HexGen: Generative Inference of Foundation Model over Heterogeneous Decentralized Environment [paper]

HexGen presents a versatile framework capable of facilitating Llama-2 inference, integrating hybrid model parallelism along with an automated mechanism for request dispatch. Key features include:

• Comprehensive support for hybrid pipeline parallelism as well as tensor parallelism.
• Ocf, a seamlessly integrated subsystem, is dedicated to coordinating and efficiently dispatching requests.

Content

• Building Environment
  • Establish A Personal Head Node Coordinator
  • Incorporate Additional Worker Nodes
• Loading Model Parameters for LlaMA Models
  • Create Separate Model State Dicts
  • Load Model Parameters
• Starting HexGen
  • Activating Head Node Coordinator
  • Activating Worker Nodes
  • Activating Independent Inference Process
• Asymmetric Parallel Group Support in HexGen
  • Tensor Model Parallelism and Pipeline Parallelism
  • Asymmetric Parallel Group Support
• Performance Results
• Acknowledgements

Building Environment

HexGen stipulates the utilization of CUDA version 11.8 and Python version 3.11 or above. The assembly of HexGen is designed to be efficient and accessible:

Only Establish A Personal Head Node Coordinator

make hexgen-head

Incorporate Additional Worker Nodes

make hexgen

Loading Model Parameters for LlaMA Models

Navigate to the hexgen/llama/load_model_parameters_utils directory. Here, you will initiate the process of setting up parameters for the model.

Create Separate Model State Dicts

For scenarios where specific custom paths are required, modifications to the create_separate_state_dicts_llama_7b.py script are necessary. In this script, locate the function call to save_model_components. You can then alter the paths according to your specific requirements. For instance:

save_model_components(
    config_path='../llama-config/',
    checkpoint_name='llama-7b',
    checkpoint_path='/path/to/Llama-2-7b-chat-hf/',
    num_layers=32,
    save_dir='./separate_state_dicts/'
)

Here, your sole requirement is to specify the checkpoint_path, as the other parameters have been pre-defined and supplied for your convenience. You can download the model checkpoints from here.

A recommended way to download is:

huggingface-cli download --resume-download meta-llama/Llama-2-7b-chat-hf --local-dir Llama-2-7b-chat-hf --token <your token>

To create the separate state dictionaries for the Llama-7b model, run the following command in the terminal:

python3 create_separate_state_dicts_llama_7b.py

This script will automatically generate and save the state dictionaries in the appropriate directory.

Load Model Parameters

In the llama_inference.py file, add the following code snippet to load the parameters for Llama-7b. Adjust the paths as per your setup:

# Load model checkpoints with respect to hetero_config
tp_ranks_whole_model = hetero_groups['tp_ranks_whole_model']
tp_group_list = hetero_groups['tp_rank_groups']
state_dicts_path = "./load_model_parameters_utils/"
load_model_parameters(model, config, state_dicts_path, tp_ranks_whole_model, tp_group_list, rank)

Starting HexGen

Activating Head Node Coordinator

HexGen can be launched in head node coordinator modes by:

bash scripts/run_head.sh

Activating Worker Nodes

HexGen can be launched in worker modes by a similar command, except that you should modify the file ./third_party/ocf/ocf-core/config/cfg.yaml, the p2p addr should be as similar format as "/ip4/{Pubilc_IP}/tcp/43905/p2p/{Peer_ID}", you could replace {Public_IP} as your own head coordinator's IP address and {Peer_ID} as its peer ID:

bash scripts/run_worker.sh

Activating Independent Inference Process

To initiate an independent inference process without involving the coordinator, navigate to the hexgen/llama directory and execute the following command:

bash scripts/run_llama_inference.sh

You have the flexibility to customize various inputs to tailor your inference task according to your specific requirements. The model_msg object can be adjusted with different parameters, as shown in the example below:

model_msg = {
    'prompt': "Do you like yourself ?",  # Define your own prompt here
    'max_new_tokens': 128,               # Set the maximum number of new tokens
    'temperature': 0.2,                  # Adjust the randomness in response generation
    'top_k': 20,                         # Specify the number of highest probability vocabulary tokens to keep for top-k sampling
    'top_p': 0.9,                        # Set the cumulative probability threshold for top-p (nucleus) sampling
}

Asymmetric Parallel Group Support in HexGen

Tensor Model Parallelism and Pipeline Parallelism

Two methods are used to distribute the workload of training large deep learning models across multiple computing units.

• Tensor Model Parallelism splits the model's layers or components across different processors.
• Pipeline Parallelism divides the process into different stages, with each stage being processed on a different set of processors.

Asymmetric Parallel Group Support

HexGen introduces a novel approach with its Asymmetric Parallel Group Support, driven by two critical parameters: --hetero_config and --pp_partition.

• --hetero_config: This parameter allows for the specification of varying TP degrees for each pipeline stage. For instance, a setting like 4 2 2 configures a three-stage pipeline with respective TP degrees of 4, 2, and 2, showing HexGen's adaptability.
• --pp_partition: This parameter complements --hetero_config by managing the distribution of model layers across the pipeline stages. A combination like 40 20 20 with a hetero_config of 4 2 2 signifies an optimized layer distribution, illustrating HexGen's capability for precision tuning according to model needs and hardware constraints.

HexGen can be launched with asymmetric parallel group by:

python3 -m torch.distributed.launch --nproc_per_node=4 --master_port 9996 llama_inference.py \
--model_size llama-7b \
--hetero_config 1 2 1 \
--pp_partition 8 16 8 \

Performance Results

The figure below shows a comprehensive comparison of cost-performance trade-offs among different configurations and implementations, focusing on Service Level Objective (SLO) attainment in various settings:

• HexGen with Asymmetric Parallel Group Support (Full Budget, Heterogeneous Setting): This configuration can achieve up to 2.3× lower latency and can handle peak request rates up to 4× higher than FlashAttention in a homogeneous setting, demonstrating significant performance enhancement.
• HexGen (Half Budget, Heterogeneous Setting): Even with a halved budget, HexGen can still slightly outperform FlashAttention in a homogeneous environment, showcasing its ability to efficiently utilize heterogeneous GPUs.
• Asymmetric vs. Symmetric Parallelism in HexGen (Full Budget, Heterogeneous Setting): The integration of asymmetric parallelism into HexGen can lead to up to 1.8× improvement in meeting lower latency deadlines and can manage peak traffic rates up to 2× higher than its symmetric parallelism counterpart.

For further information regarding the experimental methodology and technical specifications, please refer to the detailed discussion presented in our paper.

Acknowledgements

This initiative is led by the Relaxed System Lab team at HKUST. We are committed to ongoing development and support for this project. For referencing purposes, please cite HexGen as follows:

@misc{jiang2023hexgen,
      title={HexGen: Generative Inference of Foundation Model over Heterogeneous Decentralized Environment}, 
      author={Youhe Jiang and Ran Yan and Xiaozhe Yao and Beidi Chen and Binhang Yuan},
      year={2023},
      eprint={2311.11514},
      archivePrefix={arXiv},
      primaryClass={cs.DC}
}