Code

[Whisper] 🚨 Fix whisper decoding 🚨#34135

Code

Merged

eustlb merged 75 commits intohuggingface:mainhuggingface/transformers:mainfrom eustlb:fix-whisper-decodingeustlb/transformers:fix-whisper-decodingCopy head branch name to clipboardDec 18, 2024

+268-121

Commits

Commits on Oct 13, 2024

• do not remove decoder_input_ids for the first segment

  eustlb

  committed
  4c7be8c

  Copy full SHA for 4c7be8c

• do not remove eos token in generate_with_fallback

  eustlb

  committed
  805c688

  Copy full SHA for 805c688

• when removing padding tokens, do not remove eos token

  eustlb

  committed
  271014f

  Copy full SHA for 271014f

• remove eos token in generate (and not in generate_with_fallback!)

  eustlb

  committed
  d9a91b3

  Copy full SHA for d9a91b3

• reconciliate short-from/ long-form behavior

  eustlb

  committed
  b9d0ec1

  Copy full SHA for b9d0ec1

• correct avg_logprobs calculation

  eustlb

  committed
  f9d5fdd

  Copy full SHA for f9d5fdd

• handle eos token in segments

  eustlb

  committed
  f111aa3

  Copy full SHA for f111aa3

• handle decoder_input_ids and eos token in _prepare_decoder_input_ids

  eustlb

  committed
  34d690a

  Copy full SHA for 34d690a

• fix incorrect time precision

  eustlb

  committed
  16f768b

  Copy full SHA for 16f768b

Commits on Oct 18, 2024

• always remove eos token

  eustlb

  committed
  cf18e0a

  Copy full SHA for cf18e0a

• always remove decoder_input_ids

  eustlb

  committed
  ec7cd58

  Copy full SHA for ec7cd58

• no need to handle decoder_inputs_ids and eos token

  eustlb

  committed
  1530930

  Copy full SHA for 1530930

• no need to remove decoder_input_ids

  eustlb

  committed
  67865dd

  Copy full SHA for 67865dd

• no need to handle eos token

  eustlb

  committed
  eb107d9

  Copy full SHA for eb107d9

Commits on Oct 24, 2024

• fix num_beams in _retrieve_logit_processors

  eustlb

  committed
  7881928

  Copy full SHA for 7881928

Commits on Nov 4, 2024

• remove todo unconsistency

  eustlb

  committed
  031ace6

  Copy full SHA for 031ace6

• no need to add eos token

  eustlb

  committed
  eaaeec6

  Copy full SHA for eaaeec6

Commits on Nov 19, 2024

• last_timestamp_pos should indeed be timestamp token pos

  eustlb

  committed
  cdd5144

  Copy full SHA for cdd5144

Commits on Nov 20, 2024

• patch generate to enable compatibility with GenerationTesterMixin tests

  eustlb

  committed
  544f21b

  Copy full SHA for 544f21b

• adapt test_generate_continue_from_past_key_values

  eustlb

  committed
  70ffdb3

  Copy full SHA for 70ffdb3

• adapt test_prompt_lookup_decoding_matches_greedy_search

  eustlb

  committed
  79347e9

  Copy full SHA for 79347e9

Commits on Nov 22, 2024

• Merge branch 'main' into fix-whisper-decoding

  eustlb

  authored
  c01ef12

  Copy full SHA for c01ef12

• Merge branch 'main' into fix-whisper-decoding

  eustlb

  authored
  d03f87c

  Copy full SHA for d03f87c

Commits on Nov 27, 2024

• adapt generic GenerationMixin tests to whisper's generate

  eustlb

  committed
  1387919

  Copy full SHA for 1387919

• fix speculative decoding

  eustlb

  committed
  ad262c6

  Copy full SHA for ad262c6

• fix

  eustlb

  committed
  a4c2b44

  Copy full SHA for a4c2b44

• Merge branch 'main' into fix-whisper-decoding

  eustlb

  authored
  17877f3

  Copy full SHA for 17877f3

• Merge branch 'main' into fix-whisper-decoding

  eustlb

  authored
  7b2046a

  Copy full SHA for 7b2046a

Commits on Dec 5, 2024

• Merge branch 'main' into fix-whisper-decoding

  eustlb

  authored
  c9a209f

  Copy full SHA for c9a209f

• [run-slow] whisper

  

  Eustache Le Bihan

  committed
  34d5257

  Copy full SHA for 34d5257

• change HF_HUB_TOKEN for require_read_token

  eustlb

  committed
  c760d09

  Copy full SHA for c760d09

• [run-slow] whisper

  eustlb

  committed
  e7c4b85

  Copy full SHA for e7c4b85

• Merge branch 'main' into fix-whisper-decoding

  eustlb

  authored
  40bcdcc

  Copy full SHA for 40bcdcc

• prioritize kwargs over generation_config

  eustlb

  committed
  783c2f1

  Copy full SHA for 783c2f1

• remove unnecessary args

  eustlb

  committed
  88d2d38

  Copy full SHA for 88d2d38

• [run-slow] whisper

  eustlb

  committed
  3a8f739

  Copy full SHA for 3a8f739

• update tests

  eustlb

  committed
  93cf1b7

  Copy full SHA for 93cf1b7

• Merge branch 'main' into fix-whisper-decoding

  eustlb

  authored
  2afb05c

  Copy full SHA for 2afb05c

• [run-slow] whisper

  eustlb

  committed
  ecbe56b

  Copy full SHA for ecbe56b

Commits on Dec 6, 2024

• add comment

  eustlb

  committed
  c9b36c3

  Copy full SHA for c9b36c3

• update test

  eustlb

  committed
  14c280a

  Copy full SHA for 14c280a

• Merge branch 'main' into fix-whisper-decoding

  eustlb

  authored
  78ed67e

  Copy full SHA for 78ed67e

• [run-slow] whisper

  eustlb

  committed
  34adc99

  Copy full SHA for 34adc99

• update test + revert require_read_token

  eustlb

  committed
  58f0526

  Copy full SHA for 58f0526

• docstring updates

  eustlb

  committed
  367006d

  Copy full SHA for 367006d

Commits on Dec 8, 2024

• revert tokenizer decode args change

  eustlb

  committed
  2f19263

  Copy full SHA for 2f19263

• Merge branch 'main' into fix-whisper-decoding

  eustlb

  authored
  9c6fe55

  Copy full SHA for 9c6fe55

Commits on Dec 10, 2024

• do not use a patch + docstring updates

  eustlb

  committed
  3ebd12f

  Copy full SHA for 3ebd12f

• Merge branch 'main' into fix-whisper-decoding

  eustlb

  authored
  b9ba08a

  Copy full SHA for b9ba08a

• [run-slow] whisper

  eustlb

  committed
  24d9399

  Copy full SHA for 24d9399

• make

  eustlb

  committed
  9fe3011

  Copy full SHA for 9fe3011

• [run-slow] whisper

  eustlb

  committed
  3aae6b8

  Copy full SHA for 3aae6b8

Commits on Dec 11, 2024

• add a flag to force unique call to generate

  eustlb

  committed
  0ea240f

  Copy full SHA for 0ea240f

• test update

  eustlb

  committed
  0960a52

  Copy full SHA for 0960a52

• [run-slow] whisper

  eustlb

  committed
  0870ac7

  Copy full SHA for 0870ac7

• add force_unique_generate_call arg

  eustlb

  committed
  00d37e8

  Copy full SHA for 00d37e8

• do not use a patch

  eustlb

  committed
  cf75ea3

  Copy full SHA for cf75ea3

• correct the timestamps for the pad tokens

  eustlb

  committed
  2cb638d

  Copy full SHA for 2cb638d

• docstring update

  eustlb

  committed
  66290cb

  Copy full SHA for 66290cb

• docstring update

  eustlb

  committed
  b41e2c1

  Copy full SHA for b41e2c1

• docstring update

  eustlb

  committed
  221bae1

  Copy full SHA for 221bae1

• Merge branch 'main' into fix-whisper-decoding

  eustlb

  authored
  98dc7f1

  Copy full SHA for 98dc7f1

• upodate TF tests

  eustlb

  committed
  0b6687c

  Copy full SHA for 0b6687c

• add require_read_token

  eustlb

  committed
  ab910f7

  Copy full SHA for ab910f7

• [run-slow] whisper

  eustlb

  committed
  692bf14

  Copy full SHA for 692bf14

Commits on Dec 12, 2024

• test reset dynamo

  eustlb

  committed
  a15aa4a

  Copy full SHA for a15aa4a

• Merge branch 'main' into fix-whisper-decoding

  eustlb

  authored
  0011cec

  Copy full SHA for 0011cec

• [run-slow] whisper

  ydshieh

  committed
  41df6ca

  Copy full SHA for 41df6ca

• fix

  ydshieh

  committed
  aaecea4

  Copy full SHA for aaecea4

• [run-slow] whisper

  ydshieh

  committed
  58d0b90

  Copy full SHA for 58d0b90

Commits on Dec 18, 2024

• avoid iterating twice on current_segments

  eustlb

  committed
  a1f4e43

  Copy full SHA for a1f4e43

• [run-slow] whisper

  eustlb

  committed
  671b079

  Copy full SHA for 671b079

• Merge branch 'main' into fix-whisper-decoding

  eustlb

  authored
  3d21ed8

  Copy full SHA for 3d21ed8

• [run-slow] whisper

  eustlb

  committed
  dc6fbd1

  Copy full SHA for dc6fbd1

• Merge branch 'main' into fix-whisper-decoding

  eustlb

  authored
  8b7a2e8

  Copy full SHA for 8b7a2e8