inference revise

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)

# Server configuration

server:

  base_url: "http://localhost:5000"  # Environment server URL

  timeout: 600  # Request timeout (seconds)

  max_workers: 10  # Maximum number of concurrent worker threads

batch_size: 32

max_steps: 10

num_workers: 4

# Inference settings

inference:

  max_steps: 10  # Maximum steps per environment

  show_progress: true  # Display progress bar

  debug: false  # Debug mode

  save_images: true  # Save image results

  save_intermediate: false  # Save intermediate results

dataset_dir: data

use_split: test

# Generation parameters (overrides defaults in model config)

generation:

  temperature: 0.7  # Temperature parameter

  top_p: 0.95  # Nucleus sampling parameter

  max_tokens: 512  # Maximum number of tokens to generate

  stop: ["\n\nUser:", "\n\nHuman:"]  # Stop generation at these markers

server_url: http://localhost:5000

server_timeout: 600

server_max_workers: 48

# Batch processing settings

batch:

  batch_size: 4  # Batch size

  batch_size_multiple: 4  # Batch size multiple (for hardware optimization)

output_dir: inference_results

show_progress: true

# Evaluation configuration

evaluation:

  metrics:

    - score  # Overall score

    - done  # Completion rate

    - steps  # Step count statistics

# Weights & Biases logging configuration

wandb:

  project: "qwen-vl-eval"  # Project name

  experiment_name: "qwen2.5-vl-3b-eval"  # Experiment name

  entity: null  # Team/entity name (optional)

  val_generations_to_log_to_wandb: 5  # Number of generated samples to log

  log_model: false  # Whether to save the model to wandb

# Output configuration

output:

  format: "json"  # Output format

  include_metrics: true  # Include evaluation metrics

  include_trajectories: true  # Include full trajectories

  pretty_print: true  # Pretty-print the output

use_wandb: true

wandb_project: vagen-prompting-frozenlake

debug: false

 No newline at end of file

type: "vllm"

name: "Qwen2.5-0.5B-Instruct"

# Model path

path: "Qwen/Qwen2.5-0.5B-Instruct"

tokenizer_path: null  # Use the same path as the model

processor_path: null  # Use the same path as the model

# vLLM specific configuration

tensor_parallel_size: 1  # Adjust according to the number of available GPUs

dtype: "bfloat16"  # Options: auto, float16, bfloat16, float32

gpu_memory_utilization: 0.9

trust_remote_code: true

max_model_len: 4096  # Maximum sequence length

# Generation parameters

qwen_0.5b:

  provider: vllm

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

  max_tokens: 1024

  temperature: 0.7

top_p: 0.95

  top_p: 0.9

  top_k: 50

max_tokens: 512

repetition_penalty: 1.0

presence_penalty: 0.0

frequency_penalty: 0.0

  tensor_parallel_size: 1

  gpu_memory_utilization: 0.9

  dtype: bfloat16

  trust_remote_code: true

# Multimodal specific configuration

is_multimodal: true

limit_mm_per_prompt:

  image: 4  # Maximum number of images per prompt

# qwen_vl_3b:

#   provider: vllm

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

#   max_tokens: 1024

#   temperature: 0.7

#   top_p: 0.9

#   top_k: 50

#   tensor_parallel_size: 1

#   gpu_memory_utilization: 0.9

#   dtype: bfloat16

#   trust_remote_code: true

 No newline at end of file

    --force_gen

python -m vagen.inference.run_inference \

  --model_config "$SCRIPT_DIR/model_config.yaml" \

  --inference_config "$SCRIPT_DIR/inference_config.yaml" \

  --dataset "data/$EXPERIMENT_NAME/test.parquet" \

  --output_dir results

 No newline at end of file

    --inference_config_path="$SCRIPT_DIR/inference_config.yaml" \

    --model_config_path="$SCRIPT_DIR/model_config.yaml" \

    --val_files_path="data/$EXPERIMENT_NAME/test.parquet"