[docs] fix out-dated doc

.github/workflows/pr-test.yml

@@ -701,6 +701,10 @@ jobs:
               "num_gpus": 0,
               "test_file": "test_eval_config.py"
             },
+            {
+              "num_gpus": 0,
+              "test_file": "test_docs_consistency.py"
+            },
             {
               "num_gpus": 0,
               "test_file": "test_process_rollout_data.py"

.github/workflows/pr-test.yml.j2

@@ -101,6 +101,7 @@
         {'test_file': 'test_rm_deepscaler.py', 'num_gpus': 0},
         {'test_file': 'test_sample.py', 'num_gpus': 0},
         {'test_file': 'test_eval_config.py', 'num_gpus': 0},
+        {'test_file': 'test_docs_consistency.py', 'num_gpus': 0},
         {'test_file': 'test_process_rollout_data.py', 'num_gpus': 0},
         {'test_file': 'test_filter_long_prompt.py', 'num_gpus': 0},
         {'test_file': 'test_read_file_slicing.py', 'num_gpus': 0},

README.md

@@ -102,7 +102,7 @@ We also provide examples for some use cases not covered in the quick start guide
 
 For agentic RL workloads, the following examples plug into the standard rollout / Data Buffer loop through customization interfaces — they are not separate frameworks:
 
-- [`examples/multi_agent`](examples/multi_agent/README.md): Multi-agent rollout via a custom `--rollout-function-path`.
+- [`examples/multi_agent`](examples/multi_agent/README.md): Multi-agent generation via `--custom-generate-function-path` inside the standard rollout loop.
 - [`examples/search-r1`](examples/search-r1/): Search/RAG-style multi-turn generation via `--custom-generate-function-path`.
 - [`examples/fully_async`](examples/fully_async/README.md): Fully-async rollout, useful for long-tail agentic generation where some samples take much longer than others.
 - [`examples/coding_agent_rl`](examples/coding_agent_rl/README.md): End-to-end SWE coding-agent RL with sandboxed tool use, test-based rewards, and token-correct trajectory segments via `--custom-generate-function-path`.

README_zh.md

@@ -104,7 +104,7 @@ slime 被当作 RL 基础设施来开发，因为“脚本能跑起来”远远
 
 下面这些 example 通过 customization 接口接入标准的 rollout / Data Buffer 闭环，而不是独立的 framework：
 
-- [`examples/multi_agent`](examples/multi_agent/README.md)：通过自定义 `--rollout-function-path` 实现多 agent 的 rollout。
+- [`examples/multi_agent`](examples/multi_agent/README.md)：在标准 rollout loop 内通过 `--custom-generate-function-path` 实现多 agent 生成。
 - [`examples/search-r1`](examples/search-r1/)：通过 `--custom-generate-function-path` 实现 search/RAG 风格的多轮生成。
 - [`examples/fully_async`](examples/fully_async/README.md)：fully-async rollout，适合不同样本生成耗时差异较大的 long-tail agentic 场景。
 - [`examples/coding_agent_rl`](examples/coding_agent_rl/README.md)：端到端 SWE coding-agent RL，包含 sandboxed tool use、test-based reward，以及通过 `--custom-generate-function-path` 导出的 token-correct trajectory segments。

docker/README.md

@@ -26,7 +26,7 @@ just release-cu13      # CUDA 13 (cu130 base, Blackwell): publishes `latest-cu13
 
 `slimerl/slime:latest` tracks the CUDA 12 build. The tag suffixes (`-cu129` /
 `-cu130`) match the SGLang base image. `docker/Dockerfile` branches on the base
-image's CUDA version; it defaults to the cu130 SGLang base, and the cu129 base
+image's CUDA version; it defaults to the cu129 SGLang base, while the cu130 base
 is selected via build args (see `docker/justfile`).
 
 To build a single image directly without publishing:
@@ -55,4 +55,3 @@ The following components are pinned and rebuilt in the image:
 
 For a non-default GPU architecture list, pass
 `--build-arg DEEPEP_CUDA_ARCH_LIST='<torch arch list>'`.
-

docs/conf.py

@@ -4,9 +4,9 @@
 from datetime import datetime
 from pathlib import Path
 
-sys.path.insert(0, os.path.abspath("../.."))
+sys.path.insert(0, str(Path(__file__).resolve().parents[1]))
 
-__version__ = "0.0.1"
+__version__ = "0.3.1"
 
 project = "slime"
 copyright = f"2025-{datetime.now().year}, slime"
@@ -184,7 +184,7 @@ def _sync_examples(app):
             if not candidate.exists():
                 continue  # skip entirely if nothing suitable
             target_dir = out_dir / d.name
-            target_dir.mkdir(parents=True, exist_ok=True)
+            shutil.copytree(d, target_dir, ignore=shutil.ignore_patterns("README.md", "README_zh.md"))
             shutil.copy2(candidate, target_dir / "README.md")
             entries.append((d.name, f"_examples_synced/{d.name}/README.md"))
 

docs/en/advanced/megatron-config.md

@@ -74,7 +74,6 @@ megatron:
 ```bash
 python train.py \
   --advantage-estimator ppo \
-  --use-critic \
   --megatron-config-path megatron_ppo.yaml \
   --tensor-model-parallel-size 2 \
   --sequence-parallel \
@@ -89,6 +88,7 @@ python train.py \
 
 In this setup:
 
+- `--advantage-estimator ppo` enables the critic automatically; there is no separate `--use-critic` CLI flag.
 - CLI defines the shared topology and resource layout; in current PPO, critic training resources follow the actor configuration.
 - YAML defines the role-specific differences, such as `lr`, `load`, `save`, or optimizer / scheduler parameters.
 

docs/en/advanced/reproducibility.md

@@ -46,7 +46,7 @@ PYTHONPATH=/root/Megatron-LM/ python \
 Run training with:
 
 ```bash
-bash script/run-qwen2.5-0.5B-reproducibility.sh
+bash scripts/run-qwen2.5-0.5B-reproducibility.sh
 ```
 
 The wandb screenshots are recorded in [pull#370](https://github.com/THUDM/slime/pull/370).

docs/en/developer_guide/ci.md

@@ -77,7 +77,7 @@ Agent adapter tests are kept in a separate CPU job because they need extra SDK d
 Useful local commands:
 
 ```bash
-python tests/test_agent_trajectory.py
+python tests/test_agent/test_trajectory_manager_branching.py
 python -m pytest tests/test_megatron_argument_validation.py tests/plugin_contracts/test_plugin_generate_contracts.py
 ```
 

docs/en/examples/glm4.7-30B-A3B.md

@@ -32,12 +32,12 @@ Execute the training script:
 
 ```bash
 cd /root/slime
-bash scripts/run-glm4.7-30B-A3B-8gpus.sh
+bash scripts/run-glm4.7-30B-A3B.sh
 ```
 
 ### Parameter Introduction
 
-Here, we will briefly introduce the key parts in the [run-glm4.7-30B-A3B-8gpus.sh](https://github.com/THUDM/slime/blob/main/scripts/run-glm4.7-30B-A3B-8gpus.sh) script.
+Here, we will briefly introduce the key parts in [run-glm4.7-30B-A3B.sh](../../../scripts/run-glm4.7-30B-A3B.sh).
 
 #### MoE Configuration
 
@@ -123,20 +123,15 @@ SPEC_ARGS=(
 >
 > For other models with MTP training support (e.g., MiMo), see `scripts/run-mimo-7B-rl-eagle.sh` as a reference.
 
-### Multi-Node Support
-
-For multi-node training (e.g., 2×8 H100), use the multi-node script:
+### Multi-Node Adaptation
 
-```bash
-cd /root/slime
-export BASE_DIR=/shared/path  # accessible by all nodes
-bash scripts/run-glm4.7-30B-A3B.sh
-```
+The checked-in `scripts/run-glm4.7-30B-A3B.sh` launcher starts a local, single-node Ray cluster and passes `--actor-num-nodes 1`; it is not a drop-in multi-node launcher. To adapt this recipe for multi-node training (for example, 2×8 H100), start or connect all workers to the same Ray cluster and update the launcher as follows:
 
 Key modifications for multi-node:
 
   - Place the model and data on a path accessible by all nodes.
   - Set `MASTER_ADDR` to an address accessible by all nodes.
+  - Set `--actor-num-nodes` to the number of training nodes instead of `1`.
   - Remove CPU Adam configurations (distributed optimizer reduces per-GPU memory usage).
   - Adjust parallelism: e.g., TP=4, PP=2, EP=8, CP=2.
 

docs/en/examples/qwen3-4b-base-openhermes.md

@@ -45,7 +45,7 @@ Execute the training:
 
 ```bash
 cd /root/slime
-bash script/run-qwen3-4B-base-sft.sh
+bash scripts/run-qwen3-4B-base-sft.sh
 ```
 
 ### Parameter Introduction