NVIDIA / NVIDIA/physicsnemo

Move HealDA-v2 Triton kernels from experimental to nn.functional

Open
#1,824 2 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

stale
Dominant language
Python
Stars
3.3k
Forks
787
Avg merge
2d 21h
Merged PRs (30d)
27

Description

Tracked from review comment in #1758.

The fused Triton kernels currently registered under the "healda::" torch.library namespace (kernels/obs_tokenizer_film.py, kernels/pixel_attention.py) should move to physicsnemo/nn/functional/ once VideoHealDA graduates out of experimental. This would give them a stable namespace (e.g. "physicsnemo::pixel_attn_fwd") and make them reusable by models outside the healda-v2 subpackage.

Scope when addressing:

  • Move custom_op registrations to physicsnemo/nn/functional/
  • Update the "healda::" namespace to "physicsnemo::"
  • Add op schema docstrings and dtype/device constraints
  • Update PixelCrossAttention and ObsTokenizerFiLM to import from the new location

Prerequisite: VideoHealDA stable through one release cycle after #1758 merges.

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

First confirm that VideoHealDA has been stable through one release cycle after #1758. Then read kernels/obs_tokenizer_film.py, kernels/pixel_attention.py, and the PixelCrossAttention and ObsTokenizerFiLM call sites; done means the registrations live under physicsnemo/nn/functional/, use the physicsnemo namespace, document schemas and constraints, and imports work from the new location.

Written by the indexing model from the issue text.

Assessment

Tech stack
python, pytorch
Domain
machine-learning
Issue type
Refactor
Difficulty
4/5
Estimated time
3-5 days
Activity status
Active
Clarity
Mostly clear
Newbie friendliness
55/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.