Add distributed context in pytorch engine to support torchrun

[grimoire]

Thanks for your contribution and we appreciate it a lot. The following instructions would make your pull request more healthy and more easily receiving feedbacks. If you do not understand some items, don't worry, just make the pull request and seek help from maintainers.

Motivation

Please describe the motivation of this PR and the goal you want to achieve through this PR.

Modification

Please briefly describe what modification is made in this PR.

BC-breaking (Optional)

Does the modification introduce changes that break the backward-compatibility of the downstream repositories?
If so, please describe how it breaks the compatibility and how the downstream projects should modify their code to keep compatibility with this PR.

Use cases (Optional)

If this PR introduces a new feature, it is better to list some use cases here, and update the documentation.

Checklist

1. Pre-commit or other linting tools are used to fix the potential lint issues.
2. The modification is covered by complete unit tests. If not, please add more unit tests to ensure the correctness.
3. If the modification has a dependency on downstream projects of a newer version, this PR should be tested with all supported versions of downstream projects.
4. The documentation has been modified accordingly, like docstring or example tutorials.

[RunningLeon]

Can you provide a short script to show how to use torchrun with lmdeploy to test this pr?

[grimoire]

import os
import time
import argparse

import torch
import torch.distributed as dist
import torch.multiprocessing as mp

from tqdm import tqdm
from lmdeploy import pipeline, PytorchEngineConfig, GenerationConfig, ChatTemplateConfig, VisionConfig
from lmdeploy.vl import load_image as load_image

os.environ['TOKENIZERS_PARALLELISM'] = 'true'

def init_dist_pytorch(tcp_port, local_rank, backend='nccl'):
    if mp.get_start_method(allow_none=True) is None:
        mp.set_start_method('spawn')

    num_gpus = torch.cuda.device_count()
    if torch.__version__ > '1.10':
        local_rank = int(os.environ['LOCAL_RANK'])
    torch.cuda.set_device(local_rank % num_gpus)

    dist.init_process_group(
        backend=backend,
    )
    rank = dist.get_rank()
    num_gpus = dist.get_world_size()
    return num_gpus, rank

def parse_config():
    parser = argparse.ArgumentParser(description='arg parser')
    parser.add_argument('--model_path', type=str, default=None, help='checkpoint to start from')
    parser.add_argument('--tcp_port', type=int, default=18888, help='tcp port for distrbuted training')
    parser.add_argument('--local_rank', type=int, default=0, help='local rank for distributed training')
    args = parser.parse_args()
    return args

if __name__ == '__main__':
    args = parse_config()
    num_gpus, rank = init_dist_pytorch(args.tcp_port, args.local_rank)

    pipe = pipeline(model_path=args.model_path, 
                    backend_config=PytorchEngineConfig(dtype='bfloat16', cache_max_entry_count=0.1,
                                                       max_batch_size=1),
                    vision_config=VisionConfig(max_batch_size=1), log_level='INFO',
                    # chat_template_config=ChatTemplateConfig(model_name='internvl2-internlm2')
                    )

    generation_config = GenerationConfig(max_new_tokens=4096, do_sample=False, temperature=0.0)

    iteration_num = 20

    input_text = "Explain the concept of artificial intelligence in simple terms."
    if num_gpus == 1:
        start_time = time.time()

        for _ in tqdm(range(iteration_num), ncols=140, desc=f"Single GPU"):
            output = pipe([input_text], gen_config=generation_config)

        print(f"Single GPU average inference time: {time.time()-start_time:.1f} seconds")

    else:
        dist.barrier()

        start_time = time.time()
        for _ in tqdm(range(iteration_num//num_gpus), ncols=140, desc=f"Multi GPU", disable=rank!=0):
            output = pipe([input_text], gen_config=generation_config)

        dist.barrier()
        if rank == 0:
            print(f"Multi-GPU average inference time: {time.time()-start_time:.1f} seconds")

torchrun --nproc_per_node=2 test.py \
    --model_path InternVL2-1B

[RunningLeon]

LGTM