Merge branch 'main' into python-3.13

.github/workflows/tests_latest.yml

@@ -17,7 +17,7 @@ jobs:
     steps:
       - name: Git checkout
         uses: actions/checkout@v4
-        with: { ref: v0.13-release }
+        with: { ref: v0.15-release }
       - name: Set up Python 3.12
         uses: actions/setup-python@v5
         with:

CITATION.cff

@@ -31,4 +31,4 @@ keywords:
   - pytorch
   - transformers
 license: Apache-2.0
-version: 0.13
+version: 0.15

CONTRIBUTING.md

@@ -23,7 +23,7 @@ There are several ways you can contribute to TRL:
 * Contribute to the examples or the documentation.
 
 If you don't know where to start, there is a special [Good First
-Issue](https://github.com/huggingface/trl/contribute) listing. It will give you a list of
+Issue](https://github.com/huggingface/trl/labels/%F0%9F%91%B6%20good%20first%20issue) listing. It will give you a list of
 open issues that are beginner-friendly and help you start contributing to open-source. The best way to do that is to open a Pull Request and link it to the issue that you'd like to work on. We try to give priority to opened PRs as we can easily track the progress of the fix, and if the contributor does not have time anymore, someone else can take the PR over.
 
 For something slightly more challenging, you can also take a look at the [Good Second Issue](https://github.com/huggingface/trl/labels/Good%20Second%20Issue) list. In general though, if you feel like you know what you're doing, go for it and we'll help you get there! 🚀

README.md

@@ -137,39 +137,26 @@ trainer = RewardTrainer(
 trainer.train()
 ```
 
-### `RLOOTrainer`
+### `GRPOTrainer`
 
-`RLOOTrainer` implements a [REINFORCE-style optimization](https://huggingface.co/papers/2402.14740) for RLHF that is more performant and memory-efficient than PPO. Here is a basic example of how to use the `RLOOTrainer`:
+`GRPOTrainer` implements the [Group Relative Policy Optimization (GRPO) algorithm](https://huggingface.co/papers/2402.03300) that is more memory-efficient than PPO and was used to train [Deepseek AI's R1](https://huggingface.co/deepseek-ai/DeepSeek-R1).
 
 ```python
-from trl import RLOOConfig, RLOOTrainer, apply_chat_template
 from datasets import load_dataset
-from transformers import (
-    AutoModelForCausalLM,
-    AutoModelForSequenceClassification,
-    AutoTokenizer,
-)
+from trl import GRPOConfig, GRPOTrainer
 
-tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-0.5B-Instruct")
-reward_model = AutoModelForSequenceClassification.from_pretrained(
-    "Qwen/Qwen2.5-0.5B-Instruct", num_labels=1
-)
-ref_policy = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-0.5B-Instruct")
-policy = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-0.5B-Instruct")
+dataset = load_dataset("trl-lib/tldr", split="train")
 
-dataset = load_dataset("trl-lib/ultrafeedback-prompt")
-dataset = dataset.map(apply_chat_template, fn_kwargs={"tokenizer": tokenizer})
-dataset = dataset.map(lambda x: tokenizer(x["prompt"]), remove_columns="prompt")
+# Dummy reward function: rewards completions that are close to 20 characters
+def reward_len(completions, **kwargs):
+    return [-abs(20 - len(completion)) for completion in completions]
 
-training_args = RLOOConfig(output_dir="Qwen2.5-0.5B-RL")
-trainer = RLOOTrainer(
-    config=training_args,
-    processing_class=tokenizer,
-    policy=policy,
-    ref_policy=ref_policy,
-    reward_model=reward_model,
-    train_dataset=dataset["train"],
-    eval_dataset=dataset["test"],
+training_args = GRPOConfig(output_dir="Qwen2-0.5B-GRPO", logging_steps=10)
+trainer = GRPOTrainer(
+    model="Qwen/Qwen2-0.5B-Instruct",
+    reward_funcs=reward_len,
+    args=training_args,
+    train_dataset=dataset,
 )
 trainer.train()
 ```

commands/run_sft.sh

@@ -42,7 +42,7 @@ accelerate launch $EXTRA_ACCELERATE_ARGS \
     --output_dir $OUTPUT_DIR \
     --max_steps $MAX_STEPS \
     --per_device_train_batch_size $BATCH_SIZE \
-    --max_seq_length $SEQ_LEN \
+    --max_length $SEQ_LEN \
     $EXTRA_TRAINING_ARGS
 """
 

docs/source/_toctree.yml

@@ -68,6 +68,8 @@
       title: Online DPO
     - local: gkd_trainer
       title: GKD
+    - local: grpo_trainer
+      title: GRPO
     - local: kto_trainer
       title: KTO
     - local: nash_md_trainer

docs/source/alignprop_trainer.mdx → docs/source/alignprop_trainer.md

@@ -16,7 +16,7 @@ The `alignprop.py` script is a working example of using the `AlignProp` trainer
 
 **Note:** one A100 GPU is recommended to get this running. For lower memory setting, consider setting truncated_backprop_rand to False. With default settings this will do truncated backpropagation with K=1.
 
-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
+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 alignprop.py --hf_user_access_token <token>
@@ -26,7 +26,7 @@ To obtain the documentation of `stable_diffusion_tuning.py`, please run `python
 
 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 randomized truncation range (`--alignprop_config.truncated_rand_backprop_minmax=(0,50)`) the first number should be equal and greater to 0, while the second number should equal or less to the number of diffusion timesteps (sample_num_steps)
+- The configurable randomized truncation range (`--alignprop_config.truncated_rand_backprop_minmax=(0,50)`) the first number should be equal and greater than 0, while the second number should equal or less to the number of diffusion timesteps (sample_num_steps)
 - The configurable truncation backprop absolute step (`--alignprop_config.truncated_backprop_timestep=49`) the number should be less than the number of diffusion timesteps (sample_num_steps), it only matters when truncated_backprop_rand is set to False
 
 ## Setting up the image logging hook function

docs/source/bco_trainer.mdx → docs/source/bco_trainer.md

@@ -62,7 +62,7 @@ embedding_model = Accelerator().prepare_model(self.embedding_model)
 embedding_func = partial(embed_prompt, model=embedding_model)
 ```
 
-Set `prompt_sample_size` to defined how many prompts are selected to train the UDM classifier and start the training with the provided embedding function:
+Set `prompt_sample_size` to define how many prompts are selected to train the UDM classifier and start the training with the provided embedding function:
 
 ```py
 training_args = BCOConfig(
@@ -97,4 +97,4 @@ To scale how much the auxiliary loss contributes to the total loss, use the hype
 
 ## BCOConfig
 
-[[autodoc]] BCOConfig
+[[autodoc]] BCOConfig

docs/source/best_of_n.mdx → docs/source/best_of_n.md

@@ -67,6 +67,6 @@ 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
+Furthermore, at the time of initialization you can set the seed to control the repeatability of the generation process and the number of samples to generate for each query
 
 

docs/source/clis.mdx → docs/source/clis.md

@@ -7,12 +7,13 @@ Currently supported CLIs are:
 #### Training commands
 
 - `trl dpo`: fine-tune a LLM with DPO
+- `trl grpo`: fine-tune a LLM with GRPO
 - `trl kto`: fine-tune a LLM with KTO
 - `trl sft`: fine-tune a LLM with SFT
 
 #### Other commands
 
-- `trl chat`: quickly spin up a LLM fine-tuned for chatting
+- `trl chat`: quickly spin up an LLM fine-tuned for chatting
 - `trl env`: get the system information
 
 ## Fine-tuning with the CLI