feat: A100 submits for imaging/pixelization Nautilus + NSS control-cell#42
Merged
Merged
Conversation
Pair with the existing delaunay submits. RectangularAdaptImage has essentially identical memory budget to Delaunay per the vram probe (931 vs 922 MB/replica on HST, both cap at batch=16) but its mesh interpolator does NOT use jax.pure_callback. PyAutoArray's delaunay interpolator (autoarray/inversion/mesh/interpolator/delaunay.py:80,249) calls scipy.spatial.Delaunay via pure_callback; the rectangular path is pure JAX. The Nautilus pixelization cell expected to behave roughly like Nautilus Delaunay (84 ms/eval). The NSS pixelization cell is the decisive test: if it fits A100 where NSS Delaunay (322592/96/600/602) OOM'd, the failure mode for NSS Delaunay is the pure_callback × vmap pytree blow-up — not generic inversion-matrix memory — and the PyAutoPrompt autofit/nss_chunked_vmap_for_inversion_heavy_likelihoods.md should re-focus on the pure_callback path. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Pair with the existing delaunay submits. RectangularAdaptImage has essentially identical memory budget to Delaunay (`vram/config.py`: 931 vs 922 MB/replica HST, both cap at batch=16) but its mesh interpolator is pure JAX — PyAutoArray's delaunay interpolator (`autoarray/inversion/mesh/interpolator/delaunay.py:80,249`) calls scipy.spatial.Delaunay via `jax.pure_callback`, the rectangular path doesn't.
This is the decisive test for whether NSS Delaunay's OOM is the `pure_callback` × vmap pytree blow-up (per the project memory note `feedback_jax_pure_callback_const_fold.md`) or generic inversion-matrix memory pressure. If pixelization NSS fits and Delaunay NSS doesn't (which it doesn't — 322592/96/600/602 all OOM'd), the failure mode is the callback.
The Nautilus pixelization cell expected to behave roughly like Nautilus Delaunay (84 ms/eval) since the inversion is structurally similar.
Test plan
🤖 Generated with Claude Code