Skip to content

Add GR00T + PLADIS single-pass sparse cross-attention guidance (Phase 1)#6

Open
KwanjoonPark wants to merge 1 commit into
DAVIAN-Robotics:mainfrom
KwanjoonPark:feature/pladis-guidance
Open

Add GR00T + PLADIS single-pass sparse cross-attention guidance (Phase 1)#6
KwanjoonPark wants to merge 1 commit into
DAVIAN-Robotics:mainfrom
KwanjoonPark:feature/pladis-guidance

Conversation

@KwanjoonPark

Copy link
Copy Markdown

Implements PLADIS (arXiv:2503.07677) as a new test-time guidance for GR00T-N1, mirroring the ACG guidance plumbing.

  • pladis.py: PLADISAttnProcessor extrapolates between dense (softmax) and sparse (entmax1.5/sparsemax) cross-attention: attn = dense + pladis_scale*(sparse - dense). Single forward pass per denoising step (no incoherent second pass), installed on cross-attention (even-idx) DiT blocks only.
  • guidance/init.py: register name=="pladis" dispatch.
  • gr00t_config.py: add guidance.pladis_scale (1.5), .method (ent15max), .pladis_layers.
  • PLADIS_INTEGRATION_TASKS.md: phased plan (GR00T+PLADIS, GR00T+ACG+PLADIS).
  • CLAUDE.md: repo guidance for Claude Code.

Verified: pladis_scale=0 reduces exactly to standard softmax attention; scale=1.5 differs and is finite; bf16 path is NaN-free.

Implements PLADIS (arXiv:2503.07677) as a new test-time guidance for GR00T-N1,
mirroring the ACG guidance plumbing.

- pladis.py: PLADISAttnProcessor extrapolates between dense (softmax) and
  sparse (entmax1.5/sparsemax) cross-attention:
  attn = dense + pladis_scale*(sparse - dense). Single forward pass per
  denoising step (no incoherent second pass), installed on cross-attention
  (even-idx) DiT blocks only.
- guidance/__init__.py: register name=="pladis" dispatch.
- gr00t_config.py: add guidance.pladis_scale (1.5), .method (ent15max), .pladis_layers.
- PLADIS_INTEGRATION_TASKS.md: phased plan (GR00T+PLADIS, GR00T+ACG+PLADIS).
- CLAUDE.md: repo guidance for Claude Code.

Verified: pladis_scale=0 reduces exactly to standard softmax attention;
scale=1.5 differs and is finite; bf16 path is NaN-free.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
@KwanjoonPark

Copy link
Copy Markdown
Author

Cross Attention에 pladis 적용

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

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant