github
/
ColossalAI
mirror of https://github.com/hpcaitech/ColossalAI
1
0
Fork 0
Code Issues Projects Releases Wiki Activity

Merge pull request #5842 from wangbluo/dev/zero-offload 

update llama model
pull/5844/head^2
Wang Binluo 2024-06-20 15:37:03 +08:00 committed by GitHub
parent df612434c9 351a1c269b
commit 35ef72bfd1
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 281 additions and 162 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

43
colossalai/shardformer/layer/loss.py
View File

@ -15,7 +15,15 @@ class DistCrossEntropy(Function):
"""
@staticmethod
def forward(ctx, vocab_logits: torch.Tensor, target: torch.Tensor, ignore_index: int, process_group: ProcessGroup):
def forward(
ctx,
vocab_logits: torch.Tensor,
target: torch.Tensor,
ignore_index: int,
process_group: ProcessGroup,
vocab_size: int,
dtype=torch.float32,
):
r"""
Calculate the cross entropy loss before gather, the origin loss function is as follows:
loss = -log(exp(x[class])/sum(exp(x[i]))
@ -27,7 +35,7 @@ class DistCrossEntropy(Function):
Args:
vocab_logits (:class:`torch.Tensor`): The logits of the vocabulary, shape is
[batch_size, seq_len, vocab_size]
labels (:class:`torch.Tensor`): The labels of the vocabulary, shape is
target (:class:`torch.Tensor`): The labels of the vocabulary, shape is
[batch_size, seq_len]
Returns:
@ -41,15 +49,21 @@ class DistCrossEntropy(Function):
vocab_logits = vocab_logits - logits_max.unsqueeze(dim=-1)
# mask the target in the local device
partition_vocab_size = vocab_logits.size()[-1]
rank = dist.get_rank(group=process_group)
world_size = dist.get_world_size(group=process_group)
global_vocab_size = partition_vocab_size * world_size
if vocab_size == None:
partition_vocab_size = vocab_logits.size()[-1]
global_vocab_size = partition_vocab_size * world_size
else:
global_vocab_size = vocab_size
partition_vocab_size = global_vocab_size // world_size
# [down, up) => false, other device and -100 => true
delta = (global_vocab_size + world_size - 1) // world_size
down_threshold = rank * delta
up_threshold = down_threshold + delta
if up_threshold > global_vocab_size:
up_threshold = global_vocab_size
mask = (target < down_threshold) | (target >= up_threshold)
masked_target = target.clone() - down_threshold
masked_target[mask] = 0
@ -57,7 +71,8 @@ class DistCrossEntropy(Function):
# reshape the logits and target
# reshape the vocab_logits to [bath_size * seq_len, vocab_size]
# reshape the labels to [bath_size * seq_len]
logits_2d = vocab_logits.view(-1, partition_vocab_size)
self_vocab_size = vocab_logits.size()[-1]
logits_2d = vocab_logits.view(-1, self_vocab_size)
masked_target_1d = masked_target.view(-1)
# extract the x[class] and set the x[other device] to zero
@ -72,7 +87,7 @@ class DistCrossEntropy(Function):
dist.all_reduce(pred_logits, op=dist.ReduceOp.SUM, group=process_group)
exp_logits = vocab_logits
torch.exp(vocab_logits, out=exp_logits)
sum_exp_logits = torch.sum(exp_logits, dim=-1)
sum_exp_logits = torch.sum(exp_logits, dim=-1, dtype=torch.float32)
dist.all_reduce(sum_exp_logits, op=dist.ReduceOp.SUM, group=process_group)
# calculate the loss
@ -83,9 +98,10 @@ class DistCrossEntropy(Function):
loss = torch.sum(loss).div_(num_non_zero)
# calculate the softmax
exp_logits.div_(sum_exp_logits.unsqueeze(dim=-1))
exp_logits = exp_logits.div(sum_exp_logits.unsqueeze(dim=-1)).to(dtype)
exp_logits[target == ignore_index] = 0.0
ctx.save_for_backward(exp_logits, mask, masked_target_1d)
ctx.dtype = dtype
return loss
@ -100,14 +116,19 @@ class DistCrossEntropy(Function):
partion_vocab_size = grad_logits.shape[-1]
grad_logits_2d = grad_logits.view(-1, partion_vocab_size)
update = 1.0 - mask.view(-1).float()
update = 1.0 - mask.view(-1).float().to(ctx.dtype)
grad_logits_2d[torch.arange(0, grad_logits_2d.shape[0]), masked_target_1d] -= update
grad_logits.mul_(grad_output.unsqueeze(dim=-1))
return grad_logits, None, None, None
return grad_logits, None, None, None, None, None
def cross_entropy_1d(
vocab_logits: torch.Tensor, labels: torch.Tensor, ignore_index: int = -100, process_group: ProcessGroup = None
vocab_logits: torch.Tensor,
labels: torch.Tensor,
ignore_index: int = -100,
process_group: ProcessGroup = None,
vocab_size: int = None,
dtype: torch.dtype = None,
) -> torch.Tensor:
return DistCrossEntropy.apply(vocab_logits, labels, ignore_index, process_group)
return DistCrossEntropy.apply(vocab_logits, labels, ignore_index, process_group, vocab_size, dtype)

400
colossalai/shardformer/modeling/llama.py
View File

@ -1,29 +1,34 @@
import math
import warnings
from typing import List, Optional, Tuple, Union
import torch
import torch.nn.functional as F
import torch.utils.checkpoint
from torch import nn
from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
from transformers.cache_utils import Cache, DynamicCache
from transformers.modeling_outputs import (
BaseModelOutputWithPast,
CausalLMOutputWithPast,
SequenceClassifierOutputWithPast,
)
from transformers.models.llama.modeling_llama import LlamaForCausalLM, LlamaForSequenceClassification, LlamaModel
from transformers.models.llama.modeling_llama import (
LlamaForCausalLM,
LlamaForSequenceClassification,
LlamaModel,
StaticCache,
apply_rotary_pos_emb,
repeat_kv,
)
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
from ..layer import ColoAttention, cross_entropy_1d
try:
from transformers.models.llama.modeling_llama import _prepare_4d_causal_attention_mask
LATEST_VERSION = True
except ImportError:
LATEST_VERSION = False
class LlamaPipelineForwards:
"""
@ -43,6 +48,7 @@ class LlamaPipelineForwards:
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
cache_position: Optional[torch.LongTensor] = None,
stage_manager: Optional[PipelineStageManager] = None,
hidden_states: Optional[torch.FloatTensor] = None,
stage_index: Optional[List[int]] = None,
@ -55,30 +61,45 @@ class LlamaPipelineForwards:
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
if use_cache:
logger.warning_once(
"`use_cache=True` is incompatible with pipeline parallelism. Setting `use_cache=False`..."
)
use_cache = False
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
# retrieve input_ids and inputs_embeds
if stage_manager.is_first_stage():
if input_ids is not None and inputs_embeds is not None:
raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
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
batch_size, seq_length = input_ids.shape[:2]
elif inputs_embeds is not None:
batch_size, seq_length, _ = inputs_embeds.shape
batch_size, seq_length, _ = inputs_embeds.shape[:2]
else:
raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")
raise ValueError("You have to specify either input_ids or inputs_embeds")
device = input_ids.device if input_ids is not None else inputs_embeds.device
if inputs_embeds is None:
inputs_embeds = self.embed_tokens(input_ids)
hidden_states = inputs_embeds
else:
input_shape = hidden_states.shape[:-1]
batch_size, seq_length = input_shape
device = hidden_states.device
seq_length_with_past = seq_length
past_key_values_length = 0
past_seen_tokens = 0
if use_cache: # kept for BC (cache positions)
if not isinstance(past_key_values, StaticCache):
past_key_values = DynamicCache.from_legacy_cache(past_key_values)
past_seen_tokens = past_key_values.get_seq_length()
if cache_position is None:
if isinstance(past_key_values, StaticCache):
raise ValueError("cache_position is a required argument when using StaticCache.")
cache_position = torch.arange(past_seen_tokens, past_seen_tokens + hidden_states.shape[1], device=device)
seq_length_with_past = seq_length + past_seen_tokens
# TODO(jianghai): left the recording kv-value tensors as () or None type, this feature may be added in the future.
if output_attentions:
@ -91,17 +112,8 @@ class LlamaPipelineForwards:
logger.warning_once("use_cache=True is not supported for pipeline models at the moment.")
use_cache = False
if past_key_values 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 position_ids is None:
position_ids = torch.arange(
past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device
)
position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
else:
position_ids = position_ids.view(-1, seq_length).long()
position_ids = cache_position.unsqueeze(0)
# embed positions, for the first stage, hidden_states is the input embeddings,
# for the other stages, hidden_states is the output of the previous stage
@ -109,23 +121,16 @@ class LlamaPipelineForwards:
# in this case, attention_mask is a dict rather than a tensor
mask_shape = (batch_size, 1, seq_length_with_past, seq_length_with_past)
attention_mask = ColoAttention.prepare_attn_kwargs(
mask_shape, hidden_states.dtype, hidden_states.device, q_padding_mask=attention_mask, is_causal=True
mask_shape,
hidden_states.dtype,
hidden_states.device,
q_padding_mask=attention_mask,
is_causal=True,
)
else:
if attention_mask is None:
attention_mask = torch.ones(
(batch_size, seq_length_with_past), dtype=torch.bool, device=hidden_states.device
)
if LATEST_VERSION:
attention_mask = _prepare_4d_causal_attention_mask(
attention_mask, (batch_size, seq_length), hidden_states, past_key_values_length
)
else:
attention_mask = self._prepare_decoder_attention_mask(
attention_mask, (batch_size, seq_length), hidden_states, past_key_values_length
)
attention_mask = self._update_causal_mask(attention_mask, hidden_states, cache_position)
if self.gradient_checkpointing and self.training:
if self.gradient_checkpointing and self.training and use_cache:
if use_cache:
logger.warning_once(
"`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
@ -135,7 +140,7 @@ class LlamaPipelineForwards:
# decoder layers
all_hidden_states = () if output_hidden_states else None
all_self_attns = () if output_attentions else None
next_decoder_cache = () if use_cache else None
next_decoder_cache = None
start_idx, end_idx = stage_index[0], stage_index[1]
num_ckpt_layers = 0
@ -145,8 +150,10 @@ class LlamaPipelineForwards:
if shard_config.gradient_checkpoint_config is not None:
num_ckpt_layers = shard_config.gradient_checkpoint_config.get_num_ckpt_layers(
stage=stage_manager.stage,
num_stages=stage_manager.num_stages,
num_layers=end_idx - start_idx,
model_chunk_id=stage_manager.model_chunk_id if stage_manager.is_interleave else 0,
model_chunk_id=(stage_manager.model_chunk_id if stage_manager.is_interleave else 0),
num_model_chunks=stage_manager.num_model_chunks,
)
assert num_ckpt_layers <= end_idx - start_idx
@ -154,38 +161,32 @@ class LlamaPipelineForwards:
if output_hidden_states:
all_hidden_states += (hidden_states,)
past_key_value = past_key_values[idx] if past_key_values is not None else None
if idx - start_idx < num_ckpt_layers:
def create_custom_forward(module):
def custom_forward(*inputs):
# None for past_key_value
return module(*inputs, output_attentions, None)
return custom_forward
layer_outputs = torch.utils.checkpoint.checkpoint(
create_custom_forward(decoder_layer),
layer_outputs = self._gradient_checkpointing_func(
decoder_layer.__call__,
hidden_states,
attention_mask,
position_ids,
None,
past_key_values,
output_attentions,
use_cache,
cache_position,
)
else:
layer_outputs = decoder_layer(
hidden_states,
attention_mask=attention_mask,
position_ids=position_ids,
past_key_value=past_key_value,
past_key_value=past_key_values,
output_attentions=output_attentions,
use_cache=use_cache,
cache_position=cache_position,
)
hidden_states = layer_outputs[0]
if use_cache:
next_decoder_cache += (layer_outputs[2 if output_attentions else 1],)
next_decoder_cache = layer_outputs[2 if output_attentions else 1]
if output_attentions:
all_self_attns += (layer_outputs[1],)
@ -198,7 +199,16 @@ class LlamaPipelineForwards:
next_cache = next_decoder_cache if use_cache else None
if stage_manager.is_last_stage():
if not return_dict:
return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
return tuple(
v
for v in [
hidden_states,
next_cache,
all_hidden_states,
all_self_attns,
]
if v is not None
)
return BaseModelOutputWithPast(
last_hidden_state=hidden_states,
past_key_values=next_cache,
@ -221,6 +231,7 @@ class LlamaPipelineForwards:
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
cache_position: Optional[torch.LongTensor] = None,
stage_manager: Optional[PipelineStageManager] = None,
hidden_states: Optional[torch.FloatTensor] = None,
stage_index: Optional[List[int]] = None,
@ -278,6 +289,7 @@ class LlamaPipelineForwards:
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
cache_position=cache_position,
stage_manager=stage_manager,
hidden_states=hidden_states,
stage_index=stage_index,
@ -302,7 +314,11 @@ class LlamaPipelineForwards:
new_vocab_size = logits.shape[-1]
shift_logits = shift_logits.view(-1, new_vocab_size)
loss = cross_entropy_1d(
shift_logits, shift_labels, process_group=shard_config.tensor_parallel_process_group
shift_logits,
shift_labels,
process_group=shard_config.tensor_parallel_process_group,
vocab_size=self.lm_head.out_features,
dtype=self.model.dtype,
)
else:
shift_logits = shift_logits.view(-1, self.config.vocab_size)
@ -336,6 +352,7 @@ class LlamaPipelineForwards:
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
cache_position: Optional[torch.LongTensor] = None,
stage_manager: Optional[PipelineStageManager] = None,
hidden_states: Optional[torch.FloatTensor] = None,
stage_index: Optional[List[int]] = None,
@ -369,6 +386,7 @@ class LlamaPipelineForwards:
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
cache_position=cache_position,
stage_manager=stage_manager,
hidden_states=hidden_states,
stage_index=stage_index,
@ -438,71 +456,136 @@ class LlamaPipelineForwards:
return {"hidden_states": hidden_states}
def get_llama_flash_attention_forward(shard_config: ShardConfig):
from transformers.models.llama.modeling_llama import LlamaAttention, apply_rotary_pos_emb
llama_version = 2
try:
from transformers.models.llama.modeling_llama import repeat_kv
except:
warnings.warn("using llamav1, llamav1 hasn't repeat_kv function")
llama_version = 1
def get_llama_flash_attention_forward(shard_config, sp_mode=None, sp_size=None, sp_group=None):
def forward(
self: LlamaAttention,
self,
hidden_states: torch.Tensor,
attention_mask: Optional[dict] = None,
attention_mask: Optional[torch.Tensor] = None,
position_ids: Optional[torch.LongTensor] = None,
past_key_value: Optional[Tuple[torch.Tensor]] = None,
past_key_value: Optional[Cache] = None,
output_attentions: bool = False,
use_cache: bool = False,
cache_position: Optional[torch.LongTensor] = None,
**kwargs,
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
bsz, q_len, _ = hidden_states.size()
assert q_len % 4 == 0, "Flash Attention Error: The sequence length should be a multiple of 4."
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Cache]]:
if sp_mode is not None:
assert sp_mode in ["all_to_all", "split_gather", "ring"], "Invalid sp_mode"
assert (sp_size is not None) and (
sp_group is not None
), "Must specify sp_size and sp_group for sequence parallel"
if "padding_mask" in kwargs:
warnings.warn(
"Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
)
query_states = self.q_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
key_states = self.k_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
value_states = self.v_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
bsz, q_len, _ = hidden_states.size()
# sp: modify sp_len when sequence parallel mode is ring
if sp_mode in ["split_gather", "ring"]:
q_len *= sp_size
if self.config.pretraining_tp > 1:
key_value_slicing = (self.num_key_value_heads * self.head_dim) // self.config.pretraining_tp
query_slices = self.q_proj.weight.split(
(self.num_heads * self.head_dim) // self.config.pretraining_tp, dim=0
)
key_slices = self.k_proj.weight.split(key_value_slicing, dim=0)
value_slices = self.v_proj.weight.split(key_value_slicing, dim=0)
query_states = [F.linear(hidden_states, query_slices[i]) for i in range(self.config.pretraining_tp)]
query_states = torch.cat(query_states, dim=-1)
key_states = [F.linear(hidden_states, key_slices[i]) for i in range(self.config.pretraining_tp)]
key_states = torch.cat(key_states, dim=-1)
value_states = [F.linear(hidden_states, value_slices[i]) for i in range(self.config.pretraining_tp)]
value_states = torch.cat(value_states, dim=-1)
else:
query_states = self.q_proj(hidden_states)
key_states = self.k_proj(hidden_states)
value_states = self.v_proj(hidden_states)
query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
kv_seq_len = key_states.shape[-2]
if past_key_value is not None:
kv_seq_len += past_key_value[0].shape[-2]
if self.layer_idx is None:
raise ValueError(
f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} "
"for auto-regressive decoding with k/v caching, please make sure to initialize the attention class "
"with a layer index."
)
cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
cos, sin = self.rotary_emb(value_states, position_ids)
query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
if past_key_value is not None:
# reuse k, v, self_attention
key_states = torch.cat([past_key_value[0], key_states], dim=2)
value_states = torch.cat([past_key_value[1], value_states], dim=2)
past_key_value = (key_states, value_states) if use_cache else None
cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
# repeat k/v heads if n_kv_heads < n_heads
if llama_version == 2:
key_states = repeat_kv(key_states, self.num_key_value_groups)
value_states = repeat_kv(value_states, self.num_key_value_groups)
key_states = repeat_kv(key_states, self.num_key_value_groups)
value_states = repeat_kv(value_states, self.num_key_value_groups)
if shard_config.enable_flash_attention:
assert isinstance(attention_mask, dict), "Flash Attention Error: attention_mask should be a dict."
attn_output = ColoAttention.attention(query_states, key_states, value_states, **attention_mask)
else:
attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
raise ValueError(
f"Attention weights should be of size {(bsz, self.num_heads, q_len, kv_seq_len)}, but is"
f" {attn_weights.size()}"
)
if attention_mask is not None:
if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
raise ValueError(
f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
)
attn_weights = attn_weights + attention_mask
# upcast attention to fp32
attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
attn_output = torch.matmul(attn_weights, value_states)
if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
raise ValueError(
f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
f" {attn_output.size()}"
)
assert isinstance(attention_mask, dict), "Flash Attention Error: attention_mask should be a dict."
attn_output = ColoAttention.attention(query_states, key_states, value_states, **attention_mask)
attn_output = attn_output.transpose(1, 2).contiguous()
attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
# sp: all-to-all comminucation when introducing sequence parallel
if sp_mode == "all_to_all":
attn_output = attn_output.reshape(bsz, q_len, self.num_heads * self.head_dim)
# attn_output = all_to_all_comm(attn_output, sp_group, scatter_dim=1, gather_dim=2)
else:
attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
attn_output = self.o_proj(attn_output)
if self.config.pretraining_tp > 1:
attn_output = attn_output.split(self.hidden_size // self.config.pretraining_tp, dim=2)
o_proj_slices = self.o_proj.weight.split(self.hidden_size // self.config.pretraining_tp, dim=1)
attn_output = sum([F.linear(attn_output[i], o_proj_slices[i]) for i in range(self.config.pretraining_tp)])
else:
attn_output = self.o_proj(attn_output)
return attn_output, None, past_key_value
if not output_attentions:
attn_weights = None
return attn_output, attn_weights, past_key_value
return forward
def get_llama_model_forward_for_flash_attn(shard_config: ShardConfig):
def get_llama_model_forward_for_flash_attn(shard_config, sp_mode=None, sp_size=None, sp_group=None):
logger = logging.get_logger(__name__)
assert shard_config.enable_flash_attention, "Flash Attention is not enabled."
def forward(
self: LlamaModel,
self,
input_ids: torch.LongTensor = None,
attention_mask: Optional[torch.Tensor] = None,
position_ids: Optional[torch.LongTensor] = None,
@ -512,112 +595,122 @@ def get_llama_model_forward_for_flash_attn(shard_config: ShardConfig):
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
cache_position: Optional[torch.LongTensor] = None,
) -> Union[Tuple, BaseModelOutputWithPast]:
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
# retrieve input_ids and inputs_embeds
if input_ids is not None and inputs_embeds is not None:
raise ValueError("You cannot specify both decoder_input_ids and decoder_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 decoder_input_ids or decoder_inputs_embeds")
seq_length_with_past = seq_length
past_key_values_length = 0
if past_key_values 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 position_ids is None:
device = input_ids.device if input_ids is not None else inputs_embeds.device
position_ids = torch.arange(
past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device
if (input_ids is None) ^ (inputs_embeds is not None):
raise ValueError(
"You cannot specify both input_ids and inputs_embeds at the same time, and must specify either one"
)
position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
else:
position_ids = position_ids.view(-1, seq_length).long()
if inputs_embeds is None:
inputs_embeds = self.embed_tokens(input_ids)
# embed positions
hidden_states = inputs_embeds
# in this case, attention_mask is a dict rather than a tensor
mask_shape = (batch_size, 1, seq_length_with_past, seq_length_with_past)
attention_mask = ColoAttention.prepare_attn_kwargs(
mask_shape, hidden_states.dtype, hidden_states.device, q_padding_mask=attention_mask, is_causal=True
)
if self.gradient_checkpointing and self.training:
if (self.gradient_checkpointing or sp_mode in ["ring", "all_to_all"]) and self.training:
if use_cache:
logger.warning_once(
"`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
)
use_cache = False
if inputs_embeds is None:
inputs_embeds = self.embed_tokens(input_ids)
past_seen_tokens = 0
seq_len = inputs_embeds.shape[1]
if use_cache: # kept for BC (cache positions)
if not isinstance(past_key_values, StaticCache):
past_key_values = DynamicCache.from_legacy_cache(past_key_values)
past_seen_tokens = past_key_values.get_seq_length()
if cache_position is None:
if isinstance(past_key_values, StaticCache):
raise ValueError("cache_position is a required argument when using StaticCache.")
cache_position = torch.arange(past_seen_tokens, past_seen_tokens + seq_len, device=inputs_embeds.device)
if position_ids is None:
position_ids = cache_position.unsqueeze(0)
# in this case, attention_mask is a dict rather than a tensor
if shard_config.enable_flash_attention:
mask_shape = (inputs_embeds.shape[0], 1, past_seen_tokens + seq_len, past_seen_tokens + seq_len)
attention_mask = ColoAttention.prepare_attn_kwargs(
mask_shape,
inputs_embeds.dtype,
inputs_embeds.device,
q_padding_mask=attention_mask,
is_causal=True,
)
else:
attention_mask = self._update_causal_mask(attention_mask, inputs_embeds, cache_position)
if sp_mode in ["ring", "split_gather"]:
inputs_embeds = split_forward_gather_backward(inputs_embeds, 1, sp_group)
elif sp_mode == "all_to_all":
inputs_embeds = split_forward_gather_backward(inputs_embeds, 1, sp_group, 1 / sp_size)
hidden_states = inputs_embeds
# decoder layers
all_hidden_states = () if output_hidden_states else None
all_self_attns = () if output_attentions else None
next_decoder_cache = () if use_cache else None
next_decoder_cache = None
for idx, decoder_layer in enumerate(self.layers):
for decoder_layer in self.layers:
if output_hidden_states:
all_hidden_states += (hidden_states,)
past_key_value = past_key_values[idx] if past_key_values is not None else None
if self.gradient_checkpointing and self.training:
def create_custom_forward(module):
def custom_forward(*inputs):
# None for past_key_value
return module(*inputs, past_key_value, output_attentions)
return custom_forward
layer_outputs = torch.utils.checkpoint.checkpoint(
create_custom_forward(decoder_layer),
layer_outputs = self._gradient_checkpointing_func(
decoder_layer.__call__,
hidden_states,
attention_mask,
position_ids,
past_key_values,
output_attentions,
use_cache,
cache_position,
)
else:
layer_outputs = decoder_layer(
hidden_states,
attention_mask=attention_mask,
position_ids=position_ids,
past_key_value=past_key_value,
past_key_value=past_key_values,
output_attentions=output_attentions,
use_cache=use_cache,
cache_position=cache_position,
)
hidden_states = layer_outputs[0]
if use_cache:
next_decoder_cache += (layer_outputs[2 if output_attentions else 1],)
next_decoder_cache = layer_outputs[2 if output_attentions else 1]
if output_attentions:
all_self_attns += (layer_outputs[1],)
hidden_states = self.norm(hidden_states)
if sp_mode == "ring" or sp_mode == "split_gather":
hidden_states = gather_forward_split_backward(hidden_states, 1, sp_group)
elif sp_mode == "all_to_all":
hidden_states = gather_forward_split_backward(hidden_states, 1, sp_group, grad_scale=sp_size)
# add hidden states from the last decoder layer
if output_hidden_states:
all_hidden_states += (hidden_states,)
next_cache = next_decoder_cache if use_cache else None
next_cache = None
if use_cache:
next_cache = (
next_decoder_cache.to_legacy_cache() if isinstance(next_decoder_cache, Cache) else next_decoder_cache
)
if not return_dict:
return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
return BaseModelOutputWithPast(
last_hidden_state=hidden_states,
past_key_values=next_cache,
@ -643,6 +736,7 @@ def get_lm_forward_with_dist_cross_entropy(shard_config: ShardConfig):
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
cache_position: Optional[torch.LongTensor] = None,
) -> Union[Tuple, CausalLMOutputWithPast]:
r"""
Args:
@ -687,6 +781,7 @@ def get_lm_forward_with_dist_cross_entropy(shard_config: ShardConfig):
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
cache_position=cache_position,
)
hidden_states = outputs[0]
@ -706,11 +801,14 @@ def get_lm_forward_with_dist_cross_entropy(shard_config: ShardConfig):
shift_labels = shift_labels.view(-1)
# Enable model parallelism
shift_labels = shift_labels.to(shift_logits.device)
new_vocab_size = logits.shape[-1]
shift_logits = shift_logits.view(-1, new_vocab_size)
loss = cross_entropy_1d(
shift_logits, shift_labels, process_group=shard_config.tensor_parallel_process_group
shift_logits,
shift_labels,
process_group=shard_config.tensor_parallel_process_group,
vocab_size=self.lm_head.out_features,
dtype=self.model.dtype,
)
if not return_dict: