Bump transformers to 4.25.1 (#151)
- latest accelerate, transformers, huggingface_hub - rearrange attention caches to support https://github.com/huggingface/transformers/pull/18344 - remove unused code - fix edge case where session crashes when receiving seq length 0 - assert transformer version when importing WrappedBloomBlock Co-authored-by: Alexander Borzunov <borzunov.alexander@gmail.com> Co-authored-by: Max Ryabinin <mryabinin0@gmail.com>pull/153/head
parent
e4dc938dfe
commit
b04982c1a2
@ -1,2 +0,0 @@
|
||||
from petals.bloom.block import BloomBlock
|
||||
from petals.bloom.model import BloomConfig, BloomForCausalLM, BloomModel, BloomPreTrainedModel
|
@ -1,595 +0,0 @@
|
||||
"""
|
||||
PyTorch BLOOM model that implements several memory-efficient modes.
|
||||
Based on https://github.com/huggingface/transformers/commit/ca2a55e9dfb245527b5e1c954fec6ffbb7aef07b
|
||||
See commit history for authorship.
|
||||
"""
|
||||
from typing import Optional, Tuple, Union
|
||||
|
||||
import torch
|
||||
import torch.nn.functional as F
|
||||
import torch.utils.checkpoint
|
||||
from hivemind import use_hivemind_log_handler
|
||||
from torch import nn
|
||||
from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, LayerNorm, MSELoss
|
||||
from transformers.file_utils import (
|
||||
add_code_sample_docstrings,
|
||||
add_start_docstrings,
|
||||
add_start_docstrings_to_model_forward,
|
||||
)
|
||||
from transformers.modeling_outputs import (
|
||||
BaseModelOutputWithPastAndCrossAttentions,
|
||||
CausalLMOutputWithCrossAttentions,
|
||||
SequenceClassifierOutputWithPast,
|
||||
)
|
||||
from transformers.models.bloom.configuration_bloom import BloomConfig
|
||||
from transformers.models.bloom.modeling_bloom import BloomPreTrainedModel
|
||||
from transformers.utils import logging
|
||||
|
||||
from petals.bloom.block import BloomBlock
|
||||
|
||||
use_hivemind_log_handler("in_root_logger")
|
||||
logger = logging.get_logger(__file__)
|
||||
|
||||
_CHECKPOINT_FOR_DOC = "bigscience/Bloom"
|
||||
_CONFIG_FOR_DOC = "BloomConfig"
|
||||
_TOKENIZER_FOR_DOC = "BloomTokenizer"
|
||||
|
||||
|
||||
BLOOM_START_DOCSTRING = r"""
|
||||
|
||||
This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
|
||||
library implements for all its model (such as downloading or saving, resizing the input embeddings etc.)
|
||||
|
||||
This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
|
||||
Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
|
||||
and behavior.
|
||||
|
||||
Parameters:
|
||||
config ([`MemoryEfficientBloomConfig`]): Model configuration class with all the parameters of the model.
|
||||
Initializing with a config file does not load the weights associated with the model, only the
|
||||
configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
|
||||
"""
|
||||
|
||||
BLOOM_INPUTS_DOCSTRING = r"""
|
||||
Args:
|
||||
input_ids (`torch.LongTensor` of shape `(batch_size, input_ids_length)`):
|
||||
`input_ids_length` = `sequence_length` if `past_key_values` is `None` else
|
||||
`past_key_values[0][0].shape[-2]` (`sequence_length` of input past key value states). Indices of input
|
||||
sequence tokens in the vocabulary.
|
||||
|
||||
If `past_key_values` is used, only `input_ids` that do not have their past calculated should be passed as
|
||||
`input_ids`.
|
||||
|
||||
Indices can be obtained using [`BloomTokenizer`]. See [`PreTrainedTokenizer.encode`] and
|
||||
[`PreTrainedTokenizer.__call__`] for details.
|
||||
|
||||
[What are input IDs?](../glossary#input-ids)
|
||||
past_key_values (`Tuple[Tuple[torch.Tensor]]` of length `config.n_layers`):
|
||||
Contains precomputed hidden-states (key and values in the attention blocks) as computed by the model (see
|
||||
`past_key_values` output below). Can be used to speed up sequential decoding. The `input_ids` which have
|
||||
their past given to this model should not be passed as `input_ids` as they have already been computed.
|
||||
attention_mask (`torch.FloatTensor` of shape `(batch_size, sequence_length)`, *optional*):
|
||||
Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
|
||||
|
||||
- 1 for tokens that are **not masked**,
|
||||
- 0 for tokens that are **masked**.
|
||||
|
||||
[What are attention masks?](../glossary#attention-mask)
|
||||
position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
|
||||
Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
|
||||
config.max_position_embeddings - 1]`.
|
||||
|
||||
[What are position IDs?](../glossary#position-ids)
|
||||
head_mask (`torch.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*):
|
||||
Mask to nullify selected heads of the self-attention modules. Mask values selected in `[0, 1]`:
|
||||
|
||||
- 1 indicates the head is **not masked**,
|
||||
- 0 indicates the head is **masked**.
|
||||
|
||||
inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
|
||||
Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
|
||||
is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
|
||||
model's internal embedding lookup matrix.
|
||||
|
||||
If `past_key_values` is used, optionally only the last `inputs_embeds` have to be input (see
|
||||
`past_key_values`).
|
||||
use_cache (`bool`, *optional*):
|
||||
If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
|
||||
`past_key_values`).
|
||||
output_attentions (`bool`, *optional*):
|
||||
Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
|
||||
tensors for more detail.
|
||||
output_hidden_states (`bool`, *optional*):
|
||||
Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
|
||||
more detail.
|
||||
return_dict (`bool`, *optional*):
|
||||
Whether or not to return a [`~file_utils.ModelOutput`] instead of a plain tuple.
|
||||
"""
|
||||
|
||||
|
||||
class _BloomPreTrainedModelWithModifiedDefaults(BloomPreTrainedModel):
|
||||
@classmethod
|
||||
def from_pretrained(cls, *args, low_cpu_mem_usage: Optional[bool] = None, **kwargs):
|
||||
if low_cpu_mem_usage is None:
|
||||
low_cpu_mem_usage = True
|
||||
return super().from_pretrained(*args, low_cpu_mem_usage=low_cpu_mem_usage, **kwargs)
|
||||
|
||||
from_pretrained.__doc__ = BloomPreTrainedModel.from_pretrained.__doc__.replace(
|
||||
"low_cpu_mem_usage(`bool`, *optional*)",
|
||||
"low_cpu_mem_usage(`bool`, *optional*, defaults to `True` in Petals)",
|
||||
)
|
||||
|
||||
|
||||
@add_start_docstrings(
|
||||
"The bare Bloom Model transformer outputting raw hidden-states without any specific head on top.",
|
||||
BLOOM_START_DOCSTRING,
|
||||
)
|
||||
class BloomModel(_BloomPreTrainedModelWithModifiedDefaults):
|
||||
def __init__(self, config):
|
||||
super().__init__(config)
|
||||
assert not config.slow_but_exact, "slow_but_exact mode was removed for code simplicity"
|
||||
|
||||
self.embed_dim = config.hidden_size
|
||||
self.n_head = config.n_head
|
||||
|
||||
# Embedding + LN Embedding
|
||||
self.word_embeddings = nn.Embedding(config.vocab_size, self.embed_dim)
|
||||
self.word_embeddings_layernorm = LayerNorm(self.embed_dim, eps=config.layer_norm_epsilon)
|
||||
|
||||
# Transformer blocks
|
||||
self.h = nn.ModuleList([BloomBlock(config, layer_number=i) for i in range(config.num_hidden_layers)])
|
||||
|
||||
# Final Layer Norm
|
||||
self.ln_f = LayerNorm(self.embed_dim, eps=config.layer_norm_epsilon)
|
||||
|
||||
self.gradient_checkpointing = False
|
||||
|
||||
# Initialize weights and apply final processing
|
||||
self.post_init()
|
||||
|
||||
def get_input_embeddings(self):
|
||||
return self.word_embeddings
|
||||
|
||||
def set_input_embeddings(self, new_embeddings):
|
||||
self.word_embeddings = new_embeddings
|
||||
|
||||
@add_start_docstrings_to_model_forward(BLOOM_INPUTS_DOCSTRING)
|
||||
@add_code_sample_docstrings(
|
||||
processor_class=_TOKENIZER_FOR_DOC,
|
||||
checkpoint=_CHECKPOINT_FOR_DOC,
|
||||
output_type=BaseModelOutputWithPastAndCrossAttentions,
|
||||
config_class=_CONFIG_FOR_DOC,
|
||||
)
|
||||
def forward(
|
||||
self,
|
||||
input_ids=None,
|
||||
past_key_values=None,
|
||||
attention_mask=None,
|
||||
position_ids=None,
|
||||
head_mask=None,
|
||||
inputs_embeds=None,
|
||||
use_cache=None,
|
||||
output_attentions=None,
|
||||
output_hidden_states=None,
|
||||
return_dict=None,
|
||||
):
|
||||
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")
|
||||
if position_ids is not None:
|
||||
logger.warning("position_ids are ignored in this bloom implementation")
|
||||
elif input_ids is not None:
|
||||
input_shape = input_ids.size()
|
||||
input_ids = input_ids.view(-1, input_shape[-1])
|
||||
elif inputs_embeds is not None:
|
||||
input_shape = inputs_embeds.size()[:-1]
|
||||
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 bsz x n_head x N x N
|
||||
# head_mask has shape n_layer x batch x n_head 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)
|
||||
|
||||
# Note: it supports only float32 or bfloat16 inputs
|
||||
hidden_states = self.word_embeddings_layernorm(inputs_embeds)
|
||||
|
||||
output_shape = input_shape + (hidden_states.size(-1),)
|
||||
|
||||
presents = () if use_cache else None
|
||||
all_self_attentions = () if output_attentions else None
|
||||
all_hidden_states = () if output_hidden_states else None
|
||||
|
||||
# Compute alibi tensor: check build_alibi_tensor documentation
|
||||
current_sequence_length = hidden_states.shape[1]
|
||||
if past_key_values and past_key_values[0]:
|
||||
current_sequence_length += past_key_values[0][0].shape[1]
|
||||
|
||||
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:
|
||||
|
||||
if use_cache:
|
||||
logger.warning(
|
||||
"`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
|
||||
)
|
||||
use_cache = False
|
||||
|
||||
def create_custom_forward(module):
|
||||
def custom_forward(*inputs):
|
||||
# None for past_key_value
|
||||
return module(*inputs, use_cache, output_attentions, alibi=None)
|
||||
|
||||
return custom_forward
|
||||
|
||||
outputs = torch.utils.checkpoint.checkpoint(
|
||||
create_custom_forward(block),
|
||||
hidden_states,
|
||||
None,
|
||||
attention_mask,
|
||||
head_mask[i],
|
||||
)
|
||||
else:
|
||||
outputs = block(
|
||||
hidden_states,
|
||||
layer_past=layer_past,
|
||||
attention_mask=attention_mask,
|
||||
head_mask=head_mask[i],
|
||||
use_cache=use_cache,
|
||||
output_attentions=output_attentions,
|
||||
alibi=None,
|
||||
)
|
||||
|
||||
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],)
|
||||
|
||||
# Add last hidden state
|
||||
hidden_states = self.ln_f(hidden_states)
|
||||
|
||||
if output_hidden_states:
|
||||
all_hidden_states = all_hidden_states + (hidden_states,)
|
||||
|
||||
hidden_states = hidden_states.view(output_shape)
|
||||
|
||||
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,
|
||||
)
|
||||
|
||||
|
||||
@add_start_docstrings(
|
||||
"""
|
||||
The Bloom Model transformer with a language modeling head on top (linear layer with weights tied to the input
|
||||
embeddings).
|
||||
""",
|
||||
BLOOM_START_DOCSTRING,
|
||||
)
|
||||
class BloomForCausalLM(_BloomPreTrainedModelWithModifiedDefaults):
|
||||
_keys_to_ignore_on_load_missing = [r"h.*.self_attention.scale_mask_softmax.causal_mask", r"lm_head.weight"]
|
||||
|
||||
def __init__(self, config):
|
||||
super().__init__(config)
|
||||
self.transformer = BloomModel(config)
|
||||
self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
|
||||
|
||||
# Initialize weights and apply final processing
|
||||
self.post_init()
|
||||
|
||||
def get_output_embeddings(self):
|
||||
return self.lm_head
|
||||
|
||||
def set_output_embeddings(self, new_embeddings):
|
||||
self.lm_head = new_embeddings
|
||||
|
||||
def prepare_inputs_for_generation(self, input_ids, past=None, **kwargs):
|
||||
# only last token for inputs_ids if past is defined in kwargs
|
||||
if past:
|
||||
input_ids = input_ids[:, -1].unsqueeze(-1)
|
||||
|
||||
attention_mask = kwargs.get("attention_mask", None)
|
||||
position_ids = kwargs.get("position_ids", None)
|
||||
|
||||
if attention_mask is not None and position_ids is None:
|
||||
# create position_ids on the fly for batch generation
|
||||
position_ids = attention_mask.long().cumsum(-1) - 1
|
||||
position_ids.masked_fill_(attention_mask == 0, 1)
|
||||
if past:
|
||||
position_ids = position_ids[:, -1].unsqueeze(-1)
|
||||
else:
|
||||
position_ids = None
|
||||
return {
|
||||
"input_ids": input_ids,
|
||||
"past_key_values": past,
|
||||
"use_cache": kwargs.get("use_cache"),
|
||||
"position_ids": position_ids,
|
||||
"attention_mask": attention_mask,
|
||||
}
|
||||
|
||||
@add_start_docstrings_to_model_forward(BLOOM_INPUTS_DOCSTRING)
|
||||
@add_code_sample_docstrings(
|
||||
processor_class=_TOKENIZER_FOR_DOC,
|
||||
checkpoint=_CHECKPOINT_FOR_DOC,
|
||||
output_type=CausalLMOutputWithCrossAttentions,
|
||||
config_class=_CONFIG_FOR_DOC,
|
||||
)
|
||||
def forward(
|
||||
self,
|
||||
input_ids=None,
|
||||
past_key_values=None,
|
||||
attention_mask=None,
|
||||
position_ids=None,
|
||||
head_mask=None,
|
||||
inputs_embeds=None,
|
||||
labels=None,
|
||||
use_cache=None,
|
||||
output_attentions=None,
|
||||
output_hidden_states=None,
|
||||
return_dict=None,
|
||||
) -> Union[Tuple[torch.Tensor], CausalLMOutputWithCrossAttentions]:
|
||||
r"""
|
||||
labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
|
||||
Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set
|
||||
`labels = input_ids` Indices are selected in `[-100, 0, ..., config.vocab_size]` All labels set to `-100`
|
||||
are ignored (masked), the loss is only computed for labels in `[0, ..., config.vocab_size]`
|
||||
"""
|
||||
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
||||
|
||||
transformer_outputs = self.transformer(
|
||||
input_ids,
|
||||
past_key_values=past_key_values,
|
||||
attention_mask=attention_mask,
|
||||
position_ids=position_ids,
|
||||
head_mask=head_mask,
|
||||
inputs_embeds=inputs_embeds,
|
||||
use_cache=use_cache,
|
||||
output_attentions=output_attentions,
|
||||
output_hidden_states=output_hidden_states,
|
||||
return_dict=return_dict,
|
||||
)
|
||||
hidden_states = transformer_outputs[0]
|
||||
|
||||
lm_logits = self.lm_head(hidden_states)
|
||||
|
||||
loss = None
|
||||
if labels is not None:
|
||||
# Shift so that tokens < n predict n
|
||||
shift_logits = lm_logits[..., :-1, :].contiguous()
|
||||
shift_labels = labels[..., 1:].contiguous()
|
||||
# Flatten the tokens
|
||||
loss_fct = CrossEntropyLoss()
|
||||
loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
|
||||
|
||||
if not return_dict:
|
||||
output = (lm_logits,) + transformer_outputs[1:]
|
||||
return ((loss,) + output) if loss is not None else output
|
||||
|
||||
return CausalLMOutputWithCrossAttentions(
|
||||
loss=loss,
|
||||
logits=lm_logits,
|
||||
past_key_values=transformer_outputs.past_key_values,
|
||||
hidden_states=transformer_outputs.hidden_states,
|
||||
attentions=transformer_outputs.attentions,
|
||||
)
|
||||
|
||||
@staticmethod
|
||||
def _reorder_cache(past: Tuple[Tuple[torch.Tensor]], beam_idx: torch.Tensor) -> Tuple[Tuple[torch.Tensor]]:
|
||||
"""
|
||||
This function is used to re-order the `past_key_values` cache if [`~PreTrainedModel.beam_search`] or
|
||||
[`~PreTrainedModel.beam_sample`] is called. This is required to match `past_key_values` with the correct
|
||||
beam_idx at every generation step.
|
||||
"""
|
||||
return tuple(
|
||||
tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past)
|
||||
for layer_past in past
|
||||
)
|
||||
|
||||
|
||||
@add_start_docstrings(
|
||||
"""
|
||||
The modified language modeling head which does not create extra tensor for the linear layer with weights tied to the input
|
||||
embeddings. Thus, it reduces initial memory consumption which might be crucial for large dictionaries.
|
||||
In addition, it provides an effcient way to deal with half-precision word embeddings on CPU.
|
||||
""",
|
||||
BLOOM_START_DOCSTRING,
|
||||
)
|
||||
class LMHead(nn.Module):
|
||||
def __init__(self, config, word_embeddings: nn.Embedding):
|
||||
super().__init__()
|
||||
self.word_embeddings = word_embeddings
|
||||
self.chunk_size = config.chunk_size_for_efficient_fp16_on_cpu
|
||||
|
||||
@property
|
||||
def in_features(self) -> int:
|
||||
return self.word_embeddings.num_embeddings
|
||||
|
||||
@property
|
||||
def out_features(self) -> int:
|
||||
return self.word_embeddings.embedding_dim
|
||||
|
||||
@property
|
||||
def weight(self):
|
||||
return self.word_embeddings.weight
|
||||
|
||||
@property
|
||||
def bias(self):
|
||||
return None
|
||||
|
||||
def forward(self, hidden_states):
|
||||
word_embeddings = self.word_embeddings.weight
|
||||
|
||||
# We use 'chunked_forward' only when embeddings are in half-precision on CPU.
|
||||
if word_embeddings.dtype in [torch.float16, torch.bfloat16] and word_embeddings.device.type == "cpu":
|
||||
lm_logits = self.chunked_forward(hidden_states)
|
||||
else:
|
||||
# Switch dtype in case word_embeddings are fp16/bf16
|
||||
hidden_states = hidden_states.to(word_embeddings.dtype)
|
||||
lm_logits = F.linear(hidden_states, word_embeddings)
|
||||
return lm_logits
|
||||
|
||||
def chunked_forward(self, hidden_states):
|
||||
"""Splits word embeddings on chunks and iteratively casts them into fp32 to perform matmul more efficiently on CPU.
|
||||
chunk_size: provides trade-off between efficiency and extra memory consumption.
|
||||
"""
|
||||
assert self.chunk_size > 0, "Chunk size for chunked forward must be positive"
|
||||
|
||||
word_embeddings = self.word_embeddings.weight
|
||||
num_embeddings = self.word_embeddings.num_embeddings
|
||||
|
||||
hidden_states = hidden_states.float()
|
||||
output = torch.zeros(*hidden_states.shape[:-1], num_embeddings)
|
||||
|
||||
for i in range(0, num_embeddings, self.chunk_size):
|
||||
chunk = word_embeddings[i : i + self.chunk_size].float()
|
||||
output[..., i : i + self.chunk_size] = F.linear(hidden_states, chunk)
|
||||
return output
|
||||
|
||||
|
||||
@add_start_docstrings(
|
||||
"""
|
||||
The Bloom Model transformer with a sequence classification head on top (linear layer).
|
||||
[`BloomForSequenceClassification`] uses the last token in order to do the classification, as other causal models
|
||||
(e.g. GPT-1) do.
|
||||
Since it does classification on the last token, it requires to know the position of the last token. If a
|
||||
`pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each row. If
|
||||
no `pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot guess the
|
||||
padding tokens when `inputs_embeds` are passed instead of `input_ids`, it does the same (take the last value in
|
||||
each row of the batch).
|
||||
""",
|
||||
BLOOM_START_DOCSTRING,
|
||||
)
|
||||
class BloomForSequenceClassification(_BloomPreTrainedModelWithModifiedDefaults):
|
||||
_keys_to_ignore_on_load_missing = [r"h.*.self_attention.scale_mask_softmax.causal_mask", r"lm_head.weight"]
|
||||
|
||||
def __init__(self, config):
|
||||
super().__init__(config)
|
||||
self.num_labels = config.num_labels
|
||||
self.transformer = BloomModel(config)
|
||||
self.score = nn.Linear(config.hidden_size, config.num_labels, bias=False)
|
||||
|
||||
# Initialize weights and apply final processing
|
||||
self.post_init()
|
||||
|
||||
@add_start_docstrings_to_model_forward(BLOOM_INPUTS_DOCSTRING)
|
||||
@add_code_sample_docstrings(
|
||||
processor_class=_TOKENIZER_FOR_DOC,
|
||||
checkpoint=_CHECKPOINT_FOR_DOC,
|
||||
output_type=SequenceClassifierOutputWithPast,
|
||||
config_class=_CONFIG_FOR_DOC,
|
||||
)
|
||||
def forward(
|
||||
self,
|
||||
input_ids=None,
|
||||
past_key_values=None,
|
||||
attention_mask=None,
|
||||
position_ids=None,
|
||||
head_mask=None,
|
||||
inputs_embeds=None,
|
||||
labels=None,
|
||||
use_cache=None,
|
||||
output_attentions=None,
|
||||
output_hidden_states=None,
|
||||
return_dict=None,
|
||||
) -> Union[Tuple[torch.Tensor], SequenceClassifierOutputWithPast]:
|
||||
r"""
|
||||
labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
|
||||
Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
|
||||
config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
|
||||
`config.num_labels > 1` a classification loss is computed (Cross-Entropy).
|
||||
"""
|
||||
|
||||
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
||||
|
||||
transformer_outputs = self.transformer(
|
||||
input_ids,
|
||||
past_key_values=past_key_values,
|
||||
attention_mask=attention_mask,
|
||||
position_ids=position_ids,
|
||||
head_mask=head_mask,
|
||||
inputs_embeds=inputs_embeds,
|
||||
use_cache=use_cache,
|
||||
output_attentions=output_attentions,
|
||||
output_hidden_states=output_hidden_states,
|
||||
return_dict=return_dict,
|
||||
)
|
||||
|
||||
hidden_states = transformer_outputs[0]
|
||||
logits = self.score(hidden_states)
|
||||
|
||||
if input_ids is not None:
|
||||
batch_size = input_ids.shape[0]
|
||||
else:
|
||||
batch_size = inputs_embeds.shape[0]
|
||||
|
||||
if self.config.pad_token_id is None and batch_size != 1:
|
||||
raise ValueError("Cannot handle batch sizes > 1 if no padding token is defined.")
|
||||
if self.config.pad_token_id is None:
|
||||
sequence_lengths = -1
|
||||
else:
|
||||
if input_ids is not None:
|
||||
sequence_lengths = torch.ne(input_ids, self.config.pad_token_id).sum(-1) - 1
|
||||
else:
|
||||
sequence_lengths = -1
|
||||
logger.warning(
|
||||
f"{self.__class__.__name__} will not detect padding tokens in `inputs_embeds`. Results may be "
|
||||
"unexpected if using padding tokens in conjunction with `inputs_embeds.`"
|
||||
)
|
||||
|
||||
pooled_logits = logits[torch.arange(batch_size, device=logits.device), sequence_lengths]
|
||||
loss = None
|
||||
if labels is not None:
|
||||
if self.config.problem_type is None:
|
||||
if self.num_labels == 1:
|
||||
self.config.problem_type = "regression"
|
||||
elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
|
||||
self.config.problem_type = "single_label_classification"
|
||||
else:
|
||||
self.config.problem_type = "multi_label_classification"
|
||||
|
||||
if self.config.problem_type == "regression":
|
||||
loss_fct = MSELoss()
|
||||
if self.num_labels == 1:
|
||||
loss = loss_fct(pooled_logits.squeeze(), labels.squeeze())
|
||||
else:
|
||||
loss = loss_fct(pooled_logits, labels)
|
||||
elif self.config.problem_type == "single_label_classification":
|
||||
loss_fct = CrossEntropyLoss()
|
||||
loss = loss_fct(pooled_logits.view(-1, self.num_labels), labels.view(-1))
|
||||
elif self.config.problem_type == "multi_label_classification":
|
||||
loss_fct = BCEWithLogitsLoss()
|
||||
loss = loss_fct(pooled_logits, labels)
|
||||
if not return_dict:
|
||||
output = (pooled_logits,) + transformer_outputs[1:]
|
||||
return ((loss,) + output) if loss is not None else output
|
||||
|
||||
return SequenceClassifierOutputWithPast(
|
||||
loss=loss,
|
||||
logits=pooled_logits,
|
||||
past_key_values=transformer_outputs.past_key_values,
|
||||
hidden_states=transformer_outputs.hidden_states,
|
||||
attentions=transformer_outputs.attentions,
|
||||
)
|
@ -0,0 +1,74 @@
|
||||
"""
|
||||
PyTorch BLOOM model that implements several memory-efficient modes.
|
||||
Based on https://github.com/huggingface/transformers/commit/ca2a55e9dfb245527b5e1c954fec6ffbb7aef07b
|
||||
See commit history for authorship.
|
||||
"""
|
||||
|
||||
import torch
|
||||
import torch.nn.functional as F
|
||||
import torch.utils.checkpoint
|
||||
from hivemind import use_hivemind_log_handler
|
||||
from torch import nn
|
||||
from transformers import BloomConfig
|
||||
from transformers.utils import logging
|
||||
|
||||
use_hivemind_log_handler("in_root_logger")
|
||||
logger = logging.get_logger(__file__)
|
||||
|
||||
|
||||
class LMHead(nn.Module):
|
||||
"""
|
||||
The modified language modeling head which does not create extra tensor for the linear layer with weights tied to the input
|
||||
embeddings. Thus, it reduces initial memory consumption which might be crucial for large dictionaries.
|
||||
In addition, it provides an effcient way to deal with half-precision word embeddings on CPU.
|
||||
"""
|
||||
|
||||
def __init__(self, config: BloomConfig, word_embeddings: nn.Embedding):
|
||||
super().__init__()
|
||||
self.word_embeddings = word_embeddings
|
||||
self.chunk_size = config.chunk_size_for_efficient_fp16_on_cpu
|
||||
|
||||
@property
|
||||
def in_features(self) -> int:
|
||||
return self.word_embeddings.num_embeddings
|
||||
|
||||
@property
|
||||
def out_features(self) -> int:
|
||||
return self.word_embeddings.embedding_dim
|
||||
|
||||
@property
|
||||
def weight(self):
|
||||
return self.word_embeddings.weight
|
||||
|
||||
@property
|
||||
def bias(self):
|
||||
return None
|
||||
|
||||
def forward(self, hidden_states):
|
||||
word_embeddings = self.word_embeddings.weight
|
||||
|
||||
# We use 'chunked_forward' only when embeddings are in half-precision on CPU.
|
||||
if word_embeddings.dtype in [torch.float16, torch.bfloat16] and word_embeddings.device.type == "cpu":
|
||||
lm_logits = self.chunked_forward(hidden_states)
|
||||
else:
|
||||
# Switch dtype in case word_embeddings are fp16/bf16
|
||||
hidden_states = hidden_states.to(word_embeddings.dtype)
|
||||
lm_logits = F.linear(hidden_states, word_embeddings)
|
||||
return lm_logits
|
||||
|
||||
def chunked_forward(self, hidden_states):
|
||||
"""Splits word embeddings on chunks and iteratively casts them into fp32 to perform matmul more efficiently on CPU.
|
||||
chunk_size: provides trade-off between efficiency and extra memory consumption.
|
||||
"""
|
||||
assert self.chunk_size > 0, "Chunk size for chunked forward must be positive"
|
||||
|
||||
word_embeddings = self.word_embeddings.weight
|
||||
num_embeddings = self.word_embeddings.num_embeddings
|
||||
|
||||
hidden_states = hidden_states.float()
|
||||
output = torch.empty(*hidden_states.shape[:-1], num_embeddings)
|
||||
|
||||
for i in range(0, num_embeddings, self.chunk_size):
|
||||
chunk = word_embeddings[i : i + self.chunk_size].float()
|
||||
output[..., i : i + self.chunk_size] = F.linear(hidden_states, chunk)
|
||||
return output
|
@ -1,242 +0,0 @@
|
||||
"""
|
||||
Utility operations used in the the BLOOM model
|
||||
Based on https://github.com/huggingface/transformers/commit/ca2a55e9dfb245527b5e1c954fec6ffbb7aef07b
|
||||
See commit history for authorship.
|
||||
"""
|
||||
import math
|
||||
|
||||
import torch
|
||||
import torch.autograd
|
||||
import torch.nn.functional as F
|
||||
from torch import nn
|
||||
|
||||
|
||||
def split_tensor_along_last_dim(tensor, num_partitions, contiguous_split_chunks=False):
|
||||
"""Split a tensor along its last dimension.
|
||||
|
||||
Args:
|
||||
tensor: ([`torch.tensor`], *required*):
|
||||
input tensor to split
|
||||
num_partitions ([`int`], *required*):
|
||||
number of partitions to split the tensor
|
||||
contiguous_split_chunks ([`bool`], *optional*, default=`False`)::
|
||||
If True, make each chunk contiguous in memory.
|
||||
"""
|
||||
# Get the size and dimension.
|
||||
last_dim = tensor.dim() - 1
|
||||
numerator, denominator = tensor.size()[last_dim], num_partitions
|
||||
if not (numerator % denominator == 0):
|
||||
raise ValueError(f"{numerator} is not divisible by {denominator}")
|
||||
last_dim_size = numerator // denominator
|
||||
# Split.
|
||||
tensor_list = torch.split(tensor, last_dim_size, dim=last_dim)
|
||||
# Note: torch.split does not create contiguous tensors by default.
|
||||
if contiguous_split_chunks:
|
||||
return tuple(chunk.contiguous() for chunk in tensor_list)
|
||||
|
||||
return tensor_list
|
||||
|
||||
|
||||
def attention_mask_func(attention_scores, attention_mask, causal_mask):
|
||||
if attention_mask.dtype == torch.bool:
|
||||
attention_mask_bool = ~attention_mask
|
||||
else:
|
||||
attention_mask_bool = (1 - attention_mask).bool()
|
||||
|
||||
query_length, key_length, n_heads = attention_scores.size(2), attention_scores.size(3), attention_scores.size(1)
|
||||
padded_causal_mask = (
|
||||
attention_mask_bool[:, None, key_length - query_length : key_length, None]
|
||||
+ ~causal_mask[:, :, key_length - query_length : key_length, :key_length]
|
||||
).bool()
|
||||
padded_causal_mask = padded_causal_mask + attention_mask_bool[:, None, None, :key_length].bool()
|
||||
# Make use of floats
|
||||
return (
|
||||
attention_scores.masked_fill_(padded_causal_mask.expand(-1, n_heads, -1, -1), -10000.0),
|
||||
padded_causal_mask,
|
||||
)
|
||||
|
||||
|
||||
def build_alibi_tensor(
|
||||
max_seq_len: int, n_head: int, dtype: torch.dtype = torch.bfloat16, device: torch.device = torch.device("cpu")
|
||||
) -> torch.Tensor:
|
||||
"""
|
||||
Link to paper: https://arxiv.org/abs/2108.12409 Alibi tensor is not causal as the original paper mentions, it
|
||||
relies on a translation invariance of softmax for quick implementation: with l being a tensor, and a fixed value
|
||||
`softmax(l+a) = softmax(l)`. Based on
|
||||
https://github.com/ofirpress/attention_with_linear_biases/blob/a35aaca144e0eb6b789dfcb46784c4b8e31b7983/fairseq/models/transformer.py#L742
|
||||
Args:
|
||||
Returns tensor shaped (n_head, 1, max_seq_len)
|
||||
max_seq_len: (`int`, *required*):
|
||||
max sequence length
|
||||
n_head: (`int`, *required*):
|
||||
number of heads
|
||||
dtype: (`torch.dtype`, *optional*, default=`torch.bfloat16`):
|
||||
dtype of the output tensor
|
||||
device: (`torch.device`, *optional*, default=`torch.device('cpu')`):
|
||||
device of the output alibi tensor
|
||||
"""
|
||||
closest_power_of_2 = 2 ** math.floor(math.log2(n_head))
|
||||
base = torch.tensor(2 ** (-(2 ** -(math.log2(closest_power_of_2) - 3))), device=device, dtype=torch.float32)
|
||||
powers = torch.arange(1, 1 + closest_power_of_2, device=device, dtype=torch.int32)
|
||||
slopes = torch.pow(base, powers)
|
||||
|
||||
if closest_power_of_2 != n_head:
|
||||
extra_base = torch.tensor(
|
||||
2 ** (-(2 ** -(math.log2(2 * closest_power_of_2) - 3))), device=device, dtype=torch.float32
|
||||
)
|
||||
num_remaining_heads = min(closest_power_of_2, n_head - closest_power_of_2)
|
||||
extra_powers = torch.arange(1, 1 + 2 * num_remaining_heads, 2, device=device, dtype=torch.int32)
|
||||
slopes = torch.cat([slopes, torch.pow(extra_base, extra_powers)], dim=0)
|
||||
|
||||
lengths = torch.arange(max_seq_len, device=device, dtype=torch.int32)
|
||||
return (slopes.view(-1, 1, 1) * lengths.view(1, 1, -1)).to(dtype)
|
||||
|
||||
|
||||
def pre_process_alibi_for_pad(alibi: torch.Tensor, attention_mask: torch.Tensor):
|
||||
"""
|
||||
Args:
|
||||
Pre-process the alibi tensor for padding.
|
||||
alibi: ([`torch.tensor`], *required*):
|
||||
alibi tensor to pre-process
|
||||
attention_mask: ([`torch.tensor`], *required*):
|
||||
attention mask to pre-process
|
||||
"""
|
||||
assert attention_mask.ndim == 2, "mask should be [batch_size, seq_length]"
|
||||
unpadded_indices = torch.relu(attention_mask.cumsum(dim=1) - 1)
|
||||
# ^-- [batch, max_len], values correspond to element indices after removing padding
|
||||
# We shift the alibi tensor + replace all the values where attention_mask==0.0 by 0
|
||||
alibi = alibi.take_along_dim(unpadded_indices.unsqueeze(0), -1) * attention_mask.unsqueeze(0)
|
||||
return alibi.reshape(alibi.shape[0] * alibi.shape[1], 1, -1)
|
||||
|
||||
|
||||
def dropout_add(x, residual, prob, training):
|
||||
"""
|
||||
Dropout add function
|
||||
|
||||
Args:
|
||||
x (`torch.tensor`, *required*):
|
||||
input tensor
|
||||
residual (`torch.tensor`, *required*):
|
||||
esidual tensor
|
||||
prob (`float`, *required*):
|
||||
dropout probability
|
||||
training (`bool`, *required*):
|
||||
training mode
|
||||
"""
|
||||
out = nn.functional.dropout(x, p=prob, training=training)
|
||||
out = residual + out
|
||||
return out
|
||||
|
||||
|
||||
def bloom_gelu_forward(x):
|
||||
"""
|
||||
Custom bias GELU function. Adapted from Megatron-DeepSpeed code. Here we use a simple implementation (inference) to
|
||||
make the model jitable.
|
||||
|
||||
Args:
|
||||
x (`torch.tensor`, *required*):
|
||||
input hidden states
|
||||
"""
|
||||
return x * 0.5 * (1.0 + torch.tanh(0.79788456 * x * (1 + 0.044715 * x * x)))
|
||||
|
||||
|
||||
def bloom_gelu_back(g, x):
|
||||
"""
|
||||
gradient of tanh approximation of gelu gradient of actual gelu is: 0.5 * (1. + torch.erf(x * 0.70710678)) +
|
||||
0.3989423 * x * torch.exp(-0.5 * x * x)
|
||||
|
||||
Args:
|
||||
g (`torch.tensor`, *required*):
|
||||
gradient output tensor
|
||||
x (`torch.tensor`, *required*):
|
||||
input tensor
|
||||
"""
|
||||
x = x[0] # x is a tuple of 1 element, needs to unpack it first
|
||||
tanh_out = torch.tanh(0.79788456 * x * (1 + 0.044715 * x * x))
|
||||
# sqrt(2/pi) * 3 * 0.044715 -> 0.1070322243
|
||||
ff = 0.5 * x * ((1 - tanh_out * tanh_out) * (0.79788456 + 0.1070322243 * x * x)) + 0.5 * (1 + tanh_out)
|
||||
return ff * g
|
||||
|
||||
|
||||
class GeLUFunction(torch.autograd.Function):
|
||||
@staticmethod
|
||||
def forward(ctx, input):
|
||||
ctx.save_for_backward(input)
|
||||
return bloom_gelu_forward(input)
|
||||
|
||||
@staticmethod
|
||||
def backward(ctx, grad_output):
|
||||
input = ctx.saved_tensors
|
||||
tmp = bloom_gelu_back(grad_output, input)
|
||||
return tmp
|
||||
|
||||
|
||||
class BloomGelu(nn.Module):
|
||||
"""
|
||||
BloomBiasGelu wrapper function that make use of the simple function on inference mode to make the model
|
||||
torchscriptable and use the autograd function in training mode to get the accurate results of the gradients Partly
|
||||
copied from Megatron-DeepSpeed code and adapted for our needs
|
||||
|
||||
See here why autograd functions are not torchscriptable: https://github.com/pytorch/pytorch/issues/22329
|
||||
|
||||
"""
|
||||
|
||||
def __init__(self):
|
||||
super().__init__()
|
||||
|
||||
def forward(self, x):
|
||||
if self.training:
|
||||
return GeLUFunction.apply(x)
|
||||
else:
|
||||
return bloom_gelu_forward(x)
|
||||
|
||||
|
||||
class BloomScaledSoftmax(nn.Module):
|
||||
"""
|
||||
fused operation: scaling + mask + softmax
|
||||
|
||||
Args:
|
||||
scaled_masked_softmax_fusion (`bool`, *required*):
|
||||
flag to indicate user want to use softmax fusion
|
||||
mask_func (`function`, *required*):
|
||||
mask function to be applied.
|
||||
softmax_in_fp32 (`bool`, *required*):
|
||||
if true, softmax in performed at fp32 precision.
|
||||
scale (`float`, *required*):
|
||||
scaling factor used in input tensor scaling.
|
||||
"""
|
||||
|
||||
def __init__(self, scaled_masked_softmax_fusion, mask_func, softmax_in_fp32, scale):
|
||||
super().__init__()
|
||||
self.scaled_masked_softmax_fusion = scaled_masked_softmax_fusion
|
||||
self.mask_func = mask_func
|
||||
self.softmax_in_fp32 = softmax_in_fp32
|
||||
self.scale = scale
|
||||
|
||||
if not (self.scale is None or softmax_in_fp32):
|
||||
raise ValueError("softmax should be in fp32 when scaled")
|
||||
|
||||
def forward(self, input, mask, max_positions):
|
||||
input_dtype = input.dtype
|
||||
input_in_16bit = input_dtype in [torch.float16, torch.bfloat16]
|
||||
softmax_dtype = torch.float32 if self.softmax_in_fp32 else input_dtype
|
||||
|
||||
if self.scale is not None:
|
||||
input = input * self.scale
|
||||
|
||||
if mask is None:
|
||||
mask = torch.ones(input.shape[0], max_positions, dtype=torch.bool, device=input.device)
|
||||
|
||||
mask = mask.to(input.device)
|
||||
causal_mask = (
|
||||
torch.tril(torch.ones((max_positions, max_positions), dtype=torch.bool))
|
||||
.view(1, 1, max_positions, max_positions)
|
||||
.to(input.device)
|
||||
)
|
||||
mask_output, padded_causal_mask = self.mask_func(input, mask, causal_mask)
|
||||
probs = F.softmax(mask_output, dim=-1, dtype=softmax_dtype) * (~padded_causal_mask)
|
||||
|
||||
if input_in_16bit and self.softmax_in_fp32:
|
||||
probs = probs.to(dtype=input_dtype)
|
||||
|
||||
return probs
|
@ -0,0 +1,15 @@
|
||||
import pytest
|
||||
import torch
|
||||
from test_utils import MODEL_NAME
|
||||
|
||||
from petals.client import DistributedBloomConfig
|
||||
from petals.server.throughput import measure_compute_rps
|
||||
|
||||
|
||||
@pytest.mark.forked
|
||||
def test_throughput_basic():
|
||||
config = DistributedBloomConfig.from_pretrained(MODEL_NAME)
|
||||
throughput = measure_compute_rps(
|
||||
config, device=torch.device("cpu"), dtype=torch.bfloat16, load_in_8bit=False, n_steps=10
|
||||
)
|
||||
assert isinstance(throughput, float) and throughput > 0
|
Loading…
Reference in New Issue