[NLP] Access scaler only in FP16 case #7916
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
ericharper
reviewed
Nov 20, 2023
Comment on lines
777
to
781
| if self.torch_dtype == torch.float16 and not self.trainer.precision.startswith("16"): | ||
| # Make sure that trainer.precision_plugin.scaler exists for config_mapping below | ||
| raise ValueError( | ||
| "Creating a half-precision model requires setting trainer.precision to '16' or '16-mixed'" | ||
| ) |
There was a problem hiding this comment.
Should we do the same check for bf16?
There was a problem hiding this comment.
I think it is needed only for model.precision=16 which is self.torch_dtype here, see this. More specifically, the check is to make sure this will work -- loading FP16 model assumes that a scaler is available in trainer so it needs to be configured properly.
In my understanding in other cases it is the trainer precision to handle training precision if model.precision is different, and this would result in necessary type casting. But I might have not considered all the possibilities.
Which case exactly you mean for model.precision and trainer.precision, respectively?
janekl
force-pushed
the
jlasek/loading_model_precision_16
branch
from
e520963 to
6d1a5c0
Compare
janekl
changed the title
janekl
commented
Dec 6, 2023
| @@ -103,7 +103,7 @@ def __init__(self, cfg: DictConfig, trainer: Trainer, no_lm_init=True): | |||
| self.tokenizer = None | |||
|
|
|||
| with open_dict(cfg): | |||
| if cfg.get('precision', None) is None and trainer is not None: | |||
There was a problem hiding this comment.
The latter condition is guaranteed by the check in L93
Signed-off-by: Jan Lasek <janek.lasek@gmail.com>
…ecision != 16 Signed-off-by: Jan Lasek <janek.lasek@gmail.com>
for more information, see https://pre-commit.ci
Signed-off-by: Jan Lasek <janek.lasek@gmail.com>
Signed-off-by: Jan Lasek <janek.lasek@gmail.com>
janekl
force-pushed
the
jlasek/loading_model_precision_16
branch
from
eaf20c1 to
b08c123
Compare
pzelasko
pushed a commit
to pzelasko/NeMo
that referenced
this pull request
Jan 3, 2024
* Remove unused 'precision' variable Signed-off-by: Jan Lasek <janek.lasek@gmail.com> * Raise informative error when trying to load FP16 model for trainer.precision != 16 Signed-off-by: Jan Lasek <janek.lasek@gmail.com> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * Make sure scaler is available instead of raising error Signed-off-by: Jan Lasek <janek.lasek@gmail.com> * trainer != None is assured thanks to a previous check Signed-off-by: Jan Lasek <janek.lasek@gmail.com> --------- Signed-off-by: Jan Lasek <janek.lasek@gmail.com> Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> Signed-off-by: Piotr Żelasko <petezor@gmail.com>
ssh-meister
pushed a commit
to ssh-meister/NeMo
that referenced
this pull request
Feb 15, 2024
* Remove unused 'precision' variable Signed-off-by: Jan Lasek <janek.lasek@gmail.com> * Raise informative error when trying to load FP16 model for trainer.precision != 16 Signed-off-by: Jan Lasek <janek.lasek@gmail.com> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * Make sure scaler is available instead of raising error Signed-off-by: Jan Lasek <janek.lasek@gmail.com> * trainer != None is assured thanks to a previous check Signed-off-by: Jan Lasek <janek.lasek@gmail.com> --------- Signed-off-by: Jan Lasek <janek.lasek@gmail.com> Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> Signed-off-by: Sasha Meister <ameister@nvidia.com>
rohitrango
pushed a commit
to rohitrango/NeMo
that referenced
this pull request
Jun 25, 2024
* Remove unused 'precision' variable Signed-off-by: Jan Lasek <janek.lasek@gmail.com> * Raise informative error when trying to load FP16 model for trainer.precision != 16 Signed-off-by: Jan Lasek <janek.lasek@gmail.com> * [pre-commit.ci] auto fixes from pre-commit.com hooks for more information, see https://pre-commit.ci * Make sure scaler is available instead of raising error Signed-off-by: Jan Lasek <janek.lasek@gmail.com> * trainer != None is assured thanks to a previous check Signed-off-by: Jan Lasek <janek.lasek@gmail.com> --------- Signed-off-by: Jan Lasek <janek.lasek@gmail.com> Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
What does this PR do ?
When trying to load FP16 model with
trainer.precision != 16(or16-mixed) a user gets a mysterious error message like:I'm changing the code to access
trainer.precision_plugin.scaler.scaleonly whentraineris configured with the relevant precision: "16" or "16-mixed".Collection: NLP
Before your PR is "Ready for review"
Pre checks:
PR Type:
If you haven't finished some of the above items you can still open "Draft" PR.
Who can review?
Anyone in the community is free to review the PR once the checks have passed.
Contributor guidelines contains specific people who can review PRs to various areas.