Add patch to locomanipulation SDG pipeline to avoid use of flash attn

[jaybdub]

Make locomanipulation SDG GR00T flow runnable without flash-attn

Summary

Two small fixes that let users finetune and roll out the locomanipulation SDG
GR00T policy on hardware where flash-attn is unavailable (e.g. Blackwell, or
any environment where the wheel fails to build).

Changes

• scripts/imitation_learning/locomanipulation_sdg/gr00t/no_flash_attn.patch
  (new): patch against the Isaac-GR00T repo that switches the bundled Eagle 2.5
  VL model from flash_attention_2 to PyTorch SDPA, and guards the RADIO
  vision module's flash_attn imports so the package becomes importable
  without flash-attn installed. SigLIP path works; RADIO path is unsupported
  without flash-attn (documented in the patch).

• docs/source/overview/imitation-learning/humanoids_imitation.rst: adds
  a note in the GR00T install section explaining when to apply the patch
  (build failure, or RuntimeError: FlashAttention only supports Ampere GPUs or newer) and how to apply it from the sibling Isaac-GR00T checkout.

• scripts/imitation_learning/locomanipulation_sdg/gr00t/rollout_policy.py:
  override env_cfg.recorders with ActionStateRecorderManagerCfg() so the
  rollout doesn't try to record env._locomanipulation_sdg_output_data, which
  is only populated by the data-generation state machine in generate_data.py
  and is absent during policy rollout. Without this, the recorder raises
  AttributeError on the first pre-step.

Type of change

• Bug fix (non-breaking change which fixes an issue)

Screenshots

Please attach before and after screenshots of the change if applicable.

Checklist

• I have read and understood the contribution guidelines
• I have run the pre-commit checks with ./isaaclab.sh --format
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• I have updated the changelog and the corresponding version in the extension's config/extension.toml file
• I have added my name to the CONTRIBUTORS.md or my name already exists there

[isaaclab-review-bot]

Isaac Lab Review Bot 🤖

Re-review after commits 511bc8d → 15c318f

Summary: This PR adds an optional patch to bypass flash-attn dependency in the locomanipulation SDG pipeline for GR00T. Good fix for compatibility with newer GPU architectures (like Blackwell) where flash-attn prebuilt kernels don't exist yet.

📝 Changes in this push

The new commit adds a small but important fix to rollout_policy.py:

• Line 379-381: Resets env_cfg.recorders to a fresh ActionStateRecorderManagerCfg() before setting the export path. This fixes a bug where the SDG output-data recorder term would try to pull env._locomanipulation_sdg_output_data which is only populated during data generation, not during policy rollout. Good defensive fix.

✅ What looks good

1. Well-scoped change - Only 3 files modified, focused on the specific issue
2. Optional approach - The patch is opt-in, not forced on all users
3. Good documentation - Clear explanation in the RST file about when/why to use the patch
4. Graceful fallback - The patch handles missing flash-attn imports gracefully with try/except blocks
5. Uses SDPA - PyTorch's native Scaled Dot Product Attention is a reasonable fallback that works across CUDA architectures
6. Rollout fix - The new change properly isolates rollout from SDG-specific recorder state

📋 Minor suggestions (non-blocking)

1. Checklist items unchecked - The PR checklist shows all items unchecked. Consider marking the applicable ones before merge.

⚠️ One consideration (unchanged from previous review)

The patch sets flash_attn_unpadded_qkvpacked_func = None and pad_input = unpad_input = None when imports fail, but FlashAttention class still exists and may be instantiated. If RADIO vision encoder is used (rather than SigLIP), this could hit a runtime error when the None functions are called. The comment notes "RADIO vision won't work, SigLIP does" — consider adding a runtime check that raises a clear error message if RADIO is selected without flash-attn installed.

Verdict: ✅ LGTM — straightforward compatibility fix with a good rollout bugfix in the latest commit.

---

Update (e3868cf): Branch rebased onto develop. The 3 PR files are unchanged — no new review items.

[peterd-NV]

Is this git patch sensitive to the particular commit of GR00T? I.e. if someone changes to a different commit and the code changes slightly this will result in a git apply failure. Maybe worth mentioning in the docs the particular commit that this patch works against.

[jaybdub]

The gr00t checkout is already pinned to a specific version in the documentation that matches the patch

[peterd-NV]

lgtm, left a comment regarding documenting the commit the git patch works against.