pytorch / pytorch/FBGEMM

Do you plan to support heterogeneous schedules for FusedEmbeddingBag operations?

Open
#4,152 0 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

Dominant language
C++
Stars
1.6k
Forks
787
PR merge metrics
No merged PRs in 30d

Description

The RecFlex paper points out that embedding tables within a fused embedding bag collection can be heterogeneous, such as having different embedding dimensions or access patterns (e.g., one-hot vs. multi-hot). Applying the same code schedule to all tables in the fused kernel could lead to sub-optimal performance.

I’m wondering if there is any plan to support generating and compiling kernels at runtime, so that different tables can use different code schedules, for both inference and training?

Contributor guide

Open the contributing guide

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

No source files, tests, or entry points are identified in the issue. Start by reading the linked RecFlex paper and locating the fused embedding bag implementation and its inference and training kernel-generation paths; done means heterogeneous schedules can be selected and compiled for different tables in both modes.

Written by the indexing model from the issue text.

Assessment

Tech stack
cpp
Domain
machine-learning, performance
Issue type
Feature
Difficulty
5/5
Estimated time
Over a week
Activity status
Stale
Clarity
Mostly clear
Newbie friendliness
25/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.