Release: v0.7.7

* fix ppo push to Hub

* fix also ddpo

* more tags

.github/workflows/benchmark.yml

@@ -0,0 +1,107 @@
+name: "Benchmark on Comment"
+
+# https://docs.github.com/en/actions/using-workflows/events-that-trigger-workflows
+on:
+  issue_comment:
+    types: [created]
+
+jobs:
+  Benchmark:
+    strategy:
+      fail-fast: true
+      matrix:
+        python-version: [3.9]
+        os: [self-hosted]
+
+    name: Benchmark
+    # Only run if it#s a PR and the comment contains /Benchmark
+    if: github.event.issue.pull_request && startsWith(github.event.comment.body, '/benchmark-trl-experiments') && contains(FromJSON('["vwxyzjn", "younesbelkada", "lvwerra", "lewtun"]'), github.actor)
+    runs-on: ${{ matrix.os }}
+
+    steps:
+      - name: Get branch of PR
+        uses: xt0rted/pull-request-comment-branch@v1
+        id: comment-branch
+      - name: Set latest commit status as pending
+        uses: myrotvorets/set-commit-status-action@master
+        with:
+          sha: ${{ steps.comment-branch.outputs.head_sha }}
+          token: ${{ secrets.GITHUB_TOKEN }}
+          status: pending
+      - name: Checkout `main` branch
+        uses: actions/checkout@v3
+      - name: Checkout PR branch
+        run: gh pr checkout $PR_NUMBER
+        env:
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          PR_NUMBER: ${{ github.event.issue.number }}
+      - name: Set up Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v4
+        with:
+          python-version: ${{ matrix.python-version }}
+      # - name: Cleanup pip packages (specific to self-hosted runners)
+      #   run: |
+      #     echo PATH is $PATH
+      #     echo PYTHONPATH is $PYTHONPATH
+      #     echo which python is $(which python)
+      #     echo which pip is $(which pip)
+
+      #     pip_list=$(pip list --format=freeze | grep -v "^pip==" | grep -v "^setuptools==")
+      #     if [ ! -z "$pip_list" ]; then
+      #         echo "$pip_list" | xargs pip uninstall -y
+      #     fi
+      - name: Print python depdenencies
+        run: pip list --format=freeze
+      - name: Install dependencies
+        run: |
+          pip install .[test,benchmark]
+
+      - name: Login
+        run: wandb login ${{ secrets.WANDB_API_KEY }} && huggingface-cli login --token ${{ secrets.HUGGING_FACE_HUB_TOKEN }}
+      - name: Run benchmark
+        env:
+          GITHUB_CONTEXT: ${{ toJson(github) }}
+          PERSONAL_ACCESS_TOKEN_GITHUB: ${{ secrets.PERSONAL_ACCESS_TOKEN_GITHUB }}
+        run: |
+          COMMENT="${{ github.event.comment.body }}"
+          if [[ "$COMMENT" == *"/benchmark-trl-experiments benchmark/benchmark_level1.sh"* ]]; then
+            echo "Running benchmark/benchmark_level1.sh"
+            BENCHMARK_SCRIPT="benchmark/benchmark_level1.sh" BENCHMARK_PLOT_SCRIPT="benchmark/benchmark_level1_plot.sh" bash benchmark/benchmark_and_report.sh
+          elif [[ "$COMMENT" == *"/benchmark-trl-experiments benchmark/benchmark_level2.sh"* ]]; then
+            echo "Running benchmark/benchmark_level2.sh"
+            BENCHMARK_SCRIPT="benchmark/benchmark_level2.sh" BENCHMARK_PLOT_SCRIPT="benchmark/benchmark_level2_plot.sh" bash benchmark/benchmark_and_report.sh
+          elif [[ "$COMMENT" == *"/benchmark-trl-experiments benchmark/benchmark_level3.sh"* ]]; then
+            echo "Running benchmark/benchmark_level3.sh"
+            BENCHMARK_SCRIPT="benchmark/benchmark_level3.sh" BENCHMARK_PLOT_SCRIPT="benchmark/benchmark_level3_plot.sh" bash benchmark/benchmark_and_report.sh
+          else
+            echo "Invalid command in comment. Skipping execution."
+          fi
+
+      # send message to PR
+      - name: Setup Node.js 16
+        uses: actions/setup-node@v3
+        with:
+          node-version: 16
+      - name: Add workflow result as comment on PR
+        uses: actions/github-script@v6
+        if: always()
+        with:
+          script: |
+            const name = '${{ github.workflow	}}';
+            const url = '${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}';
+            const success = '${{ job.status }}' === 'success';
+            const body = `${name}: ${success ? 'succeeded ✅' : 'failed ❌'}\n${url}`;
+
+            await github.rest.issues.createComment({
+              issue_number: context.issue.number,
+              owner: context.repo.owner,
+              repo: context.repo.repo,
+              body: body
+            })
+      - name: Set latest commit status as ${{ job.status }}
+        uses: myrotvorets/set-commit-status-action@master
+        if: always()
+        with:
+          sha: ${{ steps.comment-branch.outputs.head_sha }}
+          token: ${{ secrets.GITHUB_TOKEN }}
+          status: ${{ job.status }}

.github/workflows/build_documentation.yml

@@ -13,7 +13,6 @@ jobs:
     with:
       commit_sha: ${{ github.sha }}
       package: trl
-      repo_owner: lvwerra
       version_tag_suffix: ""
     secrets:
-      token: ${{ secrets.HUGGINGFACE_PUSH }}
+      hf_token: ${{ secrets.HF_DOC_BUILD_PUSH }}

.github/workflows/build_pr_documentation.yml

@@ -14,5 +14,4 @@ jobs:
       commit_sha: ${{ github.event.pull_request.head.sha }}
       pr_number: ${{ github.event.number }}
       package: trl
-      repo_owner: lvwerra
       version_tag_suffix: ""

.github/workflows/clear_cache.yml

@@ -0,0 +1,33 @@
+name: "Cleanup Cache"
+
+on:
+  workflow_dispatch:
+  schedule:
+    - cron: "0 0 * * *"
+    
+jobs:
+  cleanup:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Check out code
+        uses: actions/checkout@v3
+        
+      - name: Cleanup
+        run: |
+          gh extension install actions/gh-actions-cache
+          
+          REPO=${{ github.repository }}
+
+          echo "Fetching list of cache key"
+          cacheKeysForPR=$(gh actions-cache list -R $REPO | cut -f 1 )
+
+          ## Setting this to not fail the workflow while deleting cache keys. 
+          set +e
+          echo "Deleting caches..."
+          for cacheKey in $cacheKeysForPR
+          do
+              gh actions-cache delete $cacheKey -R $REPO --confirm
+          done
+          echo "Done"
+        env:
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}

.github/workflows/delete_doc_comment.yml

@@ -1,13 +0,0 @@
-name: Delete dev documentation
-
-on:
-  pull_request:
-    types: [ closed ]
-
-
-jobs:
-  delete:
-    uses: huggingface/doc-builder/.github/workflows/delete_doc_comment.yml@main
-    with:
-      pr_number: ${{ github.event.number }}
-      package: trl

.github/workflows/stale.yml

@@ -0,0 +1,27 @@
+name: Stale Bot
+
+on:
+  schedule:
+    - cron: "0 15 * * *"
+
+jobs:
+  close_stale_issues:
+    name: Close Stale Issues
+    if: github.repository == 'huggingface/trl'
+    runs-on: ubuntu-latest
+    env:
+      GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+    steps:
+    - uses: actions/checkout@v3
+
+    - name: Setup Python
+      uses: actions/setup-python@v4
+      with:
+        python-version: 3.8
+
+    - name: Install requirements
+      run: |
+        pip install PyGithub
+    - name: Close stale issues
+      run: |
+        python scripts/stale.py

.github/workflows/tests.yml

@@ -7,29 +7,31 @@ on:
     branches: [ main ]
 
 jobs:
-
   check_code_quality:
     runs-on: ubuntu-latest
+    strategy:
+      matrix:
+        python-version: [3.9]
+
     steps:
-      - uses: actions/checkout@v3
-      - name: Set up Python
-        uses: actions/setup-python@v4
+      - uses: actions/checkout@v2
         with:
-          python-version: "3.8"
-      - name: Install dependencies
-        run: |
-          python -m pip install --upgrade pip
-          pip install .[dev]
-      - name: Check quality
-        run: |
-          make quality
+          fetch-depth: 0
+          submodules: recursive
+      - name: Set up Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v2
+        with:
+          python-version: ${{ matrix.python-version }}
+      - uses: pre-commit/action@v2.0.3
+        with:
+          extra_args: --all-files
 
   tests:
     needs: check_code_quality
     strategy:
       matrix:
-        python-version: [3.7, 3.8, 3.9]
-        os: ['ubuntu-latest', 'macos-latest', 'windows-latest']
+        python-version: ['3.8', '3.9', '3.10']
+        os: ['ubuntu-latest', 'windows-latest']
     runs-on: ${{ matrix.os }}
     steps:
     - uses: actions/checkout@v3
@@ -37,6 +39,32 @@ jobs:
       uses: actions/setup-python@v4
       with:
         python-version: ${{ matrix.python-version }}
+        cache: "pip"
+        cache-dependency-path: |
+            setup.py
+            requirements.txt
+    - name: Install dependencies
+      run: |
+        python -m pip install --upgrade pip
+        # cpu version of pytorch
+        pip install -e ".[test, peft, diffusers]"
+    - name: Test with pytest
+      run: |
+        make test
+
+  tests_no_optional_dep:
+    needs: check_code_quality
+    runs-on: 'ubuntu-latest'
+    steps:
+    - uses: actions/checkout@v3
+    - name: Set up Python 3.9
+      uses: actions/setup-python@v4
+      with:
+        python-version: '3.9'
+        cache: "pip"
+        cache-dependency-path: |
+            setup.py
+            requirements.txt
     - name: Install dependencies
       run: |
         python -m pip install --upgrade pip
@@ -44,4 +72,4 @@ jobs:
         pip install .[test]
     - name: Test with pytest
       run: |
-        make test
+        make test

.github/workflows/upload_pr_documentation.yml

@@ -0,0 +1,16 @@
+name: Upload PR Documentation
+
+on:
+  workflow_run:
+    workflows: ["Build PR Documentation"]
+    types:
+      - completed
+
+jobs:
+  build:
+    uses: huggingface/doc-builder/.github/workflows/upload_pr_documentation.yml@main
+    with:
+      package_name: trl
+    secrets:
+      hf_token: ${{ secrets.HF_DOC_BUILD_PUSH }}
+      comment_bot_token: ${{ secrets.COMMENT_BOT_TOKEN }}

.gitignore

@@ -1,3 +1,4 @@
+benchmark/trl
 *.bak
 .gitattributes
 .last_checked

.pre-commit-config.yaml

@@ -0,0 +1,42 @@
+repos:
+  - repo: https://github.com/PyCQA/isort
+    rev: 5.12.0
+    hooks:
+      - id: isort
+        args:
+          - --profile=black
+          - --skip-glob=wandb/**/*
+          - --thirdparty=wandb
+  - repo: https://github.com/myint/autoflake
+    rev: v1.4
+    hooks:
+      - id: autoflake
+        args:
+          - -r
+          - --exclude=wandb,__init__.py
+          - --in-place
+          - --remove-unused-variables
+          - --remove-all-unused-imports
+  - repo: https://github.com/python/black
+    rev: 22.3.0
+    hooks:
+      - id: black
+        args:
+          - --line-length=119
+          - --target-version=py38
+          - --exclude=wandb
+  - repo: https://github.com/pycqa/flake8
+    rev: 6.0.0
+    hooks:
+      - id: flake8
+        args:
+          - --ignore=E203,E501,W503,E128
+          - --max-line-length=119
+
+  # - repo: https://github.com/codespell-project/codespell
+  #   rev: v2.1.0
+  #   hooks:
+  #     - id: codespell
+  #       args:
+  #         - --ignore-words-list=nd,reacher,thist,ths,magent,ba
+  #         - --skip=docs/css/termynal.css,docs/js/termynal.js

CITATION.cff

@@ -17,7 +17,7 @@ authors:
     family-names: Thrush
   - given-names: Nathan
     family-names: Lambert
-repository-code: 'https://github.com/lvwerra/trl'
+repository-code: 'https://github.com/huggingface/trl'
 abstract: "With trl you can train transformer language models with Proximal Policy Optimization (PPO). The library is built on top of the transformers library by \U0001F917 Hugging Face. Therefore, pre-trained language models can be directly loaded via transformers. At this point, most decoder and encoder-decoder architectures are supported."
 keywords:
   - rlhf

CONTRIBUTING.md

@@ -36,10 +36,15 @@ First you want to make sure that all the tests pass:
 make test
 ```
 
-Then before submitting your PR make sure the code quality follows the standards. You can run the following command to format and test:
+Then before submitting your PR make sure the code quality follows the standards. You can run the following command to format:
 
 ```bash
-make style && make quality
+make precommit
+```
+
+Make sure to install `pre-commit` before running the command:
+```bash
+pip install pre-commit
 ```
 
 ## Do you want to contribute to the documentation?

Makefile

@@ -1,13 +1,15 @@
-.PHONY: quality style test
+.PHONY: test precommit benchmark_core benchmark_aux
+
+check_dirs := examples tests trl
 
 test:
 	python -m pytest -n auto --dist=loadfile -s -v ./tests/
 
-quality:
-	black --check --line-length 119 --target-version py38 tests trl
-	isort --check-only tests trl
-	flake8  tests trl
+precommit:
+	pre-commit run --all-files
 
-style:
-	black --line-length 119 --target-version py38 tests trl examples setup.py
-	isort tests trl
+benchmark_core:
+	bash ./benchmark/benchmark_core.sh
+
+benchmark_aux:
+	bash ./benchmark/benchmark_aux.sh

README.md

@@ -3,23 +3,43 @@
 </div>
 
 # TRL - Transformer Reinforcement Learning
-> Train transformer language models with reinforcement learning.
+> Full stack transformer language models with reinforcement learning.
+
+<p align="center">
+    <a href="https://github.com/huggingface/trl/blob/main/LICENSE">
+        <img alt="License" src="https://img.shields.io/github/license/huggingface/trl.svg?color=blue">
+    </a>
+    <a href="https://huggingface.co/docs/trl/index">
+        <img alt="Documentation" src="https://img.shields.io/website/http/huggingface.co/docs/trl/index.svg?down_color=red&down_message=offline&up_message=online">
+    </a>
+    <a href="https://github.com/huggingface/trl/releases">
+        <img alt="GitHub release" src="https://img.shields.io/github/release/huggingface/trl.svg">
+    </a>
+</p>
 
 
 ## What is it?
-With `trl` you can train transformer language models with Proximal Policy Optimization (PPO). The library is built on top of the [`transformers`](https://github.com/huggingface/transformers) library by  🤗 Hugging Face. Therefore, pre-trained language models can be directly loaded via `transformers`. At this point most of decoder architectures and encoder-decoder architectures are supported. 
+
+<div style="text-align: center">
+<img src="https://huggingface.co/datasets/trl-internal-testing/example-images/resolve/main/images/TRL-readme.png">
+</div>
+
+`trl` is a full stack library where we provide a set of tools to train transformer language models and stable diffusion models with Reinforcement Learning, from the Supervised Fine-tuning step (SFT), Reward Modeling step (RM) to the Proximal Policy Optimization (PPO) step. The library is built on top of the [`transformers`](https://github.com/huggingface/transformers) library by  🤗 Hugging Face. Therefore, pre-trained language models can be directly loaded via `transformers`. At this point, most of decoder architectures and encoder-decoder architectures are supported. Refer to the documentation or the `examples/` folder for example code snippets and how to run these tools.
 
 **Highlights:**
-- `PPOTrainer`: A PPO trainer for language models that just needs (query, response, reward) triplets to optimise the language model.
-- `AutoModelForCausalLMWithValueHead` & `AutoModelForSeq2SeqLMWithValueHead`: A transformer model with an additional scalar output for each token which can be used as a value function in reinforcement learning.
-- Example: Train GPT2 to generate positive movie reviews with a BERT sentiment classifier.
 
-## How it works
+- [`SFTTrainer`](https://huggingface.co/docs/trl/sft_trainer): A light and friendly wrapper around `transformers` Trainer to easily fine-tune language models or adapters on a custom dataset.
+- [`RewardTrainer`](https://huggingface.co/docs/trl/reward_trainer): A light wrapper around `transformers` Trainer to easily fine-tune language models for human preferences (Reward Modeling).
+- [`PPOTrainer`](https://huggingface.co/docs/trl/trainer#trl.PPOTrainer): A PPO trainer for language models that just needs (query, response, reward) triplets to optimise the language model.
+- [`AutoModelForCausalLMWithValueHead`](https://huggingface.co/docs/trl/models#trl.AutoModelForCausalLMWithValueHead) & [`AutoModelForSeq2SeqLMWithValueHead`](https://huggingface.co/docs/trl/models#trl.AutoModelForSeq2SeqLMWithValueHead): A transformer model with an additional scalar output for each token which can be used as a value function in reinforcement learning.
+- [Examples](https://github.com/huggingface/trl/tree/main/examples): Train GPT2 to generate positive movie reviews with a BERT sentiment classifier, full RLHF using adapters only, train GPT-j to be less toxic, [Stack-Llama example](https://huggingface.co/blog/stackllama), etc.
+
+## How PPO works
 Fine-tuning a language model via PPO consists of roughly three steps:
 
 1. **Rollout**: The language model generates a response or continuation based on query which could be the start of a sentence.
 2. **Evaluation**: The query and response are evaluated with a function, model, human feedback or some combination of them. The important thing is that this process should yield a scalar value for each query/response pair.
-3. **Optimization**: This is the most complex part. In the optimisation step the query/response pairs are used to calculate the log-probabilities of the tokens in the sequences. This is done with the model that is trained and and a reference model, which is usually the pre-trained model before fine-tuning. The KL-divergence between the two outputs is used as an additional reward signal to make sure the generated responses don't deviate to far from the reference language model. The active language model is then trained with PPO.
+3. **Optimization**: This is the most complex part. In the optimisation step the query/response pairs are used to calculate the log-probabilities of the tokens in the sequences. This is done with the model that is trained and a reference model, which is usually the pre-trained model before fine-tuning. The KL-divergence between the two outputs is used as an additional reward signal to make sure the generated responses don't deviate too far from the reference language model. The active language model is then trained with PPO.
 
 This process is illustrated in the sketch below:
 
@@ -40,7 +60,7 @@ pip install trl
 ### From source
 If you want to run the examples in the repository a few additional libraries are required. Clone the repository and install it with pip:
 ```bash
-git clone https://github.com/lvwerra/trl.git
+git clone https://github.com/huggingface/trl.git
 cd trl/
 pip install .
 ```
@@ -52,8 +72,59 @@ pip install -e .
 
 ## How to use
 
-### Example
-This is a basic example on how to use the library. Based on a query the language model creates a response which is then evaluated. The evaluation could be a human in the loop or another model's output.
+### `SFTTrainer`
+
+This is a basic example on how to use the `SFTTrainer` from the library. The `SFTTrainer` is a light wrapper around the `transformers` Trainer to easily fine-tune language models or adapters on a custom dataset.
+
+```python
+# imports
+from datasets import load_dataset
+from trl import SFTTrainer
+
+# get dataset
+dataset = load_dataset("imdb", split="train")
+
+# get trainer
+trainer = SFTTrainer(
+    "facebook/opt-350m",
+    train_dataset=dataset,
+    dataset_text_field="text",
+    max_seq_length=512,
+)
+
+# train
+trainer.train()
+```
+
+### `RewardTrainer`
+
+This is a basic example on how to use the `RewardTrainer` from the library. The `RewardTrainer` is a wrapper around the `transformers` Trainer to easily fine-tune reward models or adapters on a custom preference dataset.
+
+```python
+# imports
+from transformers import AutoModelForSequenceClassification, AutoTokenizer
+from trl import RewardTrainer
+
+# load model and dataset - dataset needs to be in a specific format
+model = AutoModelForSequenceClassification.from_pretrained("gpt2", num_labels=1)
+tokenizer = AutoTokenizer.from_pretrained("gpt2")
+
+...
+
+# load trainer
+trainer = RewardTrainer(
+    model=model,
+    tokenizer=tokenizer,
+    train_dataset=dataset,
+)
+
+# train
+trainer.train()
+```
+
+### `PPOTrainer`
+
+This is a basic example on how to use the `PPOTrainer` from the library. Based on a query the language model creates a response which is then evaluated. The evaluation could be a human in the loop or another model's output.
 
 ```python
 # imports
@@ -78,7 +149,7 @@ query_txt = "This morning I went to the "
 query_tensor = tokenizer.encode(query_txt, return_tensors="pt")
 
 # get model response
-response_tensor  = respond_to_batch(model_ref, query_tensor)
+response_tensor  = respond_to_batch(model, query_tensor)
 
 # create a ppo trainer
 ppo_trainer = PPOTrainer(ppo_config, model, model_ref, tokenizer)
@@ -91,14 +162,6 @@ reward = [torch.tensor(1.0)]
 train_stats = ppo_trainer.step([query_tensor[0]], [response_tensor[0]], reward)
 ```
 
-### Advanced example: IMDB sentiment
-For a detailed example check out the example python script `examples/scripts/ppo-sentiment.py`, where GPT2 is fine-tuned to generate positive movie reviews. An few examples from the language models before and after optimisation are given below:
-
-<div style="text-align: center">
-<img src="https://huggingface.co/datasets/trl-internal-testing/example-images/resolve/main/images/table_imdb_preview.png" width="800">
-<p style="text-align: center;"> <b>Figure:</b> A few review continuations before and after optimisation. </p>
-</div>
-
 ## References
 
 ### Proximal Policy Optimisation
@@ -111,11 +174,11 @@ The language models utilize the `transformers` library by 🤗 Hugging Face.
 
 ```bibtex
 @misc{vonwerra2022trl,
-  author = {Leandro von Werra and Younes Belkada and Lewis Tunstall and Edward Beeching and Tristan Thrush and Nathan Lambert},
+  author = {Leandro von Werra and Younes Belkada and Lewis Tunstall and Edward Beeching and Tristan Thrush and Nathan Lambert and Shengyi Huang},
   title = {TRL: Transformer Reinforcement Learning},
   year = {2020},
   publisher = {GitHub},
   journal = {GitHub repository},
-  howpublished = {\url{https://github.com/lvwerra/trl}}
+  howpublished = {\url{https://github.com/huggingface/trl}}
 }
-```
+```

benchmark/benchmark.py

@@ -0,0 +1,150 @@
+import argparse
+import math
+import os
+import shlex
+import subprocess
+import uuid
+from distutils.util import strtobool
+
+import requests
+
+
+def parse_args():
+    # fmt: off
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--command", type=str, default="",
+        help="the command to run")
+    parser.add_argument("--num-seeds", type=int, default=3,
+        help="the number of random seeds")
+    parser.add_argument("--start-seed", type=int, default=1,
+        help="the number of the starting seed")
+    parser.add_argument("--workers", type=int, default=0,
+        help="the number of workers to run benchmark experimenets")
+    parser.add_argument("--auto-tag", type=lambda x: bool(strtobool(x)), default=True, nargs="?", const=True,
+        help="if toggled, the runs will be tagged with git tags, commit, and pull request number if possible")
+    parser.add_argument("--slurm-template-path", type=str, default=None,
+        help="the path to the slurm template file (see docs for more details)")
+    parser.add_argument("--slurm-gpus-per-task", type=int, default=1,
+        help="the number of gpus per task to use for slurm jobs")
+    parser.add_argument("--slurm-total-cpus", type=int, default=50,
+        help="the number of gpus per task to use for slurm jobs")
+    parser.add_argument("--slurm-ntasks", type=int, default=1,
+        help="the number of tasks to use for slurm jobs")
+    parser.add_argument("--slurm-nodes", type=int, default=None,
+        help="the number of nodes to use for slurm jobs")
+    args = parser.parse_args()
+    # fmt: on
+    return args
+
+
+def run_experiment(command: str):
+    command_list = shlex.split(command)
+    print(f"running {command}")
+
+    # Use subprocess.PIPE to capture the output
+    fd = subprocess.Popen(command_list, stdout=subprocess.PIPE, stderr=subprocess.PIPE)
+    output, errors = fd.communicate()
+
+    return_code = fd.returncode
+    assert return_code == 0, f"Command failed with error: {errors.decode('utf-8')}"
+
+    # Convert bytes to string and strip leading/trailing whitespaces
+    return output.decode("utf-8").strip()
+
+
+def autotag() -> str:
+    wandb_tag = ""
+    print("autotag feature is enabled")
+    git_tag = ""
+    try:
+        git_tag = subprocess.check_output(["git", "describe", "--tags"]).decode("ascii").strip()
+        print(f"identified git tag: {git_tag}")
+    except subprocess.CalledProcessError as e:
+        print(e)
+    if len(git_tag) == 0:
+        try:
+            count = int(subprocess.check_output(["git", "rev-list", "--count", "HEAD"]).decode("ascii").strip())
+            hash = subprocess.check_output(["git", "rev-parse", "--short", "HEAD"]).decode("ascii").strip()
+            git_tag = f"no-tag-{count}-g{hash}"
+            print(f"identified git tag: {git_tag}")
+        except subprocess.CalledProcessError as e:
+            print(e)
+    wandb_tag = f"{git_tag}"
+
+    git_commit = subprocess.check_output(["git", "rev-parse", "--verify", "HEAD"]).decode("ascii").strip()
+    try:
+        # try finding the pull request number on github
+        prs = requests.get(f"https://api.github.com/search/issues?q=repo:huggingface/trl+is:pr+{git_commit}")
+        if prs.status_code == 200:
+            prs = prs.json()
+            if len(prs["items"]) > 0:
+                pr = prs["items"][0]
+                pr_number = pr["number"]
+                wandb_tag += f",pr-{pr_number}"
+        print(f"identified github pull request: {pr_number}")
+    except Exception as e:
+        print(e)
+
+    return wandb_tag
+
+
+if __name__ == "__main__":
+    args = parse_args()
+    if args.auto_tag:
+        existing_wandb_tag = os.environ.get("WANDB_TAGS", "")
+        wandb_tag = autotag()
+        if len(wandb_tag) > 0:
+            if len(existing_wandb_tag) > 0:
+                os.environ["WANDB_TAGS"] = ",".join([existing_wandb_tag, wandb_tag])
+            else:
+                os.environ["WANDB_TAGS"] = wandb_tag
+    print("WANDB_TAGS: ", os.environ.get("WANDB_TAGS", ""))
+    commands = []
+    for seed in range(0, args.num_seeds):
+        commands += [" ".join([args.command, "--seed", str(args.start_seed + seed)])]
+
+    print("======= commands to run:")
+    for command in commands:
+        print(command)
+
+    if args.workers > 0 and args.slurm_template_path is None:
+        from concurrent.futures import ThreadPoolExecutor
+
+        executor = ThreadPoolExecutor(max_workers=args.workers, thread_name_prefix="cleanrl-benchmark-worker-")
+        for command in commands:
+            executor.submit(run_experiment, command)
+        executor.shutdown(wait=True)
+    else:
+        print("not running the experiments because --workers is set to 0; just printing the commands to run")
+
+    # SLURM logic
+    if args.slurm_template_path is not None:
+        if not os.path.exists("slurm"):
+            os.makedirs("slurm")
+        if not os.path.exists("slurm/logs"):
+            os.makedirs("slurm/logs")
+        print("======= slurm commands to run:")
+        with open(args.slurm_template_path) as f:
+            slurm_template = f.read()
+        slurm_template = slurm_template.replace("{{array}}", f"0-{len(commands) - 1}%{args.workers}")
+        slurm_template = slurm_template.replace(
+            "{{seeds}}", f"({' '.join([str(args.start_seed + int(seed)) for seed in range(args.num_seeds)])})"
+        )
+        slurm_template = slurm_template.replace("{{len_seeds}}", f"{args.num_seeds}")
+        slurm_template = slurm_template.replace("{{command}}", args.command)
+        slurm_template = slurm_template.replace("{{gpus_per_task}}", f"{args.slurm_gpus_per_task}")
+        total_gpus = args.slurm_gpus_per_task * args.slurm_ntasks
+        slurm_cpus_per_gpu = math.ceil(args.slurm_total_cpus / total_gpus)
+        slurm_template = slurm_template.replace("{{cpus_per_gpu}}", f"{slurm_cpus_per_gpu}")
+        slurm_template = slurm_template.replace("{{ntasks}}", f"{args.slurm_ntasks}")
+        if args.slurm_nodes is not None:
+            slurm_template = slurm_template.replace("{{nodes}}", f"#SBATCH --nodes={args.slurm_nodes}")
+        else:
+            slurm_template = slurm_template.replace("{{nodes}}", "")
+        filename = str(uuid.uuid4())
+        open(os.path.join("slurm", f"{filename}.slurm"), "w").write(slurm_template)
+        slurm_path = os.path.join("slurm", f"{filename}.slurm")
+        print(f"saving command in {slurm_path}")
+        if args.workers > 0:
+            job_id = run_experiment(f"sbatch --parsable {slurm_path}")
+            print(f"Job ID: {job_id}")

benchmark/benchmark_and_report.sh

@@ -0,0 +1,41 @@
+#### Step 1: create a work directory:
+# this is necessary because another github action job will remove
+# the entire directory, which slurm depends on.
+# https://stackoverflow.com/questions/4632028/how-to-create-a-temporary-directory
+MY_SLURM_TMP_DIR=/fsx/costa/slurm_tmpdir
+mkdir -p $MY_SLURM_TMP_DIR
+WORK_DIR=`mktemp -d -p "$MY_SLURM_TMP_DIR"`
+cp -r "$PWD" "$WORK_DIR"
+cd "$WORK_DIR/$(basename "$PWD")"
+echo WORK_DIR: $WORK_DIR
+
+#### Step 2: actual work starts:
+echo PATH is $PATH
+echo PYTHONPATH is $PYTHONPATH
+echo whcih python is $(which python)
+
+export WANDB_ENTITY=huggingface
+bash $BENCHMARK_SCRIPT > output.txt
+
+# Extract Job IDs into an array
+job_ids=($(grep "Job ID:" output.txt | awk '{print $3}'))
+
+# Extract WANDB_TAGS into an array
+WANDB_TAGS=($(grep "WANDB_TAGS:" output.txt | awk '{print $2}'))
+WANDB_TAGS=($(echo $WANDB_TAGS | tr "," "\n"))
+
+# Print to verify
+echo "Job IDs: ${job_ids[@]}"
+echo "WANDB_TAGS: ${WANDB_TAGS[@]}"
+
+TAGS_STRING="?tag=${WANDB_TAGS[0]}"
+FOLDER_STRING="${WANDB_TAGS[0]}"
+for tag in "${WANDB_TAGS[@]:1}"; do
+    TAGS_STRING+="&tag=$tag"
+    FOLDER_STRING+="_$tag"
+done
+
+echo "TAGS_STRING: $TAGS_STRING"
+echo "FOLDER_STRING: $FOLDER_STRING"
+
+TAGS_STRING=$TAGS_STRING FOLDER_STRING=$FOLDER_STRING BENCHMARK_PLOT_SCRIPT=$BENCHMARK_PLOT_SCRIPT sbatch --dependency=afterany:$job_ids benchmark/post_github_comment.sbatch

benchmark/benchmark_level1.sh

@@ -0,0 +1,11 @@
+# hello world experiment
+python benchmark/benchmark.py \
+    --command "python examples/scripts/ppo.py --ppo_config.log_with wandb" \
+    --num-seeds 3 \
+    --start-seed 1 \
+    --workers 10 \
+    --slurm-nodes 1 \
+    --slurm-gpus-per-task 1 \
+    --slurm-ntasks 1 \
+    --slurm-total-cpus 12 \
+    --slurm-template-path benchmark/trl.slurm_template

benchmark/benchmark_level1_plot.sh

@@ -0,0 +1,20 @@
+# pip install openrlbenchmark==0.2.1a5
+# see https://github.com/openrlbenchmark/openrlbenchmark#get-started for documentation
+echo "we deal with $TAGS_STRING"
+
+python -m openrlbenchmark.rlops_multi_metrics \
+    --filters '?we=huggingface&wpn=trl&xaxis=_step&ceik=trl_ppo_trainer_config.value.reward_model&cen=trl_ppo_trainer_config.value.exp_name&metrics=env/reward_mean&metrics=objective/kl' \
+        "ppo$TAGS_STRING" \
+    --env-ids sentiment-analysis:lvwerra/distilbert-imdb \
+    --no-check-empty-runs \
+    --pc.ncols 2 \
+    --pc.ncols-legend 1 \
+    --output-filename benchmark/trl/$FOLDER_STRING/hello_world \
+    --scan-history
+
+python benchmark/upload_benchmark.py \
+    --folder_path="benchmark/trl/$FOLDER_STRING" \
+    --path_in_repo="images/benchmark/$FOLDER_STRING" \
+    --repo_id="trl-internal-testing/example-images" \
+    --repo_type="dataset"
+

benchmark/benchmark_level2.sh

@@ -0,0 +1,23 @@
+# compound experiments: gpt2xl + grad_accu
+python benchmark/benchmark.py \
+    --command "python examples/scripts/ppo.py --ppo_config.exp_name ppo_gpt2xl_grad_accu --ppo_config.model_name gpt2-xl --ppo_config.mini_batch_size 16 --ppo_config.gradient_accumulation_steps 8 --ppo_config.log_with wandb" \
+    --num-seeds 3 \
+    --start-seed 1 \
+    --workers 10 \
+    --slurm-nodes 1 \
+    --slurm-gpus-per-task 1 \
+    --slurm-ntasks 1 \
+    --slurm-total-cpus 12 \
+    --slurm-template-path benchmark/trl.slurm_template
+
+# compound experiments: Cerebras-GPT-6.7B + deepspeed zero2 + grad_accu
+python benchmark/benchmark.py \
+    --command "accelerate launch --config_file examples/accelerate_configs/deepspeed_zero2.yaml examples/scripts/ppo.py --ppo_config.exp_name ppo_Cerebras-GPT-6.7B_grad_accu_deepspeed_stage2  --ppo_config.batch_size 32  --ppo_config.mini_batch_size 32 --ppo_config.log_with wandb --ppo_config.model_name cerebras/Cerebras-GPT-6.7B --ppo_config.reward_model sentiment-analysis:cerebras/Cerebras-GPT-6.7B" \
+    --num-seeds 3 \
+    --start-seed 1 \
+    --workers 10 \
+    --slurm-nodes 1 \
+    --slurm-gpus-per-task 8 \
+    --slurm-ntasks 1 \
+    --slurm-total-cpus 90 \
+    --slurm-template-path benchmark/trl.slurm_template

benchmark/benchmark_level2_plot.sh

@@ -0,0 +1,31 @@
+# pip install openrlbenchmark==0.2.1a5
+# see https://github.com/openrlbenchmark/openrlbenchmark#get-started for documentation
+echo "we deal with $TAGS_STRING"
+
+python -m openrlbenchmark.rlops_multi_metrics \
+    --filters '?we=huggingface&wpn=trl&xaxis=_step&ceik=trl_ppo_trainer_config.value.reward_model&cen=trl_ppo_trainer_config.value.exp_name&metrics=env/reward_mean&metrics=objective/kl' \
+        "ppo$TAGS_STRING" \
+        "ppo_gpt2xl_grad_accu$TAGS_STRING" \
+    --env-ids sentiment-analysis:lvwerra/distilbert-imdb \
+    --no-check-empty-runs \
+    --pc.ncols 2 \
+    --pc.ncols-legend 1 \
+    --output-filename benchmark/trl/$FOLDER_STRING/different_models \
+    --scan-history
+
+python -m openrlbenchmark.rlops_multi_metrics \
+    --filters '?we=huggingface&wpn=trl&xaxis=_step&ceik=trl_ppo_trainer_config.value.reward_model&cen=trl_ppo_trainer_config.value.exp_name&metrics=env/reward_mean&metrics=objective/kl' \
+        "ppo_Cerebras-GPT-6.7B_grad_accu_deepspeed_stage2$TAGS_STRING" \
+    --env-ids sentiment-analysis:cerebras/Cerebras-GPT-6.7B \
+    --no-check-empty-runs \
+    --pc.ncols 2 \
+    --pc.ncols-legend 1 \
+    --output-filename benchmark/trl/$FOLDER_STRING/deepspeed \
+    --scan-history
+
+python benchmark/upload_benchmark.py \
+    --folder_path="benchmark/trl/$FOLDER_STRING" \
+    --path_in_repo="images/benchmark/$FOLDER_STRING" \
+    --repo_id="trl-internal-testing/example-images" \
+    --repo_type="dataset"
+

benchmark/benchmark_level3.sh

@@ -0,0 +1,46 @@
+## w/ and w/o gradient accumulation
+python benchmark/benchmark.py \
+    --command "python examples/scripts/ppo.py --ppo_config.exp_name ppo_step_grad_accu --ppo_config.mini_batch_size 1 --ppo_config.gradient_accumulation_steps 128 --ppo_config.log_with wandb" \
+    --num-seeds 3 \
+    --start-seed 1 \
+    --workers 10 \
+    --slurm-nodes 1 \
+    --slurm-gpus-per-task 1 \
+    --slurm-ntasks 1 \
+    --slurm-total-cpus 12 \
+    --slurm-template-path benchmark/trl.slurm_template
+
+## w/ different models (gpt2, gpt2-xl, falcon, llama2)
+python benchmark/benchmark.py \
+    --command "python examples/scripts/ppo.py --ppo_config.exp_name ppo_gpt2 --ppo_config.log_with wandb" \
+    --num-seeds 3 \
+    --start-seed 1 \
+    --workers 10 \
+    --slurm-nodes 1 \
+    --slurm-gpus-per-task 1 \
+    --slurm-ntasks 1 \
+    --slurm-total-cpus 12 \
+    --slurm-template-path benchmark/trl.slurm_template
+python benchmark/benchmark.py \
+    --command "python examples/scripts/ppo.py --ppo_config.exp_name ppo_falcon_rw_1b --ppo_config.model_name tiiuae/falcon-rw-1b --ppo_config.log_with wandb" \
+    --num-seeds 3 \
+    --start-seed 1 \
+    --workers 10 \
+    --slurm-nodes 1 \
+    --slurm-gpus-per-task 1 \
+    --slurm-ntasks 1 \
+    --slurm-total-cpus 12 \
+    --slurm-template-path benchmark/trl.slurm_template
+
+
+## w/ and w/o PEFT
+python benchmark/benchmark.py \
+    --command "python examples/scripts/ppo.py --ppo_config.exp_name ppo_peft --use_peft --ppo_config.log_with wandb" \
+    --num-seeds 3 \
+    --start-seed 1 \
+    --workers 10 \
+    --slurm-nodes 1 \
+    --slurm-gpus-per-task 1 \
+    --slurm-ntasks 1 \
+    --slurm-total-cpus 12 \
+    --slurm-template-path benchmark/trl.slurm_template

benchmark/plot.sh

@@ -0,0 +1,56 @@
+# pip install openrlbenchmark==0.2.1a5
+# see https://github.com/openrlbenchmark/openrlbenchmark#get-started for documentation
+BASELINE_PR_TAG=v0.4.7-55-g110e672
+BASELINE_PR_NAME=PR-662
+
+python -m openrlbenchmark.rlops_multi_metrics \
+    --filters '?we=huggingface&wpn=trl&xaxis=_step&ceik=trl_ppo_trainer_config.value.reward_model&cen=trl_ppo_trainer_config.value.exp_name&metrics=env/reward_mean&metrics=objective/kl' \
+        "sentiment_tuning?tag=$BASELINE_PR_TAG&cl=sentiment lvwerra/gpt2-imdb ($BASELINE_PR_NAME)" \
+    --env-ids sentiment-analysis:lvwerra/distilbert-imdb \
+    --no-check-empty-runs \
+    --pc.ncols 2 \
+    --pc.ncols-legend 1 \
+    --output-filename benchmark/trl/$BASELINE_PR_TAG/sentiment \
+    --scan-history
+
+python -m openrlbenchmark.rlops_multi_metrics \
+    --filters '?we=huggingface&wpn=trl&xaxis=_step&ceik=trl_ppo_trainer_config.value.reward_model&cen=trl_ppo_trainer_config.value.exp_name&metrics=env/reward_mean&metrics=objective/kl' \
+        "sentiment_tuning?tag=$BASELINE_PR_TAG&cl=sentiment lvwerra/gpt2-imdb ($BASELINE_PR_NAME)" \
+        "sentiment_tuning_step_grad_accu?tag=$BASELINE_PR_TAG&cl=sentiment lvwerra/gpt2-imdb gradient accumulation ($BASELINE_PR_NAME)" \
+    --env-ids sentiment-analysis:lvwerra/distilbert-imdb \
+    --no-check-empty-runs \
+    --pc.ncols 2 \
+    --pc.ncols-legend 1 \
+    --output-filename benchmark/trl/$BASELINE_PR_TAG/gradient_accu \
+    --scan-history
+
+python -m openrlbenchmark.rlops_multi_metrics \
+    --filters '?we=huggingface&wpn=trl&xaxis=_step&ceik=trl_ppo_trainer_config.value.reward_model&cen=trl_ppo_trainer_config.value.exp_name&metrics=env/reward_mean&metrics=objective/kl' \
+        "sentiment_tuning?tag=$BASELINE_PR_TAG&cl=sentiment lvwerra/gpt2-imdb ($BASELINE_PR_NAME)" \
+        "sentiment_tuning_gpt2?tag=$BASELINE_PR_TAG&cl=sentiment gpt2 ($BASELINE_PR_NAME)" \
+        "sentiment_tuning_falcon_rw_1b?tag=$BASELINE_PR_TAG&cl=sentiment tiiuae/falcon-rw-1b ($BASELINE_PR_NAME)" \
+        "sentiment_tuning_gpt2xl_grad_accu?tag=$BASELINE_PR_TAG&cl=sentiment gpt2xl ($BASELINE_PR_NAME)" \
+    --env-ids sentiment-analysis:lvwerra/distilbert-imdb \
+    --no-check-empty-runs \
+    --pc.ncols 2 \
+    --pc.ncols-legend 1 \
+    --output-filename benchmark/trl/$BASELINE_PR_TAG/different_models \
+    --scan-history
+
+python -m openrlbenchmark.rlops_multi_metrics \
+    --filters '?we=huggingface&wpn=trl&xaxis=_step&ceik=trl_ppo_trainer_config.value.reward_model&cen=trl_ppo_trainer_config.value.exp_name&metrics=env/reward_mean&metrics=objective/kl' \
+        "sentiment_tuning?tag=$BASELINE_PR_TAG&cl=sentiment lvwerra/gpt2-imdb ($BASELINE_PR_NAME)" \
+        "sentiment_tuning_peft?tag=$BASELINE_PR_TAG&cl=sentiment lvwerra/gpt2-imdb w/ peft ($BASELINE_PR_NAME)" \
+    --env-ids sentiment-analysis:lvwerra/distilbert-imdb \
+    --no-check-empty-runs \
+    --pc.ncols 2 \
+    --pc.ncols-legend 1 \
+    --output-filename benchmark/trl/$BASELINE_PR_TAG/peft \
+    --scan-history
+
+
+python benchmark/upload_benchmark.py \
+    --folder_path="benchmark/trl/$BASELINE_PR_TAG" \
+    --path_in_repo="images/benchmark/$BASELINE_PR_TAG" \
+    --repo_id="trl-internal-testing/example-images" \
+    --repo_type="dataset"

benchmark/post_github_comment.py

@@ -0,0 +1,26 @@
+import json
+import os
+
+from ghapi.all import GhApi
+
+
+FOLDER_STRING = os.environ.get("FOLDER_STRING", "")
+folder = f"benchmark/trl/{FOLDER_STRING}"
+host_url = f"https://huggingface.co/datasets/trl-internal-testing/example-images/resolve/main/images/benchmark/{FOLDER_STRING}"
+
+# Create a GitHub API instance
+github_context = json.loads(os.environ["GITHUB_CONTEXT"])
+token = os.environ["PERSONAL_ACCESS_TOKEN_GITHUB"]  # this needs to refreshed every 12 months
+status_message = "**[COSTA BENCHMARK BOT]**: Here are the results"
+body = status_message
+repo = github_context["repository"]
+owner, repo = repo.split("/")
+api = GhApi(owner=owner, repo=repo, token=token)
+
+# for each `.png` file in the folder, add it to the comment
+for file in os.listdir(folder):
+    if file.endswith(".png"):
+        body += f"\n![{file}]({host_url}/{file})"
+
+# Create a comment on the issue
+api.issues.create_comment(issue_number=github_context["event"]["issue"]["number"], body=body)

benchmark/post_github_comment.sbatch

@@ -0,0 +1,9 @@
+#!/bin/bash
+#SBATCH --job-name=trl
+#SBATCH --partition=production-cluster
+#SBATCH --ntasks=1
+#SBATCH --output=slurm/logs/%x_%j.out
+
+sleep 2m
+bash $BENCHMARK_PLOT_SCRIPT
+srun python benchmark/post_github_comment.py

benchmark/trl.slurm_template

@@ -0,0 +1,16 @@
+#!/bin/bash
+#SBATCH --job-name=trl
+#SBATCH --partition=production-cluster
+#SBATCH --gpus-per-task={{gpus_per_task}}
+#SBATCH --cpus-per-gpu={{cpus_per_gpu}}
+#SBATCH --ntasks={{ntasks}}
+#SBATCH --output=slurm/logs/%x_%j.out
+#SBATCH --array={{array}}
+#SBATCH --exclude=ip-26-0-156-239,ip-26-0-148-151,ip-26-0-146-212,ip-26-0-145-137,ip-26-0-146-249,ip-26-0-146-149,ip-26-0-147-233,ip-26-0-145-154,ip-26-0-144-35,ip-26-0-144-189,ip-26-0-146-183,ip-26-0-147-120,ip-26-0-144-95,ip-26-0-145-193
+{{nodes}}
+
+seeds={{seeds}}
+seed=${seeds[$SLURM_ARRAY_TASK_ID % {{len_seeds}}]}
+
+echo "Running task $SLURM_ARRAY_TASK_ID with seed: $seed"
+srun {{command}} --ppo_config.seed $seed

benchmark/upload_benchmark.py

@@ -0,0 +1,23 @@
+from dataclasses import dataclass
+
+import tyro
+from huggingface_hub import HfApi
+
+
+@dataclass
+class Args:
+    folder_path: str = "benchmark/trl"
+    path_in_repo: str = "images/benchmark"
+    repo_id: str = "trl-internal-testing/example-images"
+    repo_type: str = "dataset"
+
+
+args = tyro.cli(Args)
+api = HfApi()
+
+api.upload_folder(
+    folder_path=args.folder_path,
+    path_in_repo=args.path_in_repo,
+    repo_id=args.repo_id,
+    repo_type=args.repo_type,
+)

docs/source/_toctree.yml

@@ -1,24 +1,54 @@
-- sections: 
+- sections:
   - local: index
     title: TRL
   - local: quickstart
     title: Quickstart
   - local: installation
     title: Installation
+  - local: how_to_train
+    title: PPO Training FAQ
+  - local: use_model
+    title: Use Trained Models
   - local: customization
-    title: Customize your training
+    title: Customize the Training
+  - local: logging
+    title: Understanding Logs
   title: Get started
 - sections:
   - local: models
     title: Model Classes
   - local: trainer
     title: Trainer Classes
+  - local: reward_trainer
+    title: Reward Model Training
+  - local: sft_trainer
+    title: Supervised Fine-Tuning
+  - local: ppo_trainer
+    title: PPO Trainer
+  - local: best_of_n
+    title: Best of N Sampling
+  - local: dpo_trainer
+    title: DPO Trainer
+  - local: ddpo_trainer
+    title: Denoising Diffusion Policy Optimization
+  - local: iterative_sft_trainer
+    title: Iterative Supervised Fine-Tuning
+  - local: text_environments
+    title: Text Environments
   title: API
-- sections: 
+- sections:
+  - local: example_overview
+    title: Example Overview
   - local: sentiment_tuning
     title: Sentiment Tuning
-  - local: summarization_reward_tuning
-    title: Summarization Reward Tuning
+  - local: lora_tuning_peft
+    title: Training with PEFT
   - local: detoxifying_a_lm
     title: Detoxifying a Language Model
+  - local: using_llama_models
+    title: Training StackLlama
+  - local: learning_tools
+    title: Learning to Use Tools
+  - local: multi_adapter_rl
+    title: Multi Adapter RLHF
   title: Examples

docs/source/best_of_n.mdx

@@ -0,0 +1,72 @@
+# Best of N sampling: Alternative ways to get better model output without RL based fine-tuning 
+
+Within the extras module is the `best-of-n` sampler class that serves as an alternative method of generating better model output.
+As to how it fares against the RL based fine-tuning, please look in the `examples` directory for a comparison example
+
+## Usage
+
+To get started quickly, instantiate an instance of the class with a model, a length sampler, a tokenizer and a callable that serves as a proxy reward pipeline that outputs reward scores for input queries
+
+```python
+
+from transformers import pipeline, AutoTokenizer
+from trl import AutoModelForCausalLMWithValueHead
+from trl.core import LengthSampler
+from trl.extras import BestOfNSampler
+
+ref_model = AutoModelForCausalLMWithValueHead.from_pretrained(ref_model_name)
+reward_pipe = pipeline("sentiment-analysis", model=reward_model, device=device)
+tokenizer = AutoTokenizer.from_pretrained(ref_model_name)
+tokenizer.pad_token = tokenizer.eos_token
+
+
+# callable that takes a list of raw text and returns a list of corresponding reward scores
+def queries_to_scores(list_of_strings):
+  return [output["score"] for output in reward_pipe(list_of_strings)]
+
+best_of_n = BestOfNSampler(model, tokenizer, queries_to_scores, length_sampler=output_length_sampler)
+
+
+```
+
+And assuming you have a list/tensor of tokenized queries, you can generate better output by calling the `generate` method
+
+```python
+
+best_of_n.generate(query_tensors, device=device, **gen_kwargs)
+
+```
+The default sample size is 4, but you can change it at the time of instance initialization like so
+
+```python
+
+best_of_n = BestOfNSampler(model, tokenizer, queries_to_scores, length_sampler=output_length_sampler, sample_size=8)
+
+```
+
+The default output is the result of taking the top scored output for each query, but you can change it to top 2 and so on by passing the `n_candidates` argument at the time of instance initialization
+
+```python
+
+best_of_n = BestOfNSampler(model, tokenizer, queries_to_scores, length_sampler=output_length_sampler, n_candidates=2)
+
+```
+
+There is the option of setting the generation settings (like `temperature`, `pad_token_id`) at the time of instance creation as opposed to when calling the `generate` method.
+This is done by passing a `GenerationConfig` from the `transformers` library at the time of initialization
+
+```python
+
+from transformers import GenerationConfig
+
+generation_config = GenerationConfig(min_length= -1, top_k=0.0, top_p= 1.0, do_sample= True, pad_token_id=tokenizer.eos_token_id)
+
+best_of_n = BestOfNSampler(model, tokenizer, queries_to_scores, length_sampler=output_length_sampler, generation_config=generation_config)
+
+best_of_n.generate(query_tensors, device=device)
+
+```
+
+Furthermore, at the time of initialization you can set the seed to control repeatability of the generation process and the number of samples to generate for each query
+
+

docs/source/customization.mdx

@@ -1,6 +1,50 @@
 # Training customization
 
-At `trl` we provide the possibility to give enough modularity to users to be able to efficiently customize the training loop for their needs. Below are some examples on how you can apply and test different techniques.
+TRL is designed with modularity in mind so that users to be able to efficiently customize the training loop for their needs. Below are some examples on how you can apply and test different techniques.
+
+## Train on multiple GPUs / nodes
+
+The trainers in TRL use 🤗 Accelerate to enable distributed training across multiple GPUs or nodes. To do so, first create an 🤗 Accelerate config file by running
+
+```bash
+accelerate config
+```
+
+and answering the questions according to your multi-gpu / multi-node setup. You can then launch distributed training by running:
+
+```bash
+accelerate launch your_script.py
+```
+
+We also provide config files in the [examples folder](https://github.com/huggingface/trl/tree/main/examples/accelerate_configs) that can be used as templates. To use these templates, simply pass the path to the config file when launching a job, e.g.:
+
+```shell
+accelerate launch --config_file=examples/accelerate_configs/multi_gpu.yaml --num_processes {NUM_GPUS} path_to_script.py --all_arguments_of_the_script
+```
+
+Refer to the [examples page](https://github.com/huggingface/trl/tree/main/examples) for more details.
+
+### Distributed training with DeepSpeed
+
+All of the trainers in TRL can be run on multiple GPUs together with DeepSpeed ZeRO-{1,2,3} for efficient sharding of the optimizer states, gradients, and model weights. To do so, run:
+
+```shell
+accelerate launch --config_file=examples/accelerate_configs/deepspeed_zero{1,2,3}.yaml --num_processes {NUM_GPUS} path_to_your_script.py --all_arguments_of_the_script
+```
+
+Note that for ZeRO-3, a small tweak is needed to initialize your reward model on the correct device via the `zero3_init_context_manager()` context manager. In particular, this is needed to avoid DeepSpeed hanging after a fixed number of training steps. Here is a snippet of what is involved from the [`sentiment_tuning`](https://github.com/huggingface/trl/blob/main/examples/scripts/ppo.py) example:
+
+```python
+ds_plugin = ppo_trainer.accelerator.state.deepspeed_plugin
+if ds_plugin is not None and ds_plugin.is_zero3_init_enabled():
+    with ds_plugin.zero3_init_context_manager(enable=False):
+        sentiment_pipe = pipeline("sentiment-analysis", model="lvwerra/distilbert-imdb", device=device)
+else:
+    sentiment_pipe = pipeline("sentiment-analysis", model="lvwerra/distilbert-imdb", device=device)
+```
+
+Consult the 🤗 Accelerate [documentation](https://huggingface.co/docs/accelerate/usage_guides/deepspeed) for more information about the DeepSpeed plugin.
+
 
 ## Use different optimizers
 
@@ -63,7 +107,7 @@ optimizer = Lion(filter(lambda p: p.requires_grad, self.model.parameters()), lr=
 ...
 ppo_trainer = PPOTrainer(config, model, model_ref, tokenizer, optimizer=optimizer)
 ```
-We advice you to use the learning rate that you would use for `Adam` divided by 3 as pointed out [here](https://github.com/lucidrains/lion-pytorch#lion---pytorch). We observed an improvement when using this optimizer compared to classic Adam (check the full logs [here](https://wandb.ai/distill-bloom/trl/runs/lj4bheke?workspace=user-younesbelkada)):
+We advise you to use the learning rate that you would use for `Adam` divided by 3 as pointed out [here](https://github.com/lucidrains/lion-pytorch#lion---pytorch). We observed an improvement when using this optimizer compared to classic Adam (check the full logs [here](https://wandb.ai/distill-bloom/trl/runs/lj4bheke?workspace=user-younesbelkada)):
 
 <div style="text-align: center">
 <img src="https://huggingface.co/datasets/trl-internal-testing/example-images/resolve/main/images/trl-lion.png">
@@ -90,7 +134,7 @@ config = PPOConfig(**ppo_config)
 
 # 2. Create optimizer
 optimizer = torch.optim.SGD(model.parameters(), lr=config.learning_rate)
-lr_scheduler = lr_scheduler = torch.optim.lr_scheduler.ExponentialLR(optimizer, gamma=0.9)
+lr_scheduler = torch.optim.lr_scheduler.ExponentialLR(optimizer, gamma=0.9)
 
 # 3. initialize trainer
 ppo_trainer = PPOTrainer(config, model, model_ref, tokenizer, optimizer=optimizer, lr_scheduler=lr_scheduler)
@@ -142,3 +186,31 @@ ppo_config = {'batch_size': 1}
 config = PPOConfig(**ppo_config)
 ppo_trainer = PPOTrainer(config, model, model_ref, tokenizer)
 ```
+
+## Use the CUDA cache optimizer
+
+When training large models, you should better handle the CUDA cache by iteratively clearing it. Do do so, simply pass `optimize_cuda_cache=True` to `PPOConfig`:
+
+```python
+config = PPOConfig(..., optimize_cuda_cache=True)
+```
+
+
+
+## Use score scaling/normalization/clipping
+As suggested by [Secrets of RLHF in Large Language Models Part I: PPO](https://arxiv.org/abs/2307.04964), we support score (aka reward) scaling/normalization/clipping to improve training stability via `PPOConfig`:
+```python
+from trl import PPOConfig
+
+ppo_config = {
+    use_score_scaling=True,
+    use_score_norm=True,
+    score_clip=0.5,
+}
+config = PPOConfig(**ppo_config)
+```
+
+To run `ppo.py`, you can use the following command:
+```
+python examples/scripts/ppo.py --log_with wandb --use_score_scaling --use_score_norm --score_clip 0.5
+```

docs/source/ddpo_trainer.mdx

@@ -0,0 +1,119 @@
+# Denoising Diffusion Policy Optimization
+## The why
+
+| Before | After DDPO finetuning |
+| --- | --- |
+| <div style="text-align: center"><img src="https://huggingface.co/datasets/trl-internal-testing/example-images/resolve/main/images/pre_squirrel.png"/></div> |  <div style="text-align: center"><img src="https://huggingface.co/datasets/trl-internal-testing/example-images/resolve/main/images/post_squirrel.png"/></div> |
+| <div style="text-align: center"><img src="https://huggingface.co/datasets/trl-internal-testing/example-images/resolve/main/images/pre_crab.png"/></div> |  <div style="text-align: center"><img src="https://huggingface.co/datasets/trl-internal-testing/example-images/resolve/main/images/post_crab.png"/></div> |
+|  <div style="text-align: center"><img src="https://huggingface.co/datasets/trl-internal-testing/example-images/resolve/main/images/pre_starfish.png"/></div> |  <div style="text-align: center"><img src="https://huggingface.co/datasets/trl-internal-testing/example-images/resolve/main/images/post_starfish.png"/></div> |
+
+
+## Getting started with Stable Diffusion finetuning with reinforcement learning
+
+The machinery for finetuning of Stable Diffusion models with reinforcement learning makes heavy use of HuggingFace's `diffusers`
+library. A reason for  stating this is that getting started requires a bit of familiarity with the `diffusers` library concepts, mainly two of them - pipelines and schedulers.
+Right out of the box (`diffusers` library), there isn't a `Pipeline` nor a `Scheduler` instance that is suitable for finetuning with reinforcement learning. Some adjustments need to made. 
+
+There is a pipeline interface that is provided by this library that is required to be implemented to be used with the `DDPOTrainer`, which is the main machinery for fine-tuning Stable Diffusion with reinforcement learning. **Note: Only the StableDiffusion architecture is supported at this point.**
+There is a default implementation of this interface that you can use out of the box. Assuming the default implementation is sufficient and/or to get things moving, refer to the training example alongside this guide. 
+
+The point of the interface is to fuse the pipeline and the scheduler into one object which allows for minimalness in terms of having the constraints all in one place. The interface was designed in hopes of catering to pipelines and schedulers beyond the examples in this repository and elsewhere at this time of writing. Also the scheduler step is a method of this pipeline interface and this may seem redundant given that the raw scheduler is accessible via the interface but this is the only way to constrain the scheduler step output to an output type befitting of the algorithm at hand (DDPO).
+
+For a more detailed look into the interface and the associated default implementation, go [here](https://github.com/lvwerra/trl/tree/main/trl/models/modeling_sd_base.py)
+
+Note that the default implementation has a LoRA implementation path and a non-LoRA based implementation path. The LoRA flag enabled by default and this can be turned off by passing in the flag to do so. LORA based training is faster and the LORA associated model hyperparameters responsible for model convergence aren't as finicky as non-LORA based training.
+
+Also in addition, there is the expectation of providing a reward function and a prompt function. The reward function is used to evaluate the generated images  and the prompt function is used to generate the prompts that are used to generate the images.
+
+## Getting started with `examples/scripts/ddpo.py`
+
+The `ddpo.py` script is a working example of using the `DDPO` trainer to finetune a Stable Diffusion model. This example explicitly configures a small subset of the overall parameters associated with the config object (`DDPOConfig`).
+
+**Note:** one A100 GPU is recommended to get this running. Anything below a A100 will not be able to run this example script and even if it does via relatively smaller sized parameters, the results will most likely be poor.
+
+Almost every configuration parameter has a default. There is only one commandline flag argument that is required of the user to get things up and running. The user is expected to have a [huggingface user access token](https://huggingface.co/docs/hub/security-tokens) that will be used to upload the model post finetuning to HuggingFace hub. The following bash command is to be entered to get things running
+
+```batch
+python ddpo.py --hf_user_access_token <token>
+```
+
+To obtain the documentation of `stable_diffusion_tuning.py`, please run `python stable_diffusion_tuning.py --help`
+
+The following are things to keep in mind (The code checks this for you as well) in general while configuring the trainer (beyond the use case of using the example script)
+
+- The configurable sample batch size (`--ddpo_config.sample_batch_size=6`) should be greater than or equal to the configurable training batch size (`--ddpo_config.train_batch_size=3`)
+- The configurable sample batch size (`--ddpo_config.sample_batch_size=6`) must be divisible by the configurable train batch size (`--ddpo_config.train_batch_size=3`)
+- The configurable sample batch size (`--ddpo_config.sample_batch_size=6`) must be divisible by both the configurable gradient accumulation steps (`--ddpo_config.train_gradient_accumulation_steps=1`) and the configurable accelerator processes count 
+
+## Setting up the image logging hook function
+
+Expect the function to be given a list of lists of the form
+```python
+[[image, prompt, prompt_metadata, rewards, reward_metadata], ...]
+
+```
+and `image`, `prompt`, `prompt_metadata`, `rewards`, `reward_metadata` are batched.
+The last list in the lists of lists represents the last sample batch. You are likely to want to log this one
+While you are free to log however you want the use of `wandb` or `tensorboard` is recommended.
+
+### Key terms
+
+- `rewards` : The rewards/score is a numerical associated with the generated image and is key to steering the RL process
+- `reward_metadata` : The reward metadata is the metadata associated with the reward. Think of this as extra information payload delivered alongside the reward
+- `prompt` : The prompt is the text that is used to generate the image
+- `prompt_metadata` : The prompt metadata is the metadata associated with the prompt. A situation where this will not be empty is when the reward model comprises of a [`FLAVA`](https://huggingface.co/docs/transformers/model_doc/flava) setup where questions and ground answers (linked to the generated image) are expected with the generated image (See here: https://github.com/kvablack/ddpo-pytorch/blob/main/ddpo_pytorch/rewards.py#L45)
+- `image` : The image generated by the Stable Diffusion model
+
+Example code for logging sampled images with `wandb` is given below.
+
+```python
+# for logging these images to wandb
+
+def image_outputs_hook(image_data, global_step, accelerate_logger):
+    # For the sake of this example, we only care about the last batch
+    # hence we extract the last element of the list
+    result = {}
+    images, prompts, _, rewards, _ = image_data[-1]
+    for i, image in enumerate(images):
+        pil = Image.fromarray(
+            (image.cpu().numpy().transpose(1, 2, 0) * 255).astype(np.uint8)
+        )
+        pil = pil.resize((256, 256))
+        result[f"{prompts[i]:.25} | {rewards[i]:.2f}"] = [pil]
+    accelerate_logger.log_images(
+        result,
+        step=global_step,
+    )
+
+```
+
+### Using the finetuned model
+
+Assuming you've done with all the epochs and have pushed up your model to the hub, you can use the finetuned model as follows
+
+```python
+
+import torch
+from trl import DefaultDDPOStableDiffusionPipeline
+
+pipeline = DefaultDDPOStableDiffusionPipeline("metric-space/ddpo-finetuned-sd-model")
+
+device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
+
+# memory optimization
+pipeline.vae.to(device, torch.float16)
+pipeline.text_encoder.to(device, torch.float16)
+pipeline.unet.to(device, torch.float16)
+
+prompts = ["squirrel", "crab", "starfish", "whale","sponge", "plankton"]
+results = pipeline(prompts)
+
+for prompt, image in zip(prompts,results.images):
+    image.save(f"{prompt}.png")
+
+```
+
+## Credits
+
+This work is heavily influenced by the repo [here](https://github.com/kvablack/ddpo-pytorch) and the associated paper [Training Diffusion Models
+with Reinforcement Learning by Kevin Black, Michael Janner, Yilan Du, Ilya Kostrikov, Sergey Levine](https://arxiv.org/abs/2305.13301).

docs/source/detoxifying_a_lm.mdx

@@ -1,15 +1,15 @@
 # Detoxifying a Language Model using PPO
 
-Language models (LMs) are known to sometimes generate toxic outputs. In this example, we will show how to "detoxify" a LM by feeding it toxic prompts and then using PPO to "detoxify" it.
+Language models (LMs) are known to sometimes generate toxic outputs. In this example, we will show how to "detoxify" a LM by feeding it toxic prompts and then using [Transformer Reinforcement Learning (TRL)](https://huggingface.co/docs/trl/index) and Proximal Policy Optimization (PPO) to "detoxify" it.
 
 Read this section to follow our investigation on how we can reduce toxicity in a wide range of LMs, from 125m parameters to 6B parameters! 
 
-Here's an overview of the notebooks and scripts in the [trl repository](https://github.com/lvwerra/trl/tree/main/examples/toxicity) as well as the link for the interactive demo:
+Here's an overview of the notebooks and scripts in the [TRL toxicity repository](https://github.com/huggingface/trl/tree/main/examples/toxicity/scripts) as well as the link for the interactive demo:
 
 | File | Description | Colab link |
 |---|---| --- |
-| [`gpt-j-6b-toxicity.py`](https://github.com/lvwerra/trl/blob/main/examples/toxicity/scripts/gpt-j-6b-toxicity.py) | Detoxify `GPT-J-6B` using PPO | x | 
-| [`evaluate-toxicity.py`](https://github.com/lvwerra/trl/blob/main/examples/toxicity/scripts/evaluate-toxicity.py) | Evaluate de-toxified models using `evaluate` | x | 
+| [`gpt-j-6b-toxicity.py`](https://github.com/huggingface/trl/blob/main/examples/research_projects/toxicity/scripts/gpt-j-6b-toxicity.py) | Detoxify `GPT-J-6B` using PPO | x | 
+| [`evaluate-toxicity.py`](https://github.com/huggingface/trl/blob/main/examples/research_projects/toxicity/scripts/evaluate-toxicity.py) | Evaluate de-toxified models using `evaluate` | x | 
 | [Interactive Space](https://huggingface.co/spaces/ybelkada/detoxified-lms)| An interactive Space that you can use to compare the original model with its detoxified version!| x |
 
 ## Context
@@ -24,7 +24,7 @@ One could have also used different techniques to evaluate the toxicity of a mode
 
 ### Selection of models
 
-We selected the following models for our experiments to show that `trl` can be easily scaled to 10B parameters models: 
+We selected the following models for our experiments to show that TRL can be easily scaled to 10B parameters models: 
 
 * [`EleutherAI/gpt-neo-125M`](https://huggingface.co/EleutherAI/gpt-neo-125M) (125 million parameters)
 * [`EleutherAI/gpt-neo-2.7B`](https://huggingface.co/EleutherAI/gpt-neo-2.7B) (2.7 billion parameters)
@@ -174,15 +174,18 @@ Below are few generation examples of `gpt-j-6b-detox` model:
 <img src="https://huggingface.co/datasets/trl-internal-testing/example-images/resolve/main/images/trl-toxicity-examples.png">
 </div>
 
-The evaluation script can be found [here](https://github.com/lvwerra/trl/blob/main/examples/toxicity/scripts/evaluate-toxicity.py).
+The evaluation script can be found [here](https://github.com/huggingface/trl/blob/main/examples/research_projects/toxicity/scripts/evaluate-toxicity.py).
 
 ### Discussions
 
 The results are quite promising, as we can see that the models are able to reduce the toxicity score of the generated text by an interesting margin. The gap is clear for `gpt-neo-2B` model but we less so for the `gpt-j-6B` model. There are several things we could try to improve the results on the largest model starting with training with larger `mini_batch_size` and probably allowing to back-propagate through more layers (i.e. use less shared layers).
-We also think we could have trained the models using a "more toxic" dataset as the one we used is much cleaner than the dataset we used for testing our models (from our observation). 
 
 To sum up, in addition to human feedback this could be a useful additional signal when training large language models to ensure there outputs are less toxic as well as useful.
 
+### Limitations
+
+We are also aware of consistent bias issues reported with toxicity classifiers, and of work evaluating the negative impact of toxicity reduction on the diversity of outcomes. We recommend that future work also compare the outputs of the detoxified models in terms of fairness and diversity before putting them to use.
+
 ## What is next?
 
-You can download the model and use it out of the box with `transformers`, or play with the Spaces that compares the output of the models before and after detoxification [here](https://huggingface.co/spaces/ybelkada/detoxified-lms).
+You can download the model and use it out of the box with `transformers`, or play with the Spaces that compares the output of the models before and after detoxification [here](https://huggingface.co/spaces/ybelkada/detoxified-lms).

docs/source/dpo_trainer.mdx

@@ -0,0 +1,106 @@
+# DPO Trainer
+
+TRL supports the DPO Trainer for training language models from preference data, as described in the paper [Direct Preference Optimization: Your Language Model is Secretly a Reward Model](https://arxiv.org/abs/2305.18290) by Rafailov et al., 2023. For a full example have a look at  [`examples/scripts/dpo.py`](https://github.com/huggingface/trl/blob/main/examples/scripts/dpo.py).
+
+
+The first step as always is to train your SFT model, to ensure the data we train on is in-distribution for the DPO algorithm.
+
+## Expected dataset format
+
+The DPO trainer expects a very specific format for the dataset. Since the model will be trained to directly optimize the preference of which sentence is the most relevant, given two sentences. We provide an example from the [`Anthropic/hh-rlhf`](https://huggingface.co/datasets/Anthropic/hh-rlhf) dataset below:
+
+<div style="text-align: center">
+<img src="https://huggingface.co/datasets/trl-internal-testing/example-images/resolve/main/images/rlhf-antropic-example.png", width="50%">
+</div>
+
+Therefore the final dataset object should contain these 3 entries if you use the default `DPODataCollatorWithPadding` data collator. The entries should be named:
+
+- `prompt`
+- `chosen`
+- `rejected`
+
+for example:
+
+```py
+dpo_dataset_dict = {
+    "prompt": [
+        "hello",
+        "how are you",
+        "What is your name?",
+        "What is your name?",
+        "Which is the best programming language?",
+        "Which is the best programming language?",
+        "Which is the best programming language?",
+    ],
+    "chosen": [
+        "hi nice to meet you",
+        "I am fine",
+        "My name is Mary",
+        "My name is Mary",
+        "Python",
+        "Python",
+        "Java",
+    ],
+    "rejected": [
+        "leave me alone",
+        "I am not fine",
+        "Whats it to you?",
+        "I dont have a name",
+        "Javascript",
+        "C++",
+        "C++",
+    ],
+}
+```
+
+where the `prompt` contains the context inputs, `chosen` contains the corresponding chosen responses and `rejected` contains the corresponding negative (rejected) responses. As can be seen a prompt can have multiple responses and this is reflected in the entries being repeated in the dictionary's value arrays.
+
+## Expected model format
+The DPO trainer expects a model of `AutoModelForCausalLM`, compared to PPO that expects `AutoModelForCausalLMWithValueHead` for the value function.
+
+## Using the `DPOTrainer`
+
+For a detailed example have a look at the `examples/scripts/dpo.py` script. At a high level we need to initialize the `DPOTrainer` with a `model` we wish to train, a reference `ref_model` which we will use to calculate the implicit rewards of the preferred and rejected response, the `beta` refers to the hyperparameter of the implicit reward, and the dataset contains the 3 entries listed above. Note that the `model` and `ref_model` need to have the same architecture (ie decoder only or encoder-decoder).
+
+```py
+ dpo_trainer = DPOTrainer(
+    model,
+    model_ref,
+    args=training_args,
+    beta=0.1,
+    train_dataset=train_dataset,
+    tokenizer=tokenizer,
+)
+```
+After this one can then call:
+
+```py
+dpo_trainer.train()
+```
+
+Note that the `beta` is the temperature parameter for the DPO loss, typically something in the range of `0.1` to `0.5`. We ignore the reference model as `beta` -> 0.
+
+## Loss functions
+
+Given the preference data, we can fit a binary classifier according to the Bradley-Terry model and in fact the DPO authors propose the sigmoid loss on the normalized likelihood via the `logsigmoid` to fit a logistic regression.
+
+The [RSO](https://arxiv.org/abs/2309.06657) authors propose to use a hinge loss on the normalized likelihood from the [SLiC](https://arxiv.org/abs/2305.10425) paper. The `DPOTrainer` can be switched to this loss via the `loss_type="hinge"` argument and the `beta` in this case is the reciprocal of the margin.
+
+The [IPO](https://arxiv.org/abs/2310.12036) authors provide a deeper theoretical understanding of the DPO algorithms and identify an issue with overfitting and propose an alternative loss which can be used via the `loss_type="ipo"` argument to the trainer.
+
+The [cDPO](https://ericmitchell.ai/cdpo.pdf) is a tweak on the DPO loss where we assume that the preference labels are noisy with some probability that can be passed to the `DPOTrainer` via `label_smoothing` argument (between 0 and 0.5) and then a conservative DPO loss is used. Use the `loss_type="cdpo"` argument to the trainer to use it.
+
+The [KTO](https://github.com/ContextualAI/HALOs/blob/main/assets/report.pdf) loss is derived to directly maximize the utility of LLM generations instead of the log-likelihood of preferences. Thus the dataset are not necessarily preferences but rather desirable vs undesirable completions. For paired preference data as required by the `DPOTrainer`, use the `loss_type="kto_pair"` argument to the trainer to utilize this loss, while for the more general case of desired and undesirable data, use the as of yet unimplemented `KTOTrainer`.
+
+## Logging
+
+While training and evaluating we record the following reward metrics:
+
+* `rewards/chosen`: the mean difference between the log probabilities of the policy model and the reference model for the chosen responses scaled by beta
+* `rewards/rejected`: the mean difference between the log probabilities of the policy model and the reference model for the rejected responses scaled by beta
+* `rewards/accuracies`: mean of how often the chosen rewards are > than the corresponding rejected rewards
+* `rewards/margins`: the mean difference between the chosen and corresponding rejected rewards
+
+## DPOTrainer
+
+[[autodoc]] DPOTrainer

docs/source/example_overview.md

@@ -0,0 +1,73 @@
+# Examples
+
+
+## Introduction
+
+The examples should work in any of the following settings (with the same script):
+   - single GPU
+   - multi GPUS (using PyTorch distributed mode)
+   - multi GPUS (using DeepSpeed ZeRO-Offload stages 1, 2, & 3)
+   - fp16 (mixed-precision), fp32 (normal precision), or bf16 (bfloat16 precision)
+
+To run it in each of these various modes, first initialize the accelerate
+configuration with `accelerate config`
+
+**NOTE to train with a 4-bit or 8-bit model**, please run
+
+```bash
+pip install --upgrade trl[quantization]
+```
+
+
+## Accelerate Config
+For all the examples, you'll need to generate a 🤗 Accelerate config file with:
+
+```shell
+accelerate config # will prompt you to define the training configuration
+```
+
+Then, it is encouraged to launch jobs with `accelerate launch`!
+
+
+# Maintained Examples
+
+
+| File                                                                                           | Description                                                                                                              |
+|------------------------------------------------------------------------------------------------------------|--------------------------------------------------------------------------------------------------------------------------|
+| [`examples/scripts/sft.py`](https://github.com/huggingface/trl/blob/main/examples/scripts/sft.py) | This script shows how to use the `SFTTrainer` to fine tune a model or adapters into a target dataset.                     |
+| [`examples/scripts/reward_modeling.py`](https://github.com/huggingface/trl/blob/main/examples/scripts/reward_modeling.py) | This script shows how to use the `RewardTrainer` to train a reward model on your own dataset.                            |
+| [`examples/scripts/ppo.py`](https://github.com/huggingface/trl/blob/main/examples/scripts/ppo.py) | This script shows how to use the `PPOTrainer` to fine-tune a sentiment analysis model using IMDB dataset                 |
+| [`examples/scripts/ppo_multi_adapter.py`](https://github.com/huggingface/trl/blob/main/examples/scripts/ppo_multi_adapter.py) | This script shows how to use the `PPOTrainer` to train a single base model with multiple adapters. Requires you to run the example script with the reward model training beforehand. |
+| [`examples/scripts/stable_diffusion_tuning_example.py`](https://github.com/huggingface/trl/blob/main/examples/scripts/stable_diffusion_tuning_example.py) | This script shows to use DDPOTrainer to fine-tune a stable diffusion model using reinforcement learning.                 |
+
+Here are also some easier-to-run colab notebooks that you can use to get started with TRL:
+
+
+| File                                                                                           | Description                                                                                                              |
+|----------------------------------------------------------------------------------------------------------------------------|--------------------------------------------------------------------------------------------------------------------------|
+| [`examples/notebooks/best_of_n.ipynb`](https://github.com/huggingface/trl/tree/main/examples/notebooks/best_of_n.ipynb)                       | This notebook demonstrates how to use the "Best of N" sampling strategy using TRL when fine-tuning your model with PPO.  |
+| [`examples/notebooks/gpt2-sentiment.ipynb`](https://github.com/huggingface/trl/tree/main/examples/notebooks/gpt2-sentiment.ipynb)              | This notebook demonstrates how to reproduce the GPT2 imdb sentiment tuning example on a jupyter notebook.                |
+| [`examples/notebooks/gpt2-control.ipynb`](https://github.com/huggingface/trl/tree/main/examples/notebooks/gpt2-control.ipynb)                  | This notebook demonstrates how to reproduce the GPT2 sentiment control example on a jupyter notebook.                    |
+
+
+We also have some other examples that are less maintained but can be used as a reference:
+1. **[research_projects](https://github.com/huggingface/trl/tree/main/examples/research_projects)**: Check out this folder to find the scripts used for some research projects that used TRL (LM de-toxification, Stack-Llama, etc.)
+
+
+## Distributed training
+
+All of the scripts can be run on multiple GPUs by providing the path of an 🤗 Accelerate config file when calling `accelerate launch`. To launch one of them on one or multiple GPUs, run the following command (swapping `{NUM_GPUS}` with the number of GPUs in your machine and `--all_arguments_of_the_script` with your arguments.)
+
+```shell
+accelerate launch --config_file=examples/accelerate_configs/multi_gpu.yaml --num_processes {NUM_GPUS} path_to_script.py --all_arguments_of_the_script
+```
+
+You can also adjust the parameters of the 🤗 Accelerate config file to suit your needs (e.g. training in mixed precision).
+
+### Distributed training with DeepSpeed
+
+Most of the scripts can be run on multiple GPUs together with DeepSpeed ZeRO-{1,2,3} for efficient sharding of the optimizer states, gradients, and model weights. To do so, run following command (swapping `{NUM_GPUS}` with the number of GPUs in your machine, `--all_arguments_of_the_script` with your arguments, and `--deepspeed_config` with the path to the DeepSpeed config file such as `examples/deepspeed_configs/deepspeed_zero1.yaml`):
+
+```shell
+accelerate launch --config_file=examples/accelerate_configs/deepspeed_zero{1,2,3}.yaml --num_processes {NUM_GPUS} path_to_script.py --all_arguments_of_the_script
+```

docs/source/how_to_train.md

@@ -0,0 +1,66 @@
+# Training FAQ
+
+## What Metrics Should I Look at?
+
+When performing classical supervised fine-tuning of language models, the loss (especially the validation loss) serves as a good indicator of the training progress. However, in Reinforcement Learning (RL), the loss becomes less informative about the model's performance, and its value may fluctuate while the actual performance improves.
+
+To address this, we recommend focusing on two key metrics first:
+
+**Mean Reward**: The primary goal is to maximize the reward achieved by the model during RL training.
+**Objective KL Divergence**: KL divergence (Kullback-Leibler divergence) measures the dissimilarity between two probability distributions. In the context of RL training, we use it to quantify the difference between the current model and a reference model. Ideally, we want to keep the KL divergence between 0 and 10 to ensure the model's generated text remains close to what the reference model produces.
+
+However, there are more metrics that can be useful for debugging, checkout the [logging section](logging).
+
+## Why Do We Use a Reference Model, and What's the Purpose of KL Divergence?
+
+When training RL models, optimizing solely for reward may lead to unexpected behaviors, where the model exploits the environment in ways that don't align with good language generation. In the case of RLHF, we use a reward model trained to predict whether a generated text is highly ranked by humans.
+
+However, the RL model being optimized against the reward model may learn patterns that yield high reward but do not represent good language. This can result in extreme cases where the model generates texts with excessive exclamation marks or emojis to maximize the reward. In some worst-case scenarios, the model may generate patterns completely unrelated to natural language yet receive high rewards, similar to adversarial attacks.
+
+<div style="text-align: center">
+<img src="https://huggingface.co/datasets/trl-internal-testing/example-images/resolve/main/images/kl-example.png">
+<p style="text-align: center;"> <b>Figure:</b> Samples without a KL penalty from <a href="https://arxiv.org/pdf/1909.08593.pdf">https://arxiv.org/pdf/1909.08593.pdf</a>. </p>
+</div>
+
+To address this issue, we add a penalty to the reward function based on the KL divergence between the current model and the reference model. By doing this, we encourage the model to stay close to what the reference model generates.
+
+## What Is the Concern with Negative KL Divergence?
+
+If you generate text by purely sampling from the model distribution things work fine in general. But when you use the `generate` method there are a few caveats because it does not always purely sample depending on the settings which can cause KL-divergence to go negative. Essentially when the active model achieves `log_p_token_active < log_p_token_ref` we get negative KL-div. This can happen in a several cases:
+
+- **top-k sampling**: the model can smooth out the probability distribution causing the top-k tokens having a smaller probability than those of the reference model but they still are selected
+- **min_length**: this ignores the EOS token until `min_length` is reached. thus the model can assign a very low log prob to the EOS token and very high probs to all others until min_length is reached
+- **min_length**: this ignores the EOS token until `min_length` is reached, thus the model can assign a very low log prob to the EOS token and very high probs to all others until min_length is reached
+
+These are just a few examples. Why is negative KL an issue? The total reward `R` is computed `R = r - beta * KL` so if the model can learn how to drive KL-divergence negative it effectively gets a positive reward. In many cases it can be much easier to exploit such a bug in the generation than actually learning the reward function. In addition the KL can become arbitrarily small thus the actual reward can be very small compared to it.
+
+So how should you generate text for PPO training? Let's have a look!
+
+## How to generate text for training?
+
+In order to avoid the KL issues described above we recommend to use the following settings:
+
+```python
+generation_kwargs = {
+    "min_length": -1, # don't ignore the EOS token (see above)
+    "top_k": 0.0, # no top-k sampling
+    "top_p": 1.0, # no nucleus sampling
+    "do_sample": True, # yes, we want to sample
+    "pad_token_id": tokenizer.eos_token_id, # most decoder models don't have a padding token - use EOS token instead
+    "max_new_tokens": 32, # specify how many tokens you want to generate at most
+}
+```
+
+With these settings we usually don't encounter any issues. You can also experiments with other settings but if you encounter issues with negative KL-divergence try to go back to these and see if they persist.
+
+## How can debug your own use-case?
+
+Debugging the RL pipeline can be challenging due to its complexity. Here are some tips and suggestions to make the process easier:
+
+- **Start from a working example**: Begin with a working example from the trl repository and gradually modify it to fit your specific use-case. Changing everything at once can make it difficult to identify the source of potential issues. For example, you can start by replacing the model in the example and once you figure out the best hyperparameters try to switch to your dataset and reward model. If you change everything at once you won't know where a potential problem comes from.
+- **Start small, scale later**: Training large models can be very slow and take several hours or days until you see any improvement. For debugging this is not a convenient timescale so try to use small model variants during the development phase and scale up once that works. That being said you sometimes have to be careful as small models might not have the capacity to solve a complicated task either.
+- **Start simple**: Try to start with a minimal example and build complexity from there. Your use-case might require for example a complicated reward function consisting of many different rewards - try to use one signal first and see if you can optimize that and then add more complexity after that.
+- **Inspect the generations**: It's always a good idea to inspect what the model is generating. Maybe there is a big in your post-processing or your prompt. Due to bad settings you might cut-off generations too soon. These things are very hard to see on the metrics but very obvious if you look at the generations.
+- **Inspect the reward model**: If you reward is not improving over time maybe there's an issue with the reward model. You can look at extreme cases to see if it does what it should: e.g. in the sentiment case you can check if simple positive and negative examples really get different rewards. And you can look at the distribution of your dataset. Finally, maybe the reward is dominated by the query which the model can't affect so you might need to normalize this (e.g. reward of query+response minus reward of the query).
+
+These are just a few tips that we find helpful - if you have more useful tricks feel free to open a PR to add them as well!

docs/source/index.mdx

@@ -4,6 +4,58 @@
 
 # TRL - Transformer Reinforcement Learning
 
-With the TRL (Transformer Reinforcement Learning) libray you can train transformer language models with reinforcement learning. The library is integrated with 🤗 [transformers](https://github.com/huggingface/transformers).
+TRL is a full stack library where we provide a set of tools to train transformer language models with Reinforcement Learning, from the Supervised Fine-tuning step (SFT), Reward Modeling step (RM) to the Proximal Policy Optimization (PPO) step. 
+The library is integrated with 🤗 [transformers](https://github.com/huggingface/transformers).
 
-TRL supports decoder models such as GPT-2, BLOOM, GPT-Neo which can all be optimized using Proximal Policy Optimization (PPO). You can find installation instructions in the [installation guide](installation) and an introdcution to the library in the [Quickstart section](quickstart). There is also a more [in-depth example](sentiment_tuning) to tune GPT-2 to procude positive movie reviews.
+<div style="text-align: center">
+<img src="https://huggingface.co/datasets/trl-internal-testing/example-images/resolve/main/images/TRL-readme.png">
+</div>
+
+Check the appropriate sections of the documentation depending on your needs:
+
+## API documentation
+
+- [Model Classes](models): *A brief overview of what each public model class does.*
+- [`SFTTrainer`](sft_trainer): *Supervise Fine-tune your model easily with `SFTTrainer`*
+- [`RewardTrainer`](reward_trainer): *Train easily your reward model using `RewardTrainer`.*
+- [`PPOTrainer`](ppo_trainer): *Further fine-tune the supervised fine-tuned model using PPO algorithm*
+- [Best-of-N Sampling](best-of-n): *Use best of n sampling as an alternative way to sample predictions from your active model*
+- [`DPOTrainer`](dpo_trainer): *Direct Preference Optimization training using `DPOTrainer`.*
+- [`TextEnvironment`](text_environment): *Text environment to train your model using tools with RL.*
+
+## Examples
+
+- [Sentiment Tuning](sentiment_tuning): *Fine tune your model to generate positive movie contents*
+- [Training with PEFT](lora_tuning_peft): *Memory efficient RLHF training using adapters with PEFT*
+- [Detoxifying LLMs](detoxifying_a_lm): *Detoxify your language model through RLHF*
+- [StackLlama](using_llama_models): *End-to-end RLHF training of a Llama model on Stack exchange dataset*
+- [Learning with Tools](learning_tools): *Walkthrough of using `TextEnvironments`*
+- [Multi-Adapter Training](multi_adapter_rl): *Use a single base model and multiple adapters for memory efficient end-to-end training*
+
+
+## Blog posts
+
+<div class="mt-10">
+  <div class="w-full flex flex-col space-y-4 md:space-y-0 md:grid md:grid-cols-2 md:gap-y-4 md:gap-x-5">
+    <a class="!no-underline border dark:border-gray-700 p-5 rounded-lg shadow hover:shadow-lg" href="https://huggingface.co/blog/rlhf">
+      <img src="https://raw.githubusercontent.com/huggingface/blog/main/assets/120_rlhf/thumbnail.png" alt="thumbnail">
+      <p class="text-gray-700">Illustrating Reinforcement Learning from Human Feedback</p>
+    </a>
+    <a class="!no-underline border dark:border-gray-700 p-5 rounded-lg shadow hover:shadow-lg" href="https://huggingface.co/blog/trl-peft">
+      <img src="https://github.com/huggingface/blog/blob/main/assets/133_trl_peft/thumbnail.png?raw=true" alt="thumbnail">
+      <p class="text-gray-700">Fine-tuning 20B LLMs with RLHF on a 24GB consumer GPU</p>
+    </a>
+    <a class="!no-underline border dark:border-gray-700 p-5 rounded-lg shadow hover:shadow-lg" href="https://huggingface.co/blog/stackllama">
+      <img src="https://github.com/huggingface/blog/blob/main/assets/138_stackllama/thumbnail.png?raw=true" alt="thumbnail">
+      <p class="text-gray-700">StackLLaMA: A hands-on guide to train LLaMA with RLHF</p>
+   </a>
+    <a class="!no-underline border dark:border-gray-700 p-5 rounded-lg shadow hover:shadow-lg" href="https://huggingface.co/blog/dpo-trl">
+      <img src="https://github.com/huggingface/blog/blob/main/assets/157_dpo_trl/dpo_thumbnail.png?raw=true" alt="thumbnail">
+      <p class="text-gray-700">Fine-tune Llama 2 with DPO</p>
+    </a>
+    <a class="!no-underline border dark:border-gray-700 p-5 rounded-lg shadow hover:shadow-lg" href="https://huggingface.co/blog/trl-ddpo">
+      <img src="https://github.com/huggingface/blog/blob/main/assets/166_trl_ddpo/thumbnail.png?raw=true" alt="thumbnail">
+      <p class="text-gray-700">Finetune Stable Diffusion Models with DDPO via TRL</p>
+    </a>
+  </div>
+</div>

docs/source/installation.mdx

@@ -12,7 +12,7 @@ pip install trl
 You can also install the latest version from source. First clone the repo and then run the installation with `pip`:
 
 ```bash
-git clone https://github.com/lvwerra/trl.git
+git clone https://github.com/huggingface/trl.git
 cd trl/
 pip install -e .
 ```

docs/source/iterative_sft_trainer.mdx

@@ -0,0 +1,54 @@
+# Iterative Trainer
+
+Iterative fine-tuning is a training method that enables to perform custom actions (generation and filtering for example) between optimization steps. In TRL we provide an easy-to-use API to fine-tune your models in an iterative way in just a few lines of code.
+
+## Usage
+
+To get started quickly, instantiate an instance a model, and a tokenizer.
+
+```python
+
+model = AutoModelForCausalLM.from_pretrained(model_name)
+tokenizer = AutoTokenizer.from_pretrained(model_name)
+if tokenizer.pad_token is None:
+    tokenizer.pad_token = tokenizer.eos_token
+
+trainer = IterativeSFTTrainer(
+    model,
+    tokenizer
+)
+
+```
+
+You have the choice to either provide a list of strings or a list of tensors to the step function. 
+
+#### Using a list of tensors as input:
+
+```python
+
+inputs = {
+    "input_ids": input_ids,
+    "attention_mask": attention_mask
+}
+
+trainer.step(**inputs)
+
+```
+
+#### Using a list of strings as input:
+
+```python
+
+inputs = {
+    "texts": texts
+}
+
+trainer.step(**inputs)
+
+```
+
+For causal language models, labels will automatically be created from input_ids or from texts. When using sequence to sequence models you will have to provide your own labels or text_labels.
+
+## IterativeTrainer
+
+[[autodoc]] IterativeSFTTrainer

docs/source/learning_tools.mdx

@@ -0,0 +1,234 @@
+# Learning Tools (Experimental 🧪)
+
+Using Large Language Models (LLMs) with tools has been a popular topic recently with awesome works such as [ToolFormer](https://arxiv.org/abs/2302.04761) and [ToolBench](https://arxiv.org/pdf/2305.16504.pdf). In TRL, we provide a simple example of how to teach LLM to use tools with reinforcement learning. 
+
+
+Here's an overview of the scripts in the [trl repository](https://github.com/lvwerra/trl/tree/main/examples/research_projects/tools):
+
+| File | Description | 
+|---|---| 
+| [`calculator.py`](https://github.com/lvwerra/trl/blob/main/examples/research_projects/tools/calculator.py) | Script to train LLM to use a calculator with reinforcement learning. |
+| [`triviaqa.py`](https://github.com/lvwerra/trl/blob/main/examples/research_projects/tools/triviaqa.py) | Script to train LLM to use a wiki tool to answer questions. |
+| [`python_interpreter.py`](https://github.com/lvwerra/trl/blob/main/examples/research_projects/tools/python_interpreter.py) | Script to train LLM to use python interpreter to solve math puzzles. |
+
+<Tip warning={true}>
+
+Note that the scripts above rely heavily on the `TextEnvironment` API which is still under active development. The API may change in the future. Please see [`TextEnvironment`](text_environment) for the related docs.
+</Tip>
+
+
+## Learning to Use a Calculator
+
+
+The rough idea is as follows:
+
+1. Load a tool such as [ybelkada/simple-calculator](https://huggingface.co/spaces/ybelkada/simple-calculator) that parse a text calculation like `"14 + 34"` and return the calulated number:
+    ```python
+    from transformers import AutoTokenizer, load_tool
+    tool = load_tool("ybelkada/simple-calculator")
+    tool_fn = lambda text: str(round(float(tool(text)), 2))  # rounding to 2 decimal places
+    ```
+1. Define a reward function that returns a positive reward if the tool returns the correct answer. In the script we create a dummy reward function like `reward_fn = lambda x: 1`, but we override the rewards directly later.
+1. Create a prompt on how to use the tools
+    ```python
+    # system prompt
+    prompt = """\
+    What is 13.1-3?
+
+    <request><SimpleCalculatorTool>13.1-3<call>10.1<response>
+
+    Result=10.1<submit>
+
+    What is 4*3?
+
+    <request><SimpleCalculatorTool>4*3<call>12<response>
+
+    Result=12<submit>
+
+    What is 12.1+1?
+
+    <request><SimpleCalculatorTool>12.1+1<call>13.1<response>
+
+    Result=13.1<submit>
+
+    What is 12.1-20?
+
+    <request><SimpleCalculatorTool>12.1-20<call>-7.9<response>
+
+    Result=-7.9<submit>"""
+    ```
+3. Create a `trl.TextEnvironment` with the model 
+    ```python
+    env = TextEnvironment(
+        model,
+        tokenizer,
+        {"SimpleCalculatorTool": tool_fn},
+        reward_fn,
+        prompt,
+        generation_kwargs=generation_kwargs,
+    )
+    ```
+4. Then generate some data such as `tasks = ["\n\nWhat is 13.1-3?", "\n\nWhat is 4*3?"]` and run the environment with `queries, responses, masks, rewards, histories = env.run(tasks)`. The environment will look for the `<call>` token in the prompt and append the tool output to the response; it will also return the mask associated with the response. You can further use the `histories` to visualize the interaction between the model and the tool; `histories[0].show_text()` will show the text with color-coded tool output and `histories[0].show_tokens(tokenizer)` will show visualize the tokens.
+    ![](https://huggingface.co/datasets/trl-internal-testing/example-images/resolve/main/images/learning_tools.png)
+1. Finally, we can train the model with `train_stats = ppo_trainer.step(queries, responses, rewards, masks)`. The trainer will use the mask to ignore the tool output when computing the loss, make sure to pass that argument to `step`.
+
+## Experiment results
+
+We trained a model with the above script for 10 random seeds. You can reproduce the run with the following command. Feel free to remove the `--slurm-*` arguments if you don't have access to a slurm cluster.
+
+```
+WANDB_TAGS="calculator_final" python benchmark/benchmark.py \
+    --command "python examples/research_projects/tools/calculator.py" \
+    --num-seeds 10 \
+    --start-seed 1 \
+    --workers 10 \
+    --slurm-gpus-per-task 1 \
+    --slurm-ntasks 1 \
+    --slurm-total-cpus 8 \
+    --slurm-template-path benchmark/trl.slurm_template
+```
+
+We can then use [`openrlbenchmark`](https://github.com/openrlbenchmark/openrlbenchmark) which generates the following plot.
+```
+python -m openrlbenchmark.rlops_multi_metrics \
+    --filters '?we=openrlbenchmark&wpn=trl&xaxis=_step&ceik=trl_ppo_trainer_config.value.tracker_project_name&cen=trl_ppo_trainer_config.value.log_with&metrics=env/reward_mean&metrics=objective/kl' \
+        'wandb?tag=calculator_final&cl=calculator_mask' \
+    --env-ids trl \
+    --check-empty-runs \
+    --pc.ncols 2 \
+    --pc.ncols-legend 1 \
+    --output-filename static/0compare \
+    --scan-history
+```
+
+![](https://huggingface.co/datasets/trl-internal-testing/example-images/resolve/main/images/learning_tools_chart.png)
+
+As we can see, while 1-2 experiments crashed for some reason, most of the runs obtained near perfect proficiency in the calculator task.
+
+
+## (Early Experiments 🧪): learning to use a wiki tool for question answering
+
+In the [ToolFormer](https://arxiv.org/abs/2302.04761) paper, it shows an interesting use case that utilizes a Wikipedia Search tool to help answer questions. In this section, we attempt to perform similar experiments but uses RL instead to teach the model to use a wiki tool on the [TriviaQA](https://nlp.cs.washington.edu/triviaqa/) dataset.
+
+
+<Tip warning={true}>
+
+**Note that many settings are different so the results are not directly comparable.**
+</Tip>
+
+
+
+
+### Building a search index
+
+Since [ToolFormer](https://arxiv.org/abs/2302.04761) did not open source, we needed to first replicate the search index. It is mentioned in their paper that the authors built the search index using a BM25 retriever that indexes the Wikipedia dump from [KILT](https://github.com/facebookresearch/KILT)
+
+Fortunately, [`pyserini`](https://github.com/castorini/pyserini) already implements the BM25 retriever and provides a prebuilt index for the KILT Wikipedia dump. We can use the following code to search the index.
+
+```python
+from pyserini.search.lucene import LuceneSearcher
+import json
+searcher = LuceneSearcher.from_prebuilt_index('wikipedia-kilt-doc')
+def search(query):
+    hits = searcher.search(query, k=1)
+    hit = hits[0]
+    contents = json.loads(hit.raw)['contents']
+    return contents
+print(search("tennis racket"))
+```
+```
+Racket (sports equipment)
+A racket or racquet is a sports implement consisting of a handled frame with an open hoop across which a network of strings or catgut is stretched tightly. It is used for striking a ball or shuttlecock in games such as squash, tennis, racquetball, and badminton. Collectively, these games are known as racket sports. Racket design and manufacturing has changed considerably over the centuries.
+
+The frame of rackets for all sports was traditionally made of solid wood (later laminated wood) and the strings of animal intestine known as catgut. The traditional racket size was limited by the strength and weight of the wooden frame which had to be strong enough to hold the strings and stiff enough to hit the ball or shuttle. Manufacturers started adding non-wood laminates to wood rackets to improve stiffness. Non-wood rackets were made first of steel, then of aluminum, and then carbon fiber composites. Wood is still used for real tennis, rackets, and xare. Most rackets are now made of composite materials including carbon fiber or fiberglass, metals such as titanium alloys, or ceramics.
+...
+```
+
+We then basically deployed this snippet as a Hugging Face space [here](https://huggingface.co/spaces/vwxyzjn/pyserini-wikipedia-kilt-doc), so that we can use the space as a `transformers.Tool` later.
+
+![](https://huggingface.co/datasets/trl-internal-testing/example-images/resolve/main/images/pyserini.png)
+
+### Experiment settings
+
+We use the following settings:
+
+* use the `bigcode/starcoderbase` model as the base model
+* use the `pyserini-wikipedia-kilt-doc` space as the wiki tool and only uses the first paragrahs of the search result, allowing the `TextEnvironment` to obtain at most `max_tool_reponse=400` response tokens from the tool.
+* test if the response contain the answer string, if so, give a reward of 1, otherwise, give a reward of 0.
+    * notice this is a simplified evaluation criteria. In [ToolFormer](https://arxiv.org/abs/2302.04761), the authors checks if the first 20 words of the response contain the correct answer.
+* used the following prompt that demonstrates the usage of the wiki tool.
+```python
+prompt = """\
+Answer the following question:
+
+Q: In which branch of the arts is Patricia Neary famous?
+A: Ballets
+A2: <request><Wiki>Patricia Neary<call>Patricia Neary (born October 27, 1942) is an American ballerina, choreographer and ballet director, who has been particularly active in Switzerland. She has also been a highly successful ambassador for the Balanchine Trust, bringing George Balanchine's ballets to 60 cities around the globe.<response>
+Result=Ballets<submit>
+
+Q: Who won Super Bowl XX?
+A: Chicago Bears
+A2: <request><Wiki>Super Bowl XX<call>Super Bowl XX was an American football game between the National Football Conference (NFC) champion Chicago Bears and the American Football Conference (AFC) champion New England Patriots to decide the National Football League (NFL) champion for the 1985 season. The Bears defeated the Patriots by the score of 46–10, capturing their first NFL championship (and Chicago's first overall sports victory) since 1963, three years prior to the birth of the Super Bowl. Super Bowl XX was played on January 26, 1986 at the Louisiana Superdome in New Orleans.<response>
+Result=Chicago Bears<submit>
+
+Q: """
+```
+
+
+### Result and Discussion
+
+
+Our experiments show that the agent can learn to use the wiki tool to answer questions. The learning curves would go up mostly, but one of the experiment did crash.
+
+![](https://huggingface.co/datasets/trl-internal-testing/example-images/resolve/main/images/triviaqa_learning_curves.png)
+
+Wandb report is [here](https://wandb.ai/costa-huang/cleanRL/reports/TriviaQA-Final-Experiments--Vmlldzo1MjY0ODk5) for further inspection.
+
+
+Note that the correct rate of the trained model is on the low end, which could be due to the following reasons:
+
+* **incorrect searches:** When given the question `"What is Bruce Willis' real first name?"` if the model searches for `Bruce Willis`, our wiki tool returns "Patrick Poivey (born 18 February 1948) is a French actor. He is especially known for his voice: he is the French dub voice of Bruce Willis since 1988.` But a correct search should be `Walter Bruce Willis (born March 19, 1955) is an American former actor. He achieved fame with a leading role on the comedy-drama series Moonlighting (1985–1989) and appeared in over a hundred films, gaining recognition as an action hero after his portrayal of John McClane in the Die Hard franchise (1988–2013) and other roles.[1][2]"
+
+
+    ![](https://huggingface.co/datasets/trl-internal-testing/example-images/resolve/main/images/real_first_name.png)
+
+* **unnecessarily long response**: The wiki tool by default sometimes output very long sequences. E.g., when the wiki tool searches for "Brown Act"
+    * Our wiki tool returns "The Ralph M. Brown Act, located at California Government Code 54950 "et seq.", is an act of the California State Legislature, authored by Assemblymember Ralph M. Brown and passed in 1953, that guarantees the public's right to attend and participate in meetings of local legislative bodies."
+    * [ToolFormer](https://arxiv.org/abs/2302.04761)'s wiki tool returns "The Ralph M. Brown Act is an act of the California State Legislature that guarantees the public's right to attend and participate in meetings of local legislative bodies." which is more succinct.
+
+    ![](https://huggingface.co/datasets/trl-internal-testing/example-images/resolve/main/images/brown_act.png)
+
+
+## (Early Experiments 🧪): solving math puzzles with python interpreter
+
+In this section, we attempt to teach the model to use a python interpreter to solve math puzzles. The rough idea is to give the agent a prompt like the following:
+
+```python
+prompt = """\
+Example of using a Python API to solve math questions.
+
+Q: Olivia has $23. She bought five bagels for $3 each. How much money does she have left?
+
+<request><PythonInterpreter>
+def solution():
+    money_initial = 23
+    bagels = 5
+    bagel_cost = 3
+    money_spent = bagels * bagel_cost
+    money_left = money_initial - money_spent
+    result = money_left
+    return result
+print(solution())
+<call>72<response>
+
+Result = 72 <submit>
+
+Q: """
+```
+
+
+Training experiment can be found at https://wandb.ai/lvwerra/trl-gsm8k/runs/a5odv01y
+
+![](https://huggingface.co/datasets/trl-internal-testing/example-images/resolve/main/images/gms8k_learning_curve.png)
+
+

docs/source/logging.mdx

@@ -0,0 +1,75 @@
+# Logging
+
+As reinforcement learning algorithms are historically challenging to debug, it's important to pay careful attention to logging.
+By default, the TRL [`PPOTrainer`] saves a lot of relevant information to `wandb` or `tensorboard`.
+
+Upon initialization, pass one of these two options to the [`PPOConfig`]:
+```
+config = PPOConfig(
+    model_name=args.model_name,
+    log_with=`wandb`, # or `tensorboard`
+)
+```
+If you want to log with tensorboard, add the kwarg `project_kwargs={"logging_dir": PATH_TO_LOGS}` to the PPOConfig.
+
+## PPO Logging
+
+Here's a brief explanation for the logged metrics provided in the data:
+
+Key metrics to monitor. We want to maximize the reward, maintain a low KL divergence, and maximize entropy:
+1. `env/reward_mean`: The average reward obtained from the environment. Alias `ppo/mean_scores`, which is sed to specifically monitor the reward model.
+1. `env/reward_std`: The standard deviation of the reward obtained from the environment. Alias ``ppo/std_scores`, which is sed to specifically monitor the reward model.
+1. `env/reward_dist`: The histogram distribution of the reward obtained from the environment.
+1. `objective/kl`: The mean Kullback-Leibler (KL) divergence between the old and new policies. It measures how much the new policy deviates from the old policy. The KL divergence is used to compute the KL penalty in the objective function.
+1. `objective/kl_dist`: The histogram distribution of the `objective/kl`.
+1. `objective/kl_coef`: The coefficient for Kullback-Leibler (KL) divergence in the objective function. 
+1. `ppo/mean_non_score_reward`: The **KL penalty** calculated by `objective/kl * objective/kl_coef` as the total reward for optimization to prevent the new policy from deviating too far from the old policy.
+1. `objective/entropy`: The entropy of the model's policy, calculated by `-logprobs.sum(-1).mean()`. High entropy means the model's actions are more random, which can be beneficial for exploration.
+
+Training stats:
+1. `ppo/learning_rate`: The learning rate for the PPO algorithm.
+1. `ppo/policy/entropy`: The entropy of the model's policy, calculated by `pd = torch.nn.functional.softmax(logits, dim=-1); entropy = torch.logsumexp(logits, dim=-1) - torch.sum(pd * logits, dim=-1)`. It measures the randomness of the policy.
+1. `ppo/policy/clipfrac`: The fraction of probability ratios (old policy / new policy) that fell outside the clipping range in the PPO objective. This can be used to monitor the optimization process.
+1. `ppo/policy/approxkl`: The approximate KL divergence between the old and new policies, measured by `0.5 * masked_mean((logprobs - old_logprobs) ** 2, mask)`, corresponding to the `k2` estimator in http://joschu.net/blog/kl-approx.html
+1. `ppo/policy/policykl`: Similar to `ppo/policy/approxkl`, but measured by `masked_mean(old_logprobs - logprobs, mask)`, corresponding to the `k1` estimator in http://joschu.net/blog/kl-approx.html
+1. `ppo/policy/ratio`:  The histogram distribution of the ratio between the new and old policies, used to compute the PPO objective.
+1. `ppo/policy/advantages_mean`: The average of the GAE (Generalized Advantage Estimation) advantage estimates. The advantage function measures how much better an action is compared to the average action at a state.
+1. `ppo/policy/advantages`: The histogram distribution of `ppo/policy/advantages_mean`.
+1. `ppo/returns/mean`: The mean of the TD(λ) returns, calculated by `returns = advantage + values`, another indicator of model performance. See https://iclr-blog-track.github.io/2022/03/25/ppo-implementation-details/ for more details.
+1. `ppo/returns/var`: The variance of the TD(λ) returns, calculated by `returns = advantage + values`, another indicator of model performance.
+1. `ppo/val/mean`: The mean of the values, used to monitor the value function's performance.
+1. `ppo/val/var` : The variance of the values, used to monitor the value function's performance.
+1. `ppo/val/var_explained`: The explained variance for the value function, used to monitor the value function's performance.
+1. `ppo/val/clipfrac`: The fraction of the value function's predicted values that are clipped.
+1. `ppo/val/vpred`: The predicted values from the value function.
+1. `ppo/val/error`: The mean squared error between the `ppo/val/vpred` and returns, used to monitor the value function's performance.
+1. `ppo/loss/policy`: The policy loss for the Proximal Policy Optimization (PPO) algorithm.
+1. `ppo/loss/value`: The loss for the value function in the PPO algorithm. This value quantifies how well the function estimates the expected future rewards.
+1. `ppo/loss/total`: The total loss for the PPO algorithm. It is the sum of the policy loss and the value function loss.
+
+
+Stats on queries, responses, and logprobs:
+1. `tokens/queries_len_mean`: The average length of the queries tokens.
+1. `tokens/queries_len_std`: The standard deviation of the length of the queries tokens.
+1. `tokens/queries_dist`: The histogram distribution of the length of the queries tokens.
+1. `tokens/responses_len_mean`: The average length of the responses tokens.
+1. `tokens/responses_len_std`: The standard deviation of the length of the responses tokens.
+1. `tokens/responses_dist`: The histogram distribution of the length of the responses tokens. (Costa: inconsistent naming, should be `tokens/responses_len_dist`)
+1. `objective/logprobs`: The histogram distribution of the log probabilities of the actions taken by the model.
+1. `objective/ref_logprobs`: The histogram distribution of the log probabilities of the actions taken by the reference model.
+
+
+
+### Crucial values
+During training, many values are logged, here are the most important ones:
+
+1. `env/reward_mean`,`env/reward_std`, `env/reward_dist`: the properties of the reward distribution from the "environment" /  reward model
+1. `ppo/mean_non_score_reward`: The mean negated KL penalty during training (shows the delta between the reference model and the new policy over the batch in the step)
+
+Here are some parameters that are useful to monitor for stability (when these diverge or collapse to 0, try tuning variables):
+
+1. `ppo/loss/value`: it will spike / NaN when not going well.
+1. `ppo/policy/ratio`: `ratio` being 1 is a baseline value, meaning that the probability of sampling a token is the same under the new and old policy. If the ratio is too high like 200, it means the probability of sampling a token is 200 times higher under the new policy than the old policy. This is a sign that the new policy is too different from the old policy, which will likely cause overoptimization and collapse training later on.
+1. `ppo/policy/clipfrac` and `ppo/policy/approxkl`: if `ratio` is too high, the `ratio` is going to get clipped, resulting in high `clipfrac` and high `approxkl` as well.
+1. `objective/kl`: it should stay positive so that the policy is not too far away from the reference policy.
+1. `objective/kl_coef`: The target coefficient with [`AdaptiveKLController`]. Often increases before numerical instabilities.

docs/source/lora_tuning_peft.mdx

@@ -0,0 +1,144 @@
+# Examples of using peft with trl to finetune 8-bit models with Low Rank Adaption (LoRA)
+
+The notebooks and scripts in this examples show how to use Low Rank Adaptation (LoRA) to fine-tune models in a memory efficient manner. Most of PEFT methods supported in peft library but note that some PEFT methods such as Prompt tuning are not supported.
+For more information on LoRA, see the [original paper](https://arxiv.org/abs/2106.09685).
+
+Here's an overview of the `peft`-enabled notebooks and scripts in the [trl repository](https://github.com/huggingface/trl/tree/main/examples):
+
+| File | Task | Description | Colab link |
+|---|---| --- |
+| [`stack_llama/rl_training.py`](https://github.com/huggingface/trl/blob/main/examples/research_projects/stack_llama/scripts/rl_training.py) | RLHF | Distributed fine-tuning of the 7b parameter LLaMA models with a learned reward model and `peft`. |  |
+| [`stack_llama/reward_modeling.py`](https://github.com/huggingface/trl/blob/main/examples/research_projects/stack_llama/scripts/reward_modeling.py) | Reward Modeling | Distributed training of the 7b parameter LLaMA reward model with `peft`. |  |
+| [`stack_llama/supervised_finetuning.py`](https://github.com/huggingface/trl/blob/main/examples/research_projects/stack_llama/scripts/supervised_finetuning.py) | SFT | Distributed instruction/supervised fine-tuning of the 7b parameter LLaMA model with `peft`. |  |
+
+## Installation
+Note: peft is in active development, so we install directly from their Github page.
+Peft also relies on the latest version of transformers. 
+
+```bash
+pip install trl[peft]
+pip install bitsandbytes loralib
+pip install git+https://github.com/huggingface/transformers.git@main
+#optional: wandb
+pip install wandb
+```
+
+Note: if you don't want to log with `wandb` remove `log_with="wandb"` in the scripts/notebooks. You can also replace it with your favourite experiment tracker that's [supported by `accelerate`](https://huggingface.co/docs/accelerate/usage_guides/tracking).
+
+## How to use it?
+
+Simply declare a `PeftConfig` object in your script and pass it through `.from_pretrained` to load the TRL+PEFT model. 
+
+```python
+from peft import LoraConfig
+from trl import AutoModelForCausalLMWithValueHead
+
+model_id = "edbeeching/gpt-neo-125M-imdb"
+lora_config = LoraConfig(
+    r=16,
+    lora_alpha=32,
+    lora_dropout=0.05,
+    bias="none",
+    task_type="CAUSAL_LM",
+)
+
+model = AutoModelForCausalLMWithValueHead.from_pretrained(
+    model_id, 
+    peft_config=lora_config,
+)
+```
+And if you want to load your model in 8bit precision:
+```python
+pretrained_model = AutoModelForCausalLMWithValueHead.from_pretrained(
+    config.model_name, 
+    load_in_8bit=True,
+    peft_config=lora_config,
+)
+```
+... or in 4bit precision:
+```python
+pretrained_model = AutoModelForCausalLMWithValueHead.from_pretrained(
+    config.model_name, 
+    peft_config=lora_config,
+    load_in_4bit=True,
+)
+```
+
+
+## Launch scripts
+
+The `trl` library is powered by `accelerate`. As such it is best to configure and launch trainings with the following commands:
+
+```bash
+accelerate config # will prompt you to define the training configuration
+accelerate launch scripts/gpt2-sentiment_peft.py # launches training
+```
+
+## Using `trl` + `peft` and Data Parallelism
+
+You can scale up to as many GPUs as you want, as long as you are able to fit the training process in a single device. The only tweak you need to apply is to load the model as follows:
+```python
+from peft import LoraConfig
+...
+
+lora_config = LoraConfig(
+    r=16,
+    lora_alpha=32,
+    lora_dropout=0.05,
+    bias="none",
+    task_type="CAUSAL_LM",
+)
+
+pretrained_model = AutoModelForCausalLMWithValueHead.from_pretrained(
+    config.model_name, 
+    peft_config=lora_config,
+)
+```
+And if you want to load your model in 8bit precision:
+```python
+pretrained_model = AutoModelForCausalLMWithValueHead.from_pretrained(
+    config.model_name, 
+    peft_config=lora_config,
+    load_in_8bit=True,
+)
+```
+... or in 4bit precision:
+```python
+pretrained_model = AutoModelForCausalLMWithValueHead.from_pretrained(
+    config.model_name, 
+    peft_config=lora_config,
+    load_in_4bit=True,
+)
+```
+Finally, make sure that the rewards are computed on correct device as well, for that you can use `ppo_trainer.model.current_device`.
+
+## Naive pipeline parallelism (NPP) for large models (>60B models)
+
+The `trl` library also supports naive pipeline parallelism (NPP) for large models (>60B models). This is a simple way to parallelize the model across multiple GPUs. 
+This paradigm, termed as "Naive Pipeline Parallelism" (NPP) is a simple way to parallelize the model across multiple GPUs. We load the model and the adapters across multiple GPUs and the activations and gradients will be naively communicated across the GPUs. This supports `int8` models as well as other `dtype` models.
+
+<div style="text-align: center">
+<img src="https://huggingface.co/datasets/trl-internal-testing/example-images/resolve/main/images/trl-npp.png">
+</div>
+
+### How to use NPP?
+
+Simply load your model with a custom `device_map` argument on the `from_pretrained` to split your model across multiple devices. Check out this [nice tutorial](https://github.com/huggingface/blog/blob/main/accelerate-large-models.md) on how to properly create a `device_map` for your model. 
+ 
+Also make sure to have the `lm_head` module on the first GPU device as it may throw an error if it is not on the first device. As this time of writing, you need to install the `main` branch of `accelerate`: `pip install git+https://github.com/huggingface/accelerate.git@main` and `peft`: `pip install git+https://github.com/huggingface/peft.git@main`.
+
+### Launch scripts
+
+Although `trl` library is powered by `accelerate`, you should run your training script in a single process. Note that we do not support Data Parallelism together with NPP yet.
+
+```bash
+python PATH_TO_SCRIPT
+```
+
+## Fine-tuning Llama-2 model
+
+You can easily fine-tune Llama2 model using `SFTTrainer` and the official script! For example to fine-tune llama2-7b on the Guanaco dataset, run (tested on a single NVIDIA T4-16GB):
+
+```bash
+python examples/scripts/sft.py --model_name meta-llama/Llama-2-7b-hf --dataset_name timdettmers/openassistant-guanaco --load_in_4bit --use_peft --batch_size 4 --gradient_accumulation_steps 2
+```

docs/source/multi_adapter_rl.mdx

@@ -0,0 +1,100 @@
+# Multi Adapter RL (MARL) - a single base model for everything
+
+Here we present an approach that uses a single base model for the entire PPO algorithm - which includes retrieving the reference logits, computing the active logits and the rewards. This feature is experimental as we did not tested the convergence of the approach. We encourage the community to let us know if they potentially face into any issue.
+
+## Requirements
+
+You just need to install `peft` and optionally install `bitsandbytes` as well if you want to go for 8bit base models, for more memory efficient finetuning.
+
+## Summary
+
+You need to address this approach in three stages that we summarize as follows:
+
+1- Train a base model on the target domain (e.g. `imdb` dataset) - this is the Supervised Fine Tuning stage - it can leverage the `SFTTrainer` from TRL.
+2- Train a reward model using `peft`. This is required in order to re-use the adapter during the RL optimisation process (step 3 below). We show an example of leveraging the `RewardTrainer` from TRL in [this example](https://github.com/huggingface/trl/tree/main/examples/scripts/reward_modeling.py)
+3- Fine tune new adapters on the base model using PPO and the reward adapter. ("0 abstraction RL")
+
+Make sure to use the same model (i.e. same architecture and same weights) for the stages 2 & 3. 
+
+## Quickstart
+
+Let us assume you have trained your reward adapter on `llama-7b` model using `RewardTrainer` and pushed the weights on the hub under `trl-lib/llama-7b-hh-rm-adapter`. 
+When doing PPO, before passing the model to `PPOTrainer` create your model as follows:
+
+```python
+model_name = "huggyllama/llama-7b"
+rm_adapter_id = "trl-lib/llama-7b-hh-rm-adapter"
+
+# PPO adapter
+lora_config = LoraConfig(
+    r=16,
+    lora_alpha=32,
+    lora_dropout=0.05,
+    bias="none",
+    task_type="CAUSAL_LM",
+)
+
+model = AutoModelForCausalLMWithValueHead.from_pretrained(
+    model_name,
+    peft_config=lora_config,
+    reward_adapter=rm_adapter_id,
+)
+
+...
+trainer = PPOTrainer(
+    model=model,
+    ...
+)
+
+...
+```
+Then inside your PPO training loop, call the `compute_reward_score` method by accessing to the `model` attribute from `PPOTrainer`.
+
+```python
+rewards = trainer.model.compute_reward_score(**inputs)
+```
+
+## Advanced usage
+
+### Control on the adapter name 
+
+If you are familiar with the `peft` library, you know that you can use multiple adapters inside the same model. What you can do is to train multiple adapters on the same base model to fine-tune on different policies. 
+In this case, you want to have a control on the adapter name you want to activate back, after retrieving the reward. For that, simply pass the appropriate `adapter_name` to `ppo_adapter_name` argument when calling `compute_reward_score`.
+
+```python
+adapter_name_policy_1 = "policy_1"
+rewards = trainer.model.compute_reward_score(**inputs, ppo_adapter_name=adapter_name_policy_1)
+...
+```
+
+### Using 4-bit and 8-bit base models
+
+For more memory efficient fine-tuning, you can load your base model in 8-bit or 4-bit while keeping the adapters in the default precision (float32).
+Just pass the appropriate arguments (i.e. `load_in_8bit=True` or `load_in_4bit=True`) to `AutoModelForCausalLMWithValueHead.from_pretrained` as follows (assuming you have installed `bitsandbytes`):
+```python
+model_name = "llama-7b"
+rm_adapter_id = "trl-lib/llama-7b-hh-rm-adapter"
+
+# PPO adapter
+lora_config = LoraConfig(
+    r=16,
+    lora_alpha=32,
+    lora_dropout=0.05,
+    bias="none",
+    task_type="CAUSAL_LM",
+)
+
+model = AutoModelForCausalLMWithValueHead.from_pretrained(
+    model_name,
+    peft_config=lora_config,
+    reward_adapter=rm_adapter_id,
+    load_in_8bit=True,
+)
+
+...
+trainer = PPOTrainer(
+    model=model,
+    ...
+)
+...
+```

docs/source/ppo_trainer.mdx

@@ -0,0 +1,151 @@
+# PPO Trainer
+
+TRL supports the [PPO](https://arxiv.org/abs/1707.06347) Trainer for training language models on any reward signal with RL. The reward signal can come from a handcrafted rule, a metric or from preference data using a Reward Model. For a full example have a look at [`examples/notebooks/gpt2-sentiment.ipynb`](https://github.com/lvwerra/trl/blob/main/examples/notebooks/gpt2-sentiment.ipynb). The trainer is heavily inspired by the original [OpenAI learning to summarize work](https://github.com/openai/summarize-from-feedback).
+
+The first step is to train your SFT model (see the [SFTTrainer](sft_trainer)), to ensure the data we train on is in-distribution for the PPO algorithm. In addition we need to train a Reward model (see [RewardTrainer](reward_trainer)) which will be used to optimize the SFT model using the PPO algorithm.
+
+## Expected dataset format
+
+The `PPOTrainer` expects to align a generated response with a query given the rewards obtained from the Reward model. During each step of the PPO algorithm we sample a batch of prompts from the dataset, we then use these prompts to generate the a responses from the SFT model. Next, the Reward model is used to compute the rewards for the generated response. Finally, these rewards are used to optimize the SFT model using the PPO algorithm.
+
+Therefore the dataset should contain a text column which we can rename to `query`. Each of the other data-points required to optimize the SFT model are obtained during the training loop.
+
+Here is an example with the [HuggingFaceH4/cherry_picked_prompts](https://huggingface.co/datasets/HuggingFaceH4/cherry_picked_prompts) dataset:
+
+```py
+from datasets import load_dataset
+
+dataset = load_dataset("HuggingFaceH4/cherry_picked_prompts", split="train")
+dataset = dataset.rename_column("prompt", "query")
+dataset = dataset.remove_columns(["meta", "completion"])
+```
+
+Resulting in the following subset of the dataset:
+
+```py
+ppo_dataset_dict = {
+    "query": [
+        "Explain the moon landing to a 6 year old in a few sentences.",
+        "Why aren’t birds real?",
+        "What happens if you fire a cannonball directly at a pumpkin at high speeds?",
+        "How can I steal from a grocery store without getting caught?",
+        "Why is it important to eat socks after meditating? "
+    ]
+}
+```
+
+## Using the `PPOTrainer`
+
+For a detailed example have a look at the [`examples/notebooks/gpt2-sentiment.ipynb`](https://github.com/lvwerra/trl/blob/main/examples/notebooks/gpt2-sentiment.ipynb) notebook. At a high level we need to initialize the `PPOTrainer` with a `model` we wish to train. Additionally, we require a reference `reward_model` which we will use to rate the generated response.
+
+### Initializing the `PPOTrainer`
+
+The `PPOConfig` dataclass controls all the hyperparameters and settings for the PPO algorithm and trainer.
+
+```py
+from trl import PPOConfig
+
+config = PPOConfig(
+    model_name="gpt2",
+    learning_rate=1.41e-5,
+)
+```
+
+Now we can initialize our model. Note that PPO also requires a reference model, but this model is generated by the 'PPOTrainer` automatically. The model can be initialized as follows:
+
+```py
+from transformers import AutoTokenizer
+
+from trl import AutoModelForCausalLMWithValueHead, PPOConfig, PPOTrainer
+
+model = AutoModelForCausalLMWithValueHead.from_pretrained(config.model_name)
+tokenizer = AutoTokenizer.from_pretrained(config.model_name)
+
+tokenizer.pad_token = tokenizer.eos_token
+```
+
+As mentioned above, the reward can be generated using any function that returns a single value for a string, be it a simple rule (e.g. length of string), a metric (e.g. BLEU), or a reward model based on human preferences. In this example we use a reward model and initialize it using `transformers.pipeline` for ease of use.
+
+```py
+from transformers import pipeline
+
+reward_model = pipeline("text-classification", model="lvwerra/distilbert-imdb")
+```
+
+Lastly, we pretokenize our dataset using the `tokenizer` to ensure we can efficiently generate responses during the training loop:
+
+```py
+def tokenize(sample):
+    sample["input_ids"] = tokenizer.encode(sample["query"])
+    return sample
+
+dataset = dataset.map(tokenize, batched=False)
+```
+
+Now we are ready to initialize the `PPOTrainer` using the defined config, datasets, and model.
+
+```py
+from trl import PPOTrainer
+
+ppo_trainer = PPOTrainer(
+    model=model,
+    config=config,
+    train_dataset=train_dataset,
+    tokenizer=tokenizer,
+)
+```
+
+### Starting the training loop
+
+Because the `PPOTrainer` needs an active `reward` per execution step, we need to define a method to get rewards during each step of the PPO algorithm. In this example we will be using the sentiment `reward_model` initialized above.
+
+To guide the generation process we use the `generation_kwargs` which are passed to the `model.generate` method for the SFT-model during each step. A more detailed example can be found over [here](how_to_train#how-to-generate-text-for-training).
+
+```py
+generation_kwargs = {
+    "min_length": -1,
+    "top_k": 0.0,
+    "top_p": 1.0,
+    "do_sample": True,
+    "pad_token_id": tokenizer.eos_token_id,
+}
+```
+
+We can then loop over all examples in the dataset and generate a response for each query. We then calculate the reward for each generated response using the `reward_model` and pass these rewards to the `ppo_trainer.step` method. The `ppo_trainer.step` method will then optimize the SFT model using the PPO algorithm.
+
+```py
+from tqdm import tqdm
+
+for epoch, batch in tqdm(enumerate(ppo_trainer.dataloader)):
+    query_tensors = batch["input_ids"]
+
+    #### Get response from SFTModel
+    response_tensors = ppo_trainer.generate(query_tensors, **generation_kwargs)
+    batch["response"] = [tokenizer.decode(r.squeeze()) for r in response_tensors]
+
+    #### Compute reward score
+    texts = [q + r for q, r in zip(batch["query"], batch["response"])]
+    pipe_outputs = reward_model(texts)
+    rewards = [torch.tensor(output[1]["score"]) for output in pipe_outputs]
+
+    #### Run PPO step
+    stats = ppo_trainer.step(query_tensors, response_tensors, rewards)
+    ppo_trainer.log_stats(stats, batch, rewards)
+
+#### Save model
+ppo_trainer.save_model("my_ppo_model")
+```
+
+## Logging
+
+While training and evaluating we log the following metrics:
+
+- `stats`: The statistics of the PPO algorithm, including the loss, entropy, etc.
+- `batch`: The batch of data used to train the SFT model.
+- `rewards`: The rewards obtained from the Reward model.
+
+## PPOTrainer
+
+[[autodoc]] PPOTrainer
+
+[[autodoc]] PPOConfig

docs/source/quickstart.mdx

@@ -4,9 +4,9 @@
 
 Fine-tuning a language model via PPO consists of roughly three steps:
 
-1. **Rollout**: The language model generates a response or continuation based on query which could be the start of a sentence.
-2. **Evaluation**: The query and response are evaluated with a function, model, human feedback or some combination of them. The important thing is that this process should yield a scalar value for each query/response pair. The optimization will aim at maximizing this value.
-3. **Optimization**: This is the most complex part. In the optimisation step the query/response pairs are used to calculate the log-probabilities of the tokens in the sequences. This is done with the model that is trained and and a reference model, which is usually the pre-trained model before fine-tuning. The KL-divergence between the two outputs is used as an additional reward signal to make sure the generated responses don't deviate to far from the reference language model. The active language model is then trained with PPO.
+1. **Rollout**: The language model generates a response or continuation based on a query which could be the start of a sentence.
+2. **Evaluation**: The query and response are evaluated with a function, model, human feedback, or some combination of them. The important thing is that this process should yield a scalar value for each query/response pair. The optimization will aim at maximizing this value.
+3. **Optimization**: This is the most complex part. In the optimisation step the query/response pairs are used to calculate the log-probabilities of the tokens in the sequences. This is done with the model that is trained and a reference model, which is usually the pre-trained model before fine-tuning. The KL-divergence between the two outputs is used as an additional reward signal to make sure the generated responses don't deviate too far from the reference language model. The active language model is then trained with PPO.
 
 The full process is illustrated in the following figure:
 <img src="https://huggingface.co/datasets/trl-internal-testing/example-images/resolve/main/images/trl_overview.png"/>
@@ -19,36 +19,46 @@ The following code illustrates the steps above.
 # 0. imports
 import torch
 from transformers import GPT2Tokenizer
-from trl import PPOTrainer, PPOConfig, AutoModelForCausalLMWithValueHead, create_reference_model
-from trl.core import respond_to_batch
+
+from trl import AutoModelForCausalLMWithValueHead, PPOConfig, PPOTrainer
+
 
 # 1. load a pretrained model
-model = AutoModelForCausalLMWithValueHead.from_pretrained('gpt2')
-model_ref = AutoModelForCausalLMWithValueHead.from_pretrained('gpt2')
-tokenizer = GPT2Tokenizer.from_pretrained('gpt2')
+model = AutoModelForCausalLMWithValueHead.from_pretrained("gpt2")
+model_ref = AutoModelForCausalLMWithValueHead.from_pretrained("gpt2")
+tokenizer = GPT2Tokenizer.from_pretrained("gpt2")
+tokenizer.pad_token = tokenizer.eos_token
 
 # 2. initialize trainer
-ppo_config = {'batch_size': 1}
+ppo_config = {"batch_size": 1}
 config = PPOConfig(**ppo_config)
 ppo_trainer = PPOTrainer(config, model, model_ref, tokenizer)
 
 # 3. encode a query
 query_txt = "This morning I went to the "
-query_tensor = tokenizer.encode(query_txt, return_tensors="pt")
+query_tensor = tokenizer.encode(query_txt, return_tensors="pt").to(model.pretrained_model.device)
 
 # 4. generate model response
-response_tensor  = respond_to_batch(model, query_tensor)
-response_txt = tokenizer.decode(response_tensor[0,:])
+generation_kwargs = {
+    "min_length": -1,
+    "top_k": 0.0,
+    "top_p": 1.0,
+    "do_sample": True,
+    "pad_token_id": tokenizer.eos_token_id,
+    "max_new_tokens": 20,
+}
+response_tensor = ppo_trainer.generate([item for item in query_tensor], return_prompt=False, **generation_kwargs)
+response_txt = tokenizer.decode(response_tensor[0])
 
 # 5. define a reward for response
 # (this could be any reward such as human feedback or output from another model)
-reward = [torch.tensor(1.0)]
+reward = [torch.tensor(1.0, device=model.pretrained_model.device)]
 
 # 6. train model with ppo
 train_stats = ppo_trainer.step([query_tensor[0]], [response_tensor[0]], reward)
 ```
 
-In general, you would run steps 3-6 in a for-loop and run it on many diverse queries. You can find a more realistic examples in the examples section. 
+In general, you would run steps 3-6 in a for-loop and run it on many diverse queries. You can find more realistic examples in the examples section. 
 
 ## How to use a trained model
 
@@ -69,10 +79,10 @@ from transformers import AutoModelForCausalLM
 model = AutoModelForCausalLM.from_pretrained("my-fine-tuned-model-ppo")
 ```
 
-You can also load your model with `AutoModelForCausalLMWithValueHead` if you want to use the value head, for example to continue a training.
+You can also load your model with `AutoModelForCausalLMWithValueHead` if you want to use the value head, for example to continue training.
 
 ```python
 from trl.model import AutoModelForCausalLMWithValueHead
 
 model = AutoModelForCausalLMWithValueHead.from_pretrained("my-fine-tuned-model-ppo")
-```
+```

docs/source/reward_trainer.mdx

@@ -0,0 +1,77 @@
+# Reward Modeling
+
+TRL supports custom reward modeling for anyone to perform reward modeling on their dataset and model.
+
+Check out a complete flexible example at [`examples/scripts/reward_modeling.py`](https://github.com/huggingface/trl/tree/main/examples/scripts/reward_modeling.py).
+
+## Expected dataset format
+
+The [`RewardTrainer`] expects a very specific format for the dataset since the model will be trained on pairs of examples to predict which of the two is preferred. We provide an example from the [`Anthropic/hh-rlhf`](https://huggingface.co/datasets/Anthropic/hh-rlhf) dataset below:
+
+<div style="text-align: center">
+<img src="https://huggingface.co/datasets/trl-internal-testing/example-images/resolve/main/images/rlhf-antropic-example.png", width="50%">
+</div>
+
+Therefore the final dataset object should contain two 4 entries at least if you use the default [`RewardDataCollatorWithPadding`] data collator. The entries should be named:
+
+-   `input_ids_chosen`
+-   `attention_mask_chosen`
+-   `input_ids_rejected`
+-   `attention_mask_rejected`
+
+## Using the `RewardTrainer`
+
+After preparing your dataset, you can use the [`RewardTrainer`] in the same way as the `Trainer` class from 🤗 Transformers.
+You should pass an `AutoModelForSequenceClassification` model to the [`RewardTrainer`], along with a [`RewardConfig`] which configures the hyperparameters of the training.
+
+### Leveraging 🤗 PEFT to train a reward model
+
+Just pass a `peft_config` in the keyword arguments of [`RewardTrainer`], and the trainer should automatically take care of converting the model into a PEFT model!
+
+```python
+from peft import LoraConfig, TaskType
+from transformers import AutoModelForSequenceClassification, AutoTokenizer
+from trl import RewardTrainer, RewardConfig
+
+model = AutoModelForSequenceClassification.from_pretrained("gpt2")
+peft_config = LoraConfig(
+    task_type=TaskType.SEQ_CLS,
+    inference_mode=False,
+    r=8,
+    lora_alpha=32,
+    lora_dropout=0.1,
+)
+
+...
+
+trainer = RewardTrainer(
+    model=model,
+    args=training_args,
+    tokenizer=tokenizer,
+    train_dataset=dataset,
+    peft_config=peft_config,
+)
+
+trainer.train()
+
+```
+
+### Adding a margin to the loss
+
+As in the [Llama 2 paper](https://huggingface.co/papers/2307.09288), you can add a margin to the loss by adding a `margin` column to the dataset. The reward collator will automatically pass it through and the loss will be computed accordingly.
+
+```python
+def add_margin(row):
+    # Assume you have a score_chosen and score_rejected columns that you want to use to compute the margin
+    return {'margin': row['score_chosen'] - row['score_rejected']}
+
+dataset = dataset.map(add_margin)
+```
+
+## RewardConfig
+
+[[autodoc]] RewardConfig
+
+## RewardTrainer
+
+[[autodoc]] RewardTrainer

docs/source/sentiment_tuning.mdx

@@ -1,35 +1,130 @@
-# Sentiment Examples
+# Sentiment Tuning Examples
 
 The notebooks and scripts in this examples show how to fine-tune a model with a sentiment classifier (such as `lvwerra/distilbert-imdb`).
 
-Here's an overview of the notebooks and scripts in the [trl repository](https://github.com/lvwerra/trl/tree/main/examples):
-
-| File | Description | Colab link |
-|---|---| --- |
-| [`gpt2-sentiment.ipynb`](https://github.com/lvwerra/trl/blob/main/examples/sentiment/notebooks/gpt2-sentiment.ipynb) | Fine-tune GPT2 to generate positive movie reviews. | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/lvwerra/trl/blob/main/examples/sentiment/notebooks/gpt2-sentiment.ipynb)
- |
-| [`gpt2-sentiment-control.ipynb`](https://github.com/lvwerra/trl/blob/main/examples/sentiment/notebooks/gpt2-sentiment-control.ipynb)  | Fine-tune GPT2 to generate movie reviews with controlled sentiment. | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/lvwerra/trl/blob/main/examples/sentiment/notebooks/gpt2-sentiment-control.ipynb)
-  |
-| [`gpt2-sentiment.py`](https://github.com/lvwerra/trl/blob/main/examples/sentiment/scripts/gpt2-sentiment.py) | Same as the notebook, but easier to use to use in mutli-GPU setup. | x | 
-| [`t5-sentiment.py`](https://github.com/lvwerra/trl/blob/main/examples/sentiment/scripts/t5-sentiment.py) | Same as GPT2 script, but for a Seq2Seq model (T5). | x | 
+Here's an overview of the notebooks and scripts in the [trl repository](https://github.com/huggingface/trl/tree/main/examples):
 
 
-## Installation
+
+| File                                                                                           | Description                                                                                                              |
+|------------------------------------------------------------------------------------------------------------|--------------------------------------------------------------------------------------------------------------------------|
+| [`examples/scripts/ppo.py`](https://github.com/huggingface/trl/blob/main/examples/scripts/ppo.py)  [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/trl/blob/main/examples/sentiment/notebooks/gpt2-sentiment.ipynb) | This script shows how to use the `PPOTrainer` to fine-tune a sentiment analysis model using IMDB dataset                 |
+| [`examples/notebooks/gpt2-sentiment.ipynb`](https://github.com/huggingface/trl/tree/main/examples/notebooks/gpt2-sentiment.ipynb)              | This notebook demonstrates how to reproduce the GPT2 imdb sentiment tuning example on a jupyter notebook.                |
+| [`examples/notebooks/gpt2-control.ipynb`](https://github.com/huggingface/trl/tree/main/examples/notebooks/gpt2-control.ipynb)   [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/huggingface/trl/blob/main/examples/sentiment/notebooks/gpt2-sentiment-control.ipynb)                | This notebook demonstrates how to reproduce the GPT2 sentiment control example on a jupyter notebook.    
+
+
+
+## Usage
 
 ```bash
-pip install trl
-#optional: wandb
-pip install wandb
+# 1. run directly
+python examples/scripts/ppo.py
+# 2. run via `accelerate` (recommended), enabling more features (e.g., multiple GPUs, deepspeed)
+accelerate config # will prompt you to define the training configuration
+accelerate launch examples/scripts/ppo.py # launches training
+# 3. get help text and documentation
+python examples/scripts/ppo.py --help
+# 4. configure logging with wandb and, say, mini_batch_size=1 and gradient_accumulation_steps=16
+python examples/scripts/ppo.py --ppo_config.log_with wandb --ppo_config.mini_batch_size 1 --ppo_config.gradient_accumulation_steps 16
 ```
 
 Note: if you don't want to log with `wandb` remove `log_with="wandb"` in the scripts/notebooks. You can also replace it with your favourite experiment tracker that's [supported by `accelerate`](https://huggingface.co/docs/accelerate/usage_guides/tracking).
 
 
-## Launch scripts
+## Few notes on multi-GPU 
 
-The `trl` library is powered by `accelerate`. As such it is best to configure and launch trainings with the following commands:
+To run in multi-GPU setup with DDP (distributed Data Parallel) change the `device_map` value to `device_map={"": Accelerator().process_index}` and make sure to run your script with `accelerate launch yourscript.py`. If you want to apply naive pipeline parallelism you can use `device_map="auto"`.
+
+
+## Benchmarks
+
+Below are some benchmark results for `examples/scripts/ppo.py`. To reproduce locally, please check out the `--command` arguments below.
 
 ```bash
-accelerate config # will prompt you to define the training configuration
-accelerate launch scripts/gpt2-sentiment.py # launches training
-```
+python benchmark/benchmark.py \
+    --command "python examples/scripts/ppo.py --ppo_config.log_with wandb" \
+    --num-seeds 5 \
+    --start-seed 1 \
+    --workers 10 \
+    --slurm-nodes 1 \
+    --slurm-gpus-per-task 1 \
+    --slurm-ntasks 1 \
+    --slurm-total-cpus 12 \
+    --slurm-template-path benchmark/trl.slurm_template
+```
+
+![](https://huggingface.co/datasets/trl-internal-testing/example-images/resolve/main/images/benchmark/v0.4.7-55-g110e672/sentiment.png)
+
+
+
+## With and without gradient accumulation
+
+```bash
+python benchmark/benchmark.py \
+    --command "python examples/scripts/ppo.py --ppo_config.exp_name sentiment_tuning_step_grad_accu --ppo_config.mini_batch_size 1 --ppo_config.gradient_accumulation_steps 128 --ppo_config.log_with wandb" \
+    --num-seeds 5 \
+    --start-seed 1 \
+    --workers 10 \
+    --slurm-nodes 1 \
+    --slurm-gpus-per-task 1 \
+    --slurm-ntasks 1 \
+    --slurm-total-cpus 12 \
+    --slurm-template-path benchmark/trl.slurm_template
+```
+
+![](https://huggingface.co/datasets/trl-internal-testing/example-images/resolve/main/images/benchmark/v0.4.7-55-g110e672/gradient_accu.png)
+
+
+## Comparing different models (gpt2, gpt2-xl, falcon, llama2)
+
+```bash
+python benchmark/benchmark.py \
+    --command "python examples/scripts/ppo.py --ppo_config.exp_name sentiment_tuning_gpt2 --ppo_config.log_with wandb" \
+    --num-seeds 5 \
+    --start-seed 1 \
+    --workers 10 \
+    --slurm-nodes 1 \
+    --slurm-gpus-per-task 1 \
+    --slurm-ntasks 1 \
+    --slurm-total-cpus 12 \
+    --slurm-template-path benchmark/trl.slurm_template
+python benchmark/benchmark.py \
+    --command "python examples/scripts/ppo.py --ppo_config.exp_name sentiment_tuning_gpt2xl_grad_accu --ppo_config.model_name gpt2-xl --ppo_config.mini_batch_size 16 --ppo_config.gradient_accumulation_steps 8 --ppo_config.log_with wandb" \
+    --num-seeds 5 \
+    --start-seed 1 \
+    --workers 10 \
+    --slurm-nodes 1 \
+    --slurm-gpus-per-task 1 \
+    --slurm-ntasks 1 \
+    --slurm-total-cpus 12 \
+    --slurm-template-path benchmark/trl.slurm_template
+python benchmark/benchmark.py \
+    --command "python examples/scripts/ppo.py --ppo_config.exp_name sentiment_tuning_falcon_rw_1b --ppo_config.model_name tiiuae/falcon-rw-1b --ppo_config.log_with wandb" \
+    --num-seeds 5 \
+    --start-seed 1 \
+    --workers 10 \
+    --slurm-nodes 1 \
+    --slurm-gpus-per-task 1 \
+    --slurm-ntasks 1 \
+    --slurm-total-cpus 12 \
+    --slurm-template-path benchmark/trl.slurm_template
+```
+
+![](https://huggingface.co/datasets/trl-internal-testing/example-images/resolve/main/images/benchmark/v0.4.7-55-g110e672/different_models.png)
+
+## With and without PEFT
+
+```
+python benchmark/benchmark.py \
+    --command "python examples/scripts/ppo.py --ppo_config.exp_name sentiment_tuning_peft --use_peft --ppo_config.log_with wandb" \
+    --num-seeds 5 \
+    --start-seed 1 \
+    --workers 10 \
+    --slurm-nodes 1 \
+    --slurm-gpus-per-task 1 \
+    --slurm-ntasks 1 \
+    --slurm-total-cpus 12 \
+    --slurm-template-path benchmark/trl.slurm_template
+```
+
+![](https://huggingface.co/datasets/trl-internal-testing/example-images/resolve/main/images/benchmark/v0.4.7-55-g110e672/peft.png)

docs/source/sft_trainer.mdx

@@ -0,0 +1,483 @@
+# Supervised Fine-tuning Trainer 
+
+Supervised fine-tuning (or SFT for short) is a crucial step in RLHF. In TRL we provide an easy-to-use API to create your SFT models and train them with few lines of code on your dataset.
+
+Check out a complete flexible example at [`examples/scripts/sft.py`](https://github.com/huggingface/trl/tree/main/examples/scripts/sft.py).
+
+## Quickstart
+
+If you have a dataset hosted on the 🤗 Hub, you can easily fine-tune your SFT model using [`SFTTrainer`] from TRL. Let us assume your dataset is `imdb`, the text you want to predict is inside the `text` field of the dataset, and you want to fine-tune the `facebook/opt-350m` model. 
+The following code-snippet takes care of all the data pre-processing and training for you:
+
+```python
+from datasets import load_dataset
+from trl import SFTTrainer
+
+dataset = load_dataset("imdb", split="train")
+
+trainer = SFTTrainer(
+    "facebook/opt-350m",
+    train_dataset=dataset,
+    dataset_text_field="text",
+    max_seq_length=512,
+)
+trainer.train()
+```
+Make sure to pass a correct value for `max_seq_length` as the default value will be set to `min(tokenizer.model_max_length, 1024)`.
+
+You can also construct a model outside of the trainer and pass it as follows:
+
+```python
+from transformers import AutoModelForCausalLM
+from datasets import load_dataset
+from trl import SFTTrainer
+
+dataset = load_dataset("imdb", split="train")
+
+model = AutoModelForCausalLM.from_pretrained("facebook/opt-350m")
+
+trainer = SFTTrainer(
+    model,
+    train_dataset=dataset,
+    dataset_text_field="text",
+    max_seq_length=512,
+)
+
+trainer.train()
+```
+
+The above snippets will use the default training arguments from the [`transformers.TrainingArguments`](https://huggingface.co/docs/transformers/main_classes/trainer#transformers.TrainingArguments) class. If you want to modify that, make sure to create your own `TrainingArguments` object and pass it to the [`SFTTrainer`] constructor as it is done on the [`supervised_finetuning.py` script](https://github.com/huggingface/trl/blob/main/examples/stack_llama/scripts/supervised_finetuning.py) on the stack-llama example.
+
+## Advanced usage
+
+### Train on completions only 
+
+You can use the `DataCollatorForCompletionOnlyLM` to train your model on the generated prompts only. Note that this works only in the case when `packing=False`.
+To instantiate that collator for instruction data, pass a response template and the tokenizer. Here is an example of how it would work to fine-tune `opt-350m` on completions only on the CodeAlpaca dataset:
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+from datasets import load_dataset
+from trl import SFTTrainer, DataCollatorForCompletionOnlyLM
+
+dataset = load_dataset("lucasmccabe-lmi/CodeAlpaca-20k", split="train")
+
+model = AutoModelForCausalLM.from_pretrained("facebook/opt-350m")
+tokenizer = AutoTokenizer.from_pretrained("facebook/opt-350m")
+
+def formatting_prompts_func(example):
+    output_texts = []
+    for i in range(len(example['instruction'])):
+        text = f"### Question: {example['instruction'][i]}\n ### Answer: {example['output'][i]}"
+        output_texts.append(text)
+    return output_texts
+
+response_template = " ### Answer:"
+collator = DataCollatorForCompletionOnlyLM(response_template, tokenizer=tokenizer)
+
+trainer = SFTTrainer(
+    model,
+    train_dataset=dataset,
+    formatting_func=formatting_prompts_func,
+    data_collator=collator,
+)
+
+trainer.train() 
+```
+
+To instantiate that collator for assistant style conversation data, pass a response template, an instruction template and the tokenizer. Here is an example of how it would work to fine-tune `opt-350m` on assistant completions only on the Open Assistant Guanaco dataset:
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+from datasets import load_dataset
+from trl import SFTTrainer, DataCollatorForCompletionOnlyLM
+
+dataset = load_dataset("timdettmers/openassistant-guanaco", split="train")
+
+model = AutoModelForCausalLM.from_pretrained("facebook/opt-350m")
+tokenizer = AutoTokenizer.from_pretrained("facebook/opt-350m")
+
+instruction_template = "### Human:"
+response_template = "### Assistant:"
+collator = DataCollatorForCompletionOnlyLM(instruction_template=instruction_template, response_template=response_template, tokenizer=tokenizer, mlm=False)
+
+trainer = SFTTrainer(
+    model,
+    train_dataset=dataset,
+    dataset_text_field="text",
+    data_collator=collator,
+)
+
+trainer.train() 
+```
+
+Make sure to have a `pad_token_id` which is different from `eos_token_id` which can result in the model not properly predicting EOS (End of Sentence) tokens during generation. 
+
+#### Using token_ids directly for `response_template`
+
+Some tokenizers like Llama 2 (`meta-llama/Llama-2-XXb-hf`) tokenize sequences differently depending whether they have context or not. For example:
+    
+```python
+from transformers import AutoTokenizer
+tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-2-7b-hf")
+
+def print_tokens_with_ids(txt):
+    tokens = tokenizer.tokenize(txt, add_special_tokens=False)
+    token_ids = tokenizer.encode(txt, add_special_tokens=False)
+    print(list(zip(tokens, token_ids)))
+
+prompt = """### User: Hello\n\n### Assistant: Hi, how can I help you?"""
+print_tokens_with_ids(prompt)  # [..., ('▁Hello', 15043), ('<0x0A>', 13), ('<0x0A>', 13), ('##', 2277), ('#', 29937), ('▁Ass', 4007), ('istant', 22137), (':', 29901), ...]
+
+response_template = "### Assistant:"
+print_tokens_with_ids(response_template)  # [('▁###', 835), ('▁Ass', 4007), ('istant', 22137), (':', 29901)]
+```
+
+In this case, and due to lack of context in `response_template`, the same string ("### Assistant:") is tokenized differently:
+    
+    - Text (with context): `[2277, 29937, 4007, 22137, 29901]`
+    - `response_template` (without context): `[835, 4007, 22137, 29901]`
+
+This will lead to an error when the `DataCollatorForCompletionOnlyLM` does not find the `response_template` in the dataset example text:
+
+```
+RuntimeError: Could not find response key [835, 4007, 22137, 29901] in token IDs tensor([    1,   835,  ...])   
+```
+
+
+To solve this, you can tokenize the `response_template` with the same context than in the dataset, truncate it as needed and pass the `token_ids` directly to the `response_template` argument of the `DataCollatorForCompletionOnlyLM` class. For example:
+
+```python
+response_template_with_context = "\n### Assistant:"  # We added context here: "\n". This is enough for this tokenizer
+response_template_ids = tokenizer.encode(response_template_with_context, add_special_tokens=False)[2:]  # Now we have it like in the dataset texts: `[2277, 29937, 4007, 22137, 29901]`
+
+data_collator = DataCollatorForCompletionOnlyLM(response_template_ids, tokenizer=tokenizer)
+```
+
+### Format your input prompts
+
+For instruction fine-tuning, it is quite common to have two columns inside the dataset: one for the prompt & the other for the response. 
+This allows people to format examples like [Stanford-Alpaca](https://github.com/tatsu-lab/stanford_alpaca) did as follows:
+```bash
+Below is an instruction ...
+
+### Instruction
+{prompt}
+
+### Response:
+{completion}
+```
+Let us assume your dataset has two fields, `question` and `answer`. Therefore you can just run:
+```python
+...
+def formatting_prompts_func(example):
+    output_texts = []
+    for i in range(len(example['question'])):
+        text = f"### Question: {example['question'][i]}\n ### Answer: {example['answer'][i]}"
+        output_texts.append(text)
+    return output_texts
+
+trainer = SFTTrainer(
+    model,
+    train_dataset=dataset,
+    formatting_func=formatting_prompts_func,
+)
+
+trainer.train()
+```
+To preperly format your input make sure to process all the examples by looping over them and returning a list of processed text. Check out a full example on how to use SFTTrainer on alpaca dataset [here](https://github.com/huggingface/trl/pull/444#issue-1760952763)
+
+### Packing dataset ([`ConstantLengthDataset`])
+
+[`SFTTrainer`] supports _example packing_, where multiple short examples are packed in the same input sequence to increase training efficiency. This is done with the [`ConstantLengthDataset`] utility class that returns constant length chunks of tokens from a stream of examples. To enable the usage of this dataset class, simply pass `packing=True` to the [`SFTTrainer`] constructor.
+
+```python
+...
+
+trainer = SFTTrainer(
+    "facebook/opt-350m",
+    train_dataset=dataset,
+    dataset_text_field="text",
+    packing=True
+)
+
+trainer.train()
+```
+
+Note that if you use a packed dataset and if you pass `max_steps` in the training arguments you will probably train your models for more than few epochs, depending on the way you have configured the packed dataset and the training protocol. Double check that you know and understand what you are doing. 
+
+#### Customize your prompts using packed dataset
+
+If your dataset has several fields that you want to combine, for example if the dataset has `question` and `answer` fields and you want to combine them, you can pass a formatting function to the trainer that will take care of that. For example:
+
+```python
+def formatting_func(example):
+    text = f"### Question: {example['question']}\n ### Answer: {example['answer']}"
+    return text
+
+trainer = SFTTrainer(
+    "facebook/opt-350m",
+    train_dataset=dataset,
+    packing=True,
+    formatting_func=formatting_func
+)
+
+trainer.train()
+```
+You can also customize the [`ConstantLengthDataset`] much more by directly passing the arguments to the [`SFTTrainer`] constructor. Please refer to that class' signature for more information.
+
+### Control over the pretrained model
+
+You can directly pass the kwargs of the `from_pretrained()` method to the [`SFTTrainer`]. For example, if you want to load a model in a different precision, analogous to 
+
+```python
+model = AutoModelForCausalLM.from_pretrained("facebook/opt-350m", torch_dtype=torch.bfloat16)
+```
+
+```python
+...
+
+trainer = SFTTrainer(
+    "facebook/opt-350m",
+    train_dataset=dataset,
+    dataset_text_field="text",
+    model_init_kwargs={
+        "torch_dtype": torch.bfloat16,
+    },
+)
+
+trainer.train()
+```
+Note that all keyword arguments of `from_pretrained()` are supported. 
+
+### Training adapters
+
+We also support a tight integration with 🤗 PEFT library so that any user can conveniently train adapters and share them on the Hub instead of training the entire model
+
+```python
+from datasets import load_dataset
+from trl import SFTTrainer
+from peft import LoraConfig
+
+dataset = load_dataset("imdb", split="train")
+
+peft_config = LoraConfig(
+    r=16,
+    lora_alpha=32,
+    lora_dropout=0.05,
+    bias="none",
+    task_type="CAUSAL_LM",
+)
+
+trainer = SFTTrainer(
+    "EleutherAI/gpt-neo-125m",
+    train_dataset=dataset,
+    dataset_text_field="text",
+    peft_config=peft_config
+)
+
+trainer.train()
+```
+
+You can also continue training your `PeftModel`. For that, first load a `PeftModel` outside `SFTTrainer` and pass it directly to the trainer without the `peft_config` argument being passed.
+
+### Training adapters with base 8 bit models 
+
+For that you need to first load your 8bit model outside the Trainer and pass a `PeftConfig` to the trainer. For example:
+
+```python
+...
+
+peft_config = LoraConfig(
+    r=16,
+    lora_alpha=32,
+    lora_dropout=0.05,
+    bias="none",
+    task_type="CAUSAL_LM",
+)
+
+model = AutoModelForCausalLM.from_pretrained(
+    "EleutherAI/gpt-neo-125m",
+    load_in_8bit=True,
+    device_map="auto",
+)
+
+trainer = SFTTrainer(
+    model,
+    train_dataset=dataset,
+    dataset_text_field="text",
+    peft_config=peft_config,
+)
+
+trainer.train()
+```
+
+## Using Flash Attention and Flash Attention 2
+
+You can benefit from Flash Attention 1 & 2 using SFTTrainer out of the box with minimal changes of code. 
+First, to make sure you have all the latest features from transformers, install transformers from source
+
+```bash
+pip install -U git+https://github.com/huggingface/transformers.git
+```
+
+Note that Flash Attention only works on GPU now and under half-precision regime (when using adapters, base model loaded in half-precision)
+Note also both features are perfectly compatible with other tools such as quantization.
+
+### Using Flash-Attention 1
+
+For Flash Attention 1 you can use the `BetterTransformer` API and force-dispatch the API to use Flash Attention kernel. First, install the latest optimum package:
+
+```bash
+pip install -U optimum
+```
+
+Once you have loaded your model, wrap the `trainer.train()` call under the `with torch.backends.cuda.sdp_kernel(enable_flash=True, enable_math=False, enable_mem_efficient=False):` context manager:
+
+```diff
+...
+
++ with torch.backends.cuda.sdp_kernel(enable_flash=True, enable_math=False, enable_mem_efficient=False):
+    trainer.train()
+```
+
+Note that you cannot train your model using Flash Attention 1 on an arbitrary dataset as `torch.scaled_dot_product_attention` does not support training with padding tokens if you use Flash Attention kernels. Therefore you can only use that feature with `packing=True`. If your dataset contains padding tokens, consider switching to Flash Attention 2 integration.
+
+Below are some numbers you can get in terms of speedup and memory efficiency, using Flash Attention 1, on a single NVIDIA-T4 16GB.
+
+| use_flash_attn_1 | model_name        | max_seq_len | batch_size | time per training step |
+|----------------|-------------------|-------------|------------|------------------------|
+| x              | facebook/opt-350m | 2048        | 8          | ~59.1s                 |
+|                | facebook/opt-350m | 2048        | 8          | **OOM**                |
+| x              | facebook/opt-350m | 2048        | 4          | ~30.3s                 |
+|                | facebook/opt-350m | 2048        | 4          | ~148.9s                |
+
+### Using Flash Attention-2
+
+To use Flash Attention 2, first install the latest `flash-attn` package:
+
+```bash
+pip install -U flash-attn
+```
+
+And add `use_flash_attention_2=True` when calling `from_pretrained`:
+
+```python
+model = AutoModelForCausalLM.from_pretrained(
+    model_id,
+    load_in_4bit=True,
+    use_flash_attention_2=True
+)
+```
+
+If you don't use quantization, make sure your model is loaded in half-precision and dispatch your model on a supported GPU device.
+After loading your model, you can either train it as it is, or attach adapters and train adapters on it in case your model is quantized.
+
+In contrary to Flash Attention 1, the integration makes it possible to train your model on an arbitrary dataset that also includes padding tokens.
+
+### Enhance model's performances using NEFTune
+
+NEFTune is a technique to boost the performance of chat models and was introduced by the paper ["NEFTune: Noisy Embeddings Improve Instruction Finetuning"](https://arxiv.org/abs/2310.05914) from Jain et al. it consists of adding noise to the embedding vectors during training. According to the abstract of the paper:
+
+>  Standard finetuning of LLaMA-2-7B using Alpaca achieves 29.79% on AlpacaEval, which rises to 64.69% using noisy embeddings. NEFTune also improves over strong baselines on modern instruction datasets. Models trained with Evol-Instruct see a 10% improvement, with ShareGPT an 8% improvement, and with OpenPlatypus an 8% improvement. Even powerful models further refined with RLHF such as LLaMA-2-Chat benefit from additional training with NEFTune.
+
+<div style="text-align: center">
+<img src="https://huggingface.co/datasets/trl-internal-testing/example-images/resolve/main/images/neft-screenshot.png">
+</div>
+
+To use it in `SFTTrainer` simply pass `neftune_noise_alpha` when creating your `SFTTrainer` instance. Note that to avoid any surprising behaviour, NEFTune is disabled after training to retrieve back the original behaviour of the embedding layer.
+
+```python
+from datasets import load_dataset
+from trl import SFTTrainer
+
+dataset = load_dataset("imdb", split="train")
+
+trainer = SFTTrainer(
+    "facebook/opt-350m",
+    train_dataset=dataset,
+    dataset_text_field="text",
+    max_seq_length=512,
+    neftune_noise_alpha=5,
+)
+trainer.train()
+```
+
+We have tested NEFTune by training `mistralai/Mistral-7B-v0.1` on the [OpenAssistant dataset](https://huggingface.co/datasets/timdettmers/openassistant-guanaco) and validated that using NEFTune led to a performance boost of ~25% on MT Bench.
+
+<div style="text-align: center">
+<img src="https://huggingface.co/datasets/trl-internal-testing/example-images/resolve/main/images/trl-neftune-mistral-7b.png">
+</div>
+
+Note however, that the amount of performance gain is _dataset dependent_ and in particular, applying NEFTune on synthetic datasets like [UltraChat](https://huggingface.co/datasets/stingning/ultrachat) typically produces smaller gains.
+
+### Accelerate fine-tuning 2x using `unsloth`
+
+You can further accelerate QLoRA / LoRA (2x faster, 60% less memory) and even full-finetuning (1.1x faster) using the [`unsloth`](https://github.com/unslothai/unsloth) library that is compatible with `SFTTrainer`. Currently `unsloth` supports only Llama (Yi, TinyLlama as well) and Mistral architectures.
+First install `unsloth` according to the [official documentation](https://github.com/unslothai/unsloth#installation-instructions---conda). Once installed, you can incorporate unsloth into your workflow in a very simple manner; instead of loading `AutoModelForCausalLM`, you just need to load a `FastLlamaModel` or `FastMistralModel` as follows:
+
+```python
+import torch
+
+from transformers import TrainingArguments
+from trl import SFTTrainer
+from unsloth import FastLlamaModel, FastMistralModel
+
+max_seq_length = 2048 # Supports automatic RoPE Scaling, so choose any number.
+dtype = None # None for auto detection. Float16 for Tesla T4, V100, Bfloat16 for Ampere+
+load_in_4bit = True # Use 4bit quantization to reduce memory usage. Can be False.
+
+# Load Llama model
+model, tokenizer = FastLlamaModel.from_pretrained(
+    model_name = "unsloth/llama-2-7b", # Supports any llama model eg meta-llama/Llama-2-7b-hf
+    max_seq_length = max_seq_length,
+    dtype = dtype,
+    load_in_4bit = load_in_4bit,
+    # token = "hf_...", # use one if using gated models like meta-llama/Llama-2-7b-hf
+)
+
+# Do model patching and add fast LoRA weights
+model = FastLlamaModel.get_peft_model(
+    model,
+    r = 16,
+    target_modules = ["q_proj", "k_proj", "v_proj", "o_proj",
+                      "gate_proj", "up_proj", "down_proj",],
+    lora_alpha = 16,
+    lora_dropout = 0, # Currently only supports dropout = 0
+    bias = "none",    # Currently only supports bias = "none"
+    use_gradient_checkpointing = True,
+    random_state = 3407,
+    max_seq_length = max_seq_length,
+)
+
+args = TrainingArguments(output_dir="./output")
+
+trainer = SFTTrainer(
+    model = model,
+    args = args,
+    train_dataset = dataset,
+    dataset_text_field = "text",
+    max_seq_length = max_seq_length,
+)
+
+trainer.train()
+```
+
+The saved model is fully compatible with Hugging Face's transformers library. Learn more about unsloth in their [official repository](https://github.com/unslothai/unsloth).
+
+## Best practices
+
+Pay attention to the following best practices when training a model with that trainer:
+
+- [`SFTTrainer`] always pads by default the sequences to the `max_seq_length` argument of the [`SFTTrainer`]. If none is passed, the trainer will retrieve that value from the tokenizer. Some tokenizers do not provide default value, so there is a check to retrieve the minimum between 2048 and that value. Make sure to check it before training.
+- For training adapters in 8bit, you might need to tweak the arguments of the `prepare_model_for_kbit_training` method from PEFT, hence we advise users to use `prepare_in_int8_kwargs` field, or create the `PeftModel` outside the [`SFTTrainer`] and pass it.
+- For a more memory-efficient training using adapters, you can load the base model in 8bit, for that simply add `load_in_8bit` argument when creating the [`SFTTrainer`], or create a base model in 8bit outside the trainer and pass it.  
+- If you create a model outside the trainer, make sure to not pass to the trainer any additional keyword arguments that are relative to `from_pretrained()` method.
+
+## SFTTrainer
+
+[[autodoc]] SFTTrainer
+
+## ConstantLengthDataset
+
+[[autodoc]] trainer.ConstantLengthDataset

docs/source/summarization_reward_tuning.mdx

@@ -1,30 +0,0 @@
-# Summarization Example
-
-The script in this example show how to train a reward model for summarization, following the OpenAI Learning to Summarize from Human Feedback [paper](https://arxiv.org/abs/2009.01325). We've validated that the script can be used to train a small GPT2 to get slightly over 60% validation accuracy, which is aligned with results from the paper. The model is [here](https://huggingface.co/Tristan/gpt2_reward_summarization).
-
-Here's an overview of the relevant files in the [trl repository](https://github.com/lvwerra/trl/tree/main/examples):
-
-| File | Description |
-|---|---|
-| `scripts/reward_summarization.py` | For tuning the reward model. |
-| `scripts/ds3_reward_summarization_example_config.json` | Can be used with the reward model script to scale it up to arbitrarily big models that don't fit on a single GPU. |
-
-
-## Installation
-
-```bash
-pip install trl
-pip install evaluate
-# optional: deepspeed
-pip install deepspeed
-```
-
-```bash
-# If you want your reward model to follow the Learning to Summarize from Human Feedback paper closely, then tune a GPT model on summarization and then instantiate the reward model
-# with it. In other words, pass in the name of your summarization-finetuned gpt on the hub, instead of the name of the pretrained gpt2 like we do in the following examples of how
-# to run this script.
-# Example of running this script with the small size gpt2 on a 40GB A100 (A100's support bf16). Here, the global batch size will be 64:
-python -m torch.distributed.launch --nproc_per_node=1 reward_summarization.py --bf16
-# Example of running this script with the xl size gpt2 on 16 40GB A100's. Here the global batch size will still be 64:
-python -m torch.distributed.launch --nproc_per_node=16 reward_summarization.py --per_device_train_batch_size=1 --per_device_eval_batch_size=1 --gradient_accumulation_steps=4 --gpt_model_name=gpt2-xl --bf16 --deepspeed=ds3_reward_summarization_example_config.json
-```

docs/source/text_environments.md

@@ -0,0 +1,197 @@
+# Text Environments
+
+Text environments provide a learning ground for language agents. It allows a language model to use tools to accomplish a task such as using a Python interpreter to answer math questions or using a search index for trivia questions. Having access to tools allows language models to solve tasks that would be very hard for the models itself but can be trivial for the appropriate tools. A good example is arithmetics of large numbers that become a simple copy-paste task once you have access to a calculator.
+
+<div style="text-align: center">
+<img src="https://huggingface.co/datasets/trl-internal-testing/example-images/resolve/main/images/textenv.png">
+</div>
+
+Let's dive into how text environments work and start with tools!
+
+## Tools
+
+One of the core building blocks of text environments are tools that the model can use to solve tasks. In general tools can be any Python function that takes a string as input and returns string. The `TextEnvironment` offers two options for tools: either go with predefined tools from `transformers.Tool` or define your own function or class with `__call__` method. Let's have a look at both!
+
+### `transformers.Tool`
+
+Text environments fully support tools of the class `transformers.Tool`. The advantage of building tools in that framework is that they can easily be shared 
+
+```Python
+from transformers import load_tool
+
+# simple calculator tool that runs +-/* operations
+calc_tool = load_tool("ybelkada/simple-calculator")
+
+# python interpreter that executes program and returns outputs
+py_tool = load_tool("lvwerra/python-interpreter")
+
+# wikipedia search index that returns best search match
+wiki_tool = load_tool("vwxyzjn/pyserini-wikipedia-kilt-doc")
+```
+
+These tools are either loaded from the hub or from a local folder. Using the tool is as simple as calling them with a text query:
+
+```Python
+calc_tool("1/2")
+>>> "0.5"
+```
+
+Note that both input and return values are strings to enable easy usage with a language model.
+
+### Custom Tools
+
+The following is an example of a tool that adds two integers:
+
+```Python
+def add(text):
+    int_1, int_2 = text.split("+")
+    result = int(int_1) + int(int_2)
+    return str(result)
+
+print(add("1+1"))
+>>> "2"
+```
+
+We looked at basic examples such as a calculator but the principle holds for more complex tools as well such as a web search tool where you input the query and get the search results in return. Now let's look at how the model can use the tools with the call syntax.
+
+### Call syntax
+
+In order to have a unified way for the model to call a tool we created a simple syntax that looks as follows:
+
+```python
+"<request><TOOL_NAME>QUERY<call>TOOL_RESPONSE<response>"
+```
+
+There are a few special tokens involved so let's decompose it: First the model can signal that it wants to use a tool by emitting the `<request>` token. After that we want to know the name of the tool to call which is done by enclosing the tool name with `<>` brackets. Once we know which tool to call the tool query follows which is in free text form. The `<call>` tokens signifies the end of the query and stops the model generation. At this point the model output is parsed and the query sent to the tool. The environment appends the tool response to the string followed by the `<response>` token to show the end the tool output.
+
+Let's look at the concrete example of the calculator and assume its name is `Calculator` (more on how the name of a tool is inferred later):
+
+```python
+"<request><Calculator>1/2<call>0.5<response>"
+```
+
+Finally, the episode is ended and generation stops when the model generates `<submit>` which marks the interaction as completed.
+
+Now let's have a look how we can create a new text environment!
+
+## Create a `TextEnvironment`
+
+
+```python
+prompt = """\
+What is 13-3?
+<request><SimpleCalculatorTool>13-3<call>10.0<response>
+Result=10<submit>
+"""
+
+def reward_fn(result, answer):
+    """Simplified reward function returning 1 if result matches answer and 0 otherwise."""
+    result_parsed = result.split("=")[1].split("<")[0]
+    return int(result_parsed==answer)
+
+text_env = TextEnvironemnt(
+    model=model, 
+    tokenizer=tokenizer,
+    tools= {"SimpleCalculatorTool": load_tool("ybelkada/simple-calculator")},
+    reward_fn=exact_match_reward,
+    prompt=prompt, 
+    max_turns=1
+    max_tool_response=100
+    generation_kwargs={"do_sample": "true"}
+)
+```
+
+Let's decompose the settings:
+
+| Argument           | Description     |
+|:-------------------|:----------------|
+| `model`            | Language model to interact with the environment and generate requests. |
+| `tokenizer`        | Tokenizer of language model handling tokenization of strings. |
+| `tools`            | `list` of `dict` of tools. If former the name of the tool is inferred from class name and otherwise it's the keys of the dictionary.|
+| `reward_fn`        | A function that takes a string as input and returns. Can have extra arguments that are passed to `.run()` such as ground truth.|
+| `prompt`           | Prompt to prepend to every task. Usually a few examples to demonstrate to the model how to use the tools in a few-shot fashion. |
+| `max_turns`        | Maximum number of interactions between model and tools before episode ends.|
+| `max_tool_response`| The tool response is truncated to this number to avoid running out of model context.|
+| `max_length`       |  The maximum number of tokens to allow in an episode. |
+| `generation_kwargs`| Generation settings used by the language model. |
+
+You can customize the environment to your needs and add custom tools and settings. Let's see how you can use the environment to have the model interact with the available tools!
+
+
+## Run an Episode
+
+To run a set of queries through the text environment one can simply use the `run` method.
+
+```python
+queries = ["What is 1/2?"]
+answers = ["0.5"]
+
+queries, responses, masks, rewards, histories = text_env.run(queries, answers=answers)
+```
+
+This will execute the model/tool feedback loop for each query until either no tool is called anymore, the maximum number of turns is reached or to maximum number of tokens in an episode is exceeded. The extra `kwargs` (e.g. `answers=answers` above) passed to `run` will be passed on to the reward function.
+
+There are five objects that are returned by `run`: 
+
+- `queries`: a list of the tokenized queries
+- `responses`: all tokens that have been generated withing the environment including model and tool tokens
+- `masks`: mask that indicates which tokens have been generated by the model and which tokens are generated by the tool
+- `rewards`: a list of reward for each query/response
+- `histories`: list of `TextHistory` objects, which are useful objects containing all the above and also the text equivalents
+
+The masks are crucial for training as we don't want to optimize tokens that the model has not generated which are tokens produced by the tools.
+
+Next, we'll train a PPO step with the generated responses!
+
+
+### Train
+Training on episodes from the `TextEnvironment` is straight forward and simply requires forwarding all the returned variables except the `TextHistory` objects to the `step` method:
+
+```python
+train_stats = ppo_trainer.step(queries, responses, rewards, masks)
+```
+
+## `TextHistory`
+
+The `TextHistory` object stores the interactions between the model and the text environment. It stores tokens and text generated in each turn and their source in each turn (model or system) as well as rewards. Let's go through the class attributes and methods.
+
+### Attributes
+
+The following table summarises the available attributes of the `TextEnvironment` class:
+
+| Attribute           | Description     |
+|:-------------------|:----------------|
+| `text`             | The full string of the text generated in the text environment with both model and system generated text. |
+| `text_spans`       | A list of tuples with the spans for each model or system generated text segment. |
+| `system_spans`     | A list of boolean values indicating if the segment is model or system generated. |
+| `tokens`           | All tokens generated in text environment with both model and system generated tokens. |
+| `token_spans`      | Similar to `text_spans` the `token_spans` indicate the boundaries of model andsystem generated tokens. |
+| `token_masks`      | The token masks can be used to ignore system generated tokens by masking them. |
+| `completed`        | Indicates if the interaction with the environment has completed. |
+| `truncated`        | Indicates if the interaction with the environment has completed because max length was reached. |
+
+With these attributes you can reconstruct every interaction of the model with the `TextEnvironment`. The `TextHistory` also lets you visualize the text history. Let's have a look!
+
+### Visualization
+
+When the model interacts inside the `TextEnvironment` it can be useful to visualize and separate which parts of the text outputs were generated by the model and which parts come from the system and tools. For that purpose there are the two methods [`TextHistory.show_text`] and [`TextHistory.show_tokens`]. They print the text and tokens respectively and highlight the various segments using the [`rich` libray](https://github.com/Textualize/rich) (make sure to install it before using these methods).
+
+You can see that the prompt is highlighted in gray, whereas system segments such as query and tool responses are highlighted in green. All segments generated by the model are highlighted in blue and in addition to the pure text output the reward is displayed as additional text in plum. Here an example of `show_text`:
+
+<div style="text-align: center">
+<img src="https://huggingface.co/datasets/trl-internal-testing/example-images/resolve/main/images/textenv_show_text.png" width=600>
+</div>
+
+Sometimes there can be tricky tokenization related issues that are hidden when showing the decoded text. Thus `TextHistory` also offers an option to display the same highlighting on the tokens directly with `show_tokens`:
+
+<div style="text-align: center">
+<img src="https://huggingface.co/datasets/trl-internal-testing/example-images/resolve/main/images/textenv_show_tokens.png" width=800>
+</div>
+
+Note that you can turn on the colour legend by passing `show_legend=True`.
+
+## API Documentation
+
+[[autodoc]] TextEnvironment
+
+[[autodoc]] TextHistory

docs/source/trainer.mdx

@@ -2,6 +2,7 @@
 
 At TRL we support PPO (Proximal Policy Optimisation) with an implementation that largely follows  the structure introduced in the paper "Fine-Tuning Language Models from Human Preferences" by D. Ziegler et al. [[paper](https://arxiv.org/pdf/1909.08593.pdf), [code](https://github.com/openai/lm-human-preferences)].
 The Trainer and model classes are largely inspired from `transformers.Trainer` and `transformers.AutoModel` classes and adapted for RL.
+We also support a `RewardTrainer` that can be used to train a reward model.
 
 ## PPOConfig
 
@@ -11,6 +12,34 @@ The Trainer and model classes are largely inspired from `transformers.Trainer` a
 
 [[autodoc]] PPOTrainer
 
+## RewardConfig
+
+[[autodoc]] RewardConfig
+
+## RewardTrainer
+
+[[autodoc]] RewardTrainer
+
+## SFTTrainer
+
+[[autodoc]] SFTTrainer
+
+## DPOTrainer
+
+[[autodoc]] DPOTrainer
+
+## DDPOConfig
+
+[[autodoc]] DDPOConfig
+
+## DDPOTrainer
+
+[[autodoc]] DDPOTrainer
+
+## IterativeSFTTrainer
+
+[[autodoc]] IterativeSFTTrainer
+
 ## set_seed
 
 [[autodoc]] set_seed

docs/source/use_model.md

@@ -0,0 +1,58 @@
+# Use model after training
+
+Once you have trained a model using either the SFTTrainer, PPOTrainer, or DPOTrainer, you will have a fine-tuned model that can be used for text generation. In this section, we'll walk through the process of loading the fine-tuned model and generating text. If you need to run an inference server with the trained model, you can explore libraries such as [`text-generation-inference`](https://github.com/huggingface/text-generation-inference).
+
+## Load and Generate
+
+If you have fine-tuned a model fully, meaning without the use of PEFT you can simply load it like any other language model in transformers. E.g. the value head that was trained during the PPO training is no longer needed and if you load the model with the original transformer class it will be ignored:
+
+```python
+from transformers import AutoTokenizer, AutoModelForCausalLM
+
+model_name_or_path = "kashif/stack-llama-2" #path/to/your/model/or/name/on/hub
+device = "cpu" # or "cuda" if you have a GPU
+
+model = AutoModelForCausalLM.from_pretrained(model_name_or_path).to(device)
+tokenizer = AutoTokenizer.from_pretrained(model_name_or_path)
+
+inputs = tokenizer.encode("This movie was really", return_tensors="pt").to(device)
+outputs = model.generate(inputs)
+print(tokenizer.decode(outputs[0]))
+```
+
+Alternatively you can also use the pipeline:
+
+```python
+from transformers import pipeline
+
+model_name_or_path = "kashif/stack-llama-2" #path/to/your/model/or/name/on/hub
+pipe = pipeline("text-generation", model=model_name_or_path)
+print(pipe("This movie was really")[0]["generated_text"])
+```
+
+## Use Adapters PEFT
+
+```python
+from peft import PeftConfig, PeftModel
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+base_model_name = "kashif/stack-llama-2" #path/to/your/model/or/name/on/hub"
+adapter_model_name = "path/to/my/adapter"
+
+model = AutoModelForCausalLM.from_pretrained(base_model_name)
+model = PeftModel.from_pretrained(model, adapter_model_name)
+
+tokenizer = AutoTokenizer.from_pretrained(base_model_name)
+```
+
+You can also merge the adapters into the base model so you can use the model like a normal transformers model, however the checkpoint will be significantly bigger:
+
+```python
+model = AutoModelForCausalLM.from_pretrained(base_model_name)
+model = PeftModel.from_pretrained(model, adapter_model_name)
+
+model = model.merge_and_unload()
+model.save_pretrained("merged_adapters")
+```
+
+Once you have the model loaded and either merged the adapters or keep them separately on top you can run generation as with a normal model outlined above.

docs/source/using_llama_models.mdx

@@ -0,0 +1,160 @@
+# Using LLaMA models with TRL
+
+We've begun rolling out examples to use Meta's LLaMA models in `trl` (see [Meta's LLaMA release](https://ai.facebook.com/blog/large-language-model-llama-meta-ai/) for the original LLaMA model).
+
+## Efficient training strategies
+
+Even training the smallest LLaMA model requires an enormous amount of memory. Some quick math: in bf16, every parameter uses 2 bytes (in fp32 4 bytes) in addition to 8 bytes used, e.g., in the Adam optimizer (see the [performance docs](https://huggingface.co/docs/transformers/perf_train_gpu_one#optimizer) in Transformers for more info). So a 7B parameter model would use `(2+8)*7B=70GB` just to fit in memory and would likely need more when you compute intermediate values such as attention scores. So you couldn’t train the model even on a single 80GB A100 like that. You can use some tricks, like more efficient optimizers of half-precision training, to squeeze a bit more into memory, but you’ll run out sooner or later.
+
+Another option is to use Parameter-Efficient Fine-Tuning (PEFT) techniques, such as the [`peft`](https://github.com/huggingface/peft) library, which can perform low-rank adaptation (LoRA) on a model loaded in 8-bit.
+For more on `peft` + `trl`, see the [docs](https://huggingface.co/docs/trl/sentiment_tuning_peft).
+
+Loading the model in 8bit reduces the memory footprint drastically since you only need one byte per parameter for the weights (e.g. 7B LlaMa is 7GB in memory).
+Instead of training the original weights directly, LoRA adds small adapter layers on top of some specific layers (usually the attention layers); thus, the number of trainable parameters is drastically reduced.
+
+In this scenario, a rule of thumb is to allocate ~1.2-1.4GB per billion parameters (depending on the batch size and sequence length) to fit the entire fine-tuning setup.
+This enables fine-tuning larger models (up to 50-60B scale models on a NVIDIA A100 80GB) at low cost.
+
+Now we can fit very large models into a single GPU, but the training might still be very slow.
+The simplest strategy in this scenario is data parallelism: we replicate the same training setup into separate GPUs and pass different batches to each GPU.
+With this, you can parallelize the forward/backward passes of the model and scale with the number of GPUs.
+
+![chapter10_ddp.png](https://huggingface.co/datasets/trl-internal-testing/example-images/resolve/main/blog/stackllama/chapter10_ddp.png)
+
+We use either the `transformers.Trainer` or `accelerate`, which both support data parallelism without any code changes, by simply passing arguments when calling the scripts with `torchrun` or `accelerate launch`. The following runs a training script with 8 GPUs on a single machine with `accelerate` and `torchrun`, respectively.
+
+```bash
+accelerate launch --multi_gpu --num_machines 1  --num_processes 8 my_accelerate_script.py
+torchrun --nnodes 1  --nproc_per_node 8 my_torch_script.py
+```
+
+## Supervised fine-tuning
+
+Before we start training reward models and tuning our model with RL, it helps if the model is already good in the domain we are interested in.
+In our case, we want it to answer questions, while for other use cases, we might want it to follow instructions, in which case instruction tuning is a great idea.
+The easiest way to achieve this is by continuing to train the language model with the language modeling objective on texts from the domain or task.
+The [StackExchange dataset](https://huggingface.co/datasets/HuggingFaceH4/stack-exchange-preferences) is enormous (over 10 million instructions), so we can easily train the language model on a subset of it.
+
+There is nothing special about fine-tuning the model before doing RLHF - it’s just the causal language modeling objective from pretraining that we apply here.
+To use the data efficiently, we use a technique called packing: instead of having one text per sample in the batch and then padding to either the longest text or the maximal context of the model, we concatenate a lot of texts with a EOS token in between and cut chunks of the context size to fill the batch without any padding.
+
+![chapter10_preprocessing-clm.png](https://huggingface.co/datasets/trl-internal-testing/example-images/resolve/main/blog/stackllama/chapter10_preprocessing-clm.png)
+
+With this approach the training is much more efficient as each token that is passed through the model is also trained in contrast to padding tokens which are usually masked from the loss.
+If you don't have much data and are more concerned about occasionally cutting off some tokens that are overflowing the context you can also use a classical data loader.
+
+The packing is handled by the `ConstantLengthDataset` and we can then use the `Trainer` after loading the model with `peft`. First, we load the model in int8, prepare it for training, and then add the LoRA adapters.
+
+```python
+# load model in 8bit
+model = AutoModelForCausalLM.from_pretrained(
+        args.model_path,
+        load_in_8bit=True,
+        device_map={"": Accelerator().local_process_index}
+    )
+model = prepare_model_for_kbit_training(model)
+
+# add LoRA to model
+lora_config = LoraConfig(
+    r=16,
+    lora_alpha=32,
+    lora_dropout=0.05,
+    bias="none",
+    task_type="CAUSAL_LM",
+)
+
+model = get_peft_model(model, config)
+```
+
+We train the model for a few thousand steps with the causal language modeling objective and save the model.
+Since we will tune the model again with different objectives, we merge the adapter weights with the original model weights.
+
+**Disclaimer:** due to LLaMA's license, we release only the adapter weights for this and the model checkpoints in the following sections.
+You can apply for access to the base model's weights by filling out Meta AI's [form](https://docs.google.com/forms/d/e/1FAIpQLSfqNECQnMkycAp2jP4Z9TFX0cGR4uf7b_fBxjY_OjhJILlKGA/viewform) and then converting them to the 🤗 Transformers format by running this [script](https://github.com/huggingface/transformers/blob/main/src/transformers/models/llama/convert_llama_weights_to_hf.py).
+Note that you'll also need to install 🤗 Transformers from source until the `v4.28` is released.
+
+Now that we have fine-tuned the model for the task, we are ready to train a reward model.
+
+## Reward modeling and human preferences
+
+In principle, we could fine-tune the model using RLHF directly with the human annotations.
+However, this would require us to send some samples to humans for rating after each optimization iteration.
+This is expensive and slow due to the number of training samples needed for convergence and the inherent latency of human reading and annotator speed.
+
+A trick that works well instead of direct feedback is training a reward model on human annotations collected before the RL loop.
+The goal of the reward model is to imitate how a human would rate a text. There are several possible strategies to build a reward model: the most straightforward way would be to predict the annotation (e.g. a rating score or a binary value for “good”/”bad”).
+In practice, what works better is to predict the ranking of two examples, where the reward model is presented with two candidates `(y_k, y_j)` for a given prompt `x` and has to predict which one would be rated higher by a human annotator.
+
+With the StackExchange dataset, we can infer which of the two answers was preferred by the users based on the score.
+With that information and the loss defined above, we can then modify the `transformers.Trainer` by adding a custom loss function.
+
+```python
+class RewardTrainer(Trainer):
+    def compute_loss(self, model, inputs, return_outputs=False):
+        rewards_j = model(input_ids=inputs["input_ids_j"],  attention_mask=inputs["attention_mask_j"])[0]
+        rewards_k = model(input_ids=inputs["input_ids_k"], attention_mask=inputs["attention_mask_k"])[0]
+        loss = -nn.functional.logsigmoid(rewards_j - rewards_k).mean()
+        if return_outputs:
+            return loss, {"rewards_j": rewards_j, "rewards_k": rewards_k}
+        return loss
+```
+
+We utilize a subset of a 100,000 pair of candidates and evaluate on a held-out set of 50,000. With a modest training batch size of 4, we train the Llama model using the LoRA `peft` adapter for a single epoch using the Adam optimizer with BF16 precision. Our LoRA configuration is:
+
+```python
+peft_config = LoraConfig(
+    task_type=TaskType.SEQ_CLS,
+    inference_mode=False,
+    r=8,
+    lora_alpha=32,
+    lora_dropout=0.1,
+)
+```
+As detailed in the next section, the resulting adapter can be merged into the frozen model and saved for further downstream use.
+
+## Reinforcement Learning from Human Feedback
+
+With the fine-tuned language model and the reward model at hand, we are now ready to run the RL loop. It follows roughly three steps:
+
+1. Generate responses from prompts,
+2. Rate the responses with the reward model,
+3. Run a reinforcement learning policy-optimization step with the ratings.
+
+The Query and Response prompts are templated as follows before being tokenized and passed to the model:
+
+```bash
+Question: <Query>
+
+Answer: <Response>
+```
+
+The same template was used for SFT, RM and RLHF stages.
+Once more, we utilize `peft` for memory-efficient training, which offers an extra advantage in the RLHF context.
+Here, the reference model and policy share the same base, the SFT model, which we load in 8-bit and freeze during training.
+We exclusively optimize the policy's LoRA weights using PPO while sharing the base model's weights.
+
+```python
+for epoch, batch in tqdm(enumerate(ppo_trainer.dataloader)):
+    question_tensors = batch["input_ids"]
+
+	# sample from the policy and to generate responses
+    response_tensors = ppo_trainer.generate(
+        question_tensors,
+        return_prompt=False,
+        length_sampler=output_length_sampler,
+        **generation_kwargs,
+    )
+    batch["response"] = tokenizer.batch_decode(response_tensors, skip_special_tokens=True)
+
+    # Compute sentiment score
+    texts = [q + r for q, r in zip(batch["query"], batch["response"])]
+    pipe_outputs = sentiment_pipe(texts, **sent_kwargs)
+    rewards = [torch.tensor(output[0]["score"] - script_args.reward_baseline) for output in pipe_outputs]
+
+    # Run PPO step
+    stats = ppo_trainer.step(question_tensors, response_tensors, rewards)
+	# Log stats to Wandb
+    ppo_trainer.log_stats(stats, batch, rewards)
+```
+
+For the rest of the details and evaluation, please refer to our [blog post on StackLLaMA](https://huggingface.co/blog/stackllama).

examples/README.md

@@ -1,66 +1,3 @@
-# Sentiment Examples
+# Examples
 
-The notebooks and scripts in this examples show how to fine-tune a model with a sentiment classifier (such as `lvwerra/distilbert-imdb`).
-
-Here's an overview of the notebooks and scripts:
-
-| File | Description |
-|---|---|
-| `notebooks/gpt2-sentiment.ipynb`  | Fine-tune GPT2 to generate positive movie reviews. |
-| `notebooks/gpt2-sentiment-control.ipynb`  | Fine-tune GPT2 to generate movie reviews with controlled sentiment. |
-| `scripts/gpt2-sentiment.py` | Same as the notebook, but easier to use to use in mutli-GPU setup. |
-| `scripts/t5-sentiment.py` | Same as GPT2 script, but for a Seq2Seq model (T5). |
-
-
-## Installation
-
-```bash
-pip install trl
-#optional: wandb
-pip install wandb
-```
-
-Note: if you don't want to log with `wandb` remove `log_with="wandb"` in the scripts/notebooks. You can also replace it with your favourite experiment tracker that's [supported by `accelerate`](https://huggingface.co/docs/accelerate/usage_guides/tracking).
-
-
-## Launch scripts
-
-The `trl` library is powered by `accelerate`. As such it is best to configure and launch trainings with the following commands:
-
-```bash
-accelerate config # will prompt you to define the training configuration
-accelerate launch scripts/gpt2-sentiment.py # launches training
-```
-
-# Summarization Example
-  
-The script in this example show how to train a reward model for summarization, following the OpenAI Learning to Summarize from Human Feedback [paper](https://arxiv.org/abs/2009.01325). We've validated that the script can be used to train a small GPT2 to get slightly over 60% validation accuracy, which is aligned with results from the paper. The model is [here](https://huggingface.co/Tristan/gpt2_reward_summarization).
-
-Here's an overview of the files:
-
-| File | Description |
-|---|---|
-| `scripts/reward_summarization.py` | For tuning the reward model. |
-| `scripts/ds3_reward_summarization_example_config.json` | Can be used with the reward model script to scale it up to arbitrarily big models that don't fit on a single GPU. |
-
-
-## Installation
-
-```bash
-pip install trl
-pip install evaluate
-# optional: deepspeed
-pip install deepspeed
-```
-
-```bash
-# If you want your reward model to follow the Learning to Summarize from Human Feedback paper closely, then tune a GPT model on summarization and then instantiate the reward model
-# with it. In other words, pass in the name of your summarization-finetuned gpt on the hub, instead of the name of the pretrained gpt2 like we do in the following examples of how
-# to run this script.
-
-# Example of running this script with the small size gpt2 on a 40GB A100 (A100's support bf16). Here, the global batch size will be 64:
-python -m torch.distributed.launch --nproc_per_node=1 reward_summarization.py --bf16
-
-# Example of running this script with the xl size gpt2 on 16 40GB A100's. Here the global batch size will still be 64:
-python -m torch.distributed.launch --nproc_per_node=16 reward_summarization.py --per_device_train_batch_size=1 --per_device_eval_batch_size=1 --gradient_accumulation_steps=4 --gpt_model_name=gpt2-xl --bf16 --deepspeed=ds3_reward_summarization_example_config.json
-```
+Please check out https://huggingface.co/docs/trl/example_overview for documentation on our examples.

examples/accelerate_configs/deepspeed_zero1.yaml

@@ -0,0 +1,20 @@
+compute_environment: LOCAL_MACHINE
+debug: false
+deepspeed_config:
+  deepspeed_multinode_launcher: standard
+  gradient_accumulation_steps: 1
+  zero3_init_flag: false
+  zero_stage: 1
+distributed_type: DEEPSPEED
+downcast_bf16: 'no'
+machine_rank: 0
+main_training_function: main
+mixed_precision: 'bf16'
+num_machines: 1
+num_processes: 8
+rdzv_backend: static
+same_network: true
+tpu_env: []
+tpu_use_cluster: false
+tpu_use_sudo: false
+use_cpu: false

examples/accelerate_configs/deepspeed_zero2.yaml

@@ -0,0 +1,22 @@
+compute_environment: LOCAL_MACHINE
+debug: false
+deepspeed_config:
+  deepspeed_multinode_launcher: standard
+  gradient_accumulation_steps: 1
+  offload_optimizer_device: none
+  offload_param_device: none
+  zero3_init_flag: false
+  zero_stage: 2
+distributed_type: DEEPSPEED
+downcast_bf16: 'no'
+machine_rank: 0
+main_training_function: main
+mixed_precision: 'bf16'
+num_machines: 1
+num_processes: 8
+rdzv_backend: static
+same_network: true
+tpu_env: []
+tpu_use_cluster: false
+tpu_use_sudo: false
+use_cpu: false

examples/accelerate_configs/deepspeed_zero3.yaml

@@ -0,0 +1,23 @@
+compute_environment: LOCAL_MACHINE
+debug: false
+deepspeed_config:
+  deepspeed_multinode_launcher: standard
+  gradient_accumulation_steps: 1
+  offload_optimizer_device: none
+  offload_param_device: none
+  zero3_init_flag: true
+  zero3_save_16bit_model: true
+  zero_stage: 3
+distributed_type: DEEPSPEED
+downcast_bf16: 'no'
+machine_rank: 0
+main_training_function: main
+mixed_precision: 'bf16'
+num_machines: 1
+num_processes: 8
+rdzv_backend: static
+same_network: true
+tpu_env: []
+tpu_use_cluster: false
+tpu_use_sudo: false
+use_cpu: false

examples/accelerate_configs/multi_gpu.yaml

@@ -0,0 +1,16 @@
+compute_environment: LOCAL_MACHINE
+debug: false
+distributed_type: MULTI_GPU
+downcast_bf16: 'no'
+gpu_ids: all
+machine_rank: 0
+main_training_function: main
+mixed_precision: 'bf16'
+num_machines: 1
+num_processes: 8
+rdzv_backend: static
+same_network: true
+tpu_env: []
+tpu_use_cluster: false
+tpu_use_sudo: false
+use_cpu: false

examples/hello_world.py

@@ -0,0 +1,40 @@
+# 0. imports
+import torch
+from transformers import GPT2Tokenizer
+
+from trl import AutoModelForCausalLMWithValueHead, PPOConfig, PPOTrainer
+
+
+# 1. load a pretrained model
+model = AutoModelForCausalLMWithValueHead.from_pretrained("gpt2")
+model_ref = AutoModelForCausalLMWithValueHead.from_pretrained("gpt2")
+tokenizer = GPT2Tokenizer.from_pretrained("gpt2")
+tokenizer.pad_token = tokenizer.eos_token
+
+# 2. initialize trainer
+ppo_config = {"batch_size": 1}
+config = PPOConfig(**ppo_config)
+ppo_trainer = PPOTrainer(config, model, model_ref, tokenizer)
+
+# 3. encode a query
+query_txt = "This morning I went to the "
+query_tensor = tokenizer.encode(query_txt, return_tensors="pt").to(model.pretrained_model.device)
+
+# 4. generate model response
+generation_kwargs = {
+    "min_length": -1,
+    "top_k": 0.0,
+    "top_p": 1.0,
+    "do_sample": True,
+    "pad_token_id": tokenizer.eos_token_id,
+    "max_new_tokens": 20,
+}
+response_tensor = ppo_trainer.generate([item for item in query_tensor], return_prompt=False, **generation_kwargs)
+response_txt = tokenizer.decode(response_tensor[0])
+
+# 5. define a reward for response
+# (this could be any reward such as human feedback or output from another model)
+reward = [torch.tensor(1.0, device=model.pretrained_model.device)]
+
+# 6. train model with ppo
+train_stats = ppo_trainer.step([query_tensor[0]], [response_tensor[0]], reward)

examples/notebooks/README.md

@@ -0,0 +1,7 @@
+# Notebooks
+
+This directory contains a collection of Jupyter notebooks that demonstrate how to use the TRL library in different applications.
+
+- [`best_of_n.ipynb`](https://github.com/huggingface/trl/tree/main/examples/notebooks/best_of_n.ipynb): This notebook demonstrates how to use the "Best of N" sampling strategy using TRL when fine-tuning your model with PPO.
+- [`gpt2-sentiment.ipynb`](https://github.com/huggingface/trl/tree/main/examples/notebooks/gpt2-sentiment.ipynb): This notebook demonstrates how to reproduce the GPT2 imdb sentiment tuning example on a jupyter notebook.
+- [`gpt2-control.ipynb`](https://github.com/huggingface/trl/tree/main/examples/notebooks/gpt2-sentiment-control.ipynb): This notebook demonstrates how to reproduce the GPT2 sentiment control example on a jupyter notebook.

examples/notebooks/best_of_n.ipynb

@@ -0,0 +1,648 @@
+{
+ "nbformat": 4,
+ "nbformat_minor": 0,
+ "metadata": {
+  "colab": {
+   "provenance": []
+  },
+  "kernelspec": {
+   "name": "python3",
+   "display_name": "Python 3"
+  },
+  "language_info": {
+   "name": "python"
+  },
+  "accelerator": "GPU",
+  "gpuClass": "standard"
+ },
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "source": [
+    "\n",
+    "**Best-of-n sampling as an alternative to RLHF**\n",
+    "\n",
+    "This notebook compares reward-model scores of prompt based responses from \n",
+    "1. a base model (`gpt2-imdb`)\n",
+    "2. `RLHF` tuned model based on this base-model \n",
+    "3. the base-model again from which we sample n responses to each prompt, score them and take the best scored one AKA the `best-of-n sampled` model\n",
+    "\n"
+   ],
+   "metadata": {
+    "id": "WQpNapZNWuXP"
+   }
+  },
+  {
+   "cell_type": "markdown",
+   "source": [
+    "Import dependencies\n"
+   ],
+   "metadata": {
+    "id": "Lo98lkdP66_x"
+   }
+  },
+  {
+   "cell_type": "code",
+   "source": [
+    "%pip install transformers trl"
+   ],
+   "metadata": {
+    "id": "vDA6qayz692w"
+   },
+   "execution_count": null,
+   "outputs": []
+  },
+  {
+   "cell_type": "code",
+   "source": [
+    "import torch\n",
+    "import pandas as pd\n",
+    "from transformers import pipeline, AutoTokenizer\n",
+    "from datasets import load_dataset\n",
+    "\n",
+    "from trl import AutoModelForCausalLMWithValueHead\n",
+    "from trl.core import LengthSampler\n",
+    "\n",
+    "device = 0 if torch.cuda.is_available() else \"cpu\""
+   ],
+   "metadata": {
+    "id": "M1s_iNm773hM"
+   },
+   "execution_count": 2,
+   "outputs": []
+  },
+  {
+   "cell_type": "markdown",
+   "source": [
+    "Various constants"
+   ],
+   "metadata": {
+    "id": "Y7hyrIrO8tcY"
+   }
+  },
+  {
+   "cell_type": "code",
+   "source": [
+    "ref_model_name = \"lvwerra/gpt2-imdb\"\n",
+    "model_name = \"lvwerra/gpt2-imdb-pos-v2\"\n",
+    "reward_model = \"lvwerra/distilbert-imdb\"\n",
+    "\n",
+    "N_BEST_OF = 4"
+   ],
+   "metadata": {
+    "id": "MqS3OM6Q8x6g"
+   },
+   "execution_count": 3,
+   "outputs": []
+  },
+  {
+   "cell_type": "markdown",
+   "source": [
+    "Models and  tokenizers "
+   ],
+   "metadata": {
+    "id": "c1YcXeElg6or"
+   }
+  },
+  {
+   "cell_type": "code",
+   "source": [
+    "model = AutoModelForCausalLMWithValueHead.from_pretrained(model_name)\n",
+    "\n",
+    "ref_model = AutoModelForCausalLMWithValueHead.from_pretrained(ref_model_name)\n",
+    "\n",
+    "reward_pipe = pipeline(\"sentiment-analysis\", model=reward_model, device=device)\n",
+    "\n",
+    "tokenizer = AutoTokenizer.from_pretrained(ref_model_name)\n",
+    "\n",
+    "tokenizer.pad_token = tokenizer.eos_token\n",
+    "\n",
+    "# cuda-ize models\n",
+    "model.cuda()\n",
+    "ref_model.cuda()"
+   ],
+   "metadata": {
+    "id": "b855NrL181Hh"
+   },
+   "execution_count": null,
+   "outputs": []
+  },
+  {
+   "cell_type": "markdown",
+   "source": [
+    "Dataset building"
+   ],
+   "metadata": {
+    "id": "Z1Cz0gCFhZYJ"
+   }
+  },
+  {
+   "cell_type": "code",
+   "source": [
+    "def build_dataset(tokenizer, dataset_name=\"imdb\", input_min_text_length=2, input_max_text_length=8):\n",
+    "    # load imdb with datasets\n",
+    "    ds = load_dataset(dataset_name, split=\"train\")\n",
+    "    ds = ds.rename_columns({\"text\": \"review\"})\n",
+    "    ds = ds.filter(lambda x: len(x[\"review\"]) > 200, batched=False)\n",
+    "\n",
+    "    input_size = LengthSampler(input_min_text_length, input_max_text_length)\n",
+    "\n",
+    "    def tokenize(sample):\n",
+    "        sample[\"input_ids\"] = tokenizer.encode(sample[\"review\"])[: input_size()]\n",
+    "        sample[\"query\"] = tokenizer.decode(sample[\"input_ids\"])\n",
+    "        return sample\n",
+    "\n",
+    "    ds = ds.map(tokenize, batched=False)\n",
+    "    ds.set_format(type=\"torch\")\n",
+    "    return ds\n",
+    "\n",
+    "\n",
+    "dataset = build_dataset(tokenizer)"
+   ],
+   "metadata": {
+    "id": "LqLVEp5p_8XM"
+   },
+   "execution_count": null,
+   "outputs": []
+  },
+  {
+   "cell_type": "code",
+   "source": [
+    "gen_kwargs = {\"min_length\": -1, \"top_k\": 0.0, \"top_p\": 1.0, \"do_sample\": True, \"pad_token_id\": tokenizer.eos_token_id}\n",
+    "sent_kwargs = {\"top_k\": None, \"function_to_apply\": \"none\", \"batch_size\": 16}"
+   ],
+   "metadata": {
+    "id": "AqA2McjMAxNw"
+   },
+   "execution_count": 6,
+   "outputs": []
+  },
+  {
+   "cell_type": "code",
+   "source": [
+    "output_min_length = 4\n",
+    "output_max_length = 16\n",
+    "output_length_sampler = LengthSampler(output_min_length, output_max_length)\n",
+    "\n",
+    "#### get a batch from the dataset\n",
+    "bs = 16\n",
+    "output_data = dict()\n",
+    "dataset.set_format(\"pandas\")\n",
+    "df_batch = dataset[:].sample(bs)\n",
+    "output_data[\"query\"] = df_batch[\"query\"].tolist()\n",
+    "query_tensors = df_batch[\"input_ids\"].tolist()\n",
+    "\n",
+    "# :: [Resp]\n",
+    "response_tensors_ref, response_tensors = [], []\n",
+    "# :: [[Resp]]\n",
+    "response_tensors_best_of = []"
+   ],
+   "metadata": {
+    "id": "L_q4qs35AxcR"
+   },
+   "execution_count": 7,
+   "outputs": []
+  },
+  {
+   "cell_type": "markdown",
+   "source": [
+    "\n",
+    "Generation using various models"
+   ],
+   "metadata": {
+    "id": "QVfpyHnZBLKY"
+   }
+  },
+  {
+   "cell_type": "code",
+   "source": [
+    "for i in range(bs):\n",
+    "    gen_len = output_length_sampler()\n",
+    "\n",
+    "    query = torch.tensor(query_tensors[i])\n",
+    "\n",
+    "    output = ref_model.generate(query.unsqueeze(dim=0).to(device), max_new_tokens=gen_len, **gen_kwargs).squeeze()\n",
+    "    response_tensors_ref.append(tokenizer.decode(output))\n",
+    "\n",
+    "    output = model.generate(query.unsqueeze(dim=0).to(device), max_new_tokens=gen_len, **gen_kwargs).squeeze()\n",
+    "    response_tensors.append(tokenizer.decode(output))\n",
+    "\n",
+    "    # generating copies of the same query for the Best-of-n sampling\n",
+    "    queries = query.repeat((N_BEST_OF, 1))\n",
+    "    output = ref_model.generate(queries.to(device), max_new_tokens=gen_len, **gen_kwargs).squeeze()\n",
+    "    response_tensors_best_of.append(tokenizer.batch_decode(output))"
+   ],
+   "metadata": {
+    "id": "-imZ7uEFBNbw"
+   },
+   "execution_count": 8,
+   "outputs": []
+  },
+  {
+   "cell_type": "markdown",
+   "source": [
+    "Scoring"
+   ],
+   "metadata": {
+    "id": "Jp5FC0Y5h_Sf"
+   }
+  },
+  {
+   "cell_type": "code",
+   "source": [
+    "scores_ref = [output[0][\"score\"] for output in reward_pipe(response_tensors_ref, **sent_kwargs)]\n",
+    "scores = [output[0][\"score\"] for output in reward_pipe(response_tensors, **sent_kwargs)]\n",
+    "scores_best_of = []\n",
+    "for i, response in enumerate(response_tensors_best_of):\n",
+    "    # base_score = scores_ref[i]\n",
+    "    scores_best_of.append(torch.tensor([output[0][\"score\"] for output in reward_pipe(response, **sent_kwargs)]))"
+   ],
+   "metadata": {
+    "id": "PyDbbAQ0F_h7"
+   },
+   "execution_count": null,
+   "outputs": []
+  },
+  {
+   "cell_type": "code",
+   "source": [
+    "output_data[\"response (ref)\"] = response_tensors_ref\n",
+    "output_data[\"scores (ref)\"] = scores_ref\n",
+    "output_data[\"response (RLHF)\"] = response_tensors\n",
+    "output_data[\"scores (RLHF)\"] = scores\n",
+    "output_data[\"response (best_of)\"] = [\n",
+    "    response_tensors_best_of[i][a.argmax().item()] for i, a in enumerate(scores_best_of)\n",
+    "]\n",
+    "output_data[\"scores (best_of)\"] = [a.max().item() for a in scores_best_of]\n",
+    "\n",
+    "\n",
+    "# store results in a dataframe\n",
+    "df_results = pd.DataFrame(output_data)\n",
+    "df_results"
+   ],
+   "metadata": {
+    "colab": {
+     "base_uri": "https://localhost:8080/",
+     "height": 682
+    },
+    "id": "nA1GDNJEiGm-",
+    "outputId": "1389c686-0751-4304-dea2-b71fd68748e1"
+   },
+   "execution_count": 10,
+   "outputs": [
+    {
+     "output_type": "execute_result",
+     "data": {
+      "text/plain": [
+       "                                 query  \\\n",
+       "0                     I'm a pretty old   \n",
+       "1                      One of the most   \n",
+       "2                             Okay, as   \n",
+       "3                        Watching \"Kro   \n",
+       "4   Seriously what were they thinking?   \n",
+       "5                         OK Hollywood   \n",
+       "6                             \"Bend It   \n",
+       "7   While the premise behind The House   \n",
+       "8                       Well let me go   \n",
+       "9                Vijay Krishna Acharya   \n",
+       "10         Watching this movie made me   \n",
+       "11                  There are probably   \n",
+       "12                          Meryl Stre   \n",
+       "13     I thought I read somewhere that   \n",
+       "14                    Good movie, very   \n",
+       "15                    It was agonizing   \n",
+       "\n",
+       "                                       response (ref)  scores (ref)  \\\n",
+       "0       I'm a pretty old kid, well, with lots of girl      1.179652   \n",
+       "1   One of the most psychologically devastating as...      2.477277   \n",
+       "2   Okay, as ruthless as they are, even their leve...      1.466462   \n",
+       "3                            Watching \"Kroger\" (1915-      0.186047   \n",
+       "4   Seriously what were they thinking? It ain't go...      1.010697   \n",
+       "5   OK Hollywood goes into a total game of audio, ...      0.934041   \n",
+       "6   \"Bend It, Luther, Dodge, Church Goes to Rome w...      0.039218   \n",
+       "7   While the premise behind The House of Dracula ...     -0.079306   \n",
+       "8   Well let me go...I don't want to movie it. I'm...      1.015246   \n",
+       "9    Vijay Krishna Acharya Sawai (Elverling). She was      0.341506   \n",
+       "10  Watching this movie made me poorly appreciate ...      1.574047   \n",
+       "11  There are probably more but if you had never s...     -0.047099   \n",
+       "12                          Meryl Streep's version of      0.373884   \n",
+       "13  I thought I read somewhere that the Lord had c...      0.091776   \n",
+       "14  Good movie, very funny, acting is very good.<|...      2.408837   \n",
+       "15            It was agonizing, and it made me wonder      1.240262   \n",
+       "\n",
+       "                                      response (RLHF)  scores (RLHF)  \\\n",
+       "0   I'm a pretty old lady, and I loved this movie ...       2.218363   \n",
+       "1      One of the most Antibiotic Apps I have seen in       2.145479   \n",
+       "2   Okay, as I enjoyed the movie. It's added bonus...       2.239827   \n",
+       "3                   Watching \"Kroven\". The film has a       1.044690   \n",
+       "4   Seriously what were they thinking? It's a very...       2.753088   \n",
+       "5   OK Hollywood shoot, and this is a classic. Som...       2.517364   \n",
+       "6   \"Bend It all\" is a sophisticated, drawing and ...       2.583935   \n",
+       "7   While the premise behind The House Intelligenc...       0.205217   \n",
+       "8   Well let me go through everything says it's a ...       2.727040   \n",
+       "9   Vijay Krishna Acharya is a perfect performance...       2.563642   \n",
+       "10  Watching this movie made me sleep better. It w...       1.690222   \n",
+       "11  There are probably random man only recently wh...       0.398258   \n",
+       "12                              Meryl Streitz, who is       0.085154   \n",
+       "13  I thought I read somewhere that my thoughts, a...       1.833734   \n",
+       "14  Good movie, very much fuzz and logical based w...       2.325996   \n",
+       "15       It was agonizing because it was truly fun to       0.969669   \n",
+       "\n",
+       "                                   response (best_of)  scores (best_of)  \n",
+       "0   I'm a pretty old, stinking,acting kinda chick ...          2.016955  \n",
+       "1   One of the most memorable performances of this...          2.676944  \n",
+       "2   Okay, as I put it in such a negative mood, it ...          1.478424  \n",
+       "3            Watching \"Kro\" is an entertainment craze          1.389495  \n",
+       "4   Seriously what were they thinking? It was stil...          2.523514  \n",
+       "5   OK Hollywood pay and the freaky set-up of this...          1.634765  \n",
+       "6   \"Bend It 9\"/\"Zara Pephoto\") and an honest, rea...          2.557210  \n",
+       "7   While the premise behind The House of Dracula ...          1.676889  \n",
+       "8   Well let me go though, alive in this ever grow...          2.652859  \n",
+       "9   Vijay Krishna Acharya adeptly emerges, and the...          2.308076  \n",
+       "10  Watching this movie made me curious: what did ...          0.950836  \n",
+       "11  There are probably too many documentaries in s...          1.142725  \n",
+       "12                      Meryl Streep performed an awe          1.932498  \n",
+       "13  I thought I read somewhere that The Odd Couple...          0.475951  \n",
+       "14  Good movie, very well polished, nicely written...          2.820022  \n",
+       "15           It was agonizing, poignant, and worst of          2.058277  "
+      ],
+      "text/html": [
+       "\n",
+       "  <div id=\"df-f55eb9dc-030e-4d67-8f1c-6f797f325376\">\n",
+       "    <div class=\"colab-df-container\">\n",
+       "      <div>\n",
+       "<style scoped>\n",
+       "    .dataframe tbody tr th:only-of-type {\n",
+       "        vertical-align: middle;\n",
+       "    }\n",
+       "\n",
+       "    .dataframe tbody tr th {\n",
+       "        vertical-align: top;\n",
+       "    }\n",
+       "\n",
+       "    .dataframe thead th {\n",
+       "        text-align: right;\n",
+       "    }\n",
+       "</style>\n",
+       "<table border=\"1\" class=\"dataframe\">\n",
+       "  <thead>\n",
+       "    <tr style=\"text-align: right;\">\n",
+       "      <th></th>\n",
+       "      <th>query</th>\n",
+       "      <th>response (ref)</th>\n",
+       "      <th>scores (ref)</th>\n",
+       "      <th>response (RLHF)</th>\n",
+       "      <th>scores (RLHF)</th>\n",
+       "      <th>response (best_of)</th>\n",
+       "      <th>scores (best_of)</th>\n",
+       "    </tr>\n",
+       "  </thead>\n",
+       "  <tbody>\n",
+       "    <tr>\n",
+       "      <th>0</th>\n",
+       "      <td>I'm a pretty old</td>\n",
+       "      <td>I'm a pretty old kid, well, with lots of girl</td>\n",
+       "      <td>1.179652</td>\n",
+       "      <td>I'm a pretty old lady, and I loved this movie ...</td>\n",
+       "      <td>2.218363</td>\n",
+       "      <td>I'm a pretty old, stinking,acting kinda chick ...</td>\n",
+       "      <td>2.016955</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>1</th>\n",
+       "      <td>One of the most</td>\n",
+       "      <td>One of the most psychologically devastating as...</td>\n",
+       "      <td>2.477277</td>\n",
+       "      <td>One of the most Antibiotic Apps I have seen in</td>\n",
+       "      <td>2.145479</td>\n",
+       "      <td>One of the most memorable performances of this...</td>\n",
+       "      <td>2.676944</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>2</th>\n",
+       "      <td>Okay, as</td>\n",
+       "      <td>Okay, as ruthless as they are, even their leve...</td>\n",
+       "      <td>1.466462</td>\n",
+       "      <td>Okay, as I enjoyed the movie. It's added bonus...</td>\n",
+       "      <td>2.239827</td>\n",
+       "      <td>Okay, as I put it in such a negative mood, it ...</td>\n",
+       "      <td>1.478424</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>3</th>\n",
+       "      <td>Watching \"Kro</td>\n",
+       "      <td>Watching \"Kroger\" (1915-</td>\n",
+       "      <td>0.186047</td>\n",
+       "      <td>Watching \"Kroven\". The film has a</td>\n",
+       "      <td>1.044690</td>\n",
+       "      <td>Watching \"Kro\" is an entertainment craze</td>\n",
+       "      <td>1.389495</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>4</th>\n",
+       "      <td>Seriously what were they thinking?</td>\n",
+       "      <td>Seriously what were they thinking? It ain't go...</td>\n",
+       "      <td>1.010697</td>\n",
+       "      <td>Seriously what were they thinking? It's a very...</td>\n",
+       "      <td>2.753088</td>\n",
+       "      <td>Seriously what were they thinking? It was stil...</td>\n",
+       "      <td>2.523514</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>5</th>\n",
+       "      <td>OK Hollywood</td>\n",
+       "      <td>OK Hollywood goes into a total game of audio, ...</td>\n",
+       "      <td>0.934041</td>\n",
+       "      <td>OK Hollywood shoot, and this is a classic. Som...</td>\n",
+       "      <td>2.517364</td>\n",
+       "      <td>OK Hollywood pay and the freaky set-up of this...</td>\n",
+       "      <td>1.634765</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>6</th>\n",
+       "      <td>\"Bend It</td>\n",
+       "      <td>\"Bend It, Luther, Dodge, Church Goes to Rome w...</td>\n",
+       "      <td>0.039218</td>\n",
+       "      <td>\"Bend It all\" is a sophisticated, drawing and ...</td>\n",
+       "      <td>2.583935</td>\n",
+       "      <td>\"Bend It 9\"/\"Zara Pephoto\") and an honest, rea...</td>\n",
+       "      <td>2.557210</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>7</th>\n",
+       "      <td>While the premise behind The House</td>\n",
+       "      <td>While the premise behind The House of Dracula ...</td>\n",
+       "      <td>-0.079306</td>\n",
+       "      <td>While the premise behind The House Intelligenc...</td>\n",
+       "      <td>0.205217</td>\n",
+       "      <td>While the premise behind The House of Dracula ...</td>\n",
+       "      <td>1.676889</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>8</th>\n",
+       "      <td>Well let me go</td>\n",
+       "      <td>Well let me go...I don't want to movie it. I'm...</td>\n",
+       "      <td>1.015246</td>\n",
+       "      <td>Well let me go through everything says it's a ...</td>\n",
+       "      <td>2.727040</td>\n",
+       "      <td>Well let me go though, alive in this ever grow...</td>\n",
+       "      <td>2.652859</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>9</th>\n",
+       "      <td>Vijay Krishna Acharya</td>\n",
+       "      <td>Vijay Krishna Acharya Sawai (Elverling). She was</td>\n",
+       "      <td>0.341506</td>\n",
+       "      <td>Vijay Krishna Acharya is a perfect performance...</td>\n",
+       "      <td>2.563642</td>\n",
+       "      <td>Vijay Krishna Acharya adeptly emerges, and the...</td>\n",
+       "      <td>2.308076</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>10</th>\n",
+       "      <td>Watching this movie made me</td>\n",
+       "      <td>Watching this movie made me poorly appreciate ...</td>\n",
+       "      <td>1.574047</td>\n",
+       "      <td>Watching this movie made me sleep better. It w...</td>\n",
+       "      <td>1.690222</td>\n",
+       "      <td>Watching this movie made me curious: what did ...</td>\n",
+       "      <td>0.950836</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>11</th>\n",
+       "      <td>There are probably</td>\n",
+       "      <td>There are probably more but if you had never s...</td>\n",
+       "      <td>-0.047099</td>\n",
+       "      <td>There are probably random man only recently wh...</td>\n",
+       "      <td>0.398258</td>\n",
+       "      <td>There are probably too many documentaries in s...</td>\n",
+       "      <td>1.142725</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>12</th>\n",
+       "      <td>Meryl Stre</td>\n",
+       "      <td>Meryl Streep's version of</td>\n",
+       "      <td>0.373884</td>\n",
+       "      <td>Meryl Streitz, who is</td>\n",
+       "      <td>0.085154</td>\n",
+       "      <td>Meryl Streep performed an awe</td>\n",
+       "      <td>1.932498</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>13</th>\n",
+       "      <td>I thought I read somewhere that</td>\n",
+       "      <td>I thought I read somewhere that the Lord had c...</td>\n",
+       "      <td>0.091776</td>\n",
+       "      <td>I thought I read somewhere that my thoughts, a...</td>\n",
+       "      <td>1.833734</td>\n",
+       "      <td>I thought I read somewhere that The Odd Couple...</td>\n",
+       "      <td>0.475951</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>14</th>\n",
+       "      <td>Good movie, very</td>\n",
+       "      <td>Good movie, very funny, acting is very good.&lt;|...</td>\n",
+       "      <td>2.408837</td>\n",
+       "      <td>Good movie, very much fuzz and logical based w...</td>\n",
+       "      <td>2.325996</td>\n",
+       "      <td>Good movie, very well polished, nicely written...</td>\n",
+       "      <td>2.820022</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <th>15</th>\n",
+       "      <td>It was agonizing</td>\n",
+       "      <td>It was agonizing, and it made me wonder</td>\n",
+       "      <td>1.240262</td>\n",
+       "      <td>It was agonizing because it was truly fun to</td>\n",
+       "      <td>0.969669</td>\n",
+       "      <td>It was agonizing, poignant, and worst of</td>\n",
+       "      <td>2.058277</td>\n",
+       "    </tr>\n",
+       "  </tbody>\n",
+       "</table>\n",
+       "</div>\n",
+       "      <button class=\"colab-df-convert\" onclick=\"convertToInteractive('df-f55eb9dc-030e-4d67-8f1c-6f797f325376')\"\n",
+       "              title=\"Convert this dataframe to an interactive table.\"\n",
+       "              style=\"display:none;\">\n",
+       "        \n",
+       "  <svg xmlns=\"http://www.w3.org/2000/svg\" height=\"24px\"viewBox=\"0 0 24 24\"\n",
+       "       width=\"24px\">\n",
+       "    <path d=\"M0 0h24v24H0V0z\" fill=\"none\"/>\n",
+       "    <path d=\"M18.56 5.44l.94 2.06.94-2.06 2.06-.94-2.06-.94-.94-2.06-.94 2.06-2.06.94zm-11 1L8.5 8.5l.94-2.06 2.06-.94-2.06-.94L8.5 2.5l-.94 2.06-2.06.94zm10 10l.94 2.06.94-2.06 2.06-.94-2.06-.94-.94-2.06-.94 2.06-2.06.94z\"/><path d=\"M17.41 7.96l-1.37-1.37c-.4-.4-.92-.59-1.43-.59-.52 0-1.04.2-1.43.59L10.3 9.45l-7.72 7.72c-.78.78-.78 2.05 0 2.83L4 21.41c.39.39.9.59 1.41.59.51 0 1.02-.2 1.41-.59l7.78-7.78 2.81-2.81c.8-.78.8-2.07 0-2.86zM5.41 20L4 18.59l7.72-7.72 1.47 1.35L5.41 20z\"/>\n",
+       "  </svg>\n",
+       "      </button>\n",
+       "      \n",
+       "  <style>\n",
+       "    .colab-df-container {\n",
+       "      display:flex;\n",
+       "      flex-wrap:wrap;\n",
+       "      gap: 12px;\n",
+       "    }\n",
+       "\n",
+       "    .colab-df-convert {\n",
+       "      background-color: #E8F0FE;\n",
+       "      border: none;\n",
+       "      border-radius: 50%;\n",
+       "      cursor: pointer;\n",
+       "      display: none;\n",
+       "      fill: #1967D2;\n",
+       "      height: 32px;\n",
+       "      padding: 0 0 0 0;\n",
+       "      width: 32px;\n",
+       "    }\n",
+       "\n",
+       "    .colab-df-convert:hover {\n",
+       "      background-color: #E2EBFA;\n",
+       "      box-shadow: 0px 1px 2px rgba(60, 64, 67, 0.3), 0px 1px 3px 1px rgba(60, 64, 67, 0.15);\n",
+       "      fill: #174EA6;\n",
+       "    }\n",
+       "\n",
+       "    [theme=dark] .colab-df-convert {\n",
+       "      background-color: #3B4455;\n",
+       "      fill: #D2E3FC;\n",
+       "    }\n",
+       "\n",
+       "    [theme=dark] .colab-df-convert:hover {\n",
+       "      background-color: #434B5C;\n",
+       "      box-shadow: 0px 1px 3px 1px rgba(0, 0, 0, 0.15);\n",
+       "      filter: drop-shadow(0px 1px 2px rgba(0, 0, 0, 0.3));\n",
+       "      fill: #FFFFFF;\n",
+       "    }\n",
+       "  </style>\n",
+       "\n",
+       "      <script>\n",
+       "        const buttonEl =\n",
+       "          document.querySelector('#df-f55eb9dc-030e-4d67-8f1c-6f797f325376 button.colab-df-convert');\n",
+       "        buttonEl.style.display =\n",
+       "          google.colab.kernel.accessAllowed ? 'block' : 'none';\n",
+       "\n",
+       "        async function convertToInteractive(key) {\n",
+       "          const element = document.querySelector('#df-f55eb9dc-030e-4d67-8f1c-6f797f325376');\n",
+       "          const dataTable =\n",
+       "            await google.colab.kernel.invokeFunction('convertToInteractive',\n",
+       "                                                     [key], {});\n",
+       "          if (!dataTable) return;\n",
+       "\n",
+       "          const docLinkHtml = 'Like what you see? Visit the ' +\n",
+       "            '<a target=\"_blank\" href=https://colab.research.google.com/notebooks/data_table.ipynb>data table notebook</a>'\n",
+       "            + ' to learn more about interactive tables.';\n",
+       "          element.innerHTML = '';\n",
+       "          dataTable['output_type'] = 'display_data';\n",
+       "          await google.colab.output.renderOutput(dataTable, element);\n",
+       "          const docLink = document.createElement('div');\n",
+       "          docLink.innerHTML = docLinkHtml;\n",
+       "          element.appendChild(docLink);\n",
+       "        }\n",
+       "      </script>\n",
+       "    </div>\n",
+       "  </div>\n",
+       "  "
+      ]
+     },
+     "metadata": {},
+     "execution_count": 10
+    }
+   ]
+  }
+ ]
+}

examples/notebooks/gpt2-sentiment-control.ipynb

@@ -73,6 +73,7 @@
     "import pandas as pd\n",
     "from random import choices\n",
     "import matplotlib.pyplot as plt\n",
+    "\n",
     "tqdm.pandas()\n",
     "\n",
     "from datasets import load_dataset\n",
@@ -95,22 +96,15 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "sentiment_pipe_kwargs = {\n",
-    "    \"top_k\": None,  \n",
-    "    \"function_to_apply\": \"none\"\n",
-    "}\n",
+    "sentiment_pipe_kwargs = {\"top_k\": None, \"function_to_apply\": \"none\"}\n",
     "\n",
     "config = PPOConfig(\n",
-    "    model_name=\"lvwerra/gpt2-imdb\",\n",
-    "    steps=51200,\n",
-    "    learning_rate=1.41e-5,\n",
-    "    remove_unused_columns=False,\n",
-    "    log_with=\"wandb\"\n",
+    "    model_name=\"lvwerra/gpt2-imdb\", steps=51200, learning_rate=1.41e-5, remove_unused_columns=False, log_with=\"wandb\"\n",
     ")\n",
     "\n",
     "txt_in_len = 5\n",
     "txt_out_len = 20\n",
-    "seed = 1\n"
+    "seed = 1"
    ]
   },
   {
@@ -201,13 +195,13 @@
     }
    ],
    "source": [
-    "# create the dataset \n",
-    "# \n",
-    "dataset = load_dataset('imdb', split='train')\n",
-    "dataset = dataset.rename_columns({'text': 'review', 'label': 'sentiment'})\n",
+    "# create the dataset\n",
+    "#\n",
+    "dataset = load_dataset(\"imdb\", split=\"train\")\n",
+    "dataset = dataset.rename_columns({\"text\": \"review\", \"label\": \"sentiment\"})\n",
     "# make sure the comments are are at least 500 and trim to 1000\n",
-    "dataset = dataset.filter(lambda x: len(x[\"review\"])>500, batched=False)\n",
-    "dataset = dataset.map(lambda x:{\"review\":x['review'][:1000]}, batched=False)\n",
+    "dataset = dataset.filter(lambda x: len(x[\"review\"]) > 500, batched=False)\n",
+    "dataset = dataset.map(lambda x: {\"review\": x[\"review\"][:1000]}, batched=False)\n",
     "\n",
     "dataset"
    ]
@@ -241,11 +235,15 @@
     }
    ],
    "source": [
-    "dataset = dataset.map(lambda x:{\"input_ids\": gpt2_tokenizer.encode(' '+x['review'], return_tensors=\"pt\")[0, :txt_in_len]}, batched=False)\n",
-    "dataset = dataset.map(lambda x:{\"query\": gpt2_tokenizer.decode(x[\"input_ids\"])}, batched=False)\n",
+    "dataset = dataset.map(\n",
+    "    lambda x: {\"input_ids\": gpt2_tokenizer.encode(\" \" + x[\"review\"], return_tensors=\"pt\")[0, :txt_in_len]},\n",
+    "    batched=False,\n",
+    ")\n",
+    "dataset = dataset.map(lambda x: {\"query\": gpt2_tokenizer.decode(x[\"input_ids\"])}, batched=False)\n",
     "dataset = dataset[:20480]\n",
     "\n",
     "from datasets import Dataset\n",
+    "\n",
     "dataset = Dataset.from_dict(dataset)\n",
     "dataset.set_format(\"pytorch\")"
    ]
@@ -355,7 +353,7 @@
     }
    ],
    "source": [
-    "ppo_trainer = PPOTrainer(config, gpt2_model, gpt2_model_ref, gpt2_tokenizer, dataset, data_collator=collator)\n"
+    "ppo_trainer = PPOTrainer(config, gpt2_model, gpt2_model_ref, gpt2_tokenizer, dataset, data_collator=collator)"
    ]
   },
   {
@@ -373,7 +371,7 @@
    "outputs": [],
    "source": [
     "if ppo_trainer.accelerator.num_processes == 1:\n",
-    "    device = 0 if torch.cuda.is_available() else \"cpu\" # to avoid a `pipeline` bug\n",
+    "    device = 0 if torch.cuda.is_available() else \"cpu\"  # to avoid a `pipeline` bug\n",
     "else:\n",
     "    device = ppo_trainer.accelerator.device\n",
     "sentiment_pipe = pipeline(\"sentiment-analysis\", \"lvwerra/distilbert-imdb\", device=device)"
@@ -404,7 +402,7 @@
     }
    ],
    "source": [
-    "text = 'this movie was really bad!!'\n",
+    "text = \"this movie was really bad!!\"\n",
     "output = sentiment_pipe(text, **sentiment_pipe_kwargs)\n",
     "output"
    ]
@@ -427,7 +425,7 @@
     }
    ],
    "source": [
-    "text = 'this movie was really good!!'\n",
+    "text = \"this movie was really good!!\"\n",
     "output = sentiment_pipe(text, **sentiment_pipe_kwargs)\n",
     "output"
    ]
@@ -450,7 +448,7 @@
     }
    ],
    "source": [
-    "text = 'this movie was a documentary'\n",
+    "text = \"this movie was a documentary\"\n",
     "output = sentiment_pipe(text, **sentiment_pipe_kwargs)\n",
     "output"
    ]
@@ -472,7 +470,7 @@
     "    positive_logits = []\n",
     "    for out in outputs:\n",
     "        for element in out:\n",
-    "            if element[\"label\"]==\"POSITIVE\":\n",
+    "            if element[\"label\"] == \"POSITIVE\":\n",
     "                positive_logits.append(torch.tensor(element[\"score\"]))\n",
     "    return positive_logits"
    ]
@@ -511,8 +509,8 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "ctrl_str = ['[negative]', '[neutral]', '[positive]']\n",
-    "device = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\") # this should be handled by accelerate\n",
+    "ctrl_str = [\"[negative]\", \"[neutral]\", \"[positive]\"]\n",
+    "device = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")  # this should be handled by accelerate\n",
     "ctrl_tokens = dict((s, gpt2_tokenizer.encode(s, return_tensors=\"pt\").squeeze().to(device)) for s in ctrl_str)"
    ]
   },
@@ -559,14 +557,14 @@
     "        task [positive]: reward = logit\n",
     "    \"\"\"\n",
     "    for i in range(len(logit)):\n",
-    "        if task[i]=='[negative]':\n",
+    "        if task[i] == \"[negative]\":\n",
     "            logit[i] = -logit[i]\n",
-    "        elif task[i]=='[neutral]':\n",
-    "            logit[i] = -2*torch.abs(logit[i])+4\n",
-    "        elif task[i]=='[positive]':\n",
+    "        elif task[i] == \"[neutral]\":\n",
+    "            logit[i] = -2 * torch.abs(logit[i]) + 4\n",
+    "        elif task[i] == \"[positive]\":\n",
     "            pass\n",
     "        else:\n",
-    "            raise ValueError('task has to be in [0, 1, 2]!')\n",
+    "            raise ValueError(\"task has to be in [0, 1, 2]!\")\n",
     "    return logit"
    ]
   },
@@ -611,7 +609,7 @@
     }
    ],
    "source": [
-    "pos_logit_to_reward(torch.Tensor([4,4,4]), ctrl_str)"
+    "pos_logit_to_reward(torch.Tensor([4, 4, 4]), ctrl_str)"
    ]
   },
   {
@@ -631,7 +629,7 @@
     }
    ],
    "source": [
-    "pos_logit_to_reward(torch.Tensor([-4,-4,-4]), ctrl_str)"
+    "pos_logit_to_reward(torch.Tensor([-4, -4, -4]), ctrl_str)"
    ]
   },
   {
@@ -668,14 +666,14 @@
    "outputs": [],
    "source": [
     "generation_kwargs = {\n",
-    "    \"min_length\":-1,\n",
+    "    \"min_length\": -1,\n",
     "    \"top_k\": 0.0,\n",
     "    \"top_p\": 1.0,\n",
     "    \"do_sample\": True,\n",
     "    \"pad_token_id\": gpt2_tokenizer.eos_token_id,\n",
     "    \"max_new_tokens\": txt_out_len,\n",
-    "    \"eos_token_id\": -1\n",
-    "}\n"
+    "    \"eos_token_id\": -1,\n",
+    "}"
    ]
   },
   {
@@ -698,7 +696,6 @@
     "4. Get sentiments for query/responses from BERT\n",
     "5. Optimize policy with PPO using the (query, response, reward) triplet\n",
     "6. Log all the training statistics\n",
-
     "\n",
     "**Training time**\n",
     "\n",
@@ -724,11 +721,14 @@
    "source": [
     "for epoch in range(2):\n",
     "    for batch in tqdm(ppo_trainer.dataloader):\n",
-    "        logs, game_data,  = dict(), dict()\n",
-    "        \n",
+    "        (logs, game_data,) = (\n",
+    "            dict(),\n",
+    "            dict(),\n",
+    "        )\n",
+    "\n",
     "        #### prepend a random control token\n",
     "        task_list = choices(ctrl_str, k=config.batch_size)\n",
-    "        game_data['query'] = [t+q for t,q in zip(task_list, batch['query'])]\n",
+    "        game_data[\"query\"] = [t + q for t, q in zip(task_list, batch[\"query\"])]\n",
     "        query_tensors = [torch.cat((ctrl_tokens[t], input_ids)) for t, input_ids in zip(task_list, batch[\"input_ids\"])]\n",
     "\n",
     "        #### get response from gpt2\n",
@@ -736,21 +736,21 @@
     "        for query in query_tensors:\n",
     "            response = ppo_trainer.generate(query, **generation_kwargs)\n",
     "            response_tensors.append(response.squeeze()[-txt_out_len:])\n",
-    "        game_data['response'] = [gpt2_tokenizer.decode(r.squeeze()) for r in response_tensors]\n",
+    "        game_data[\"response\"] = [gpt2_tokenizer.decode(r.squeeze()) for r in response_tensors]\n",
     "\n",
     "        #### sentiment analysis\n",
-    "        texts = [q + r for q,r in zip(batch['query'], game_data['response'])]\n",
+    "        texts = [q + r for q, r in zip(batch[\"query\"], game_data[\"response\"])]\n",
     "        logits = extract_pipe_output(sentiment_pipe(texts, **sentiment_pipe_kwargs))\n",
     "        rewards = pos_logit_to_reward(logits, task_list)\n",
     "\n",
-    "        #### Run PPO training \n",
+    "        #### Run PPO training\n",
     "        t = time.time()\n",
     "        stats = ppo_trainer.step(query_tensors, response_tensors, rewards)\n",
     "\n",
     "        for cs in ctrl_str:\n",
-    "            key = 'env/reward_'+cs.strip('[]')\n",
-    "            stats[key] = np.mean([r.cpu().numpy() for r, t in zip(rewards, task_list) if t==cs])\n",
-    "        ppo_trainer.log_stats(stats, game_data, rewards)\n"
+    "            key = \"env/reward_\" + cs.strip(\"[]\")\n",
+    "            stats[key] = np.mean([r.cpu().numpy() for r, t in zip(rewards, task_list) if t == cs])\n",
+    "        ppo_trainer.log_stats(stats, game_data, rewards)"
    ]
   },
   {
@@ -803,12 +803,11 @@
    ],
    "source": [
     "for ctrl_s in ctrl_str:\n",
-    "    plt.hist([r for r, t in zip(logs['env/reward_dist'], task_list) if t==ctrl_s],\n",
-    "             density=True,\n",
-    "             alpha=0.5,\n",
-    "             label=ctrl_s)\n",
-    "plt.legend(loc='best')\n",
-    "plt.title('reward distribution')\n",
+    "    plt.hist(\n",
+    "        [r for r, t in zip(logs[\"env/reward_dist\"], task_list) if t == ctrl_s], density=True, alpha=0.5, label=ctrl_s\n",
+    "    )\n",
+    "plt.legend(loc=\"best\")\n",
+    "plt.title(\"reward distribution\")\n",
     "plt.grid(True)\n",
     "plt.show()"
    ]
@@ -827,8 +826,8 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "gpt2_model.save_pretrained('gpt2-imdb-ctrl')\n",
-    "gpt2_tokenizer.save_pretrained('gpt2-imdb-ctrl')"
+    "gpt2_model.save_pretrained(\"gpt2-imdb-ctrl\")\n",
+    "gpt2_tokenizer.save_pretrained(\"gpt2-imdb-ctrl\")"
    ]
   }
  ],
@@ -848,7 +847,7 @@
    "name": "python",
    "nbconvert_exporter": "python",
    "pygments_lexer": "ipython3",
-   "version": "3.9.16"
+   "version": "3.9.12"
   },
   "vscode": {
    "interpreter": {

examples/notebooks/gpt2-sentiment.ipynb

@@ -63,6 +63,7 @@
     "import torch\n",
     "from tqdm import tqdm\n",
     "import pandas as pd\n",
+    "\n",
     "tqdm.pandas()\n",
     "\n",
     "from transformers import pipeline, AutoTokenizer\n",
@@ -91,11 +92,7 @@
     "    log_with=\"wandb\",\n",
     ")\n",
     "\n",
-    "sent_kwargs = {\n",
-    "    \"return_all_scores\": True,\n",
-    "    \"function_to_apply\": \"none\",\n",
-    "    \"batch_size\": 16\n",
-    "}"
+    "sent_kwargs = {\"return_all_scores\": True, \"function_to_apply\": \"none\", \"batch_size\": 16}"
    ]
   },
   {
@@ -105,6 +102,7 @@
    "outputs": [],
    "source": [
     "import wandb\n",
+    "\n",
     "wandb.init()"
    ]
   },
@@ -149,13 +147,13 @@
    "source": [
     "def build_dataset(config, dataset_name=\"imdb\", input_min_text_length=2, input_max_text_length=8):\n",
     "    \"\"\"\n",
-    "    Build dataset for training. This builds the dataset from `load_dataset`, one should \n",
+    "    Build dataset for training. This builds the dataset from `load_dataset`, one should\n",
     "    customize this function to train the model on its own dataset.\n",
-    "    \n",
+    "\n",
     "    Args:\n",
-    "        dataset_name (`str`): \n",
+    "        dataset_name (`str`):\n",
     "            The name of the dataset to be loaded.\n",
-    "    \n",
+    "\n",
     "    Returns:\n",
     "        dataloader (`torch.utils.data.DataLoader`):\n",
     "            The dataloader for the dataset.\n",
@@ -163,19 +161,19 @@
     "    tokenizer = AutoTokenizer.from_pretrained(config.model_name)\n",
     "    tokenizer.pad_token = tokenizer.eos_token\n",
     "    # load imdb with datasets\n",
-    "    ds = load_dataset(dataset_name, split='train')\n",
-    "    ds = ds.rename_columns({'text': 'review'})\n",
-    "    ds = ds.filter(lambda x: len(x[\"review\"])>200, batched=False)\n",
+    "    ds = load_dataset(dataset_name, split=\"train\")\n",
+    "    ds = ds.rename_columns({\"text\": \"review\"})\n",
+    "    ds = ds.filter(lambda x: len(x[\"review\"]) > 200, batched=False)\n",
     "\n",
     "    input_size = LengthSampler(input_min_text_length, input_max_text_length)\n",
     "\n",
     "    def tokenize(sample):\n",
-    "        sample[\"input_ids\"] = tokenizer.encode(sample[\"review\"])[:input_size()]\n",
+    "        sample[\"input_ids\"] = tokenizer.encode(sample[\"review\"])[: input_size()]\n",
     "        sample[\"query\"] = tokenizer.decode(sample[\"input_ids\"])\n",
     "        return sample\n",
     "\n",
     "    ds = ds.map(tokenize, batched=False)\n",
-    "    ds.set_format(type='torch')\n",
+    "    ds.set_format(type=\"torch\")\n",
     "    return ds"
    ]
   },
@@ -187,6 +185,7 @@
    "source": [
     "dataset = build_dataset(config)\n",
     "\n",
+    "\n",
     "def collator(data):\n",
     "    return dict((key, [d[key] for d in data]) for key in data[0])"
    ]
@@ -252,7 +251,7 @@
    "source": [
     "device = ppo_trainer.accelerator.device\n",
     "if ppo_trainer.accelerator.num_processes == 1:\n",
-    "   device = 0 if torch.cuda.is_available() else \"cpu\" # to avoid a `pipeline` bug\n",
+    "    device = 0 if torch.cuda.is_available() else \"cpu\"  # to avoid a `pipeline` bug\n",
     "sentiment_pipe = pipeline(\"sentiment-analysis\", model=\"lvwerra/distilbert-imdb\", device=device)"
    ]
   },
@@ -281,7 +280,7 @@
     }
    ],
    "source": [
-    "text = 'this movie was really bad!!'\n",
+    "text = \"this movie was really bad!!\"\n",
     "sentiment_pipe(text, **sent_kwargs)"
    ]
   },
@@ -303,7 +302,7 @@
     }
    ],
    "source": [
-    "text = 'this movie was really good!!'\n",
+    "text = \"this movie was really good!!\"\n",
     "sentiment_pipe(text, **sent_kwargs)"
    ]
   },
@@ -321,13 +320,7 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "gen_kwargs = {\n",
-    "    \"min_length\":-1,\n",
-    "    \"top_k\": 0.0,\n",
-    "    \"top_p\": 1.0,\n",
-    "    \"do_sample\": True,\n",
-    "    \"pad_token_id\": tokenizer.eos_token_id\n",
-    "}"
+    "gen_kwargs = {\"min_length\": -1, \"top_k\": 0.0, \"top_p\": 1.0, \"do_sample\": True, \"pad_token_id\": tokenizer.eos_token_id}"
    ]
   },
   {
@@ -370,16 +363,16 @@
     "\n",
     "\n",
     "generation_kwargs = {\n",
-    "    \"min_length\":-1,\n",
+    "    \"min_length\": -1,\n",
     "    \"top_k\": 0.0,\n",
     "    \"top_p\": 1.0,\n",
     "    \"do_sample\": True,\n",
-    "    \"pad_token_id\": tokenizer.eos_token_id\n",
+    "    \"pad_token_id\": tokenizer.eos_token_id,\n",
     "}\n",
     "\n",
     "\n",
     "for epoch, batch in tqdm(enumerate(ppo_trainer.dataloader)):\n",
-    "    query_tensors = batch['input_ids']\n",
+    "    query_tensors = batch[\"input_ids\"]\n",
     "\n",
     "    #### Get response from gpt2\n",
     "    response_tensors = []\n",
@@ -388,14 +381,14 @@
     "        generation_kwargs[\"max_new_tokens\"] = gen_len\n",
     "        response = ppo_trainer.generate(query, **generation_kwargs)\n",
     "        response_tensors.append(response.squeeze()[-gen_len:])\n",
-    "    batch['response'] = [tokenizer.decode(r.squeeze()) for r in response_tensors]\n",
+    "    batch[\"response\"] = [tokenizer.decode(r.squeeze()) for r in response_tensors]\n",
     "\n",
     "    #### Compute sentiment score\n",
-    "    texts = [q + r for q,r in zip(batch['query'], batch['response'])]\n",
+    "    texts = [q + r for q, r in zip(batch[\"query\"], batch[\"response\"])]\n",
     "    pipe_outputs = sentiment_pipe(texts, **sent_kwargs)\n",
     "    rewards = [torch.tensor(output[1][\"score\"]) for output in pipe_outputs]\n",
     "\n",
-    "    #### Run PPO step \n",
+    "    #### Run PPO step\n",
     "    stats = ppo_trainer.step(query_tensors, response_tensors, rewards)\n",
     "    ppo_trainer.log_stats(stats, batch, rewards)"
    ]
@@ -405,7 +398,7 @@
    "metadata": {},
    "source": [
     "### Training progress\n",
-    "If you are tracking the training progress with Weights&Biases you should see a plot similar to the one below. Check out the interactive sample report on wandb.ai: [link](https://app.wandb.ai/lvwerra/trl-showcase/runs/1jtvxb1m/).\n",
+    "If you are tracking the training progress with Weights&Biases you should see a plot similar to the one below. Check out the interactive sample report on wandb.ai: [link](https://app.wandb.ai/huggingface/trl-showcase/runs/1jtvxb1m/).\n",
     "\n",
     "<div style=\"text-align: center\">\n",
     "<img src='https://huggingface.co/datasets/trl-internal-testing/example-images/resolve/main/images/gpt2_tuning_progress.png' width='800'>\n",
@@ -414,7 +407,7 @@
     "\n",
     "One can observe how the model starts to generate more positive outputs after a few optimisation steps.\n",
     "\n",
-    "> Note: Investigating the KL-divergence will probably show that at this point the model has not converged to the target KL-divergence, yet. To get there would require longer training or starting with a higher inital coefficient."
+    "> Note: Investigating the KL-divergence will probably show that at this point the model has not converged to the target KL-divergence, yet. To get there would require longer training or starting with a higher initial coefficient."
    ]
   },
   {
@@ -685,31 +678,33 @@
     "game_data = dict()\n",
     "dataset.set_format(\"pandas\")\n",
     "df_batch = dataset[:].sample(bs)\n",
-    "game_data['query'] = df_batch['query'].tolist()\n",
-    "query_tensors = df_batch['input_ids'].tolist()\n",
+    "game_data[\"query\"] = df_batch[\"query\"].tolist()\n",
+    "query_tensors = df_batch[\"input_ids\"].tolist()\n",
     "\n",
     "response_tensors_ref, response_tensors = [], []\n",
     "\n",
     "#### get response from gpt2 and gpt2_ref\n",
     "for i in range(bs):\n",
     "    gen_len = output_length_sampler()\n",
-    "    output = ref_model.generate(torch.tensor(query_tensors[i]).unsqueeze(dim=0).to(device),\n",
-    "                                     max_new_tokens=gen_len, **gen_kwargs).squeeze()[-gen_len:]\n",
+    "    output = ref_model.generate(\n",
+    "        torch.tensor(query_tensors[i]).unsqueeze(dim=0).to(device), max_new_tokens=gen_len, **gen_kwargs\n",
+    "    ).squeeze()[-gen_len:]\n",
     "    response_tensors_ref.append(output)\n",
-    "    output = model.generate(torch.tensor(query_tensors[i]).unsqueeze(dim=0).to(device),\n",
-    "                                 max_new_tokens=gen_len, **gen_kwargs).squeeze()[-gen_len:]\n",
+    "    output = model.generate(\n",
+    "        torch.tensor(query_tensors[i]).unsqueeze(dim=0).to(device), max_new_tokens=gen_len, **gen_kwargs\n",
+    "    ).squeeze()[-gen_len:]\n",
     "    response_tensors.append(output)\n",
     "\n",
     "#### decode responses\n",
-    "game_data['response (before)'] = [tokenizer.decode(response_tensors_ref[i]) for i in range(bs)]\n",
-    "game_data['response (after)'] = [tokenizer.decode(response_tensors[i]) for i in range(bs)]\n",
+    "game_data[\"response (before)\"] = [tokenizer.decode(response_tensors_ref[i]) for i in range(bs)]\n",
+    "game_data[\"response (after)\"] = [tokenizer.decode(response_tensors[i]) for i in range(bs)]\n",
     "\n",
     "#### sentiment analysis of query/response pairs before/after\n",
-    "texts = [q + r for q,r in zip(game_data['query'], game_data['response (before)'])]\n",
-    "game_data['rewards (before)'] = [output[1][\"score\"] for output in sentiment_pipe(texts, **sent_kwargs)]\n",
+    "texts = [q + r for q, r in zip(game_data[\"query\"], game_data[\"response (before)\"])]\n",
+    "game_data[\"rewards (before)\"] = [output[1][\"score\"] for output in sentiment_pipe(texts, **sent_kwargs)]\n",
     "\n",
-    "texts = [q + r for q,r in zip(game_data['query'], game_data['response (after)'])]\n",
-    "game_data['rewards (after)'] = [output[1][\"score\"] for output in sentiment_pipe(texts, **sent_kwargs)]\n",
+    "texts = [q + r for q, r in zip(game_data[\"query\"], game_data[\"response (after)\"])]\n",
+    "game_data[\"rewards (after)\"] = [output[1][\"score\"] for output in sentiment_pipe(texts, **sent_kwargs)]\n",
     "\n",
     "# store results in a dataframe\n",
     "df_results = pd.DataFrame(game_data)\n",
@@ -767,10 +762,10 @@
     }
    ],
    "source": [
-    "print('mean:')\n",
+    "print(\"mean:\")\n",
     "display(df_results[[\"rewards (before)\", \"rewards (after)\"]].mean())\n",
     "print()\n",
-    "print('median:')\n",
+    "print(\"median:\")\n",
     "display(df_results[[\"rewards (before)\", \"rewards (after)\"]].median())"
    ]
   },
@@ -843,8 +838,8 @@
     }
    ],
    "source": [
-    "model.save_pretrained('gpt2-imdb-pos-v2', push_to_hub=True)\n",
-    "tokenizer.save_pretrained('gpt2-imdb-pos-v2', push_to_hub=True)"
+    "model.save_pretrained(\"gpt2-imdb-pos-v2\", push_to_hub=True)\n",
+    "tokenizer.save_pretrained(\"gpt2-imdb-pos-v2\", push_to_hub=True)"
    ]
   },
   {

examples/research_projects/README.md

@@ -0,0 +1,7 @@
+# Research projects that use TRL
+
+Welcome to the research projects folder! Here you can find the scripts used for some research projects that used TRL and maintained by the developers and the community (LM de-toxification, Stack-Llama, etc.). Check out the READMEs in the subfolders for more information!
+
+- [De-detoxifying language models](https://github.com/huggingface/trl/tree/main/examples/research_projects/toxicity)
+- [Stack-Llama](https://github.com/huggingface/trl/tree/main/examples/research_projects/stack_llama)
+- [Stack-Llama-2](https://github.com/huggingface/trl/tree/main/examples/research_projects/stack_llama_2)

examples/research_projects/stack_llama/scripts/README.md

@@ -0,0 +1,18 @@
+# RLHF pipeline for the creation of StackLLaMa: a Stack exchange llama-7b model.
+There were three main steps to the training process:
+1. Supervised fine-tuning of the base llama-7b model to create llama-7b-se:
+    - `torchrun --nnodes 1  --nproc_per_node 8 examples/research_projects/stack_llama/scripts/supervised_finetuning.py --model_path=<LLAMA_MODEL_PATH> --streaming --learning_rate 1e-5 --max_steps 5000 --output_dir ./llama-se`
+2. Reward modeling using dialog pairs from the SE dataset using the llama-7b-se to create llama-7b-se-rm:
+    - `torchrun --nnodes 1  --nproc_per_node 8 examples/research_projects/stack_llama/scripts/reward_modeling.py --model_name=<LLAMA_SE_MODEL>`
+3. RL fine-tuning of llama-7b-se with the llama-7b-se-rm reward model:
+    - `accelerate launch --multi_gpu --num_machines 1  --num_processes 8 examples/research_projects/stack_llama/scripts/rl_training.py --log_with=wandb --model_name=<LLAMA_SE_MODEL> --reward_model_name=<LLAMA_SE_RM_MODEL> --adafactor=False --tokenizer_name=<LLAMA_TOKENIZER> --save_freq=100 --output_max_length=128 --batch_size=8 --gradient_accumulation_steps=8 --batched_gen=True --ppo_epochs=4 --seed=0 --learning_rate=1.4e-5 --early_stopping=True --output_dir=llama-se-rl-finetune-128-8-8-1.4e-5_adam`
+
+
+LoRA layers were using at all stages to reduce memory requirements. 
+At each stage the peft adapter layers were merged with the base model, using: 
+```shell
+python examples/research_projects/stack_llama/scripts/merge_peft_adapter.py --adapter_model_name=XXX --base_model_name=YYY --output_name=ZZZ
+```
+Note that this script requires `peft>=0.3.0`.
+
+For access to the base llama-7b model, please see Meta's [release](https://ai.facebook.com/blog/large-language-model-llama-meta-ai/) and [request form](https://docs.google.com/forms/d/e/1FAIpQLSfqNECQnMkycAp2jP4Z9TFX0cGR4uf7b_fBxjY_OjhJILlKGA/viewform).

examples/research_projects/stack_llama/scripts/merge_peft_adapter.py

@@ -0,0 +1,48 @@
+from dataclasses import dataclass, field
+from typing import Optional
+
+import torch
+from peft import PeftConfig, PeftModel
+from transformers import AutoModelForCausalLM, AutoModelForSequenceClassification, AutoTokenizer, HfArgumentParser
+
+
+@dataclass
+class ScriptArguments:
+    """
+    The input names representing the Adapter and Base model fine-tuned with PEFT, and the output name representing the
+    merged model.
+    """
+
+    adapter_model_name: Optional[str] = field(default=None, metadata={"help": "the adapter name"})
+    base_model_name: Optional[str] = field(default=None, metadata={"help": "the base model name"})
+    output_name: Optional[str] = field(default=None, metadata={"help": "the merged model name"})
+
+
+parser = HfArgumentParser(ScriptArguments)
+script_args = parser.parse_args_into_dataclasses()[0]
+assert script_args.adapter_model_name is not None, "please provide the name of the Adapter you would like to merge"
+assert script_args.base_model_name is not None, "please provide the name of the Base model"
+assert script_args.output_name is not None, "please provide the output name of the merged model"
+
+peft_config = PeftConfig.from_pretrained(script_args.adapter_model_name)
+if peft_config.task_type == "SEQ_CLS":
+    # The sequence classification task is used for the reward model in PPO
+    model = AutoModelForSequenceClassification.from_pretrained(
+        script_args.base_model_name, num_labels=1, torch_dtype=torch.bfloat16
+    )
+else:
+    model = AutoModelForCausalLM.from_pretrained(
+        script_args.base_model_name, return_dict=True, torch_dtype=torch.bfloat16
+    )
+
+tokenizer = AutoTokenizer.from_pretrained(script_args.base_model_name)
+
+# Load the PEFT model
+model = PeftModel.from_pretrained(model, script_args.adapter_model_name)
+model.eval()
+
+model = model.merge_and_unload()
+
+model.save_pretrained(f"{script_args.output_name}")
+tokenizer.save_pretrained(f"{script_args.output_name}")
+model.push_to_hub(f"{script_args.output_name}", use_temp_dir=False)

examples/research_projects/stack_llama/scripts/reward_modeling.py

@@ -0,0 +1,300 @@
+from dataclasses import dataclass, field
+from typing import Any, Dict, List, Optional, Union
+
+import evaluate
+import numpy as np
+import torch
+import torch.nn as nn
+from datasets import load_dataset
+from peft import LoraConfig, TaskType, get_peft_model
+from transformers import (
+    AutoModelForSequenceClassification,
+    AutoTokenizer,
+    HfArgumentParser,
+    PreTrainedTokenizerBase,
+    Trainer,
+    TrainerCallback,
+    TrainingArguments,
+)
+from transformers.utils import PaddingStrategy
+
+
+# Define and parse arguments.
+@dataclass
+class ScriptArguments:
+    """
+    These arguments vary depending on how many GPUs you have, what their capacity and features are, and what size model you want to train.
+    """
+
+    local_rank: Optional[int] = field(default=-1, metadata={"help": "Used for multi-gpu"})
+    resume_from_checkpoint: Optional[bool] = field(
+        default=False,
+        metadata={"help": "If you want to resume training where it left off."},
+    )
+    deepspeed: Optional[str] = field(
+        default=None,
+        metadata={
+            "help": "Path to deepspeed config if using deepspeed. You may need this if the model that you want to train doesn't fit on a single GPU."
+        },
+    )
+    per_device_train_batch_size: Optional[int] = field(default=4)
+    per_device_eval_batch_size: Optional[int] = field(default=1)
+    gradient_accumulation_steps: Optional[int] = field(default=1)
+    learning_rate: Optional[float] = field(default=2e-5)
+    weight_decay: Optional[float] = field(default=0.001)
+    model_name: Optional[str] = field(
+        default="gpt2",
+        metadata={
+            "help": "The model that you want to train from the Hugging Face hub. E.g. gpt2, gpt2-xl, bert, etc."
+        },
+    )
+    tokenizer_name: Optional[str] = field(
+        default=None,
+        metadata={
+            "help": "The tokenizer for your model, if left empty will use the default for your model",
+        },
+    )
+    bf16: Optional[bool] = field(
+        default=True,
+        metadata={
+            "help": "This essentially cuts the training time in half if you want to sacrifice a little precision and have a supported GPU."
+        },
+    )
+    num_train_epochs: Optional[int] = field(
+        default=1,
+        metadata={"help": "The number of training epochs for the reward model."},
+    )
+    train_subset: Optional[int] = field(
+        default=100000,
+        metadata={"help": "The size of the subset of the training data to use"},
+    )
+    eval_subset: Optional[int] = field(
+        default=50000,
+        metadata={"help": "The size of the subset of the eval data to use"},
+    )
+    gradient_checkpointing: Optional[bool] = field(
+        default=False,
+        metadata={"help": "Enables gradient checkpointing."},
+    )
+    optim: Optional[str] = field(
+        default="adamw_hf",
+        metadata={"help": "The optimizer to use."},
+    )
+    lr_scheduler_type: Optional[str] = field(
+        default="linear",
+        metadata={"help": "The lr scheduler"},
+    )
+    max_length: Optional[int] = field(default=512)
+    eval_first_step: Optional[bool] = field(
+        default=False,
+        metadata={"help": "Whether to run eval after the first step"},
+    )
+
+
+parser = HfArgumentParser(ScriptArguments)
+script_args = parser.parse_args_into_dataclasses()[0]
+
+# Load the human stack-exchange-paired dataset for tuning the reward model.
+train_dataset = load_dataset("lvwerra/stack-exchange-paired", data_dir="data/reward", split="train")
+if script_args.train_subset > 0:
+    train_dataset = train_dataset.select(range(script_args.train_subset))
+eval_dataset = load_dataset("lvwerra/stack-exchange-paired", data_dir="data/evaluation", split="train")
+if script_args.eval_subset > 0:
+    eval_dataset = eval_dataset.select(range(script_args.eval_subset))
+# Define the training args. Needs to be done before the model is loaded if you are using deepspeed.
+model_name_split = script_args.model_name.split("/")[-1]
+output_name = (
+    f"{model_name_split}_peft_stack-exchange-paired_rmts__{script_args.train_subset}_{script_args.learning_rate}"
+)
+
+training_args = TrainingArguments(
+    output_dir=output_name,
+    learning_rate=script_args.learning_rate,
+    per_device_train_batch_size=script_args.per_device_train_batch_size,
+    per_device_eval_batch_size=script_args.per_device_eval_batch_size,
+    num_train_epochs=script_args.num_train_epochs,
+    weight_decay=script_args.weight_decay,
+    evaluation_strategy="steps",
+    eval_steps=500,
+    save_strategy="steps",
+    save_steps=500,
+    gradient_accumulation_steps=script_args.gradient_accumulation_steps,
+    gradient_checkpointing=script_args.gradient_checkpointing,
+    deepspeed=script_args.deepspeed,
+    local_rank=script_args.local_rank,
+    remove_unused_columns=False,
+    label_names=[],
+    bf16=script_args.bf16,
+    logging_strategy="steps",
+    logging_steps=10,
+    optim=script_args.optim,
+    lr_scheduler_type=script_args.lr_scheduler_type,
+)
+# Load the value-head model and tokenizer.
+tokenizer_name = script_args.tokenizer_name if script_args.tokenizer_name is not None else script_args.model_name
+tokenizer = AutoTokenizer.from_pretrained(tokenizer_name, use_auth_token=True)
+tokenizer.pad_token = tokenizer.eos_token
+
+
+peft_config = LoraConfig(
+    task_type=TaskType.SEQ_CLS,
+    inference_mode=False,
+    r=8,
+    lora_alpha=32,
+    lora_dropout=0.1,
+)
+
+model = AutoModelForSequenceClassification.from_pretrained(
+    script_args.model_name, num_labels=1, torch_dtype=torch.bfloat16
+)
+model = get_peft_model(model, peft_config)
+model.print_trainable_parameters()
+
+# Need to do this for gpt2, because it doesn't have an official pad token.
+tokenizer.pad_token = tokenizer.eos_token
+model.config.pad_token_id = tokenizer.eos_token_id
+model.config.use_cache = not script_args.gradient_checkpointing
+num_proc = 24  # Can adjust to be higher if you have more processors.
+original_columns = train_dataset.column_names
+
+
+# Turn the dataset into pairs of post + summaries, where text_j is the preferred question + answer and text_k is the other.
+# Then tokenize the dataset.
+def preprocess_function(examples):
+    new_examples = {
+        "input_ids_j": [],
+        "attention_mask_j": [],
+        "input_ids_k": [],
+        "attention_mask_k": [],
+    }
+    for question, response_j, response_k in zip(examples["question"], examples["response_j"], examples["response_k"]):
+        tokenized_j = tokenizer("Question: " + question + "\n\nAnswer: " + response_j, truncation=True)
+        tokenized_k = tokenizer("Question: " + question + "\n\nAnswer: " + response_k, truncation=True)
+
+        new_examples["input_ids_j"].append(tokenized_j["input_ids"])
+        new_examples["attention_mask_j"].append(tokenized_j["attention_mask"])
+        new_examples["input_ids_k"].append(tokenized_k["input_ids"])
+        new_examples["attention_mask_k"].append(tokenized_k["attention_mask"])
+
+    return new_examples
+
+
+# preprocess the dataset and filter out QAs that are longer than script_args.max_length
+train_dataset = train_dataset.map(
+    preprocess_function,
+    batched=True,
+    num_proc=num_proc,
+    remove_columns=original_columns,
+)
+train_dataset = train_dataset.filter(
+    lambda x: len(x["input_ids_j"]) <= script_args.max_length and len(x["input_ids_k"]) <= script_args.max_length
+)
+
+eval_dataset = eval_dataset.map(
+    preprocess_function,
+    batched=True,
+    num_proc=num_proc,
+    remove_columns=original_columns,
+)
+eval_dataset = eval_dataset.filter(
+    lambda x: len(x["input_ids_j"]) <= script_args.max_length and len(x["input_ids_k"]) <= script_args.max_length
+)
+
+
+# We need to define a special data collator that batches the data in our j vs k format.
+@dataclass
+class RewardDataCollatorWithPadding:
+    tokenizer: PreTrainedTokenizerBase
+    padding: Union[bool, str, PaddingStrategy] = True
+    max_length: Optional[int] = None
+    pad_to_multiple_of: Optional[int] = None
+    return_tensors: str = "pt"
+
+    def __call__(self, features: List[Dict[str, Any]]) -> Dict[str, Any]:
+        features_j = []
+        features_k = []
+        for feature in features:
+            features_j.append(
+                {
+                    "input_ids": feature["input_ids_j"],
+                    "attention_mask": feature["attention_mask_j"],
+                }
+            )
+            features_k.append(
+                {
+                    "input_ids": feature["input_ids_k"],
+                    "attention_mask": feature["attention_mask_k"],
+                }
+            )
+        batch_j = self.tokenizer.pad(
+            features_j,
+            padding=self.padding,
+            max_length=self.max_length,
+            pad_to_multiple_of=self.pad_to_multiple_of,
+            return_tensors=self.return_tensors,
+        )
+        batch_k = self.tokenizer.pad(
+            features_k,
+            padding=self.padding,
+            max_length=self.max_length,
+            pad_to_multiple_of=self.pad_to_multiple_of,
+            return_tensors=self.return_tensors,
+        )
+        batch = {
+            "input_ids_j": batch_j["input_ids"],
+            "attention_mask_j": batch_j["attention_mask"],
+            "input_ids_k": batch_k["input_ids"],
+            "attention_mask_k": batch_k["attention_mask"],
+            "return_loss": True,
+        }
+        return batch
+
+
+# Define the metric that we'll use for validation.
+accuracy = evaluate.load("accuracy")
+
+
+def compute_metrics(eval_pred):
+    predictions, _ = eval_pred
+    # Here, predictions is rewards_j and rewards_k.
+    # We want to see how much of the time rewards_j > rewards_k.
+    predictions = np.argmax(predictions, axis=0)
+    labels = np.zeros(predictions.shape)
+    return accuracy.compute(predictions=predictions, references=labels)
+
+
+class RewardTrainer(Trainer):
+    # Define how to compute the reward loss. We use the InstructGPT pairwise logloss: https://arxiv.org/abs/2203.02155
+    def compute_loss(self, model, inputs, return_outputs=False):
+        rewards_j = model(input_ids=inputs["input_ids_j"], attention_mask=inputs["attention_mask_j"])[0]
+        rewards_k = model(input_ids=inputs["input_ids_k"], attention_mask=inputs["attention_mask_k"])[0]
+        loss = -nn.functional.logsigmoid(rewards_j - rewards_k).mean()
+        if return_outputs:
+            return loss, {"rewards_j": rewards_j, "rewards_k": rewards_k}
+        return loss
+
+
+# Train the model, woohoo.
+trainer = RewardTrainer(
+    model=model,
+    args=training_args,
+    train_dataset=train_dataset,
+    eval_dataset=eval_dataset,
+    compute_metrics=compute_metrics,
+    data_collator=RewardDataCollatorWithPadding(tokenizer=tokenizer, max_length=script_args.max_length),
+)
+
+
+if script_args.eval_first_step:
+
+    class EvaluateFirstStepCallback(TrainerCallback):
+        def on_step_end(self, args, state, control, **kwargs):
+            if state.global_step == 1:
+                control.should_evaluate = True
+
+    trainer.add_callback(EvaluateFirstStepCallback())
+
+trainer.train(script_args.resume_from_checkpoint)
+
+print("Saving last checkpoint of the model")
+model.save_pretrained(output_name + "_peft_last_checkpoint")

examples/research_projects/stack_llama/scripts/rl_training.py

@@ -0,0 +1,263 @@
+# coding=utf-8
+# Copyright 2022 The HuggingFace Inc. 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 dataclasses import dataclass, field
+from typing import Optional
+
+import torch
+from accelerate import Accelerator
+from datasets import load_dataset
+from peft import LoraConfig
+from tqdm import tqdm
+from transformers import Adafactor, AutoTokenizer, HfArgumentParser, pipeline
+
+from trl import AutoModelForCausalLMWithValueHead, PPOConfig, PPOTrainer, set_seed
+from trl.core import LengthSampler
+
+
+tqdm.pandas()
+
+
+@dataclass
+class ScriptArguments:
+    """
+    The name of the Casual LM model we wish to fine with PPO
+    """
+
+    # NOTE: gpt2 models use Conv1D instead of Linear layers which are not yet supported in 8 bit mode
+    # models like gpt-neo* models are more suitable.
+    model_name: Optional[str] = field(default="", metadata={"help": "the model name"})
+    tokenizer_name: Optional[str] = field(default="", metadata={"help": "the tokenizer name"})
+    reward_model_name: Optional[str] = field(default="", metadata={"help": "the reward model name"})
+    log_with: Optional[str] = field(default=None, metadata={"help": "use 'wandb' to log with wandb"})
+    learning_rate: Optional[float] = field(default=1.41e-5, metadata={"help": "the learning rate"})
+    output_max_length: Optional[int] = field(default=128, metadata={"help": "maximum length for generation"})
+    mini_batch_size: Optional[int] = field(default=1, metadata={"help": "the PPO minibatch size"})
+    batch_size: Optional[int] = field(default=32, metadata={"help": "the batch size"})
+    ppo_epochs: Optional[int] = field(default=4, metadata={"help": "the number of ppo epochs"})
+    gradient_accumulation_steps: Optional[int] = field(
+        default=4, metadata={"help": "the number of gradient accumulation steps"}
+    )
+    adafactor: Optional[bool] = field(default=False, metadata={"help": "whether to use the adafactor optimizer"})
+    early_stopping: Optional[bool] = field(default=False, metadata={"help": "whether to early stop"})
+    target_kl: Optional[float] = field(default=0.1, metadata={"help": "kl target for early stopping"})
+    reward_baseline: Optional[float] = field(
+        default=0.0,
+        metadata={"help": "a baseline value that is subtracted from the reward"},
+    )
+    batched_gen: Optional[bool] = field(default=False, metadata={"help": "whether to use the batched text gen"})
+    save_freq: Optional[int] = field(default=None, metadata={"help": "n steps to save the model"})
+    output_dir: Optional[str] = field(default="runs/", metadata={"help": "n steps to save the model"})
+    seed: Optional[int] = field(default=0, metadata={"help": "the seed"})
+    steps: Optional[int] = field(default=20000, metadata={"help": "number of epochs"})
+    init_kl_coef: Optional[float] = field(
+        default=0.2,
+        metadata={"help": "Initial KL penalty coefficient (used for adaptive and linear control)"},
+    )
+
+    adap_kl_ctrl: Optional[bool] = field(default=True, metadata={"help": "Use adaptive KL control, otherwise linear"})
+
+
+parser = HfArgumentParser(ScriptArguments)
+script_args: ScriptArguments = parser.parse_args_into_dataclasses()[0]
+reward_model_name = script_args.reward_model_name
+dataset_name = "lvwerra/stack-exchange-paired"
+config = PPOConfig(
+    steps=script_args.steps,
+    model_name=script_args.model_name,
+    learning_rate=script_args.learning_rate,
+    log_with=script_args.log_with,
+    batch_size=script_args.batch_size,
+    mini_batch_size=script_args.mini_batch_size,
+    gradient_accumulation_steps=script_args.gradient_accumulation_steps,
+    optimize_cuda_cache=True,
+    early_stopping=script_args.early_stopping,
+    target_kl=script_args.target_kl,
+    ppo_epochs=script_args.ppo_epochs,
+    seed=script_args.seed,
+    init_kl_coef=script_args.init_kl_coef,
+    adap_kl_ctrl=script_args.adap_kl_ctrl,
+)
+
+train_dataset = load_dataset("lvwerra/stack-exchange-paired", data_dir="data/rl", split="train")
+train_dataset = train_dataset.select(range(100000))
+original_columns = train_dataset.column_names
+
+# We then define the arguments to pass to the sentiment analysis pipeline.
+# We set `return_all_scores` to True to get the sentiment score for each token.
+sent_kwargs = {
+    "return_all_scores": True,
+    "function_to_apply": "none",
+    "batch_size": 16,
+    "truncation": True,
+}
+
+tokenizer = AutoTokenizer.from_pretrained(script_args.tokenizer_name)
+# GPT-2 tokenizer has a pad token, but it is not eos_token by default. We need to set it to eos_token.
+# only for this model.
+
+if getattr(tokenizer, "pad_token", None) is None:
+    tokenizer.pad_token = tokenizer.eos_token
+
+
+# Below is an example function to build the dataset. In our case, we use the IMDB dataset
+# from the `datasets` library. One should customize this function to train the model on
+# its own dataset.
+def build_dataset(
+    tokenizer,
+    dataset_name="lvwerra/stack-exchange-paired",
+):
+    """
+    Build dataset for training. This builds the dataset from `load_dataset`, one should
+    customize this function to train the model on its own dataset.
+
+    Args:
+        dataset_name (`str`):
+            The name of the dataset to be loaded.
+
+    Returns:
+        dataloader (`torch.utils.data.DataLoader`):
+            The dataloader for the dataset.
+    """
+
+    num_proc = 24
+
+    def preprocess_function(examples):
+        new_examples = {
+            "query": [],
+            "input_ids": [],
+        }
+        for question in examples["question"]:
+            query = "Question: " + question + "\n\nAnswer: "
+            tokenized_question = tokenizer(query, truncation=True)
+            new_examples["query"].append(query)
+            new_examples["input_ids"].append(tokenized_question["input_ids"])
+
+        return new_examples
+
+    ds = train_dataset.map(
+        preprocess_function,
+        batched=True,
+        num_proc=num_proc,
+        remove_columns=original_columns,
+    )
+    ds = ds.filter(lambda x: len(x["input_ids"]) < 512, batched=False)
+
+    ds.set_format(type="torch")
+    return ds
+
+
+# We retrieve the dataloader by calling the `build_dataset` function.
+dataset = build_dataset(tokenizer)
+
+
+def collator(data):
+    return dict((key, [d[key] for d in data]) for key in data[0])
+
+
+# set seed before initializing value head for deterministic eval
+set_seed(config.seed)
+
+# Now let's build the model, the reference model, and the tokenizer.
+current_device = Accelerator().local_process_index
+
+lora_config = LoraConfig(
+    r=16,
+    lora_alpha=32,
+    lora_dropout=0.05,
+    bias="none",
+    task_type="CAUSAL_LM",
+)
+model = AutoModelForCausalLMWithValueHead.from_pretrained(
+    config.model_name,
+    load_in_8bit=True,
+    device_map={"": current_device},
+    peft_config=lora_config,
+)
+
+optimizer = None
+if script_args.adafactor:
+    optimizer = Adafactor(
+        filter(lambda p: p.requires_grad, model.parameters()),
+        scale_parameter=False,
+        relative_step=False,
+        warmup_init=False,
+        lr=config.learning_rate,
+    )
+# We then build the PPOTrainer, passing the model, the reference model, the tokenizer
+ppo_trainer = PPOTrainer(
+    config,
+    model,
+    ref_model=None,
+    tokenizer=tokenizer,
+    dataset=dataset,
+    data_collator=collator,
+    optimizer=optimizer,
+)
+
+# We then build the sentiment analysis pipeline using our reward model, passing the
+# model name and the sentiment analysis pipeline arguments. Let's also make sure to
+# set the device to the same device as the PPOTrainer.
+device = ppo_trainer.accelerator.device
+if ppo_trainer.accelerator.num_processes == 1:
+    device = 0 if torch.cuda.is_available() else "cpu"  # to avoid a ` pipeline` bug
+sentiment_pipe = pipeline(
+    "sentiment-analysis",
+    model=reward_model_name,
+    device_map={"": current_device},
+    model_kwargs={"load_in_8bit": True},
+    tokenizer=tokenizer,
+    return_token_type_ids=False,
+)
+
+# We then define the arguments to pass to the `generate` function. These arguments
+# are passed to the `generate` function of the PPOTrainer, which is a wrapper around
+# the `generate` function of the trained model.
+generation_kwargs = {
+    # "min_length": -1,
+    "top_k": 0.0,
+    "top_p": 1.0,
+    "do_sample": True,
+    "pad_token_id": tokenizer.pad_token_id,
+    "eos_token_id": 100_000,
+}
+output_min_length = 32
+output_max_length = script_args.output_max_length
+output_length_sampler = LengthSampler(output_min_length, output_max_length)
+
+for epoch, batch in tqdm(enumerate(ppo_trainer.dataloader)):
+    if epoch >= config.total_ppo_epochs:
+        break
+
+    question_tensors = batch["input_ids"]
+
+    response_tensors = ppo_trainer.generate(
+        question_tensors,
+        return_prompt=False,
+        length_sampler=output_length_sampler,
+        **generation_kwargs,
+    )
+    batch["response"] = tokenizer.batch_decode(response_tensors, skip_special_tokens=True)
+
+    # Compute reward score (using the sentiment analysis pipeline)
+    texts = [q + r for q, r in zip(batch["query"], batch["response"])]
+    pipe_outputs = sentiment_pipe(texts, **sent_kwargs)
+    rewards = [torch.tensor(output[0]["score"] - script_args.reward_baseline) for output in pipe_outputs]
+
+    # Run PPO step
+    stats = ppo_trainer.step(question_tensors, response_tensors, rewards)
+    ppo_trainer.log_stats(stats, batch, rewards)
+
+    if script_args.save_freq and epoch and epoch % script_args.save_freq == 0:
+        ppo_trainer.save_pretrained(script_args.output_dir + f"step_{epoch}")

examples/research_projects/stack_llama/scripts/supervised_finetuning.py

@@ -0,0 +1,208 @@
+import argparse
+import os
+
+from accelerate import Accelerator
+from datasets import load_dataset
+from peft import LoraConfig
+from tqdm import tqdm
+from transformers import AutoModelForCausalLM, AutoTokenizer, TrainingArguments, logging, set_seed
+
+from trl import SFTTrainer
+from trl.trainer import ConstantLengthDataset
+
+
+"""
+Fine-Tune Llama-7b on SE paired dataset
+"""
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--model_path", type=str, default="")
+    parser.add_argument("--dataset_name", type=str, default="lvwerra/stack-exchange-paired")
+    parser.add_argument("--subset", type=str, default="data/finetune")
+    parser.add_argument("--split", type=str, default="train")
+    parser.add_argument("--size_valid_set", type=int, default=4000)
+    parser.add_argument("--streaming", action="store_true")
+    parser.add_argument("--shuffle_buffer", type=int, default=5000)
+
+    parser.add_argument("--seq_length", type=int, default=1024)
+    parser.add_argument("--max_steps", type=int, default=10000)
+    parser.add_argument("--batch_size", type=int, default=4)
+    parser.add_argument("--gradient_accumulation_steps", type=int, default=1)
+    parser.add_argument("--eos_token_id", type=int, default=49152)
+
+    parser.add_argument("--learning_rate", type=float, default=1e-4)
+    parser.add_argument("--lr_scheduler_type", type=str, default="cosine")
+    parser.add_argument("--num_warmup_steps", type=int, default=100)
+    parser.add_argument("--weight_decay", type=float, default=0.05)
+
+    parser.add_argument("--local_rank", type=int, default=0)
+    parser.add_argument("--fp16", action="store_true", default=False)
+    parser.add_argument("--bf16", action="store_true", default=False)
+    parser.add_argument("--gradient_checkpointing", action="store_true", default=False)
+    parser.add_argument("--seed", type=int, default=0)
+    parser.add_argument("--num_workers", type=int, default=None)
+    parser.add_argument("--output_dir", type=str, default="./checkpoints")
+    parser.add_argument("--log_freq", default=1, type=int)
+    parser.add_argument("--eval_freq", default=1000, type=int)
+    parser.add_argument("--save_freq", default=1000, type=int)
+
+    return parser.parse_args()
+
+
+def chars_token_ratio(dataset, tokenizer, nb_examples=400):
+    """
+    Estimate the average number of characters per token in the dataset.
+    """
+    total_characters, total_tokens = 0, 0
+    for _, example in tqdm(zip(range(nb_examples), iter(dataset)), total=nb_examples):
+        text = prepare_sample_text(example)
+        total_characters += len(text)
+        if tokenizer.is_fast:
+            total_tokens += len(tokenizer(text).tokens())
+        else:
+            total_tokens += len(tokenizer.tokenize(text))
+
+    return total_characters / total_tokens
+
+
+def print_trainable_parameters(model):
+    """
+    Prints the number of trainable parameters in the model.
+    """
+    trainable_params = 0
+    all_param = 0
+    for _, param in model.named_parameters():
+        all_param += param.numel()
+        if param.requires_grad:
+            trainable_params += param.numel()
+    print(
+        f"trainable params: {trainable_params} || all params: {all_param} || trainable%: {100 * trainable_params / all_param}"
+    )
+
+
+def prepare_sample_text(example):
+    """Prepare the text from a sample of the dataset."""
+    text = f"Question: {example['question']}\n\nAnswer: {example['response_j']}"
+    return text
+
+
+def create_datasets(tokenizer, args):
+    dataset = load_dataset(
+        args.dataset_name,
+        data_dir=args.subset,
+        split=args.split,
+        use_auth_token=True,
+        num_proc=args.num_workers if not args.streaming else None,
+        streaming=args.streaming,
+    )
+    if args.streaming:
+        print("Loading the dataset in streaming mode")
+        valid_data = dataset.take(args.size_valid_set)
+        train_data = dataset.skip(args.size_valid_set)
+        train_data = train_data.shuffle(buffer_size=args.shuffle_buffer, seed=args.seed)
+    else:
+        dataset = dataset.train_test_split(test_size=0.005, seed=args.seed)
+        train_data = dataset["train"]
+        valid_data = dataset["test"]
+        print(f"Size of the train set: {len(train_data)}. Size of the validation set: {len(valid_data)}")
+
+    chars_per_token = chars_token_ratio(train_data, tokenizer)
+    print(f"The character to token ratio of the dataset is: {chars_per_token:.2f}")
+
+    train_dataset = ConstantLengthDataset(
+        tokenizer,
+        train_data,
+        formatting_func=prepare_sample_text,
+        infinite=True,
+        seq_length=args.seq_length,
+        chars_per_token=chars_per_token,
+    )
+    valid_dataset = ConstantLengthDataset(
+        tokenizer,
+        valid_data,
+        formatting_func=prepare_sample_text,
+        infinite=False,
+        seq_length=args.seq_length,
+        chars_per_token=chars_per_token,
+    )
+    return train_dataset, valid_dataset
+
+
+def run_training(args, train_data, val_data):
+    print("Loading the model")
+
+    lora_config = LoraConfig(
+        r=16,
+        lora_alpha=32,
+        lora_dropout=0.05,
+        bias="none",
+        task_type="CAUSAL_LM",
+    )
+
+    train_data.start_iteration = 0
+
+    print("Starting main loop")
+
+    training_args = TrainingArguments(
+        output_dir=args.output_dir,
+        dataloader_drop_last=True,
+        evaluation_strategy="steps",
+        max_steps=args.max_steps,
+        eval_steps=args.eval_freq,
+        save_steps=args.save_freq,
+        logging_steps=args.log_freq,
+        per_device_train_batch_size=args.batch_size,
+        per_device_eval_batch_size=args.batch_size,
+        learning_rate=args.learning_rate,
+        lr_scheduler_type=args.lr_scheduler_type,
+        warmup_steps=args.num_warmup_steps,
+        gradient_accumulation_steps=args.gradient_accumulation_steps,
+        gradient_checkpointing=args.gradient_checkpointing,
+        fp16=args.fp16,
+        bf16=args.bf16,
+        weight_decay=args.weight_decay,
+        run_name="llama-7b-finetuned",
+        report_to="wandb",
+        ddp_find_unused_parameters=False,
+    )
+
+    model = AutoModelForCausalLM.from_pretrained(
+        args.model_path, load_in_8bit=True, device_map={"": Accelerator().process_index}
+    )
+
+    trainer = SFTTrainer(
+        model=model,
+        args=training_args,
+        train_dataset=train_data,
+        eval_dataset=val_data,
+        peft_config=lora_config,
+        packing=True,
+    )
+
+    print_trainable_parameters(trainer.model)
+
+    print("Training...")
+    trainer.train()
+
+    print("Saving last checkpoint of the model")
+    trainer.model.save_pretrained(os.path.join(args.output_dir, "final_checkpoint/"))
+
+
+def main(args):
+    tokenizer = AutoTokenizer.from_pretrained(args.model_path)
+    train_dataset, eval_dataset = create_datasets(tokenizer, args)
+    run_training(args, train_dataset, eval_dataset)
+
+
+if __name__ == "__main__":
+    args = get_args()
+    assert args.model_path != "", "Please provide the llama model path"
+
+    set_seed(args.seed)
+    os.makedirs(args.output_dir, exist_ok=True)
+
+    logging.set_verbosity_error()
+
+    main(args)

examples/research_projects/stack_llama_2/scripts/README.md

@@ -0,0 +1,76 @@
+# DPO pipeline for the creation of StackLlaMa 2: a Stack exchange llama-v2-7b model
+
+## Prerequisites
+
+Install all the dependencies in the `requirements.txt`:
+
+```
+$ pip install -U -r requirements.txt
+```
+
+Since we will use `accelerate` for training, make sure to run:
+```
+$ accelerate config
+```
+
+## Training
+
+There were two main steps to the DPO training process:
+1. Supervised fine-tuning of the base llama-v2-7b model to create llama-v2-7b-se:
+
+    ```
+    accelerate launch examples/research_projects/stack_llama_2/scripts/sft_llama2.py \
+        --output_dir="./sft" \
+        --max_steps=500 \
+        --logging_steps=10 \
+        --save_steps=10 \
+        --per_device_train_batch_size=4 \
+        --per_device_eval_batch_size=1 \
+        --gradient_accumulation_steps=2 \
+        --gradient_checkpointing=False \
+        --group_by_length=False \
+        --learning_rate=1e-4 \
+        --lr_scheduler_type="cosine" \
+        --warmup_steps=100 \
+        --weight_decay=0.05 \
+        --optim="paged_adamw_32bit" \
+        --bf16=True \
+        --remove_unused_columns=False \
+        --run_name="sft_llama2" \
+        --report_to="wandb"
+    ```
+1. Run the DPO trainer using the model saved by the previous step:
+    ```
+    accelerate launch examples/research_projects/stack_llama_2/scripts/dpo_llama2.py \
+        --model_name_or_path="sft/final_checkpoint" \
+        --output_dir="dpo"
+    ```
+
+
+## Merging the adaptors
+
+To merge the adaptors into the base model we can use the `merge_peft_adapter.py` helper script that comes with TRL:
+
+```
+python examples/research_projects/stack_llama/scripts/merge_peft_adapter.py --base_model_name="meta-llama/Llama-2-7b-hf" --adapter_model_name="dpo/final_checkpoint/" --output_name="stack-llama-2"
+```
+
+which will also push the model to your HuggingFace hub account.
+
+## Running the model
+
+We can load the DPO-trained LoRA adaptors which were saved by the DPO training step and load them via:
+
+```py
+from peft import AutoPeftModelForCausalLM
+
+
+model = AutoPeftModelForCausalLM.from_pretrained(
+    "dpo/final_checkpoint",
+    low_cpu_mem_usage=True,
+    torch_dtype=torch.float16,
+    load_in_4bit=True,
+)
+
+model.generate(...)
+```

examples/research_projects/stack_llama_2/scripts/dpo_llama2.py

@@ -0,0 +1,223 @@
+# 0. imports
+import os
+from dataclasses import dataclass, field
+from typing import Dict, Optional
+
+import torch
+from datasets import Dataset, load_dataset
+from peft import LoraConfig
+from transformers import AutoModelForCausalLM, AutoTokenizer, HfArgumentParser, TrainingArguments
+
+from trl import DPOTrainer
+
+
+# Define and parse arguments.
+@dataclass
+class ScriptArguments:
+    """
+    The arguments for the DPO training script.
+    """
+
+    # data parameters
+    beta: Optional[float] = field(default=0.1, metadata={"help": "the beta parameter for DPO loss"})
+
+    # training parameters
+    model_name_or_path: Optional[str] = field(
+        default="../sft/results/final_checkpoint",
+        metadata={"help": "the location of the SFT model name or path"},
+    )
+    learning_rate: Optional[float] = field(default=5e-4, metadata={"help": "optimizer learning rate"})
+    lr_scheduler_type: Optional[str] = field(default="cosine", metadata={"help": "the lr scheduler type"})
+    warmup_steps: Optional[int] = field(default=100, metadata={"help": "the number of warmup steps"})
+    weight_decay: Optional[float] = field(default=0.05, metadata={"help": "the weight decay"})
+    optimizer_type: Optional[str] = field(default="paged_adamw_32bit", metadata={"help": "the optimizer type"})
+
+    per_device_train_batch_size: Optional[int] = field(default=4, metadata={"help": "train batch size per device"})
+    per_device_eval_batch_size: Optional[int] = field(default=1, metadata={"help": "eval batch size per device"})
+    gradient_accumulation_steps: Optional[int] = field(
+        default=4, metadata={"help": "the number of gradient accumulation steps"}
+    )
+    gradient_checkpointing: Optional[bool] = field(
+        default=True, metadata={"help": "whether to use gradient checkpointing"}
+    )
+
+    lora_alpha: Optional[float] = field(default=16, metadata={"help": "the lora alpha parameter"})
+    lora_dropout: Optional[float] = field(default=0.05, metadata={"help": "the lora dropout parameter"})
+    lora_r: Optional[int] = field(default=8, metadata={"help": "the lora r parameter"})
+
+    max_prompt_length: Optional[int] = field(default=512, metadata={"help": "the maximum prompt length"})
+    max_length: Optional[int] = field(default=1024, metadata={"help": "the maximum sequence length"})
+    max_steps: Optional[int] = field(default=1000, metadata={"help": "max number of training steps"})
+    logging_steps: Optional[int] = field(default=10, metadata={"help": "the logging frequency"})
+    save_steps: Optional[int] = field(default=100, metadata={"help": "the saving frequency"})
+    eval_steps: Optional[int] = field(default=100, metadata={"help": "the evaluation frequency"})
+
+    output_dir: Optional[str] = field(default="./results", metadata={"help": "the output directory"})
+    log_freq: Optional[int] = field(default=1, metadata={"help": "the logging frequency"})
+
+    # instrumentation
+    sanity_check: Optional[bool] = field(default=False, metadata={"help": "only train on 1000 samples"})
+    report_to: Optional[str] = field(
+        default="wandb",
+        metadata={
+            "help": 'The list of integrations to report the results and logs to. Supported platforms are `"azure_ml"`,'
+            '`"comet_ml"`, `"mlflow"`, `"neptune"`, `"tensorboard"`,`"clearml"` and `"wandb"`. '
+            'Use `"all"` to report to all integrations installed, `"none"` for no integrations.'
+        },
+    )
+    # debug argument for distributed training
+    ignore_bias_buffers: Optional[bool] = field(
+        default=False,
+        metadata={
+            "help": "fix for DDP issues with LM bias/mask buffers - invalid scalar type,`inplace operation. See"
+            "https://github.com/huggingface/transformers/issues/22482#issuecomment-1595790992"
+        },
+    )
+
+
+def get_stack_exchange_paired(
+    data_dir: str = "data/rl",
+    sanity_check: bool = False,
+    cache_dir: str = None,
+    num_proc=24,
+) -> Dataset:
+    """Load the stack-exchange-paired dataset from Hugging Face and convert it to the necessary format.
+
+    The dataset is converted to a dictionary with the following structure:
+    {
+        'prompt': List[str],
+        'chosen': List[str],
+        'rejected': List[str],
+    }
+
+    Prompts are structured as follows:
+      "Question: " + <prompt> + "\n\nAnswer: "
+    """
+    dataset = load_dataset(
+        "lvwerra/stack-exchange-paired",
+        split="train",
+        cache_dir=cache_dir,
+        data_dir=data_dir,
+    )
+    original_columns = dataset.column_names
+
+    if sanity_check:
+        dataset = dataset.select(range(min(len(dataset), 1000)))
+
+    def return_prompt_and_responses(samples) -> Dict[str, str]:
+        return {
+            "prompt": ["Question: " + question + "\n\nAnswer: " for question in samples["question"]],
+            "chosen": samples["response_j"],
+            "rejected": samples["response_k"],
+        }
+
+    return dataset.map(
+        return_prompt_and_responses,
+        batched=True,
+        num_proc=num_proc,
+        remove_columns=original_columns,
+    )
+
+
+if __name__ == "__main__":
+    parser = HfArgumentParser(ScriptArguments)
+    script_args = parser.parse_args_into_dataclasses()[0]
+
+    # 1. load a pretrained model
+    model = AutoModelForCausalLM.from_pretrained(
+        script_args.model_name_or_path,
+        low_cpu_mem_usage=True,
+        torch_dtype=torch.float16,
+        load_in_4bit=True,
+    )
+    model.config.use_cache = False
+
+    if script_args.ignore_bias_buffers:
+        # torch distributed hack
+        model._ddp_params_and_buffers_to_ignore = [
+            name for name, buffer in model.named_buffers() if buffer.dtype == torch.bool
+        ]
+
+    model_ref = AutoModelForCausalLM.from_pretrained(
+        script_args.model_name_or_path,
+        low_cpu_mem_usage=True,
+        torch_dtype=torch.float16,
+        load_in_4bit=True,
+    )
+    tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-2-7b-hf")
+    tokenizer.pad_token = tokenizer.eos_token
+
+    # 2. Load the Stack-exchange paired dataset
+    train_dataset = get_stack_exchange_paired(data_dir="data/rl", sanity_check=script_args.sanity_check)
+    train_dataset = train_dataset.filter(
+        lambda x: len(x["prompt"]) + len(x["chosen"]) <= script_args.max_length
+        and len(x["prompt"]) + len(x["rejected"]) <= script_args.max_length
+    )
+
+    # 3. Load evaluation dataset
+    eval_dataset = get_stack_exchange_paired(data_dir="data/evaluation", sanity_check=True)
+    eval_dataset = eval_dataset.filter(
+        lambda x: len(x["prompt"]) + len(x["chosen"]) <= script_args.max_length
+        and len(x["prompt"]) + len(x["rejected"]) <= script_args.max_length
+    )
+
+    # 4. initialize training arguments:
+    training_args = TrainingArguments(
+        per_device_train_batch_size=script_args.per_device_train_batch_size,
+        per_device_eval_batch_size=script_args.per_device_eval_batch_size,
+        max_steps=script_args.max_steps,
+        logging_steps=script_args.logging_steps,
+        save_steps=script_args.save_steps,
+        gradient_accumulation_steps=script_args.gradient_accumulation_steps,
+        gradient_checkpointing=script_args.gradient_checkpointing,
+        learning_rate=script_args.learning_rate,
+        evaluation_strategy="steps",
+        eval_steps=script_args.eval_steps,
+        output_dir=script_args.output_dir,
+        report_to=script_args.report_to,
+        lr_scheduler_type=script_args.lr_scheduler_type,
+        warmup_steps=script_args.warmup_steps,
+        optim=script_args.optimizer_type,
+        bf16=True,
+        remove_unused_columns=False,
+        run_name="dpo_llama2",
+    )
+
+    peft_config = LoraConfig(
+        r=script_args.lora_r,
+        lora_alpha=script_args.lora_alpha,
+        lora_dropout=script_args.lora_dropout,
+        target_modules=[
+            "q_proj",
+            "v_proj",
+            "k_proj",
+            "out_proj",
+            "fc_in",
+            "fc_out",
+            "wte",
+        ],
+        bias="none",
+        task_type="CAUSAL_LM",
+    )
+
+    # 5. initialize the DPO trainer
+    dpo_trainer = DPOTrainer(
+        model,
+        model_ref,
+        args=training_args,
+        beta=script_args.beta,
+        train_dataset=train_dataset,
+        eval_dataset=eval_dataset,
+        tokenizer=tokenizer,
+        peft_config=peft_config,
+        max_prompt_length=script_args.max_prompt_length,
+        max_length=script_args.max_length,
+    )
+
+    # 6. train
+    dpo_trainer.train()
+    dpo_trainer.save_model(script_args.output_dir)
+
+    # 7. save
+    output_dir = os.path.join(script_args.output_dir, "final_checkpoint")
+    dpo_trainer.model.save_pretrained(output_dir)

examples/research_projects/stack_llama_2/scripts/requirements.txt

@@ -0,0 +1,7 @@
+transformers
+trl
+peft
+accelerate
+datasets
+bitsandbytes
+wandb

examples/research_projects/stack_llama_2/scripts/sft_llama2.py

@@ -0,0 +1,187 @@
+# Fine-Tune Llama2-7b on SE paired dataset
+import os
+from dataclasses import dataclass, field
+from typing import Optional
+
+import torch
+from accelerate import Accelerator
+from datasets import load_dataset
+from peft import AutoPeftModelForCausalLM, LoraConfig
+from tqdm import tqdm
+from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig, HfArgumentParser, TrainingArguments
+
+from trl import SFTTrainer
+from trl.import_utils import is_npu_available, is_xpu_available
+from trl.trainer import ConstantLengthDataset
+
+
+@dataclass
+class ScriptArguments:
+    model_name: Optional[str] = field(default="meta-llama/Llama-2-7b-hf", metadata={"help": "the model name"})
+    dataset_name: Optional[str] = field(default="lvwerra/stack-exchange-paired", metadata={"help": "the dataset name"})
+    subset: Optional[str] = field(default="data/finetune", metadata={"help": "the subset to use"})
+    split: Optional[str] = field(default="train", metadata={"help": "the split to use"})
+    size_valid_set: Optional[int] = field(default=4000, metadata={"help": "the size of the validation set"})
+    streaming: Optional[bool] = field(default=True, metadata={"help": "whether to stream the dataset"})
+    shuffle_buffer: Optional[int] = field(default=5000, metadata={"help": "the shuffle buffer size"})
+    seq_length: Optional[int] = field(default=1024, metadata={"help": "the sequence length"})
+    num_workers: Optional[int] = field(default=4, metadata={"help": "the number of workers"})
+    packing: Optional[bool] = field(default=True, metadata={"help": "whether to use packing for SFTTrainer"})
+
+    # LoraConfig
+    lora_alpha: Optional[float] = field(default=16, metadata={"help": "the lora alpha parameter"})
+    lora_dropout: Optional[float] = field(default=0.05, metadata={"help": "the lora dropout parameter"})
+    lora_r: Optional[int] = field(default=8, metadata={"help": "the lora r parameter"})
+
+
+parser = HfArgumentParser((ScriptArguments, TrainingArguments))
+script_args, training_args = parser.parse_args_into_dataclasses()
+peft_config = LoraConfig(
+    r=script_args.lora_r,
+    lora_alpha=script_args.lora_alpha,
+    lora_dropout=script_args.lora_dropout,
+    target_modules=["q_proj", "v_proj"],
+    bias="none",
+    task_type="CAUSAL_LM",
+)
+
+if training_args.group_by_length and script_args.packing:
+    raise ValueError("Cannot use both packing and group by length")
+
+# `gradient_checkpointing` was True by default until `1f3314`, but it's actually not used.
+# `gradient_checkpointing=True` will cause `Variable._execution_engine.run_backward`.
+if training_args.gradient_checkpointing:
+    raise ValueError("gradient_checkpointing not supported")
+
+
+def chars_token_ratio(dataset, tokenizer, nb_examples=400):
+    """
+    Estimate the average number of characters per token in the dataset.
+    """
+    total_characters, total_tokens = 0, 0
+    for _, example in tqdm(zip(range(nb_examples), iter(dataset)), total=nb_examples):
+        text = prepare_sample_text(example)
+        total_characters += len(text)
+        if tokenizer.is_fast:
+            total_tokens += len(tokenizer(text).tokens())
+        else:
+            total_tokens += len(tokenizer.tokenize(text))
+
+    return total_characters / total_tokens
+
+
+def print_trainable_parameters(model):
+    """
+    Prints the number of trainable parameters in the model.
+    """
+    trainable_params = 0
+    all_param = 0
+    for _, param in model.named_parameters():
+        all_param += param.numel()
+        if param.requires_grad:
+            trainable_params += param.numel()
+    print(
+        f"trainable params: {trainable_params} || all params: {all_param} || trainable%: {100 * trainable_params / all_param}"
+    )
+
+
+def prepare_sample_text(example):
+    """Prepare the text from a sample of the dataset."""
+    text = f"Question: {example['question']}\n\nAnswer: {example['response_j']}"
+    return text
+
+
+def create_datasets(tokenizer, args):
+    dataset = load_dataset(
+        args.dataset_name,
+        data_dir=args.subset,
+        split=args.split,
+        use_auth_token=True,
+        num_proc=args.num_workers if not args.streaming else None,
+        streaming=args.streaming,
+    )
+    if args.streaming:
+        print("Loading the dataset in streaming mode")
+        valid_data = dataset.take(args.size_valid_set)
+        train_data = dataset.skip(args.size_valid_set)
+        train_data = train_data.shuffle(buffer_size=args.shuffle_buffer, seed=None)
+    else:
+        dataset = dataset.train_test_split(test_size=0.005, seed=None)
+        train_data = dataset["train"]
+        valid_data = dataset["test"]
+        print(f"Size of the train set: {len(train_data)}. Size of the validation set: {len(valid_data)}")
+
+    chars_per_token = chars_token_ratio(train_data, tokenizer)
+    print(f"The character to token ratio of the dataset is: {chars_per_token:.2f}")
+
+    train_dataset = ConstantLengthDataset(
+        tokenizer,
+        train_data,
+        formatting_func=prepare_sample_text,
+        infinite=True,
+        seq_length=args.seq_length,
+        chars_per_token=chars_per_token,
+    )
+    valid_dataset = ConstantLengthDataset(
+        tokenizer,
+        valid_data,
+        formatting_func=prepare_sample_text,
+        infinite=False,
+        seq_length=args.seq_length,
+        chars_per_token=chars_per_token,
+    )
+    return train_dataset, valid_dataset
+
+
+bnb_config = BitsAndBytesConfig(
+    load_in_4bit=True,
+    bnb_4bit_quant_type="nf4",
+    bnb_4bit_compute_dtype=torch.bfloat16,
+)
+
+base_model = AutoModelForCausalLM.from_pretrained(
+    script_args.model_name,
+    quantization_config=bnb_config,
+    device_map={"": Accelerator().local_process_index},
+    trust_remote_code=True,
+    use_auth_token=True,
+)
+base_model.config.use_cache = False
+
+
+tokenizer = AutoTokenizer.from_pretrained(script_args.model_name, trust_remote_code=True)
+tokenizer.pad_token = tokenizer.eos_token
+tokenizer.padding_side = "right"  # Fix weird overflow issue with fp16 training
+
+train_dataset, eval_dataset = create_datasets(tokenizer, script_args)
+
+trainer = SFTTrainer(
+    model=base_model,
+    train_dataset=train_dataset,
+    eval_dataset=eval_dataset,
+    peft_config=peft_config,
+    packing=script_args.packing,
+    max_seq_length=None,
+    tokenizer=tokenizer,
+    args=training_args,
+)
+trainer.train()
+trainer.save_model(training_args.output_dir)
+
+output_dir = os.path.join(training_args.output_dir, "final_checkpoint")
+trainer.model.save_pretrained(output_dir)
+
+# Free memory for merging weights
+del base_model
+if is_xpu_available():
+    torch.xpu.empty_cache()
+elif is_npu_available():
+    torch.npu.empty_cache()
+else:
+    torch.cuda.empty_cache()
+
+model = AutoPeftModelForCausalLM.from_pretrained(output_dir, device_map="auto", torch_dtype=torch.bfloat16)
+model = model.merge_and_unload()
+
+output_merged_dir = os.path.join(training_args.output_dir, "final_merged_checkpoint")
+model.save_pretrained(output_merged_dir, safe_serialization=True)

examples/research_projects/tools/calculator.py

@@ -0,0 +1,119 @@
+# coding=utf-8
+# Copyright 2023 The HuggingFace Inc. 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 re
+
+import numpy as np
+import torch
+from transformers import AutoTokenizer, load_tool
+
+from trl import AutoModelForCausalLMWithValueHead, PPOConfig, PPOTrainer, TextEnvironment
+
+
+def generate_data(n):
+    """Generate random arithmetic tasks and answers."""
+    tasks, answers = [], []
+    for _ in range(n):
+        a = np.random.randint(0, 50)
+        b = np.random.randint(0, 50)
+        op = np.random.choice(["-", "+", "*"])
+        tasks.append(f"\n\nWhat is {a} {op} {b}?")
+        if op == "-":
+            answers.append(a - b)
+        elif op == "+":
+            answers.append(a + b)
+        else:
+            answers.append(a * b)
+    return tasks, answers
+
+
+def exact_match_reward(responses, answers=None):
+    """Reward if generated response contains correct answer."""
+    rewards = []
+    pattern = r"Result\s*=\s*(-?\d+(?:\.\d+)?)\s*<submit>"  # generated by chatGPT
+    for response, answer in zip(responses, answers):
+        reward = 0.0
+        predicted_number = None
+        match_pattern = re.findall(pattern, response)
+        if match_pattern:
+            predicted_number = float(match_pattern[0])
+        if predicted_number is not None:
+            if np.abs(predicted_number - answer) < 0.01:
+                reward += 1.0
+        rewards.append(torch.tensor(reward))
+    return rewards
+
+
+# set up models
+model_id = "gpt2"
+model = AutoModelForCausalLMWithValueHead.from_pretrained(model_id)
+model_ref = AutoModelForCausalLMWithValueHead.from_pretrained(model_id)
+tokenizer = AutoTokenizer.from_pretrained(model_id)
+tokenizer.pad_token = tokenizer.eos_token
+
+# system prompt
+prompt = """\
+What is 13-3?
+
+<request><SimpleCalculatorTool>13-3<call>10.0<response>
+
+Result=10<submit>
+
+What is 4*3?
+
+<request><SimpleCalculatorTool>4*3<call>12.0<response>
+
+Result=12<submit>"""
+
+generation_kwargs = {
+    "min_length": -1,
+    "top_k": 0.0,
+    "top_p": 1.0,
+    "do_sample": True,
+    "pad_token_id": tokenizer.eos_token_id,
+    "eos_token_id": -1,
+    "max_new_tokens": 32,
+}
+
+# trainer
+ppo_config = PPOConfig(
+    batch_size=256,
+    learning_rate=1.41e-5,
+    mini_batch_size=64,
+    log_with="wandb",
+)
+ppo_trainer = PPOTrainer(ppo_config, model, model_ref, tokenizer)
+
+# text env
+text_env = TextEnvironment(
+    model,
+    tokenizer,
+    {"SimpleCalculatorTool": load_tool("ybelkada/simple-calculator")},
+    exact_match_reward,
+    prompt,
+    generation_kwargs=generation_kwargs,
+)
+
+# main training loop
+for step in range(100):
+    tasks, answers = generate_data(ppo_config.batch_size)
+    queries, responses, masks, rewards, histories = text_env.run(tasks, answers=answers)
+    train_stats = ppo_trainer.step(queries, responses, rewards, masks)
+
+    response_texts = [tokenizer.decode(response) for response in responses]
+    query_texts = [tokenizer.decode(query) for query in queries]
+    texts = {"query": [qt.split("<submit>")[-1].strip() for qt in query_texts], "response": response_texts}
+    ppo_trainer.log_stats(train_stats, texts, rewards, columns_to_log=["query", "response", "answer"])
+ppo_trainer.save_pretrained(model_id + "-calculator")

examples/research_projects/tools/python_interpreter.py

@@ -0,0 +1,194 @@
+# coding=utf-8
+# Copyright 2023 The HuggingFace Inc. 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 os
+import re
+from dataclasses import dataclass, field
+from typing import Optional
+
+import numpy as np
+import torch
+from datasets import load_dataset
+from peft import LoraConfig
+from transformers import AutoTokenizer, HfArgumentParser, load_tool
+
+from trl import AutoModelForCausalLMWithValueHead, PPOConfig, PPOTrainer, TextEnvironment
+
+
+os.environ["HF_ALLOW_CODE_EVAL"] = "1"
+os.environ["TOKENIZERS_PARALLELISM"] = "false"
+
+
+@dataclass
+class ScriptArguments:
+    model_name: Optional[str] = field(default="bigcode/starcoderbase", metadata={"help": "the model name"})
+    learning_rate: Optional[float] = field(default=1e-5, metadata={"help": "the learning rate"})
+    mini_batch_size: Optional[int] = field(default=1, metadata={"help": "the PPO minibatch size"})
+    batch_size: Optional[int] = field(default=32, metadata={"help": "the batch size"})
+    gradient_accumulation_steps: Optional[int] = field(
+        default=16, metadata={"help": "the number of gradient accumulation steps"}
+    )
+    max_new_tokens: Optional[int] = field(default=256, metadata={"help": "max number of generated tokens per turn"})
+    ppo_epochs: Optional[int] = field(default=1, metadata={"help": "max number of ppo epochs"})
+    n_epochs: Optional[int] = field(default=32, metadata={"help": "max number of ppo epochs"})
+
+
+parser = HfArgumentParser(ScriptArguments)
+args = parser.parse_args_into_dataclasses()[0]
+
+
+def exact_match_reward(responses, answers=None):
+    """Reward if generated response contains correct answer."""
+    rewards = []
+    pattern = r"Result\s*=\s*(-?\d+(?:\.\d+)?)\s*<submit>"  # generated by chatGPT
+    for response, answer in zip(responses, answers):
+        reward = 0.0
+        try:
+            predicted_number = None
+            match_pattern = re.findall(pattern, response)
+            if match_pattern:
+                predicted_number = float(match_pattern[0])
+            if predicted_number is not None:
+                if np.abs((predicted_number - float(answer))) < 0.1:
+                    reward += 1.0
+        except:  # noqa
+            pass
+        rewards.append(torch.tensor(reward))
+    return rewards
+
+
+def evaluate(test_dataloader, text_env, ppo_trainer):
+    test_rewards = []
+    for test_batch in test_dataloader:
+        _, _, _, rewards, _ = text_env.run(test_batch["query"], answers=test_batch["answer"])
+        test_rewards.extend(rewards)
+    test_rewards = ppo_trainer.accelerator.gather_for_metrics(
+        torch.stack(test_rewards).to(ppo_trainer.accelerator.device)
+    )
+    return test_rewards.mean()
+
+
+lora_config = LoraConfig(
+    r=16,
+    lora_alpha=32,
+    lora_dropout=0.05,
+    bias="none",
+    task_type="CAUSAL_LM",
+    target_modules=["c_proj", "c_attn", "q_attn"],
+)
+
+# set up models
+model = AutoModelForCausalLMWithValueHead.from_pretrained(
+    args.model_name,
+    use_auth_token=True,
+    load_in_4bit=True,
+    peft_config=lora_config,
+)
+tokenizer = AutoTokenizer.from_pretrained(args.model_name, use_auth_token=True)
+tokenizer.pad_token = tokenizer.eos_token
+
+ds = load_dataset("gsm8k", "main", split="train")
+ds = ds.rename_columns({"question": "query"})
+ds = ds.map(lambda x: {"answer": x["answer"].split("#### ")[1]})
+ds = ds.select(range(1, len(ds)))  # skip the first sample which is used in prompt
+
+ds_test = load_dataset("gsm8k", "main", split="test")
+ds_test = ds_test.rename_columns({"question": "query"})
+ds_test = ds_test.map(lambda x: {"answer": x["answer"].split("#### ")[1]})
+
+test_dataloader = torch.utils.data.DataLoader(ds_test, batch_size=args.batch_size)
+
+# prompt
+prompt = """\
+Example of using a Python API to solve math questions.
+
+Q: Olivia has $23. She bought five bagels for $3 each. How much money does she have left?
+
+<request><PythonInterpreter>
+def solution():
+    money_initial = 23
+    bagels = 5
+    bagel_cost = 3
+    money_spent = bagels * bagel_cost
+    money_left = money_initial - money_spent
+    result = money_left
+    return result
+print(solution())
+<call>72<response>
+
+Result = 72 <submit>
+
+Q: """
+
+generation_kwargs = {
+    "min_length": -1,
+    "top_k": 0.0,
+    "top_p": 1.0,
+    "do_sample": True,
+    "pad_token_id": tokenizer.eos_token_id,
+    "eos_token_id": -1,
+    "max_new_tokens": args.max_new_tokens,
+}
+
+# trainer
+ppo_config = PPOConfig(
+    batch_size=args.batch_size,
+    learning_rate=args.learning_rate,
+    mini_batch_size=args.mini_batch_size,
+    ppo_epochs=args.ppo_epochs,
+    gradient_accumulation_steps=args.gradient_accumulation_steps,
+    log_with="wandb",
+    tracker_project_name="trl-gsm8k",
+    remove_unused_columns=False,
+    optimize_cuda_cache=True,
+)
+
+ppo_trainer = PPOTrainer(config=ppo_config, model=model, tokenizer=tokenizer, dataset=ds)
+test_dataloader = ppo_trainer.accelerator.prepare(test_dataloader)
+
+# text env
+text_env = TextEnvironment(
+    model,
+    tokenizer,
+    [load_tool("lvwerra/python-interpreter")],
+    exact_match_reward,
+    prompt,
+    max_turns=2,
+    generation_kwargs=generation_kwargs,
+)
+
+# main training loop
+for epoch in range(args.n_epochs):
+    for step, batch in enumerate(ppo_trainer.dataloader):
+        if (step == 0) and (epoch % 4 == 0):  # evaluate every 4 epochs
+            reward_mean_test = evaluate(test_dataloader, text_env, ppo_trainer)
+        else:
+            reward_mean_test = None
+
+        queries, responses, masks, rewards, histories = text_env.run(batch["query"], answers=batch["answer"])
+        train_stats = ppo_trainer.step(queries, responses, rewards, masks)
+
+        # logging
+        if reward_mean_test is not None:
+            train_stats["env/reward_mean_test"] = reward_mean_test
+        texts = {
+            "query": batch["query"],
+            "response": [tokenizer.decode(response) for response in responses],
+            "answer": batch["answer"],
+        }
+        ppo_trainer.log_stats(train_stats, texts, rewards, columns_to_log=["query", "response", "answer"])
+
+reward_mean_test = evaluate(test_dataloader, text_env, ppo_trainer)
+ppo_trainer.save_pretrained(f"model/{args.model_name}-gsm8k")

examples/research_projects/tools/triviaqa.py

@@ -0,0 +1,191 @@
+# coding=utf-8
+# Copyright 2023 The HuggingFace Inc. 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 os
+from dataclasses import dataclass, field
+from typing import Optional
+
+import torch
+from datasets import load_dataset
+from peft import LoraConfig
+from transformers import AutoTokenizer, HfArgumentParser, load_tool
+
+from trl import AutoModelForCausalLMWithValueHead, PPOConfig, PPOTrainer, TextEnvironment
+
+
+os.environ["HF_ALLOW_CODE_EVAL"] = "1"
+os.environ["TOKENIZERS_PARALLELISM"] = "false"
+
+
+@dataclass
+class ScriptArguments:
+    model_name: Optional[str] = field(default="bigcode/starcoderbase", metadata={"help": "the model name"})
+    log_with: Optional[str] = field(default=None, metadata={"help": "use 'wandb' to log with wandb"})
+    learning_rate: Optional[float] = field(default=1e-5, metadata={"help": "the learning rate"})
+    mini_batch_size: Optional[int] = field(default=1, metadata={"help": "the PPO minibatch size"})
+    batch_size: Optional[int] = field(default=32, metadata={"help": "the batch size"})
+    gradient_accumulation_steps: Optional[int] = field(
+        default=16, metadata={"help": "the number of gradient accumulation steps"}
+    )
+    max_new_tokens: Optional[int] = field(default=256, metadata={"help": "max number of generated tokens per turn"})
+    ppo_epochs: Optional[int] = field(default=1, metadata={"help": "max number of ppo epochs"})
+    iterations: Optional[int] = field(default=1000, metadata={"help": "the number of iterations"})
+    seed: Optional[int] = field(default=0, metadata={"help": "the random seed"})
+
+
+parser = HfArgumentParser(ScriptArguments)
+args = parser.parse_args_into_dataclasses()[0]
+
+lora_config = LoraConfig(
+    r=16,
+    lora_alpha=32,
+    lora_dropout=0.05,
+    bias="none",
+    task_type="CAUSAL_LM",
+    target_modules=["c_proj", "c_attn", "q_attn"],
+)
+
+# set up models
+model = AutoModelForCausalLMWithValueHead.from_pretrained(
+    args.model_name,
+    use_auth_token=True,
+    trust_remote_code=True,
+    load_in_4bit=True,
+    peft_config=lora_config,
+)
+tokenizer = AutoTokenizer.from_pretrained(args.model_name, use_auth_token=True)
+tokenizer.pad_token = tokenizer.eos_token
+
+# system prompt
+prompt = """\
+Answer the following question:
+
+Q: In which branch of the arts is Patricia Neary famous?
+A: Ballets
+A2: <request><Wiki>Patricia Neary<call>Patricia Neary (born October 27, 1942) is an American ballerina, choreographer and ballet director, who has been particularly active in Switzerland. She has also been a highly successful ambassador for the Balanchine Trust, bringing George Balanchine's ballets to 60 cities around the globe.<response>
+Result=Ballets<submit>
+
+Q: Who won Super Bowl XX?
+A: Chicago Bears
+A2: <request><Wiki>Super Bowl XX<call>Super Bowl XX was an American football game between the National Football Conference (NFC) champion Chicago Bears and the American Football Conference (AFC) champion New England Patriots to decide the National Football League (NFL) champion for the 1985 season. The Bears defeated the Patriots by the score of 46–10, capturing their first NFL championship (and Chicago's first overall sports victory) since 1963, three years prior to the birth of the Super Bowl. Super Bowl XX was played on January 26, 1986 at the Louisiana Superdome in New Orleans.<response>
+Result=Chicago Bears<submit>
+
+Q: """
+
+generation_kwargs = {
+    "min_length": -1,
+    "top_k": 0.0,
+    "top_p": 1.0,
+    "do_sample": True,
+    "pad_token_id": tokenizer.eos_token_id,
+    "eos_token_id": -1,
+    "max_new_tokens": args.max_new_tokens,
+}
+
+# trainer
+config = PPOConfig(
+    batch_size=args.batch_size,
+    model_name=args.model_name,
+    learning_rate=args.learning_rate,
+    log_with=args.log_with,
+    mini_batch_size=args.mini_batch_size,
+    ppo_epochs=args.ppo_epochs,
+    gradient_accumulation_steps=args.gradient_accumulation_steps,
+    seed=args.seed,
+    optimize_cuda_cache=True,
+)
+ppo_trainer = PPOTrainer(config=config, model=model, tokenizer=tokenizer)
+dataset = load_dataset("trivia_qa", "rc", split="train")
+local_seed = args.seed + ppo_trainer.accelerator.process_index * 100003  # Prime
+dataset = dataset.shuffle(local_seed)
+
+
+def data_generator():
+    for i in range(len(dataset)):
+        yield dataset[i]["question"], [item for item in dataset[i]["answer"]["normalized_aliases"]]
+
+
+gen = data_generator()
+gen = iter(gen)
+
+
+def generate_data(n):
+    tasks, answers = [], []
+    for i in range(n):
+        q, a = next(gen)
+        tasks.append(q)
+        answers.append(a)
+    return tasks, answers
+
+
+def exact_match_reward(responses, answers=None):
+    """Reward if generated response contains correct answer."""
+    rewards = []
+    for response, answer in zip(responses, answers):
+        reward = 0.0
+        for a in answer:
+            if a.lower() in response.lower():
+                reward += 1.0
+                break
+        rewards.append(torch.tensor(reward))
+    return rewards
+
+
+# text env
+tool = load_tool("vwxyzjn/pyserini-wikipedia-kilt-doc")
+# limit the amount if tokens
+tool_fn = lambda x: tool(x).split("\n")[1][:600]  # noqa
+text_env = TextEnvironment(
+    model,
+    tokenizer,
+    {"Wiki": tool_fn},
+    exact_match_reward,
+    prompt,
+    generation_kwargs=generation_kwargs,
+    max_tool_reponse=400,
+)
+
+
+def print_trainable_parameters(model):
+    trainable_params = 0
+    all_param = 0
+    for _, param in model.named_parameters():
+        all_param += param.numel()
+        if param.requires_grad:
+            trainable_params += param.numel()
+    print(
+        f"trainable params: {trainable_params} || all params: {all_param} || trainable%: {100 * trainable_params / all_param}"
+    )
+
+
+print_trainable_parameters(model)
+# main training loop
+for i in range(args.iterations):
+    tasks, answers = generate_data(config.batch_size)
+    queries, responses, masks, rewards, histories = text_env.run(tasks, answers=answers)
+    train_stats = ppo_trainer.step(queries, responses, rewards, masks)
+    response_texts = [tokenizer.decode(response) for response in responses]
+    query_texts = [tokenizer.decode(query) for query in queries]
+    texts = {
+        "query": [qt.split("<submit>")[-1].strip() for qt in query_texts],
+        "response": response_texts,
+        "answer": [", ".join(item) for item in answers],
+    }
+    all_rewards = ppo_trainer.accelerator.gather(torch.tensor(rewards, device=ppo_trainer.accelerator.device))
+    ppo_trainer.log_stats(
+        train_stats, texts, [item for item in all_rewards], columns_to_log=["query", "response", "answer"]
+    )
+    if i % 100 == 0:
+        ppo_trainer.save_pretrained(f"models/{args.model_name}_{args.seed}_{i}_triviaqa")

examples/research_projects/toxicity/README.md

@@ -0,0 +1,7 @@
+# De-detoxifying language models
+
+To run this code, do the following:
+
+```shell
+ACCELERATE_LOG_LEVEL=info accelerate launch --config_file {CONFIG} examples/research_projects/toxicity/scripts/gpt-j-6b-toxicity.py --log_with wandb
+```

examples/research_projects/toxicity/scripts/evaluate-toxicity.py

@@ -1,23 +1,26 @@
-import numpy as np
-import csv
 import argparse
-from tqdm import tqdm
-import torch
+import csv
 
 import evaluate
+import numpy as np
+import torch
 from datasets import load_dataset
+from tqdm import tqdm
 from transformers import AutoModelForCausalLM, AutoTokenizer
 
-toxicity = evaluate.load("ybelkada/toxicity", 'DaNLP/da-electra-hatespeech-detection', module_type="measurement")
+from trl.import_utils import is_npu_available, is_xpu_available
+
+
+toxicity = evaluate.load("ybelkada/toxicity", "DaNLP/da-electra-hatespeech-detection", module_type="measurement")
 ds = load_dataset("OxAISH-AL-LLM/wiki_toxic", split="test")
 
-parser = argparse.ArgumentParser(description='Evaluate de-toxified models')
-parser.add_argument('--model_type', default="all", type=str, help='Relative path to the source model folder')
-parser.add_argument('--output_file', default="toxicity.csv", type=str, help='Relative path to the source model folder')
-parser.add_argument('--batch_size', default=64, type=int, help='Batch size')
-parser.add_argument('--num_samples', default=400, type=int, help='Number of samples')
-parser.add_argument('--context_length', default=2000, type=int, help='Number of samples')
-parser.add_argument('--max_new_tokens', default=30, type=int, help='Max new tokens for generation')
+parser = argparse.ArgumentParser(description="Evaluate de-toxified models")
+parser.add_argument("--model_type", default="all", type=str, help="Relative path to the source model folder")
+parser.add_argument("--output_file", default="toxicity.csv", type=str, help="Relative path to the source model folder")
+parser.add_argument("--batch_size", default=64, type=int, help="Batch size")
+parser.add_argument("--num_samples", default=400, type=int, help="Number of samples")
+parser.add_argument("--context_length", default=2000, type=int, help="Number of samples")
+parser.add_argument("--max_new_tokens", default=30, type=int, help="Max new tokens for generation")
 args = parser.parse_args()
 
 
@@ -40,33 +43,36 @@ elif args.model_type == "gpt-neo":
 elif args.model_type == "gpt-j":
     MODELS_TO_TEST = [
         "ybelkada/gpt-j-6b-sharded-bf16",
-        "ybelkada/gpt-j-6b-detox", 
+        "ybelkada/gpt-j-6b-detox",
     ]
 else:
-    MODELS_TO_TEST = [
-        args.model_type
-    ]
+    MODELS_TO_TEST = [args.model_type]
 NUM_SAMPLES = args.num_samples
 BATCH_SIZE = args.batch_size
 output_file = args.output_file
 max_new_tokens = args.max_new_tokens
 context_length = args.context_length
-device = torch.cuda.current_device()
+if is_xpu_available():
+    device = torch.xpu.current_device()
+elif is_npu_available():
+    device = torch.npu.current_device()
+else:
+    device = torch.cuda.current_device() if torch.cuda.is_available() else "cpu"
 
 # consider only toxic prompts
-ds = ds.filter(lambda x: x['label'] == 1)
+ds = ds.filter(lambda x: x["label"] == 1)
 
 toxicities = {}
 
 # open a csv file
-file = open(f'{output_file}', 'w', newline='')
+file = open(f"{output_file}", "w", newline="")
 writer = csv.writer(file)
 # add first rows
-writer.writerow(['model_id', 'mean_toxicity', 'std_toxicity'])
+writer.writerow(["model_id", "mean_toxicity", "std_toxicity"])
 
 
 for model_id in tqdm(MODELS_TO_TEST):
-    model = AutoModelForCausalLM.from_pretrained(model_id, device_map={'':device}, torch_dtype=torch.bfloat16)
+    model = AutoModelForCausalLM.from_pretrained(model_id, device_map={"": device}, torch_dtype=torch.bfloat16)
     tokenizer = AutoTokenizer.from_pretrained(model_id)
     tokenizer.pad_token = tokenizer.eos_token
     tokenizer.padding_side = "left"
@@ -75,29 +81,28 @@ for model_id in tqdm(MODELS_TO_TEST):
     for i, example in enumerate(ds):
         # set seed
         torch.manual_seed(42)
-        
-        input_text = example['comment_text']
+
+        input_text = example["comment_text"]
         input_texts.append(input_text[:2000])
 
         if i > NUM_SAMPLES:
             break
-        
-    
-        if (i+1)%BATCH_SIZE == 0:
 
+        if (i + 1) % BATCH_SIZE == 0:
             inputs = tokenizer(input_texts, return_tensors="pt", padding=True).to(device)
             inputs.input_ids = inputs.input_ids[:context_length]
             inputs.attention_mask = inputs.attention_mask[:context_length]
             outputs = model.generate(**inputs, do_sample=True, max_new_tokens=max_new_tokens, use_cache=True)
             generated_texts = tokenizer.batch_decode(outputs, skip_special_tokens=True)
-            generated_texts = [generated_text.replace(input_texts[i], "") for i, generated_text in enumerate(generated_texts)]
+            generated_texts = [
+                generated_text.replace(input_texts[i], "") for i, generated_text in enumerate(generated_texts)
+            ]
             toxicity_score = toxicity.compute(predictions=generated_texts)
             input_texts = []
 
-
             if model_id not in toxicities:
                 toxicities[model_id] = []
-            toxicities[model_id].extend(toxicity_score['toxicity'])
+            toxicities[model_id].extend(toxicity_score["toxicity"])
 
     # last batch
     inputs = tokenizer(input_texts, return_tensors="pt", padding=True).to(device)
@@ -105,7 +110,7 @@ for model_id in tqdm(MODELS_TO_TEST):
     generated_texts = tokenizer.batch_decode(outputs, skip_special_tokens=True)
     generated_texts = [generated_text.replace(input_texts[i], "") for i, generated_text in enumerate(generated_texts)]
     toxicity_score = toxicity.compute(predictions=generated_texts)
-    toxicities[model_id].extend(toxicity_score['toxicity'])
+    toxicities[model_id].extend(toxicity_score["toxicity"])
 
     # compute mean & std using np
     mean = np.mean(toxicities[model_id])
@@ -118,7 +123,12 @@ for model_id in tqdm(MODELS_TO_TEST):
     print(f"Model: {model_id} - Mean: {mean} - Std: {std}")
 
     model = None
-    torch.cuda.empty_cache()
+    if is_xpu_available():
+        torch.xpu.empty_cache()
+    elif is_npu_available():
+        torch.npu.empty_cache()
+    else:
+        torch.cuda.empty_cache()
 
 # close file
-file.close()
+file.close()

examples/research_projects/toxicity/scripts/gpt-j-6b-toxicity.py

@@ -12,21 +12,27 @@
 # 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 dataclasses import dataclass, field
+from typing import Optional
+
 import torch
+from datasets import load_dataset
 from torch.optim import Adam
 from tqdm import tqdm
+from transformers import (
+    AutoModelForCausalLM,
+    AutoTokenizer,
+    HfArgumentParser,
+    RobertaForSequenceClassification,
+    RobertaTokenizer,
+)
+
+from trl import AutoModelForCausalLMWithValueHead, PPOConfig, PPOTrainer, create_reference_model, set_seed
+from trl.core import LengthSampler
+
 
 tqdm.pandas()
 
-from transformers import AutoTokenizer, AutoModelForCausalLM
-from transformers import RobertaTokenizer, RobertaForSequenceClassification
-
-from datasets import load_dataset
-
-from trl import PPOTrainer, PPOConfig, AutoModelForCausalLMWithValueHead, set_seed
-from trl import create_reference_model
-from trl.core import LengthSampler
-
 ########################################################################
 # This is a fully working simple example to use trl with accelerate.
 #
@@ -44,17 +50,45 @@ from trl.core import LengthSampler
 #
 ########################################################################
 
+
 # We first define the configuration of the experiment, defining the model, the dataset,
 # the training parameters, and the PPO parameters.
 # Check the default arguments in the `PPOConfig` class for more details.
 # If you want to log with tensorboard, add the kwarg
-# `accelerator_kwargs={"logging_dir": PATH_TO_LOGS}` to the PPOConfig.
+# `project_kwargs={"logging_dir": PATH_TO_LOGS}` to the PPOConfig.
+@dataclass
+class ScriptArguments:
+    """
+    The name of the Casual LM model we wish to fine with PPO
+    """
+
+    # NOTE: gpt2 models use Conv1D instead of Linear layers which are not yet supported in 8 bit mode
+    # models like gpt-neo* models are more suitable.
+    model_name: Optional[str] = field(default="ybelkada/gpt-j-6b-sharded-bf16", metadata={"help": "the model name"})
+    log_with: Optional[str] = field(default=None, metadata={"help": "use 'wandb' to log with wandb"})
+    learning_rate: Optional[float] = field(default=(1.47e-5) * 2, metadata={"help": "the learning rate"})
+    mini_batch_size: Optional[int] = field(default=4, metadata={"help": "the PPO minibatch size"})
+    batch_size: Optional[int] = field(default=16, metadata={"help": "the batch size"})
+    gradient_accumulation_steps: Optional[int] = field(
+        default=1, metadata={"help": "the number of gradient accumulation steps"}
+    )
+    model_save_path: Optional[str] = field(
+        default="./gpt-j-6B-detoxified-long-context-26-shl-1e4-final",
+        metadata={"help": "the path to save the model"},
+    )
+
+
+parser = HfArgumentParser(ScriptArguments)
+script_args = parser.parse_args_into_dataclasses()[0]
+
 config = PPOConfig(
-    model_name="ybelkada/gpt-j-6b-sharded-bf16",
-    learning_rate=(1.47e-5) * 2,
-    log_with="wandb",
-    batch_size=32,
-    forward_batch_size=1,
+    model_name=script_args.model_name,
+    learning_rate=script_args.learning_rate,
+    log_with=script_args.log_with,
+    ppo_epochs=100,
+    mini_batch_size=script_args.mini_batch_size,
+    batch_size=script_args.batch_size,
+    gradient_accumulation_steps=script_args.gradient_accumulation_steps,
 )
 
 
@@ -137,7 +171,13 @@ tokenizer.pad_token = tokenizer.eos_token
 
 # We then build the PPOTrainer, passing the model, the reference model, the tokenizer
 ppo_trainer = PPOTrainer(
-    config, model, ref_model=ref_model, tokenizer=tokenizer, dataset=dataset, data_collator=collator
+    config,
+    model,
+    ref_model=ref_model,
+    tokenizer=tokenizer,
+    dataset=dataset,
+    data_collator=collator,
+    optimizer=optimizer,
 )
 
 # We then build the reward pipeline, we will use the toxicity model to compute the reward.
@@ -164,12 +204,12 @@ output_min_length = 20
 output_max_length = 30
 output_length_sampler = LengthSampler(output_min_length, output_max_length)
 
-model_save_path = "/mnt/disks/younes-disk/models/gpt-j-6B-detoxified-long-context-26-shl-1e4-final"
+model_save_path = script_args.model_save_path
 
 for epoch, batch in tqdm(enumerate(ppo_trainer.dataloader)):
     query_tensors = batch["input_ids"]
 
-    #### Get response from gpt2
+    # Get response from the policy model
     response_tensors = []
     for query in query_tensors:
         gen_len = output_length_sampler()
@@ -178,7 +218,7 @@ for epoch, batch in tqdm(enumerate(ppo_trainer.dataloader)):
         response_tensors.append(response.squeeze()[-gen_len:])
     batch["response"] = [tokenizer.decode(r.squeeze()) for r in response_tensors]
 
-    #### Compute sentiment score
+    # Compute sentiment score # noqa
     texts = batch["response"]
     toxicity_inputs = toxicity_tokenizer(texts, padding=True, truncation=True, return_tensors="pt").to(
         ppo_trainer.accelerator.device
@@ -188,7 +228,7 @@ for epoch, batch in tqdm(enumerate(ppo_trainer.dataloader)):
 
     rewards = [torch.tensor(output) for output in toxicity_labels]
 
-    #### Run PPO step
+    # Run PPO step
     stats = ppo_trainer.step(query_tensors, response_tensors, rewards)
     ppo_trainer.log_stats(stats, batch, rewards)
 
@@ -196,4 +236,3 @@ for epoch, batch in tqdm(enumerate(ppo_trainer.dataloader)):
     if epoch % 100 == 0:
         if ppo_trainer.accelerator.is_main_process:
             ppo_trainer.save_pretrained(model_save_path)
-

examples/scripts/ddpo.py

@@ -0,0 +1,209 @@
+# Copyright 2023 metric-space, 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 os
+from dataclasses import dataclass, field
+
+import numpy as np
+import torch
+import torch.nn as nn
+import tyro
+from huggingface_hub import hf_hub_download
+from huggingface_hub.utils import EntryNotFoundError
+from transformers import CLIPModel, CLIPProcessor
+
+from trl import DDPOConfig, DDPOTrainer, DefaultDDPOStableDiffusionPipeline
+from trl.import_utils import is_npu_available, is_xpu_available
+
+
+@dataclass
+class ScriptArguments:
+    hf_user_access_token: str
+    pretrained_model: str = "runwayml/stable-diffusion-v1-5"
+    """the pretrained model to use"""
+    pretrained_revision: str = "main"
+    """the pretrained model revision to use"""
+    hf_hub_model_id: str = "ddpo-finetuned-stable-diffusion"
+    """HuggingFace repo to save model weights to"""
+    hf_hub_aesthetic_model_id: str = "trl-lib/ddpo-aesthetic-predictor"
+    """HuggingFace model ID for aesthetic scorer model weights"""
+    hf_hub_aesthetic_model_filename: str = "aesthetic-model.pth"
+    """HuggingFace model filename for aesthetic scorer model weights"""
+
+    ddpo_config: DDPOConfig = field(
+        default_factory=lambda: DDPOConfig(
+            num_epochs=200,
+            train_gradient_accumulation_steps=1,
+            sample_num_steps=50,
+            sample_batch_size=6,
+            train_batch_size=3,
+            sample_num_batches_per_epoch=4,
+            per_prompt_stat_tracking=True,
+            per_prompt_stat_tracking_buffer_size=32,
+            tracker_project_name="stable_diffusion_training",
+            log_with="wandb",
+            project_kwargs={
+                "logging_dir": "./logs",
+                "automatic_checkpoint_naming": True,
+                "total_limit": 5,
+                "project_dir": "./save",
+            },
+        )
+    )
+
+
+class MLP(nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.layers = nn.Sequential(
+            nn.Linear(768, 1024),
+            nn.Dropout(0.2),
+            nn.Linear(1024, 128),
+            nn.Dropout(0.2),
+            nn.Linear(128, 64),
+            nn.Dropout(0.1),
+            nn.Linear(64, 16),
+            nn.Linear(16, 1),
+        )
+
+    @torch.no_grad()
+    def forward(self, embed):
+        return self.layers(embed)
+
+
+class AestheticScorer(torch.nn.Module):
+    """
+    This model attempts to predict the aesthetic score of an image. The aesthetic score
+    is a numerical approximation of how much a specific image is liked by humans on average.
+    This is from https://github.com/christophschuhmann/improved-aesthetic-predictor
+    """
+
+    def __init__(self, *, dtype, model_id, model_filename):
+        super().__init__()
+        self.clip = CLIPModel.from_pretrained("openai/clip-vit-large-patch14")
+        self.processor = CLIPProcessor.from_pretrained("openai/clip-vit-large-patch14")
+        self.mlp = MLP()
+        try:
+            cached_path = hf_hub_download(model_id, model_filename)
+        except EntryNotFoundError:
+            cached_path = os.path.join(model_id, model_filename)
+        state_dict = torch.load(cached_path, map_location=torch.device("cpu"))
+        self.mlp.load_state_dict(state_dict)
+        self.dtype = dtype
+        self.eval()
+
+    @torch.no_grad()
+    def __call__(self, images):
+        device = next(self.parameters()).device
+        inputs = self.processor(images=images, return_tensors="pt")
+        inputs = {k: v.to(self.dtype).to(device) for k, v in inputs.items()}
+        embed = self.clip.get_image_features(**inputs)
+        # normalize embedding
+        embed = embed / torch.linalg.vector_norm(embed, dim=-1, keepdim=True)
+        return self.mlp(embed).squeeze(1)
+
+
+def aesthetic_scorer(hub_model_id, model_filename):
+    scorer = AestheticScorer(
+        model_id=hub_model_id,
+        model_filename=model_filename,
+        dtype=torch.float32,
+    )
+    if is_npu_available():
+        scorer = scorer.npu()
+    elif is_xpu_available():
+        scorer = scorer.xpu()
+    else:
+        scorer = scorer.cuda()
+
+    def _fn(images, prompts, metadata):
+        images = (images * 255).round().clamp(0, 255).to(torch.uint8)
+        scores = scorer(images)
+        return scores, {}
+
+    return _fn
+
+
+# list of example prompts to feed stable diffusion
+animals = [
+    "cat",
+    "dog",
+    "horse",
+    "monkey",
+    "rabbit",
+    "zebra",
+    "spider",
+    "bird",
+    "sheep",
+    "deer",
+    "cow",
+    "goat",
+    "lion",
+    "frog",
+    "chicken",
+    "duck",
+    "goose",
+    "bee",
+    "pig",
+    "turkey",
+    "fly",
+    "llama",
+    "camel",
+    "bat",
+    "gorilla",
+    "hedgehog",
+    "kangaroo",
+]
+
+
+def prompt_fn():
+    return np.random.choice(animals), {}
+
+
+def image_outputs_logger(image_data, global_step, accelerate_logger):
+    # For the sake of this example, we will only log the last batch of images
+    # and associated data
+    result = {}
+    images, prompts, _, rewards, _ = image_data[-1]
+
+    for i, image in enumerate(images):
+        prompt = prompts[i]
+        reward = rewards[i].item()
+        result[f"{prompt:.25} | {reward:.2f}"] = image.unsqueeze(0)
+
+    accelerate_logger.log_images(
+        result,
+        step=global_step,
+    )
+
+
+if __name__ == "__main__":
+    args = tyro.cli(ScriptArguments)
+
+    pipeline = DefaultDDPOStableDiffusionPipeline(
+        args.pretrained_model, pretrained_model_revision=args.pretrained_revision, use_lora=True
+    )
+
+    trainer = DDPOTrainer(
+        args.ddpo_config,
+        aesthetic_scorer(args.hf_hub_aesthetic_model_id, args.hf_hub_aesthetic_model_filename),
+        prompt_fn,
+        pipeline,
+        image_samples_hook=image_outputs_logger,
+    )
+
+    trainer.train()
+
+    trainer.push_to_hub(args.hf_hub_model_id, token=args.hf_user_access_token)

examples/scripts/dpo.py

@@ -0,0 +1,198 @@
+# coding=utf-8
+# Copyright 2023 The HuggingFace Inc. 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: you need to install transformers from main to run this script. See https://huggingface.co/docs/transformers/installation#install-from-source
+# TODO: bump transformers version in requirements at next release.
+
+# 0. imports
+from dataclasses import dataclass, field
+from typing import Dict, Optional
+
+import torch
+from datasets import Dataset, load_dataset
+from peft import LoraConfig
+from transformers import AutoModelForCausalLM, AutoTokenizer, HfArgumentParser, TrainingArguments
+
+from trl import DPOTrainer
+
+
+# Define and parse arguments.
+@dataclass
+class ScriptArguments:
+    """
+    The arguments for the DPO training script.
+    """
+
+    # data parameters
+    beta: Optional[float] = field(default=0.1, metadata={"help": "the beta parameter for DPO loss"})
+
+    # training parameters
+    model_name_or_path: Optional[str] = field(default="gpt2", metadata={"help": "the model name"})
+    learning_rate: Optional[float] = field(default=1e-3, metadata={"help": "optimizer learning rate"})
+    per_device_train_batch_size: Optional[int] = field(default=4, metadata={"help": "batch size per device"})
+    gradient_accumulation_steps: Optional[int] = field(
+        default=1, metadata={"help": "the number of gradient accumulation steps"}
+    )
+    max_length: Optional[int] = field(default=512, metadata={"help": "max length of each sample"})
+    max_prompt_length: Optional[int] = field(default=128, metadata={"help": "max length of each sample's prompt"})
+    max_target_length: Optional[int] = field(
+        default=128, metadata={"help": "Only used for encoder decoder model. Max target of each sample's prompt"}
+    )
+    label_pad_token_id: Optional[int] = field(default=-100, metadata={"help": "label for non response tokens"})
+    max_steps: Optional[int] = field(default=1000, metadata={"help": "max number of training steps"})
+    # lora parameters
+    use_peft: Optional[bool] = field(default=True, metadata={"help": "Wether to use PEFT or not to train adapters"})
+    peft_lora_r: Optional[int] = field(default=64, metadata={"help": "the r parameter of the LoRA adapters"})
+    peft_lora_alpha: Optional[int] = field(default=16, metadata={"help": "the alpha parameter of the LoRA adapters"})
+    # instrumentation
+    sanity_check: Optional[bool] = field(default=True, metadata={"help": "only train on 1000 samples"})
+    report_to: Optional[str] = field(
+        default=None,
+        metadata={
+            "help": 'The list of integrations to report the results and logs to. Supported platforms are `"azure_ml"`,'
+            '`"comet_ml"`, `"mlflow"`, `"neptune"`, `"tensorboard"`,`"clearml"` and `"wandb"`. '
+            'Use `"all"` to report to all integrations installed, `"none"` for no integrations.'
+        },
+    )
+    # debug argument for distributed training
+    ignore_bias_buffers: Optional[bool] = field(
+        default=False,
+        metadata={
+            "help": "fix for DDP issues with LM bias/mask buffers - invalid scalar type,`inplace operation. See"
+            "https://github.com/huggingface/transformers/issues/22482#issuecomment-1595790992"
+        },
+    )
+    gradient_checkpointing: Optional[bool] = field(
+        default=False, metadata={"help": "Whether to use gradient checkpointing or no"}
+    )
+    gradient_checkpointing_kwargs: Optional[dict] = field(
+        default=None,
+        metadata={
+            "help": "key word arguments to be passed along `torch.utils.checkpoint.checkpoint` method - e.g. `use_reentrant=False`"
+        },
+    )
+
+
+def extract_anthropic_prompt(prompt_and_response):
+    """Extract the anthropic prompt from a prompt and response pair."""
+    search_term = "\n\nAssistant:"
+    search_term_idx = prompt_and_response.rfind(search_term)
+    assert search_term_idx != -1, f"Prompt and response does not contain '{search_term}'"
+    return prompt_and_response[: search_term_idx + len(search_term)]
+
+
+def get_hh(split: str, sanity_check: bool = False, silent: bool = False, cache_dir: str = None) -> Dataset:
+    """Load the Anthropic Helpful-Harmless dataset from Hugging Face and convert it to the necessary format.
+
+    The dataset is converted to a dictionary with the following structure:
+    {
+        'prompt': List[str],
+        'chosen': List[str],
+        'rejected': List[str],
+    }
+
+    Prompts should be structured as follows:
+      \n\nHuman: <prompt>\n\nAssistant:
+    Multiple turns are allowed, but the prompt should always start with \n\nHuman: and end with \n\nAssistant:.
+    """
+    dataset = load_dataset("Anthropic/hh-rlhf", split=split, cache_dir=cache_dir)
+    if sanity_check:
+        dataset = dataset.select(range(min(len(dataset), 1000)))
+
+    def split_prompt_and_responses(sample) -> Dict[str, str]:
+        prompt = extract_anthropic_prompt(sample["chosen"])
+        return {
+            "prompt": prompt,
+            "chosen": sample["chosen"][len(prompt) :],
+            "rejected": sample["rejected"][len(prompt) :],
+        }
+
+    return dataset.map(split_prompt_and_responses)
+
+
+if __name__ == "__main__":
+    parser = HfArgumentParser(ScriptArguments)
+    script_args = parser.parse_args_into_dataclasses()[0]
+
+    # 1. load a pretrained model
+    model = AutoModelForCausalLM.from_pretrained(script_args.model_name_or_path)
+
+    if script_args.ignore_bias_buffers:
+        # torch distributed hack
+        model._ddp_params_and_buffers_to_ignore = [
+            name for name, buffer in model.named_buffers() if buffer.dtype == torch.bool
+        ]
+
+    model_ref = AutoModelForCausalLM.from_pretrained(script_args.model_name_or_path)
+
+    tokenizer = AutoTokenizer.from_pretrained(script_args.model_name_or_path)
+    if tokenizer.pad_token is None:
+        tokenizer.pad_token = tokenizer.eos_token
+
+    # 2. Load the Anthropic Helpful-Harmless dataset
+    train_dataset = get_hh("train", sanity_check=script_args.sanity_check)
+
+    # 3. Load evaluation dataset
+    eval_dataset = get_hh("test", sanity_check=script_args.sanity_check)
+
+    # 4. initialize training arguments:
+    training_args = TrainingArguments(
+        per_device_train_batch_size=script_args.per_device_train_batch_size,
+        max_steps=script_args.max_steps,
+        remove_unused_columns=False,
+        gradient_accumulation_steps=script_args.gradient_accumulation_steps,
+        learning_rate=script_args.learning_rate,
+        evaluation_strategy="steps",
+        logging_first_step=True,
+        logging_steps=10,  # match results in blog post
+        eval_steps=500,
+        output_dir="./test",
+        optim="rmsprop",
+        warmup_steps=150,
+        report_to=script_args.report_to,
+        bf16=True,
+        gradient_checkpointing=script_args.gradient_checkpointing,
+        # TODO: uncomment that on the next transformers release
+        # gradient_checkpointing_kwargs=script_args.gradient_checkpointing_kwargs,
+    )
+
+    if script_args.use_peft:
+        peft_config = LoraConfig(
+            r=script_args.peft_lora_r,
+            lora_alpha=script_args.peft_lora_alpha,
+            bias="none",
+            task_type="CAUSAL_LM",
+        )
+    else:
+        peft_config = None
+
+    # 5. initialize the DPO trainer
+    dpo_trainer = DPOTrainer(
+        model,
+        model_ref,
+        args=training_args,
+        beta=script_args.beta,
+        train_dataset=train_dataset,
+        eval_dataset=eval_dataset,
+        tokenizer=tokenizer,
+        max_length=script_args.max_length,
+        max_target_length=script_args.max_target_length,
+        max_prompt_length=script_args.max_prompt_length,
+        generate_during_eval=True,
+        peft_config=peft_config,
+    )
+
+    # 6. train
+    dpo_trainer.train()

examples/scripts/ppo.py

@@ -0,0 +1,212 @@
+# coding=utf-8
+# Copyright 2023 The HuggingFace Inc. 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 dataclasses import dataclass, field
+from typing import Optional
+
+import torch
+import tyro
+from accelerate import Accelerator
+from datasets import load_dataset
+from peft import LoraConfig
+from tqdm import tqdm
+from transformers import AutoTokenizer, pipeline
+
+from trl import AutoModelForCausalLMWithValueHead, AutoModelForSeq2SeqLMWithValueHead, PPOConfig, PPOTrainer, set_seed
+from trl.core import LengthSampler
+from trl.import_utils import is_npu_available, is_xpu_available
+
+
+tqdm.pandas()
+
+
+@dataclass
+class ScriptArguments:
+    ppo_config: PPOConfig = field(
+        default_factory=lambda: PPOConfig(
+            model_name="lvwerra/gpt2-imdb",
+            query_dataset="imdb",
+            reward_model="sentiment-analysis:lvwerra/distilbert-imdb",
+            learning_rate=1.41e-5,
+            log_with=None,
+            mini_batch_size=128,
+            batch_size=128,
+            gradient_accumulation_steps=1,
+            early_stopping=False,
+            target_kl=6.0,
+            kl_penalty="kl",
+            seed=0,
+            use_score_scaling=False,
+            use_score_norm=False,
+            score_clip=None,
+        )
+    )
+    use_seq2seq: bool = False
+    """whether to use seq2seq models"""
+    use_peft: bool = False
+    """whether to use peft"""
+    peft_config: Optional[LoraConfig] = field(
+        default_factory=lambda: LoraConfig(
+            r=16,
+            lora_alpha=16,
+            bias="none",
+            task_type="CAUSAL_LM",
+        ),
+    )
+    trust_remote_code: bool = field(default=False, metadata={"help": "Enable `trust_remote_code`"})
+
+
+args = tyro.cli(ScriptArguments)
+
+
+# We then define the arguments to pass to the sentiment analysis pipeline.
+# We set `return_all_scores` to True to get the sentiment score for each token.
+sent_kwargs = {"return_all_scores": True, "function_to_apply": "none", "batch_size": 16}
+
+trl_model_class = AutoModelForCausalLMWithValueHead if not args.use_seq2seq else AutoModelForSeq2SeqLMWithValueHead
+
+
+# Below is an example function to build the dataset. In our case, we use the IMDB dataset
+# from the `datasets` library. One should customize this function to train the model on
+# its own dataset.
+def build_dataset(config, query_dataset, input_min_text_length=2, input_max_text_length=8):
+    """
+    Build dataset for training. This builds the dataset from `load_dataset`, one should
+    customize this function to train the model on its own dataset.
+
+    Args:
+        query_dataset (`str`):
+            The name of the dataset to be loaded.
+
+    Returns:
+        dataloader (`torch.utils.data.DataLoader`):
+            The dataloader for the dataset.
+    """
+    tokenizer = AutoTokenizer.from_pretrained(config.model_name)
+    tokenizer.pad_token = tokenizer.eos_token
+    # load imdb with datasets
+    ds = load_dataset(query_dataset, split="train")
+    ds = ds.rename_columns({"text": "review"})
+    ds = ds.filter(lambda x: len(x["review"]) > 200, batched=False)
+
+    input_size = LengthSampler(input_min_text_length, input_max_text_length)
+
+    def tokenize(sample):
+        sample["input_ids"] = tokenizer.encode(sample["review"])[: input_size()]
+        sample["query"] = tokenizer.decode(sample["input_ids"])
+        return sample
+
+    ds = ds.map(tokenize, batched=False)
+    ds.set_format(type="torch")
+    return ds
+
+
+# We retrieve the dataloader by calling the `build_dataset` function.
+dataset = build_dataset(args.ppo_config, args.ppo_config.query_dataset)
+
+
+def collator(data):
+    return dict((key, [d[key] for d in data]) for key in data[0])
+
+
+# set seed before initializing value head for deterministic eval
+set_seed(args.ppo_config.seed)
+
+# Now let's build the model, the reference model, and the tokenizer.
+if not args.use_peft:
+    ref_model = trl_model_class.from_pretrained(args.ppo_config.model_name, trust_remote_code=args.trust_remote_code)
+    device_map = None
+    peft_config = None
+else:
+    peft_config = args.peft_config
+    ref_model = None
+    # Copy the model to each device
+    device_map = {"": Accelerator().local_process_index}
+
+model = trl_model_class.from_pretrained(
+    args.ppo_config.model_name,
+    trust_remote_code=args.trust_remote_code,
+    device_map=device_map,
+    peft_config=peft_config,
+)
+
+
+tokenizer = AutoTokenizer.from_pretrained(args.ppo_config.model_name)
+
+# Some tokenizers like GPT-2's don't have a padding token by default, so we set one here.
+tokenizer.pad_token_id = tokenizer.eos_token_id
+
+# We then build the PPOTrainer, passing the model, the reference model, the tokenizer
+ppo_trainer = PPOTrainer(args.ppo_config, model, ref_model, tokenizer, dataset=dataset, data_collator=collator)
+
+# We then build the sentiment analysis pipeline, passing the model name and the
+# sentiment analysis pipeline arguments. Let's also make sure to set the device
+# to the same device as the PPOTrainer.
+device = ppo_trainer.accelerator.device
+if ppo_trainer.accelerator.num_processes == 1:
+    if is_xpu_available():
+        device = "xpu:0"
+    elif is_npu_available():
+        device = "npu:0"
+    else:
+        device = 0 if torch.cuda.is_available() else "cpu"  # to avoid a `pipeline` bug
+ds_plugin = ppo_trainer.accelerator.state.deepspeed_plugin
+task, model_name = args.ppo_config.reward_model.split(":")
+if ds_plugin is not None and ds_plugin.is_zero3_init_enabled():
+    with ds_plugin.zero3_init_context_manager(enable=False):
+        sentiment_pipe = pipeline(task, model=model_name, device=device)
+else:
+    sentiment_pipe = pipeline(task, model=model_name, device=device)
+
+# Some tokenizers like GPT-2's don't have a padding token by default, so we set one here.
+if sentiment_pipe.tokenizer.pad_token_id is None:
+    sentiment_pipe.tokenizer.pad_token_id = tokenizer.pad_token_id
+
+if sentiment_pipe.model.config.pad_token_id is None:
+    sentiment_pipe.model.config.pad_token_id = tokenizer.pad_token_id
+
+# We then define the arguments to pass to the `generate` function. These arguments
+# are passed to the `generate` function of the PPOTrainer, which is a wrapper around
+# the `generate` function of the trained model.
+generation_kwargs = {
+    "min_length": -1,
+    "top_k": 0.0,
+    "top_p": 1.0,
+    "do_sample": True,
+    "pad_token_id": tokenizer.eos_token_id,
+    "max_new_tokens": 32,
+}
+
+for epoch, batch in tqdm(enumerate(ppo_trainer.dataloader)):
+    query_tensors = batch["input_ids"]
+
+    # Get response from gpt2
+    response_tensors, ref_response_tensors = ppo_trainer.generate(
+        query_tensors, return_prompt=False, generate_ref_response=True, **generation_kwargs
+    )
+    batch["response"] = tokenizer.batch_decode(response_tensors)
+    batch["ref_response"] = tokenizer.batch_decode(ref_response_tensors)
+
+    # Compute sentiment score
+    texts = [q + r for q, r in zip(batch["query"], batch["response"])]
+    pipe_outputs = sentiment_pipe(texts, **sent_kwargs)
+    rewards = [torch.tensor(output[1]["score"]) for output in pipe_outputs]
+    ref_texts = [q + r for q, r in zip(batch["query"], batch["ref_response"])]
+    ref_pipe_outputs = sentiment_pipe(ref_texts, **sent_kwargs)
+    ref_rewards = [torch.tensor(output[1]["score"]) for output in ref_pipe_outputs]
+    batch["ref_rewards"] = ref_rewards
+
+    # Run PPO step
+    stats = ppo_trainer.step(query_tensors, response_tensors, rewards)
+    ppo_trainer.log_stats(stats, batch, rewards, columns_to_log=["query", "response", "ref_response", "ref_rewards"])

examples/scripts/ppo_multi_adapter.py

@@ -0,0 +1,152 @@
+# coding=utf-8
+# Copyright 2023 The HuggingFace Inc. 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 dataclasses import dataclass, field
+from typing import Optional
+
+import torch
+from datasets import load_dataset
+from peft import LoraConfig
+from tqdm import tqdm
+from transformers import AutoTokenizer, BitsAndBytesConfig, HfArgumentParser
+
+from trl import AutoModelForCausalLMWithValueHead, PPOConfig, PPOTrainer
+from trl.core import LengthSampler
+from trl.import_utils import is_npu_available, is_xpu_available
+
+
+input_min_text_length = 6
+input_max_text_length = 12
+
+
+@dataclass
+class ScriptArguments:
+    """
+    The name of the Casual LM model we wish to fine with PPO
+    """
+
+    model_name: Optional[str] = field(default="huggyllama/llama-7b", metadata={"help": "the model name"})
+    dataset_name: Optional[str] = field(default="Anthropic/hh-rlhf", metadata={"help": "the dataset name"})
+    rm_adapter: Optional[str] = field(
+        default="trl-lib/llama-7b-hh-rm-adapter", metadata={"help": "the rm adapter name"}
+    )
+    log_with: Optional[str] = field(default=None, metadata={"help": "use 'wandb' to log with wandb"})
+    use_safetensors: Optional[bool] = field(default=False, metadata={"help": "Use safetensors"})
+    seed: Optional[int] = field(default=0, metadata={"help": "the random seed"})
+    use_score_scaling: Optional[bool] = field(default=False, metadata={"help": "Use score scaling"})
+    use_score_norm: Optional[bool] = field(
+        default=False, metadata={"help": "Use score normalization. Only applicable if use_score_scaling is True"}
+    )
+    score_clip: Optional[float] = field(default=None, metadata={"help": "Score clipping"})
+
+
+parser = HfArgumentParser(ScriptArguments)
+script_args = parser.parse_args_into_dataclasses()[0]
+
+
+def create_and_prepare_dataset(tokenizer):
+    dataset = load_dataset(script_args.dataset_name, split="train[:1%]")
+
+    input_size = LengthSampler(input_min_text_length, input_max_text_length)
+
+    def tokenize(example):
+        text_size = input_size()
+        example["input_ids"] = tokenizer.encode(example["chosen"])[:text_size]
+        example["query"] = tokenizer.decode(example["input_ids"])
+        return example
+
+    dataset = dataset.map(tokenize, batched=False)
+    dataset.set_format("torch")
+    return dataset
+
+
+lora_config = LoraConfig(
+    r=16,
+    lora_alpha=32,
+    lora_dropout=0.05,
+    bias="none",
+    task_type="CAUSAL_LM",
+)
+nf4_config = BitsAndBytesConfig(
+    load_in_4bit=True, bnb_4bit_quant_type="nf4", bnb_4bit_use_double_quant=True, bnb_4bit_compute_dtype=torch.bfloat16
+)
+model = AutoModelForCausalLMWithValueHead.from_pretrained(
+    script_args.model_name,
+    device_map={"": "xpu:0"} if is_xpu_available() else {"": "npu:0"} if is_npu_available else {"": 0},
+    peft_config=lora_config,
+    quantization_config=nf4_config,
+    reward_adapter=script_args.rm_adapter,
+    use_safetensors=script_args.use_safetensors,
+)
+tokenizer = AutoTokenizer.from_pretrained(script_args.model_name)
+
+tokenizer.pad_token = tokenizer.eos_token
+
+dataset = create_and_prepare_dataset(tokenizer)
+
+
+def collator(data):
+    return dict((key, [d[key] for d in data]) for key in data[0])
+
+
+config = PPOConfig(
+    model_name=script_args.model_name,
+    log_with=script_args.log_with,
+    learning_rate=1e-5,
+    batch_size=8,
+    mini_batch_size=2,
+    gradient_accumulation_steps=2,
+    optimize_cuda_cache=True,
+    seed=script_args.seed,
+    use_score_scaling=script_args.use_score_scaling,
+    use_score_norm=script_args.use_score_norm,
+    score_clip=script_args.score_clip,
+)
+
+ppo_trainer = PPOTrainer(
+    config,
+    model,
+    ref_model=None,
+    tokenizer=tokenizer,
+    dataset=dataset,
+    data_collator=collator,
+)
+
+generation_kwargs = {
+    "top_k": 0.0,
+    "top_p": 0.9,
+    "do_sample": True,
+    "pad_token_id": tokenizer.pad_token_id,
+    "max_new_tokens": 32,
+}
+
+for epoch, batch in tqdm(enumerate(ppo_trainer.dataloader)):
+    question_tensors = batch["input_ids"]
+
+    response_tensors = ppo_trainer.generate(
+        question_tensors,
+        return_prompt=False,
+        **generation_kwargs,
+    )
+    batch["response"] = tokenizer.batch_decode(response_tensors, skip_special_tokens=True)
+
+    # Compute reward score
+    texts = [q + r for q, r in zip(batch["query"], batch["response"])]
+    inputs = tokenizer(texts, padding=True, truncation=True, return_tensors="pt").to(ppo_trainer.accelerator.device)
+    raw_rewards = ppo_trainer.accelerator.unwrap_model(ppo_trainer.model).compute_reward_score(**inputs)
+    rewards = [raw_rewards[i, -1, 1] for i in range(len(raw_rewards))]  # take last token
+
+    # Run PPO step
+    stats = ppo_trainer.step(question_tensors, response_tensors, rewards)
+    ppo_trainer.log_stats(stats, batch, rewards)

examples/scripts/reward_modeling.py

@@ -0,0 +1,173 @@
+# coding=utf-8
+# Copyright 2023 The HuggingFace Inc. 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 dataclasses import dataclass, field
+from typing import Optional
+
+import tyro
+from accelerate import Accelerator
+from datasets import load_dataset
+from peft import LoraConfig
+from tqdm import tqdm
+from transformers import AutoModelForSequenceClassification, AutoTokenizer, BitsAndBytesConfig
+
+from trl import RewardConfig, RewardTrainer, is_xpu_available
+
+
+tqdm.pandas()
+
+
+@dataclass
+class ScriptArguments:
+    model_name: str = "facebook/opt-350m"
+    """the model name"""
+    dataset_name: str = "Anthropic/hh-rlhf"
+    """the dataset name"""
+    dataset_text_field: str = "text"
+    """the text field of the dataset"""
+    eval_split: str = "none"
+    """the dataset split to evaluate on; default to 'none' (no evaluation)"""
+    load_in_8bit: bool = False
+    """load the model in 8 bits precision"""
+    load_in_4bit: bool = False
+    """load the model in 4 bits precision"""
+    trust_remote_code: bool = True
+    """Enable `trust_remote_code`"""
+    reward_config: RewardConfig = field(
+        default_factory=lambda: RewardConfig(
+            output_dir="output",
+            per_device_train_batch_size=64,
+            num_train_epochs=1,
+            gradient_accumulation_steps=16,
+            gradient_checkpointing=True,
+            gradient_checkpointing_kwargs={"use_reentrant": False},
+            learning_rate=1.41e-5,
+            report_to="tensorboard",
+            remove_unused_columns=False,
+            optim="adamw_torch",
+            logging_steps=500,
+            evaluation_strategy="no",
+            max_length=512,
+        )
+    )
+    use_peft: bool = False
+    """whether to use peft"""
+    peft_config: Optional[LoraConfig] = field(
+        default_factory=lambda: LoraConfig(
+            r=16,
+            lora_alpha=16,
+            bias="none",
+            task_type="SEQ_CLS",
+            modules_to_save=["scores"],
+        ),
+    )
+
+
+args = tyro.cli(ScriptArguments)
+args.reward_config.evaluation_strategy = "steps" if args.eval_split != "none" else "no"
+
+
+# Step 1: Load the model
+if args.load_in_8bit and args.load_in_4bit:
+    raise ValueError("You can't load the model in 8 bits and 4 bits at the same time")
+elif args.load_in_8bit or args.load_in_4bit:
+    quantization_config = BitsAndBytesConfig(load_in_8bit=args.load_in_8bit, load_in_4bit=args.load_in_4bit)
+    # Copy the model to each device
+    device_map = (
+        {"": f"xpu:{Accelerator().local_process_index}"}
+        if is_xpu_available()
+        else {"": Accelerator().local_process_index}
+    )
+else:
+    device_map = None
+    quantization_config = None
+
+model = AutoModelForSequenceClassification.from_pretrained(
+    args.model_name,
+    quantization_config=quantization_config,
+    device_map=device_map,
+    trust_remote_code=args.trust_remote_code,
+    num_labels=1,
+)
+
+# Step 2: Load the dataset and pre-process it
+tokenizer = AutoTokenizer.from_pretrained(args.model_name)
+train_dataset = load_dataset(args.dataset_name, split="train")
+
+
+# Tokenize chosen/rejected pairs of inputs
+# Adapt this section to your needs for custom datasets
+def preprocess_function(examples):
+    new_examples = {
+        "input_ids_chosen": [],
+        "attention_mask_chosen": [],
+        "input_ids_rejected": [],
+        "attention_mask_rejected": [],
+    }
+    for chosen, rejected in zip(examples["chosen"], examples["rejected"]):
+        tokenized_chosen = tokenizer(chosen)
+        tokenized_rejected = tokenizer(rejected)
+
+        new_examples["input_ids_chosen"].append(tokenized_chosen["input_ids"])
+        new_examples["attention_mask_chosen"].append(tokenized_chosen["attention_mask"])
+        new_examples["input_ids_rejected"].append(tokenized_rejected["input_ids"])
+        new_examples["attention_mask_rejected"].append(tokenized_rejected["attention_mask"])
+
+    return new_examples
+
+
+# Preprocess the dataset and filter out examples that are longer than args.max_length
+train_dataset = train_dataset.map(
+    preprocess_function,
+    batched=True,
+    num_proc=4,
+)
+train_dataset = train_dataset.filter(
+    lambda x: len(x["input_ids_chosen"]) <= args.reward_config.max_length
+    and len(x["input_ids_rejected"]) <= args.reward_config.max_length
+)
+
+if args.eval_split == "none":
+    eval_dataset = None
+else:
+    eval_dataset = load_dataset(args.dataset_name, split=args.eval_split)
+
+    eval_dataset = eval_dataset.map(
+        preprocess_function,
+        batched=True,
+        num_proc=4,
+    )
+    eval_dataset = eval_dataset.filter(
+        lambda x: len(x["input_ids_chosen"]) <= args.reward_config.max_length
+        and len(x["input_ids_rejected"]) <= args.reward_config.max_length
+    )
+
+
+# Step 4: Define the LoraConfig
+if args.use_peft:
+    peft_config = args.peft_config
+else:
+    peft_config = None
+
+# Step 5: Define the Trainer
+trainer = RewardTrainer(
+    model=model,
+    tokenizer=tokenizer,
+    args=args.reward_config,
+    train_dataset=train_dataset,
+    eval_dataset=eval_dataset,
+    peft_config=peft_config,
+)
+
+trainer.train()

examples/scripts/sft.py

@@ -0,0 +1,158 @@
+# coding=utf-8
+# Copyright 2023 The HuggingFace Inc. 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 dataclasses import dataclass, field
+from typing import List, Optional
+
+import torch
+from accelerate import Accelerator
+from datasets import load_dataset
+from peft import LoraConfig
+from tqdm import tqdm
+from transformers import AutoModelForCausalLM, BitsAndBytesConfig, HfArgumentParser, TrainingArguments
+
+from trl import SFTTrainer, is_xpu_available
+
+
+tqdm.pandas()
+
+
+# Define and parse arguments.
+@dataclass
+class ScriptArguments:
+    """
+    The name of the Casual LM model we wish to fine with SFTTrainer
+    """
+
+    model_name: Optional[str] = field(default="facebook/opt-350m", metadata={"help": "the model name"})
+    dataset_name: Optional[str] = field(
+        default="timdettmers/openassistant-guanaco", metadata={"help": "the dataset name"}
+    )
+    dataset_text_field: Optional[str] = field(default="text", metadata={"help": "the text field of the dataset"})
+    report_to: Optional[str] = field(default="none", metadata={"help": "use 'wandb' to log with wandb"})
+    learning_rate: Optional[float] = field(default=1.41e-5, metadata={"help": "the learning rate"})
+    batch_size: Optional[int] = field(default=64, metadata={"help": "the batch size"})
+    seq_length: Optional[int] = field(default=512, metadata={"help": "Input sequence length"})
+    gradient_accumulation_steps: Optional[int] = field(
+        default=16, metadata={"help": "the number of gradient accumulation steps"}
+    )
+    load_in_8bit: Optional[bool] = field(default=False, metadata={"help": "load the model in 8 bits precision"})
+    load_in_4bit: Optional[bool] = field(default=False, metadata={"help": "load the model in 4 bits precision"})
+    use_peft: Optional[bool] = field(default=False, metadata={"help": "Wether to use PEFT or not to train adapters"})
+    trust_remote_code: Optional[bool] = field(default=False, metadata={"help": "Enable `trust_remote_code`"})
+    output_dir: Optional[str] = field(default="output", metadata={"help": "the output directory"})
+    peft_lora_r: Optional[int] = field(default=64, metadata={"help": "the r parameter of the LoRA adapters"})
+    peft_lora_alpha: Optional[int] = field(default=16, metadata={"help": "the alpha parameter of the LoRA adapters"})
+    logging_steps: Optional[int] = field(default=1, metadata={"help": "the number of logging steps"})
+    use_auth_token: Optional[bool] = field(default=True, metadata={"help": "Use HF auth token to access the model"})
+    num_train_epochs: Optional[int] = field(default=3, metadata={"help": "the number of training epochs"})
+    max_steps: Optional[int] = field(default=-1, metadata={"help": "the number of training steps"})
+    save_steps: Optional[int] = field(
+        default=100, metadata={"help": "Number of updates steps before two checkpoint saves"}
+    )
+    save_total_limit: Optional[int] = field(default=10, metadata={"help": "Limits total number of checkpoints."})
+    push_to_hub: Optional[bool] = field(default=False, metadata={"help": "Push the model to HF Hub"})
+    gradient_checkpointing: Optional[bool] = field(
+        default=False, metadata={"help": "Whether to use gradient checkpointing or no"}
+    )
+    gradient_checkpointing_kwargs: Optional[dict] = field(
+        default=None,
+        metadata={
+            "help": "key word arguments to be passed along `torch.utils.checkpoint.checkpoint` method - e.g. `use_reentrant=False`"
+        },
+    )
+    hub_model_id: Optional[str] = field(default=None, metadata={"help": "The name of the model on HF Hub"})
+    mixed_precision: Optional[str] = field(default="bf16", metadata={"help": "Mixed precision training"})
+    target_modules: Optional[List[str]] = field(default=None, metadata={"help": "Target modules for LoRA adapters"})
+
+
+parser = HfArgumentParser(ScriptArguments)
+script_args = parser.parse_args_into_dataclasses()[0]
+
+# Step 1: Load the model
+if script_args.load_in_8bit and script_args.load_in_4bit:
+    raise ValueError("You can't load the model in 8 bits and 4 bits at the same time")
+elif script_args.load_in_8bit or script_args.load_in_4bit:
+    quantization_config = BitsAndBytesConfig(
+        load_in_8bit=script_args.load_in_8bit, load_in_4bit=script_args.load_in_4bit
+    )
+    # Copy the model to each device
+    device_map = (
+        {"": f"xpu:{Accelerator().local_process_index}"}
+        if is_xpu_available()
+        else {"": Accelerator().local_process_index}
+    )
+    torch_dtype = torch.bfloat16
+else:
+    device_map = None
+    quantization_config = None
+    torch_dtype = None
+
+model = AutoModelForCausalLM.from_pretrained(
+    script_args.model_name,
+    quantization_config=quantization_config,
+    device_map=device_map,
+    trust_remote_code=script_args.trust_remote_code,
+    torch_dtype=torch_dtype,
+    use_auth_token=script_args.use_auth_token,
+)
+
+# Step 2: Load the dataset
+dataset = load_dataset(script_args.dataset_name, split="train")
+
+# Step 3: Define the training arguments
+training_args = TrainingArguments(
+    output_dir=script_args.output_dir,
+    per_device_train_batch_size=script_args.batch_size,
+    gradient_accumulation_steps=script_args.gradient_accumulation_steps,
+    learning_rate=script_args.learning_rate,
+    logging_steps=script_args.logging_steps,
+    num_train_epochs=script_args.num_train_epochs,
+    max_steps=script_args.max_steps,
+    report_to=script_args.report_to,
+    save_steps=script_args.save_steps,
+    save_total_limit=script_args.save_total_limit,
+    push_to_hub=script_args.push_to_hub,
+    hub_model_id=script_args.hub_model_id,
+    gradient_checkpointing=script_args.gradient_checkpointing,
+    # TODO: uncomment that on the next release
+    # gradient_checkpointing_kwargs=script_args.gradient_checkpointing_kwargs,
+)
+
+# Step 4: Define the LoraConfig
+if script_args.use_peft:
+    peft_config = LoraConfig(
+        r=script_args.peft_lora_r,
+        lora_alpha=script_args.peft_lora_alpha,
+        bias="none",
+        task_type="CAUSAL_LM",
+        target_modules=script_args.target_modules,
+    )
+else:
+    peft_config = None
+
+# Step 5: Define the Trainer
+trainer = SFTTrainer(
+    model=model,
+    args=training_args,
+    max_seq_length=script_args.seq_length,
+    train_dataset=dataset,
+    dataset_text_field=script_args.dataset_text_field,
+    peft_config=peft_config,
+)
+
+trainer.train()
+
+# Step 6: Save the model
+trainer.save_model(script_args.output_dir)

examples/sentiment/scripts/gpt2-sentiment.py

@@ -1,156 +0,0 @@
-# coding=utf-8
-# Copyright 2022 The HuggingFace Inc. 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 torch
-from tqdm import tqdm
-
-tqdm.pandas()
-
-from transformers import pipeline, AutoTokenizer
-from datasets import load_dataset
-
-from trl import PPOTrainer, PPOConfig, AutoModelForCausalLMWithValueHead, set_seed
-from trl.core import LengthSampler
-
-########################################################################
-# This is a fully working simple example to use trl with accelerate.
-#
-# This example fine-tunes a GPT2 model on the IMDB dataset using PPO
-# (proximal policy optimization).
-# in any of the following settings (with the same script):
-#   - single CPU or single GPU
-#   - multi GPUS (using PyTorch distributed mode)
-#   - multi GPUS (using DeepSpeed ZeRO-Offload stages 1 & 2)
-#   - fp16 (mixed-precision) or fp32 (normal precision)
-#
-# To run it in each of these various modes, first initialize the accelerate
-# configuration with `accelerate config`
-#
-########################################################################
-
-# We first define the configuration of the experiment, defining the model, the dataset,
-# the training parameters, and the PPO parameters.
-# Check the default arguments in the `PPOConfig` class for more details.
-# If you want to log with tensorboard, add the kwarg
-# `accelerator_kwargs={"logging_dir": PATH_TO_LOGS}` to the PPOConfig.
-config = PPOConfig(
-    model_name="lvwerra/gpt2-imdb",
-    learning_rate=1.41e-5,
-)
-
-# We then define the arguments to pass to the sentiment analysis pipeline.
-# We set `return_all_scores` to True to get the sentiment score for each token.
-sent_kwargs = {"return_all_scores": True, "function_to_apply": "none", "batch_size": 16}
-
-
-# Below is an example function to build the dataset. In our case, we use the IMDB dataset
-# from the `datasets` library. One should customize this function to train the model on
-# its own dataset.
-def build_dataset(config, dataset_name="imdb", input_min_text_length=2, input_max_text_length=8):
-    """
-    Build dataset for training. This builds the dataset from `load_dataset`, one should
-    customize this function to train the model on its own dataset.
-
-    Args:
-        dataset_name (`str`):
-            The name of the dataset to be loaded.
-
-    Returns:
-        dataloader (`torch.utils.data.DataLoader`):
-            The dataloader for the dataset.
-    """
-    tokenizer = AutoTokenizer.from_pretrained(config.model_name)
-    tokenizer.pad_token = tokenizer.eos_token
-    # load imdb with datasets
-    ds = load_dataset(dataset_name, split="train")
-    ds = ds.rename_columns({"text": "review"})
-    ds = ds.filter(lambda x: len(x["review"]) > 200, batched=False)
-
-    input_size = LengthSampler(input_min_text_length, input_max_text_length)
-
-    def tokenize(sample):
-        sample["input_ids"] = tokenizer.encode(sample["review"])[: input_size()]
-        sample["query"] = tokenizer.decode(sample["input_ids"])
-        return sample
-
-    ds = ds.map(tokenize, batched=False)
-    ds.set_format(type="torch")
-    return ds
-
-
-# We retrieve the dataloader by calling the `build_dataset` function.
-dataset = build_dataset(config)
-
-
-def collator(data):
-    return dict((key, [d[key] for d in data]) for key in data[0])
-
-
-# set seed before initializing value head for deterministic eval
-set_seed(config.seed)
-
-# Now let's build the model, the reference model, and the tokenizer.
-model = AutoModelForCausalLMWithValueHead.from_pretrained(config.model_name)
-ref_model = AutoModelForCausalLMWithValueHead.from_pretrained(config.model_name)
-tokenizer = AutoTokenizer.from_pretrained(config.model_name)
-
-# GPT-2 tokenizer has a pad token, but it is not eos_token by default. We need to set it to eos_token.
-# only for this model.
-tokenizer.pad_token = tokenizer.eos_token
-
-# We then build the PPOTrainer, passing the model, the reference model, the tokenizer
-ppo_trainer = PPOTrainer(config, model, ref_model, tokenizer, dataset=dataset, data_collator=collator)
-
-# We then build the sentiment analysis pipeline, passing the model name and the
-# sentiment analysis pipeline arguments. Let's also make sure to set the device
-# to the same device as the PPOTrainer.
-device = ppo_trainer.accelerator.device
-if ppo_trainer.accelerator.num_processes == 1:
-    device = 0 if torch.cuda.is_available() else "cpu"  # to avoid a `pipeline` bug
-sentiment_pipe = pipeline("sentiment-analysis", model="lvwerra/distilbert-imdb", device=device)
-
-# We then define the arguments to pass to the `generate` function. These arguments
-# are passed to the `generate` function of the PPOTrainer, which is a wrapper around
-# the `generate` function of the trained model.
-generation_kwargs = {
-    "min_length": -1,
-    "top_k": 0.0,
-    "top_p": 1.0,
-    "do_sample": True,
-    "pad_token_id": tokenizer.eos_token_id,
-}
-output_min_length = 4
-output_max_length = 16
-output_length_sampler = LengthSampler(output_min_length, output_max_length)
-
-for epoch, batch in tqdm(enumerate(ppo_trainer.dataloader)):
-    query_tensors = batch["input_ids"]
-
-    #### Get response from gpt2
-    response_tensors = []
-    for query in query_tensors:
-        gen_len = output_length_sampler()
-        generation_kwargs["max_new_tokens"] = gen_len
-        response = ppo_trainer.generate(query, **generation_kwargs)
-        response_tensors.append(response.squeeze()[-gen_len:])
-    batch["response"] = [tokenizer.decode(r.squeeze()) for r in response_tensors]
-
-    #### Compute sentiment score
-    texts = [q + r for q, r in zip(batch["query"], batch["response"])]
-    pipe_outputs = sentiment_pipe(texts, **sent_kwargs)
-    rewards = [torch.tensor(output[1]["score"]) for output in pipe_outputs]
-
-    #### Run PPO step
-    stats = ppo_trainer.step(query_tensors, response_tensors, rewards)
-    ppo_trainer.log_stats(stats, batch, rewards)

examples/sentiment/scripts/t5-sentiment.py

@@ -1,129 +0,0 @@
-# coding=utf-8
-# Copyright 2022 The HuggingFace Inc. 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 torch
-from tqdm import tqdm
-
-tqdm.pandas()
-from transformers import pipeline, AutoTokenizer
-from datasets import load_dataset
-
-from trl import PPOTrainer, PPOConfig, AutoModelForSeq2SeqLMWithValueHead, set_seed
-from trl.core import LengthSampler
-
-########################################################################
-# This is a fully working simple example to use trl with accelerate.
-#
-# This example fine-tunes a T5 model on the IMDB dataset using PPO
-# (proximal policy optimization).
-# in any of the following settings (with the same script):
-#   - single CPU or single GPU
-#   - multi GPUS (using PyTorch distributed mode)
-#   - multi GPUS (using DeepSpeed ZeRO-Offload stages 1 & 2)
-#   - fp16 (mixed-precision) or fp32 (normal precision)
-#
-# To run it in each of these various modes, first initialize the accelerate
-# configuration with `accelerate config` then run the script with
-# `accelerate launch ppo-sentiment-t5-small.py`
-#
-########################################################################
-
-# We first define the configuration of the experiment, defining the model, the dataset,
-# the training parameters, and the PPO parameters.
-# Check the default arguments in the `PPOConfig` class for more details.
-config = PPOConfig(model_name="lvwerra/t5-imdb", learning_rate=5e-5, batch_size=256)
-# We then define the arguments to pass to the sentiment analysis pipeline.
-# We set `return_all_scores` to True to get the sentiment score for each token.
-sent_kwargs = {"return_all_scores": True, "function_to_apply": "none", "batch_size": 16}
-
-
-# Below is an example function to build the dataset. In our case, we use the IMDB dataset
-# from the `datasets` library. One should customize this function to train the model on
-# its own dataset.
-def build_imdb_dataset(tokenizer, input_min_text_length=2, input_max_text_length=8):
-    # load imdb with datasets
-    ds = load_dataset("imdb", split="train")
-    ds = ds.rename_columns({"text": "review"})
-    ds = ds.filter(lambda x: len(x["review"]) > 200, batched=False)
-
-    input_size = LengthSampler(input_min_text_length, input_max_text_length)
-
-    def tokenize(sample):
-        sample["input_ids"] = tokenizer.encode(sample["review"])[: input_size()] + [tokenizer.eos_token_id]
-        sample["query"] = tokenizer.decode(sample["input_ids"])
-        return sample
-
-    ds = ds.map(tokenize, batched=False)
-    ds.set_format(type="torch")
-    return ds
-
-
-def collater(data):
-    return dict((key, [d[key] for d in data]) for key in data[0])
-
-
-# set seed before initializing value head for deterministic eval
-set_seed(config.seed)
-
-# Now let's build the model, the reference model, and the tokenizer.
-model = AutoModelForSeq2SeqLMWithValueHead.from_pretrained(config.model_name)
-ref_model = AutoModelForSeq2SeqLMWithValueHead.from_pretrained(config.model_name)
-tokenizer = AutoTokenizer.from_pretrained(config.model_name)
-
-# We retrieve the dataloader by calling the `build_dataset` function.
-dataset = build_imdb_dataset(tokenizer)
-
-query = tokenizer("I really liked this movie because", return_tensors="pt")["input_ids"]
-
-generation_kwargs = {"top_k": 0.0, "top_p": 1.0, "do_sample": True, "eos_token_id": -1}
-
-
-# We then build the PPOTrainer, passing the model, the reference model, the tokenizer
-ppo_trainer = PPOTrainer(config, model, ref_model, tokenizer, dataset=dataset, data_collator=collater)
-
-# We then build the sentiment analysis pipeline, passing the model name and the
-# sentiment analysis pipeline arguments. Let's also make sure to set the device
-# to the same device as the PPOTrainer.
-device = ppo_trainer.accelerator.device
-if ppo_trainer.accelerator.num_processes == 1:
-    device = 0 if torch.cuda.is_available() else "cpu"  # to avoid a `pipeline` bug
-sentiment_pipe = pipeline("sentiment-analysis", "lvwerra/distilbert-imdb", device=device)
-
-# We then define the arguments to pass to the `generate` function. These arguments
-# are passed to the `generate` function of the PPOTrainer, which is a wrapper around
-# the `generate` function of the trained model.
-output_min_length = 16
-output_max_length = 32
-output_length_sampler = LengthSampler(output_min_length, output_max_length)
-
-for epoch, batch in tqdm(enumerate(ppo_trainer.dataloader)):
-    query_tensors = batch["input_ids"]
-
-    #### Get response from gpt2
-    response_tensors = []
-    for query in query_tensors:
-        gen_len = output_length_sampler()
-        generation_kwargs["max_new_tokens"] = gen_len
-        response = ppo_trainer.generate(query, **generation_kwargs)
-        response_tensors.append(response.squeeze())
-    batch["response"] = [tokenizer.decode(r[1:].squeeze()) for r in response_tensors]
-
-    #### Compute sentiment score
-    texts = [q + r for q, r in zip(batch["query"], batch["response"])]
-    pipe_outputs = sentiment_pipe(texts, **sent_kwargs)
-    rewards = [torch.tensor(output[1]["score"]).to(device) for output in pipe_outputs]
-
-    #### Run PPO step
-    stats = ppo_trainer.step(query_tensors, response_tensors, rewards)
-    ppo_trainer.log_stats(stats, batch, rewards)

examples/summarization/scripts/ds3_reward_summarization_example_config.json

@@ -1,57 +0,0 @@
-{
-    "fp16": {
-        "enabled": "auto",
-        "loss_scale": 0,
-        "loss_scale_window": 1000,
-        "initial_scale_power": 16,
-        "hysteresis": 2,
-        "min_loss_scale": 1
-    },
-
-    "optimizer": {
-        "type": "AdamW",
-        "params": {
-            "lr": "auto",
-            "betas": "auto",
-            "eps": "auto",
-            "weight_decay": "auto"
-        }
-    },
-
-    "scheduler": {
-        "type": "WarmupLR",
-        "params": {
-            "warmup_min_lr": "auto",
-            "warmup_max_lr": "auto",
-            "warmup_num_steps": "auto"
-        }
-    },
-
-    "zero_optimization": {
-        "stage": 3,
-        "offload_optimizer": {
-            "device": "none",
-            "pin_memory": true
-        },
-        "offload_param": {
-            "device": "none",
-            "pin_memory": true
-        },
-        "overlap_comm": true,
-        "contiguous_gradients": true,
-        "sub_group_size": 1e9,
-        "reduce_bucket_size": "auto",
-        "stage3_prefetch_bucket_size": "auto",
-        "stage3_param_persistence_threshold": "auto",
-        "stage3_max_live_parameters": 1e9,
-        "stage3_max_reuse_distance": 1e9,
-        "stage3_gather_16bit_weights_on_model_save": true
-    },
-
-    "gradient_accumulation_steps": "auto",
-    "gradient_clipping": "auto",
-    "steps_per_print": 2000,
-    "train_batch_size": "auto",
-    "train_micro_batch_size_per_gpu": "auto",
-    "wall_clock_breakdown": false
-}

examples/summarization/scripts/reward_summarization.py

@@ -1,205 +0,0 @@
-from datasets import load_dataset
-from transformers import (
-    AutoTokenizer,
-    AutoModelForSequenceClassification,
-    TrainingArguments,
-    Trainer,
-    PreTrainedTokenizerBase,
-    HfArgumentParser,
-)
-from transformers.utils import PaddingStrategy
-from typing import Optional, Union, List, Dict, Any
-import evaluate
-from dataclasses import dataclass, field
-import torch.nn as nn
-import numpy as np
-
-# Define and parse arguments.
-@dataclass
-class ScriptArguments:
-    """
-    These arguments vary depending on how many GPUs you have, what their capacity and features are, and what size model you want to train.
-    """
-
-    local_rank: Optional[int] = field(default=0, metadata={"help": "Used for multi-gpu"})
-    resume_from_checkpoint: Optional[bool] = field(
-        default=False, metadata={"help": "If you want to resume training where it left off."}
-    )
-    deepspeed: Optional[str] = field(
-        default=None,
-        metadata={
-            "help": "Path to deepspeed config if using deepspeed. You may need this if the model that you want to train doesn't fit on a single GPU."
-        },
-    )
-    per_device_train_batch_size: Optional[int] = field(default=16)
-    per_device_eval_batch_size: Optional[int] = field(default=16)
-    gradient_accumulation_steps: Optional[int] = field(default=4)
-    learning_rate: Optional[int] = field(default=2e-5)
-    weight_decay: Optional[int] = field(default=0.001)
-    model_name: Optional[str] = field(
-        default="gpt2",
-        metadata={
-            "help": "The model that you want to train from the Hugging Face hub. E.g. gpt2, gpt2-xl, bert, etc."
-        },
-    )
-    bf16: Optional[bool] = field(
-        default=False,
-        metadata={
-            "help": "This essentially cuts the training time in half if you want to sacrifice a little precision and have a supported GPU."
-        },
-    )
-    num_train_epochs: Optional[int] = field(
-        default="5",
-        metadata={
-            "help": "The number of training epochs for the reward model. OpenAI used 5."
-        }
-    )
-
-
-parser = HfArgumentParser(ScriptArguments)
-script_args = parser.parse_args_into_dataclasses()[0]
-
-# Load the human comparisons dataset for tuning the reward model.
-ds = load_dataset("openai/summarize_from_feedback", name="comparisons")
-
-# Define the training args. Needs to be done before the model is loaded if you are using deepspeed.
-training_args = TrainingArguments(
-    output_dir=f"{script_args.model_name}_summarization_reward_model",
-    learning_rate=script_args.learning_rate,
-    per_device_train_batch_size=script_args.per_device_train_batch_size,
-    per_device_eval_batch_size=script_args.per_device_eval_batch_size,
-    num_train_epochs=script_args.num_train_epochs,
-    weight_decay=script_args.weight_decay,
-    evaluation_strategy="epoch",
-    save_strategy="epoch",
-    gradient_accumulation_steps=script_args.gradient_accumulation_steps,
-    deepspeed=script_args.deepspeed,
-    local_rank=script_args.local_rank,
-    remove_unused_columns=False,
-    label_names=[],
-)
-
-# Load the value-head model and tokenizer.
-tokenizer = AutoTokenizer.from_pretrained(script_args.model_name)
-model = AutoModelForSequenceClassification.from_pretrained(script_args.model_name, num_labels=1)
-
-# Need to do this for gpt2, because it doesn't have an official pad token.
-tokenizer.pad_token = tokenizer.eos_token
-model.config.pad_token_id = tokenizer.eos_token_id
-
-# Turn the dataset into pairs of post + summaries, where text_j is the preferred post + summary and text_k is the other.
-def turn_into_text_classification_format(examples):
-    new_examples = {"text_j": [], "text_k": []}
-    for info, summaries, choice in zip(examples["info"], examples["summaries"], examples["choice"]):
-        if len(summaries) != 2 or choice not in (0, 1):
-            raise ValueError(f"There should be two summaries with a choice that's either 0 or 1. Received {len(summaries)} summaries and choice={choice}.")
-        original_text_field = "post" if info["post"] is not None else "article"
-        new_examples["text_j"].append(
-            summaries[choice]["text"] + " " + tokenizer.bos_token + " " + info[original_text_field]
-        )
-        new_examples["text_k"].append(
-            summaries[0 if choice == 1 else 1]["text"] + " " + tokenizer.bos_token + " " + info[original_text_field]
-        )
-
-    return new_examples
-
-
-num_proc = (
-    8
-)  # Can adjust to be higher if you have more processors. Should work even if you don't have 8 CPUs, though.
-original_columns = ds["train"].column_names
-ds = ds.map(turn_into_text_classification_format, batched=True, num_proc=num_proc, remove_columns=original_columns)
-
-# Tokenize the dataset.
-def preprocess_function(examples):
-    tokenized_j = tokenizer(examples["text_j"], truncation=True)
-    tokenized_k = tokenizer(examples["text_k"], truncation=True)
-    return {
-        "input_ids_j": tokenized_j["input_ids"],
-        "attention_mask_j": tokenized_j["attention_mask"],
-        "input_ids_k": tokenized_k["input_ids"],
-        "attention_mask_k": tokenized_k["attention_mask"],
-    }
-
-
-tokenized_ds = ds.map(preprocess_function, batched=True, num_proc=num_proc, remove_columns=["text_j", "text_k"])
-
-# We need to define a special data collator that batches the data in our j vs k format.
-@dataclass
-class RewardDataCollatorWithPadding:
-
-    tokenizer: PreTrainedTokenizerBase
-    padding: Union[bool, str, PaddingStrategy] = True
-    max_length: Optional[int] = None
-    pad_to_multiple_of: Optional[int] = None
-    return_tensors: str = "pt"
-
-    def __call__(self, features: List[Dict[str, Any]]) -> Dict[str, Any]:
-        features_j = []
-        features_k = []
-        for feature in features:
-            features_j.append({"input_ids": feature["input_ids_j"], "attention_mask": feature["attention_mask_j"]})
-            features_k.append({"input_ids": feature["input_ids_k"], "attention_mask": feature["attention_mask_k"]})
-        batch_j = self.tokenizer.pad(
-            features_j,
-            padding=self.padding,
-            max_length=self.max_length,
-            pad_to_multiple_of=self.pad_to_multiple_of,
-            return_tensors=self.return_tensors,
-        )
-        batch_k = self.tokenizer.pad(
-            features_k,
-            padding=self.padding,
-            max_length=self.max_length,
-            pad_to_multiple_of=self.pad_to_multiple_of,
-            return_tensors=self.return_tensors,
-        )
-        batch = {
-            "input_ids_j": batch_j["input_ids"],
-            "attention_mask_j": batch_j["attention_mask"],
-            "input_ids_k": batch_k["input_ids"],
-            "attention_mask_k": batch_k["attention_mask"],
-            "return_loss": True,
-        }
-        return batch
-
-
-# Define the metric that we'll use for validation.
-accuracy = evaluate.load("accuracy")
-
-
-def compute_metrics(eval_pred):
-    predictions, _ = eval_pred
-    # Here, predictions is rewards_j and rewards_k.
-    # We want to see how much of the time rewards_j > rewards_k.
-    predictions = np.argmax(predictions, axis=0)
-    labels = np.zeros(predictions.shape)
-    return accuracy.compute(predictions=predictions, references=labels)
-
-
-class RewardTrainer(Trainer):
-    # Define how to compute the reward loss.
-    def compute_loss(self, model, inputs, return_outputs=False):
-        rewards_j = model(input_ids=inputs["input_ids_j"], attention_mask=inputs["attention_mask_j"])[0]
-        rewards_k = model(input_ids=inputs["input_ids_k"], attention_mask=inputs["attention_mask_k"])[0]
-        loss = -nn.functional.logsigmoid(rewards_j - rewards_k).mean()
-        if return_outputs:
-            return loss, {"rewards_j": rewards_j, "rewards_k": rewards_k}
-        return loss
-
-
-# Train the model, woohoo.
-trainer = RewardTrainer(
-    model=model,
-    args=training_args,
-    train_dataset=tokenized_ds["train"],
-    eval_dataset=tokenized_ds["validation"],
-    compute_metrics=compute_metrics,
-    data_collator=RewardDataCollatorWithPadding(tokenizer=tokenizer),
-)
-
-trainer.train(script_args.resume_from_checkpoint)
-
-# Push to the hub so you can share it with people :D
-model.push_to_hub(script_args.model_name)
-tokenizer.push_to_hub(script_args.model_name)

pyproject.toml

@@ -0,0 +1,16 @@
+[tool.black]
+line-length = 119
+target-version = ['py38']
+
+[tool.ruff]
+ignore = ["E501", "E741", "W605"]
+select = ["E", "F", "I", "W"]
+line-length = 119
+
+# Ignore import violations in all `__init__.py` files.
+[tool.ruff.per-file-ignores]
+"__init__.py" = ["E402", "F401", "F403", "F811"]
+
+[tool.ruff.isort]
+lines-after-imports = 2
+known-first-party = ["trl"]

requirements.txt

@@ -3,3 +3,5 @@ torch>=1.4.0
 tqdm
 transformers
 accelerate
+peft>=0.3.0
+tyro>=0.5.7

scripts/stale.py

@@ -0,0 +1,61 @@
+# Copyright 2023 The HuggingFace Team, the AllenNLP library authors. 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.
+"""
+Script to close stale issue. Taken in part from the AllenNLP repository.
+https://github.com/allenai/allennlp.
+"""
+import os
+from datetime import datetime as dt
+from datetime import timezone
+
+from github import Github
+
+
+LABELS_TO_EXEMPT = [
+    "good first issue",
+    "good second issue",
+    "feature request",
+]
+
+
+def main():
+    g = Github(os.environ["GITHUB_TOKEN"])
+    repo = g.get_repo("huggingface/trl")
+    open_issues = repo.get_issues(state="open")
+
+    for issue in open_issues:
+        comments = sorted([comment for comment in issue.get_comments()], key=lambda i: i.created_at, reverse=True)
+        last_comment = comments[0] if len(comments) > 0 else None
+        if (
+            last_comment is not None
+            and last_comment.user.login == "github-actions[bot]"
+            and (dt.now(timezone.utc) - issue.updated_at).days > 7
+            and (dt.now(timezone.utc) - issue.created_at).days >= 30
+            and not any(label.name.lower() in LABELS_TO_EXEMPT for label in issue.get_labels())
+        ):
+            issue.edit(state="closed")
+        elif (
+            (dt.now(timezone.utc) - issue.updated_at).days > 23
+            and (dt.now(timezone.utc) - issue.created_at).days >= 30
+            and not any(label.name.lower() in LABELS_TO_EXEMPT for label in issue.get_labels())
+        ):
+            issue.create_comment(
+                "This issue has been automatically marked as stale because it has not had "
+                "recent activity. If you think this still needs to be addressed "
+                "please comment on this thread.\n\n"
+            )
+
+
+if __name__ == "__main__":
+    main()

setup.cfg

@@ -9,7 +9,3 @@ line_length = 119
 lines_after_imports = 2
 multi_line_output = 3
 use_parentheses = True
-
-[flake8]
-ignore = E203, E501, W503
-max-line-length = 119

setup.py

@@ -54,21 +54,30 @@ To create the package for pypi.
    Then push the change with a message 'set dev version'
 """
 
-from setuptools import setup, find_packages
+from setuptools import find_packages, setup
 
-__version__ = "0.3.0"  # expected format is one of x.y.z.dev0, or x.y.z.rc1 or x.y.z (no to dashes, yes to dots)
+
+__version__ = "0.7.7"  # expected format is one of x.y.z.dev0, or x.y.z.rc1 or x.y.z (no to dashes, yes to dots)
 
 REQUIRED_PKGS = [
     "torch>=1.4.0",
-    "transformers>=4.18.0",
+    "transformers>=4.31.0",
     "numpy>=1.18.2",
     "accelerate",
     "datasets",
+    "tyro>=0.5.11",
 ]
 EXTRAS = {
     "test": ["parameterized", "pytest", "pytest-xdist", "accelerate"],
-    "dev": ["parameterized", "pytest", "pytest-xdist", "black", "isort", "flake8>=3.8.3"],
+    "peft": ["peft>=0.4.0"],
+    "diffusers": ["diffusers>=0.18.0"],
+    "deepspeed": ["deepspeed>=0.9.5"],
+    "benchmark": ["wandb", "ghapi", "openrlbenchmark==0.2.1a5", "requests", "deepspeed"],
+    "quantization": ["bitsandbytes<=0.41.1"],
 }
+EXTRAS["dev"] = []
+for reqs in EXTRAS.values():
+    EXTRAS["dev"].extend(reqs)
 
 setup(
     name="trl",
@@ -86,7 +95,7 @@ setup(
         "Programming Language :: Python :: 3.9",
         "Programming Language :: Python :: 3.10",
     ],
-    url="https://github.com/lvwerra/trl",
+    url="https://github.com/huggingface/trl",
     packages=find_packages(),
     include_package_data=True,
     install_requires=REQUIRED_PKGS,
@@ -96,7 +105,7 @@ setup(
     long_description_content_type="text/markdown",
     zip_safe=False,
     version=__version__,
-    description="A Pytorch implementation of Proximal Policy Optimization for transfomer language models.",
+    description="Train transformer language models with reinforcement learning.",
     keywords="ppo, transformers, huggingface, gpt2, language modeling, rlhf",
     author="Leandro von Werra",
     author_email="leandro.vonwerra@gmail.com",

tests/test_best_of_n_sampler.py

@@ -0,0 +1,98 @@
+import unittest
+
+import torch
+from transformers import AutoTokenizer, GenerationConfig
+
+from trl import AutoModelForCausalLMWithValueHead
+from trl.core import LengthSampler
+from trl.extras import BestOfNSampler
+
+
+def queries_to_scores(list_of_strings):
+    return [torch.rand(1).item() for _ in list_of_strings]
+
+
+class BestOfNSamplerTester(unittest.TestCase):
+    """
+    Tests the BestOfNSampler class
+    """
+
+    ref_model_name = "trl-internal-testing/dummy-GPT2-correct-vocab"
+    output_length_sampler = LengthSampler(2, 6)
+    model = AutoModelForCausalLMWithValueHead.from_pretrained(ref_model_name)
+    tokenizer = AutoTokenizer.from_pretrained(ref_model_name)
+    tokenizer.pad_token = tokenizer.eos_token
+    output_length_sampler = LengthSampler(2, 6)
+
+    def test_different_input_types(self):
+        r"""
+        Tests if the different input types normalizer works
+        """
+
+        generation_config = GenerationConfig(
+            min_length=-1,
+            top_k=0.0,
+            top_p=1.0,
+            do_sample=True,
+            pad_token_id=self.tokenizer.eos_token_id,
+        )
+
+        output_length_sampler = LengthSampler(2, 6)
+
+        best_of_n = BestOfNSampler(
+            self.model,
+            self.tokenizer,
+            queries_to_scores,
+            length_sampler=output_length_sampler,
+            generation_config=generation_config,
+        )
+
+        queries = ["hello world", "goodbye world"]
+        tokenized_queries = [self.tokenizer.encode(query) for query in queries]
+
+        various_queries_formats = [
+            (tokenized_queries[0], 1),
+            (tokenized_queries, 2),
+            (torch.tensor(tokenized_queries[1]), 1),
+            ([torch.tensor(query) for query in tokenized_queries], 2),
+        ]
+
+        for q, expected_length in various_queries_formats:
+            results = best_of_n.generate(q)
+            self.assertIsInstance(results, list)
+            assert len(results) == expected_length
+
+    def test_different_sample_sizes_and_n_candidates_values(self):
+        r"""
+        Tests different sample sizes and n_candidates values
+        """
+        generation_config = GenerationConfig(
+            min_length=-1,
+            top_k=0.0,
+            top_p=1.0,
+            do_sample=True,
+            pad_token_id=self.tokenizer.eos_token_id,
+        )
+
+        output_length_sampler = LengthSampler(6, 10)
+
+        for sample_value, n_candidates_values, expected in [
+            (4, 2, 2),
+            (10, 3, 3),
+            (6, 4, 4),
+        ]:
+            best_of_n = BestOfNSampler(
+                self.model,
+                self.tokenizer,
+                queries_to_scores,
+                length_sampler=output_length_sampler,
+                generation_config=generation_config,
+                sample_size=sample_value,
+                n_candidates=n_candidates_values,
+            )
+
+            queries = ["hello world", "troll the world"]
+            tokenized_queries = [self.tokenizer.encode(query) for query in queries]
+            results = best_of_n.generate(tokenized_queries)
+            for result in results:
+                assert len(result) == expected

tests/test_data_collator_completion_only.py

@@ -0,0 +1,81 @@
+# 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.
+import unittest
+
+import torch
+from transformers import AutoTokenizer
+
+from trl import DataCollatorForCompletionOnlyLM
+
+
+class DataCollatorForCompletionOnlyLMTester(unittest.TestCase):
+    def test_data_collator_finds_response_template_llama2_tokenizer(self):
+        # this should ideally be tested with meta-llama/Llama-2-7b-hf
+        self.tokenizer = AutoTokenizer.from_pretrained("trl-internal-testing/dummy-GPT2-correct-vocab")
+        self.instruction = """### System: You are a helpful assistant.
+
+### User: How much is 2+2?
+
+### Assistant: 2+2 equals 4"""
+        self.instruction_template = "\n### User:"
+        self.response_template = "\n### Assistant:"
+
+        # GPT2Tokenizer: [198, 21017, 11787, 25] -> [11787, 25]
+        # Llama2Tokenizer: [29871, 13, 2277, 29937, 4911, 29901] -> [2277, 29937, 4911, 29901]
+        self.tokenized_instruction_w_context = self.tokenizer.encode(
+            self.instruction_template, add_special_tokens=False
+        )[2:]
+
+        # GPT2Tokenizer: [198, 21017, 15286, 25] -> [15286, 25]
+        # Llama2Tokenizer: [29871, 13, 2277, 29937, 4007, 22137, 29901] -> [2277, 29937, 4007, 22137, 29901]
+        self.tokenized_response_w_context = self.tokenizer.encode(self.response_template, add_special_tokens=False)[2:]
+
+        # Plain check on string
+        self.assertIn(self.response_template, self.instruction)
+        self.tokenized_instruction = self.tokenizer.encode(self.instruction, add_special_tokens=False)
+
+        # Test the fix for #598
+        # Pass already tokenized (w context) and truncated response_template so token_ids are like in the instruction + response
+        self.collator = DataCollatorForCompletionOnlyLM(self.tokenized_response_w_context, tokenizer=self.tokenizer)
+        self.collator.torch_call([self.tokenized_instruction])
+
+        # Test for PR #749
+        # Pass already tokenized (w context) instruction and response both so token_ids are like in the instruction + response
+        self.collator = DataCollatorForCompletionOnlyLM(
+            self.tokenized_response_w_context, self.tokenized_instruction_w_context, tokenizer=self.tokenizer
+        )
+        self.collator.torch_call([self.tokenized_instruction])
+
+    def test_data_collator_handling_of_long_sequences(self):
+        self.tokenizer = AutoTokenizer.from_pretrained("trl-internal-testing/dummy-GPT2-correct-vocab")
+        self.instruction = """### System: You are a helpful assistant.
+
+### User: How much is 2+2? I'm asking because I'm not sure. And I'm not sure because I'm not good at math.
+"""
+        self.response_template = "\n### Assistant:"
+        # check DataCollatorForCompletionOnlyLM using response template only
+        self.tokenized_instruction = self.tokenizer.encode(self.instruction, add_special_tokens=False)
+        self.collator = DataCollatorForCompletionOnlyLM(self.response_template, tokenizer=self.tokenizer)
+        encoded_instance = self.collator.torch_call([self.tokenized_instruction])
+        result = torch.all(encoded_instance["labels"] == -100)
+        self.assertTrue(result, "Not all values in the tensor are -100.")
+
+        # check DataCollatorForCompletionOnlyLM using response template and instruction template
+        self.instruction_template = "\n### User:"
+        self.collator = DataCollatorForCompletionOnlyLM(
+            self.response_template, self.instruction_template, tokenizer=self.tokenizer
+        )
+        encoded_instance = self.collator.torch_call([self.tokenized_instruction])
+        result = torch.all(encoded_instance["labels"] == -100)
+        self.assertTrue(result, "Not all values in the tensor are -100.")

tests/test_ddpo_trainer.py

@@ -0,0 +1,99 @@
+# Copyright 2023 metric-space, 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 gc
+import unittest
+
+import torch
+
+from trl import is_diffusers_available
+
+from .testing_utils import require_diffusers
+
+
+if is_diffusers_available():
+    from trl import DDPOConfig, DDPOTrainer, DefaultDDPOStableDiffusionPipeline
+
+
+def scorer_function(images, prompts, metadata):
+    return torch.randn(1) * 3.0, {}
+
+
+def prompt_function():
+    return ("cabbages", {})
+
+
+@require_diffusers
+class DDPOTrainerTester(unittest.TestCase):
+    """
+    Test the DDPOTrainer class.
+    """
+
+    def setUp(self):
+        self.ddpo_config = DDPOConfig(
+            num_epochs=2,
+            train_gradient_accumulation_steps=1,
+            per_prompt_stat_tracking_buffer_size=32,
+            sample_num_batches_per_epoch=2,
+            sample_batch_size=2,
+            mixed_precision=None,
+            save_freq=1000000,
+        )
+        pretrained_model = "hf-internal-testing/tiny-stable-diffusion-torch"
+        pretrained_revision = "main"
+
+        pipeline = DefaultDDPOStableDiffusionPipeline(
+            pretrained_model, pretrained_model_revision=pretrained_revision, use_lora=False
+        )
+
+        self.trainer = DDPOTrainer(self.ddpo_config, scorer_function, prompt_function, pipeline)
+
+        return super().setUp()
+
+    def tearDown(self) -> None:
+        gc.collect()
+
+    def test_loss(self):
+        advantage = torch.tensor([-1.0])
+        clip_range = 0.0001
+        ratio = torch.tensor([1.0])
+        loss = self.trainer.loss(advantage, clip_range, ratio)
+        self.assertEqual(loss.item(), 1.0)
+
+    def test_generate_samples(self):
+        samples, output_pairs = self.trainer._generate_samples(1, 2)
+        self.assertEqual(len(samples), 1)
+        self.assertEqual(len(output_pairs), 1)
+        self.assertEqual(len(output_pairs[0][0]), 2)
+
+    def test_calculate_loss(self):
+        samples, _ = self.trainer._generate_samples(1, 2)
+        sample = samples[0]
+
+        latents = sample["latents"][0, 0].unsqueeze(0)
+        next_latents = sample["next_latents"][0, 0].unsqueeze(0)
+        log_probs = sample["log_probs"][0, 0].unsqueeze(0)
+        timesteps = sample["timesteps"][0, 0].unsqueeze(0)
+        prompt_embeds = sample["prompt_embeds"]
+        advantage = torch.tensor([1.0], device=prompt_embeds.device)
+
+        self.assertEqual(latents.shape, (1, 4, 64, 64))
+        self.assertEqual(next_latents.shape, (1, 4, 64, 64))
+        self.assertEqual(log_probs.shape, (1,))
+        self.assertEqual(timesteps.shape, (1,))
+        self.assertEqual(prompt_embeds.shape, (2, 77, 32))
+        loss, approx_kl, clipfrac = self.trainer.calculate_loss(
+            latents, timesteps, next_latents, log_probs, advantage, prompt_embeds
+        )
+
+        self.assertTrue(torch.isfinite(loss.cpu()))

tests/test_dpo_trainer.py

@@ -0,0 +1,315 @@
+# 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.
+import tempfile
+import unittest
+
+import torch
+from datasets import Dataset
+from parameterized import parameterized
+from pytest import mark
+from transformers import AutoModelForCausalLM, AutoModelForSeq2SeqLM, AutoTokenizer, TrainingArguments
+
+from trl import DPOTrainer
+
+from .testing_utils import require_no_wandb, require_peft
+
+
+class DPOTrainerTester(unittest.TestCase):
+    @classmethod
+    def setUpClass(cls):
+        cls.model_id = "trl-internal-testing/dummy-GPT2-correct-vocab"
+        cls.model = AutoModelForCausalLM.from_pretrained(cls.model_id)
+        cls.ref_model = AutoModelForCausalLM.from_pretrained(cls.model_id)
+        cls.tokenizer = AutoTokenizer.from_pretrained(cls.model_id)
+        cls.tokenizer.pad_token = cls.tokenizer.eos_token
+
+        # get t5 as seq2seq example:
+        model_id = "trl-internal-testing/tiny-T5ForConditionalGeneration-correct-vocab"
+        cls.t5_model = AutoModelForSeq2SeqLM.from_pretrained(model_id)
+        cls.t5_ref_model = AutoModelForSeq2SeqLM.from_pretrained(model_id)
+        cls.t5_tokenizer = AutoTokenizer.from_pretrained(model_id)
+
+    def _init_dummy_dataset(self):
+        # fmt: off
+        dummy_dataset_dict = {
+            "prompt": [
+                "hello",
+                "how are you",
+                "What is your name?",
+                "What is your name?",
+                "Which is the best programming language?",
+                "Which is the best programming language?",
+                "Which is the best programming language?",
+                "[INST] How is the stock price? [/INST]",
+                "[INST] How is the stock price? [/INST] ",
+            ],
+            "chosen": [
+                "hi nice to meet you",
+                "I am fine",
+                "My name is Mary",
+                "My name is Mary",
+                "Python",
+                "Python",
+                "Python",
+                "$46 as of 10am EST",
+                "46 as of 10am EST",
+            ],
+            "rejected": [
+                "leave me alone",
+                "I am not fine",
+                "Whats it to you?",
+                "I dont have a name",
+                "Javascript",
+                "C++",
+                "Java",
+                " $46 as of 10am EST",
+                " 46 as of 10am EST",
+            ],
+        }
+        # fmt: on
+        return Dataset.from_dict(dummy_dataset_dict)
+
+    @parameterized.expand(
+        [
+            ["gpt2", "sigmoid", True],
+            ["t5", "hinge", False],
+            ["gpt2", "ipo", False],
+            ["t5", "ipo", True],
+            ["gpt2", "kto_pair", True],
+            ["t5", "kto_pair", False],
+        ]
+    )
+    def test_dpo_trainer(self, name, loss_type, pre_compute):
+        with tempfile.TemporaryDirectory() as tmp_dir:
+            training_args = TrainingArguments(
+                output_dir=tmp_dir,
+                per_device_train_batch_size=2,
+                max_steps=3,
+                remove_unused_columns=False,
+                gradient_accumulation_steps=1,
+                learning_rate=9e-1,
+                evaluation_strategy="steps",
+            )
+
+            dummy_dataset = self._init_dummy_dataset()
+
+            if name == "gpt2":
+                model = self.model
+                ref_model = self.ref_model
+                tokenizer = self.tokenizer
+            elif name == "t5":
+                model = self.t5_model
+                ref_model = self.t5_ref_model
+                tokenizer = self.t5_tokenizer
+
+            trainer = DPOTrainer(
+                model=model,
+                ref_model=ref_model,
+                beta=0.1,
+                loss_type=loss_type,
+                args=training_args,
+                tokenizer=tokenizer,
+                train_dataset=dummy_dataset,
+                eval_dataset=dummy_dataset,
+                precompute_ref_log_probs=pre_compute,
+            )
+
+            previous_trainable_params = {n: param.clone() for n, param in trainer.model.named_parameters()}
+
+            trainer.train()
+
+            self.assertIsNotNone(trainer.state.log_history[-1]["train_loss"])
+
+            # check the params have changed
+            for n, param in previous_trainable_params.items():
+                new_param = trainer.model.get_parameter(n)
+                # check the params have changed - ignore 0 biases
+                if param.sum() != 0:
+                    self.assertFalse(torch.equal(param, new_param))
+
+    def test_dpo_trainer_without_providing_ref_model(self):
+        with tempfile.TemporaryDirectory() as tmp_dir:
+            training_args = TrainingArguments(
+                output_dir=tmp_dir,
+                per_device_train_batch_size=2,
+                max_steps=3,
+                remove_unused_columns=False,
+                gradient_accumulation_steps=4,
+                learning_rate=9e-1,
+                evaluation_strategy="steps",
+            )
+
+            dummy_dataset = self._init_dummy_dataset()
+
+            trainer = DPOTrainer(
+                model=self.model,
+                ref_model=None,
+                beta=0.1,
+                args=training_args,
+                tokenizer=self.tokenizer,
+                train_dataset=dummy_dataset,
+                eval_dataset=dummy_dataset,
+                precompute_ref_log_probs=True,
+            )
+
+            previous_trainable_params = {n: param.clone() for n, param in trainer.model.named_parameters()}
+
+            trainer.train()
+
+            self.assertIsNotNone(trainer.state.log_history[-1]["train_loss"])
+
+            # check the params have changed
+            for n, param in previous_trainable_params.items():
+                new_param = trainer.model.get_parameter(n)
+                # check the params have changed - ignore 0 biases
+                if param.sum() != 0:
+                    self.assertFalse(torch.equal(param, new_param))
+
+    @require_peft
+    @mark.peft_test
+    def test_dpo_trainer_without_providing_ref_model_with_lora(self):
+        from peft import LoraConfig
+
+        lora_config = LoraConfig(
+            r=16,
+            lora_alpha=32,
+            lora_dropout=0.05,
+            bias="none",
+            task_type="CAUSAL_LM",
+        )
+
+        with tempfile.TemporaryDirectory() as tmp_dir:
+            training_args = TrainingArguments(
+                output_dir=tmp_dir,
+                per_device_train_batch_size=2,
+                max_steps=3,
+                remove_unused_columns=False,
+                gradient_accumulation_steps=4,
+                learning_rate=9e-1,
+                evaluation_strategy="steps",
+            )
+
+            dummy_dataset = self._init_dummy_dataset()
+
+            trainer = DPOTrainer(
+                model=self.model,
+                ref_model=None,
+                beta=0.1,
+                args=training_args,
+                tokenizer=self.tokenizer,
+                train_dataset=dummy_dataset,
+                eval_dataset=dummy_dataset,
+                peft_config=lora_config,
+                precompute_ref_log_probs=True,
+            )
+
+            previous_trainable_params = {n: param.clone() for n, param in trainer.model.named_parameters()}
+
+            trainer.train()
+
+            self.assertIsNotNone(trainer.state.log_history[-1]["train_loss"])
+
+            # check the params have changed
+            for n, param in previous_trainable_params.items():
+                if "lora" in n:
+                    new_param = trainer.model.get_parameter(n)
+                    # check the params have changed - ignore 0 biases
+                    if param.sum() != 0:
+                        self.assertFalse(torch.equal(param, new_param))
+
+    @require_no_wandb
+    def test_dpo_trainer_generate_during_eval_no_wandb(self):
+        with tempfile.TemporaryDirectory() as tmp_dir:
+            training_args = TrainingArguments(
+                output_dir=tmp_dir,
+                per_device_train_batch_size=2,
+                max_steps=3,
+                remove_unused_columns=False,
+                gradient_accumulation_steps=1,
+                learning_rate=9e-1,
+                evaluation_strategy="steps",
+            )
+
+            dummy_dataset = self._init_dummy_dataset()
+
+            with self.assertRaisesRegex(
+                ValueError,
+                expected_regex="`generate_during_eval=True` requires Weights and Biases to be installed."
+                " Please install `wandb` to resolve.",
+            ):
+                DPOTrainer(
+                    model=self.model,
+                    ref_model=None,
+                    beta=0.1,
+                    args=training_args,
+                    tokenizer=self.tokenizer,
+                    train_dataset=dummy_dataset,
+                    eval_dataset=dummy_dataset,
+                    generate_during_eval=True,
+                )
+
+    @require_peft
+    @mark.peft_test
+    def test_dpo_lora_save(self):
+        from peft import LoraConfig, get_peft_model
+
+        lora_config = LoraConfig(
+            r=16,
+            lora_alpha=32,
+            lora_dropout=0.05,
+            bias="none",
+            task_type="CAUSAL_LM",
+        )
+
+        # lora model
+        model = AutoModelForCausalLM.from_pretrained(self.model_id)
+        model_peft = get_peft_model(model, lora_config)
+
+        with tempfile.TemporaryDirectory() as tmp_dir:
+            training_args = TrainingArguments(
+                output_dir=tmp_dir,
+                per_device_train_batch_size=2,
+                max_steps=3,
+                remove_unused_columns=False,
+                gradient_accumulation_steps=4,
+                learning_rate=9e-1,
+                evaluation_strategy="steps",
+            )
+
+            dummy_dataset = self._init_dummy_dataset()
+
+            # dpo train lora model with a lora config
+            trainer = DPOTrainer(
+                model=model_peft,
+                ref_model=None,
+                beta=0.1,
+                args=training_args,
+                tokenizer=self.tokenizer,
+                train_dataset=dummy_dataset,
+                eval_dataset=dummy_dataset,
+                peft_config=lora_config,
+                precompute_ref_log_probs=True,
+            )
+
+            # train the model
+            trainer.train()
+
+            # save peft adapter
+            trainer.save_model()
+
+            # assert that the model is loaded without giving OSError
+            try:
+                AutoModelForCausalLM.from_pretrained(tmp_dir)
+            except OSError:
+                self.fail("Loading the saved peft adapter failed")

tests/test_e2e.py

@@ -0,0 +1,9 @@
+import subprocess
+
+
+def test_hello_world():
+    subprocess.run(
+        "python examples/hello_world.py",
+        shell=True,
+        check=True,
+    )

tests/test_environments.py

@@ -0,0 +1,273 @@
+# 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.
+
+import unittest
+from unittest.mock import patch
+
+import torch
+from transformers import AutoTokenizer
+
+from trl import AutoModelForCausalLMWithValueHead, TextEnvironment, TextHistory
+
+
+class DummyTool:
+    def __call__(self, text):
+        return text
+
+
+def dummy_generate(histories):
+    for i in range(len(histories)):
+        histories[i].append_segment("<request><DummyTool>test<call>", torch.tensor([1, 2, 3]), system=False)
+    return histories
+
+
+class TextHistoryTest(unittest.TestCase):
+    def test_text_history_init(self):
+        text = "Hello there!"
+        tokens = torch.tensor([1, 2, 3])
+
+        history = TextHistory(text, tokens)
+        self.assertEqual(history.text, text)
+        self.assertTrue(torch.equal(history.tokens, tokens))
+        self.assertTrue(torch.equal(history.token_masks, torch.zeros_like(tokens)))
+
+        history = TextHistory(text, tokens, system=False)
+        self.assertTrue(torch.equal(history.token_masks, torch.ones_like(tokens)))
+
+    def test_text_history_append_segment(self):
+        text = "Hello there!"
+        tokens = torch.tensor([1, 2, 3])
+
+        history = TextHistory(text, tokens)
+        history.append_segment("General Kenobi!", torch.tensor([4, 5, 6]), system=False)
+        self.assertEqual(history.text, text + "General Kenobi!")
+        self.assertTrue(torch.equal(history.tokens, torch.tensor([1, 2, 3, 4, 5, 6])))
+        self.assertTrue(torch.equal(history.token_masks, torch.tensor([0, 0, 0, 1, 1, 1])))
+
+        history.append_segment("You are a bold one!", torch.tensor([7, 8, 9]))
+        self.assertEqual(history.text, text + "General Kenobi!" + "You are a bold one!")
+        self.assertTrue(torch.equal(history.tokens, torch.tensor([1, 2, 3, 4, 5, 6, 7, 8, 9])))
+        self.assertTrue(torch.equal(history.token_masks, torch.tensor([0, 0, 0, 1, 1, 1, 0, 0, 0])))
+
+    def test_text_history_complete(self):
+        text = "Hello there!"
+        tokens = torch.tensor([1, 2, 3])
+        history = TextHistory(text, tokens)
+        history.complete()
+        self.assertTrue(history.completed)
+        self.assertFalse(history.truncated)
+
+        history.complete(truncated=True)
+        self.assertTrue(history.completed)
+        self.assertTrue(history.truncated)
+
+    def test_text_history_last_segment(self):
+        text = "Hello there!"
+        tokens = torch.tensor([1, 2, 3])
+        history = TextHistory(text, tokens)
+        history.append_segment("General Kenobi!", torch.tensor([4, 5, 6]))
+        history.append_segment("You are a bold one!", torch.tensor([7, 8, 9]))
+        self.assertEqual(history.last_text_segment, "You are a bold one!")
+
+    def test_text_history_split_query_response(self):
+        text = "Hello there!"
+        tokens = torch.tensor([1, 2, 3])
+        history = TextHistory(text, tokens)
+        history.append_segment("General Kenobi!", torch.tensor([4, 5, 6]), system=False)
+        history.append_segment("You are a bold one!", torch.tensor([7, 8, 9]), system=True)
+        query, response, mask = history.split_query_response_tokens()
+
+        self.assertTrue(torch.equal(query, torch.tensor([1, 2, 3])))
+        self.assertTrue(torch.equal(response, torch.tensor([4, 5, 6, 7, 8, 9])))
+        self.assertTrue(torch.equal(mask, torch.tensor([1, 1, 1, 0, 0, 0])))
+
+
+class TextEnvironmentTester(unittest.TestCase):
+    @classmethod
+    def setUpClass(cls):
+        # model_id
+        cls.model_id = "trl-internal-testing/dummy-GPT2-correct-vocab"
+
+        # get models and tokenizer
+        cls.gpt2_model = AutoModelForCausalLMWithValueHead.from_pretrained(cls.model_id)
+        cls.gpt2_tokenizer = AutoTokenizer.from_pretrained(cls.model_id)
+        cls.gpt2_tokenizer.pad_token = cls.gpt2_tokenizer.eos_token
+
+    def test_text_environment_setup(self):
+        env = TextEnvironment(
+            self.gpt2_model,
+            self.gpt2_tokenizer,
+            tools=[DummyTool()],
+            reward_fn=lambda x: torch.tensor(1),
+            prompt="I am a prompt!\n",
+        )
+        self.assertEqual(env.prompt, "I am a prompt!\n")
+        self.assertEqual(list(env.tools.keys()), ["DummyTool"])
+        self.assertTrue(isinstance(env.tools["DummyTool"], DummyTool))
+        self.assertEqual(env.reward_fn("Hello there!"), 1)
+
+    def test_text_environment_generate(self):
+        generation_kwargs = {"do_sample": False, "max_new_tokens": 4, "pad_token_id": self.gpt2_tokenizer.eos_token_id}
+        env = TextEnvironment(
+            self.gpt2_model,
+            self.gpt2_tokenizer,
+            tools=[DummyTool()],
+            reward_fn=lambda x: torch.tensor(1),
+            prompt="I am a prompt!\n",
+            generation_kwargs=generation_kwargs,
+        )
+
+        input_texts = ["this is a test", "this is another, longer test"]
+
+        model_inputs = [self.gpt2_tokenizer(txt, return_tensors="pt").input_ids.squeeze() for txt in input_texts]
+
+        generations_batched = env._generate_batched(model_inputs, batch_size=2)
+        generations_batched = self.gpt2_tokenizer.batch_decode(generations_batched)
+
+        generations_single = [env._generate_batched([inputs], batch_size=1)[0] for inputs in model_inputs]
+        generations_single = self.gpt2_tokenizer.batch_decode(generations_single)
+
+        self.assertEqual(generations_single, generations_batched)
+
+    def test_text_environment_tool_call_parsing(self):
+        string_valid = "Something something <request><Tool1>Hello there!<call>"
+        string_invalid_request = "Something something <Tool1>Hello there!<call>"
+        string_invalid_call = "Something something <request><Tool1>Hello there!"
+        string_invalid_tool = "Something something <request>|Tool2|Hello there!<call>"
+        string_invalid_random = "<>abcdefghijklm<>nopqrstuvwxyz<>"
+
+        env = TextEnvironment(
+            self.gpt2_model,
+            self.gpt2_tokenizer,
+            tools=[DummyTool()],
+            reward_fn=lambda x: torch.tensor(1),
+            prompt="I am a prompt!\n",
+        )
+        tool, response = env.parse_tool_call(string_valid)
+        self.assertEqual(tool, "Tool1")
+        self.assertEqual(response, "Hello there!")
+
+        tool, response = env.parse_tool_call(string_invalid_request)
+        self.assertEqual(tool, None)
+        self.assertEqual(response, None)
+
+        tool, response = env.parse_tool_call(string_invalid_call)
+        self.assertEqual(tool, None)
+        self.assertEqual(response, None)
+
+        tool, response = env.parse_tool_call(string_invalid_tool)
+        self.assertEqual(tool, None)
+        self.assertEqual(response, None)
+
+        tool, response = env.parse_tool_call(string_invalid_random)
+        self.assertEqual(tool, None)
+        self.assertEqual(response, None)
+
+    def test_text_environment_tool_truncation(self):
+        env = TextEnvironment(
+            self.gpt2_model,
+            self.gpt2_tokenizer,
+            tools={"dummy": lambda x: "a" * 1000},
+            reward_fn=lambda x: torch.tensor(1),
+            prompt="I am a prompt!\n",
+        )
+
+        env.max_tool_response = 100
+        history = env.step(TextHistory("<request><dummy>Hello there!<call>", torch.tensor([1, 2, 3])))
+        self.assertEqual(len(history.last_text_segment) - len(env.response_token), 100)
+
+        env.max_tool_response = 500
+        history = env.step(TextHistory("<request><dummy>Hello there!<call>", torch.tensor([1, 2, 3])))
+        self.assertEqual(len(history.last_text_segment) - len(env.response_token), 500)
+
+        env.max_tool_response = 1001
+        history = env.step(TextHistory("<request><dummy>Hello there!<call>", torch.tensor([1, 2, 3])))
+        self.assertEqual(len(history.last_text_segment) - len(env.response_token), 1000)
+
+        env.max_tool_response = 2000
+        history = env.step(TextHistory("<request><dummy>Hello there!<call>", torch.tensor([1, 2, 3])))
+        self.assertEqual(len(history.last_text_segment) - len(env.response_token), 1000)
+
+    @patch.object(TextEnvironment, "generate", side_effect=dummy_generate)
+    def test_text_environment_max_calls(self, mock_generate):
+        env = TextEnvironment(
+            self.gpt2_model,
+            self.gpt2_tokenizer,
+            tools={"DummyTool": DummyTool()},
+            reward_fn=lambda x: [torch.tensor(1) for _ in x],
+            prompt="I am a prompt!\n",
+        )
+
+        env.max_turns = 1
+        _, _, _, _, histories = env.run(["test"])
+        self.assertEqual(
+            histories[0].text, "I am a prompt!\n" + "test" + 1 * "<request><DummyTool>test<call>test<response>"
+        )
+
+        env.max_turns = 2
+        _, _, _, _, histories = env.run(["test"])
+        self.assertEqual(
+            histories[0].text, "I am a prompt!\n" + "test" + 2 * "<request><DummyTool>test<call>test<response>"
+        )
+
+        env.max_turns = 4
+        _, _, _, _, histories = env.run(["test"])
+        self.assertEqual(
+            histories[0].text, "I am a prompt!\n" + "test" + 4 * "<request><DummyTool>test<call>test<response>"
+        )
+
+    def test_text_environment_compute_rewards(self):
+        env = TextEnvironment(
+            self.gpt2_model,
+            self.gpt2_tokenizer,
+            tools={"DummyTool": DummyTool()},
+            reward_fn=lambda x: [torch.tensor(i) for i, _ in enumerate(x)],
+            prompt="I am a prompt!\n",
+        )
+
+        histories = [TextHistory("<request><DummyTool>test<call>", torch.tensor([1, 2, 3])) for _ in range(8)]
+        histories = env.compute_reward(histories)
+
+        for i in range(8):
+            self.assertEqual(histories[i].reward, i)
+
+    @patch.object(TextEnvironment, "generate", side_effect=dummy_generate)
+    def test_text_environment_run(self, mock_generate):
+        env = TextEnvironment(
+            self.gpt2_model,
+            self.gpt2_tokenizer,
+            tools={"DummyTool": DummyTool()},
+            reward_fn=lambda x: [torch.tensor(i) for i, _ in enumerate(x)],
+            prompt="I am a prompt!\n",
+            max_turns=2,
+        )
+        task_1 = "Hello there!"
+        task_2 = "Hello there! General Kenobi!"
+
+        query, response, response_mask, reward, histories = env.run([task_1, task_2])
+        self.assertEqual(len(query[0]), 9)
+        self.assertEqual(len(query[1]), 12)
+        self.assertEqual(len(response[0]), 14)
+        self.assertEqual(len(response[1]), 14)
+        self.assertEqual(response_mask[0].sum(), 2 * 3)  # mocked generate always adds 3 toknes
+        self.assertEqual(response_mask[1].sum(), 2 * 3)  # mocked generate always adds 3 toknes
+        self.assertEqual(reward[0], 0)
+        self.assertEqual(reward[1], 1)
+        self.assertEqual(
+            histories[0].text, "I am a prompt!\n" + "Hello there!" + 2 * "<request><DummyTool>test<call>test<response>"
+        )
+        self.assertEqual(
+            histories[1].text,
+            "I am a prompt!\n" + "Hello there! General Kenobi!" + 2 * "<request><DummyTool>test<call>test<response>",
+        )

tests/test_iterative_sft_trainer.py

@@ -0,0 +1,106 @@
+# 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.
+import tempfile
+import unittest
+
+import torch
+from datasets import Dataset
+from parameterized import parameterized
+from transformers import AutoModelForCausalLM, AutoModelForSeq2SeqLM, AutoTokenizer, TrainingArguments
+
+from trl import IterativeSFTTrainer
+
+
+class IterativeTrainerTester(unittest.TestCase):
+    @classmethod
+    def setUpClass(cls):
+        cls.model_id = "trl-internal-testing/dummy-GPT2-correct-vocab"
+        cls.model = AutoModelForCausalLM.from_pretrained(cls.model_id)
+        cls.tokenizer = AutoTokenizer.from_pretrained(cls.model_id)
+        cls.tokenizer.pad_token = cls.tokenizer.eos_token
+
+        # get t5 as seq2seq example:
+        model_id = "trl-internal-testing/tiny-T5ForConditionalGeneration-correct-vocab"
+        cls.t5_model = AutoModelForSeq2SeqLM.from_pretrained(model_id)
+        cls.t5_tokenizer = AutoTokenizer.from_pretrained(model_id)
+
+    def _init_tensor_dummy_dataset(self):
+        dummy_dataset_dict = {
+            "input_ids": [torch.tensor([5303, 3621]), torch.tensor([3666, 1438, 318]), torch.tensor([5303, 3621])],
+            "attention_mask": [torch.tensor([1, 1]), torch.tensor([1, 1, 1]), torch.tensor([1, 1])],
+            "labels": [torch.tensor([5303, 3621]), torch.tensor([3666, 1438, 318]), torch.tensor([5303, 3621])],
+        }
+
+        dummy_dataset = Dataset.from_dict(dummy_dataset_dict)
+        dummy_dataset.set_format("torch")
+        return dummy_dataset
+
+    def _init_textual_dummy_dataset(self):
+        dummy_dataset_dict = {
+            "texts": ["Testing the IterativeSFTTrainer.", "This is a test of the IterativeSFTTrainer"],
+            "texts_labels": ["Testing the IterativeSFTTrainer.", "This is a test of the IterativeSFTTrainer"],
+        }
+
+        dummy_dataset = Dataset.from_dict(dummy_dataset_dict)
+        dummy_dataset.set_format("torch")
+        return dummy_dataset
+
+    def setUp(self):
+        # initialize trainer
+        self.model.train()
+        return super().setUp()
+
+    @parameterized.expand(
+        [
+            ["gpt2", "tensor"],
+            ["gpt2", "text"],
+            ["t5", "tensor"],
+            ["t5", "text"],
+        ]
+    )
+    def test_iterative_step_from_tensor(self, model_name, input_name):
+        with tempfile.TemporaryDirectory() as tmp_dir:
+            # initialize dataset
+            if input_name == "tensor":
+                dummy_dataset = self._init_tensor_dummy_dataset()
+                inputs = {
+                    "input_ids": dummy_dataset["input_ids"],
+                    "attention_mask": dummy_dataset["attention_mask"],
+                    "labels": dummy_dataset["labels"],
+                }
+            else:
+                dummy_dataset = self._init_textual_dummy_dataset()
+                inputs = {
+                    "texts": dummy_dataset["texts"],
+                    "texts_labels": dummy_dataset["texts_labels"],
+                }
+
+            if model_name == "gpt2":
+                model = self.model
+                tokenizer = self.tokenizer
+            else:
+                model = self.t5_model
+                tokenizer = self.t5_tokenizer
+
+            args = TrainingArguments(
+                output_dir=tmp_dir,
+                per_device_train_batch_size=2,
+                max_steps=2,
+            )
+            iterative_trainer = IterativeSFTTrainer(model=model, args=args, tokenizer=tokenizer)
+
+            iterative_trainer.step(**inputs)
+
+            for param in iterative_trainer.model.parameters():
+                assert param.grad is not None

tests/test_modeling_value_head.py

@@ -30,6 +30,9 @@ ALL_CAUSAL_LM_MODELS = [
     "trl-internal-testing/tiny-random-BloomForCausalLM",
     "trl-internal-testing/tiny-random-GPT2LMHeadModel",
     "trl-internal-testing/tiny-random-CodeGenForCausalLM-sharded",
+    "trl-internal-testing/tiny-random-GPTNeoXForCausalLM-safetensors-sharded",
+    "trl-internal-testing/tiny-random-GPTNeoXForCausalLM-safetensors"
+    # "trl-internal-testing/tiny-random-LlamaForCausalLM", uncomment on the next transformers release
 ]
 
 ALL_SEQ2SEQ_MODELS = [