Skip to content

Add flash_attn_3 as first option for FA3 import - #2010

Merged
santhnm2 merged 7 commits into
NVIDIA:mainfrom
santhnm2:flash_attn_3
Nov 5, 2025
Merged

Add flash_attn_3 as first option for FA3 import#2010
santhnm2 merged 7 commits into
NVIDIA:mainfrom
santhnm2:flash_attn_3

Conversation

@santhnm2

Copy link
Copy Markdown
Contributor

What does this PR do ?

Makes flash_attn_3 the first option for importing FlashAttention 3 as this is the default location.

Contribution process

flowchart LR
    A[Pre-checks] --> B[PR Tests]
    subgraph Code Review/Approval
        C1[Expert Review] --> C2[Final Review]
    end
    B --> C1
    C2 --> D[Merge]
Loading

Pre-checks

  • I want this PR in a versioned release and have added the appropriate Milestone (e.g., Core 0.8)
  • I have added relevant unit tests
  • I have added relevant functional tests
  • I have added proper typing to my code Typing guidelines
  • I have added relevant documentation
  • I have run the autoformatter.sh on my PR

Code review

The following process is enforced via the CODEOWNERS file for changes into megatron/core. For changes outside of megatron/core, it is up to the PR author whether or not to tag the Final Reviewer team.

For MRs into `main` branch

(Step 1): Add PR label Expert Review

(Step 2): Collect the expert reviewers reviews

  1. Attach the Expert Review label when your PR is ready for review.
  2. GitHub auto-assigns expert reviewers based on your changes. They will get notified and pick up your PR soon.

⚠️ Only proceed to the next step once all reviewers have approved, merge-conflict are resolved and the CI is passing.
Final Review might get declined if these requirements are not fulfilled.

(Step 3): Final Review

  1. Add Final Review label
  2. GitHub auto-assigns final reviewers based on your changes. They will get notified and pick up your PR soon.

(Optional Step 4): Cherry-pick into release branch

If this PR also needs to be merged into core_r* release branches, after this PR has been merged, select Cherry-pick to open a new PR into the release branch.

For MRs into `dev` branch The proposed review process for `dev` branch is under active discussion.

MRs are mergable after one approval by either eharper@nvidia.com or zijiey@nvidia.com.

Merging your PR

Any member of core-adlr and core-nemo will be able to merge your PR.

@santhnm2 santhnm2 self-assigned this Oct 28, 2025
@santhnm2
santhnm2 requested review from a team as code owners October 28, 2025 20:45
@copy-pr-bot

copy-pr-bot Bot commented Oct 28, 2025

Copy link
Copy Markdown

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

@Skylion007

Copy link
Copy Markdown
Contributor

I was shocked that FA3 is not in the NGC pytorch containers. Is it installed under this alternate names?

try:
from flashattn_hopper.flash_attn_interface import _flash_attn_forward
from flashattn_hopper.flash_attn_interface import (
from flash_attn_3.flash_attn_interface import _flash_attn_forward

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Can we get rid of the nested try/except clause loops?

It would look a bit nicer if we have smth like:

HAVE_FA3=False
try:
  # import old way
  HAVE_FA3=True
except:
  pass # ideally log
try:
  # import new way
  HAVE_FA3=True
except:
  pass # ideally log

Or maybe add a branch depending on the fa version or smth. TransformerEngine has FlashAttentionUtils that can give you the FA version, we could use that to unify the behaviour between the training and inference side.

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The second try block would have to be wrapped in a if not HAVE_FA3 clause though, so it seems like it would be equally messy? Open to either option though

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Agreed! Still looks more linear/simpler to me.

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Updated to not nest the imports

@ko3n1g

ko3n1g commented Oct 30, 2025

Copy link
Copy Markdown
Contributor

/ok to test a4410fc

@tdene

tdene commented Nov 5, 2025

Copy link
Copy Markdown
Contributor

/ok to test 3eae66f

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

7 participants