Skip to content

GitLab

Unverified Commit efba0f44 authored by Hongxin Liu's avatar Hongxin Liu Committed by GitHub
Browse files

Merge pull request #4612 from hpcaitech/feature/shardformer 

[shardformer] update hybrid parallel plugin and fix bugs
parents ac178ca5 fae6c92e
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 1839 additions and 245 deletions
colossalai/shardformer/layer/linear.py
View file @ efba0f44
......@@ -24,6 +24,8 @@ from colossalai.tensor.d_tensor.api import (
from ._operation import (
gather_forward_split_backward,
linear_gather_forward_reducescatter_backward,
linear_reducescatter_forward_gather_backward,
linear_with_async_comm,
reduce_forward,
split_forward_gather_backward,
......@@ -50,6 +52,8 @@ class Linear1D_Col(ParallelModule):
gather_output (bool, optional): If true, call all-gather on output and make Y available
to all GPUs, otherwise, every GPU will have its output
which is :math:`Y_i = XA_i`, defaults to False
seq_parallel (`bool`): If set to ``True``, it will use sequence parallel, defaults to False.
overlap (`bool`): If set to ``True``, it will overlap input all-gather with gradient computation during backward, defaults to False.
skip_bias_add (bool): If set to ``True``, it will skip bias add for linear layer,
which is preserved for kernel fusion, defaults to False
weight_initializer (`typing.Callable`):
......@@ -69,6 +73,9 @@ class Linear1D_Col(ParallelModule):
device: torch.device = None,
process_group: ProcessGroup = None,
gather_output: bool = False,
seq_parallel: bool = False,
seq_parallel_dim: int = 1,
overlap: torch.cuda.Stream = None,
skip_bias_add: bool = False,
weight: Optional[Parameter] = None,
bias_: Optional[Parameter] = None,
......@@ -80,6 +87,9 @@ class Linear1D_Col(ParallelModule):
self.in_features = in_features
self.out_features = out_features
self.gather_output = gather_output
self.seq_parallel = seq_parallel
self.seq_parallel_dim = seq_parallel_dim
self.overlap = overlap
self.skip_bias_add = skip_bias_add
self.device = device
self.process_group = process_group
......@@ -180,7 +190,12 @@ class Linear1D_Col(ParallelModule):
# Matrix multiply.
bias = self.bias if not self.skip_bias_add else None
output_parallel = linear_with_async_comm(input_parallel, self.weight, bias, self.process_group, True)
if self.seq_parallel:
output_parallel = linear_gather_forward_reducescatter_backward(input_parallel, self.weight, bias,
self.process_group, True,
self.seq_parallel_dim, self.overlap)
else:
output_parallel = linear_with_async_comm(input_parallel, self.weight, bias, self.process_group, True)
if self.gather_output:
# All-gather across the partitions.
......@@ -203,6 +218,8 @@ class Linear1D_Row(ParallelModule):
bias (bool, optional): If set to ``False``, the layer will not learn an additive bias, defaults to ``True``.
dtype (`torch.dtype`): The dtype of parameters, defaults to None.
parallel_input (bool): If set to ``True``, it's assumed that the input is split, defaults to False.
process_group (`torch.distributed.ProcessGroup`): The process group to be used for weight sharding and communication, defaults to None.
seq_parallel (`bool`): If set to ``True``, it will use sequence parallel, defaults to False.
skip_bias_add (bool): If set to ``True``, it will skip bias add for linear layer,
which is preserved for kernel fusion, defaults to False
weight_initializer (:class:`typing.Callable`, optional):
......@@ -221,6 +238,8 @@ class Linear1D_Row(ParallelModule):
dtype: torch.dtype = None,
device: torch.device = None,
process_group: ProcessGroup = None,
seq_parallel: bool = False,
seq_parallel_dim: int = 1,
parallel_input: bool = True,
skip_bias_add: bool = False,
weight: Optional[Parameter] = None,
......@@ -238,6 +257,8 @@ class Linear1D_Row(ParallelModule):
self.parallel_input = parallel_input
self.skip_bias_add = skip_bias_add
self.process_group = process_group
self.seq_parallel = seq_parallel
self.seq_parallel_dim = seq_parallel_dim
self.num_partitions = dist.get_world_size(self.process_group)
if skip_bias_add and not bias:
......@@ -373,7 +394,11 @@ class Linear1D_Row(ParallelModule):
output = torch.cat(output_parallel_list, dim=-1)
else:
output_parallel = F.linear(input_, self.weight)
output = reduce_forward(output_parallel, self.process_group)
if self.seq_parallel:
output = linear_reducescatter_forward_gather_backward(output_parallel, self.process_group,
self.seq_parallel_dim)
else:
output = reduce_forward(output_parallel, self.process_group)
if not self.skip_bias_add:
if self.bias is not None:
......
colossalai/shardformer/layer/parallel_module.py
View file @ efba0f44
......@@ -10,6 +10,7 @@ import torch.nn as nn
from torch.distributed import ProcessGroup
from torch.nn.modules.module import _EXTRA_STATE_KEY_SUFFIX, Module
from colossalai.checkpoint_io.utils import gather_distributed_param
from colossalai.tensor.d_tensor import (
distribute_tensor,
distribute_tensor_with_customization,
......@@ -56,13 +57,7 @@ class ParallelModule(nn.Module, ABC):
"""
for name, param in self._parameters.items():
if param is not None:
param_ = param if keep_vars else param.detach()
if is_distributed_tensor(param_):
destination[prefix + name] = to_global(param_)
elif is_customized_distributed_tensor(param_):
destination[prefix + name] = to_global_for_customized_distributed_tensor(param_)
else:
destination[prefix + name] = param_
destination[prefix + name] = gather_distributed_param(param, keep_vars=keep_vars)
for name, buf in self._buffers.items():
if buf is not None and name not in self._non_persistent_buffers_set:
......
colossalai/shardformer/layer/qkv_fused_linear.py
View file @ efba0f44
......@@ -25,7 +25,9 @@ from colossalai.tensor.d_tensor.api import (
from ._operation import (
gather_forward_split_backward,
linear_reducescatter_forward_gather_backward,
linear_with_async_comm,
matmul_gather_forward_reducescatter_backward,
matmul_with_async_comm,
reduce_backward,
reduce_forward,
......@@ -150,6 +152,7 @@ class GPT2FusedLinearConv1D_Col(ParallelModule):
device (`torch.device`): The device of parameters, defaults to None.
n_fused (int): The number items fused, defaults to 3 (QKV).
process_group (`torch.distributed.ProcessGroup`): The process group to be used for weight sharding and communication, defaults to None.
seq_parallel (`bool`): If set to ``True``, it will use sequence parallel, defaults to False.
gather_output (bool, optional): If true, call all-gather on output and make Y available
to all GPUs, otherwise, every GPU will have its output
which is :math:`Y_i = XA_i`, defaults to False
......@@ -173,6 +176,8 @@ class GPT2FusedLinearConv1D_Col(ParallelModule):
process_group: ProcessGroup = None,
async_communication: bool = False,
gather_output: bool = False,
seq_parallel: bool = False,
overlap: bool = False,
skip_bias_add: bool = False,
n_fused: int = 3,
weight: Optional[Parameter] = None,
......@@ -185,6 +190,8 @@ class GPT2FusedLinearConv1D_Col(ParallelModule):
self.in_features = in_features
self.out_features = out_features
self.gather_output = gather_output
self.seq_parallel = seq_parallel
self.overlap = overlap
self.skip_bias_add = skip_bias_add
self.device = device
self.n_fused = n_fused
......@@ -296,15 +303,19 @@ class GPT2FusedLinearConv1D_Col(ParallelModule):
assert input_.shape[-1] == self.weight.shape[0], \
'Invalid shapes in Linear1D_Col forward: input={}, weight={}. Expected last dim of input {}.'.format(
input_.shape, self.weight.shape, self.weight.shape[-1])
# Set up backprop all-reduce.
input_parallel = reduce_backward(input_, self.process_group)
# input_parallel = input_
# Matrix multiply.
bias = self.bias if not self.skip_bias_add else None
output_parallel = matmul_with_async_comm(input_parallel, self.weight, bias, self.process_group,
self.async_communication)
if self.seq_parallel:
input_parallel = input_
output_parallel = matmul_gather_forward_reducescatter_backward(input_parallel, self.weight, bias,
self.process_group, True, 1, self.overlap)
else:
# Set up backprop all-reduce.
input_parallel = reduce_backward(input_, self.process_group)
output_parallel = matmul_with_async_comm(input_parallel, self.weight, bias, self.process_group,
self.async_communication)
if self.gather_output:
# All-gather across the partitions.
......@@ -329,6 +340,7 @@ class GPT2FusedLinearConv1D_Row(ParallelModule):
dtype (`torch.dtype`): The dtype of parameters, defaults to None.
parallel_input (bool): If set to ``True``, it's assumed that the input is split, defaults to False.
skip_bias_add (bool): If set to ``True``, it will skip bias add for linear layer,
seq_parallel (`bool`): If set to ``True``, it will use sequence parallel, defaults to False.
which is preserved for kernel fusion, defaults to False
weight_initializer (:class:`typing.Callable`, optional):
The initializer of weight, defaults to kaiming uniform initializer.
......@@ -346,6 +358,7 @@ class GPT2FusedLinearConv1D_Row(ParallelModule):
dtype: torch.dtype = None,
device: torch.device = None,
process_group: ProcessGroup = None,
seq_parallel: bool = False,
parallel_input: bool = True,
skip_bias_add: bool = False,
weight: Optional[Parameter] = None,
......@@ -363,6 +376,7 @@ class GPT2FusedLinearConv1D_Row(ParallelModule):
self.parallel_input = parallel_input
self.skip_bias_add = skip_bias_add
self.process_group = process_group
self.seq_parallel = seq_parallel
self.num_partitions = dist.get_world_size(self.process_group)
if skip_bias_add and not bias:
......@@ -499,7 +513,10 @@ class GPT2FusedLinearConv1D_Row(ParallelModule):
output = torch.cat(output_parallel_list, dim=-1)
else:
output_parallel = torch.matmul(input_, self.weight)
output = reduce_forward(output_parallel, self.process_group)
if self.seq_parallel:
output = linear_reducescatter_forward_gather_backward(output_parallel, self.process_group, 1)
else:
output = reduce_forward(output_parallel, self.process_group)
if not self.skip_bias_add:
if self.bias is not None:
......
colossalai/shardformer/modeling/bert.py
View file @ efba0f44
This diff is collapsed.
colossalai/shardformer/modeling/bloom.py
View file @ efba0f44
......@@ -23,6 +23,10 @@ from transformers.models.bloom.modeling_bloom import (
from transformers.utils import logging
from colossalai.pipeline.stage_manager import PipelineStageManager
from colossalai.shardformer.layer._operation import gather_forward_split_backward, split_forward_gather_backward
from colossalai.shardformer.shard import ShardConfig
logger = logging.get_logger(__name__)
def build_bloom_alibi_tensor_fn(process_group: ProcessGroup) -> torch.Tensor:
......@@ -111,6 +115,7 @@ class BloomPipelineForwards:
stage_manager: Optional[PipelineStageManager] = None,
hidden_states: Optional[torch.FloatTensor] = None,
stage_index: Optional[List[int]] = None,
shard_config: ShardConfig = None,
**deprecated_arguments,
) -> Union[Tuple[torch.Tensor, ...], 'BaseModelOutputWithPastAndCrossAttentions']:
......@@ -205,6 +210,13 @@ class BloomPipelineForwards:
past_key_values_length=past_key_values_length,
)
# split the input tensor along sequence dimension
# [batch_size, seq_len, hidden_size] -> [batch_size, seq_len/TP_size, hidden_size]
if shard_config.enable_sequence_parallelism:
hidden_states = split_forward_gather_backward(hidden_states,
dim=1,
process_group=shard_config.tensor_parallel_process_group)
start_idx, end_idx = stage_index[0], stage_index[1]
for i, (block, layer_past) in enumerate(zip(self.h[start_idx:end_idx], past_key_values[start_idx:end_idx]),
start=start_idx):
......@@ -248,6 +260,12 @@ class BloomPipelineForwards:
all_self_attentions = all_self_attentions + \
(outputs[2 if use_cache else 1],)
# When sequence parallelism done, gather the output tensor in forward and split it in backward
if shard_config.enable_sequence_parallelism:
hidden_states = gather_forward_split_backward(hidden_states,
dim=1,
process_group=shard_config.tensor_parallel_process_group)
if stage_manager.is_last_stage():
# Add last hidden state
hidden_states = self.ln_f(hidden_states)
......@@ -287,6 +305,7 @@ class BloomPipelineForwards:
stage_manager: Optional[PipelineStageManager] = None,
hidden_states: Optional[torch.FloatTensor] = None,
stage_index: Optional[List[int]] = None,
shard_config: ShardConfig = None,
**deprecated_arguments):
r"""
labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
......@@ -327,7 +346,8 @@ class BloomPipelineForwards:
return_dict=return_dict,
stage_manager=stage_manager,
hidden_states=hidden_states,
stage_index=stage_index)
stage_index=stage_index,
shard_config=shard_config)
past_key_values = None
all_hidden_states = None
all_self_attentions = None
......@@ -380,6 +400,7 @@ class BloomPipelineForwards:
stage_manager: Optional[PipelineStageManager] = None,
hidden_states: Optional[torch.FloatTensor] = None,
stage_index: Optional[List[int]] = None,
shard_config: ShardConfig = None,
**deprecated_arguments,
):
r"""
......@@ -424,6 +445,7 @@ class BloomPipelineForwards:
stage_manager=stage_manager,
hidden_states=hidden_states,
stage_index=stage_index,
shard_config=shard_config,
)
past_key_values = None
all_hidden_states = None
......@@ -503,6 +525,7 @@ class BloomPipelineForwards:
stage_manager: Optional[PipelineStageManager] = None,
hidden_states: Optional[torch.FloatTensor] = None,
stage_index: Optional[List[int]] = None,
shard_config: ShardConfig = None,
**deprecated_arguments,
):
r"""
......@@ -547,6 +570,7 @@ class BloomPipelineForwards:
stage_manager=stage_manager,
hidden_states=hidden_states,
stage_index=stage_index,
shard_config=shard_config,
)
past_key_values = None
all_hidden_states = None
......@@ -597,6 +621,7 @@ class BloomPipelineForwards:
stage_manager: Optional[PipelineStageManager] = None,
hidden_states: Optional[torch.FloatTensor] = None,
stage_index: Optional[List[int]] = None,
shard_config: ShardConfig = None,
):
r"""
start_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
......@@ -632,6 +657,7 @@ class BloomPipelineForwards:
stage_manager=stage_manager,
hidden_states=hidden_states,
stage_index=stage_index,
shard_config=shard_config,
)
past_key_values = None
all_hidden_states = None
......@@ -700,8 +726,7 @@ def get_bloom_flash_attention_forward(enabel_jit_fused=False):
fused_qkv = self.query_key_value(hidden_states)
(query_layer, key_layer, value_layer) = self._split_heads(fused_qkv)
batch_size, tgt_len, _ = hidden_states.size()
assert tgt_len % 4 == 0, "Flash Attention Error: The sequence length should be a multiple of 4."
batch_size, tgt_len, _ = query_layer.size()
_, kv_length, _, _ = key_layer.size()
......@@ -896,3 +921,156 @@ def get_jit_fused_bloom_gelu_forward():
return self.bloom_gelu_forward(x, bias)
return forward
def get_bloom_sequence_parallel_forward_fn(shard_config: ShardConfig):
from transformers import BloomModel
def forward(
self: BloomModel,
input_ids: Optional[torch.LongTensor] = None,
past_key_values: Optional[Tuple[Tuple[torch.Tensor, torch.Tensor], ...]] = None,
attention_mask: Optional[torch.Tensor] = None,
head_mask: Optional[torch.LongTensor] = None,
inputs_embeds: Optional[torch.LongTensor] = None,
use_cache: Optional[bool] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
**deprecated_arguments,
) -> Union[Tuple[torch.Tensor, ...], BaseModelOutputWithPastAndCrossAttentions]:
if deprecated_arguments.pop("position_ids", False) is not False:
# `position_ids` could have been `torch.Tensor` or `None` so defaulting pop to `False` allows to detect if users were passing explicitly `None`
warnings.warn(
"`position_ids` have no functionality in BLOOM and will be removed in v5.0.0. You can safely ignore"
" passing `position_ids`.",
FutureWarning,
)
if len(deprecated_arguments) > 0:
raise ValueError(f"Got unexpected arguments: {deprecated_arguments}")
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
output_hidden_states = (output_hidden_states
if output_hidden_states is not None else self.config.output_hidden_states)
use_cache = use_cache if use_cache is not None else self.config.use_cache
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
if input_ids is not None and inputs_embeds is not None:
raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
elif input_ids is not None:
batch_size, seq_length = input_ids.shape
elif inputs_embeds is not None:
batch_size, seq_length, _ = inputs_embeds.shape
else:
raise ValueError("You have to specify either input_ids or inputs_embeds")
if past_key_values is None:
past_key_values = tuple([None] * len(self.h))
# Prepare head mask if needed
# 1.0 in head_mask indicate we keep the head
# attention_probs has shape batch_size x num_heads x N x N
# head_mask has shape n_layer x batch x num_heads x N x N
head_mask = self.get_head_mask(head_mask, self.config.n_layer)
if inputs_embeds is None:
inputs_embeds = self.word_embeddings(input_ids)
hidden_states = self.word_embeddings_layernorm(inputs_embeds)
presents = () if use_cache else None
all_self_attentions = () if output_attentions else None
all_hidden_states = () if output_hidden_states else None
if self.gradient_checkpointing and self.training:
if use_cache:
logger.warning_once(
"`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`...")
use_cache = False
# Compute alibi tensor: check build_alibi_tensor documentation
seq_length_with_past = seq_length
past_key_values_length = 0
if past_key_values[0] is not None:
past_key_values_length = past_key_values[0][0].shape[2]
seq_length_with_past = seq_length_with_past + past_key_values_length
if attention_mask is None:
attention_mask = torch.ones((batch_size, seq_length_with_past), device=hidden_states.device)
else:
attention_mask = attention_mask.to(hidden_states.device)
alibi = self.build_alibi_tensor(attention_mask, self.num_heads, dtype=hidden_states.dtype)
causal_mask = self._prepare_attn_mask(
attention_mask,
input_shape=(batch_size, seq_length),
past_key_values_length=past_key_values_length,
)
# split the input tensor along sequence dimension
# [batch_size, seq_len, hidden_size] -> [batch_size, seq_len/TP_size, hidden_size]
hidden_states = split_forward_gather_backward(hidden_states,
dim=1,
process_group=shard_config.tensor_parallel_process_group)
for i, (block, layer_past) in enumerate(zip(self.h, past_key_values)):
if output_hidden_states:
all_hidden_states = all_hidden_states + (hidden_states,)
if self.gradient_checkpointing and self.training:
def create_custom_forward(module):
def custom_forward(*inputs):
# None for past_key_value
return module(*inputs, use_cache=use_cache, output_attentions=output_attentions)
return custom_forward
outputs = torch.utils.checkpoint.checkpoint(
create_custom_forward(block),
hidden_states,
alibi,
causal_mask,
layer_past,
head_mask[i],
)
else:
outputs = block(
hidden_states,
layer_past=layer_past,
attention_mask=causal_mask,
head_mask=head_mask[i],
use_cache=use_cache,
output_attentions=output_attentions,
alibi=alibi,
)
hidden_states = outputs[0]
if use_cache is True:
presents = presents + (outputs[1],)
if output_attentions:
all_self_attentions = all_self_attentions + (outputs[2 if use_cache else 1],)
# When sequence parallelism done, gather the output tensor in forward and split it in backward
hidden_states = gather_forward_split_backward(hidden_states,
dim=1,
process_group=shard_config.tensor_parallel_process_group)
# Add last hidden state
hidden_states = self.ln_f(hidden_states)
if output_hidden_states:
all_hidden_states = all_hidden_states + (hidden_states,)
if not return_dict:
return tuple(v for v in [hidden_states, presents, all_hidden_states, all_self_attentions] if v is not None)
return BaseModelOutputWithPastAndCrossAttentions(
last_hidden_state=hidden_states,
past_key_values=presents,
hidden_states=all_hidden_states,
attentions=all_self_attentions,
)
return forward
colossalai/shardformer/modeling/chatglm.py → colossalai/shardformer/modeling/chatglm2.py
View file @ efba0f44
......@@ -9,6 +9,8 @@ from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutpu
from transformers.utils import logging
from colossalai.pipeline.stage_manager import PipelineStageManager
from colossalai.shardformer import ShardConfig
from colossalai.shardformer.layer._operation import gather_forward_split_backward, split_forward_gather_backward
from colossalai.shardformer.modeling.chatglm2_6b.configuration_chatglm import ChatGLMConfig
from colossalai.shardformer.modeling.chatglm2_6b.modeling_chatglm import (
ChatGLMForConditionalGeneration,
......@@ -146,6 +148,7 @@ class ChatGLMPipelineForwards:
stage_manager: Optional[PipelineStageManager] = None,
hidden_states: Optional[torch.FloatTensor] = None,
stage_index: Optional[List[int]] = None,
shard_config: ShardConfig = None,
):
logger = logging.get_logger(__name__)
output_hidden_states = (output_hidden_states
......@@ -198,6 +201,11 @@ class ChatGLMPipelineForwards:
all_self_attentions = None
all_hidden_states = () if output_hidden_states else None
start_idx, end_idx = stage_index[0], stage_index[1]
if shard_config.enable_sequence_parallelism:
hidden_states = split_forward_gather_backward(hidden_states,
dim=0,
process_group=shard_config.tensor_parallel_process_group)
for idx in range(start_idx, end_idx):
layer = self.encoder._get_layer(idx)
if output_hidden_states:
......@@ -214,6 +222,11 @@ class ChatGLMPipelineForwards:
hidden_states, kv_cache = layer_ret
if use_cache:
presents = presents + (kv_cache,)
if shard_config.enable_sequence_parallelism:
hidden_states = gather_forward_split_backward(hidden_states,
dim=0,
process_group=shard_config.tensor_parallel_process_group)
if output_hidden_states:
all_hidden_states = all_hidden_states + (hidden_states,)
if stage_manager.is_last_stage():
......@@ -233,23 +246,22 @@ class ChatGLMPipelineForwards:
return {'hidden_states': hidden_states}
@staticmethod
def chatglm_for_conditional_generation_forward(
self: ChatGLMForConditionalGeneration,
input_ids: Optional[torch.Tensor] = None,
position_ids: Optional[torch.Tensor] = None,
attention_mask: Optional[torch.Tensor] = None,
past_key_values: Optional[Tuple[torch.FloatTensor]] = None,
inputs_embeds: Optional[torch.Tensor] = None,
labels: Optional[torch.Tensor] = None,
use_cache: Optional[bool] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
return_last_logit: Optional[bool] = False,
stage_manager: Optional[PipelineStageManager] = None,
hidden_states: Optional[torch.FloatTensor] = None,
stage_index: Optional[List[int]] = None,
):
def chatglm_for_conditional_generation_forward(self: ChatGLMForConditionalGeneration,
input_ids: Optional[torch.Tensor] = None,
position_ids: Optional[torch.Tensor] = None,
attention_mask: Optional[torch.Tensor] = None,
past_key_values: Optional[Tuple[torch.FloatTensor]] = None,
inputs_embeds: Optional[torch.Tensor] = None,
labels: Optional[torch.Tensor] = None,
use_cache: Optional[bool] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
return_last_logit: Optional[bool] = False,
stage_manager: Optional[PipelineStageManager] = None,
hidden_states: Optional[torch.FloatTensor] = None,
stage_index: Optional[List[int]] = None,
shard_config: ShardConfig = None):
logger = logging.get_logger(__name__)
use_cache = use_cache if use_cache is not None else self.config.use_cache
return_dict = (return_dict if return_dict is not None else self.config.use_return_dict)
......@@ -266,6 +278,7 @@ class ChatGLMPipelineForwards:
stage_manager=stage_manager,
hidden_states=hidden_states,
stage_index=stage_index,
shard_config=shard_config,
)
if stage_manager.is_last_stage():
hidden_states = transformer_outputs[0]
......@@ -296,3 +309,91 @@ class ChatGLMPipelineForwards:
)
else:
return transformer_outputs
def get_chatglm_sequence_parallel_forward_fn(shard_config: ShardConfig):
def forward(
self,
input_ids,
position_ids: Optional[torch.Tensor] = None,
attention_mask: Optional[torch.BoolTensor] = None,
full_attention_mask: Optional[torch.BoolTensor] = None,
past_key_values: Optional[Tuple[Tuple[torch.Tensor, torch.Tensor], ...]] = None,
inputs_embeds: Optional[torch.Tensor] = None,
use_cache: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
):
output_hidden_states = (output_hidden_states
if output_hidden_states is not None else self.config.output_hidden_states)
use_cache = use_cache if use_cache is not None else self.config.use_cache
return_dict = (return_dict if return_dict is not None else self.config.use_return_dict)
batch_size, seq_length = input_ids.shape
if inputs_embeds is None:
inputs_embeds = self.embedding(input_ids)
if self.pre_seq_len is not None:
if past_key_values is None:
past_key_values = self.get_prompt(
batch_size=batch_size,
device=input_ids.device,
dtype=inputs_embeds.dtype,
)
if attention_mask is not None:
attention_mask = torch.cat(
[
attention_mask.new_ones((batch_size, self.pre_seq_len)),
attention_mask,
],
dim=-1,
)
if full_attention_mask is None:
if (attention_mask is not None and not attention_mask.all()) or (past_key_values and seq_length != 1):
full_attention_mask = self.get_masks(input_ids, past_key_values, padding_mask=attention_mask)
# Rotary positional embeddings
rotary_pos_emb = self.rotary_pos_emb(self.seq_length)
if position_ids is not None:
rotary_pos_emb = rotary_pos_emb[position_ids]
else:
rotary_pos_emb = rotary_pos_emb[None, :seq_length]
rotary_pos_emb = rotary_pos_emb.transpose(0, 1).contiguous()
# Run encoder.
# [seq_len, batch_size, hidden_size] -> [seq_len/TP_size, batch_size, hidden_size]
inputs_embeds = split_forward_gather_backward(inputs_embeds,
dim=0,
process_group=shard_config.tensor_parallel_process_group)
hidden_states, presents, all_hidden_states, all_self_attentions = self.encoder(
inputs_embeds,
full_attention_mask,
rotary_pos_emb=rotary_pos_emb,
kv_caches=past_key_values,
use_cache=use_cache,
output_hidden_states=output_hidden_states,
)
hidden_states = gather_forward_split_backward(hidden_states,
dim=0,
process_group=shard_config.tensor_parallel_process_group)
if not return_dict:
return tuple(v for v in [
hidden_states,
presents,
all_hidden_states,
all_self_attentions,
] if v is not None)
return BaseModelOutputWithPast(
last_hidden_state=hidden_states,
past_key_values=presents,
hidden_states=all_hidden_states,
attentions=all_self_attentions,
)
return forward
colossalai/shardformer/modeling/gpt2.py
View file @ efba0f44
This diff is collapsed.
colossalai/shardformer/modeling/whisper.py
View file @ efba0f44
This diff is collapsed.
colossalai/shardformer/policies/auto_policy.py
View file @ efba0f44
......@@ -125,9 +125,9 @@ _POLICY_LIST = {
# ChatGLM
"colossalai.shardformer.modeling.chatglm2_6b.modeling_chatglm.ChatGLMModel":
PolicyLocation(file_name="chatglm", class_name="ChatGLMModelPolicy"),
PolicyLocation(file_name="chatglm2", class_name="ChatGLMModelPolicy"),
"colossalai.shardformer.modeling.chatglm2_6b.modeling_chatglm.ChatGLMForConditionalGeneration":
PolicyLocation(file_name="chatglm", class_name="ChatGLMForConditionalGenerationPolicy"),
PolicyLocation(file_name="chatglm2", class_name="ChatGLMForConditionalGenerationPolicy"),
}
......
colossalai/shardformer/policies/base_policy.py
View file @ efba0f44
......@@ -11,17 +11,12 @@ from torch.nn import Module
from colossalai.pipeline.stage_manager import PipelineStageManager
from ..layer.parallel_module import ParallelModule
from ..shard.shard_config import ShardConfig
__all__ = ["ParallelModule", "SubModuleReplacementDescription", "ModulePolicyDescription", "Policy"]
class ParallelModule():
def __init__(self):
pass
@dataclass
class SubModuleReplacementDescription:
r"""
......
colossalai/shardformer/policies/bert.py
View file @ efba0f44
......@@ -10,6 +10,7 @@ import colossalai.shardformer.layer as col_nn
from .._utils import getattr_, setattr_
from ..modeling.bert import (
BertPipelineForwards,
bert_sequence_parallel_forward_fn,
get_bert_flash_attention_forward,
get_jit_fused_bert_output_forward,
get_jit_fused_bert_self_output_forward,
......@@ -47,13 +48,15 @@ class BertPolicy(Policy):
from transformers.models.bert.modeling_bert import (
BertEmbeddings,
BertLayer,
BertModel,
BertOutput,
BertSelfAttention,
BertSelfOutput,
)
policy = {}
use_sequence_parallel = self.shard_config.enable_sequence_parallelism
overlap = self.shard_config.enable_sequence_overlap
if self.shard_config.enable_tensor_parallelism:
policy[BertLayer] = ModulePolicyDescription(attribute_replacement={
"attention.self.all_head_size":
......@@ -69,14 +72,26 @@ class BertPolicy(Policy):
SubModuleReplacementDescription(
suffix="attention.self.query",
target_module=col_nn.Linear1D_Col,
kwargs={
"seq_parallel": use_sequence_parallel,
"overlap": overlap
},
),
SubModuleReplacementDescription(
suffix="attention.self.key",
target_module=col_nn.Linear1D_Col,
kwargs={
"seq_parallel": use_sequence_parallel,
"overlap": overlap
},
),
SubModuleReplacementDescription(
suffix="attention.self.value",
target_module=col_nn.Linear1D_Col,
kwargs={
"seq_parallel": use_sequence_parallel,
"overlap": overlap
},
),
SubModuleReplacementDescription(
suffix="attention.self.dropout",
......@@ -85,6 +100,7 @@ class BertPolicy(Policy):
SubModuleReplacementDescription(
suffix="attention.output.dense",
target_module=col_nn.Linear1D_Row,
kwargs={"seq_parallel": use_sequence_parallel},
),
SubModuleReplacementDescription(
suffix="attention.output.dropout",
......@@ -93,10 +109,15 @@ class BertPolicy(Policy):
SubModuleReplacementDescription(
suffix="intermediate.dense",
target_module=col_nn.Linear1D_Col,
kwargs={
"seq_parallel": use_sequence_parallel,
"overlap": overlap
},
),
SubModuleReplacementDescription(
suffix="output.dense",
target_module=col_nn.Linear1D_Row,
kwargs={"seq_parallel": use_sequence_parallel},
),
SubModuleReplacementDescription(
suffix="output.dropout",
......@@ -115,6 +136,12 @@ class BertPolicy(Policy):
)
])
if use_sequence_parallel:
self.append_or_create_method_replacement(
description={'forward': bert_sequence_parallel_forward_fn(self.shard_config)},
policy=policy,
target_key=BertModel)
# optimization configuration
if self.shard_config.enable_fused_normalization:
# Handle bert layer
......@@ -141,20 +168,26 @@ class BertPolicy(Policy):
# use flash attention
if self.shard_config.enable_flash_attention:
policy[BertSelfAttention] = ModulePolicyDescription(method_replacement={
self.append_or_create_method_replacement(description={
'forward': get_bert_flash_attention_forward(),
})
},
policy=policy,
target_key=BertSelfAttention)
# use jit operator
if self.shard_config.enable_jit_fused:
policy[BertSelfOutput] = ModulePolicyDescription(method_replacement={
self.append_or_create_method_replacement(description={
'forward': get_jit_fused_bert_self_output_forward(),
'dropout_add': get_jit_fused_dropout_add_func(),
})
policy[BertOutput] = ModulePolicyDescription(method_replacement={
},
policy=policy,
target_key=BertSelfOutput)
self.append_or_create_method_replacement(description={
'forward': get_jit_fused_bert_output_forward(),
'dropout_add': get_jit_fused_dropout_add_func(),
})
},
policy=policy,
target_key=BertOutput)
return policy
......@@ -205,7 +238,13 @@ class BertPolicy(Policy):
layers_per_stage = Policy.distribute_layers(len(module.encoder.layer), stage_manager.num_stages)
stage_index = Policy.get_stage_index(layers_per_stage, stage_manager.stage)
method_replacement = {'forward': partial(new_forward, stage_manager=stage_manager, stage_index=stage_index)}
method_replacement = {
'forward':
partial(new_forward,
stage_manager=stage_manager,
stage_index=stage_index,
shard_config=self.shard_config)
}
self.append_or_create_method_replacement(description=method_replacement,
policy=policy,
target_key=model_cls)
......
colossalai/shardformer/policies/blip2.py
View file @ efba0f44
This diff is collapsed.
colossalai/shardformer/policies/bloom.py
View file @ efba0f44
This diff is collapsed.
colossalai/shardformer/policies/gpt2.py
View file @ efba0f44
This diff is collapsed.
colossalai/shardformer/policies/llama.py
View file @ efba0f44
import warnings
from functools import partial
from typing import Callable, Dict, List, Union
......@@ -35,6 +36,10 @@ class LlamaPolicy(Policy):
policy = {}
if self.shard_config.enable_sequence_parallelism:
self.shard_config.enable_sequence_parallelism = False
warnings.warn("Llama dosen't support sequence parallelism now, will ignore the sequence parallelism flag.")
if self.shard_config.enable_tensor_parallelism:
policy[LlamaDecoderLayer] = ModulePolicyDescription(
attribute_replacement={
......@@ -105,9 +110,11 @@ class LlamaPolicy(Policy):
target_key=LlamaModel)
if self.shard_config.enable_flash_attention:
policy[LlamaAttention] = ModulePolicyDescription(method_replacement={
self.append_or_create_method_replacement(description={
'forward': get_llama_flash_attention_forward(),
})
},
policy=policy,
target_key=LlamaAttention)
return policy
......
colossalai/shardformer/policies/opt.py
View file @ efba0f44
This diff is collapsed.
colossalai/shardformer/policies/sam.py
View file @ efba0f44
This diff is collapsed.
colossalai/shardformer/policies/t5.py
View file @ efba0f44
This diff is collapsed.
colossalai/shardformer/policies/vit.py
View file @ efba0f44
This diff is collapsed.
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment