2022-04-28 09:45:06 +00:00
|
|
|
import torch
|
|
|
|
from colossalai.tensor.op_wrapper import colo_op_impl
|
|
|
|
from colossalai.context import ParallelMode
|
|
|
|
from colossalai.nn.layer.parallel_1d._utils import split_forward_gather_backward, reduce_input, \
|
|
|
|
gather_forward_split_backward, reduce_grad
|
|
|
|
from colossalai.nn.layer.utils import divide
|
|
|
|
from colossalai.core import global_context as gpc
|
|
|
|
from packaging import version
|
2022-05-13 07:13:52 +00:00
|
|
|
from colossalai.tensor import ComputePattern, TensorSpec, ComputePattern, ParallelAction, ColoTensor, dist_spec
|
|
|
|
|
2022-04-28 09:45:06 +00:00
|
|
|
|
|
|
|
def colo_embedding_1Dcol(input_tensor: ColoTensor, weight: ColoTensor, args, kwargs) -> ColoTensor:
|
2022-04-29 06:10:05 +00:00
|
|
|
# embedding_1Dcol split the weight(lookup table) to (num_embeddings, embedding_dim/P)
|
2022-04-28 09:45:06 +00:00
|
|
|
# Gather splitted lookup table
|
2022-05-13 07:13:52 +00:00
|
|
|
parallel_action = weight.spec.get_action_by_compute_pattern(ComputePattern.TP1DCol)
|
|
|
|
input_tensor.to_dist_spec(dist_spec.replicate(weight.spec.get_process_group()))
|
2022-04-28 09:45:06 +00:00
|
|
|
|
2022-05-13 07:13:52 +00:00
|
|
|
output_parallel = torch.nn.functional.embedding(input_tensor.torch_tensor(), weight.torch_tensor(), *args, **kwargs)
|
|
|
|
output_spec = TensorSpec(
|
|
|
|
dist_spec.shard(weight.spec.get_process_group(), [-1], [weight.spec.get_process_group().size()]),
|
|
|
|
[ParallelAction(priority=1, parallel_mode=parallel_action.parallel_mode)])
|
|
|
|
output = ColoTensor.init_from_torch_tensor(output_parallel, spec=output_spec)
|
|
|
|
output.to_dist_spec(dist_spec.replicate(weight.spec.get_process_group()))
|
2022-04-28 09:45:06 +00:00
|
|
|
return output
|
|
|
|
|
2022-05-13 07:13:52 +00:00
|
|
|
|
2022-04-29 06:10:05 +00:00
|
|
|
def colo_embedding_1Drow(input_tensor: ColoTensor, weight: ColoTensor, args, kwargs) -> ColoTensor:
|
|
|
|
# embedding_1Drow split the weight(lookup table) to (num_embeddings/P, embedding_dim)
|
|
|
|
# Find index in this shard and mask those not here
|
|
|
|
# Reduce all
|
2022-05-13 07:13:52 +00:00
|
|
|
parallel_action = weight.spec.get_action_by_compute_pattern(ComputePattern.TP1DRow)
|
|
|
|
input_tensor.to_dist_spec(dist_spec.replicate(weight.spec.get_process_group()))
|
|
|
|
|
2022-04-29 06:10:05 +00:00
|
|
|
tensor_parallel_rank = gpc.get_local_rank(parallel_action.parallel_mode)
|
|
|
|
num_embeddings_per_partition = weight.size(0)
|
|
|
|
vocab_start_index = tensor_parallel_rank * num_embeddings_per_partition
|
|
|
|
vocab_end_index = vocab_start_index + num_embeddings_per_partition
|
|
|
|
|
|
|
|
# Build the mask.
|
|
|
|
input_mask = (input_tensor.torch_tensor() < vocab_start_index) | \
|
|
|
|
(input_tensor.torch_tensor() >= vocab_end_index)
|
|
|
|
# Mask the input.
|
|
|
|
# TODO(jzy) masked_input may be an activation managed by ColoTensor.
|
|
|
|
masked_input = input_tensor.torch_tensor().clone() - vocab_start_index
|
|
|
|
masked_input[input_mask] = 0
|
|
|
|
|
2022-05-13 07:13:52 +00:00
|
|
|
partial_output = torch.nn.functional.embedding(masked_input, weight.torch_tensor(), *args, **kwargs)
|
2022-04-29 06:10:05 +00:00
|
|
|
|
|
|
|
# Mask the output embedding.
|
|
|
|
partial_output[input_mask, :] = 0.
|
|
|
|
# Reduce across all the model parallel GPUs.
|
|
|
|
output = reduce_input(partial_output, parallel_action.parallel_mode)
|
2022-05-13 07:13:52 +00:00
|
|
|
output = ColoTensor.init_from_torch_tensor(output,
|
|
|
|
spec=TensorSpec(dist_spec.replicate(weight.spec.get_process_group())))
|
2022-04-29 06:10:05 +00:00
|
|
|
return output
|
|
|
|
|
2022-05-13 07:13:52 +00:00
|
|
|
|
2022-04-28 09:45:06 +00:00
|
|
|
@colo_op_impl(torch.nn.functional.embedding)
|
|
|
|
def colo_embedding(types, args, kwargs, pg):
|
|
|
|
"""Handles ``__torch_function__`` dispatch for ``torch.nn.functional.embedding``.
|
|
|
|
This method looks up an embedding table.
|
|
|
|
"""
|
|
|
|
input_tensor = args[0]
|
|
|
|
weight = args[1]
|
|
|
|
args = args[2:]
|
|
|
|
|
|
|
|
if not isinstance(input_tensor, ColoTensor):
|
|
|
|
input_tensor = ColoTensor.init_from_torch_tensor(input_tensor)
|
|
|
|
|
|
|
|
if not isinstance(weight, ColoTensor):
|
|
|
|
weight = ColoTensor.init_from_torch_tensor(weight)
|
2022-05-13 07:13:52 +00:00
|
|
|
|
2022-04-28 09:45:06 +00:00
|
|
|
# Handle differen parallel actions.
|
2022-05-13 07:13:52 +00:00
|
|
|
if not weight.has_spec(): # No Model Parallel Applied
|
2022-04-28 09:45:06 +00:00
|
|
|
input_tensor = input_tensor.torch_tensor()
|
|
|
|
weight = weight.torch_tensor()
|
|
|
|
output = torch.nn.functional.embedding(input_tensor, weight, *args, **kwargs)
|
|
|
|
return ColoTensor.init_from_torch_tensor(output)
|
2022-05-13 07:13:52 +00:00
|
|
|
elif weight.spec.num_action == 1: # Single Model Parallel Applied
|
|
|
|
compute_patterns = weight.spec.compute_patterns
|
|
|
|
if ComputePattern.TP1DRow in compute_patterns:
|
2022-04-29 06:10:05 +00:00
|
|
|
return colo_embedding_1Drow(input_tensor, weight, args, kwargs)
|
2022-05-13 07:13:52 +00:00
|
|
|
elif ComputePattern.TP1DCol in compute_patterns:
|
2022-04-28 09:45:06 +00:00
|
|
|
return colo_embedding_1Dcol(input_tensor, weight, args, kwargs)
|
|
|
|
else:
|
|
|
|
raise NotImplementedError
|
|
|
|
else:
|
|
|
|
raise NotImplementedError
|