Source code for torch.distributed.algorithms.ddp_comm_hooks.powerSGD_hook
import logging
import math
import numpy as np
import torch
import torch.distributed as dist
from . import default_hooks as default
def _orthogonalize(matrix, epsilon=1e-8):
"""
Applies Gram-Schmidt procedure to orthogonalize a given 2D tensor.
If epsilon is 0, this is equivalent to `torch.qr(matrix, out=(matrix, _))`,
but `torch.qr` is very slow, probably because it is not optimized for a matrix that has a small number of columns.
"""
num_cols = matrix.shape[1]
for i in range(num_cols):
# Normalize the i'th column.
col = matrix[:, i : i + 1]
# If no epsilon is added here, division by zero may be caused by vanishing gradients.
# This epsilon is not needed if the input matrix covers the gradients of at least one entire layer in the neural network.
if epsilon == 0:
# Note that col ** 2 can underflow/overflow if we use FP16.
# May need to consider multiplying a scaling factor and dividing it later, or using bfloat16 instead.
col /= torch.norm(col)
else:
col /= torch.norm(col) + epsilon
# Project it on the rest and remove it.
if i + 1 < num_cols:
rest = matrix[:, i + 1 :]
rest -= torch.sum(col * rest, dim=0) * col
[docs]class PowerSGDState(object):
r"""
Stores both the algorithm's hyperparameters and the internal state for all the gradients during the training.
Particularly, ``matrix_approximation_rank`` and ``start_powerSGD_iter`` are the main hyperparameters that should be tuned by the user.
For performance, we suggest to keep binary hyperparameters ``use_error_feedback`` and ``warm_start`` on.
1. ``matrix_approximation_rank`` controls the size of compressed low-rank tensors, which determines the compression rate. The lower the rank, the stronger the compression.
1.1. If ``matrix_approximation_rank`` is too low, the full model quality will need more training steps to reach or will never reach and yield loss in accuracy.
1.2. The increase of ``matrix_approximation_rank`` can substantially increase the computation costs of the compression, and the accuracy may not be futher improved beyond a certain ``matrix_approximation_rank`` threshold.
To tune ``matrix_approximation_rank``, we suggest to start from 1 and increase by factors of 2 (like an expoential grid search, 1, 2, 4, ...), until a satisfactory accuracy is reached. Typically only a small value 1-4 is used. For some NLP tasks (as shown in Appendix D of the original paper), this value has been increased to 32.
2. ``start_powerSGD_iter`` defers PowerSGD compression util step ``start_powerSGD_iter``, and vanilla allreduce runs prior to step ``start_powerSGD_iter``. This hybrid scheme of **vanilla allreduce + PowerSGD** can effectively improve the accuracy, even a relatively small ``matrix_approximation_rank`` is used. This is because that, the beginning of training phase is usually very sensitive to inaccurate gradients, and compressing gradients too early may make the training quickly take a suboptimal trajectory, which can result in an irrecoverable impact on the accuracy.
To tune ``start_powerSGD_iter``, we suggest to start with 10% of total training steps, and increase it until a satisfactory accuracy is reached.
.. warning ::
If error feedback or warm-up is enabled, the minimum value of ``start_powerSGD_iter`` allowed in DDP is 2.
This is because there is another internal optimization that rebuilds buckets at iteration 1 in DDP,
and this can conflict with any tensor memorized before the rebuild process.
""" # noqa
__slots__ = [
"process_group",
# The two fields below are the hyperparameters that should be tuned by the user.
"matrix_approximation_rank",
"start_powerSGD_iter",
# The two fields below are the binary hyperparameters recommended to be turned on for performance.
"use_error_feedback",
"warm_start",
# The fields below are internal state.
"rng",
"error_dict",
"p_memory_dict",
"q_memory_dict",
"iter",
]
def __init__(
self,
process_group,
matrix_approximation_rank=1,
start_powerSGD_iter=10,
use_error_feedback=True,
warm_start=True,
random_seed=0,
):
logging.info(
"PowerSGD config: matrix_approximation_rank = {}; "
"start_powerSGD_iter = {}; use_error_feedback = {}; warm_start = {}.".format(
matrix_approximation_rank,
start_powerSGD_iter,
use_error_feedback,
warm_start,
)
)
self.process_group = process_group
self.matrix_approximation_rank = matrix_approximation_rank
# Deferring PowerSGD compression util step 'start_powerSGD_iter' can have two advantages:
# 1) It turns out that PowerSGD may lead to a non-trivial accuracy loss,
# even if the matrix approximation rank is increased to a large value.
# To mitigate the accuracy loss, a simple yet effective way is mixing vanilla allreduce
# (or a more conservative compression such as FP16 compression) with PowerSGD.
# 2) There is an internal optimization of rebuilding buckets process in DDP,
# in order to save the memory space.
# This step takes place after the first iteration.
# However, this means that the shape of input bucketized tensors is subject to change,
# which will complicate the implementations of error feedback and warm-up.
# Running vanilla allreduce in the first few iterations can avoid this complexity.
if (use_error_feedback or warm_start) and start_powerSGD_iter <= 1:
raise ValueError(
"Expect `start_powerSGD_iter` > 1 if `use_error_feedback` or `warm_start` is enabled, "
"because PowerSGD can only be applied after the first two iterations in DDP."
)
self.start_powerSGD_iter = start_powerSGD_iter
# Error feedback is usually crucial for both for convergence and generalization,
# because PowerSGD is a biased compressor,
# i.e., compressing and decompressing a random gradient does not yield the original in expectation.
# This mechanism requires a temporary copy of the input gradients,
# so it increases the peak memory consumption by the size of the gradient tensor.
# However, if the target matrices are known to be exactly low-ranked (instead of just low stable rank),
# sometimes it is possible to converge to the optima without error feedback.
# See: http://proceedings.mlr.press/v54/yurtsever17a/yurtsever17a.pdf
self.use_error_feedback = use_error_feedback
# Warm-start reuses P(s) and Q(s) from the previous iteration.
# This can improve the approximation quality and hence improve the accuracy.
# Additionally, by avoiding the initialization of these low-rank tensors at every step,
# this can also accelerate training.
# However, this is at the cost of extra memory.
self.warm_start = warm_start
# The purpose of this RNG is to generate different random seeds for initializing Q across iterations,
# but in the same order for all the DDP replicas.
# Different random seeds across iterations indicate different 'projections' of the gradients at different SGD steps.
# If the same random projection is used,
# there will be differences between the gradients that are never synchronized.
self.rng = np.random.RandomState(random_seed)
# Since there is only a single state instance for all the input buckets,
# need to maintain a dictionary that maps each bucket index to the local error.
self.error_dict = {}
self.p_memory_dict = {}
self.q_memory_dict = {}
# Iteration/step in the training loop.
self.iter = 0
def maybe_increase_iter(self, bucket):
# Since bucket 0 is the last bucket to allreduce in an iteration.
# Only increase `iter` when bucket 0 is processed.
if bucket.get_index() == 0:
self.iter += 1
if self.iter == self.start_powerSGD_iter:
logging.info(
"Start to apply PowerSGD after {} iterations.".format(self.iter)
)
[docs]def powerSGD_hook(state: PowerSGDState, bucket) -> torch.futures.Future:
r"""
This DDP communication hook implements PowerSGD gradient compression
algorithm described in the `paper <https://arxiv.org/abs/1905.13727>`_.
Once gradient tensors are aggregated across all workers, this hook applies
compression as follows:
1. Views the input flattened 1D gradient tensor as two groups of per-parameter tensors: high-rank tensors and vector-like rank-1 tensors (for biases).
2. Handles rank-1 tensors by allreducing them without compression:
2.1. Allocate contiguous memory for those rank-1 tensors, and allreduces all the rank-1 tensors as a batch, without compression;
2.2. Copies the individual rank-1 tensors from the contiguous memory back to the input tensor.
3. Handles high-rank tensors by PowerSGD compression:
3.1. For each high-rank tensor M, creates two low-rank tensors P and Q for decomposing M,
such that M = PQ^T, where Q is initialized from a standard normal distribution and orthogonalized;
3.2. Computes each P in Ps, which is equal to MQ;
3.3. Allreduces Ps as a batch;
3.4. Orthogonalizes each P in Ps;
3.5. Computes each Q in Qs, which is approximately equal to M^TP;
3.6. Allreduces Qs as a batch;
3.7. Computes each M among all the high-rank tensors, which is approximately equal to PQ^T.
Note that this communication hook enforces vanilla allreduce for the first ``state.start_powerSGD_iter`` iterations.
This not only gives the user more control over the tradeoff between speedup and accuracy,
but also helps abstract away some complexity of the internal optimization of DDP for future communication hook developers.
Args:
state (PowerSGDState): State information to configure the compression rate and support error feedback, warm start, etc.
To tune the compression configs, mainly need to tune `matrix_approximation_rank`` and ``start_powerSGD_iter``.
bucket (dist._GradBucket): Bucket that stores a 1D flattened gradient tensor that batches multiple per-variable tensors.
Note that since DDP comm hook only supports single process single device mode at this time,
only exactly one tensor is stored in this bucket.
Returns:
Future handler of the communication, which updates the gradients in place.
Example::
>>> state = PowerSGDState(process_group=process_group, matrix_approximation_rank=1, start_powerSGD_iter=10)
>>> ddp_model.register_comm_hook(state, powerSGD_hook)
""" # noqa
process_group = state.process_group
group_to_use = process_group if process_group is not None else dist.group.WORLD
world_size = group_to_use.size()
# The input tensor is a flattened 1D tensor.
input_tensor = bucket.get_tensors()[0]
# Run vanilla allreduce in the first `start_powerSGD_iter` iterations.
if state.iter < state.start_powerSGD_iter:
state.maybe_increase_iter(bucket)
return default._allreduce_fut(group_to_use, input_tensor)
# Apply PowerSGD after `start_powerSGD_iter` iterations.
device = input_tensor.device
dtype = input_tensor.dtype
# Incorporate the error from the previous state into the gradients.
bucket_index = bucket.get_index()
input_tensor_cp = None
total_length = input_tensor.shape[0]
if state.use_error_feedback:
if bucket_index in state.error_dict:
input_tensor.add_(state.error_dict[bucket_index])
else:
logging.info(
"A zero tensor of length {} that represents local error is created.".format(
total_length
)
)
state.error_dict[bucket_index] = torch.zeros(
total_length, device=device, dtype=dtype
)
# Keep a copy of the input tensor,
# so that we can compute the local error caused by compression later,
# by comparing this copy and the input tensor updated after decompression.
input_tensor_cp = torch.clone(input_tensor).detach()
# Unflatten the input tensor into per-parameter tensors, for layer-wise compression.
tensors = [
input_tensor[offset : offset + length].view(sizes)
for offset, length, sizes in zip(
bucket.get_offsets(), bucket.get_lengths(), bucket.get_sizes_list()
)
]
# Step I: Handle rank-1 tensors.
# Allocate contiguous memory for rank-1 tensors to allreduce them without compression efficiently.
rank1_tensors = [tensor for tensor in tensors if tensor.ndimension() <= 1]
rank1_tensors_memory = (
torch.cat([tensor.view(-1) for tensor in rank1_tensors])
if rank1_tensors
else torch.tensor([], device=device, dtype=dtype)
)
# Step II: Handle high-rank tensors.
# Allocate contiguous memory for Ps and Qs to allreduce compressed high-rank tensors efficiently.
high_rank_tensors = [
tensor.view(tensor.shape[0], -1)
for tensor in tensors
if tensor.ndimension() > 1
]
total_Ps_size = 0
total_Qs_size = 0
for tensor in high_rank_tensors:
n, m = tensor.shape
matrix_approximation_rank = min(n, m, state.matrix_approximation_rank)
total_Ps_size += n * matrix_approximation_rank
total_Qs_size += m * matrix_approximation_rank
# If warm-start is enabled, reuse Ps and Qs from the previous iteration if possible.
# The memory spaces of Ps and Qs need to be allocated in the first iteration when PowerSGD is applied.
need_randomize_qs = False
if not state.warm_start or bucket_index not in state.p_memory_dict:
need_randomize_qs = True
# If warm-start is disabled, low-rank tensors will be initialized at every step.
# Only log this if warm-start to avoid spamming.
if state.warm_start:
logging.info(
"Allocating contiguous memory of length {} for Ps, and of length {} for Qs, respectively.".format(
total_Ps_size, total_Qs_size
)
)
state.p_memory_dict[bucket_index] = torch.empty(
total_Ps_size, device=device, dtype=dtype
)
state.q_memory_dict[bucket_index] = torch.empty(
total_Qs_size, device=device, dtype=dtype
)
# Create Ps and Qs that point to the allocated memory.
ps = []
qs = []
p_idx = 0
q_idx = 0
for tensor in high_rank_tensors:
n, m = tensor.shape
matrix_approximation_rank = min(n, m, state.matrix_approximation_rank)
ps.append(
state.p_memory_dict[bucket_index][
p_idx : p_idx + n * matrix_approximation_rank
].view(n, matrix_approximation_rank)
)
qs.append(
state.q_memory_dict[bucket_index][
q_idx : q_idx + m * matrix_approximation_rank
].view(m, matrix_approximation_rank)
)
p_idx += n * matrix_approximation_rank
q_idx += m * matrix_approximation_rank
# If warm-start is enabled, reuse Qs from the previous iteration if possible and skip filling random values.
# The exception is the first iteration when PowerSGD is applied.
if not need_randomize_qs:
for q in qs:
_orthogonalize(q)
else:
with torch.random.fork_rng(devices=[]):
# Fork this RNG to avoid changing the seed globally and affecting the random sampling anywhere else in the training.
# The seed makes sure that the initial random values are the same across all the DDP replicas.
# This seed should differ at every step.
# Since it is very slow to fork RNG state across all the CUDA devices,
# only fork on CPU and then move the generated tensor to the CUDA device (by overwriting q).
torch.manual_seed(state.rng.randint(1_000_000_000))
for q in qs:
q.copy_(
torch.randn(
*q.shape,
device="cpu",
dtype=dtype,
)
)
_orthogonalize(q)
# Compute Ps.
for tensor, q, p in zip(high_rank_tensors, qs, ps):
torch.matmul(tensor, q, out=p)
# This allreduce is only applied to rank-1 tensors,
# so it should have been kicked off before the above computation on the high-rank tensors to hide more communication costs.
# However, this somehow requires a separate future chain at this time.
allreduce_contiguous_rank1_tensors_fut = dist.all_reduce(
rank1_tensors_memory, group=group_to_use, async_op=True
).get_future()
def unpack_rank1_tensors_and_allreduce_ps(fut):
rank1_tensors_memory = fut.value()[0].div_(world_size)
idx = 0
for tensor in rank1_tensors:
tensor.copy_(rank1_tensors_memory[idx : idx + tensor.shape[0]])
idx += tensor.shape[0]
# Since these Ps will be orthogonalized later, no need to divide them by world size.
return [
dist.all_reduce(
state.p_memory_dict[bucket_index], group=group_to_use, async_op=True
)
.get_future()
.wait()[0]
]
def compute_qs(fut):
state.p_memory_dict[bucket_index] = fut.value()[0]
for p in ps:
_orthogonalize(p)
# Compute Qs.
for tensor, p, q in zip(high_rank_tensors, ps, qs):
torch.matmul(tensor.t(), p, out=q)
# TODO: The above procedure does two matmul+allreduce steps per iteration --
# one left multiplication and one right multiplication.
# For warm-start, can take one such step at a time, and alternate between them.
# Allreduce Qs.
return [
dist.all_reduce(
state.q_memory_dict[bucket_index], group=group_to_use, async_op=True
)
.get_future()
.wait()[0]
]
def decompress(fut):
state.q_memory_dict[bucket_index] = fut.value()[0].div_(world_size)
for p, q, tensor in zip(ps, qs, high_rank_tensors):
torch.matmul(p, q.t(), out=tensor)
if torch.cuda.is_available():
torch.cuda.synchronize(device)
if state.use_error_feedback:
# Memorize the local errors.
state.error_dict[bucket_index] = input_tensor_cp - input_tensor
if not state.warm_start:
state.p_memory_dict.clear()
state.q_memory_dict.clear()
state.maybe_increase_iter(bucket)
return [input_tensor]
return (
allreduce_contiguous_rank1_tensors_fut.then(
unpack_rank1_tensors_and_allreduce_ps
)
.then(compute_qs)
.then(decompress)
)
[docs]def batched_powerSGD_hook(state: PowerSGDState, bucket) -> torch.futures.Future:
r"""
This DDP communication hook implements a simplified PowerSGD gradient compression
algorithm described in the `paper <https://arxiv.org/abs/1905.13727>`_.
This variant does not compress the gradients layer by layer,
but instead compresses the flattened input tensor that batches all the gradients.
Therefore, it is **faster** than :meth:`powerSGD_hook`,
but usually results in a **much lower accuracy**, unless ``matrix_approximation_rank`` is 1.
.. warning ::
Increasing ``matrix_approximation_rank`` here may not necessarily increase the accuracy,
because batching per-parameter tensors without column/row alignment can destroy low-rank structure.
Therefore, the user should always consider :meth:`powerSGD_hook` first,
and only consider this variant when a satisfactory accuracy can be achieved when ``matrix_approximation_rank`` is 1.
Once gradient tensors are aggregated across all workers, this hook applies
compression as follows:
1. Views the input flattened 1D gradient tensor as a square-shaped tensor M with 0 paddings;
2. Creates two low-rank tensors P and Q for decomposing M, such that M = PQ^T, where Q is initialized from a standard normal distribution and orthogonalized;
3. Computes P, which is equal to MQ;
4. Allreduces P;
5. Orthogonalizes P;
6. Computes Q, which is approximately equal to M^TP;
7. Allreduces Q;
8. Computes M, which is approximately equal to PQ^T.
9. Truncates the input tensor to the original length.
Note that this communication hook enforces vanilla allreduce for the first ``state.start_powerSGD_iter`` iterations.
This not only gives the user more control over the tradeoff between speedup and accuracy,
but also helps abstract away some complexity of the internal optimization of DDP for future communication hook developers.
Args:
state (PowerSGDState): State information to configure the compression rate and support error feedback, warm start, etc.
To tune the compression configs, mainly need to tune ``matrix_approximation_rank`` and ``start_powerSGD_iter``.
bucket (dist._GradBucket): Bucket that stores a 1D flattened gradient tensor that batches multiple per-variable tensors.
Note that since DDP comm hook only supports single process single device mode at this time,
only exactly one tensor is stored in this bucket.
Returns:
Future handler of the communication, which updates the gradients in place.
Example::
>>> state = PowerSGDState(process_group=process_group, matrix_approximation_rank=1)
>>> ddp_model.register_comm_hook(state, batched_powerSGD_hook)
""" # noqa
process_group = state.process_group
group_to_use = process_group if process_group is not None else dist.group.WORLD
world_size = group_to_use.size()
# The input tensor is a flattened 1D tensor.
input_tensor = bucket.get_tensors()[0]
# Run vanilla allreduce in the first `start_powerSGD_iter` iterations.
if state.iter < state.start_powerSGD_iter:
state.maybe_increase_iter(bucket)
return default._allreduce_fut(group_to_use, input_tensor)
# Apply PowerSGD after `start_powerSGD_iter` iterations.
device = input_tensor.device
total_length = input_tensor.shape[0]
# View the input tensor as a 2D square-shape tensor, and pad 0s if necessary.
square_side_length = math.ceil(math.sqrt(total_length))
padded_total_length = square_side_length ** 2
input_tensor.resize_(padded_total_length)
input_tensor[total_length:padded_total_length].fill_(0)
# Incorporate the error from the previous state into the gradients.
bucket_index = bucket.get_index()
input_tensor_cp = None
if state.use_error_feedback:
if bucket_index in state.error_dict:
input_tensor.add_(state.error_dict[bucket_index])
else:
logging.info(
"A zero tensor of length {} that represents local error is created.".format(
padded_total_length
)
)
state.error_dict[bucket_index] = torch.zeros(
padded_total_length, device=device, dtype=input_tensor.dtype
)
# Keep a copy of the input tensor,
# so that we can compute the local error caused by compression later,
# by comparing this copy and the input tensor updated after decompression.
input_tensor_cp = torch.clone(input_tensor).detach()
matrix = input_tensor.view(square_side_length, square_side_length)
# Reuse P and Q from the previous iteration if possible.
# The memory spaces of P and Q need to be allocated in the first iteration when PowerSGD is applied.
if not state.warm_start or bucket_index not in state.p_memory_dict:
# If warm-start is disabled, low-rank tensors will be initialized at every step.
# Only log this if warm-start to avoid spamming.
if state.warm_start:
logging.info(
"Initializing low-rank tensors P and Q, each of which has a shape of {} x {}.".format(
square_side_length, state.matrix_approximation_rank
)
)
def create_low_rank_tensor(fill_random_values, rng):
"Returns a low-rank 2D tensor of square_side_length * matrix_approximation_rank."
if fill_random_values:
with torch.random.fork_rng(devices=[]):
# Fork this RNG to avoid changing the seed globally and affecting the random sampling
# anywhere else in the training.
# The seed makes sure that the initial random values are the same across all the DDP replicas.
# This seed should differ at every step.
# Since it is very slow to fork RNG state across all the CUDA devices,
# only fork on CPU and then move the generated tensor to the CUDA device.
torch.manual_seed(rng.randint(1_000_000_000))
return torch.randn(
square_side_length,
state.matrix_approximation_rank,
device="cpu",
dtype=input_tensor.dtype,
).to(device)
else:
return torch.empty(
square_side_length,
state.matrix_approximation_rank,
device=device,
dtype=input_tensor.dtype,
)
state.p_memory_dict[bucket_index] = create_low_rank_tensor(
fill_random_values=False, rng=state.rng
)
state.q_memory_dict[bucket_index] = create_low_rank_tensor(
fill_random_values=True, rng=state.rng
)
_orthogonalize(state.q_memory_dict[bucket_index], 0)
torch.matmul(
matrix, state.q_memory_dict[bucket_index], out=state.p_memory_dict[bucket_index]
)
allreduce_p_fut = dist.all_reduce(
state.p_memory_dict[bucket_index], group=group_to_use, async_op=True
).get_future()
def compute_q(fut):
state.p_memory_dict[bucket_index] = fut.value()[0]
_orthogonalize(state.p_memory_dict[bucket_index], 0)
torch.matmul(
matrix.t(),
state.p_memory_dict[bucket_index],
out=state.q_memory_dict[bucket_index],
)
# TODO: The above procedure does two matmul+allreduce steps per iteration --
# one left multiplication and one right multiplication.
# For warm-start, can take one such step at a time, and alternate between them.
return [
dist.all_reduce(
state.q_memory_dict[bucket_index], group=group_to_use, async_op=True
)
.get_future()
.wait()[0]
]
def decompress(fut):
state.q_memory_dict[bucket_index] = fut.value()[0].div_(world_size)
torch.matmul(
state.p_memory_dict[bucket_index],
state.q_memory_dict[bucket_index].t(),
out=matrix,
)
if state.use_error_feedback:
# Memorize the local errors.
state.error_dict[bucket_index] = input_tensor_cp - input_tensor
if torch.cuda.is_available():
torch.cuda.synchronize(device)
if not state.warm_start:
state.p_memory_dict.clear()
state.q_memory_dict.clear()
ret = input_tensor.resize_(total_length)
state.maybe_increase_iter(bucket)
return [ret]
return allreduce_p_fut.then(compute_q).then(decompress)
