float8 delayed scaling: private API to fix user overriding buffers #1292
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
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/ao/1292
Note: Links to docs will display an error until the docs builds have been completed. ❗ 1 Active SEVsThere are 1 currently active SEVs. If your PR is affected, please view them below: ✅ No FailuresAs of commit faa1593 with merge base 56bf2e8 ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
facebook-github-bot
added
the
CLA Signed
Summary: Context: pytorch/torchtitan#654 If the user has delayed scaling and FSDP float8 all-gather on, there is a subtle bug that can happen if the user calls `model.to_empty(device="cuda")`: 1. to_empty recreates the buffers for tracking weight amax and scale 2. (1) leaves the buffers pointed to by Float8Linear.weight._amax_buffer, etc orphaned, because they don't participate in `to_empty` I couldn't think of an easy and clean way to auto-fix this since we can't expect `torch.nn.Module` to know that our logic has multiple references to the same buffer, so exposing a private API for now until we can think of something better. With the current fix, the user can then call `_maybe_fixup_delayed_scaling_buffers` manually to relink the buffers to the correct new versions. Test Plan: CI Reviewers: Subscribers: Tasks: Tags:
vkuzo
force-pushed
the
20241115_fix_to_empty
branch
from
9457814 to
faa1593
Compare
vkuzo
added
the
topic: not user facing
weifengpy
reviewed
Nov 21, 2024
| assert m_fp8[0].fp8_scale_weight is m_fp8[0].weight._scale_buffer | ||
|
|
||
| m_fp8.to_empty(device="cuda") | ||
| m_fp8[0]._maybe_fixup_delayed_scaling_buffers() |
There was a problem hiding this comment.
we would need to call this inside torchtitan’s training loop?
There was a problem hiding this comment.
yeah, which is definitely not ideal
yanbing-j
pushed a commit
to yanbing-j/ao
that referenced
this pull request
Dec 9, 2024
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.
Summary:
Context: pytorch/torchtitan#654
If the user has delayed scaling and FSDP float8 all-gather on, there is a subtle bug that can happen if the user calls
model.to_empty(device="cuda"):to_emptyI couldn't think of an easy and clean way to auto-fix this since we can't expect
torch.nn.Moduleto know that our logic has multiple references to the same buffer, so exposing a private API for now until we can think of something better.With the current fix, the user can then call
_maybe_fixup_delayed_scaling_buffersmanually to relink the buffers to the correct new versions.Test Plan: CI
Reviewers:
Subscribers:
Tasks:
Tags: