2022-11-04 11:03:04 -06:00
|
|
|
import torch
|
|
|
|
|
|
|
|
from dataclasses import dataclass
|
2023-02-13 05:02:45 -07:00
|
|
|
from opentelemetry import trace
|
2023-01-17 01:10:22 -07:00
|
|
|
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, PreTrainedTokenizerBase
|
2023-04-20 03:07:40 -06:00
|
|
|
from typing import Optional, Tuple, List, Type, Dict
|
2022-11-04 11:03:04 -06:00
|
|
|
|
2023-03-07 10:52:22 -07:00
|
|
|
from text_generation_server.models import Model
|
|
|
|
from text_generation_server.models.types import (
|
|
|
|
GeneratedText,
|
|
|
|
Batch,
|
|
|
|
Generation,
|
|
|
|
PrefillTokens,
|
|
|
|
)
|
|
|
|
from text_generation_server.pb import generate_pb2
|
|
|
|
from text_generation_server.utils import NextTokenChooser, StoppingCriteria, Sampling
|
2022-11-04 11:03:04 -06:00
|
|
|
|
2023-02-13 05:02:45 -07:00
|
|
|
tracer = trace.get_tracer(__name__)
|
|
|
|
|
2022-11-04 11:03:04 -06:00
|
|
|
|
|
|
|
@dataclass
|
2023-01-17 01:10:22 -07:00
|
|
|
class Seq2SeqLMBatch(Batch):
|
2022-11-04 11:03:04 -06:00
|
|
|
batch_id: int
|
|
|
|
requests: List[generate_pb2.Request]
|
2023-04-20 03:07:40 -06:00
|
|
|
requests_idx_mapping: Dict[int, int]
|
2022-11-04 11:03:04 -06:00
|
|
|
|
2022-11-07 04:53:56 -07:00
|
|
|
# Encoder values
|
2023-04-24 06:15:42 -06:00
|
|
|
input_ids: Optional[torch.Tensor]
|
2022-11-04 11:03:04 -06:00
|
|
|
attention_mask: torch.Tensor
|
|
|
|
|
2022-11-07 04:53:56 -07:00
|
|
|
# Decoder values
|
2022-11-04 11:03:04 -06:00
|
|
|
decoder_input_ids: torch.Tensor
|
|
|
|
decoder_attention_mask: Optional[torch.Tensor]
|
|
|
|
encoder_last_hidden_state: Optional[torch.Tensor]
|
|
|
|
|
2023-04-20 03:07:40 -06:00
|
|
|
# All tokens
|
|
|
|
all_decoder_input_ids: List[torch.Tensor]
|
|
|
|
|
2022-11-07 04:53:56 -07:00
|
|
|
# Seq2SeqLM keeps track of both encoder and decoder attention keys and values
|
2022-11-04 11:03:04 -06:00
|
|
|
past_key_values: Optional[List[Tuple]]
|
|
|
|
|
2022-11-07 04:53:56 -07:00
|
|
|
# Lengths of all generations present in the batch
|
2022-11-04 11:03:04 -06:00
|
|
|
input_lengths: List[int]
|
|
|
|
decoder_input_lengths: List[int]
|
2023-04-11 08:38:22 -06:00
|
|
|
offsets: List[Optional[int]]
|
|
|
|
token_offsets: List[Optional[int]]
|
2022-11-04 11:03:04 -06:00
|
|
|
|
2022-11-07 04:53:56 -07:00
|
|
|
# Generation helpers
|
2022-11-04 11:03:04 -06:00
|
|
|
next_token_choosers: List[NextTokenChooser]
|
|
|
|
stopping_criterias: List[StoppingCriteria]
|
|
|
|
|
2022-11-07 04:53:56 -07:00
|
|
|
# Metadata used for padding
|
2022-11-04 11:03:04 -06:00
|
|
|
max_input_length: int
|
|
|
|
max_decoder_input_length: int
|
2023-02-24 04:49:21 -07:00
|
|
|
padding_right_offset: int
|
2022-11-04 11:03:04 -06:00
|
|
|
|
2023-04-24 09:59:00 -06:00
|
|
|
# Maximum number of tokens this batch will grow to
|
|
|
|
max_tokens: int
|
|
|
|
|
2023-01-17 01:10:22 -07:00
|
|
|
def to_pb(self) -> generate_pb2.Batch:
|
2023-03-07 10:52:22 -07:00
|
|
|
"""Convert a Seq2SeqLMBatch to a text_generation_server.v1.Batch protobuf"""
|
2022-11-04 11:03:04 -06:00
|
|
|
return generate_pb2.Batch(
|
2023-04-24 09:59:00 -06:00
|
|
|
id=self.batch_id,
|
|
|
|
requests=self.requests,
|
|
|
|
size=len(self),
|
|
|
|
max_tokens=self.max_tokens,
|
2022-11-04 11:03:04 -06:00
|
|
|
)
|
|
|
|
|
|
|
|
@classmethod
|
|
|
|
def from_pb(
|
2023-01-20 04:24:39 -07:00
|
|
|
cls,
|
|
|
|
pb: generate_pb2.Batch,
|
|
|
|
tokenizer: PreTrainedTokenizerBase,
|
|
|
|
device: torch.device,
|
2022-11-04 11:03:04 -06:00
|
|
|
) -> "Seq2SeqLMBatch":
|
2023-03-07 10:52:22 -07:00
|
|
|
"""Convert a text_generation_server.v1.Batch protobuf to a Seq2SeqLMBatch"""
|
2022-11-04 11:03:04 -06:00
|
|
|
inputs = []
|
|
|
|
next_token_choosers = []
|
|
|
|
stopping_criterias = []
|
|
|
|
|
|
|
|
decoder_input_lengths = []
|
2023-04-11 08:38:22 -06:00
|
|
|
offsets = []
|
|
|
|
token_offsets = []
|
2023-04-20 03:07:40 -06:00
|
|
|
requests_idx_mapping = {}
|
2022-11-04 11:03:04 -06:00
|
|
|
|
|
|
|
# Parse batch
|
2023-04-09 12:22:27 -06:00
|
|
|
max_truncation = 0
|
2023-02-24 04:49:21 -07:00
|
|
|
padding_right_offset = 0
|
2023-04-24 09:59:00 -06:00
|
|
|
max_decode_tokens = 0
|
2023-04-20 03:07:40 -06:00
|
|
|
for i, r in enumerate(pb.requests):
|
2022-11-04 11:03:04 -06:00
|
|
|
inputs.append(r.inputs)
|
2023-04-20 03:07:40 -06:00
|
|
|
requests_idx_mapping[r.id] = i
|
2022-11-04 11:03:04 -06:00
|
|
|
decoder_input_lengths.append(1)
|
2023-04-11 08:38:22 -06:00
|
|
|
offsets.append(None)
|
|
|
|
token_offsets.append(None)
|
2023-03-15 06:12:49 -06:00
|
|
|
next_token_choosers.append(NextTokenChooser.from_pb(r.parameters, device))
|
2023-02-24 04:49:21 -07:00
|
|
|
stopping_criteria = StoppingCriteria.from_pb(
|
|
|
|
r.stopping_parameters, tokenizer
|
|
|
|
)
|
|
|
|
stopping_criterias.append(stopping_criteria)
|
2023-04-09 12:22:27 -06:00
|
|
|
max_truncation = max(max_truncation, r.truncate)
|
2023-04-24 09:59:00 -06:00
|
|
|
max_decode_tokens += stopping_criteria.max_new_tokens
|
2023-02-24 04:49:21 -07:00
|
|
|
padding_right_offset = max(
|
|
|
|
padding_right_offset, stopping_criteria.max_new_tokens
|
2022-11-04 11:03:04 -06:00
|
|
|
)
|
|
|
|
|
2022-11-07 04:53:56 -07:00
|
|
|
# Tokenize batch
|
2022-11-04 11:03:04 -06:00
|
|
|
tokenized_inputs = tokenizer(
|
2022-12-12 10:25:22 -07:00
|
|
|
inputs,
|
|
|
|
return_tensors="pt",
|
|
|
|
padding=True,
|
2023-01-20 04:24:39 -07:00
|
|
|
return_token_type_ids=False,
|
2023-04-09 12:22:27 -06:00
|
|
|
truncation=True,
|
|
|
|
max_length=max_truncation,
|
2022-11-04 11:03:04 -06:00
|
|
|
).to(device)
|
2023-03-16 05:12:26 -06:00
|
|
|
|
|
|
|
input_lengths = tokenized_inputs["attention_mask"].sum(1)
|
|
|
|
max_input_length = input_lengths.max()
|
|
|
|
|
2023-04-20 03:07:40 -06:00
|
|
|
# Decoder sequence only contains the bos_token
|
|
|
|
decoder_input_ids = (
|
|
|
|
torch.tensor(tokenizer.bos_token_id, device=device)
|
|
|
|
.repeat(len(pb.requests))
|
|
|
|
.view(-1, 1)
|
|
|
|
)
|
|
|
|
all_decoder_input_ids = decoder_input_ids.view(-1).split(1)
|
2022-11-04 11:03:04 -06:00
|
|
|
|
2023-04-24 09:59:00 -06:00
|
|
|
max_tokens = len(inputs) * max_input_length + max_decode_tokens
|
|
|
|
|
2022-11-04 11:03:04 -06:00
|
|
|
return cls(
|
|
|
|
batch_id=pb.id,
|
|
|
|
requests=pb.requests,
|
2023-04-20 03:07:40 -06:00
|
|
|
requests_idx_mapping=requests_idx_mapping,
|
2022-11-04 11:03:04 -06:00
|
|
|
input_ids=tokenized_inputs["input_ids"],
|
|
|
|
attention_mask=tokenized_inputs["attention_mask"],
|
|
|
|
decoder_input_ids=decoder_input_ids,
|
2023-04-20 03:07:40 -06:00
|
|
|
all_decoder_input_ids=list(all_decoder_input_ids),
|
2022-11-04 11:03:04 -06:00
|
|
|
decoder_attention_mask=None,
|
|
|
|
encoder_last_hidden_state=None,
|
|
|
|
past_key_values=None,
|
2023-03-16 05:12:26 -06:00
|
|
|
input_lengths=input_lengths.tolist(),
|
2022-11-04 11:03:04 -06:00
|
|
|
decoder_input_lengths=decoder_input_lengths,
|
2023-04-11 08:38:22 -06:00
|
|
|
offsets=offsets,
|
|
|
|
token_offsets=token_offsets,
|
2022-11-04 11:03:04 -06:00
|
|
|
next_token_choosers=next_token_choosers,
|
|
|
|
stopping_criterias=stopping_criterias,
|
2023-03-16 05:12:26 -06:00
|
|
|
max_input_length=max_input_length.item(),
|
2022-11-04 11:03:04 -06:00
|
|
|
max_decoder_input_length=1,
|
2023-02-24 04:49:21 -07:00
|
|
|
padding_right_offset=padding_right_offset,
|
2023-04-24 09:59:00 -06:00
|
|
|
max_tokens=max_tokens,
|
2022-11-04 11:03:04 -06:00
|
|
|
)
|
|
|
|
|
2023-04-20 03:07:40 -06:00
|
|
|
@tracer.start_as_current_span("filter")
|
|
|
|
def filter(
|
|
|
|
self, requests: List[generate_pb2.Request]
|
|
|
|
) -> Optional["Seq2SeqLMBatch"]:
|
|
|
|
if len(requests) == 0:
|
|
|
|
raise ValueError("Batch must have at least one request")
|
|
|
|
if len(requests) == len(self):
|
|
|
|
return self
|
|
|
|
|
|
|
|
keep_indices = []
|
|
|
|
|
|
|
|
# New values after filtering
|
|
|
|
requests_idx_mapping = {}
|
|
|
|
input_lengths = []
|
|
|
|
decoder_input_lengths = []
|
|
|
|
offsets = []
|
|
|
|
token_offsets = []
|
|
|
|
|
|
|
|
all_decoder_input_ids = []
|
|
|
|
|
|
|
|
next_token_choosers = []
|
|
|
|
stopping_criterias = []
|
|
|
|
|
|
|
|
max_input_length = 0
|
|
|
|
max_decoder_input_length = 0
|
2023-04-24 06:15:42 -06:00
|
|
|
padding_right_offset = 0
|
2023-04-20 03:07:40 -06:00
|
|
|
|
2023-04-27 01:57:28 -06:00
|
|
|
total_remaining_decode_tokens = 0
|
2023-04-24 09:59:00 -06:00
|
|
|
|
2023-04-20 03:07:40 -06:00
|
|
|
for i, r in enumerate(requests):
|
|
|
|
idx = self.requests_idx_mapping[r.id]
|
|
|
|
requests_idx_mapping[r.id] = i
|
|
|
|
keep_indices.append(idx)
|
|
|
|
|
|
|
|
offsets.append(self.offsets[idx])
|
|
|
|
token_offsets.append(self.token_offsets[idx])
|
|
|
|
|
|
|
|
all_decoder_input_ids.append(self.all_decoder_input_ids[idx])
|
|
|
|
|
|
|
|
request_input_length = self.input_lengths[idx]
|
|
|
|
input_lengths.append(request_input_length)
|
|
|
|
max_input_length = max(max_input_length, request_input_length)
|
|
|
|
|
|
|
|
request_decoder_input_length = self.decoder_input_lengths[idx]
|
|
|
|
decoder_input_lengths.append(request_decoder_input_length)
|
|
|
|
max_decoder_input_length = max(
|
|
|
|
max_decoder_input_length, request_decoder_input_length
|
|
|
|
)
|
|
|
|
|
|
|
|
next_token_choosers.append(self.next_token_choosers[idx])
|
2023-04-24 09:59:00 -06:00
|
|
|
stopping_criteria = self.stopping_criterias[idx]
|
|
|
|
stopping_criterias.append(stopping_criteria)
|
2023-04-27 01:57:28 -06:00
|
|
|
remaining_decode_tokens = (
|
2023-04-24 09:59:00 -06:00
|
|
|
stopping_criteria.max_new_tokens - stopping_criteria.current_tokens
|
|
|
|
)
|
2023-04-27 01:57:28 -06:00
|
|
|
total_remaining_decode_tokens += remaining_decode_tokens
|
|
|
|
padding_right_offset = max(padding_right_offset, remaining_decode_tokens)
|
2023-04-20 03:07:40 -06:00
|
|
|
|
|
|
|
# Apply indices to input_ids, attention mask, past key values and other items that need to be cached
|
2023-04-24 06:15:42 -06:00
|
|
|
self.decoder_input_ids = self.decoder_input_ids[keep_indices]
|
|
|
|
self.attention_mask = self.attention_mask[keep_indices, -max_input_length:]
|
2023-04-20 03:07:40 -06:00
|
|
|
if self.decoder_attention_mask is not None:
|
2023-04-24 06:15:42 -06:00
|
|
|
self.decoder_attention_mask = self.decoder_attention_mask[
|
2023-04-24 09:59:00 -06:00
|
|
|
keep_indices,
|
|
|
|
-(self.padding_right_offset + max_decoder_input_length) : (
|
|
|
|
self.decoder_attention_mask.shape[1] - self.padding_right_offset
|
|
|
|
)
|
|
|
|
+ padding_right_offset,
|
2023-04-24 06:15:42 -06:00
|
|
|
]
|
|
|
|
|
2023-04-24 09:59:00 -06:00
|
|
|
self.encoder_last_hidden_state = self.encoder_last_hidden_state[
|
|
|
|
keep_indices, -max_input_length:
|
|
|
|
]
|
2023-04-24 06:15:42 -06:00
|
|
|
|
|
|
|
# Ensure that past_key_values tensors can be updated in-place
|
|
|
|
if type(self.past_key_values[0]) == tuple:
|
2023-04-24 09:59:00 -06:00
|
|
|
self.past_key_values = [
|
|
|
|
[t for t in layer] for layer in self.past_key_values
|
|
|
|
]
|
2023-04-24 06:15:42 -06:00
|
|
|
|
|
|
|
decoder_past_seq_len = max_decoder_input_length - 1
|
|
|
|
for layer in self.past_key_values:
|
|
|
|
layer[0] = layer[0][keep_indices, :, -decoder_past_seq_len:]
|
|
|
|
layer[1] = layer[1][keep_indices, :, -decoder_past_seq_len:]
|
|
|
|
layer[2] = layer[2][keep_indices, :, -max_input_length:]
|
|
|
|
layer[3] = layer[3][keep_indices, :, -max_input_length:]
|
|
|
|
|
2023-04-24 09:59:00 -06:00
|
|
|
max_tokens = (
|
|
|
|
len(requests) * (max_input_length + max_decoder_input_length)
|
|
|
|
+ remaining_decode_tokens
|
|
|
|
)
|
|
|
|
|
2023-04-24 06:15:42 -06:00
|
|
|
self.requests = requests
|
|
|
|
self.requests_idx_mapping = requests_idx_mapping
|
|
|
|
self.input_ids = None
|
|
|
|
self.all_decoder_input_ids = all_decoder_input_ids
|
|
|
|
self.input_lengths = input_lengths
|
|
|
|
self.decoder_input_lengths = decoder_input_lengths
|
|
|
|
self.offsets = offsets
|
|
|
|
self.token_offsets = token_offsets
|
|
|
|
self.next_token_choosers = next_token_choosers
|
|
|
|
self.stopping_criterias = stopping_criterias
|
|
|
|
self.max_input_length = max_input_length
|
|
|
|
self.max_decoder_input_length = max_decoder_input_length
|
|
|
|
self.padding_right_offset = padding_right_offset
|
2023-04-24 09:59:00 -06:00
|
|
|
self.max_tokens = max_tokens
|
2023-04-24 06:15:42 -06:00
|
|
|
|
|
|
|
return self
|
2023-04-20 03:07:40 -06:00
|
|
|
|
2022-11-04 11:03:04 -06:00
|
|
|
@classmethod
|
2023-02-13 05:02:45 -07:00
|
|
|
@tracer.start_as_current_span("concatenate")
|
2022-11-04 11:03:04 -06:00
|
|
|
def concatenate(cls, batches: List["Seq2SeqLMBatch"]) -> "Seq2SeqLMBatch":
|
2022-11-07 04:53:56 -07:00
|
|
|
"""Concatenate multiple batches together by padding internal torch tensors"""
|
|
|
|
|
2022-11-04 11:03:04 -06:00
|
|
|
# Used for padding
|
2023-02-24 04:49:21 -07:00
|
|
|
total_batch_size = 0
|
|
|
|
max_input_length = 0
|
|
|
|
max_decoder_input_length = 0
|
|
|
|
padding_right_offset = 0
|
|
|
|
for batch in batches:
|
2023-04-20 03:07:40 -06:00
|
|
|
total_batch_size += len(batch)
|
2023-02-24 04:49:21 -07:00
|
|
|
max_input_length = max(max_input_length, batch.max_input_length)
|
|
|
|
max_decoder_input_length = max(
|
|
|
|
max_decoder_input_length, batch.max_decoder_input_length
|
|
|
|
)
|
|
|
|
padding_right_offset = max(padding_right_offset, batch.padding_right_offset)
|
2022-11-04 11:03:04 -06:00
|
|
|
|
|
|
|
# Batch attributes
|
|
|
|
requests = []
|
2023-04-20 03:07:40 -06:00
|
|
|
requests_idx_mapping = {}
|
|
|
|
all_decoder_input_ids = []
|
2022-11-04 11:03:04 -06:00
|
|
|
input_lengths = []
|
|
|
|
decoder_input_lengths = []
|
2023-04-11 08:38:22 -06:00
|
|
|
offsets = []
|
|
|
|
token_offsets = []
|
2022-11-04 11:03:04 -06:00
|
|
|
next_token_choosers = []
|
|
|
|
stopping_criterias = []
|
2023-04-24 09:59:00 -06:00
|
|
|
max_tokens = 0
|
2022-11-04 11:03:04 -06:00
|
|
|
|
2022-11-07 04:53:56 -07:00
|
|
|
# Batch tensors
|
2022-11-04 11:03:04 -06:00
|
|
|
attention_mask = None
|
|
|
|
decoder_input_ids = None
|
|
|
|
decoder_attention_mask = None
|
|
|
|
encoder_last_hidden_state = None
|
|
|
|
past_key_values = []
|
|
|
|
|
|
|
|
# Used for slicing correctly inside the tensors
|
|
|
|
# Equivalent to a cumsum on batch sizes
|
|
|
|
start_index = 0
|
2022-11-07 04:53:56 -07:00
|
|
|
|
2022-11-04 11:03:04 -06:00
|
|
|
for i, batch in enumerate(batches):
|
2022-11-07 04:53:56 -07:00
|
|
|
# Extend all list attributes
|
2022-11-04 11:03:04 -06:00
|
|
|
requests.extend(batch.requests)
|
2023-04-20 03:07:40 -06:00
|
|
|
all_decoder_input_ids.extend(batch.all_decoder_input_ids)
|
2022-11-04 11:03:04 -06:00
|
|
|
input_lengths.extend(batch.input_lengths)
|
|
|
|
decoder_input_lengths.extend(batch.decoder_input_lengths)
|
2023-04-11 08:38:22 -06:00
|
|
|
offsets.extend(batch.offsets)
|
|
|
|
token_offsets.extend(batch.token_offsets)
|
2022-11-04 11:03:04 -06:00
|
|
|
next_token_choosers.extend(batch.next_token_choosers)
|
|
|
|
stopping_criterias.extend(batch.stopping_criterias)
|
|
|
|
|
2023-04-20 03:07:40 -06:00
|
|
|
if i == 0:
|
|
|
|
requests_idx_mapping = batch.requests_idx_mapping
|
|
|
|
else:
|
|
|
|
# We need to offset the mapping for each batch by the cumulative batch size
|
|
|
|
for k, v in batch.requests_idx_mapping.items():
|
|
|
|
requests_idx_mapping[k] = v + start_index
|
|
|
|
|
2022-11-04 11:03:04 -06:00
|
|
|
# Slicing end index for this batch
|
2023-04-20 03:07:40 -06:00
|
|
|
end_index = start_index + len(batch)
|
2022-11-04 11:03:04 -06:00
|
|
|
|
|
|
|
# We only concatenate batches that did at least one step
|
|
|
|
if batch.encoder_last_hidden_state is None:
|
|
|
|
raise ValueError("Batch encoder_last_hidden_state cannot be None")
|
|
|
|
|
2022-11-07 04:53:56 -07:00
|
|
|
# Create padded tensor
|
2022-11-04 11:03:04 -06:00
|
|
|
if attention_mask is None:
|
2023-01-17 01:10:22 -07:00
|
|
|
attention_mask = batch.attention_mask.new_zeros(
|
2022-11-04 11:03:04 -06:00
|
|
|
(total_batch_size, max_input_length),
|
|
|
|
)
|
2022-11-07 04:53:56 -07:00
|
|
|
# Copy to correct indices
|
2022-11-04 11:03:04 -06:00
|
|
|
attention_mask[
|
|
|
|
start_index:end_index, -batch.max_input_length :
|
|
|
|
] = batch.attention_mask[:, -batch.max_input_length :]
|
|
|
|
|
2022-11-07 04:53:56 -07:00
|
|
|
# Create padded tensor
|
2022-11-04 11:03:04 -06:00
|
|
|
if decoder_input_ids is None:
|
2023-01-17 01:10:22 -07:00
|
|
|
decoder_input_ids = batch.decoder_input_ids.new_zeros(
|
2023-04-20 03:07:40 -06:00
|
|
|
(total_batch_size, 1),
|
2022-11-04 11:03:04 -06:00
|
|
|
)
|
2022-11-07 04:53:56 -07:00
|
|
|
# Copy to correct indices
|
2023-04-20 03:07:40 -06:00
|
|
|
decoder_input_ids[start_index:end_index] = batch.decoder_input_ids
|
2022-11-04 11:03:04 -06:00
|
|
|
|
2022-11-07 04:53:56 -07:00
|
|
|
# Create padded tensor
|
2022-11-04 11:03:04 -06:00
|
|
|
if decoder_attention_mask is None:
|
2023-01-17 01:10:22 -07:00
|
|
|
# As decoder_attention_mask might not exist, we use `batch.attention_mask` for device here
|
|
|
|
decoder_attention_mask = batch.attention_mask.new_zeros(
|
2023-02-24 04:49:21 -07:00
|
|
|
(total_batch_size, max_decoder_input_length + padding_right_offset),
|
2022-11-04 11:03:04 -06:00
|
|
|
)
|
2022-11-07 04:53:56 -07:00
|
|
|
# If the decoder mask does not exist yet, all generations started at the same time and we never concatenated
|
|
|
|
# this batch. All generations are of length `batch.max_decoder_input_length`.
|
2023-02-24 04:49:21 -07:00
|
|
|
left_offset = max_decoder_input_length - batch.max_decoder_input_length
|
2022-11-04 11:03:04 -06:00
|
|
|
if batch.decoder_attention_mask is None:
|
|
|
|
decoder_attention_mask[
|
2023-02-24 04:49:21 -07:00
|
|
|
start_index:end_index,
|
|
|
|
left_offset:-padding_right_offset,
|
2022-11-04 11:03:04 -06:00
|
|
|
] = 1
|
2022-11-07 04:53:56 -07:00
|
|
|
# If it exists, we need to index
|
2022-11-04 11:03:04 -06:00
|
|
|
else:
|
2023-02-24 04:49:21 -07:00
|
|
|
batch_left_offset = (
|
|
|
|
batch.decoder_attention_mask.shape[1]
|
2023-02-24 07:55:57 -07:00
|
|
|
- batch.max_decoder_input_length
|
|
|
|
- batch.padding_right_offset
|
2023-02-24 04:49:21 -07:00
|
|
|
)
|
2022-11-04 11:03:04 -06:00
|
|
|
decoder_attention_mask[
|
2023-02-24 04:49:21 -07:00
|
|
|
start_index:end_index,
|
|
|
|
left_offset:-padding_right_offset,
|
|
|
|
] = batch.decoder_attention_mask[
|
|
|
|
:,
|
|
|
|
batch_left_offset : -batch.padding_right_offset,
|
|
|
|
]
|
2022-11-04 11:03:04 -06:00
|
|
|
|
2022-11-07 04:53:56 -07:00
|
|
|
# Create padded tensor
|
2022-11-04 11:03:04 -06:00
|
|
|
if encoder_last_hidden_state is None:
|
2023-01-17 01:10:22 -07:00
|
|
|
encoder_last_hidden_state = batch.encoder_last_hidden_state.new_zeros(
|
2022-11-04 11:03:04 -06:00
|
|
|
(
|
|
|
|
total_batch_size,
|
|
|
|
max_input_length,
|
|
|
|
batch.encoder_last_hidden_state.shape[-1],
|
|
|
|
),
|
|
|
|
)
|
|
|
|
|
2022-11-07 04:53:56 -07:00
|
|
|
# Copy to correct indices
|
2022-11-04 11:03:04 -06:00
|
|
|
encoder_last_hidden_state[
|
2022-12-08 10:49:33 -07:00
|
|
|
start_index:end_index, -batch.max_input_length :, :
|
|
|
|
] = batch.encoder_last_hidden_state[:, -batch.max_input_length :, :]
|
2023-04-24 06:15:42 -06:00
|
|
|
batch.encoder_last_hidden_state = None
|
2022-11-04 11:03:04 -06:00
|
|
|
|
2023-04-24 06:15:42 -06:00
|
|
|
# Ensure that we can update tensors in-place
|
|
|
|
if type(batch.past_key_values[0]) == tuple:
|
2023-04-24 09:59:00 -06:00
|
|
|
batch.past_key_values = [
|
|
|
|
[t for t in layer] for layer in batch.past_key_values
|
|
|
|
]
|
2022-11-04 11:03:04 -06:00
|
|
|
|
2023-04-24 09:59:00 -06:00
|
|
|
# Add eventual padding tokens that were added while concatenating
|
|
|
|
max_tokens += batch.max_tokens + (
|
|
|
|
max_input_length
|
|
|
|
- batch.max_input_length
|
|
|
|
+ max_decoder_input_length
|
|
|
|
- batch.max_decoder_input_length
|
|
|
|
) * len(batch)
|
2022-11-04 11:03:04 -06:00
|
|
|
|
2023-04-27 01:57:28 -06:00
|
|
|
start_index = end_index
|
|
|
|
|
2023-04-24 06:15:42 -06:00
|
|
|
# Determine shapes for new past kv tensors
|
|
|
|
first_past_kvs = batches[0].past_key_values
|
|
|
|
_, num_heads, _, head_dim = first_past_kvs[0][0].shape
|
2022-11-04 11:03:04 -06:00
|
|
|
|
2023-04-24 06:15:42 -06:00
|
|
|
padded_dec_t_shape = (
|
|
|
|
total_batch_size,
|
|
|
|
num_heads,
|
|
|
|
(max_decoder_input_length - 1),
|
|
|
|
head_dim,
|
|
|
|
)
|
2022-11-04 11:03:04 -06:00
|
|
|
|
2023-04-24 06:15:42 -06:00
|
|
|
padded_enc_t_shape = (
|
|
|
|
total_batch_size,
|
|
|
|
num_heads,
|
|
|
|
max_input_length,
|
|
|
|
head_dim,
|
|
|
|
)
|
2022-11-04 11:03:04 -06:00
|
|
|
|
2023-04-24 06:15:42 -06:00
|
|
|
# Iterate over attention layers
|
|
|
|
for j in range(len(first_past_kvs)):
|
|
|
|
past_key_values.append([])
|
|
|
|
|
|
|
|
# Decoder past
|
|
|
|
for k in range(0, 2):
|
|
|
|
# Initialize tensors
|
|
|
|
padded_past_values = first_past_kvs[j][k].new_zeros(padded_dec_t_shape)
|
|
|
|
past_key_values[j].append(padded_past_values)
|
|
|
|
|
|
|
|
start_index = 0
|
|
|
|
for batch in batches:
|
|
|
|
t = batch.past_key_values[j][k]
|
|
|
|
# Clear reference to the original tensor
|
|
|
|
batch.past_key_values[j][k] = None
|
|
|
|
# Slicing end index for this batch
|
|
|
|
end_index = start_index + len(batch)
|
|
|
|
# We slice the past keys and values to remove the padding from previous batches
|
|
|
|
past_seq_len = batch.max_decoder_input_length - 1
|
2023-04-24 09:59:00 -06:00
|
|
|
padded_past_values[start_index:end_index, :, -past_seq_len:, :] = t[
|
|
|
|
:, :, -past_seq_len:, :
|
|
|
|
]
|
2023-04-24 06:15:42 -06:00
|
|
|
del t
|
|
|
|
|
|
|
|
start_index = end_index
|
|
|
|
|
|
|
|
# Encoder past
|
|
|
|
for k in range(2, 4):
|
|
|
|
# Initialize tensors
|
|
|
|
padded_past_values = first_past_kvs[j][k].new_zeros(padded_enc_t_shape)
|
|
|
|
past_key_values[j].append(padded_past_values)
|
|
|
|
|
|
|
|
start_index = 0
|
|
|
|
for batch in batches:
|
|
|
|
t = batch.past_key_values[j][k]
|
|
|
|
# Clear reference to the original tensor
|
|
|
|
batch.past_key_values[j][k] = None
|
|
|
|
# Slicing end index for this batch
|
|
|
|
end_index = start_index + len(batch)
|
|
|
|
# We slice the past keys and values to remove the padding from previous batches
|
|
|
|
padded_past_values[
|
2023-04-24 09:59:00 -06:00
|
|
|
start_index:end_index, :, -batch.max_input_length :, :
|
|
|
|
] = t[:, :, -batch.max_input_length :, :]
|
2023-04-24 06:15:42 -06:00
|
|
|
del t
|
2022-11-04 11:03:04 -06:00
|
|
|
|
2023-04-24 06:15:42 -06:00
|
|
|
start_index = end_index
|
2022-11-04 11:03:04 -06:00
|
|
|
|
|
|
|
return cls(
|
|
|
|
batch_id=batches[0].batch_id,
|
|
|
|
requests=requests,
|
2023-04-20 03:07:40 -06:00
|
|
|
requests_idx_mapping=requests_idx_mapping,
|
2023-02-24 04:49:21 -07:00
|
|
|
input_ids=None,
|
2022-11-04 11:03:04 -06:00
|
|
|
attention_mask=attention_mask,
|
|
|
|
decoder_input_ids=decoder_input_ids,
|
2023-04-20 03:07:40 -06:00
|
|
|
all_decoder_input_ids=all_decoder_input_ids,
|
2022-11-04 11:03:04 -06:00
|
|
|
decoder_attention_mask=decoder_attention_mask,
|
|
|
|
encoder_last_hidden_state=encoder_last_hidden_state,
|
|
|
|
past_key_values=past_key_values,
|
|
|
|
input_lengths=input_lengths,
|
|
|
|
decoder_input_lengths=decoder_input_lengths,
|
2023-04-11 08:38:22 -06:00
|
|
|
offsets=offsets,
|
|
|
|
token_offsets=token_offsets,
|
2022-11-04 11:03:04 -06:00
|
|
|
next_token_choosers=next_token_choosers,
|
|
|
|
stopping_criterias=stopping_criterias,
|
|
|
|
max_input_length=max_input_length,
|
|
|
|
max_decoder_input_length=max_decoder_input_length,
|
2023-02-24 04:49:21 -07:00
|
|
|
padding_right_offset=padding_right_offset,
|
2023-04-24 09:59:00 -06:00
|
|
|
max_tokens=max_tokens,
|
2022-11-04 11:03:04 -06:00
|
|
|
)
|
|
|
|
|
2023-01-31 09:04:00 -07:00
|
|
|
def __len__(self):
|
|
|
|
return len(self.requests)
|
|
|
|
|
2022-11-04 11:03:04 -06:00
|
|
|
|
|
|
|
class Seq2SeqLM(Model):
|
2023-04-12 04:03:10 -06:00
|
|
|
def __init__(
|
|
|
|
self,
|
|
|
|
model_id: str,
|
|
|
|
revision: Optional[str] = None,
|
|
|
|
quantize: bool = False,
|
|
|
|
decode_buffer: int = 3,
|
|
|
|
):
|
2022-11-04 11:03:04 -06:00
|
|
|
if torch.cuda.is_available():
|
|
|
|
device = torch.device("cuda")
|
|
|
|
dtype = torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float32
|
|
|
|
else:
|
2022-12-08 10:49:33 -07:00
|
|
|
if quantize:
|
|
|
|
raise ValueError("quantization is not available on CPU")
|
|
|
|
|
2022-11-04 11:03:04 -06:00
|
|
|
device = torch.device("cpu")
|
|
|
|
dtype = torch.float32
|
|
|
|
|
|
|
|
self.model = AutoModelForSeq2SeqLM.from_pretrained(
|
2023-02-03 04:43:37 -07:00
|
|
|
model_id,
|
2023-01-31 10:53:56 -07:00
|
|
|
revision=revision,
|
2022-11-04 11:03:04 -06:00
|
|
|
torch_dtype=dtype,
|
|
|
|
device_map="auto" if torch.cuda.is_available() else None,
|
2022-11-07 04:53:56 -07:00
|
|
|
load_in_8bit=quantize,
|
2022-11-04 11:03:04 -06:00
|
|
|
).eval()
|
2023-01-31 10:53:56 -07:00
|
|
|
tokenizer = AutoTokenizer.from_pretrained(
|
2023-04-11 08:38:22 -06:00
|
|
|
model_id, revision=revision, padding_side="left", truncation_side="left"
|
2023-01-31 10:53:56 -07:00
|
|
|
)
|
2022-11-04 11:03:04 -06:00
|
|
|
tokenizer.bos_token_id = self.model.config.decoder_start_token_id
|
|
|
|
|
|
|
|
super(Seq2SeqLM, self).__init__(
|
2023-04-21 07:36:29 -06:00
|
|
|
tokenizer=tokenizer,
|
|
|
|
requires_padding=True,
|
|
|
|
dtype=dtype,
|
|
|
|
device=device,
|
|
|
|
decode_buffer=decode_buffer,
|
2022-11-04 11:03:04 -06:00
|
|
|
)
|
|
|
|
|
|
|
|
@property
|
|
|
|
def batch_type(self) -> Type[Seq2SeqLMBatch]:
|
|
|
|
return Seq2SeqLMBatch
|
|
|
|
|
2023-01-20 04:24:39 -07:00
|
|
|
def decode(self, decoder_ids: List[int]) -> str:
|
2023-03-06 05:22:58 -07:00
|
|
|
return self.tokenizer.decode(
|
|
|
|
decoder_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False
|
|
|
|
)
|
2023-01-20 04:24:39 -07:00
|
|
|
|
2022-11-04 11:03:04 -06:00
|
|
|
def forward(
|
|
|
|
self,
|
|
|
|
input_ids,
|
|
|
|
attention_mask,
|
|
|
|
decoder_input_ids,
|
|
|
|
decoder_attention_mask: Optional,
|
|
|
|
encoder_last_hidden_state: Optional,
|
|
|
|
past_key_values: Optional = None,
|
|
|
|
) -> Tuple[
|
|
|
|
torch.Tensor,
|
|
|
|
torch.Tensor,
|
|
|
|
List[Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]],
|
|
|
|
]:
|
|
|
|
# Model Forward
|
|
|
|
outputs = self.model.forward(
|
|
|
|
input_ids=input_ids,
|
|
|
|
attention_mask=attention_mask,
|
|
|
|
decoder_input_ids=decoder_input_ids,
|
|
|
|
decoder_attention_mask=decoder_attention_mask,
|
2022-11-07 04:53:56 -07:00
|
|
|
encoder_outputs=encoder_last_hidden_state,
|
2022-11-04 11:03:04 -06:00
|
|
|
past_key_values=past_key_values,
|
|
|
|
use_cache=True,
|
|
|
|
)
|
|
|
|
return (
|
|
|
|
outputs.logits,
|
|
|
|
outputs.encoder_last_hidden_state,
|
|
|
|
outputs.past_key_values,
|
|
|
|
)
|
|
|
|
|
2023-02-13 05:02:45 -07:00
|
|
|
@tracer.start_as_current_span("generate_token")
|
2022-11-04 11:03:04 -06:00
|
|
|
def generate_token(
|
|
|
|
self, batch: Seq2SeqLMBatch
|
2023-01-31 09:04:00 -07:00
|
|
|
) -> Tuple[List[Generation], Optional[Seq2SeqLMBatch]]:
|
2023-02-24 04:49:21 -07:00
|
|
|
if batch.decoder_attention_mask is not None:
|
|
|
|
# slice to the correct shape
|
|
|
|
decoder_attention_mask = batch.decoder_attention_mask[
|
|
|
|
:, : -batch.padding_right_offset
|
|
|
|
]
|
|
|
|
else:
|
|
|
|
decoder_attention_mask = None
|
|
|
|
|
|
|
|
# Wrap `encoder_last_hidden_state` because for some reason, Transformers does a `encoder_last_hidden_state[0]`
|
|
|
|
# internally...
|
|
|
|
if batch.encoder_last_hidden_state is not None:
|
|
|
|
encoder_last_hidden_state = [batch.encoder_last_hidden_state]
|
|
|
|
else:
|
2023-04-20 03:07:40 -06:00
|
|
|
encoder_last_hidden_state = None
|
2023-02-24 04:49:21 -07:00
|
|
|
|
2023-02-07 07:38:22 -07:00
|
|
|
logits, encoder_last_hidden_state, past = self.forward(
|
|
|
|
batch.input_ids,
|
|
|
|
batch.attention_mask,
|
2023-04-20 03:07:40 -06:00
|
|
|
batch.decoder_input_ids,
|
2023-02-24 04:49:21 -07:00
|
|
|
decoder_attention_mask,
|
|
|
|
encoder_last_hidden_state,
|
2023-02-07 07:38:22 -07:00
|
|
|
batch.past_key_values,
|
2022-11-04 11:03:04 -06:00
|
|
|
)
|
|
|
|
|
|
|
|
# Finished requests
|
2023-01-31 09:04:00 -07:00
|
|
|
generations: List[Generation] = []
|
2023-04-20 03:07:40 -06:00
|
|
|
stopped = True
|
2022-11-04 11:03:04 -06:00
|
|
|
|
|
|
|
# Zipped iterator
|
|
|
|
iterator = zip(
|
|
|
|
batch.requests,
|
|
|
|
batch.input_lengths,
|
2023-04-11 08:38:22 -06:00
|
|
|
batch.offsets,
|
|
|
|
batch.token_offsets,
|
2022-11-04 11:03:04 -06:00
|
|
|
batch.decoder_input_lengths,
|
|
|
|
logits,
|
|
|
|
batch.next_token_choosers,
|
|
|
|
batch.stopping_criterias,
|
2023-04-20 03:07:40 -06:00
|
|
|
batch.all_decoder_input_ids,
|
2022-11-04 11:03:04 -06:00
|
|
|
)
|
|
|
|
|
|
|
|
# For each member of the batch
|
|
|
|
for i, (
|
|
|
|
request,
|
|
|
|
input_length,
|
2023-04-11 08:38:22 -06:00
|
|
|
offset,
|
|
|
|
token_offset,
|
2022-11-04 11:03:04 -06:00
|
|
|
decoder_input_length,
|
|
|
|
logits,
|
|
|
|
next_token_chooser,
|
|
|
|
stopping_criteria,
|
2023-04-20 03:07:40 -06:00
|
|
|
all_decoder_input_ids,
|
2022-11-04 11:03:04 -06:00
|
|
|
) in enumerate(iterator):
|
|
|
|
# Select next token
|
2023-02-01 07:58:42 -07:00
|
|
|
next_token_id, logprobs = next_token_chooser(
|
2023-04-20 03:07:40 -06:00
|
|
|
all_decoder_input_ids.view(1, -1), logits
|
2023-02-01 07:58:42 -07:00
|
|
|
)
|
2022-11-04 11:03:04 -06:00
|
|
|
|
|
|
|
# Append next token to decoder tokens
|
2023-04-20 03:07:40 -06:00
|
|
|
all_decoder_input_ids = torch.cat(
|
|
|
|
[all_decoder_input_ids, next_token_id.squeeze(1)]
|
|
|
|
)
|
2022-12-15 09:03:56 -07:00
|
|
|
new_decoder_input_length = decoder_input_length + 1
|
|
|
|
|
2023-01-31 09:04:00 -07:00
|
|
|
# Generated token
|
|
|
|
next_token_logprob = logprobs[-1, next_token_id]
|
|
|
|
next_token_id_squeezed = next_token_id.squeeze()
|
2023-04-11 08:38:22 -06:00
|
|
|
next_token_text, offset, token_offset = self.decode_token(
|
2023-04-20 03:07:40 -06:00
|
|
|
all_decoder_input_ids, offset, token_offset
|
2023-01-31 09:04:00 -07:00
|
|
|
)
|
2022-11-04 11:03:04 -06:00
|
|
|
|
|
|
|
# Evaluate stopping criteria
|
2023-01-31 09:04:00 -07:00
|
|
|
stop, reason = stopping_criteria(next_token_id, next_token_text)
|
|
|
|
|
2022-12-12 10:25:22 -07:00
|
|
|
if stop:
|
2022-12-15 09:03:56 -07:00
|
|
|
# Slice with decoder_input_length to remove padding
|
|
|
|
# Decode all tokens
|
2023-04-20 03:07:40 -06:00
|
|
|
output_text = self.decode(all_decoder_input_ids[-decoder_input_length:])
|
2023-01-30 07:36:16 -07:00
|
|
|
|
|
|
|
# Get seed
|
|
|
|
if isinstance(next_token_chooser.choice, Sampling):
|
|
|
|
seed = next_token_chooser.choice.seed
|
|
|
|
else:
|
|
|
|
seed = None
|
|
|
|
|
2023-01-31 09:04:00 -07:00
|
|
|
generated_text = GeneratedText(
|
|
|
|
output_text, stopping_criteria.current_tokens, reason, seed
|
2022-11-04 11:03:04 -06:00
|
|
|
)
|
|
|
|
else:
|
2023-01-31 09:04:00 -07:00
|
|
|
# Keep request in the batch
|
|
|
|
generated_text = None
|
2023-04-20 03:07:40 -06:00
|
|
|
stopped = False
|
2022-11-04 11:03:04 -06:00
|
|
|
|
2023-01-31 09:04:00 -07:00
|
|
|
# Prefill
|
|
|
|
if stopping_criteria.current_tokens == 1:
|
|
|
|
prefill_tokens = PrefillTokens(
|
2023-02-24 07:55:57 -07:00
|
|
|
[self.tokenizer.bos_token_id],
|
|
|
|
[float("nan")],
|
|
|
|
[self.tokenizer.bos_token],
|
2023-01-31 09:04:00 -07:00
|
|
|
)
|
|
|
|
else:
|
|
|
|
prefill_tokens = None
|
|
|
|
|
|
|
|
generation = Generation(
|
|
|
|
request.id,
|
|
|
|
prefill_tokens,
|
|
|
|
next_token_id_squeezed,
|
|
|
|
next_token_logprob,
|
|
|
|
next_token_text,
|
2023-02-24 09:20:00 -07:00
|
|
|
next_token_id_squeezed.item() in self.all_special_ids,
|
2023-01-31 09:04:00 -07:00
|
|
|
generated_text,
|
|
|
|
)
|
|
|
|
|
|
|
|
generations.append(generation)
|
|
|
|
|
2023-04-20 03:07:40 -06:00
|
|
|
# Update values
|
|
|
|
batch.decoder_input_ids[i] = next_token_id
|
|
|
|
batch.all_decoder_input_ids[i] = all_decoder_input_ids
|
|
|
|
batch.input_lengths[i] = input_length
|
|
|
|
batch.decoder_input_lengths[i] = new_decoder_input_length
|
|
|
|
batch.offsets[i] = offset
|
|
|
|
batch.token_offsets[i] = token_offset
|
|
|
|
batch.max_input_length = max(batch.max_input_length, input_length)
|
|
|
|
batch.max_decoder_input_length = max(
|
|
|
|
batch.max_decoder_input_length, new_decoder_input_length
|
|
|
|
)
|
|
|
|
|
2022-11-04 11:03:04 -06:00
|
|
|
# We finished all generations in the batch; there is no next batch
|
2023-04-20 03:07:40 -06:00
|
|
|
if stopped:
|
2023-01-31 09:04:00 -07:00
|
|
|
return generations, None
|
2022-11-04 11:03:04 -06:00
|
|
|
|
2023-04-20 03:07:40 -06:00
|
|
|
# We don't need input_ids after the prefill forward
|
|
|
|
batch.input_ids = None
|
|
|
|
batch.encoder_last_hidden_state = encoder_last_hidden_state
|
|
|
|
batch.past_key_values = past
|
2023-02-24 04:49:21 -07:00
|
|
|
# Update decoder_attention_mask as we added a new token to input_ids
|
2023-04-20 03:07:40 -06:00
|
|
|
if batch.decoder_attention_mask is not None:
|
|
|
|
batch.decoder_attention_mask[:, -batch.padding_right_offset] = 1
|
|
|
|
batch.padding_right_offset -= 1
|
2022-11-04 11:03:04 -06:00
|
|
|
|
2023-04-20 03:07:40 -06:00
|
|
|
return generations, batch
|