import get_accelerator from deepspeed.accelerator directly

benchmarks/communication/all_gather.py

@@ -1,6 +1,6 @@
 from benchmarks.communication.utils import *
 from benchmarks.communication.constants import *
-from deepspeed.accelerator.real_accelerator import get_accelerator
+from deepspeed.accelerator import get_accelerator
 
 import time
 

benchmarks/communication/all_reduce.py

@@ -1,6 +1,6 @@
 from benchmarks.communication.utils import *
 from benchmarks.communication.constants import *
-from deepspeed.accelerator.real_accelerator import get_accelerator
+from deepspeed.accelerator import get_accelerator
 
 import time
 

benchmarks/communication/all_to_all.py

@@ -1,6 +1,6 @@
 from benchmarks.communication.utils import *
 from benchmarks.communication.constants import *
-from deepspeed.accelerator.real_accelerator import get_accelerator
+from deepspeed.accelerator import get_accelerator
 
 import time
 

benchmarks/communication/broadcast.py

@@ -1,7 +1,7 @@
 import torch
 from benchmarks.communication.utils import *
 from benchmarks.communication.constants import *
-from deepspeed.accelerator.real_accelerator import get_accelerator
+from deepspeed.accelerator import get_accelerator
 
 import time
 

benchmarks/communication/constants.py

@@ -1,4 +1,4 @@
-from deepspeed.accelerator.real_accelerator import get_accelerator
+from deepspeed.accelerator import get_accelerator
 
 DEFAULT_WARMUPS = 5
 DEFAULT_TRIALS = 50

benchmarks/communication/pt2pt.py

@@ -1,6 +1,6 @@
 from benchmarks.communication.utils import *
 from benchmarks.communication.constants import *
-from deepspeed.accelerator.real_accelerator import get_accelerator
+from deepspeed.accelerator import get_accelerator
 
 import time
 

benchmarks/communication/utils.py

@@ -3,7 +3,7 @@
 import math
 import argparse
 from benchmarks.communication.constants import *
-from deepspeed.accelerator.real_accelerator import get_accelerator
+from deepspeed.accelerator import get_accelerator
 
 global dist
 

benchmarks/inference/bert-bench.py

@@ -3,7 +3,7 @@
 import deepspeed
 import argparse
 from transformers import pipeline
-from deepspeed.accelerator.real_accelerator import get_accelerator
+from deepspeed.accelerator import get_accelerator
 
 parser = argparse.ArgumentParser()
 parser.add_argument("--model", "-m", type=str, help="hf model name")

benchmarks/inference/gpt-bench.py

@@ -4,7 +4,7 @@
 import deepspeed
 import argparse
 from transformers import pipeline
-from deepspeed.accelerator.real_accelerator import get_accelerator
+from deepspeed.accelerator import get_accelerator
 
 parser = argparse.ArgumentParser()
 parser.add_argument("--model", "-m", type=str, help="hf model name")

csrc/aio/py_test/aio_bench_perf_sweep.py

@@ -15,7 +15,7 @@
 from test_ds_aio_utils import refine_integer_value
 from perf_sweep_utils import READ_OP_DESC, WRITE_OP_DESC, BENCH_LOG_DIR, \
     READ_IO_DIR, WRITE_IO_DIR, READ_LOG_DIR, WRITE_LOG_DIR
-from deepspeed.accelerator.real_accelerator import get_accelerator
+from deepspeed.accelerator import get_accelerator
 
 OTHER_OPTIONS = '--handle'
 PERF_SCRIPT = 'test_ds_aio.py'

csrc/aio/py_test/ds_aio_basic.py

@@ -10,7 +10,7 @@
 import time
 from multiprocessing import Pool, Barrier
 from test_ds_aio_utils import report_results, task_log, task_barrier
-from deepspeed.accelerator.real_accelerator import get_accelerator
+from deepspeed.accelerator import get_accelerator
 
 
 def pre_basic(args, tid, read_op):

csrc/aio/py_test/ds_aio_handle.py

@@ -10,7 +10,7 @@
 import time
 from multiprocessing import Pool, Barrier
 from test_ds_aio_utils import report_results, task_log, task_barrier
-from deepspeed.accelerator.real_accelerator import get_accelerator
+from deepspeed.accelerator import get_accelerator
 
 
 def pre_handle(args, tid, read_op):

csrc/aio/py_test/validate_async_io.py

@@ -4,5 +4,5 @@
 
 Functionality of swapping optimizer tensors to/from (NVMe) storage devices.
 """
-from deepspeed.accelerator.real_accelerator import get_accelerator
+from deepspeed.accelerator import get_accelerator
 assert get_accelerator().create_op_builder("AsyncIOBuilder").is_compatible()

deepspeed/accelerator/real_accelerator.py

@@ -38,7 +38,7 @@ def set_accelerator(accel_obj):
 
 '''
 -----------[code] test_get.py -----------
-from deepspeed.accelerator.real_accelerator import get_accelerator
+from deepspeed.accelerator import get_accelerator
 my_accelerator = get_accelerator()
 print(f'{my_accelerator._name=}')
 print(f'{my_accelerator._communication_backend=}')
@@ -58,7 +58,7 @@ def set_accelerator(accel_obj):
 from deepspeed.accelerator.cuda_accelerator import CUDA_Accelerator
 cu_accel = CUDA_Accelerator()
 print(f'{id(cu_accel)=}')
-from deepspeed.accelerator.real_accelerator import set_accelerator, get_accelerator
+from deepspeed.accelerator import set_accelerator, get_accelerator
 set_accelerator(cu_accel)
 
 my_accelerator = get_accelerator()

deepspeed/autotuning/autotuner.py

@@ -15,7 +15,7 @@
 from .scheduler import ResourceManager
 from .tuner import GridSearchTuner, RandomTuner, ModelBasedTuner
 from .utils import *
-from deepspeed.accelerator.real_accelerator import get_accelerator
+from deepspeed.accelerator import get_accelerator
 
 try:
     from tabulate import tabulate

deepspeed/comm/comm.py

@@ -31,7 +31,7 @@
 
 from ..constants import TORCH_DISTRIBUTED_DEFAULT_PORT, default_pg_timeout
 from .constants import *
-from deepspeed.accelerator.real_accelerator import get_accelerator
+from deepspeed.accelerator import get_accelerator
 
 
 class ReduceOp(Enum):

deepspeed/env_report.py

@@ -4,7 +4,7 @@
 import argparse
 from .ops.op_builder import ALL_OPS
 from .git_version_info import installed_ops, torch_info
-from deepspeed.accelerator.real_accelerator import get_accelerator
+from deepspeed.accelerator import get_accelerator
 
 GREEN = '\033[92m'
 RED = '\033[91m'

deepspeed/inference/engine.py

@@ -20,7 +20,7 @@
 from ..moe.utils import has_moe_layers
 from ..runtime.zero import GatheredParameters
 from ..module_inject import LinearAllreduce, LinearLayer, Normalize, ReplaceWithTensorSlicing
-from deepspeed.accelerator.real_accelerator import get_accelerator
+from deepspeed.accelerator import get_accelerator
 from ..module_inject.replace_policy import DSPolicy
 
 DS_INFERENCE_ENABLED = False

deepspeed/launcher/multinode_runner.py

@@ -5,7 +5,7 @@
 import warnings
 from shlex import split
 from abc import ABC, abstractmethod
-from deepspeed.accelerator.real_accelerator import get_accelerator
+from deepspeed.accelerator import get_accelerator
 from ..utils import logger
 from .constants import PDSH_MAX_FAN_OUT, MVAPICH_TMP_HOSTFILE
 

deepspeed/launcher/runner.py

@@ -24,7 +24,7 @@
 from ..utils import logger
 
 from ..autotuning import Autotuner
-from deepspeed.accelerator.real_accelerator import get_accelerator
+from deepspeed.accelerator import get_accelerator
 
 DLTS_HOSTFILE = "/job/hostfile"
 EXPORT_ENVS = ['NCCL', 'PYTHON', 'MV2', 'UCX']

deepspeed/module_inject/encoder.py

@@ -2,7 +2,7 @@
 Copyright 2022 The Microsoft DeepSpeed Team
 '''
 import torch
-from deepspeed.accelerator.real_accelerator import get_accelerator
+from deepspeed.accelerator import get_accelerator
 
 
 class DSClipEncoder(torch.nn.Module):

deepspeed/module_inject/layers.py

@@ -4,7 +4,7 @@
 from torch.nn import functional as F
 
 from torch.nn.parameter import Parameter
-from deepspeed.accelerator.real_accelerator import get_accelerator
+from deepspeed.accelerator import get_accelerator
 
 
 class LinearAllreduce(nn.Module):

deepspeed/module_inject/load_checkpoint.py

@@ -4,7 +4,7 @@
 from .layers import LinearLayer, Normalize, EmbeddingLayer
 import torch
 import gc
-from deepspeed.accelerator.real_accelerator import get_accelerator
+from deepspeed.accelerator import get_accelerator
 
 
 def load_model_with_checkpoint(r_module,

deepspeed/module_inject/replace_module.py

@@ -3,7 +3,7 @@
 import tqdm
 import deepspeed
 import deepspeed.ops.transformer as transformer_inference
-from deepspeed.accelerator.real_accelerator import get_accelerator
+from deepspeed.accelerator import get_accelerator
 from .replace_policy import HFBertLayerPolicy, HFGPT2LayerPolicy, BLOOMLayerPolicy
 from .replace_policy import replace_policies, generic_policies
 #from ..runtime.weight_quantizer import WeightQuantization

deepspeed/ops/adagrad/cpu_adagrad.py

@@ -3,7 +3,7 @@
 '''
 
 import torch
-from deepspeed.accelerator.real_accelerator import get_accelerator
+from deepspeed.accelerator import get_accelerator
 from deepspeed.utils.logging import should_log_le
 
 

deepspeed/ops/adam/fused_adam.py

@@ -9,7 +9,7 @@
 from .multi_tensor_apply import MultiTensorApply
 
 multi_tensor_applier = MultiTensorApply(2048 * 32)
-from deepspeed.accelerator.real_accelerator import get_accelerator
+from deepspeed.accelerator import get_accelerator
 
 
 class FusedAdam(torch.optim.Optimizer):

deepspeed/ops/lamb/fused_lamb.py

@@ -6,7 +6,7 @@
 '''
 import types
 import torch
-from deepspeed.accelerator.real_accelerator import get_accelerator
+from deepspeed.accelerator import get_accelerator
 
 
 class FusedLamb(torch.optim.Optimizer):

deepspeed/ops/quantizer/quantizer.py

@@ -3,7 +3,7 @@
 '''
 import torch
 
-from deepspeed.accelerator.real_accelerator import get_accelerator
+from deepspeed.accelerator import get_accelerator
 
 # Cuda modules will be imported if needed
 quantizer_cuda_module = None

deepspeed/ops/sparse_attention/matmul.py

@@ -6,7 +6,7 @@
 import triton
 import triton.language as tl
 import triton._C.libtriton as libtriton
-from deepspeed.accelerator.real_accelerator import get_accelerator
+from deepspeed.accelerator import get_accelerator
 
 
 @triton.jit

deepspeed/ops/transformer/inference/attention.py

@@ -7,7 +7,7 @@
 import torch.nn as nn
 from packaging import version as pkg_version
 from deepspeed.utils.logging import log_dist
-from deepspeed.accelerator.real_accelerator import get_accelerator
+from deepspeed.accelerator import get_accelerator
 
 # Cuda modules will be imported if needed
 inference_cuda_module = None

deepspeed/ops/transformer/inference/moe_inference.py

@@ -13,7 +13,7 @@
 from .transformer_inference import DeepSpeedSelfAttention, DeepSpeedInferenceConfig
 from ....moe.sharded_moe import TopKGate
 from deepspeed import comm as dist
-from deepspeed.accelerator.real_accelerator import get_accelerator
+from deepspeed.accelerator import get_accelerator
 
 
 class DeepSpeedMoEInferenceConfig(DeepSpeedInferenceConfig):

deepspeed/ops/transformer/inference/transformer_inference.py

@@ -8,7 +8,7 @@
 import torch.nn as nn
 from deepspeed import comm as dist
 from deepspeed.utils.logging import log_dist
-from deepspeed.accelerator.real_accelerator import get_accelerator
+from deepspeed.accelerator import get_accelerator
 from deepspeed.utils.types import ActivationFuncType
 
 # Cuda modules will be imported if needed

deepspeed/ops/transformer/transformer.py

@@ -6,7 +6,7 @@
 import torch
 from torch import nn
 from torch.autograd import Function
-from deepspeed.accelerator.real_accelerator import get_accelerator
+from deepspeed.accelerator import get_accelerator
 
 # Cuda modules will be imported if needed
 transformer_cuda_module = None

deepspeed/profiling/flops_profiler/profiler.py

@@ -6,7 +6,7 @@
 from typing import List, Optional
 from collections import OrderedDict
 import numpy as np
-from deepspeed.accelerator.real_accelerator import get_accelerator
+from deepspeed.accelerator import get_accelerator
 
 Tensor = torch.Tensor
 

deepspeed/runtime/activation_checkpointing/checkpointing.py

@@ -25,7 +25,7 @@
 from deepspeed.utils import logger
 from deepspeed.runtime.utils import copy_to_device, move_to_device, see_memory_usage, bwc_tensor_model_parallel_rank
 from deepspeed.utils.timer import SynchronizedWallClockTimer as Timers
-from deepspeed.accelerator.real_accelerator import get_accelerator
+from deepspeed.accelerator import get_accelerator
 
 # DeepSpeed Checkpointing Enabled or Disabled
 deepspeed_checkpointing_enabled = False

deepspeed/runtime/comm/nccl.py

@@ -8,7 +8,7 @@
 import numpy as np
 
 from deepspeed.runtime.compression.cupy import CupyBackend
-from deepspeed.accelerator.real_accelerator import get_accelerator
+from deepspeed.accelerator import get_accelerator
 
 
 class NcclBackend(object):

deepspeed/runtime/dataloader.py

@@ -4,7 +4,7 @@
 
 from torch.utils.data import DataLoader, RandomSampler
 from torch.utils.data.distributed import DistributedSampler
-from deepspeed.accelerator.real_accelerator import get_accelerator
+from deepspeed.accelerator import get_accelerator
 
 
 class RepeatingLoader:

deepspeed/runtime/engine.py

@@ -77,7 +77,7 @@
 from deepspeed.profiling.flops_profiler.profiler import FlopsProfiler
 from deepspeed.utils.logging import print_json_dist
 
-from deepspeed.accelerator.real_accelerator import get_accelerator
+from deepspeed.accelerator import get_accelerator
 
 # Set to torch's distributed package or deepspeed.comm based inside DeepSpeedEngine init
 dist = None

deepspeed/runtime/fp16/fused_optimizer.py

@@ -14,7 +14,7 @@
 from deepspeed.utils import groups, logger, log_dist
 from deepspeed import comm as dist
 from deepspeed.checkpoint.constants import OPTIMIZER_STATE_DICT, CLIP_GRAD
-from deepspeed.accelerator.real_accelerator import get_accelerator
+from deepspeed.accelerator import get_accelerator
 
 
 class FP16_Optimizer(DeepSpeedOptimizer):

deepspeed/runtime/fp16/onebit/adam.py

@@ -4,7 +4,7 @@
 import types
 import torch
 import numpy as np
-from deepspeed.accelerator.real_accelerator import get_accelerator
+from deepspeed.accelerator import get_accelerator
 from deepspeed import comm as dist
 
 

deepspeed/runtime/fp16/onebit/lamb.py

@@ -6,7 +6,7 @@
 import numpy as np
 from deepspeed import comm as dist
 from torch._utils import _flatten_dense_tensors, _unflatten_dense_tensors
-from deepspeed.accelerator.real_accelerator import get_accelerator
+from deepspeed.accelerator import get_accelerator
 
 
 class OnebitLamb(torch.optim.Optimizer):

deepspeed/runtime/fp16/onebit/zoadam.py

@@ -4,7 +4,7 @@
 import types
 import torch
 import numpy as np
-from deepspeed.accelerator.real_accelerator import get_accelerator
+from deepspeed.accelerator import get_accelerator
 from deepspeed import comm as dist
 
 

deepspeed/runtime/fp16/unfused_optimizer.py

@@ -14,7 +14,7 @@
 from deepspeed.runtime.fp16.loss_scaler import INITIAL_LOSS_SCALE, SCALE_WINDOW, MIN_LOSS_SCALE
 from deepspeed.utils import logger
 from deepspeed.checkpoint.constants import OPTIMIZER_STATE_DICT
-from deepspeed.accelerator.real_accelerator import get_accelerator
+from deepspeed.accelerator import get_accelerator
 from deepspeed import comm as dist
 
 

deepspeed/runtime/pipe/engine.py

@@ -7,7 +7,7 @@
 
 from deepspeed.utils import logger
 from deepspeed.utils.timer import ThroughputTimer
-from deepspeed.accelerator.real_accelerator import get_accelerator
+from deepspeed.accelerator import get_accelerator
 
 from ..engine import DeepSpeedEngine, MEMORY_OPT_ALLREDUCE_SIZE
 from ..utils import PartitionedTensor