[benchmarks] Default to bfloat16 (inference) and AMP (training) precision.#6518
Merged
[benchmarks] Default to bfloat16 (inference) and AMP (training) precision.#6518
bfloat16 (inference) and AMP (training) precision.#6518Conversation
This comment was marked as outdated.
This comment was marked as outdated.
- Pick data type based on `test`. - Create `cast_to_dtype` function.
f83895a to
4864f13
Compare
This was referenced Feb 12, 2024
Collaborator
Author
|
I'm still running the benchmarks in order to grasp the regressions this PR introduces. |
frgossen
reviewed
Feb 15, 2024
| "hf_T5_generate", | ||
| } | ||
|
|
||
| FORCE_AMP_FOR_FP16_BF16_MODELS = { |
Collaborator
There was a problem hiding this comment.
Do you know how this is configured on PyTorch HUD? Maintaining a list like this feels prone to divergence.
Collaborator
Author
There was a problem hiding this comment.
These lists are taken from the scripts in the PyTorch main repo. They are used for generating the PyTorch HUD results.
Collaborator
Author
There was a problem hiding this comment.
I will leave a comment.
Collaborator
There was a problem hiding this comment.
Could you share the link to the scripts in the PyTorch main repo that are used to generate the HUD result?
Collaborator
Author
There was a problem hiding this comment.
frgossen
approved these changes
Feb 15, 2024
| "hf_T5_generate", | ||
| } | ||
|
|
||
| FORCE_AMP_FOR_FP16_BF16_MODELS = { |
golechwierowicz
approved these changes
Feb 16, 2024
Collaborator
golechwierowicz
left a comment
There was a problem hiding this comment.
Can you post BERT_pytorch kernel profile after the change?
Collaborator
Author
|
The profiling results (in the following posts) were generated with the following command: python xla/benchmarks/experiment_runner.py \
--suite-name torchbench --accelerator cuda --dump-pytorch-profiles \
--xla PJRT --dynamo openxla --test eval --repeat 8 --iterations-per-run 1 \
-k BERT_pytorch |
Collaborator
Author
BERT_pytorch (before) |
Collaborator
Author
BERT_pytorch (after) |
Collaborator
|
Thank you! Looks good to push forward. |
amithrm
pushed a commit
to amithrm/xla
that referenced
this pull request
Mar 1, 2024
bhavya01
pushed a commit
that referenced
this pull request
Apr 22, 2024
This file contains hidden or 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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 participants
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.
Fix: #6483
This PR makes
bfloat16the default data-type for inference, and AMP the default execution mode for training. This follows the execution found in the PyTorch HUD.cc @miladm