fix: incorrect paths in document and bugs in dynamic scheduling (RLinf#303)

Signed-off-by: Lin-xs <1833080950@qq.com>

Author: Lin-xs

.github/CODEOWNERS

@@ -2,7 +2,7 @@
 /docker @andylin-hao
 
 /examples/embodiment @guozhen1997
-/examples/math @guozhen1997 @andylin-hao @Lin-xs
+/examples/reasoning @guozhen1997 @andylin-hao @Lin-xs
 
 /ray_utils @Lin-xs
 /requirements @andylin-hao

docs/source-en/rst_source/examples/reasoning.rst

@@ -102,8 +102,8 @@ Before launching, check the configuration file. Key fields include:
 
 Recommended configurations can be found in:
 
-- ``examples/math/config/qwen2.5-1.5b-grpo-megatron.yaml``  
-- ``examples/math/config/qwen2.5-7b-grpo-megatron.yaml``  
+- ``examples/reasoning/config/math/qwen2.5-1.5b-grpo-megatron.yaml``  
+- ``examples/reasoning/config/math/qwen2.5-7b-grpo-megatron.yaml``  
 
 **3. Launch Command**
 
@@ -118,7 +118,7 @@ Run the following commands to start the Ray cluster and begin training:
    if [ "$RANK" -eq 0 ]; then
        bash check_ray.sh 128; # set to number of accelerators/GPUs in the cluster
        cd /path_to_RLinf;
-       bash examples/math/qwen2.5/run_main_math_grpo_megatron.sh grpo-1.5b-megatron # change config file
+       bash examples/reasoning/run_main_grpo_math.sh qwen2.5-1.5b-grpo-megatron # change config file
    else
      if [ "$RANK" -eq 1 ]; then
          sleep 3m

docs/source-en/rst_source/start/distribute.rst

@@ -84,7 +84,7 @@ Edit the sample YAML:
 
 .. code-block:: yaml
 
-   # examples/math/config/qwen2.5-1.5b-grpo-megatron.yaml
+   # examples/reasoning/config/math/qwen2.5-1.5b-grpo-megatron.yaml
    cluster:
      num_nodes: 4          # adapt to your cluster
      component_placement:
@@ -94,19 +94,19 @@ Launch from the head node:
 
 .. code-block:: bash
 
-   bash examples/math/run_main_math_grpo_megatron.sh \
+   bash examples/reasoning/run_main_grpo_math.sh \
         qwen2.5-1.5b-grpo-megatron
 
 
 Disaggregated
 ^^^^^^^^^^^^^^^^^^
 
 Different stages receive disjoint GPU ranges,
-allowing fine-grained pipeliningng. Edit the pipeline YAML:
+allowing fine-grained pipelining. Edit the pipeline YAML:
 
 .. code-block:: yaml
 
-   # examples/math/config/qwen2.5-1.5b-grpo-megatron-pipeline.yaml
+   # examples/reasoning/config/math/qwen2.5-1.5b-grpo-megatron-pipeline.yaml
    cluster:
      num_nodes: 4
      component_placement:
@@ -122,5 +122,5 @@ Start the job:
 
 .. code-block:: bash
 
-   bash examples/math/run_main_math_pipeline_grpo_megatron.sh \
+   bash examples/reasoning/run_main_grpo_math.sh \
         qwen2.5-1.5b-grpo-megatron-pipeline

docs/source-en/rst_source/start/llm.rst

@@ -49,7 +49,7 @@ Launch Training
 For user convenience, our configuration file is set up to run with a single GPU by default.  
 However, if you have multiple GPUs and wish to accelerate the quickstart process,  
 we highly recommend updating the following configuration option in  
-``./examples/math/config/qwen2.5-1.5b-single-gpu.yaml``:  
+``./examples/reasoning/config/math/qwen2.5-1.5b-single-gpu.yaml``:  
 ``cluster.component_placement``.
 
 
@@ -75,7 +75,7 @@ After these modifications, launch the following script to start training!
 
 .. code-block:: bash
 
-   bash examples/math/run_main_math_grpo_megatron.sh qwen2.5-1.5b-single-gpu
+   bash examples/reasoning/run_main_grpo_math.sh qwen2.5-1.5b-single-gpu
 
 **Step 3: View the results:**
 

docs/source-en/rst_source/tutorials/advance/version.rst

@@ -6,7 +6,7 @@ reinforcement-learning pipeline. For the current release **SGLang and vLLM** is
 
 .. note::
 
-   RLinf is compatible with **SGLang 0.4.4 → 0.4.9**, **vLLM 0.8.5  → 0.8.5.post1**.  
+   RLinf is compatible with **SGLang 0.4.4 → 0.5.2**, **vLLM 0.8.5  → 0.8.5.post1**.  
    No manual patching is required – the framework detects the installed
    version and loads the matching shim automatically.
 
@@ -35,7 +35,7 @@ Install via pip
    pip install sglang==0.4.8
 
    # Latest supported
-   pip install sglang==0.4.9
+   pip install sglang==0.5.2
 
    # Install vLLM
    pip install vllm==0.8.5
@@ -106,43 +106,3 @@ Install from Source
         cuda_graph_max_bs: 128 # the maximum batch size for cuda graph. If the batch size is larger than this, cuda graph will not be used.
 
     ...
-
-
-Internal Version Routing
-------------------------
-
-Directory layout::
-
-   rlinf/hybrid_engines/sglang/
-   ├── __init__.py               # Version detection and routing
-   ├── sglang_worker.py          # Main worker implementation
-   ├── sglang_0_4_4/             # SGLang 0.4.4 specific implementation
-   │   ├── __init__.py
-   │   ├── io_struct.py          # I/O structures for 0.4.4
-   │   ├── sgl_engine.py         # Engine implementation for 0.4.4
-   │   ├── sgl_scheduler.py      # Scheduler for 0.4.4
-   │   └── tokenizer_manager.py  # Tokenizer management for 0.4.4
-   └── sglang_0_4_x/             # Future version implementations
-       └── ...
-
-The loader in ``__init__.py`` resolves the installed package:
-
-.. code-block:: python
-
-   from importlib.metadata import PackageNotFoundError, version
-
-   def get_version(pkg):
-       try:
-           return version(pkg)
-       except PackageNotFoundError:
-           return None
-
-   package_name = "sglang"
-   package_version = get_version(package_name)
-   
-   if package_version == "0.4.4":
-       sglang_version = "0.4.4"
-       from .sglang_0_4_4 import io_struct
-       from .sglang_0_4_4.sgl_engine import Engine
-   else:
-       raise ValueError(f"sglang version {package_version} not supported")

docs/source-en/rst_source/tutorials/extend/new_model_megatron.rst

@@ -468,7 +468,7 @@ Below is an example YAML configuration file for the qwen2.5 model family.
 
 After adapting your new model, you can refer to this YAML configuration file and make appropriate modifications.
 
-**File:** ``examples/math/config/qwen2.5-1.5b-grpo-megatron.yaml``  
+**File:** ``examples/reasoning/config/math/qwen2.5-1.5b-grpo-megatron.yaml``  
 
 Set Megatron parameters used by RLinf.
 

docs/source-en/rst_source/tutorials/mode/collocated.rst

@@ -59,4 +59,4 @@ Given the above placement configuration, users can use proper `ComponentPlacemen
     )
 
 `ModelParallelComponentPlacement` supports two types of placement: collocated and disaggregated. More importantly, it deals with rank arrangement that allows efficient model weight update from training to rollout. It parses the configuration and generates placements for different components. The generated placement is then enforced during worker launching.
-Refer to `Math RL training python script <https://github.com/RLinf/RLinf/blob/main/examples/math/main_math.py>`_ for the complete code.
+Refer to `Math RL training python script <https://github.com/RLinf/RLinf/blob/main/examples/reasoning/main_grpo.py>`_ for the complete code.

docs/source-en/rst_source/tutorials/mode/disaggregated.rst

@@ -35,4 +35,4 @@ Currently, whether the execution is pipelined is decided by the underlying code
 
 **ComponentPlacement programming**
 
-As described in :doc:`collocated`, the placement configuration in the yaml file can be parsed by `ComponentPlacement` and enforced on workers. Refer to `Math RL training with pipelining <https://github.com/RLinf/RLinf/blob/main/examples/math/main_math_pipeline.py>`_ for the complete code.
+As described in :doc:`collocated`, the placement configuration in the yaml file can be parsed by `ComponentPlacement` and enforced on workers. Refer to `Math RL training with pipelining <https://github.com/RLinf/RLinf/blob/main/examples/reasoning/main_grpo.py>`_ for the complete code.

docs/source-en/rst_source/tutorials/scheduler/auto-placement.rst

@@ -45,7 +45,7 @@ Use the provided shell script to run the auto placement tool:
 
 .. code-block:: bash
 
-   cd examples/math
+   cd examples/reasoning
    ./run_placement_autotune.sh [config_name]
 
 Where ``config_name`` is the name of your configuration file.

docs/source-en/rst_source/tutorials/user/flow.rst

@@ -19,7 +19,7 @@ For example:
 - Configs for training a **VLA** agent in embodied tasks live under
   ``examples/embodiment/config``.
 - Configs for training an **LLM** on math reasoning live under
-  ``examples/math/config``.
+  ``examples/reasoning/config/math``.
 
 As a starting point, we recommend getting familiar with the YAML structure of
 these examples, then iterating toward your custom task. Key options include