you must include ALL decorators that were applied to that function/class in the parent model. If you only override some decorators, the others won't be included in the generated modeling file.
**Note**: The `check_auto_docstrings` tool doesn't check modular files directly, but it will check (and modify when using `--fix_and_overwrite`) the generated modeling files. If issues are found in the generated files, you'll need to update your modular files accordingly.
---
## ✅ Checking Your Docstrings with `check_auto_docstrings`
The library includes a utility script to validate docstrings. This check is typically run during Continuous Integration (CI).
#### What it Checks:
* **Decorator Presence:** Ensures `@auto_docstring` is applied to relevant model classes and public methods. (TODO)
* **Argument Completeness & Consistency:**
* Flags arguments in the signature that are not known standard arguments and lack a local description.
* Ensures documented arguments exist in the signature. (TODO)
* Verifies that types and default values in the docstring match the signature. (TODO)
* **Placeholder Detection:** Reminds you to complete placeholders like `<fill_type>` or `<fill_docstring>`.
* **Formatting:** Adherence to the expected docstring style.
#### Running the Check Locally:
Run this check locally before committing. The common command is:
```bash
make fix-copies
```
`make fix-copies` runs several other checks as well. If you don't need those checks, run the command below to only perform docstring and auto-docstring checks.
Alternatively, to only perform docstrings and auto-docstring checks, you can use:
```bash
python utils/check_docstrings.py # to only check files included in the diff without fixing them
# python utils/check_docstrings.py --fix_and_overwrite # to fix and overwrite the files in the diff
# python utils/check_docstrings.py --fix_and_overwrite --check_all # to fix and overwrite all files
# Or: python utils/check_docstrings.py --fix_and_overwrite # to fix and overwrite the files in the diff
# Or: python utils/check_docstrings.py --fix_and_overwrite --check_all # to fix and overwrite all files
```
## modular_model.py files
#### Workflow with the Checker:
When working with modular files (`modular_model.py`), follow the guidelines below for applying `@auto_docstring`.
1. Add `@auto_docstring(...)` to the class or method.
2. For new, custom, or overridden arguments, add descriptions in an `r""" """` block.
3. Run `make fix-copies` (or the `check_docstrings.py` utility).
* For unrecognized arguments lacking documentation, the utility will create placeholder entries.
4. Manually edit these placeholders with accurate types and descriptions.
5. Re-run the check to ensure all issues are resolved.
- For standalone models in modular files, apply `@auto_docstring` like you would in a `modeling_model.py` file.
- For models that inherit from other library models, `@auto_docstring` is automatically carried over to the generated modeling file. You don't need to add `@auto_docstring` in your modular file.
---
If you need to modify the `@auto_docstring` behavior, apply the customized decorator in your modular file. Make sure to **include all other decorators** that are present in the original function or class.
## 🔑 Key Takeaways & Best Practices
> [!WARNING]
> When overriding any decorator in a modular file, you must include **all** decorators that were applied to that function or class in the parent model. If you only override some decorators, the others won't be included in the generated modeling file.
## How it works
The `@auto_docstring` decorator automatically generates docstrings by:
1. Inspecting the signature (arguments, types, defaults) of the decorated class' `__init__` method or the decorated function.
2. Retrieving the predefined docstrings for common arguments (`input_ids`, `attention_mask`, etc.) from internal library sources like [`ModelArgs`], [`ImageProcessorArgs`], and the `auto_docstring.py` file.
3. Adding argument descriptions in one of two ways as shown below.
| method | description | usage |
|---|---|---|
| `r""" """` | add custom docstring content directly to a method signature or within the `__init__` docstring | document new arguments or override standard descriptions |
| `custom_args` | add custom docstrings for specific arguments directly in `@auto_docstring` | define docstring for new arguments once if they're repeated in multiple places in the modeling file |
4. Adding class and function descriptions. For model classes with standard naming patterns, like `ModelForCausalLM`, or if it belongs to a pipeline, `@auto_docstring` automatically generates the appropriate descriptions with `ClassDocstring` from `auto_docstring.py`.
`@auto_docstring` also accepts the `custom_intro` argument to describe a class or function.
5. Using a templating system to allow predefined docstrings to include dynamic information from Transformers' [auto_modules](https://github.com/huggingface/transformers/tree/main/src/transformers/models/auto) such as `{{processor_class}}` and `{{config_class}}`.
6. Finding appropriate usage examples based on the model's task or pipeline compatibility. It extracts checkpoint information form the model's configuration class to provide concrete examples with real model identifiers.
7. Adding return values to the docstring. For methods like `forward`, the decorator automatically generates the `Returns` field in the docstring based on the method's return type annotation.
For example, if a method returns a [`~transformers.utils.ModelOutput`] subclass, `@auto_docstring` extracts the field descriptions from the class' docstring to create a comprehensive return value description. You can also manually specify a custom `Returns` field in a functions docstring.
8. Unrolling kwargs typed with the unpack operator. For specific methods (defined in `UNROLL_KWARGS_METHODS`) or classes (defined in `UNROLL_KWARGS_CLASSES`), the decorator processes `**kwargs` parameters that are typed with `Unpack[KwargsTypedDict]`. It extracts the documentations from the `TypedDict` and adds each parameter to the function's docstring.
Currently only supported for [`FastImageProcessorKwargs`].
## Best practices
Follow the best practices below to help maintain consistent and informative documentation for Transformers!
* Use `@auto_docstring` for new PyTorch model classes ([`PreTrainedModel`] subclasses) and their primary methods like `forward` or `get_text_features`.
* For classes, `@auto_docstring` retrieves parameter descriptions from the `__init__` method's docstring.
* Rely on standard docstrings and do not redefine common arguments unless their behavior is different in your model.
* Use `@auto_docstring` for new PyTorch model classes (`PreTrainedModel` subclasses) and their primary for methods (e.g., `forward`, `get_text_features` etc.).
* For classes, the `__init__` method's docstring is the main source for parameter descriptions when using `@auto_docstring` on the class.
* Rely on standard docstrings; do not redefine common arguments unless their behavior is different in your specific model.
* Document new or custom arguments clearly.
* Run `check_docstrings` locally and iteratively.
By following these guidelines, you help maintain consistent and informative documentation for the Hugging Face Transformers library 🤗.

View File

@ -82,37 +82,41 @@ When you use Transformers' [`Cache`] class, the self-attention module performs s
## Cache storage implementation
Caches are structured as a list of layers, where each layer contains a key and value cache. The key and value caches are tensors with the shape `[batch_size, num_heads, seq_len, head_dim]`.
The actual storage of key-value pairs varies between cache implementations. As an example, consider the [`DynamicCache`].
Layers can be of different types (e.g. `DynamicLayer`, `StaticLayer`, `SlidingWindowLayer`), which mostly changes how sequence length is handled and how the cache is updated.
The simplest is a `DynamicLayer` that grows as more tokens are processed. The sequence length dimension (`seq_len`) increases with each new token:
In [`DynamicCache`], the key-value pairs are stored as two lists of tensors. Each tensor in the lists have the shape `[batch_size, num_heads, seq_len, head_dim]`.
- `key_cache`: A list of tensors, one for each layer.
- `value_cache`: A list of tensors, one for each layer.
When new tokens are processed:
1. For each layer, the new key and value states are concatenated with the existing cache.
```py
cache.layers[idx].keys = torch.cat([cache.layers[idx].keys, key_states], dim=-2)
cache.layers[idx].values = torch.cat([cache.layers[idx].values, value_states], dim=-2)
self.key_cache[layer_idx] = torch.cat([self.key_cache[layer_idx], key_states], dim=-2)
self.value_cache[layer_idx] = torch.cat([self.value_cache[layer_idx], value_states], dim=-2)
```
Other layer types like `StaticLayer` and `SlidingWindowLayer` have a fixed sequence length that is set when the cache is created. This makes them compatible with `torch.compile`. In the case of `SlidingWindowLayer`, existing tokens are shifted out of the cache when a new token is added.
2. The cache grows dynamically as more tokens are processed. The sequence length dimension (`seq_len`) increases with each new token.
3. The cache maintains a count of seen tokens through `self._seen_tokens`. This is updated when the first layer processes a new token.
The example below demonstrates how to create a generation loop with [`DynamicCache`]. As discussed, the attention mask is a concatenation of past and current token values and `1` is added to the cache position for the next token.
```py
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, DynamicCache, infer_device
device = f"{infer_device()}:0"
from transformers import AutoTokenizer, AutoModelForCausalLM, DynamicCache
model_id = "meta-llama/Llama-2-7b-chat-hf"
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.bfloat16, device_map=device)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.bfloat16, device_map="cuda:0")
tokenizer = AutoTokenizer.from_pretrained(model_id)
past_key_values = DynamicCache()
messages = [{"role": "user", "content": "Hello, what's your name."}]
inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt", return_dict=True).to(model.device)
inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt", return_dict=True).to("cuda:0")
generated_ids = inputs.input_ids
cache_position = torch.arange(inputs.input_ids.shape[1], dtype=torch.int64, device=model.device)
cache_position = torch.arange(inputs.input_ids.shape[1], dtype=torch.int64, device="cuda:0")
max_new_tokens = 10
for _ in range(max_new_tokens):
@ -130,36 +134,6 @@ for _ in range(max_new_tokens):
print(tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0])
"[INST] Hello, what's your name. [/INST] Hello! My name is LLaMA,"
```
## Cache position
The cache position tracks where to insert new tokens in the attention cache. It represents the *absolute* position of each token in the context, independent of padding or batch structure. Suppose you already cached `N` tokens and are now processing `K` new tokens. The cache position for the new tokens will range from `N` to `N + K - 1`. In other words, you're processing tokens at positions - `[N, N + 1, N + 2, ..., N + K - 1]`.
Cache position is used internally for two purposes:
1. Selecting new tokens to process in the input sequence and ensuring only tokens that havent been cached yet are passed to the model's `forward`.
2. Storing key/value pairs at the correct positions in the cache. This is especially important for fixed-size caches, like [`StaticCache`], that pre-allocates a specific cache length.
The generation loop usually takes care of the cache position, but if you're writing a custom generation method, it is important that cache positions are accurate since they are used to write and read key/value states into fixed slots.
```py
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, DynamicCache, infer_device
device = f"{infer_device()}:0"
model_id = "meta-llama/Llama-2-7b-chat-hf"
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.bfloat16, device_map=device)
tokenizer = AutoTokenizer.from_pretrained(model_id)
messages = [{"role": "user", "content": "You are a helpful assistant."}]
inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt", return_dict=True).to(model.device)
generated_ids = model.generate(**inputs, use_cache=True, max_new_tokens=10)
```
## Legacy cache format
Before the [`Cache`] class, the cache used to be stored as a tuple of tuples of tensors. This format is dynamic because it grows as text is generated, similar to [`DynamicCache`].
@ -169,7 +143,7 @@ The legacy format is essentially the same data structure but organized different
- The tensors have the same shape `[batch_size, num_heads, seq_len, head_dim]`.
- The format is less flexible and doesn't support features like quantization or offloading.
If your project depends on this legacy format, we recommend to convert to [`DynamicCache`] with [`~DynamicCache.from_legacy_cache`]. Note that legacy cache format is deprecated and not used anymore in `Transformers`. You can convert back to tuple format with [`DynamicCache.to_legacy_cache`] functions, which is helpful if you have custom logic for manipulating a cache in a specific format.
If your project depends on this legacy format, you can convert between [`DynamicCache`] and a tuple of tuples as shown below with the [`~DynamicCache.from_legacy_cache`] and [`DynamicCache.to_legacy_cache`] functions. This is helpful if you have custom logic for manipulating a cache in a specific format.
```py
import torch
@ -185,4 +159,4 @@ generation_outputs = model.generate(**inputs, return_dict_in_generate=True, retu
cache = DynamicCache.from_legacy_cache(generation_outputs.past_key_values)
legacy_format_cache = cache.to_legacy_cache()
```
```

View File

@ -161,7 +161,7 @@ It can be confusing to manage multiple templates though, so we recommend using a
It is important to set a chat template format that matches the template format a model was pretrained on, otherwise performance may suffer. Even if youre training the model further, performance is best if the chat tokens are kept constant.
But if youre training a model from scratch or finetuning a model for chat, you have more options to select a template. For example, [ChatML](https://github.com/openai/openai-python/blob/release-v0.28.0/chatml.md) is a popular format that is flexible enough to handle many use cases. It even includes support for [generation prompts](#add_generation_prompt), but it doesnt add beginning-of-string (`BOS`) or end-of-string (`EOS`) tokens. If your model expects `BOS` and `EOS` tokens, set `add_special_tokens=True` and make sure to add them to your template.
But if youre training a model from scratch or finetuning a model for chat, you have more options to select a template. For example, [ChatML](https://github.com/openai/openai-python/blob/release-v0.28.0/chatml.md) is a popular format that is flexbile enough to handle many use cases. It even includes support for [generation prompts](#add_generation_prompt), but it doesnt add beginning-of-string (`BOS`) or end-of-string (`EOS`) tokens. If your model expects `BOS` and `EOS` tokens, set `add_special_tokens=True` and make sure to add them to your template.
```py
{%- for message in messages %}

View File

@ -56,7 +56,7 @@ Create a [`ImageTextToTextPipeline`] and pass the chat to it. For large models,
import torch
from transformers import pipeline
pipeline = pipeline("image-text-to-text", model="llava-hf/llava-onevision-qwen2-0.5b-ov-hf", device_map="auto", torch_dtype=torch.float16)
pipeline = pipeline("image-text-to-text", model="llava-hf/llava-onevision-qwen2-0.5b-ov-hf", device="cuda", torch_dtype=torch.float16)
pipeline(text=messages, max_new_tokens=50, return_full_text=False)
[{'input_text': [{'role': 'system',
'content': [{'type': 'text',
@ -111,7 +111,6 @@ Some vision models also support video inputs. The message format is very similar
- The content `"type"` should be `"video"` to indicate the content is a video.
- For videos, it can be a link to the video (`"url"`) or it could be a file path (`"path"`). Videos loaded from a URL can only be decoded with [PyAV](https://pyav.basswood-io.com/docs/stable/) or [Decord](https://github.com/dmlc/decord).
- In addition to loading videos from a URL or file path, you can also pass decoded video data directly. This is useful if youve already preprocessed or decoded video frames elsewhere in memory (e.g., using OpenCV, decord, or torchvision). You don't need to save to files or store it in an URL.
> [!WARNING]
> Loading a video from `"url"` is only supported by the PyAV or Decord backends.
@ -138,52 +137,6 @@ messages = [
]
```
### Example: Passing decoded video objects
```python
import numpy as np
video_object1 = np.random.randint(0, 255, size=(16, 224, 224, 3), dtype=np.uint8),
messages = [
{
"role": "system",
"content": [{"type": "text", "text": "You are a friendly chatbot who always responds in the style of a pirate"}],
},
{
"role": "user",
"content": [
{"type": "video", "video": video_object1},
{"type": "text", "text": "What do you see in this video?"}
],
},
]
```
You can also use existing (`"load_video()"`) function to load a video, edit the video in memory and pass it in the messages.
```python
# Make sure a video backend library (pyav, decord, or torchvision) is available.
from transformers.video_utils import load_video
# load a video file in memory for testing
video_object2, _ = load_video(
"https://test-videos.co.uk/vids/bigbuckbunny/mp4/h264/720/Big_Buck_Bunny_720_10s_10MB.mp4"
)
messages = [
{
"role": "system",
"content": [{"type": "text", "text": "You are a friendly chatbot who always responds in the style of a pirate"}],
},
{
"role": "user",
"content": [
{"type": "video", "video": video_object2},
{"type": "text", "text": "What do you see in this video?"}
],
},
]
```
Pass `messages` to [`~ProcessorMixin.apply_chat_template`] to tokenize the input content. There are a few extra parameters to include in [`~ProcessorMixin.apply_chat_template`] that controls the sampling process.
The `video_load_backend` parameter refers to a specific framework to load a video. It supports [PyAV](https://pyav.basswood-io.com/docs/stable/), [Decord](https://github.com/dmlc/decord), [OpenCV](https://github.com/opencv/opencv), and [torchvision](https://pytorch.org/vision/stable/index.html).
@ -222,7 +175,7 @@ processed_chat = processor.apply_chat_template(
add_generation_prompt=True,
tokenize=True,
return_dict=True,
video_fps=16,
video_fps=32,
video_load_backend="decord",
)
print(processed_chat.keys())

View File

@ -25,9 +25,9 @@ Check model leaderboards like [OpenLLM](https://hf.co/spaces/HuggingFaceH4/open_
This guide shows you how to quickly start chatting with Transformers from the command line, how build and format a conversation, and how to chat using the [`TextGenerationPipeline`].
## chat CLI
## transformers CLI
After you've [installed Transformers](./installation), chat with a model directly from the command line as shown below. It launches an interactive session with a model, with a few base commands listed at the start of the session.
After you've [installed Transformers](./installation.md), chat with a model directly from the command line as shown below. It launches an interactive session with a model, with a few base commands listed at the start of the session.
```bash
transformers chat Qwen/Qwen2.5-0.5B-Instruct
@ -49,8 +49,7 @@ For a full list of options, run the command below.
transformers chat -h
```
The chat is implemented on top of the [AutoClass](./model_doc/auto), using tooling from [text generation](./llm_tutorial) and [chat](./chat_templating). It uses the `transformers serve` CLI under the hood ([docs](./serving.md#serve-cli)).
The chat is implemented on top of the [AutoClass](./model_doc/auto), using tooling from [text generation](./llm_tutorial) and [chat](./chat_templating).
## TextGenerationPipeline
@ -158,4 +157,4 @@ The easiest solution for improving generation speed is to either quantize a mode
You can also try techniques like [speculative decoding](./generation_strategies#speculative-decoding), where a smaller model generates candidate tokens that are verified by the larger model. If the candidate tokens are correct, the larger model can generate more than one token per `forward` pass. This significantly alleviates the bandwidth bottleneck and improves generation speed.
> [!TIP]
> Parameters may not be active for every generated token in MoE models such as [Mixtral](./model_doc/mixtral), [Qwen2MoE](./model_doc/qwen2_moe), and [DBRX](./model_doc/dbrx). As a result, MoE models generally have much lower memory bandwidth requirements and can be faster than a regular LLM of the same size. However, techniques like speculative decoding are ineffective with MoE models because parameters become activated with each new speculated token.
> Parameters may not be active for every generated token in MoE models such as [Mixtral](./model_doc/mixtral), [Qwen2MoE](./model_doc/qwen2_moe.md), and [DBRX](./model_doc/dbrx). As a result, MoE models generally have much lower memory bandwidth requirements and can be faster than a regular LLM of the same size. However, techniques like speculative decoding are ineffective with MoE models because parameters become activated with each new speculated token.

View File

@ -1,42 +0,0 @@
# Using Cursor as a client of transformers serve
This example shows how to use `transformers serve` as a local LLM provider for [Cursor](https://cursor.com/), the popular IDE. In this particular case, requests to `transformers serve` will come from an external IP (Cursor's server IPs), which requires some additional setup. Furthermore, some of Cursor's requests require [CORS](https://developer.mozilla.org/en-US/docs/Web/HTTP/Guides/CORS), which is disabled by default for security reasons.
To launch a server with CORS enabled, run
```shell
transformers serve --enable-cors
```
You'll also need to expose your server to external IPs. A potential solution is to use [`ngrok`](https://ngrok.com/), which has a permissive free tier. After setting up your `ngrok` account and authenticating on your server machine, you run
```shell
ngrok http [port]
```
where `port` is the port used by `transformers serve` (`8000` by default). On the terminal where you launched `ngrok`, you'll see a https address in the "Forwarding" row, as in the image below. This is the address to send requests to.
<h3 align="center">
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/transformers_serve_ngrok.png"/>
</h3>
You're now ready to set things up on the app side! In Cursor, while you can't set a new provider, you can change the endpoint for OpenAI requests in the model selection settings. First, navigate to "Settings" > "Cursor Settings", "Models" tab, and expand the "API Keys" collapsible. To set your `transformers serve` endpoint, follow this order:
1. Unselect ALL models in the list above (e.g. `gpt4`, ...);
2. Add and select the model you want to use (e.g. `Qwen/Qwen3-4B`)
3. Add some random text to OpenAI API Key. This field won't be used, but it cant be empty;
4. Add the https address from `ngrok` to the "Override OpenAI Base URL" field, appending `/v1` to the address (i.e. `https://(...).ngrok-free.app/v1`);
5. Hit "Verify".
After you follow these steps, your "Models" tab should look like the image below. Your server should also have received a few requests from the verification step.
<h3 align="center">
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/transformers_serve_cursor.png"/>
</h3>
You are now ready to use your local model in Cursor! For instance, if you toggle the AI Pane, you can select the model you added and ask it questions about your local files.
<h3 align="center">
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/transformers_serve_cursor_chat.png"/>
</h3>

View File

@ -47,7 +47,7 @@ class ResnetConfig(PretrainedConfig):
def __init__(
self,
block_type="bottleneck",
layers: list[int] = [3, 4, 6, 3],
layers: List[int] = [3, 4, 6, 3],
num_classes: int = 1000,
input_channels: int = 3,
cardinality: int = 1,
@ -271,7 +271,7 @@ The model is ready to be pushed to the Hub now. Log in to your Hugging Face acco
<hfoption id="huggingface-CLI">
```bash
hf auth login
huggingface-cli login
```
</hfoption>

View File

@ -26,7 +26,6 @@ Pass the audio signal, typically stored in `array`, to the feature extractor and
from transformers import AutoFeatureExtractor
feature_extractor = AutoFeatureExtractor.from_pretrained("facebook/wav2vec2-base")
dataset = load_dataset("PolyAI/minds14", name="en-US", split="train")
processed_sample = feature_extractor(dataset[0]["audio"]["array"], sampling_rate=16000)
processed_sample
{'input_values': [array([ 9.4472744e-05, 3.0777880e-03, -2.8888427e-03, ...,

View File

@ -315,37 +315,37 @@ tokenizer.batch_decode(outputs, skip_special_tokens=True)
```
## Custom generation methods
## Custom decoding methods
Custom generation methods enable specialized behavior such as:
Custom decoding methods enable specialized generation behavior such as the following:
- have the model continue thinking if it is uncertain;
- roll back generation if the model gets stuck;
- handle special tokens with custom logic;
- use specialized KV caches;
- enhanced input preparation for advanced models;
We enable custom generation methods through model repositories, assuming a specific model tag and file structure (see subsection below). This feature is an extension of [custom modeling code](./models.md#custom-models) and, like such, requires setting `trust_remote_code=True`.
We enable custom decoding methods through model repositories, assuming a specific model tag and file structure (see subsection below). This feature is an extension of [custom modeling code](./models.md#custom-models) and, like such, requires setting `trust_remote_code=True`.
If a model repository holds a custom generation method, the easiest way to try it out is to load the model and generate with it:
If a model repository holds a custom decoding method, the easiest way to try it out is to load the model and generate with it:
```py
from transformers import AutoModelForCausalLM, AutoTokenizer
# `transformers-community/custom_generate_example` holds a copy of `Qwen/Qwen2.5-0.5B-Instruct`, but
# with custom generation code -> calling `generate` uses the custom generation method!
# with custom generation code -> calling `generate` uses the custom decoding method!
tokenizer = AutoTokenizer.from_pretrained("transformers-community/custom_generate_example")
model = AutoModelForCausalLM.from_pretrained(
"transformers-community/custom_generate_example", device_map="auto", trust_remote_code=True
)
inputs = tokenizer(["The quick brown"], return_tensors="pt").to(model.device)
# The custom generation method is a minimal greedy decoding implementation. It also prints a custom message at run time.
# The custom decoding method is a minimal greedy decoding implementation. It also prints a custom message at run time.
gen_out = model.generate(**inputs)
# you should now see its custom message, "✨ using a custom generation method ✨"
print(tokenizer.batch_decode(gen_out, skip_special_tokens=True))
'The quick brown fox jumps over a lazy dog, and the dog is a type of animal. Is'
```
Model repositories with custom generation methods have a special property: their generation method can be loaded from **any** model through [`~GenerationMixin.generate`]'s `custom_generate` argument. This means anyone can create and share their custom generation method to potentially work with any Transformers model, without requiring users to install additional Python packages.
Model repositories with custom decoding methods have a special property: their decoding method can be loaded from **any** model through [`~GenerationMixin.generate`]'s `custom_generate` argument. This means anyone can create and share their custom generation method to potentially work with any Transformers model, without requiring users to install additional Python packages.
```py
from transformers import AutoModelForCausalLM, AutoTokenizer
@ -354,7 +354,7 @@ tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-0.5B-Instruct")
model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-0.5B-Instruct", device_map="auto")
inputs = tokenizer(["The quick brown"], return_tensors="pt").to(model.device)
# `custom_generate` replaces the original `generate` by the custom generation method defined in
# `custom_generate` replaces the original `generate` by the custom decoding method defined in
# `transformers-community/custom_generate_example`
gen_out = model.generate(**inputs, custom_generate="transformers-community/custom_generate_example", trust_remote_code=True)
print(tokenizer.batch_decode(gen_out, skip_special_tokens=True)[0])
@ -364,7 +364,7 @@ print(tokenizer.batch_decode(gen_out, skip_special_tokens=True)[0])
You should read the `README.md` file of the repository containing the custom generation strategy to see what the new arguments and output type differences are, if they exist. Otherwise, you can assume it works like the base [`~GenerationMixin.generate`] method.
> [!TIP]
> You can find all custom generation methods by [searching for their custom tag.](https://huggingface.co/models?other=custom_generate), `custom_generate`.
> You can find all custom decoding methods by [searching for their custom tag.](https://huggingface.co/models?other=custom_generate), `custom_generate`
Consider the Hub repository [transformers-community/custom_generate_example](https://huggingface.co/transformers-community/custom_generate_example) as an example. The `README.md` states that it has an additional input argument, `left_padding`, which adds a number of padding tokens before the prompt.
@ -387,11 +387,11 @@ torch>=99.0 (installed: 2.6.0)
Updating your Python requirements accordingly will remove this error message.
### Creating a custom generation method
### Creating a custom decoding method
To create a new generation method, you need to create a new [**Model**](https://huggingface.co/new) repository and push a few files into it.
1. The model you've designed your generation method with.
2. `custom_generate/generate.py`, which contains all the logic for your custom generation method.
To create a new decoding method, you need to create a new [**Model**](https://huggingface.co/new) repository and push a few files into it.
1. The model you've designed your decoding method with.
2. `custom_generate/generate.py`, which contains all the logic for your custom decoding method.
3. `custom_generate/requirements.txt`, used to optionally add new Python requirements and/or lock specific versions to correctly use your method.
4. `README.md`, where you should add the `custom_generate` tag and document any new arguments or output type differences of your custom method here.
@ -409,7 +409,7 @@ your_repo/
#### Adding the base model
The starting point for your custom generation method is a model repository just like any other. The model to add to this repository should be the model you've designed your method with, and it is meant to be part of a working self-contained model-generate pair. When the model in this repository is loaded, your custom generation method will override `generate`. Don't worry -- your generation method can still be loaded with any other Transformers model, as explained in the section above.
The starting point for your custom decoding method is a model repository just like any other. The model to add to this repository should be the model you've designed your method with, and it is meant to be part of a working self-contained model-generate pair. When the model in this repository is loaded, your custom decoding method will override `generate`. Don't worry -- your decoding method can still be loaded with any other Transformers model, as explained in the section above.
If you simply want to copy an existing model, you can do
@ -418,13 +418,13 @@ from transformers import AutoModelForCausalLM, AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained("source/model_repo")
model = AutoModelForCausalLM.from_pretrained("source/model_repo")
tokenizer.save_pretrained("your/generation_method", push_to_hub=True)
model.save_pretrained("your/generation_method", push_to_hub=True)
tokenizer.save_pretrained("your/decoding_method", push_to_hub=True)
model.save_pretrained("your/decoding_method", push_to_hub=True)
```
#### generate.py
This is the core of your generation method. It *must* contain a method named `generate`, and this method *must* contain a `model` argument as its first argument. `model` is the model instance, which means you have access to all attributes and methods in the model, including the ones defined in [`GenerationMixin`] (like the base `generate` method).
This is the core of your decoding method. It *must* contain a method named `generate`, and this method *must* contain a `model` argument as its first argument. `model` is the model instance, which means you have access to all attributes and methods in the model, including the ones defined in [`GenerationMixin`] (like the base `generate` method).
> [!WARNING]
> `generate.py` must be placed in a folder named `custom_generate`, and not at the root level of the repository. The file paths for this feature are hardcoded.
@ -465,27 +465,19 @@ def generate(model, input_ids, generation_config=None, left_padding=None, **kwar
return input_ids
```
Follow the recommended practices below to ensure your custom generation method works as expected.
Follow the recommended practices below to ensure your custom decoding method works as expected.
- Feel free to reuse the logic for validation and input preparation in the original [`~GenerationMixin.generate`].
- Pin the `transformers` version in the requirements if you use any private method/attribute in `model`.
- You can add other files in the `custom_generate` folder, and use relative imports.
- Consider adding model validation, input validation, or even a separate test file to help users sanity-check your code in their environment.
Your custom `generate` method can relative import code from the `custom_generate` folder. For example, if you have a `utils.py` file, you can import it like this:
```py
from .utils import some_function
```
Only relative imports from the same-level `custom_generate` folder are supported. Parent/sibling folder imports are not valid. The `custom_generate` argument also works locally with any directory that contains a `custom_generate` structure. This is the recommended workflow for developing your custom generation method.
#### requirements.txt
You can optionally specify additional Python requirements in a `requirements.txt` file inside the `custom_generate` folder. These are checked at runtime and an exception will be thrown if they're missing, nudging users to update their environment accordingly.
#### README.md
The root level `README.md` in the model repository usually describes the model therein. However, since the focus of the repository is the custom generation method, we highly recommend to shift its focus towards describing the custom generation method. In addition to a description of the method, we recommend documenting any input and/or output differences to the original [`~GenerationMixin.generate`]. This way, users can focus on what's new, and rely on Transformers docs for generic implementation details.
The root level `README.md` in the model repository usually describes the model therein. However, since the focus of the repository is the custom decoding method, we highly recommend to shift its focus towards describing the custom decoding method. In addition to a description of the method, we recommend documenting any input and/or output differences to the original [`~GenerationMixin.generate`]. This way, users can focus on what's new, and rely on Transformers docs for generic implementation details.
For discoverability, we highly recommend you to add the `custom_generate` tag to your repository. To do so, the top of your `README.md` file should look like the example below. After you push the file, you should see the tag in your repository!
@ -504,36 +496,6 @@ Recommended practices:
- Add self-contained examples to enable quick experimentation.
- Describe soft-requirements such as if the method only works well with a certain family of models.
### Reusing `generate`s input preparation
If you're adding a new decoding loop, you might want to preserve the input preparation present in `generate` (batch expansion, attention masks, logits processors, stopping criteria, etc.). You can also pass a **callable** to `custom_generate` to reuse [`~GenerationMixin.generate`]s full preparation pipeline while overriding only the decoding loop.
```py
def custom_loop(model, input_ids, attention_mask, logits_processor, stopping_criteria, generation_config, **model_kwargs):
next_tokens = input_ids
while input_ids.shape[1] < stopping_criteria[0].max_length:
logits = model(next_tokens, attention_mask=attention_mask, **model_kwargs).logits
next_token_logits = logits_processor(input_ids, logits[:, -1, :])
next_tokens = torch.argmax(next_token_logits, dim=-1)[:, None]
input_ids = torch.cat((input_ids, next_tokens), dim=-1)
attention_mask = torch.cat((attention_mask, torch.ones_like(next_tokens)), dim=-1)
return input_ids
output = model.generate(
**inputs,
custom_generate=custom_loop,
max_new_tokens=10,
)
```
> [!TIP]
> If you publish a `custom_generate` repository, your `generate` implementation can itself define a callable and pass it to `model.generate()`. This lets you customize the decoding loop while still benefiting from Transformers built-in input preparation logic.
### Finding custom generation methods
You can find all custom generation methods by [searching for their custom tag.](https://huggingface.co/models?other=custom_generate), `custom_generate`. In addition to the tag, we curate two collections of `custom_generate` methods:
- [Custom generation methods - Community](https://huggingface.co/collections/transformers-community/custom-generation-methods-community-6888fb1da0efbc592d3a8ab6) -- a collection of powerful methods contributed by the community;
- [Custom generation methods - Tutorials](https://huggingface.co/collections/transformers-community/custom-generation-methods-tutorials-6823589657a94940ea02cfec) -- a collection of reference implementations for methods that previously were part of `transformers`, as well as tutorials for `custom_generate`.
## Resources

View File

@ -163,7 +163,7 @@ The intermediate embedding size of the feed forward layers is often bigger than
For an input of size `[batch_size, sequence_length]`, the memory required to store the intermediate feed forward
embeddings `[batch_size, sequence_length, config.intermediate_size]` can account for a large fraction of the memory
use. The authors of [Reformer: The Efficient Transformer](https://huggingface.co/papers/2001.04451) noticed that since the
use. The authors of [Reformer: The Efficient Transformer](https://arxiv.org/abs/2001.04451) noticed that since the
computation is independent of the `sequence_length` dimension, it is mathematically equivalent to compute the output
embeddings of both feed forward layers `[batch_size, config.hidden_size]_0, ..., [batch_size, config.hidden_size]_n`
individually and concat them afterward to `[batch_size, sequence_length, config.hidden_size]` with `n = sequence_length`, which trades increased computation time against reduced memory use, but yields a mathematically
@ -207,7 +207,7 @@ numerical representations of tokens building the sequences that will be used as
<Youtube id="VFp38yj8h3A"/>
Each tokenizer works differently but the underlying mechanism remains the same. Here's an example using the BERT
tokenizer, which is a [WordPiece](https://huggingface.co/papers/1609.08144) tokenizer:
tokenizer, which is a [WordPiece](https://arxiv.org/pdf/1609.08144.pdf) tokenizer:
```python
>>> from transformers import BertTokenizer

View File

@ -15,25 +15,9 @@ rendered properly in your Markdown viewer.
# Transformers
<h3 align="center">
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/transformers_as_a_model_definition.png"/>
</h3>
Transformers is a library of pretrained natural language processing, computer vision, audio, and multimodal models for inference and training. Use Transformers to train models on your data, build inference applications, and generate text with large language models.
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
frameworks (Axolotl, Unsloth, DeepSpeed, FSDP, PyTorch-Lightning, ...), inference engines (vLLM, SGLang, TGI, ...),
and adjacent modeling libraries (llama.cpp, mlx, ...) which leverage the model definition from `transformers`.
We pledge to help support new state-of-the-art models and democratize their usage by having their model definition be
simple, customizable, and efficient.
There are over 1M+ Transformers [model checkpoints](https://huggingface.co/models?library=transformers&sort=trending) on the [Hugging Face Hub](https://huggingface.com/models) you can use.
Explore the [Hub](https://huggingface.com/) today to find a model and use Transformers to help you get started right away.
Explore the [Hugging Face Hub](https://huggingface.com) today to find a model and use Transformers to help you get started right away.
## Features
@ -59,6 +43,3 @@ Transformers is designed for developers and machine learning engineers and resea
</a>
</div>
## Learn
If you're new to Transformers or want to learn more about transformer models, we recommend starting with the [LLM course](https://huggingface.co/learn/llm-course/chapter1/1?fw=pt). This comprehensive course covers everything from the fundamentals of how transformer models work to practical applications across various tasks. You'll learn the complete workflow, from curating high-quality datasets to fine-tuning large language models and implementing reasoning capabilities. The course contains both theoretical and hands-on exercises to build a solid foundational knowledge of transformer models as you learn.

View File

@ -48,4 +48,3 @@ Most of those are only useful if you are studying the general code in the librar
## Other Utilities
[[autodoc]] utils._LazyModule
[[autodoc]] pytorch_utils.infer_device

View File

@ -356,76 +356,65 @@ A [`Constraint`] can be used to force the generation to include specific tokens
## Caches
[[autodoc]] CacheLayerMixin
- update
- get_seq_length
- get_mask_sizes
- get_max_cache_shape
- reset
- reorder_cache
- lazy_initialization
[[autodoc]] DynamicLayer
- update
- lazy_initialization
- crop
- batch_repeat_interleave
- batch_select_indices
[[autodoc]] StaticLayer
- update
- lazy_initialization
[[autodoc]] SlidingWindowLayer
- update
- lazy_initialization
[[autodoc]] QuantoQuantizedLayer
- update
- lazy_initialization
[[autodoc]] HQQQuantizedLayer
- update
- lazy_initialization
[[autodoc]] Cache
- update
- early_initialization
- get_seq_length
- get_mask_sizes
- get_max_cache_shape
- reset
- reorder_cache
- crop
- batch_repeat_interleave
- batch_select_indices
[[autodoc]] CacheConfig
- update
[[autodoc]] QuantizedCacheConfig
- validate
[[autodoc]] DynamicCache
- update
- get_seq_length
- reorder_cache
- to_legacy_cache
- from_legacy_cache
[[autodoc]] QuantizedCache
- update
- get_seq_length
[[autodoc]] QuantoQuantizedCache
[[autodoc]] HQQQuantizedCache
[[autodoc]] OffloadedCache
- update
- prefetch_layer
- evict_previous_layer
[[autodoc]] StaticCache
- update
- get_seq_length
- reset
[[autodoc]] OffloadedStaticCache
- update
- get_seq_length
- reset
[[autodoc]] HybridCache
[[autodoc]] HybridChunkedCache
- update
- get_seq_length
- reset
[[autodoc]] SlidingWindowCache
- update
- reset
[[autodoc]] EncoderDecoderCache
- get_seq_length
- to_legacy_cache
- from_legacy_cache
- reset
- reorder_cache
[[autodoc]] MambaCache
- update_conv_state
- update_ssm_state
- reset
## Watermark Utils

View File

@ -38,7 +38,7 @@ However, no method can be called on that object:
```python
>>> DetrImageProcessorFast.from_pretrained()
ImportError:
DetrImageProcessorFast requires the Torchvision library but it was not found in your environment. Check out the instructions on the
DetrImageProcessorFast requires the Torchvision library but it was not found in your environment. Checkout the instructions on the
installation page: https://pytorch.org/get-started/locally/ and follow the ones that match your environment.
Please note that you may need to restart your runtime after installation.
```

View File

@ -51,7 +51,7 @@ torch.random.manual_seed(673)
# load pretrained model and processor
model_id = "llava-hf/llava-1.5-7b-hf"
processor = LlavaProcessor.from_pretrained(model_id)
model = LlavaForConditionalGeneration.from_pretrained(model_id)
model = LlavaForConditionalGeneration.from_pretrained(model_id, low_cpu_mem_usage=True)
# create random image input
random_image = Image.fromarray(torch.randint(0, 256, (224, 224, 3), dtype=torch.uint8).numpy())
@ -247,114 +247,3 @@ first and last layer will be shown. This is useful when some layers (typically c
layers.
[[autodoc]] model_addition_debugger_context
## Analyzer of skipped tests
### Scan skipped tests - for model adders and maintainers
This small util is a power user tool intended for model adders and maintainers. It lists all test methods
existing in `test_modeling_common.py`, inherited by all model tester classes, and scans the repository to measure
how many tests are being skipped and for which models.
### Rationale
When porting models to transformers, tests fail as they should, and sometimes `test_modeling_common` feels irreconcilable with the peculiarities of our brand new model. But how can we be sure we're not breaking everything by adding a seemingly innocent skip?
This utility:
- scans all test_modeling_common methods
- looks for times where a method is skipped
- returns a summary json you can load as a DataFrame/inspect
**For instance test_inputs_embeds is skipped in a whooping 39% proportion at the time of writing this util.**
![download-icon](https://huggingface.co/datasets/huggingface/documentation-images/resolve/f7f671f69b88ce4967e19179172c248958d35742/transformers/tests_skipped_visualisation.png)
### Usage
You can run the skipped test analyzer in two ways:
#### Full scan (default)
From the root of `transformers` repo, scans all common test methods and outputs the results to a JSON file (default: `all_tests_scan_result.json`).
```bash
python utils/scan_skipped_tests.py --output_dir path/to/output
```
- `--output_dir` (optional): Directory where the JSON results will be saved. Defaults to the current directory.
**Example output:**
```
🔬 Parsing 331 model test files once each...
📝 Aggregating 224 tests...
(224/224) test_update_candidate_strategy_with_matches_1es_3d_is_nonecodet_schedule_fa_kwargs
✅ Scan complete.
📄 JSON saved to /home/pablo/git/transformers/all_tests_scan_result.json
```
And it will generate `all_tests_scan_result.json` file that you can inspect. The JSON is indexed by method name, and each entry follows this schema, indicating the origin as well (from `common`or `GenerationMixin`.)
```json
{
"<method_name>": {
"origin": "<test suite>"
"models_ran": ["<model_name>", ...],
"models_skipped": ["<model_name>", ...],
"skipped_proportion": <float>,
"reasons_skipped": ["<model_name>: <reason>",
...
]
},
...
}
```
Which you can visualise as above with e.g. `pandas`
```python
df = pd.read_json('all_tests_scan_result.json').T
df.sort_values(by=['skipped_proportion'], ascending=False)
```
### Scan a single test method
You can focus on a specific test method using `--test_method_name`:
```bash
$ python utils/scan_skipped_tests.py --test_method_name test_inputs_embeds --output_dir path/to/output
```
- `--test_method_name`: Name of the test method to scan (e.g., `test_inputs_embeds`).
- `--output_dir` (optional): Directory where the JSON result will be saved.
**Example output:**
```bash
$ python utils/scan_skipped_tests.py --test_method_name test_inputs_embeds
🔬 Parsing 331 model test files once each...
== test_inputs_embeds ==
Ran : 199/323
Skipped : 124/323 (38.4%)
- aimv2: Aimv2 does not use inputs_embeds
- align: Inputs_embeds is tested in individual model tests
- altclip: Inputs_embeds is tested in individual model tests
- audio_spectrogram_transformer: AST does not use inputs_embeds
- beit: BEiT does not use inputs_embeds
- bit: Bit does not use inputs_embeds
- blip: Blip does not use inputs_embeds
- blip_2: Inputs_embeds is tested in individual model tests
- bridgetower:
- canine: CANINE does not have a get_input_embeddings() method.
- ...
📄 JSON saved to /home/pablo/git/transformers/scan_test_inputs_embeds.json
```

View File

@ -1,32 +0,0 @@
# Jan: using the serving API as a local LLM provider
This example shows how to use `transformers serve` as a local LLM provider for the [Jan](https://jan.ai/) app. Jan is a ChatGPT-alternative graphical interface, fully running on your machine. The requests to `transformers serve` come directly from the local app -- while this section focuses on Jan, you can extrapolate some instructions to other apps that make local requests.
## Running models locally
To connect `transformers serve` with Jan, you'll need to set up a new model provider ("Settings" > "Model Providers"). Click on "Add Provider", and set a new name. In your new model provider page, all you need to set is the "Base URL" to the following pattern:
```shell
http://[host]:[port]/v1
```
where `host` and `port` are the `transformers serve` CLI parameters (`localhost:8000` by default). After setting this up, you should be able to see some models in the "Models" section, hitting "Refresh". Make sure you add some text in the "API key" text field too -- this data is not actually used, but the field can't be empty. Your custom model provider page should look like this:
<h3 align="center">
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/transformers_serve_jan_model_providers.png"/>
</h3>
You are now ready to chat!
> [!TIP]
> You can add any `transformers`-compatible model to Jan through `transformers serve`. In the custom model provider you created, click on the "+" button in the "Models" section and add its Hub repository name, e.g. `Qwen/Qwen3-4B`.
## Running models on a separate machine
To conclude this example, let's look into a more advanced use-case. If you have a beefy machine to serve models with, but prefer using Jan on a different device, you need to add port forwarding. If you have `ssh` access from your Jan machine into your server, this can be accomplished by typing the following to your Jan machine's terminal
```
ssh -N -f -L 8000:localhost:8000 your_server_account@your_server_IP -p port_to_ssh_into_your_server
```
Port forwarding is not Jan-specific: you can use it to connect `transformers serve` running in a different machine with an app of your choice.

Some files were not shown because too many files have changed in this diff Show More