transformers
Reducing memory usage: removing useless logits computation in generate()
#31292
Merged

Reducing memory usage: removing useless logits computation in generate() #31292

gante merged 35 commits into huggingface:main from Cyrilvallez:logits-dtype
Cyrilvallez
amyeroberts
Cyrilvallez Cyrilvallez force pushed 1 year ago
ArthurZucker
ArthurZucker commented on 2024-06-18
HuggingFaceDocBuilderDev
gante
gante commented on 2024-06-18
gante
gante
gante approved these changes on 2024-06-20
Cyrilvallez Cyrilvallez force pushed 1 year ago
Cyrilvallez
ArthurZucker
ArthurZucker commented on 2024-07-12
ArthurZucker
Cyrilvallez
Cyrilvallez Cyrilvallez force pushed 1 year ago
Cyrilvallez
Cyrilvallez Cyrilvallez force pushed 1 year ago
Cyrilvallez Cyrilvallez force pushed 1 year ago
Cyrilvallez Cyrilvallez force pushed 1 year ago
Cyrilvallez
ringohoffman
ArthurZucker
ArthurZucker ArthurZucker requested a review from ArthurZucker ArthurZucker 1 year ago
Oxi84
ArthurZucker
ArthurZucker approved these changes on 2024-08-05
ringohoffman
Cyrilvallez
ghost
Cyrilvallez
Cyrilvallez Cyrilvallez force pushed 1 year ago
Cyrilvallez
ArthurZucker
Cyrilvallez Cyrilvallez force pushed 1 year ago
Cyrilvallez
gante
Cyrilvallez
ArthurZucker
ArthurZucker approved these changes on 2024-08-22
gante
Cyrilvallez Add .float() in all generation methods logit outputs
1748ff17
Cyrilvallez Switch float-casting of logits to training only for main models
3f4f4e8d
Cyrilvallez Add `num_logits_to_keep` in Llama and add it by default in generate
727c7e46
Cyrilvallez Apply style
222017d4
Cyrilvallez Add num_logits_to_keep as arg in prepare_input_for_generation
dc709c68
Cyrilvallez Add support for Mistral
d2f1566f
Cyrilvallez Revert models except llama and mistral
f2ef90cd
Cyrilvallez Fix default None value in _supports_num_logits_to_keep()
ce7b980c
Cyrilvallez Fix dimension of dummy input
d4201f42
Cyrilvallez Add exception for prophetnet in _supports_num_logits_to_keep()
b15b5dec
Cyrilvallez Update _supports_num_logits_to_keep() to use inspect.signature()
95e0807a
Cyrilvallez Add deprecation cycle + remove modification with pretraining_tp
12db0457
Cyrilvallez Apply style
b224e24c
Cyrilvallez Add most used models
f0e1034b
Cyrilvallez Apply style
9ac57db6
Cyrilvallez Make `num_logits_to_keep` an int in all cases to remove if-else clause
f7421b69
Cyrilvallez Add compile check for the warning
c8f91776
Cyrilvallez Fix torch versions
5e1589e1
Cyrilvallez style
7998b650
Cyrilvallez Add gemma2
8fa80181
Cyrilvallez Update warning version
b49fe767
Cyrilvallez Add comment about .float operations in generation utils
cf9378a4
Cyrilvallez Add tests in GenerationTesterMixin and ModelTesterMixin
66e3e9d8
Cyrilvallez Fix batch size for assisted decoding in tests
e4c5a71b
Cyrilvallez fix small issues in test
b68ee166
Cyrilvallez refacor test
e8374252
Cyrilvallez fix slicing removing dim issue
26863ca3
Cyrilvallez Add nemotron support (should fix check-copy issue in CIs)
3c3eeaa9
Cyrilvallez Trigger new CIs
c4008655
Cyrilvallez Trigger new CIs
802eca83
Cyrilvallez Bump version
4d6fae65
Cyrilvallez Bump version in TODO
f12f172f
Cyrilvallez Trigger CIs
7b1a26cc
Cyrilvallez Cyrilvallez force pushed to 7b1a26cc 1 year ago
Cyrilvallez remove blank space
b11b048f
Cyrilvallez Trigger CIs
f03adfb1
Cyrilvallez
gante gante merged 22e6f145 into main 1 year ago
ringohoffman
gante

Login to write a write a comment.

Login via GitHub

Reviewers
Assignees
No one assigned
Labels
Milestone