hf_text-generation-inference/server/text_generation/models/causal_lm.py

480 lines
18 KiB
Python
Raw Normal View History

import torch
from dataclasses import dataclass
from transformers import AutoTokenizer, AutoModelForCausalLM, PreTrainedTokenizerBase
2022-11-07 04:53:56 -07:00
from typing import Optional, Tuple, List, Type
from text_generation.models import Model
from text_generation.models.types import Batch, PrefillTokens, Generation, GeneratedText
from text_generation.pb import generate_pb2
from text_generation.utils import NextTokenChooser, StoppingCriteria, Sampling
@dataclass
class CausalLMBatch(Batch):
batch_id: int
requests: List[generate_pb2.Request]
2022-11-07 04:53:56 -07:00
# Decoder values
input_ids: torch.Tensor
attention_mask: torch.Tensor
2023-01-20 07:35:22 -07:00
position_ids: torch.Tensor
2022-11-07 04:53:56 -07:00
past_key_values: Optional[List[Tuple]]
# All tokens
all_input_ids: List[torch.Tensor]
2022-11-07 04:53:56 -07:00
# Lengths of all generations present in the batch
input_lengths: List[int]
# Generation helpers
next_token_choosers: List[NextTokenChooser]
stopping_criterias: List[StoppingCriteria]
2022-11-07 04:53:56 -07:00
# Metadata used for padding
size: int
max_sequence_length: int
2022-12-08 10:49:33 -07:00
# Past metadata
keys_head_dim_last: bool = True
def to_pb(self) -> generate_pb2.Batch:
return generate_pb2.Batch(
id=self.batch_id,
requests=self.requests,
size=self.size,
)
@classmethod
def from_pb(
2023-01-20 04:24:39 -07:00
cls,
pb: generate_pb2.Batch,
tokenizer: PreTrainedTokenizerBase,
device: torch.device,
) -> "CausalLMBatch":
inputs = []
next_token_choosers = []
stopping_criterias = []
2022-11-07 04:53:56 -07:00
input_lengths = []
# Parse batch
for r in pb.requests:
inputs.append(r.inputs)
2022-11-07 04:53:56 -07:00
input_lengths.append(r.input_length)
2023-01-31 06:30:33 -07:00
next_token_choosers.append(NextTokenChooser.from_pb(r.parameters, device))
stopping_criterias.append(
2022-12-12 10:25:22 -07:00
StoppingCriteria.from_pb(r.stopping_parameters, tokenizer)
)
pad_to_multiple_of = 8 if device.type == "cuda" else None
2022-11-07 04:53:56 -07:00
tokenized_inputs = tokenizer(
2022-12-12 10:25:22 -07:00
inputs,
return_tensors="pt",
padding=True,
pad_to_multiple_of=pad_to_multiple_of,
2023-01-20 04:24:39 -07:00
return_token_type_ids=False,
).to(device)
2023-01-20 07:35:22 -07:00
position_ids = tokenized_inputs["attention_mask"].long().cumsum(-1) - 1
position_ids.masked_fill_(tokenized_inputs["attention_mask"] == 0, 1)
2022-11-07 04:53:56 -07:00
all_input_ids = tokenized_inputs["input_ids"].unsqueeze(-1)
return cls(
batch_id=pb.id,
requests=pb.requests,
2022-11-07 04:53:56 -07:00
input_ids=tokenized_inputs["input_ids"],
attention_mask=tokenized_inputs["attention_mask"],
2023-01-20 07:35:22 -07:00
position_ids=position_ids,
2022-11-07 04:53:56 -07:00
past_key_values=None,
all_input_ids=all_input_ids,
2022-11-07 04:53:56 -07:00
input_lengths=input_lengths,
next_token_choosers=next_token_choosers,
stopping_criterias=stopping_criterias,
size=pb.size,
2022-11-07 04:53:56 -07:00
max_sequence_length=max(input_lengths),
)
@classmethod
def concatenate(cls, batches: List["CausalLMBatch"]) -> "CausalLMBatch":
# Used for padding
total_batch_size = sum(batch.size for batch in batches)
max_sequence_length = max(batch.max_sequence_length for batch in batches)
# Batch attributes
requests = []
2022-11-07 04:53:56 -07:00
input_lengths = []
all_input_ids = []
next_token_choosers = []
stopping_criterias = []
2022-11-07 04:53:56 -07:00
# Batch tensors
input_ids = None
attention_mask = None
2023-01-20 07:35:22 -07:00
position_ids = None
2022-11-07 04:53:56 -07:00
past_key_values = []
# Used for slicing correctly inside the tensors
# Equivalent to a cumsum on batch sizes
start_index = 0
for i, batch in enumerate(batches):
requests.extend(batch.requests)
2022-11-07 04:53:56 -07:00
input_lengths.extend(batch.input_lengths)
all_input_ids.extend(batch.all_input_ids)
next_token_choosers.extend(batch.next_token_choosers)
stopping_criterias.extend(batch.stopping_criterias)
# Slicing end index for this batch
end_index = start_index + batch.size
# We only concatenate batches that did at least one step
2022-12-12 10:25:22 -07:00
if batch.past_key_values is None:
raise ValueError("only concatenate prefilled batches")
2022-11-07 04:53:56 -07:00
# Create empty tensor
# input_ids is always of shape [batch_size, 1]
# We do not need to pad it
if input_ids is None:
input_ids = batch.input_ids.new_empty((total_batch_size, 1))
2022-11-07 04:53:56 -07:00
# Copy to correct indices
input_ids[start_index:end_index] = batch.input_ids
# Create padded tensor
if attention_mask is None:
attention_mask = batch.attention_mask.new_zeros(
(total_batch_size, max_sequence_length),
)
# We need to slice the attention mask to remove padding from previous steps
2022-11-07 04:53:56 -07:00
attention_mask[
start_index:end_index, -batch.max_sequence_length :
2022-11-07 04:53:56 -07:00
] = batch.attention_mask[:, -batch.max_sequence_length :]
2023-01-20 07:35:22 -07:00
# Create empty tensor
# position_ids is always of shape [batch_size, 1]
if position_ids is None:
position_ids = batch.position_ids.new_empty((total_batch_size, 1))
position_ids[start_index:end_index] = batch.position_ids
2022-11-07 04:53:56 -07:00
for j, past in enumerate(batch.past_key_values):
past_keys, past_values = past
# Shenanigans to get dimensions because BLOOM outputs a past with a different shape
# BLOOM Keys: [batch_size * num_heads, head_dim, seq_length]
# BLOOM Values: [batch_size * num_heads, seq_length, head_dim]
past_keys = past_keys.view(batch.size, -1, *past_keys.shape[-2:])
past_values = past_values.view(batch.size, -1, *past_values.shape[-2:])
2022-12-01 11:31:54 -07:00
_, num_heads, padded_sequence_length, head_dim = past_values.shape
2022-12-01 11:31:54 -07:00
padded_past_values_shape = (
total_batch_size,
num_heads,
max_sequence_length - 1,
2022-12-01 11:31:54 -07:00
head_dim,
)
2022-12-08 10:49:33 -07:00
if batch.keys_head_dim_last:
padded_past_keys_shape = padded_past_values_shape
else:
# seq_length is last for BLOOM
2022-12-01 11:31:54 -07:00
padded_past_keys_shape = (
total_batch_size,
num_heads,
head_dim,
2022-12-01 11:31:54 -07:00
max_sequence_length - 1,
)
# This will run only once per layer
2022-11-07 04:53:56 -07:00
if j == len(past_key_values):
padded_past_keys = past_keys.new_zeros(padded_past_keys_shape)
padded_past_values = past_values.new_zeros(padded_past_values_shape)
past_key_values.append((padded_past_keys, padded_past_values))
# We slice the past keys and values to remove the padding from previous batches
2022-12-08 10:49:33 -07:00
if batch.keys_head_dim_last:
2022-12-01 11:31:54 -07:00
past_key_values[j][0][
start_index:end_index,
:,
-(batch.max_sequence_length - 1) :,
:,
2022-12-01 11:31:54 -07:00
] = past_keys[:, :, -(batch.max_sequence_length - 1) :, :]
else:
2022-12-01 11:31:54 -07:00
past_key_values[j][0][
start_index:end_index,
:,
:,
-(batch.max_sequence_length - 1) :,
2022-12-01 11:31:54 -07:00
] = past_keys[:, :, :, -(batch.max_sequence_length - 1) :]
past_key_values[j][1][
start_index:end_index, :, -(batch.max_sequence_length - 1) :, :
] = past_values[:, :, -(batch.max_sequence_length - 1) :, :]
start_index += batch.size
return cls(
batch_id=batches[0].batch_id,
requests=requests,
input_ids=input_ids,
2022-11-07 04:53:56 -07:00
attention_mask=attention_mask,
2023-01-20 07:35:22 -07:00
position_ids=position_ids,
2022-11-07 04:53:56 -07:00
past_key_values=past_key_values,
all_input_ids=all_input_ids,
2022-11-07 04:53:56 -07:00
input_lengths=input_lengths,
next_token_choosers=next_token_choosers,
stopping_criterias=stopping_criterias,
size=total_batch_size,
max_sequence_length=max_sequence_length,
2022-12-08 10:49:33 -07:00
keys_head_dim_last=batches[0].keys_head_dim_last,
)
def __len__(self):
return len(self.requests)
class CausalLM(Model):
2023-01-31 10:53:56 -07:00
def __init__(self, model_name: str, revision: Optional[str] = None, quantize=False):
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")
device = torch.device("cpu")
dtype = torch.float32
2023-01-31 10:53:56 -07:00
tokenizer = AutoTokenizer.from_pretrained(
model_name, revision=revision, padding_side="left"
)
self.model = AutoModelForCausalLM.from_pretrained(
model_name,
2023-01-31 10:53:56 -07:00
revision=revision,
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,
).eval()
2022-12-08 10:49:33 -07:00
tokenizer.pad_token_id = (
self.model.config.pad_token_id
if self.model.config.pad_token_id is not None
else self.model.config.eos_token_id
)
super(CausalLM, self).__init__(
tokenizer=tokenizer,
device=device,
)
@property
def batch_type(self) -> Type[CausalLMBatch]:
return CausalLMBatch
2023-01-20 04:24:39 -07:00
def decode(self, generated_ids: List[int]) -> str:
return self.tokenizer.decode(
generated_ids, skip_special_tokens=True, cleanup_tokenization_spaces=False
)
def forward(
2023-01-20 07:35:22 -07:00
self, input_ids, attention_mask, position_ids, past_key_values: Optional = None
) -> Tuple[torch.Tensor, List[Tuple[torch.Tensor, torch.Tensor]]]:
# Model Forward
outputs = self.model.forward(
input_ids=input_ids,
attention_mask=attention_mask,
2023-01-20 07:35:22 -07:00
position_ids=position_ids,
past_key_values=past_key_values,
use_cache=True,
)
return outputs.logits, outputs.past_key_values
def generate_token(
self, batch: CausalLMBatch
) -> Tuple[List[Generation], Optional[CausalLMBatch]]:
# For some reason, inference_mode does not work well with GLOO which we use on CPU
context_manager = (
torch.no_grad if self.device.type == "cpu" else torch.inference_mode
)
with context_manager():
2022-11-07 04:53:56 -07:00
logits, past = self.forward(
batch.input_ids,
batch.attention_mask,
batch.position_ids,
batch.past_key_values,
2022-11-07 04:53:56 -07:00
)
# List of indices to cache
next_batch_keep_indices = []
2022-11-07 04:53:56 -07:00
# New values for next forward
next_batch_input_lengths = []
next_batch_input_ids = []
next_batch_all_input_ids = []
2022-11-07 04:53:56 -07:00
# Metadata
next_batch_size = 0
next_batch_max_sequence_length = 0
# Results
generations: List[Generation] = []
# Zipped iterator
iterator = zip(
batch.requests,
2022-11-07 04:53:56 -07:00
batch.input_lengths,
logits,
batch.next_token_choosers,
batch.stopping_criterias,
batch.all_input_ids,
)
# For each member of the batch
for i, (
request,
input_length,
logits,
next_token_chooser,
stopping_criteria,
2022-12-15 09:03:56 -07:00
all_input_ids,
) in enumerate(iterator):
# Select next token
2022-12-15 09:03:56 -07:00
tokens, logprobs = next_token_chooser(all_input_ids, logits)
next_token_id = tokens[-1].view(1, 1)
# Append next token to all tokens
all_input_ids = torch.cat([all_input_ids, next_token_id])
2022-12-15 09:03:56 -07:00
new_input_length = input_length + 1
# Generated token
next_token_logprob = logprobs[-1, next_token_id]
next_token_id_squeezed = next_token_id.squeeze()
next_token_text = self.tokenizer.decode(
next_token_id_squeezed,
clean_up_tokenization_spaces=False,
skip_special_tokens=False,
)
# Evaluate stopping criteria
2022-12-16 08:03:39 -07:00
stop, reason = stopping_criteria(
next_token_id_squeezed,
next_token_text,
2022-12-16 08:03:39 -07:00
)
2022-12-12 10:25:22 -07:00
if stop:
2023-01-20 04:24:39 -07:00
# Decode generated tokens
generated_text = self.decode(
all_input_ids[-stopping_criteria.current_tokens :, 0]
)
2023-01-20 04:24:39 -07:00
output_text = request.inputs + generated_text
2022-12-15 09:03:56 -07:00
# Get seed
if isinstance(next_token_chooser.choice, Sampling):
seed = next_token_chooser.choice.seed
else:
seed = None
generated_text = GeneratedText(
output_text, stopping_criteria.current_tokens, reason, seed
)
else:
# Keep request in the batch
generated_text = None
next_batch_keep_indices.append(i)
next_batch_input_ids.append(next_token_id)
2022-12-15 09:03:56 -07:00
next_batch_all_input_ids.append(all_input_ids)
next_batch_size += 1
2022-11-07 04:53:56 -07:00
next_batch_input_lengths.append(new_input_length)
next_batch_max_sequence_length = max(
next_batch_max_sequence_length, new_input_length
)
# Prefill
if stopping_criteria.current_tokens == 1:
# Remove generated token to only have prefill and add nan for first prompt token
prefill_logprobs = [float("nan")] + logprobs.gather(
1, all_input_ids[1:]
).squeeze(1)[-new_input_length:-1].tolist()
prefill_token_ids = all_input_ids[-new_input_length:-1]
prefill_texts = self.tokenizer.batch_decode(
prefill_token_ids,
clean_up_tokenization_spaces=False,
skip_special_tokens=False,
)
prefill_tokens = PrefillTokens(
prefill_token_ids, prefill_logprobs, prefill_texts
)
else:
prefill_tokens = None
generation = Generation(
request.id,
prefill_tokens,
next_token_id_squeezed,
next_token_logprob,
next_token_text,
generated_text,
)
generations.append(generation)
# We finished all generations in the batch; there is no next batch
if not next_batch_keep_indices:
return generations, None
2022-11-07 04:53:56 -07:00
next_batch_input_ids = torch.cat(next_batch_input_ids, dim=0)
# If we finished at least one generation, we need to evict the indices of the generations that finished
# from the values of the next batch
if len(next_batch_keep_indices) != len(batch):
# Apply indices to attention mask, past key values and other items that need to be cached
2022-11-07 04:53:56 -07:00
next_batch_attention_mask = batch.attention_mask[next_batch_keep_indices]
2023-01-20 07:35:22 -07:00
next_batch_position_ids = batch.position_ids[next_batch_keep_indices]
# Force past to be of dim [batch_size, num_heads, ...] for easy indexing
2022-11-07 04:53:56 -07:00
next_batch_past_key_values = [
[
t.view(batch.size, -1, *t.shape[-2:])[next_batch_keep_indices]
for t in layer
]
for layer in past
]
next_batch_requests = [batch.requests[i] for i in next_batch_keep_indices]
next_batch_next_token_choosers = [
batch.next_token_choosers[i] for i in next_batch_keep_indices
]
next_batch_stopping_criterias = [
batch.stopping_criterias[i] for i in next_batch_keep_indices
]
else:
2022-11-07 04:53:56 -07:00
next_batch_attention_mask = batch.attention_mask
2023-01-20 07:35:22 -07:00
next_batch_position_ids = batch.position_ids
2022-11-07 04:53:56 -07:00
next_batch_past_key_values = past
next_batch_requests = batch.requests
next_batch_next_token_choosers = batch.next_token_choosers
next_batch_stopping_criterias = batch.stopping_criterias
# Update attention_mask with padding as we added a new token to input_ids
2022-11-07 04:53:56 -07:00
next_batch_attention_mask = torch.cat(
[
2022-11-07 04:53:56 -07:00
next_batch_attention_mask,
next_batch_attention_mask.new_ones(next_batch_size, 1),
],
dim=1,
)
2023-01-20 07:35:22 -07:00
# Update position_ids
next_batch_position_ids = next_batch_position_ids[:, -1:] + 1
next_batch = CausalLMBatch(
batch_id=batch.batch_id,
requests=next_batch_requests,
input_ids=next_batch_input_ids,
2022-11-07 04:53:56 -07:00
attention_mask=next_batch_attention_mask,
2023-01-20 07:35:22 -07:00
position_ids=next_batch_position_ids,
2022-11-07 04:53:56 -07:00
past_key_values=next_batch_past_key_values,
all_input_ids=next_batch_all_input_ids,
2022-11-07 04:53:56 -07:00
input_lengths=next_batch_input_lengths,
next_token_choosers=next_batch_next_token_choosers,
stopping_criterias=next_batch_stopping_criterias,
size=next_batch_size,
max_sequence_length=next_batch_max_sequence_length,
2022-12-08 10:49:33 -07:00
keys_head_dim_last=batch.keys_head_dim_last,
)
return generations, next_batch