Merge pull request #22 from RAGEN-AI/prompt_dev

temp/

benchmark_results/

env_benchmark_results/

results/

 No newline at end of file

- New service architecture for efficient distributed training

- Check out our new guides:

  - [Creating Environments](./docs/envs/create-env.md): New environment protocal.

  - [Creating Services](./docs/envs/create-service.md): We now support hosting environments in a seperate process

  - [Creating Services](./docs/envs/create-service.md): We now support hosting environments in a separate process

## Installation

## How to Add New Model

1. Refer to [VERL](https://verl.readthedocs.io/en/latest/index.html) for adding new MLLM.

2. Refer to [QwenVLRolloutManager](./vagen/mllm_agent/rollout.py) to understand how rollout works. In most cases, you can use QwenVLRolloutManager directly with only minor modifications to the model's special tokens

2. Refer to [QwenVLRolloutManager](vagen/mllm_agent/training_rollout/rollout.py) to understand how rollout works. In most cases, you can use QwenVLRolloutManager directly with only minor modifications to the model's special tokens

## Experimental Results

> To reproduce our experiment, please refer to document: [Reproduce Experiments](docs/reproduce-exp.md)

env1:

  env_name: frozenlake

  env_config:

    render_mode: vision

    prompt_format: free_think

    use_accuracy_reward: false

  train_size: 10000

  test_size: 2

 No newline at end of file

# Server configuration

server_url: "http://localhost:5000"

server_timeout: 600

server_max_workers: 48

# Inference parameters

batch_size: 32

max_steps: 3

split: "test"

debug: false

# Output configuration

output_dir: "inference_outputs"

# WandB configuration

use_wandb: true

wandb_project: "vagen-inference"

# Display settings

show_progress: true

val_generations_to_log_to_wandb: 10

 No newline at end of file

models:

  # qwen_0.5b:

    # provider: vllm

    # model_name: Qwen/Qwen2.5-0.5B-Instruct

    # max_tokens: 1024

    # temperature: 0.7

    # tensor_parallel_size: 1

    # gpu_memory_utilization: 0.9

  # qwen_vl_3b:

  #   provider: vllm

  #   model_name: Qwen/Qwen2.5-VL-3B-Instruct

  #   max_tokens: 1024

  #   temperature: 0.7

  #   tensor_parallel_size: 2

  #   gpu_memory_utilization: 0.9

  # gpt4o:

  #   provider: openai

  #   model_name: gpt-4o

  #   max_tokens: 1024

  #   temperature: 0.7

  #   presence_penalty: 0.0

  #   frequency_penalty: 0.0

  #   max_retries: 3

  #   timeout: 60

  # gpt4_vision:

  #   provider: openai

  #   model_name: gpt-4-vision-preview

  #   max_tokens: 1024

  #   temperature: 0.7

  #   presence_penalty: 0.0

  #   frequency_penalty: 0.0

  #   max_retries: 3

  #   timeout: 60

  claude_3_haiku:

    provider: claude

    model_name: claude-3-haiku-20240307

    max_tokens: 1024

    temperature: 0.7

  # # Batch processing configuration example

  # claude_3_haiku_batch:

  #   provider: claude

  #   model_name: claude-3-haiku-20240307

  #   max_tokens: 1024

  #   temperature: 0.7

  #   use_batch_api: true  # Use batch API for cost optimization

  #   batch_poll_interval: 10  # Poll every 10 seconds

  #   batch_max_wait_time: 7200  # Wait up to 2 hours

 No newline at end of file