123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225 |
- # Copyright (c) Microsoft Corporation.
- # SPDX-License-Identifier: Apache-2.0
- # DeepSpeed Team
- import torch
- from .abstract_accelerator import DeepSpeedAccelerator
- # During setup stage torch may not be installed, pass on no torch will
- # allow op builder related API to be executed.
- try:
- import torch.mps
- except ImportError:
- pass
- class MPS_Accelerator(DeepSpeedAccelerator):
- def __init__(self):
- self._name = "mps"
- self._communication_backend_name = None
- def is_synchronized_device(self):
- return False
- # Device APIs
- def device_name(self, device_index=None):
- if device_index == None:
- return "mps"
- return "mps:{}".format(device_index)
- def device(self, device_index):
- return torch.device("mps", index=0)
- def set_device(self, device_index):
- return
- def current_device(self):
- return torch.device("mps", index=0)
- def current_device_name(self):
- return "mps:0"
- def device_count(self):
- return 1
- def synchronize(self, device_index=None):
- return torch.mps.synchronize()
- # RNG APIs
- def random(self):
- return torch.random
- def set_rng_state(self, new_state, device_index=None):
- return torch.mps.set_rng_state(new_state)
- def get_rng_state(self, device_index=None):
- return torch.mps.get_rng_state()
- def manual_seed(self, seed):
- return torch.mps.manual_seed(seed)
- def manual_seed_all(self, seed):
- return torch.mps.manual_seed(seed)
- def seed(self):
- return torch.mps.seed()
- def initial_seed(self, seed):
- return
- def default_generator(self, device_index):
- return
- # Streams/Events
- @property
- def Stream(self):
- return None
- def stream(self, stream):
- return None
- def current_stream(self, device_index=None):
- return None
- def default_stream(self, device_index=None):
- return None
- @property
- def Event(self):
- return None
- # Memory management
- def empty_cache(self):
- return torch.mps.empty_cache()
- def memory_allocated(self, device_index=None):
- return torch.mps.current_allocated_memory()
- def max_memory_allocated(self, device_index=None):
- return torch.mps.driver_allocated_memory()
- def set_per_process_memory_fraction(self, fraction):
- return torch.mps.set_per_process_memory_fraction(fraction)
- def reset_max_memory_allocated(self, device_index=None):
- return
- def memory_cached(self, device_index=None):
- return
- def max_memory_cached(self, device_index=None):
- return
- def reset_max_memory_cached(self, device_index=None):
- return
- def memory_stats(self, device_index=None):
- return
- def reset_peak_memory_stats(self, device_index=None):
- return
- def memory_reserved(self, device_index=None):
- return
- def max_memory_reserved(self, device_index=None):
- return
- def total_memory(self, device_index=None):
- return
- # Data types
- def is_bf16_supported(self):
- return False
- def is_fp16_supported(self):
- return False
- # Misc
- def amp(self):
- return
- def is_available(self):
- return hasattr(torch.backends, "mps") and torch.backends.mps.is_available()
- def range_push(self, msg):
- return
- def range_pop(self):
- return
- def lazy_call(self, callback):
- return
- def communication_backend_name(self):
- return self._communication_backend_name
- # Tensor operations
- @property
- def BFloat16Tensor(self):
- return
- @property
- def ByteTensor(self):
- return
- @property
- def DoubleTensor(self):
- return
- @property
- def FloatTensor(self):
- return
- @property
- def HalfTensor(self):
- return
- @property
- def IntTensor(self):
- return
- @property
- def LongTensor(self):
- return
- def pin_memory(self, tensor):
- return tensor.pin_memory()
- def on_accelerator(self, tensor):
- device_str = str(tensor.device)
- if device_str.startswith("mps"):
- return True
- else:
- return False
- def op_builder_dir(self):
- try:
- # is op_builder from deepspeed or a 3p version? this should only succeed if it's deepspeed
- # if successful this also means we're doing a local install and not JIT compile path
- from op_builder import __deepspeed__ # noqa: F401
- return "op_builder"
- except ImportError:
- return "deepspeed.ops.op_builder"
- # create an instance of op builder, specified by class_name
- def create_op_builder(self, op_name):
- builder_class = self.get_op_builder(op_name)
- if builder_class != None:
- return builder_class()
- return None
- # return an op builder class, specified by class_name
- def get_op_builder(self, class_name):
- from deepspeed.ops.op_builder.cpu import NotImplementedBuilder
- return NotImplementedBuilder
- def build_extension(self):
- from torch.utils.cpp_extension import BuildExtension
- return BuildExtension
|