Added the README and script files for training sql_agent on NPU

docs/how-to/ascend_npu.md

@@ -0,0 +1,96 @@
+# agent-lightning x Ascend
+
+We have added support for **Huawei Ascend NPUs** in **agent-lightning**, and provided an example of training a SQL agent based on the **Spider dataset**.
+
+## Hardware Support
+
+- Atlas 200T A2 Box16
+- Atlas 900 A2 PODc
+- Atlas 800T A3
+
+At least **a single 40GB NPU** is required to run the Qwen2.5-Coder-1.5B-Instruct model.
+
+## Environment Setup
+
+### Basic Environment
+
+- Python: 3.11.13
+- CANN: 8.2.RC1
+- torch: 2.7.1+cpu
+- torch_npu: 2.7.1.dev20250724
+
+> For basic environment preparation, please refer to this [document](https://gitcode.com/Ascend/pytorch).
+
+### Configure Mirror Sources
+
+Before installing dependencies, it is recommended to configure pip mirrors:
+
+```
+pip config set global.index-url http://repo.huaweicloud.com/repository/pypi/simple
+pip config set global.extra-index-url "https://download.pytorch.org/whl/cpu/ https://mirrors.huaweicloud.com/ascend/repos/pypi"
+
+# Mirrors:
+# http://repo.huaweicloud.com/repository/pypi/simple
+# https://download.pytorch.org/whl/cpu/
+# https://mirrors.huaweicloud.com/ascend/repos/pypi
+```
+
+### Install vLLM & vLLM-Ascend
+
+```
+pip install vllm==0.10.0 --trusted-host repo.huaweicloud.com
+pip install vllm-Ascend==0.10.0rc1 --trusted-host repo.huaweicloud.com
+```
+
+### Install VERL
+
+```
+pip install verl==0.5.0
+```
+
+> ⚠️ Note: To ensure the VERL framework runs correctly on NPU, add the following two lines to the file `verl/utils/vllm_utils.py`:
+
+```
+from vllm_ascend.patch import platform
+from vllm_ascend.patch import worker
+```
+
+### Install agent-lightning
+
+```
+pip install agentlightning==0.2.1
+```
+
+### Install Other Dependencies
+
+```
+pip install autogen-agentchat autogen-ext mcp
+pip install langgraph "langchain[openai]" langchain-community langchain-text-splitters
+pip install sqlparse nltk
+```
+
+## Model
+
+We use the [Qwen2.5-Coder-1.5B-Instruct](https://huggingface.co/Qwen/Qwen2.5-Coder-1.5B) model to train the SQL agent. Running requires at least **one 40GB NPU**.
+
+## Dataset
+
+We use the Spider 1.0 dataset, which contains about 8,000 samples, including natural language questions, database schemas, and corresponding standard SQL queries.
+
+Training requires the following three Parquet files:
+
+- `train_spider.parquet`
+- `test_dev_500.parquet`
+- `test_dev.parquet`
+
+## Training Workflow
+
+1. **Prepare the dataset**: Convert the Spider dataset into Parquet format and place it in the `data/` directory.
+
+2. **Configure the environment**: Ensure vLLM-Ascend, VERL, and agent-lightning are correctly installed.
+
+3. **Start training**: Run the following command to begin training the SQL agent:
+
+   ```
+   python train_sql_agent_npu.py qwen
+   ```

docs/how-to/train-sql-agent.md

@@ -320,6 +320,92 @@ For the LLaMA profile, export an `HF_TOKEN` before running so VERL can download
     ```bash
     env RAY_DEBUG=legacy HYDRA_FULL_ERROR=1 VLLM_USE_V1=1 ray start --head --dashboard-host=0.0.0.0
     ```
+### Launch Training with NPUS
+
+We have added support for **Huawei Ascend NPUs** in **agent-lightning**, and create a  function `config_train_npu` in the script .
+
+#### Hardware Support
+
+- **Atlas 200T A2 Box16**
+- **Atlas 900 A2 PODc**
+- **Atlas 800T A3**
+
+At least **a single 40GB NPU** is required to run the **Qwen2.5-Coder-1.5B-Instruct** model.
+
+#### Environment Setup
+
+##### Basic Environment
+
+- **Python:** 3.11.13
+- **CANN:** 8.2.RC1
+- **torch:** 2.7.1+cpu
+- **torch_npu:** 2.7.1.dev20250724
+
+> For basic environment preparation, please refer to this [document](https://gitcode.com/Ascend/pytorch).
+
+##### Configure Mirror Sources
+
+Before installing dependencies, configure the following pip mirrors:
+
+```
+pip config set global.index-url http://repo.huaweicloud.com/repository/pypi/simple
+pip config set global.extra-index-url "https://download.pytorch.org/whl/cpu/ https://mirrors.huaweicloud.com/ascend/repos/pypi"
+```
+
+##### Install vLLM & vLLM-Ascend
+
+```
+pip install vllm==0.10.0 --trusted-host repo.huaweicloud.com
+pip install vllm-Ascend==0.10.0rc1 --trusted-host repo.huaweicloud.com
+```
+
+##### Install VERL
+
+```
+pip install verl==0.5.0
+```
+
+> ⚠️ **Note:** To ensure the VERL framework runs correctly on NPU, add the following lines to
+>  `verl/utils/vllm_utils.py`:
+
+```
+from vllm_ascend.patch import platform
+from vllm_ascend.patch import worker
+```
+
+##### Install Agent-Lightning
+
+```
+pip install agentlightning==0.2.1
+```
+
+##### Install Other Dependencies
+
+```
+pip install autogen-agentchat autogen-ext mcp
+pip install langgraph "langchain[openai]" langchain-community langchain-text-splitters
+pip install sqlparse nltk
+```
+
+#### Model
+
+We use the [**Qwen2.5-Coder-1.5B-Instruct**](https://huggingface.co/Qwen/Qwen2.5-Coder-1.5B) model to train the SQL agent.
+
+#### Dataset
+
+Refer to the method above for obtaining the dataset.
+
+#### Training Workflow
+
+1. **Prepare the dataset**: Convert the Spider dataset into Parquet format and place it in the `data/` directory.
+
+2. **Configure the environment**: Ensure vLLM-Ascend, VERL, and agent-lightning are correctly installed.
+
+3. **Start training**: Run the following command to begin training the SQL agent:
+
+```
+python train_sql_agent_npu.py npu
+```
 
 ### Debugging the Agent without VERL
 

examples/spider/README.md

@@ -36,6 +36,7 @@ Train a SQL agent using the Qwen2.5-Coder-1.5B-Instruct model with the following
 ```bash
 python train_sql_agent.py qwen
 ```
+If you want to use an NPU for training, please refer to the **Launch Training with NPUS** section in [How to Train a SQL Agent](../../docs/how-to/train-sql-agent.md).
 
 ### Debugging
 

examples/spider/train_sql_agent.py

@@ -132,11 +132,24 @@ def config_train_fast() -> Dict[str, Any]:
 
 def config_train_qwen() -> Dict[str, Any]:
     """A configuration for training with Qwen-2.5B."""
-
+    
     config = deepcopy(RL_TRAINING_CONFIG)
     return config
 
 
+def config_train_npu() -> Dict[str, Any]:
+    """A configuration for training with NPU."""
+
+    config = deepcopy(RL_TRAINING_CONFIG)
+    del config["actor_rollout_ref"]["rollout"]["engine_kwargs"]["vllm"]["enable_auto_tool_choice"]
+    del config["actor_rollout_ref"]["rollout"]["engine_kwargs"]["vllm"]["tool_call_parser"]
+    del config["trainer"]["logger"][1]
+    config["actor_rollout_ref"]["actor"]["use_torch_compile"] = False
+    config["trainer"]["val_before_train"] = False
+    config["trainer"]["save_freq"] = 256
+    config["trainer"]["device"] = "npu"
+    return config
+
 def config_train_llama() -> Dict[str, Any]:
     """A configuration for training with LLaMA-3.2-1B-Instruct.
 
@@ -171,21 +184,19 @@ def main() -> None:
 
     parser.add_argument(
         "config",
-        choices=["fast", "qwen", "llama"],
-        help="Training configuration: 'fast' (CI testing), 'qwen' (Qwen-2.5-Coder-1.5B), 'llama' (LLaMA-3.2-3B)",
+        choices=["fast","qwen","llama", "npu"],
+        help="Training configuration: 'fast' (CI testing), 'qwen' (Qwen-2.5-Coder-1.5B), 'llama' (LLaMA-3.2-3B),'npu' (Train with NPU)",
     )
 
     parser.add_argument(
         "--active-agent", type=str, help="Override the active agent name (default: auto-generated based on config)"
     )
-
+    
     args = parser.parse_args()
 
     # Get the appropriate configuration
-    config_functions = {"fast": config_train_fast, "qwen": config_train_qwen, "llama": config_train_llama}
-
+    config_functions = {"fast": config_train_fast,"qwen": config_train_qwen,"llama": config_train_llama,"npu": config_train_npu}
     config = config_functions[args.config]()
-
     # Set active agent - use provided value or default based on config choice
     active_agent = args.active_agent
 
@@ -196,4 +207,4 @@ def main() -> None:
 
 
 if __name__ == "__main__":
-    main()
+    main()