Lightning-AI / Lightning-AI/lightning-thunder

Transformer Engine's recipes UX

Open
#2,080 1 comment 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

enhancement TransformerEngine ux
Dominant language
Python
Stars
1.5k
Forks
121
PR merge metrics
No merged PRs in 30d

Description

🚀 Feature

Branching the conversation from #1908(comment) there are multiple possibilities for the user to specify a recipe for TE computation.

However on the high level, there are two approaches:

  1. Thunder takes care of deciding which TE recipe is best for your and uses that throughout the program and advanced users that want to customize the choice are given the opportunity to do so by adding an option to the instantiation of the executor, or
  2. Transformer Engine fp8_autocast context manager is fully integrated with Thunder and the user has to create and decide which recipe and when to use.

In either case I believe that there should be complete commitment to the choice, for example: committing to point 2 means matching TE behavior, and allowing high precision computation outside the context manager.

cc. @kshitij12345 @IvanYashchuk

cc @borda

Contributor guide

No contributing guide indexed for this repository

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

The issue names no files, tests, or entry points. Start by reviewing the Transformer Engine fp8_autocast behavior and the two proposed recipe approaches, then establish the chosen user-facing contract. Done means the selected approach is implemented consistently and its behavior is covered by the relevant tests, which are not identified here.

Written by the indexing model from the issue text.

Assessment

Tech stack
python
Domain
machine-learning
Issue type
Feature
Difficulty
5/5
Estimated time
Over a week
Activity status
Stale
Clarity
Needs clarification
Newbie friendliness
20/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.