Code

Terminator strings for generate()#28932

Code

Merged

Rocketknight1 merged 68 commits intomainhuggingface/transformers:mainfrom terminator_strings_for_generatehuggingface/transformers:terminator_strings_for_generateCopy head branch name to clipboardApr 22, 2024

+529-4

Commits

Commits on Apr 11, 2024

• stash commit (will discard all of this)

  Rocketknight1

  committed
  262537f

• stash commit

  Rocketknight1

  committed
  cfa538b

• First commit - needs a lot of testing!

  Rocketknight1

  committed
  127182a

• Add a test

  Rocketknight1

  committed
  8cd6059

• Fix imports and make the tests actually test something

  Rocketknight1

  committed
  5fde7ae

• Tests pass!

  Rocketknight1

  committed
  ff02b0c

• Rearrange test

  Rocketknight1

  committed
  4ce1aba

• Add comments (but it's still a bit confusing)

  Rocketknight1

  committed
  1742b68

• Stop storing the tokenizer

  Rocketknight1

  committed
  9fb77e3

• Comment fixup

  Rocketknight1

  committed
  667d6d8

• Fix for input_ids with a single sequence

  Rocketknight1

  committed
  070a76e

• Update tests to test single sequences

  Rocketknight1

  committed
  4c436f2

• make fixup

  Rocketknight1

  committed
  78b0f24

• Fix incorrect use of isin()

  Rocketknight1

  committed
  8ee5762

• Expand tests to catch more cases

  Rocketknight1

  committed
  9f43a2a

• Expand tests to catch more cases

  Rocketknight1

  committed
  f0fa707

• make fixup

  Rocketknight1

  committed
  5bcf5e4

• Fix length calculation and update tests

  Rocketknight1

  committed
  8cca9a4

• Handle Ġ as a space replacement too

  Rocketknight1

  committed
  ec6f726

• Update src/transformers/generation/stopping_criteria.py

  Show description for 0e632c2
  

  Rocketknight1

  and

  gante

  committed
  0e632c2

• Add optimizations from Joao's suggestion

  Rocketknight1

  committed
  ac1135c

• Remove TODO

  Rocketknight1

  committed
  2731827

• Update src/transformers/generation/stopping_criteria.py

  Show description for 9213298
  

  Rocketknight1

  and

  gante

  committed
  9213298

• Update tests/generation/test_stopping_criteria.py

  Show description for f48522e
  

  Rocketknight1

  and

  gante

  committed
  f48522e

• make fixup

  Rocketknight1

  committed
  7a772b8

• Rename some variables and remove some debugging clauses for clarity

  Rocketknight1

  committed
  c604a2b

• Add tests for the sub-methods

  Rocketknight1

  committed
  7dd346a

• Clarify one test slightly

  Rocketknight1

  committed
  641ba72

• Add stop_strings to GenerationConfig

  Rocketknight1

  committed
  f6721a5

• generate() supports stop_string arg, asks for tokenizer if not provided

  Rocketknight1

  committed
  8772bcb

• make fixup

  Rocketknight1

  committed
  e423417

• Cleanup code and rename variables for clarity

  Rocketknight1

  committed
  398a799

• Update tokenizer error

  Rocketknight1

  committed
  e3140a6

• Update tokenizer passing, handle generation on GPU

  Rocketknight1

  committed
  0008722

• Slightly more explanation cleanup

  Rocketknight1

  committed
  a29c131

• More comment cleanup

  Rocketknight1

  committed
  9c359ff

• Factor out the token cleanup so it's more obvious what we're doing, and we can change it later

  Rocketknight1

  committed
  602222d

• Careful with that cleanup!

  Rocketknight1

  committed
  4c7a777

• Cleanup + optimizations to _get_matching_positions

  Rocketknight1

  committed
  b6e0163

• More minor performance tweaks

  Rocketknight1

  committed
  43d9e08

• Implement caching and eliminate some expensive ops (startup time: 200ms -> 9ms)

  Rocketknight1

  committed
  60eb576

• Remove the pin_memory call

  Rocketknight1

  committed
  ff42211

• Parallelize across all stop strings!

  Rocketknight1

  committed
  ae800a6

• Quick fix for tensor devices

  Rocketknight1

  committed
  46c0a9c

• Update embeddings test for the new format

  Rocketknight1

  committed
  b9a066d

• Fix test imports

  Rocketknight1

  committed
  692523c

• Manual patching for BERT-like tokenizers

  Rocketknight1

  committed
  2ba7f8e

• Return a bool vector instead of a single True/False

  Rocketknight1

  committed
  8b95ec1

• Better comment

  Rocketknight1

  committed
  1b46b20

• Better comment

  Rocketknight1

  committed
  350a850

• Add tests from @zucchini-nlp

  Rocketknight1

  committed
  0b85c6c

• Amy's list creation nit

  Rocketknight1

  committed
  e8c769d

• tok_list -> token_list

  Rocketknight1

  committed
  14de1c3

• Push a big expanded docstring (should we put it somewhere else?)

  Rocketknight1

  committed
  b8961e8

• Expand docstrings

  Rocketknight1

  committed
  7ed55ad

• Docstring fixups

  Rocketknight1

  committed
  cbb9d14

• Rebase

  Rocketknight1

  committed
  7db95c1

• make fixup

  Rocketknight1

  committed
  49b0f21

• Make a properly general method for figuring out token strings

  Rocketknight1

  committed
  c9aefe6

• Fix naming throughout the functions

  Rocketknight1

  committed
  443cd5d

• Move cache, refactor, fix tests

  Rocketknight1

  committed
  e1c9c0e

• Add comment

  Rocketknight1

  committed
  f49ec00

• Remove finished TODO

  Rocketknight1

  committed
  e90aaba

• Remove finished TODO

  Rocketknight1

  committed
  bb27d82

• make fixup

  Rocketknight1

  committed
  4317019

• Update src/transformers/generation/stopping_criteria.py

  Show description for 19df6a8
  

  Rocketknight1

  and

  amyeroberts

  committed
  19df6a8

• Update and shorten docstring

  Rocketknight1

  committed
  8b52039

• Update tests to be shorter/clearer and test specific cases

  Rocketknight1

  committed
  0aa201c