test_custom_allreduce.py

# SPDX-License-Identifier: MIT

import argparse
import logging
import os
from multiprocessing import Pool, freeze_support, set_start_method
from typing_extensions import Optional

import torch
import torch.distributed as dist
import pandas as pd

from aiter import dtypes
from aiter.dist.communication_op import tensor_model_parallel_all_reduce
from aiter.dist.parallel_state import (
    destroy_distributed_environment,
    destroy_model_parallel,
    ensure_model_parallel_initialized,
    get_tp_group,
    graph_capture,
    init_distributed_environment,
    set_custom_all_reduce,
)
from aiter.dist.utils import get_distributed_init_method, get_ip, get_open_port
from aiter.test_common import benchmark, checkAllclose, perftest

logger = logging.getLogger("aiter")

set_start_method("spawn", force=True)


def allreduce_custom(
    tp_size,
    pp_size,
    rankID,
    x,
    withGraph=False,
    distributed_init_method: Optional[str] = None,
    enable_register_for_capturing: bool = True,
):
    device = torch.device(f"cuda:{rankID}")
    torch.cuda.set_device(device)
    # init
    # Forward the user-requested capturing-registration policy down to the
    # CustomAllreduce constructor via the env var consumed inside
    # CudaCommunicator. Must be set BEFORE init_distributed_environment so
    # the worker process picks it up when the communicator is built.
    os.environ["AITER_AR_ENABLE_REG_CAPTURE"] = (
        "1" if enable_register_for_capturing else "0"
    )
    logger.info(f"RANK: {rankID} {tp_size} init_process_group...")
    set_custom_all_reduce(True)
    init_distributed_environment(
        world_size=tp_size,
        rank=rankID,
        distributed_init_method=distributed_init_method,
    )
    ensure_model_parallel_initialized(tp_size, pp_size)
    x = x.to(device)
    # dist.barrier(device_ids=[i for i in range(tp_size)])

    # warmup and align all gpu
    group = get_tp_group().device_group
    dist.all_reduce(torch.zeros(1).cuda(), group=group)
    torch.cuda.synchronize()

    if withGraph:
        graph = torch.cuda.CUDAGraph()
        with graph_capture() as gc:
            with torch.cuda.graph(graph, stream=gc.stream):
                out = tensor_model_parallel_all_reduce(x)
        out.fill_(0)

        @perftest()
        def run_ca():
            graph.replay()

        _, us = run_ca()
        out = (out, us)
    else:

        @perftest()
        def run_ca(x):
            return tensor_model_parallel_all_reduce(x)

        out = run_ca(x)

    # destroy
    if dist.is_initialized():
        destroy_model_parallel()
        destroy_distributed_environment()
        torch.cuda.empty_cache()
    return out


@benchmark()
def test_allreduce_custom(
    tp_size,
    pp_size,
    shape,
    dtype,
    withGraph=False,
    distributed_init_method: Optional[str] = None,
    enable_register_for_capturing: bool = True,
):
    os.environ["MASTER_ADDR"] = "127.0.0.1"
    os.environ["MASTER_PORT"] = "49373"
    pool = Pool(processes=tp_size)
    ref = torch.zeros(shape, dtype=dtype)
    rets = []
    for i in range(tp_size):
        x = torch.randn(shape, dtype=dtype)
        ref += x
        rets.append(
            pool.apply_async(
                allreduce_custom,
                args=(
                    tp_size,
                    pp_size,
                    i,
                    x,
                    withGraph,
                    distributed_init_method,
                    enable_register_for_capturing,
                ),
            )
        )
    pool.close()
    pool.join()
    rets = [el.get() for el in rets]
    all_us = [us for _, us in rets]
    max_err = 0.0
    for out, us in rets:
        msg = (
            f"test_allreduce_custom: {shape=} {dtype=} "
            f"{withGraph=} reg_cap={enable_register_for_capturing} {us:>8.2f}"
        )
        err = checkAllclose(ref, out.to(ref), msg=msg)
        max_err = max(max_err, err)
    return {
        "min_us": min(all_us),
        "max_us": max(all_us),
        "err": max_err,
    }


l_dtype = ["fp16", "bf16"]
l_shape = [(2, 7168), (128, 8192)]

parser = argparse.ArgumentParser(description="config input of test")
parser.add_argument(
    "-d",
    "--dtype",
    type=str,
    choices=l_dtype,
    nargs="?",
    const=None,
    default=None,
    help="data type",
)
parser.add_argument(
    "-s",
    "--shape",
    type=dtypes.str2tuple,
    nargs="?",
    const=None,
    default=None,
    help="shape. e.g. -s 128,8192",
)
parser.add_argument(
    "-g",
    "--with-graph",
    type=lambda x: str(x).lower() in ["true", "1", "yes"],
    default=True,
    help="use CUDA graph (default: True). e.g. -g true or -g false",
)
parser.add_argument(
    "--reg-capturing",
    type=str,
    choices=["true", "false", "both"],
    default="both",
    help=(
        "whether CustomAllreduce.enable_register_for_capturing is True/False. "
        "'both' (default) exercises both paths as a regression sweep."
    ),
)


if __name__ == "__main__":
    freeze_support()
    args = parser.parse_args()
    if args.dtype is None:
        l_dtype = [dtypes.d_dtypes[key] for key in l_dtype]
    else:
        l_dtype = [dtypes.d_dtypes[args.dtype]]
    if args.shape is not None:
        l_shape = [args.shape]
    if args.reg_capturing == "true":
        l_reg = [True]
    elif args.reg_capturing == "false":
        l_reg = [False]
    else:
        l_reg = [True, False]

    df = []
    for dtype in l_dtype:
        for shape in l_shape:
            for reg in l_reg:
                ret = test_allreduce_custom(
                    8,
                    1,
                    shape,
                    dtype,
                    withGraph=args.with_graph,
                    distributed_init_method=get_distributed_init_method(
                        get_ip(), get_open_port()
                    ),
                    enable_register_for_capturing=reg,
                )
                df.append(ret)
    df = pd.DataFrame(df)
    show_cols = [
        "tp_size",
        "shape",
        "dtype",
        "withGraph",
        "enable_register_for_capturing",
        "min_us",
        "max_us",
        "err",
    ]
    show_cols = [c for c in show_cols if c in df.columns]
    df[show_cols].to_csv("test_custom_allreduce.csv", index=False)
    logger.info(
        "custom allreduce summary (markdown):\n%s",
        df[show_cols].to_markdown(index=False),
    )