[Docs] Add SGLang backend documentation

Update LLM documentation with SGLang support:

llms_modules.rst:
- Add SGLangWrapper to Wrappers section
- Add AsyncSGLang to Wrappers section

llms.rst:
- Update Key Components to mention SGLang
- Add SGLang Quick Example alongside vLLM example
- Add SGLang weight sync classes to autosummary:
  - SGLangWeightSyncScheme
  - SGLangWeightSender
  - SGLangCollectiveTransport
  - get_sglang_model_metadata

Example shows both managed and external server modes:
- AsyncSGLang.from_pretrained() for managed server
- AsyncSGLang.connect() for external server

ghstack-source-id: 7b20c92
Co-authored-by: Cursor <cursoragent@cursor.com>
ghstack-source-id: 7b20c92
Pull-Request: #3435
Co-authored-by: Cursor <cursoragent@cursor.com>

Author: vmoens

.github/unittest/llm/scripts_llm/environment.yml

@@ -25,5 +25,6 @@ dependencies:
     - transformers
     - datasets
     - vllm
+    - sglang[all]
     - mcp
     - langdetect

docs/source/reference/llms.rst

@@ -12,14 +12,16 @@ Key Components
 --------------
 
 1. **Data Structures**: History class for conversation management, structured output classes
-2. **LLM Wrappers**: Unified interfaces for Transformers, vLLM, and AsyncVLLM  
+2. **LLM Wrappers**: Unified interfaces for Transformers, vLLM, SGLang, and async variants  
 3. **Environments**: ChatEnv, task-specific environments, and transforms
 4. **Collectors**: LLMCollector and RayLLMCollector for data collection
 5. **Objectives**: GRPOLoss, SFTLoss for training
 
 Quick Example
 -------------
 
+**Using vLLM backend:**
+
 .. code-block:: python
 
     from torchrl.modules.llm import vLLMWrapper, AsyncVLLM
@@ -36,6 +38,26 @@ Quick Example
     # Create collector
     collector = LLMCollector(env, policy, dialog_turns_per_batch=256)
 
+**Using SGLang backend:**
+
+.. code-block:: python
+
+    from torchrl.modules.llm import SGLangWrapper, AsyncSGLang
+    from torchrl.envs.llm import ChatEnv
+    from torchrl.collectors.llm import LLMCollector
+    
+    # Create SGLang engine (connects to server or launches managed server)
+    engine = AsyncSGLang.from_pretrained("Qwen/Qwen2.5-7B", tp_size=2)
+    # Or connect to existing server:
+    # engine = AsyncSGLang.connect("http://localhost:30000")
+    policy = SGLangWrapper(engine, tokenizer=tokenizer, input_mode="history")
+    
+    # Create environment
+    env = ChatEnv(tokenizer=tokenizer)
+    
+    # Create collector
+    collector = LLMCollector(env, policy, dialog_turns_per_batch=256)
+
 .. warning:: The LLM API is still under development and may change in the future. 
     Feedback, issues and PRs are welcome!
 
@@ -85,6 +107,10 @@ transform, or a boolean to the collector constructor.
     VLLMDoubleBufferWeightReceiver
     VLLMDoubleBufferTransport
     get_model_metadata
+    SGLangWeightSyncScheme
+    SGLangWeightSender
+    SGLangCollectiveTransport
+    get_sglang_model_metadata
 
 Legacy Weight Updaters (Deprecated)
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

docs/source/reference/llms_modules.rst

@@ -18,8 +18,10 @@ Wrappers
     LLMWrapperBase
     TransformersWrapper
     vLLMWrapper
+    SGLangWrapper
     RemoteTransformersWrapper
     AsyncVLLM
+    AsyncSGLang
 
 Data Structure Classes
 ----------------------