Repository navigation
fix: construct BF16 checkpoints in FP16 on pre-Ampere GPUs - #200
Open
Matt Van Horn (mvanhorn) wants to merge 1 commit into
Open
Matt Van Horn (mvanhorn) wants to merge 1 commit into
Matt Van Horn (mvanhorn) wants to merge 1 commit into
Conversation
For checkpoints configured as bf16 or bfloat16, from_pretrained now substitutes float16 before model construction when CUDA is available and the current device has compute capability below 8.0. The fallback emits a warning about FP16 overflow and numerical instability. The README explains the fallback, its dependence on the current CUDA device on mixed-GPU systems, and the separate backend failures older GPUs may still encounter. Fixes microsoft#23
This branch has not been deployed
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
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
For checkpoints configured as bf16 or bfloat16, from_pretrained now substitutes float16 before model construction when CUDA is available and the current device has compute capability below 8.0. The fallback emits a warning about FP16 overflow and numerical instability. The README explains the fallback, its dependence on the current CUDA device on mixed-GPU systems, and the separate backend failures older GPUs may still encounter. Loading a BF16 checkpoint on a pre-Ampere GPU can fail during attention because the available backend cannot process BF16 inputs. from_pretrained passed the checkpoint's configured dtype directly to the model constructor without checking the current CUDA device's compute capability.
Testing: The test check found no tests for the touched modules, so no runtime behavior was verified.
Fixes #23