docs: add hf ckpt to faq, and include verl apis in the website (#427)

Now APIs can be displayed: 


![image](https://github.com/user-attachments/assets/6592ce68-7bf6-46cb-8dd3-a5fa6cd99f3e)

.readthedocs.yaml

@@ -6,7 +6,8 @@ version: 2
 build:
   os: ubuntu-22.04
   tools:
-    python: "3.8"
+    python: "3.11"
+    rust: "1.70"
 
 sphinx:
   configuration: docs/conf.py
@@ -14,3 +15,5 @@ sphinx:
 python:
   install:
     - requirements: docs/requirements-docs.txt
+    - method: pip
+      path: .

README.md

@@ -118,18 +118,20 @@ If you find the project helpful, please cite:
 verl is inspired by the design of Nemo-Aligner, Deepspeed-chat and OpenRLHF. The project is adopted and supported by Anyscale, Bytedance, LMSys.org, Shanghai AI Lab, Tsinghua University, UC Berkeley, UCLA, UIUC, and University of Hong Kong.
 
 ## Awesome work using verl
-- [Enhancing Multi-Step Reasoning Abilities of Language Models through Direct Q-Function Optimization](https://arxiv.org/abs/2410.09302)
-- [Flaming-hot Initiation with Regular Execution Sampling for Large Language Models](https://arxiv.org/abs/2410.21236)
-- [Process Reinforcement Through Implicit Rewards](https://github.com/PRIME-RL/PRIME/)
-- [TinyZero](https://github.com/Jiayi-Pan/TinyZero): a reproduction of DeepSeek R1 Zero recipe for reasoning tasks
-- [RAGEN](https://github.com/ZihanWang314/ragen): a general-purpose reasoning agent training framework
-- [Logic R1](https://github.com/Unakar/Logic-RL): a reproduced DeepSeek R1 Zero on 2K Tiny Logic Puzzle Dataset.
+- [TinyZero](https://github.com/Jiayi-Pan/TinyZero): a reproduction of **DeepSeek R1 Zero** recipe for reasoning tasks
+- [PRIME](https://github.com/PRIME-RL/PRIME): Process reinforcement through implicit rewards
+- [RAGEN](https://github.com/ZihanWang314/ragen): a general-purpose reasoning **agent** training framework
+- [Logic-RL](https://github.com/Unakar/Logic-RL): a reproduction of DeepSeek R1 Zero on 2K Tiny Logic Puzzle Dataset.
 - [deepscaler](https://github.com/agentica-project/deepscaler): iterative context scaling with GRPO
-- [critic-rl](https://github.com/HKUNLP/critic-rl): Teaching Language Models to Critique via Reinforcement Learning
-- [Easy-R1](https://github.com/hiyouga/EasyR1): Multi-Modality RL
+- [critic-rl](https://github.com/HKUNLP/critic-rl): LLM critics for code generation
+- [Easy-R1](https://github.com/hiyouga/EasyR1): **Multi-modal** RL training framework
+- [self-rewarding-reasoning-LLM](https://arxiv.org/pdf/2502.19613): self-rewarding and correction with **generative reward models**
+- [Search-R1](https://github.com/PeterGriffinJin/Search-R1): RL with reasoning and **searching (tool-call)** interleaved LLMs
+- [DQO](https://arxiv.org/abs/2410.09302): Enhancing multi-Step reasoning abilities of language models through direct Q-function optimization
+- [FIRE](https://arxiv.org/abs/2410.21236): Flaming-hot initiation with regular execution sampling for large language models
 
 ## Contribution Guide
-Contributions from the community are welcome!
+Contributions from the community are welcome! Please checkout our [roadmap](https://github.com/volcengine/verl/issues/22) and [release plan](https://github.com/volcengine/verl/issues/354).
 
 ### Code formatting
 We use yapf (Google style) to enforce strict code formatting when reviewing PRs. To reformat you code locally, make sure you installed **latest** `yapf`

docs/faq/faq.rst

@@ -55,3 +55,8 @@ Please set the following environment variable. The env var must be set before th
     export VLLM_ATTENTION_BACKEND=XFORMERS
 
 If in doubt, print this env var in each rank to make sure it is properly set.
+
+Checkpoints
+------------------------
+
+If you want to convert the model checkpoint into huggingface safetensor format, please refer to ``scripts/model_merger.py``.

docs/requirements-docs.txt

@@ -7,3 +7,6 @@ sphinx-markdown-tables
 
 # crate-docs-theme
 sphinx-rtd-theme
+
+# pin tokenizers version to avoid env_logger version req
+tokenizers==0.19.1

verl/protocol.py

@@ -84,7 +84,7 @@ def union_tensor_dict(tensor_dict1: TensorDict, tensor_dict2: TensorDict) -> Ten
     return tensor_dict1
 
 
-def union_numpy_dict(tensor_dict1: dict[np.ndarray], tensor_dict2: dict[np.ndarray]) -> dict[np.ndarray]:
+def union_numpy_dict(tensor_dict1: dict[str, np.ndarray], tensor_dict2: dict[str, np.ndarray]) -> dict[str, np.ndarray]:
     for key, val in tensor_dict2.items():
         if key in tensor_dict1:
             assert isinstance(tensor_dict2[key], np.ndarray)
@@ -448,19 +448,17 @@ class DataProto:
         return self
 
     def make_iterator(self, mini_batch_size, epochs, seed=None, dataloader_kwargs=None):
-        """Make an iterator from the DataProto. This is built upon that TensorDict can be used as a normal Pytorch
+        r"""Make an iterator from the DataProto. This is built upon that TensorDict can be used as a normal Pytorch
         dataset. See https://pytorch.org/tensordict/tutorials/data_fashion for more details.
 
+
         Args:
-            mini_batch_size (int): mini-batch size when iterating the dataset. We require that
-                ``batch.batch_size[0] % mini_batch_size == 0``
+            mini_batch_size (int): mini-batch size when iterating the dataset. We require that ``batch.batch_size[0] % mini_batch_size == 0``.
             epochs (int): number of epochs when iterating the dataset.
-            dataloader_kwargs: internally, it returns a DataLoader over the batch.
-                The dataloader_kwargs is the kwargs passed to the DataLoader
+            dataloader_kwargs (Any): internally, it returns a DataLoader over the batch. The dataloader_kwargs is the kwargs passed to the DataLoader.
 
         Returns:
-            Iterator: an iterator that yields a mini-batch data at a time. The total number of iteration steps is
-            ``self.batch.batch_size * epochs // mini_batch_size``
+            Iterator: an iterator that yields a mini-batch data at a time. The total number of iteration steps is ``self.batch.batch_size * epochs // mini_batch_size``
         """
         assert self.batch.batch_size[0] % mini_batch_size == 0, f"{self.batch.batch_size[0]} % {mini_batch_size} != 0"
         # we can directly create a dataloader from TensorDict