feat: vllm speedup evaluation and evluation guide

Author: realtmxi

docs/EVALUATION_GUIDE_EN.md

@@ -0,0 +1,107 @@
+# Evaluation Guide
+
+This guide explains how to quickly set up and run WebShop environment evaluations using vLLM for accelerated throughput, as well as fallback options for the classic (slower) evaluation scripts.
+
+---
+
+## Prerequisites
+
+1. **Python 3.8+** installed on your system.
+2. **pip** package manager.
+3. **vLLM Server** (OpenAI-compatible) or local vLLM endpoint.
+4. **AgentENV–WebShop** service.
+
+---
+
+## 1. Install and Prepare the Package
+
+From the root of the repository:
+
+```bash
+# Install the agentgym package in editable mode
+cd openmanus_rl/agentgym/agentenv/
+pip install -e .
+```
+
+This ensures `agentenv` is available on your `PYTHONPATH` without extra hacks.
+
+---
+
+## 2. Start the WebShop Environment Server
+
+1. Navigate to the WebShop service directory:
+
+   ```bash
+   cd agentenv/agentenv-webshop
+   ```
+
+2. Launch the server on port **36001** (or your preferred port):
+
+   ```bash
+   webshop --host 0.0.0.0 --port 36001
+   ```
+
+Leave this process running in the background or in a separate terminal.
+
+---
+
+## 3. Run vLLM‑Accelerated Evaluation
+
+1. Execute the helper script to start or configure your vLLM server:
+
+   ```bash
+   bash openmanus_rl/evaluation/run_vllm.sh
+   ```
+
+2. Ensure your `PYTHONPATH` includes the `agentenv` package:
+
+   ```bash
+   export PYTHONPATH=./openmanus_rl/agentgym/agentenv:$PYTHONPATH
+   ```
+
+3. Launch the evaluation driver:
+
+   ```bash
+   python openmanus_rl/evaluation/vllm_eval_webshop.py
+   ```
+
+This will run tasks against the WebShop environment via your vLLM endpoint for maximum speed.
+
+---
+
+## 4. Legacy Evaluation Scripts (Slower)
+
+If you prefer the classic evaluation (single‑model, non‑vLLM), use one of these scripts:
+
+* **Basic (single‑threaded):**
+
+  ```bash
+  bash agentgym/agentenv/examples/basic/base_eval_webshop.sh
+  ```
+
+* **Distributed (multi‑worker):**
+
+  ```bash
+  bash agentgym/agentenv/examples/distributed_eval_scripts/distributed_eval_webshop.sh
+  ```
+
+Expect these to run significantly slower than the vLLM‑driven workflow.
+
+---
+
+## 5. Troubleshooting
+
+* **ModuleNotFoundError:**
+  Make sure you ran `pip install -e .` in `openmanus_rl/agentgym` and removed any stale `PYTHONPATH` overrides.
+
+* **Port conflicts on 36001:**
+  Either kill the process using that port or choose a different port and update both the WebShop server and evaluation script arguments.
+
+* **vLLM connection errors:**
+  Verify that your vLLM server is up (`run_vllm.sh` logs) and that the `base_url` in your script matches its endpoint.
+
+---
+
+## License
+
+This project is licensed under the MIT License. See [LICENSE](../../LICENSE) for details.

openmanus_rl/agentgym/agentenv/agentenv/controller/__init__.py

@@ -1,4 +1,5 @@
 from .agent import Agent
 from .env import BaseEnvClient, StepOutput
-from .task import BaseTask, ConversationMessage, TokenizedConversationOutput
+# from .task import BaseTask, ConversationMessage, TokenizedConversationOutput
+from .task import BaseTask, ConversationMessage
 from .utils import Evaluator