def bf16_compress_hook(
process_group: dist.ProcessGroup,
bucket: dist.GradBucket) -> torch.futures.Future[torch.Tensor]:
"""
Warning: This API is experimental, and it requires NCCL version later than 2.9.6.
This DDP communication hook implements a simple gradient compression
approach that casts ``GradBucket`` tensor to half-precision
`Brain floating point format <https://en.wikipedia.org/wiki/Bfloat16_floating-point_format>`_ (``torch.bfloat16``)
and then divides it by the process group size.
It allreduces those ``bfloat16`` gradient tensors. Once compressed gradient
tensors are allreduced, the chained callback ``decompress`` casts it back to the input data type (such as ``float32``).
Example::
>>> ddp_model.register_comm_hook(process_group, bf16_compress_hook)
"""
group_to_use = process_group if process_group is not None else dist.group.WORLD
world_size = group_to_use.size()
compressed_tensor = bucket.buffer().to(torch.bfloat16).div_(world_size)
fut = dist.all_reduce(compressed_tensor, group=group_to_use,
async_op=True).get_future()
def decompress(fut):
decompressed_tensor = bucket.buffer()
# Decompress in place to reduce the peak memory.
# See: https://github.com/pytorch/pytorch/issues/45968
decompressed_tensor.copy_(fut.value()[0])
return decompressed_tensor
return fut.then(decompress)
def scale_by_grad_accum_steps_wrapper_hook(
hook_state, bucket: dist.GradBucket
) -> torch.futures.Future[torch.Tensor]:
bucket.set_buffer(bucket.buffer().div_(
args.gradient_accumulation_steps))
fut = hook(hook_state, bucket)
return fut
def _simple_hook(
self, state: object,
bucket: dist.GradBucket) -> torch.futures.Future[torch.Tensor]:
fut = torch.futures.Future()
fut.set_result(torch.ones_like(bucket.buffer()))
def fut_then(fut):
# Add ones to fut's result.
t = fut.value()
return t + torch.ones_like(t)
return fut.then(fut_then)
def post_localSGD_hook(
state: PostLocalSGDState,
bucket: dist.GradBucket) -> torch.futures.Future[torch.Tensor]:
"""
This DDP communication hook is used for running post-localSGD algorithm,
by combining with a model averaging component (e.g.,
:class:`~torch.distributed.algorithms.model_averaging.averagers.PeriodicModelAverager`)
that runs after the optimizer step.
Args:
state (PostLocalSGDState): State information to run post-localSGD.
Users mainly need to tune ``start_localSGD_iter`` to determine when to start local SGD.
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,
only exactly one tensor is stored in this bucket.
Returns:
Future handler of the communication, which updates the gradients in place.
Example::
>>> # xdoctest: +SKIP
>>> state = PostLocalSGDState(process_group=process_group, subgroup=subgroup,
start_localSGD_iter=10)
>>> ddp_model.register_comm_hook(state, post_localSGD_hook)
>>> # Also need to establish a model averaging module and run model averaging after ``optimizer.step()``.
>>> # Please refer to the examples in ``torch.distributed.algorithms.model_averaging.averagers`` module.
"""
global_group_to_use = (state.process_group if state.process_group
is not None else dist.group.WORLD)
# The input tensor is a flattened 1D tensor.
input_tensor = bucket.buffer()
# Run allreduce using `global_group_to_use` in the first `start_localSGD_iter` iterations.
if state.iter < state.start_localSGD_iter:
state.maybe_increase_iter(bucket)
return default._allreduce_fut(global_group_to_use, input_tensor)
# If `post_local_gradient_allreduce` is not set,
# then no gradient synchronization after the first `start_localSGD_iter` iterations.
if not state.post_local_gradient_allreduce:
fut: torch.futures.Future[torch.Tensor] = torch.futures.Future()
fut.set_result(input_tensor)
return fut
# Run allreduce using `subgroup` after the first `start_localSGD_iter` iterations.
# Note that by default, a separate subgroup for each node is created which
# causes an intra-node allreduce to be done at each training step.
# From this moment, model averaging should run after the optimizer step,
# to globally allreduce all the parameters.
if state.subgroup is None:
state.subgroup, _ = dist.new_subgroups()
return default._allreduce_fut(state.subgroup, input_tensor)
def allreduce_hook(
process_group: dist.ProcessGroup,
bucket: dist.GradBucket) -> torch.futures.Future[torch.Tensor]:
"""
This DDP communication hook just calls ``allreduce`` using ``GradBucket``
tensors. Once gradient tensors are aggregated across all workers, its ``then``
callback takes the mean and returns the result. If user registers this hook,
DDP results is expected to be same as the case where no hook was registered.
Hence, this won't change behavior of DDP and user can use this as a reference
or modify this hook to log useful information or any other purposes while
unaffecting DDP behavior.
Example::
>>> ddp_model.register_comm_hook(process_group, allreduce_hook)
"""
return _allreduce_fut(process_group, bucket.buffer())
def bf16_compress_wrapper_hook(
hook_state,
bucket: dist.GradBucket) -> torch.futures.Future[torch.Tensor]:
# Cast bucket tensor to BF16.
bucket.set_buffer(bucket.buffer().to(torch.bfloat16))
fut = hook(hook_state, bucket)
def decompress(fut):
decompressed_tensor = bucket.buffer()
# Decompress in place to reduce the peak memory.
# See: https://github.com/pytorch/pytorch/issues/45968
decompressed_tensor.copy_(fut.value())
return decompressed_tensor
# Decompress after hook has run.
return fut.then(decompress)
def noop_hook(_: Any,
bucket: dist.GradBucket) -> torch.futures.Future[torch.Tensor]:
"""
This DDP communication hook returns a future that wraps the input,
so it is a noop that does not incur any communication overheads.
This hook should **only** be used for headroom analysis of allreduce optimization,
instead of the normal gradient synchronization.
For example, if only less than 10% speedup of training time can be observed after this hook is registered,
it usually implies that allreduce is not a performance bottleneck for this case.
Such instrumentation can be particularly useful
if GPU traces cannot be easily retrieved or the trace analysis is complicated
some factors such as the overlap between allreduce and computation or the desynchronization across ranks.
Example::
>>> ddp_model.register_comm_hook(None, noop_hook)
"""
fut: torch.futures.Future[torch.Tensor] = torch.futures.Future()
fut.set_result(bucket.buffer())
return fut
def quantization_pertensor_hook(
process_group: dist.ProcessGroup,
bucket: dist.GradBucket) -> torch.futures.Future[torch.Tensor]:
"""
Applies the ``torch.quantize_per_tensor`` logic to DDP using ``allgather``
protocol. Workers first allgather the scale and zero point of their own
``GradBucket`` prior to the quantization. After all workers have that information,
the first ``then`` callback called ``quantize_and_allgather`` quantizes worker's
own gradient tensor, and uses ``allgather`` to communicate these accross all workers.
The final ``then`` callback called ``dequantize_and_aggregate``, dequantizes and
aggregates each quantized gradient tensor locally and returns the mean.
.. warning ::
This is experimental, and uses ``allgather`` protocol which is considerably slower than
``allreduce`` protocol. It works only with flattened grads.
Example::
>>> ddp_model.register_comm_hook(process_group, quantization_pertensor_hook)
"""
group_to_use = process_group if process_group is not None else dist.group.WORLD
rank = process_group.rank(
) if process_group is not None else dist.get_rank()
world_size = group_to_use.size()
tensor = bucket.buffer()
myObserver = torch.quantization.MinMaxObserver().cuda(tensor.device)
myObserver(tensor)
s, z = myObserver.calculate_qparams()
s_and_z = torch.FloatTensor([s, z]).cuda(tensor.device)
all_ranks_s_and_z = _get_allgather_out_list(s_and_z, world_size)
# First, allgather scale and zeros.
fut = dist.all_gather(all_ranks_s_and_z,
s_and_z,
group=group_to_use,
async_op=True).get_future()
def quantize_and_allgather(fut):
# Store scale and zeros accross all workers.
all_ranks_s_and_z = fut.wait()[0]
# All workers quantize their own ``GradBucket`` tensors.
quantized_tensor = _quantize_per_tensor_cuda(
tensor, all_ranks_s_and_z[rank][0], all_ranks_s_and_z[rank][1])
# Allgather quantized tensors.
fut = dist.all_gather(
_get_allgather_out_list(quantized_tensor, world_size),
quantized_tensor,
group=group_to_use,
async_op=True,
).get_future()
return fut.wait()
def dequantize_and_aggregate(fut):
all_ranks_quantized_tensor = fut.wait()[0]
aggregated_dequantized_tensor = torch.zeros_like(
all_ranks_quantized_tensor[0],
device=tensor.device,
dtype=torch.float32)
# Using previously allgathered scales and zeros, dequantize gradient tensors
# locally and then aggregate them.
for r, quantized_tensor in enumerate(all_ranks_quantized_tensor):
aggregated_dequantized_tensor += _dequantize_per_tensor_cuda(
quantized_tensor, all_ranks_s_and_z[r][0],
all_ranks_s_and_z[r][1])
return aggregated_dequantized_tensor / world_size
return fut.then(quantize_and_allgather).then(dequantize_and_aggregate)
def quantization_perchannel_hook(
process_group: dist.ProcessGroup,
bucket: dist.GradBucket,
bucket_size=512) -> torch.futures.Future[torch.Tensor]:
"""
Applies the ``torch.quantize_per_channel`` logic to DDP using ``allgather``
protocol. Compared to pertensor, the main motivation of perchannel is
for considerably large tensors such as a tensor that contains 6 million
elements quantizing per a bucket size of 512 (or 128) elements may significantly
increase the resolution.
It first splits ``GradBucket`` tensor into multiple chunks (channels) of ``bucket_size``
elements. Then, workers allgather the scales and zero points of their own
``GradBucket`` prior to the quantization. After all workers have that information,
the first ``then`` callback called ``quantize_and_allgather`` quantizes worker's
own gradient tensor, and uses ``allgather`` to communicate these accross all workers.
The final ``then`` callback called ``dequantize_and_aggregate``, dequantizes, flattens, and
aggregates each quantized gradient tensor locally and returns the mean.
.. warning ::
This is experimental, and uses ``allgather`` protocol which is considerably slower than
``allreduce`` protocol. It works only with flattened grads.
Example::
>>> ddp_model.register_comm_hook(process_group, quantization_perchannel_hook)
"""
group_to_use = process_group if process_group is not None else dist.group.WORLD
rank = process_group.rank(
) if process_group is not None else dist.get_rank()
world_size = group_to_use.size()
tensor = bucket.buffer()
tensor_in_channels = (nn.functional.pad(
input=tensor,
pad=(0, bucket_size - len(tensor) % bucket_size),
mode="constant",
value=0,
).view(-1, bucket_size).cuda(tensor.device))
myPerChannelObserver = torch.quantization.PerChannelMinMaxObserver().cuda(
tensor.device)
myPerChannelObserver(tensor_in_channels)
s_ch, z_ch = myPerChannelObserver.calculate_qparams()
s_and_z = torch.stack((s_ch, z_ch)).cuda(tensor.device)
all_ranks_s_and_z = _get_allgather_out_list(s_and_z, world_size)
# First, allgather scale and zeros.
fut = dist.all_gather(all_ranks_s_and_z,
s_and_z,
group=group_to_use,
async_op=True).get_future()
def quantize_and_allgather(fut):
# Store scale and zeros accross all workers.
all_ranks_s_and_z = fut.wait()[0]
# All workers quantize their corresponding ``GradBucket`` tensors.
quantized_tensor = _quantize_per_channel_cuda(
tensor_in_channels,
all_ranks_s_and_z[rank, 0, :],
all_ranks_s_and_z[rank, 1, :],
)
# Allgather quantized tensors.
fut = dist.all_gather(
_get_allgather_out_list(quantized_tensor, world_size),
quantized_tensor,
group=group_to_use,
async_op=True,
).get_future()
return fut.wait()
def dequantize_and_aggregate(fut):
all_ranks_quantized_tensor = fut.wait()[0]
aggregated_dequantized_tensor = torch.zeros_like(
all_ranks_quantized_tensor[0],
device=tensor.device,
dtype=torch.float32)
# Using previously allgathered scales and zeros, dequantize gradient tensors
# locally and then aggregate them.
for r, quantized_tensor in enumerate(all_ranks_quantized_tensor):
aggregated_dequantized_tensor += _dequantize_per_channel_cuda(
quantized_tensor, all_ranks_s_and_z[r][0],
all_ranks_s_and_z[r][1])
return (torch.flatten(aggregated_dequantized_tensor).cuda(
tensor.device)[:tensor.size()[0]] / world_size)
return fut.then(quantize_and_allgather).then(dequantize_and_aggregate)
def batched_powerSGD_hook(
state: PowerSGDState,
bucket: dist.GradBucket) -> torch.futures.Future[torch.Tensor]:
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,
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: B950
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.buffer()
# 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]
state.total_numel_before_compression += total_length
# View the input tensor as a 2D square-shape tensor, and pad 0s if necessary.
square_side_length = math.ceil(math.sqrt(total_length))
state.total_numel_after_compression += (square_side_length *
state.matrix_approximation_rank *
2)
padded_total_length = square_side_length**2
input_tensor.resize_(padded_total_length)
input_tensor[total_length:padded_total_length].fill_(0)
_report_compression_stats(bucket, state)
# Incorporate the error from the previous state into the gradients.
bucket_index = bucket.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])
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])
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().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
# Removing this seemingly unnecessary sync somehow may cause faliures.
# See: https://github.com/pytorch/pytorch/pull/54838
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)
def powerSGD_hook(
state: PowerSGDState,
bucket: dist.GradBucket) -> torch.futures.Future[torch.Tensor]:
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 a list of per-parameter tensors, and divides all the tensors into two groups:
1.1 The tensors that should be compressed before allreduce, because the compression can give enough saving in bandwidth.
1.2 Rest of the tensors will be directly allreduced without compression, including all the vector tensors (for biases).
2. Handles uncompressed tensors:
2.1. Allocate contiguous memory for those uncompressed tensors, and allreduces all the uncompressed tensors as a batch, without compression;
2.2. Copies the individual uncompressed tensors from the contiguous memory back to the input tensor.
3. Handles the tensors that should be compressed by PowerSGD compression:
3.1. For each 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 compressed 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``, ``start_powerSGD_iter``
and ``min_compression_rate``.
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,
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, min_compression_rate=0.5)
>>> ddp_model.register_comm_hook(state, powerSGD_hook)
""" # noqa: B950
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.buffer()
# 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.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 = bucket.gradients()
# Step I: Divide all the tensors into two groups,
# one will be compressed before allreduce and the other will be directly allreduced without compression.
tensors_to_compress, uncompressed_tensors = [], []
total_Ps_size = 0
total_Qs_size = 0
for tensor in tensors:
matrix = tensor.view(tensor.shape[0], -1)
n, m = matrix.shape
matrix_approximation_rank = min(n, m, state.matrix_approximation_rank)
compress_test = _should_compress(n, m, matrix_approximation_rank,
state.min_compression_rate)
state.total_numel_before_compression += compress_test[1]
if compress_test[0]:
tensors_to_compress.append(matrix)
total_Ps_size += n * matrix_approximation_rank
total_Qs_size += m * matrix_approximation_rank
state.total_numel_after_compression += compress_test[2]
else:
uncompressed_tensors.append(tensor)
state.total_numel_after_compression += compress_test[1]
_report_compression_stats(bucket, state)
# Step II: Handle uncompressed tensors.
# Allocate contiguous memory for these tensors to allreduce efficiently.
uncompressed_tensors_memory = (
torch.cat([tensor.view(-1) for tensor in uncompressed_tensors]) if
uncompressed_tensors else torch.tensor([], device=device, dtype=dtype))
# Step III: Handle the tensors that should be compressed.
# Allocate contiguous memory for Ps and Qs to allreduce efficiently.
# 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 tensors_to_compress:
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, state.orthogonalization_epsilon)
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, state.orthogonalization_epsilon)
# Compute Ps.
for tensor, q, p in zip(tensors_to_compress, qs, ps):
torch.matmul(tensor, q, out=p)
# This allreduce is only applied to uncompressed tensors,
# so it should have been kicked off before the above computation on the compressed tensors to hide more communication costs.
# However, this somehow requires a separate future chain at this time.
allreduce_contiguous_uncompressed_tensors_fut = dist.all_reduce(
uncompressed_tensors_memory, group=group_to_use,
async_op=True).get_future()
def unpack_uncompressed_tensors_and_allreduce_ps(fut):
uncompressed_tensors_memory = fut.value()[0].div_(world_size)
idx = 0
for tensor in uncompressed_tensors:
tensor.copy_(
uncompressed_tensors_memory[idx:idx +
tensor.numel()].view_as(tensor))
idx += tensor.numel()
# 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()
for p in ps:
_orthogonalize(p, state.orthogonalization_epsilon)
# Compute Qs.
for tensor, p, q in zip(tensors_to_compress, 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().div_(world_size)
for p, q, tensor in zip(ps, qs, tensors_to_compress):
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_uncompressed_tensors_fut.then(
unpack_uncompressed_tensors_and_allreduce_ps).then(compute_qs).then(
decompress))
