TDATR / TDATR_utils /utils.py
aoyama's picture
Deploy TDATR ZeroGPU Space
a27dedf verified
Raw
History Blame Contribute Delete
31 kB
import logging
import math
import os
import sys
import importlib
import torch
import torch.nn.init as init
from omegaconf import DictConfig, OmegaConf, open_dict, _utils
from dataclasses import _MISSING_TYPE, MISSING, is_dataclass
from TDATR_utils.call_main import infer_init_method
from TDATR_utils.global_context import global_context as gpc
from typing import Dict, List, Tuple, Union, Optional,Callable
import torch.distributed as dist
from torch import Tensor
logger = logging.getLogger(__name__)
def split_tensor(tensor: torch.Tensor,
num_partitions: int,
dim: int,
contiguous_split_chunks: Optional[bool]=False) -> List[torch.Tensor]:
"""Split a tensor along given dim.
Arguments:
tensor: input tensor.
num_partitions: number of partitions to split the tensor
dim: split dim
contiguous_split_chunks: If True, make each chunk contiguous
in memory.
"""
# Get the size and dimension.
dim_size = divide(tensor.size()[dim], num_partitions)
# Split.
tensor_list = torch.split(tensor, dim_size, dim=dim)
# Note: torch.split does not create contiguous tensors by default.
if contiguous_split_chunks:
return tuple(chunk.contiguous() for chunk in tensor_list)
return tensor_list
def modify_logits_for_top_p_filtering(logits, top_p):
"""Set the logits for none top-p values to -inf."""
# First sort and calculate cumulative sum of probabilities.
sorted_logits, sorted_indices = torch.sort(logits, descending=True)
cumulative_probs = sorted_logits.softmax(dim=-1).cumsum(dim=-1)
# Filteration based on the cumulative sum.
filter_ = cumulative_probs > top_p
# This shift by 1 is weird and I cannot justify it. This existed
# in the original implementation:
# https://github.com/ari-holtzman/degen/blob/master/gen.py
# and I guess it is needed so keeping it for now.
filter_[:, 1:] = filter_[:, :-1].clone()
# Make sure we at least have one token to select from.
filter_[..., 0] = 0
# Fill in the filtered part
filter_ = filter_.scatter(1, sorted_indices, filter_)
logits.masked_fill_(filter_, float('-Inf'))
def modify_logits_for_top_k_filtering(logits, top_k):
"""Set the logits for none top-k values to -inf."""
filter_ = logits < torch.topk(logits, top_k)[0][..., -1, None]
logits.masked_fill_(filter_, float('-Inf'))
def sample(logits, top_k=0, top_p=0.0, temperature=1.0, vocab_size=None):
""" Sample and generate a token.
Note: logits has the dimension [b, v] where b is the batch size
and v is the vocabulary size.
If vocab_size is provided, we will make sure the sample that is
generated is in [0, vocab-size). This will avoid out of vocabulary
generations due to padding.
"""
# Check logits for consistency.
assert logits.ndim == 2, 'expected the logits to be of [b, v] shape.'
# assert logits.type() == 'torch.cuda.FloatTensor', \
# 'input logits should be floats.'
# Greedy is just simple argmax.
if top_k == 1:
assert top_p == 0.0, 'cannot set both greedy and top-p samplings.'
samples = torch.argmax(logits, dim=-1)
# Top-k or top-p sampling.
else:
# Clone so we do not modify the inputs,
logits = logits.clone()
# Apply temperature in place.
if temperature != 1.0:
logits.div_(temperature)
if top_k > 1:
assert top_p == 0.0, 'cannot set both top-k and top-p samplings.'
assert top_k <= logits.size(1), 'top-k is larger than logit size.'
if vocab_size:
assert top_k < vocab_size, 'top-k is larger than vocab size.'
modify_logits_for_top_k_filtering(logits, top_k)
elif top_p > 0.0:
assert top_p <= 1.0, 'top-p should be in (0, 1].'
modify_logits_for_top_p_filtering(logits, top_p)
# After filtering, we need to recalculate the distribution.
probs = logits.softmax(dim=-1)
samples = torch.multinomial(probs.float(), num_samples=1).view(-1)
# If vocab size is provided, make sure the samples are in
# in the range [0, vocab-size).
if vocab_size:
samples = torch.clamp(samples, min=0, max=(vocab_size - 1))
return samples
def broadcast_from_last_to_first_pipeline_stage(size, dtype, tensor=None):
"""Broadcast tensor values from last stage into the first stage."""
is_last_stage = gpc.is_pipeline_last_stage()
is_first_stage = gpc.is_pipeline_first_stage()
# If first stage and last state are the same, then there is no
# pipeline parallelism and no need to communicate.
if is_first_stage and is_last_stage:
return tensor
# Only first and last stage pipeline stages need to be involved.
if is_last_stage or is_first_stage:
if is_last_stage:
_is_cuda_contiguous(tensor)
else:
tensor = torch.empty(size,
dtype=dtype,
device=torch.cuda.current_device())
src = gpc.get_ranks_in_group(ParallelMode.PIPELINE)[-1]
group = gpc.get_group(ParallelMode.PIPELINE)
# Broadcast from last stage into the first stage.
dist.broadcast(tensor, src, group)
else:
tensor = None
return tensor
def broadcast_from_last_pipeline_stage(size, dtype, tensor=None):
"""Broadcast a tensor from last pipeline stage to all ranks."""
is_last_stage = gpc.is_pipeline_last_stage()
# If first stage and last state are the same, then there is no
# pipeline parallelism and no need to communicate.
if gpc.is_pipeline_first_stage() and is_last_stage:
return tensor
if is_last_stage:
_is_cuda_contiguous(tensor)
else:
tensor = torch.empty(size,
dtype=dtype,
device=torch.cuda.current_device())
# Get the group and corresponding source rank.
src = gpc.get_ranks_in_group(ParallelMode.PIPELINE)[-1]
group = gpc.get_group(ParallelMode.PIPELINE)
dist.broadcast(tensor, src, group)
return tensor
def copy_from_last_to_first_pipeline_stage(size, dtype, tensor=None):
"""Copy tensor values from last stage into the first stage.
Note that the input tensor is updated in place."""
is_last_stage = gpc.is_pipeline_last_stage()
is_first_stage = gpc.is_pipeline_first_stage()
# If first stage and last state are the same, then there is no
# pipeline parallelism and no need to communicate.
if is_first_stage and is_last_stage:
return
# Only first and last stage pipeline stages need to be involved.
if is_last_stage or is_first_stage:
_is_cuda(tensor)
is_contiguous = tensor.is_contiguous()
src = gpc.get_ranks_in_group(ParallelMode.PIPELINE)[-1]
group = gpc.get_group(ParallelMode.PIPELINE)
if is_contiguous:
tensor_ = tensor
else:
if is_last_stage:
tensor_ = tensor.contiguous()
else:
tensor_ = torch.empty(size,
dtype=dtype,
device=torch.cuda.current_device())
# Broadcast from last stage into the first stage.
dist.broadcast(tensor_, src, group)
# Update the first stage tensor
if is_first_stage and not is_contiguous:
tensor[...] = tensor_
def broadcast_float_list(size, float_list=None, rank=0):
"""Broadcast a list of float values."""
return broadcast_list(size, torch.float32, list_values=float_list, rank=rank)
def broadcast_list(size, dtype, list_values=None, rank=0):
"""Broadcast a list of values with a given type."""
world_size = gpc.get_world_size(ParallelMode.MODEL)
if world_size == 1:
return torch.tensor(list_values, dtype=dtype,
device=torch.cuda.current_device())
tensor = None
if gpc.get_local_rank(ParallelMode.MODEL) == rank:
tensor = torch.tensor(list_values, dtype=dtype,
device=torch.cuda.current_device())
return broadcast_tensor(size, dtype, tensor=tensor, rank=rank)
def _is_cuda(tensor):
"""Check if a tensor is not none and is cuda."""
assert tensor is not None
assert tensor.is_cuda
def _is_cuda_contiguous(tensor):
"""Check if a tensor is not none, is cuda, and is contiguous."""
_is_cuda(tensor)
assert tensor.is_contiguous()
def broadcast_tensor(size, dtype, tensor=None, rank=0):
""" Given size and type of a tensor on all ranks and the tensor value
only on a specific rank, broadcast from that rank to all other ranks.
"""
group = gpc.get_group(ParallelMode.MODEL)
world_size = gpc.get_world_size(ParallelMode.MODEL)
if world_size == 1:
return tensor
if gpc.get_local_rank(ParallelMode.MODEL) == rank:
_is_cuda_contiguous(tensor)
else:
tensor = torch.empty(size,
dtype=dtype,
device=torch.cuda.current_device())
src = gpc.get_ranks_in_group(ParallelMode.MODEL)[rank]
dist.broadcast(tensor, src, group=group)
return tensor
class VocabUtility(object):
"""Split the vocabulary into `world_size` chunks amd return the
first and last index of the vocabulary belonging to the `rank`
partition: Note that indecies in [fist, last)"""
@staticmethod
def vocab_range_from_per_partition_vocab_size(per_partition_vocab_size: int,
rank: int,
world_size: int) -> Tuple[int, int]:
index_f = rank * per_partition_vocab_size
index_l = index_f + per_partition_vocab_size
return index_f, index_l
@staticmethod
def vocab_range_from_global_vocab_size(global_vocab_size: int,
rank: int,
world_size: int) -> Tuple[int, int]:
per_partition_vocab_size = divide(global_vocab_size, world_size)
return VocabUtility.vocab_range_from_per_partition_vocab_size(
per_partition_vocab_size, rank, world_size)
def ensure_divisibility(numerator: int, denominator: int) -> None:
"""Ensure that numerator is divisible by the denominator."""
assert numerator % denominator == 0, \
f'{numerator} is not divisible by {denominator}'
def divide(numerator: int, denominator: int) -> int:
"""Ensure that numerator is divisible by the denominator and return
the division value."""
ensure_divisibility(numerator, denominator)
return numerator // denominator
def nn_conv_weight_init_(weight: torch.Tensor):
init.kaiming_uniform_(weight, a=math.sqrt(5))
def nn_conv_bias_init_(bias: torch.Tensor,
weight: torch.Tensor) -> None:
if bias is not None:
fan_in, _ = init._calculate_fan_in_and_fan_out(weight)
if fan_in != 0:
bound = 1 / math.sqrt(fan_in)
init.uniform_(bias, -bound, bound)
def import_func(tasks_dir, namespace):
for file in os.listdir(tasks_dir):
path = os.path.join(tasks_dir, file)
if (
not file.startswith("_")
and not file.startswith(".")
and (file.endswith(".py") or os.path.isdir(path))
):
task_name = file[: file.find(".py")] if file.endswith(".py") else file
importlib.import_module(namespace + "." + task_name)
def import_user_module(args):
module_path = getattr(args, "user_dir", None)
if module_path is not None:
module_path = os.path.abspath(args.user_dir)
# ensure that user modules are only imported once
import_user_module.memo = getattr(import_user_module, "memo", set())
if module_path not in import_user_module.memo:
import_user_module.memo.add(module_path)
module_parent, module_name = os.path.split(module_path)
if module_name not in sys.modules:
sys.path.insert(0, module_parent)
importlib.import_module(module_name)
models_path = os.path.join(module_path, "models")
if os.path.exists(models_path):
import_func(models_path, f"{module_name}.models")
tokenizers_path = os.path.join(module_path, "tokenizers")
if os.path.exists(tokenizers_path):
import_func(tokenizers_path, f"{module_name}.tokenizers")
else:
raise ImportError(
"Failed to import --user-dir={} because the corresponding module name "
"({}) is not globally unique. Please rename the directory to "
"something unique and try again.".format(module_path, module_name)
)
def merge_with_parent(dc , cfg: DictConfig, remove_missing=True):
if remove_missing:
if is_dataclass(dc):
target_keys = set(dc.__dataclass_fields__.keys())
else:
target_keys = set(dc.keys())
with open_dict(cfg):
for k in list(cfg.keys()):
if k not in target_keys:
del cfg[k]
merged_cfg = OmegaConf.merge(dc, cfg)
merged_cfg.__dict__["_parent"] = cfg.__dict__["_parent"]
OmegaConf.set_struct(merged_cfg, True)
return merged_cfg
def add_defaults(cfg: DictConfig) -> None:
"""This function adds default values that are stored in dataclasses that hydra doesn't know about """
from typing import Any
OmegaConf.set_struct(cfg, False)
for k, v in {"model":None, "tokenizer":None}.items():
field_cfg = cfg.get(k)
if field_cfg is not None:
dc = None
if isinstance(field_cfg, str):
field_cfg = DictConfig({"_name": field_cfg})
field_cfg.__dict__["_parent"] = field_cfg.__dict__["_parent"]
name = getattr(field_cfg, "_name", None)
if k=="model":
from TDATR.models.mini_gpt4_ipt_v2 import MiniGPT4Config
dc = MiniGPT4Config
if dc is not None:
cfg[k] = merge_with_parent(dc, field_cfg)
else:
from TDATR.tokenizers.bbox_tokenizer import BboxTokenConfig
dc = BboxTokenConfig
if dc is not None:
cfg[k] = merge_with_parent(dc, field_cfg)
assert cfg.model is not None, 'Missing model config!'
from enum import Enum
from TDATR_utils.global_variables import ParallelMode
def distributed_main(i, main, cfg, kwargs):
cfg.distributed_training.device_id = i
#if torch.cuda.is_available() and not cfg.common.cpu:
# torch.cuda.set_device(cfg.distributed_training.device_id)
if cfg.distributed_training.distributed_rank is None: # torch.multiprocessing.spawn
cfg.distributed_training.distributed_rank = kwargs.pop("start_rank", 0) + i
after_distributed_init_fn = kwargs.pop("after_distributed_init_fn", None)
if after_distributed_init_fn:
cfg = after_distributed_init_fn(cfg)
main(cfg, **kwargs)
if torch.distributed.is_initialized():
torch.distributed.barrier(gpc.get_group(ParallelMode.GLOBAL))
def call_main(cfg, main, **kwargs):
if cfg.distributed_training.distributed_init_method is None:
infer_init_method(cfg.distributed_training)
if cfg.distributed_training.distributed_init_method is not None:
# distributed training
if not cfg.distributed_training.distributed_no_spawn:
start_rank = cfg.distributed_training.distributed_rank
cfg.distributed_training.distributed_rank = None # assign automatically
kwargs["start_rank"] = start_rank
torch.multiprocessing.spawn(
fn=distributed_main,
args=(main, cfg, kwargs),
nprocs=min(
torch.cuda.device_count(),
cfg.distributed_training.distributed_world_size,
),
join=True,
)
else:
distributed_main(cfg.distributed_training.device_id, main, cfg, kwargs)
else:
# single GPU main
main(cfg, **kwargs)
try:
from hydra import compose, initialize
except ImportError:
from hydra.experimental import compose, initialize
from hydra.core.global_hydra import GlobalHydra
from omegaconf import DictConfig, OmegaConf, open_dict, _utils
from argparse import ArgumentError, ArgumentParser, Namespace
class omegaconf_no_object_check:
def __init__(self):
self.old_is_primitive = _utils.is_primitive_type
def __enter__(self):
_utils.is_primitive_type = lambda _: True
def __exit__(self, type, value, traceback):
_utils.is_primitive_type = self.old_is_primitive
def _set_legacy_defaults(args, cls):
"""Helper to set default arguments based on *add_args*."""
if not hasattr(cls, "add_args"):
return
import argparse
parser = argparse.ArgumentParser(
argument_default=argparse.SUPPRESS, allow_abbrev=False
)
cls.add_args(parser)
# copied from argparse.py:
defaults = argparse.Namespace()
for action in parser._actions:
if action.dest is not argparse.SUPPRESS:
if not hasattr(defaults, action.dest):
if action.default is not argparse.SUPPRESS:
setattr(defaults, action.dest, action.default)
for key, default_value in vars(defaults).items():
if not hasattr(args, key):
setattr(args, key, default_value)
def override_module_args(args: Namespace) -> Tuple[List[str], List[str]]:
"""use the field in args to overrides those in cfg"""
overrides = []
deletes = []
return overrides, deletes
def convert_namespace_to_omegaconf(args: Namespace) -> DictConfig:
"""Convert a flat argparse.Namespace to a structured DictConfig."""
# Here we are using field values provided in args to override counterparts inside config object
overrides, deletes = override_module_args(args)
# configs will be in hulk/config after installation
config_path = os.path.join("..", "config")
GlobalHydra.instance().clear()
with initialize(config_path=config_path):
try:
composed_cfg = compose("config", overrides=overrides, strict=False)
except:
logger.error("Error when composing. Overrides: " + str(overrides))
raise
for k in deletes:
composed_cfg[k] = None
cfg = OmegaConf.create(
OmegaConf.to_container(composed_cfg, resolve=True, enum_to_str=True)
)
# hack to be able to set Namespace in dict config. this should be removed when we update to newer
# omegaconf version that supports object flags, or when we migrate all existing models
from omegaconf import _utils
OmegaConf.set_struct(cfg, True)
return cfg
def gen_parser_from_dataclass(
parser: ArgumentParser,
dataclass_instance,
delete_default: bool = False,
with_prefix: Optional[str] = None,
) -> None:
"""
convert a dataclass instance to tailing parser arguments.
If `with_prefix` is provided, prefix all the keys in the resulting parser with it. It means that we are
building a flat namespace from a structured dataclass (see transformer_config.py for example).
"""
def argparse_name(name: str):
if name == "data" and (with_prefix is None or with_prefix == ''):
# normally data is positional args, so we don't add the -- nor the prefix
return name
if name == "_name":
# private member, skip
return None
full_name = "--" + name.replace("_", "-")
if with_prefix is not None and with_prefix != '':
# if a prefix is specified, construct the prefixed arg name
full_name = with_prefix + "-" + full_name[2:] # strip -- when composing
return full_name
def _is_tensor_parallel(partition_dim: int) -> bool:
if partition_dim is None:
return False
if gpc.get_world_size(ParallelMode.TENSOR) <= 1:
return False
return True
from contextlib import contextmanager, nullcontext
@contextmanager
def init_tensor_parallel_parameters(is_gpu_initializer=True):
torch_calc_fan_in_fan_out = torch.nn.init._calculate_fan_in_and_fan_out
def _calculate_fan_in_and_fan_out(tensor):
fan_in, fan_out = torch_calc_fan_in_fan_out(tensor)
if not getattr(tensor, _TENSOR_PARALLEL, False):
return fan_in, fan_out # 1, 0
tensor_dim = tensor.dim()
world_size = gpc.get_world_size(ParallelMode.TENSOR)
partition_dim = getattr(tensor, _PARTITION_DIM)
if not isinstance(partition_dim, int) or not (0 <= partition_dim < tensor_dim):
raise ValueError(f"Detected partition_dim={partition_dim}, but tensor dim={tensor_dim}.")
if partition_dim == 0:
fan_out *= world_size
elif partition_dim == 1:
fan_in *= world_size
elif tensor_dim > 2 and partition_dim >= 2:
fan_out *= world_size
fan_in *= world_size
return fan_in, fan_out
torch.nn.init._calculate_fan_in_and_fan_out = _calculate_fan_in_and_fan_out
context = nullcontext()
with context:
yield
torch.nn.init._calculate_fan_in_and_fan_out = torch_calc_fan_in_fan_out
def broadcast(tensor: Tensor, src: int, parallel_mode: ParallelMode, async_op: bool = False):
r"""Broadcast tensors to whole parallel group. Tensor must have the same
number of elements in all processes participating in the collective.
Note:
The parallel_mode should be concluded in ``ParallelMode``. More details about ``ParallelMode`` could be found
in `parallel_mode <https://github.com/hpcaitech/ColossalAI/blob/main/colossalai/context/parallel_mode.py>`_.
Args:
tensor (:class:`torch.Tensor`): Tensor to be broadcast.
src (int): Source rank.
parallel_mode (:class:`colossalai.context.ParallelMode`): Parallel group mode used in this communication.
async_op (bool, optional): Whether operations are asynchronous.
Returns:
Union[tuple(:class:`torch.Tensor`, work handle), :class:`torch.Tensor`]: The tensor need to be broadcast only,
if async_op is set to False. A tuple of output of all-gather and Async work handle, if async_op is set to True.
"""
depth = gpc.get_world_size(parallel_mode)
if depth == 1:
out = tensor
work = None
else:
out = tensor.contiguous()
if out.data.dtype == torch.bfloat16:
assert not async_op, "BF16 does not support async broadcast"
if gpc.get_world_size(parallel_mode) != 1:
out = out.to(torch.cuda.current_device())
dist.broadcast(
out,
src=src,
group=gpc.get_group(parallel_mode)
)
out = out.cpu()
else:
group = gpc.get_cpu_group(parallel_mode) if tensor.device.type == "cpu" else gpc.get_group(parallel_mode)
work = dist.broadcast(out, src=src, group=group, async_op=async_op)
if async_op:
return out, work
else:
return out
_TENSOR_PARALLEL = "tensor_model_parallel"
_PARTITION_DIM = "partition_dim"
_PARTITION_STRIDE = "partition_stride"
_MODEL_PARALLEL_ATTRIBUTE_DEFAULTS = {_TENSOR_PARALLEL: False,
_PARTITION_DIM: -1,
_PARTITION_STRIDE: 1}
def set_tensor_model_parallel_attributes(tensor: torch.Tensor,
is_parallel: int,
dim: int,
partition_stride: int):
# Make sure the attributes are not set.
for attribute in _MODEL_PARALLEL_ATTRIBUTE_DEFAULTS:
assert not hasattr(tensor, attribute)
# Set the attributes.
setattr(tensor, _TENSOR_PARALLEL, is_parallel)
setattr(tensor, _PARTITION_DIM, dim)
setattr(tensor, _PARTITION_STRIDE, partition_stride)
@torch.no_grad()
def initialize_weight_gpu(weight: torch.Tensor,
init_method: Callable,
partition_dim: Optional[int]=None,
partition_stride:Optional[int]=1):
assert weight.device.type == 'cuda' or weight.device.type == 'npu', f"Expected a cuda tensor, but got {weight.device}."
is_parallel = _is_tensor_parallel(partition_dim)
if not is_parallel:
partition_dim = None
set_tensor_model_parallel_attributes(tensor=weight,
is_parallel=is_parallel,
dim=partition_dim,
partition_stride=partition_stride)
with init_tensor_parallel_parameters():
init_method(weight)
if not is_parallel: # is a global tensor
broadcast(weight, gpc.get_ranks_in_group(ParallelMode.TENSOR)[0], ParallelMode.TENSOR)
def get_1d_parallel_split_size(dim_size:int,
partition_stride: int=1,
world_size: Optional[int]=None) -> Tuple[int, int]:
"""If we split tensor by world size and stride along a dim,
this function will return the number of chunks and size of each chunk."""
if world_size is None:
world_size = gpc.get_world_size(ParallelMode.TENSOR)
per_partition_size = divide(dim_size, world_size)
per_partition_per_stride_size = divide(per_partition_size, partition_stride)
num_chunks = partition_stride * world_size
return num_chunks, per_partition_per_stride_size
def get_global_tensor_size(dim_size:int, world_size: Optional[int]=None) -> int:
"""Compute the global tensor dim size"""
if world_size is None:
world_size = gpc.get_world_size(ParallelMode.TENSOR)
return dim_size * world_size
def get_1d_parallel_split_indices(rank:Optional[int]=None,
partition_stride: int=1,
world_size: Optional[int]=None) -> Tuple[int]:
"""If we split tensor by world size and stride along a dim,
this function will return the indices of splited chunks on
the current member in TP group."""
if rank is None:
rank = gpc.get_local_rank(ParallelMode.TENSOR)
if world_size is None:
world_size = gpc.get_world_size(ParallelMode.TENSOR)
num_chunks = partition_stride * world_size
return tuple(range(rank, num_chunks, world_size))
def split_tensor_to_local(tensor: torch.Tensor,
partition_dim:int,
partition_stride: Optional[int],
out:Optional[torch.Tensor]=None,
world_size: Optional[int]=None,
cur_rank: Optional[int]=None) -> torch.Tensor:
"""Split a tensor by tensor parallel size and stride along a dim.
If we want to split a global tensor along `dim` to each member of tensor parallel group,
we first split weight to `world_size*stride` partitions, each partition size equal to
`dim_size//(world_size*stride)`.
with world_size=4, stride=1, each member of tensor parallel groups gets the indices of partitions:
rank=0: [0]
rank=1: [1]
rank=2: [2]
rank=3: [3]
with world_size=4, stride=2, each member of tensor parallel groups gets the indices of partitions:
rank=0: [0, 4]
rank=1: [1, 5]
rank=2: [2, 6]
rank=3: [3, 7]
with world_size=4, stride=4, each member of tensor parallel groups gets the indices of partitions:
parallel rank=0: [0, 4, 8, 12]
parallel rank=1: [1, 5, 9, 13]
parallel rank=2: [2, 6, 10, 14]
parallel rank=3: [3, 7, 11, 15]
"""
if world_size is None:
world_size = gpc.get_world_size(ParallelMode.TENSOR)
if cur_rank is None:
cur_rank = gpc.get_local_rank(ParallelMode.TENSOR)
if world_size == 1:
return tensor if out is None else out.data.copy_(tensor)
num_chunks, per_partition_per_stride_size = \
get_1d_parallel_split_size(dim_size=tensor.size(partition_dim),
partition_stride=partition_stride,
world_size=world_size)
assert per_partition_per_stride_size * num_chunks == tensor.size(partition_dim)
tensor_chunks = torch.chunk(tensor.data, num_chunks, dim=partition_dim)
partition_indices = get_1d_parallel_split_indices(rank=cur_rank,
partition_stride=partition_stride,
world_size=world_size)
if out is not None and isinstance(out, torch.Tensor):
return torch.cat([tensor_chunks[i] for i in partition_indices], dim=partition_dim, out=out)
tensor_partition = torch.cat([tensor_chunks[i] for i in partition_indices], dim=partition_dim)
return tensor_partition.contiguous()
@torch.no_grad()
def initialize_weight_cpu(weight: torch.Tensor,
partition_dim: Optional[int]=None,
per_partition_size: Optional[int]=0,
init_method: Optional[Callable]=None,
partition_stride: Optional[int]=1) -> Union[None, torch.Tensor]:
assert weight.device.type == 'cpu', f"Expected a cpu tensor, but got {weight.device}."
is_parallel = _is_tensor_parallel(partition_dim)
if not is_parallel:
partition_dim = None
set_tensor_model_parallel_attributes(tensor=weight,
is_parallel=is_parallel,
dim=partition_dim,
partition_stride=partition_stride)
shape = list(weight.shape)
if is_parallel:
tp_world_size = gpc.get_world_size(ParallelMode.TENSOR)
shape[partition_dim] = shape[partition_dim] * tp_world_size
master_weight = torch.empty(shape, dtype=torch.float32, requires_grad=False)
with init_tensor_parallel_parameters(is_gpu_initializer=False):
init_method(master_weight)
master_weight = master_weight.to(dtype=weight.dtype)
if is_parallel:
assert per_partition_size != 0, "per_partition_size can't be 0 if partition_dim is not None"
# Split and copy
split_tensor_to_local(master_weight, partition_dim, partition_stride, weight)
else: # global tensor
weight.data = master_weight.data.clone()
broadcast(weight, gpc.get_ranks_in_group(ParallelMode.TENSOR)[0], ParallelMode.TENSOR)