Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

format code with black, yapf, autopep8 and isort #100

Open
wants to merge 1 commit into
base: master
Choose a base branch
from
Open

format code with black, yapf, autopep8 and isort #100

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
101 changes: 65 additions & 36 deletions src/grouped_sampling/batch_end_to_end_pipeline.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,7 +1,7 @@
from typing import List, Optional

import torch
from torch import Tensor, long, inference_mode, full, argmax, int8, eq, ones_like
from torch import Tensor, argmax, eq, full, inference_mode, int8, long, ones_like
from transformers import GenerationConfig

from src.grouped_sampling.logits_vec_to_token import LogitVectorToTokenPipeLine
Expand All @@ -10,12 +10,13 @@


class BatchEndToEndSingleSequencePipeLine:

def __init__(
self,
model_name: str,
load_in_8bit: bool = False,
model_kwargs: Optional[dict] = None,
generation_config: Optional[GenerationConfig] = None,
self,
model_name: str,
load_in_8bit: bool = False,
model_kwargs: Optional[dict] = None,
generation_config: Optional[GenerationConfig] = None,
):
"""
Create a new BatchEndToEndSingleSequencePipeLine.
Expand All @@ -33,15 +34,22 @@ def __init__(
TypeError: If one of the arguments is of the wrong type.
"""
if not isinstance(model_name, str):
raise TypeError(f"model_name should be a string, got {type(model_name)}")
raise TypeError(
f"model_name should be a string, got {type(model_name)}")
if not isinstance(load_in_8bit, bool):
raise TypeError(f"load_in_8bit should be a bool, got {type(load_in_8bit)}")
raise TypeError(
f"load_in_8bit should be a bool, got {type(load_in_8bit)}")
if model_kwargs is not None and not isinstance(model_kwargs, dict):
raise TypeError(f"model_kwargs should be a dict or None, got {type(model_kwargs)}")
if generation_config is not None and not isinstance(generation_config, GenerationConfig):
raise TypeError(f"generation_config should be a GenerationConfig or None, got {type(generation_config)}")
raise TypeError(
f"model_kwargs should be a dict or None, got {type(model_kwargs)}"
)
if generation_config is not None and not isinstance(
generation_config, GenerationConfig):
raise TypeError(
f"generation_config should be a GenerationConfig or None, got {type(generation_config)}"
)
if not load_in_8bit:
torch.set_float32_matmul_precision('high')
torch.set_float32_matmul_precision("high")
self.tokenizer = get_tokenizer(model_name)
if model_kwargs is None:
self.model = get_model(
Expand All @@ -57,18 +65,23 @@ def __init__(
self.device: torch.device = self.model.device
self.max_total_len = self.model.config.max_position_embeddings
if generation_config is None:
generation_config = GenerationConfig.from_model_config(self.model.config)
self.logit_to_token_pipeline = LogitVectorToTokenPipeLine(generation_config=generation_config)
generation_config = GenerationConfig.from_model_config(
self.model.config)
self.logit_to_token_pipeline = LogitVectorToTokenPipeLine(
generation_config=generation_config)

def tokenize_and_pad(
self,
prompts: List[str],
output_length: int,
self,
prompts: List[str],
output_length: int,
) -> Tensor:
"""A helper function that converts a list of strings to a padded tensor of tokens."""
prompt_tokens_list = self.tokenizer.batch_encode_plus(
prompts, add_special_tokens=True, padding=True, return_attention_mask=False,
)['input_ids']
prompts,
add_special_tokens=True,
padding=True,
return_attention_mask=False,
)["input_ids"]
prompt_tokens = torch.tensor(
prompt_tokens_list,
device=self.device,
Expand All @@ -77,7 +90,9 @@ def tokenize_and_pad(
max_input_length = prompt_tokens.shape[1]
padding_length: int = max_input_length + output_length - 1
if padding_length > self.max_total_len:
raise ValueError(f"padding_length should be at most {self.max_total_len}, got {padding_length}")
raise ValueError(
f"padding_length should be at most {self.max_total_len}, got {padding_length}"
)
batch_size = prompt_tokens.shape[0]
extra_padding = full(
fill_value=self.tokenizer.pad_token_id,
Expand All @@ -88,31 +103,39 @@ def tokenize_and_pad(
return torch.cat([prompt_tokens, extra_padding], dim=1)

def tokens_batch_to_logit_matrices(
self,
padded_tokens: Tensor,
output_length: int,
self,
padded_tokens: Tensor,
output_length: int,
) -> List[Tensor]:
"""
Given a batch of prompts where each prompt is a sequence of tokens, and an output_length,
returns the logits matrices of shape (batch_size, output_length, vocab_size)
where logits[i] is the logits matrix of the i-th prompt.
"""
if padded_tokens.dim() != 2:
raise ValueError(f"tokens should be a 2D tensor, got {padded_tokens.dim()}D tensor")
raise ValueError(
f"tokens should be a 2D tensor, got {padded_tokens.dim()}D tensor"
)
if padded_tokens.requires_grad:
raise ValueError("tokens should not require grad")
if not isinstance(output_length, int):
raise TypeError(f"output_length should be an int, got {type(output_length)}")
raise TypeError(
f"output_length should be an int, got {type(output_length)}")
if output_length <= 0:
raise ValueError(f"output_length should be positive, got {output_length}")
attenction_mask = ones_like(padded_tokens, dtype=torch.long, device=self.device, requires_grad=False)
raise ValueError(
f"output_length should be positive, got {output_length}")
attenction_mask = ones_like(padded_tokens,
dtype=torch.long,
device=self.device,
requires_grad=False)
all_logits = self.model(
output_attentions=False,
output_hidden_states=False,
input_ids=padded_tokens,
attention_mask=attenction_mask,
).logits
padding_int_tokens = eq(padded_tokens, self.tokenizer.pad_token_id).to(int8)
padding_int_tokens = eq(padded_tokens,
self.tokenizer.pad_token_id).to(int8)
last_non_pad_indices = argmax(padding_int_tokens, dim=1) - 1
relavent_logits = [
all_logits[i, index:index + output_length]
Expand All @@ -122,11 +145,15 @@ def tokens_batch_to_logit_matrices(

def _validate_output_length(self, output_length: int) -> None:
if not isinstance(output_length, int):
raise TypeError(f"output_length should be an int, got {type(output_length)}")
raise TypeError(
f"output_length should be an int, got {type(output_length)}")
if output_length <= 0:
raise ValueError(f"output_length should be positive, got {output_length}")
raise ValueError(
f"output_length should be positive, got {output_length}")
if output_length >= self.max_total_len:
raise ValueError(f"output_length should be smaller than {self.max_total_len}, got {output_length}")
raise ValueError(
f"output_length should be smaller than {self.max_total_len}, got {output_length}"
)

@staticmethod
def _validate_prompts(prompts: List[str]):
Expand All @@ -137,9 +164,9 @@ def _validate_prompts(prompts: List[str]):

@inference_mode()
def genearte_batch(
self,
prompts: List[str] | str,
output_length: int,
self,
prompts: List[str] | str,
output_length: int,
) -> List[str]:
"""
Given a batch of prompts and output length, generates a list of output strings.
Expand All @@ -160,9 +187,11 @@ def genearte_batch(
return []
self._validate_prompts(prompts)
padded_tokens = self.tokenize_and_pad(prompts, output_length)
logits = self.tokens_batch_to_logit_matrices(padded_tokens, output_length)
logits = self.tokens_batch_to_logit_matrices(padded_tokens,
output_length)
output_tokens = self.logit_to_token_pipeline.batch_to_tokens(
input_ids=padded_tokens,
batch=logits,
)
return self.tokenizer.batch_decode(output_tokens, skip_special_tokens=True)
return self.tokenizer.batch_decode(output_tokens,
skip_special_tokens=True)
Loading