Subtensor lift: reason about advanced indices jointly when gating#2251
Merged
Conversation
ricardoV94
force-pushed
the
proper_idx_gating
branch
from
June 21, 2026 22:22
30aa289 to
c27a1c3
Compare
ricardoV94
force-pushed
the
proper_idx_gating
branch
3 times, most recently
from
June 25, 2026 12:38
483c1b9 to
f63da68
Compare
ricardoV94
marked this pull request as ready for review
June 25, 2026 13:06
ricardoV94
force-pushed
the
proper_idx_gating
branch
3 times, most recently
from
June 27, 2026 12:58
f7ceb7d to
fe342d6
Compare
ricardoV94
commented
Jun 27, 2026
| continue | ||
| if not _index_provably_not_larger(idx, inp.type.shape[axis], fgraph): | ||
| return None | ||
| adv_indices = [ |
Member
Author
There was a problem hiding this comment.
this should be called adv_indices and input_static_dim
ricardoV94
force-pushed
the
proper_idx_gating
branch
4 times, most recently
from
July 2, 2026 14:17
da593b2 to
a727559
Compare
ricardoV94
marked this pull request as draft
July 5, 2026 08:23
ricardoV94
force-pushed
the
proper_idx_gating
branch
3 times, most recently
from
July 7, 2026 09:00
557c648 to
756c14b
Compare
ricardoV94
marked this pull request as ready for review
July 7, 2026 09:21
Generalize the duplicate-free reasoning used to gate the advanced-index write rewrites from a per-axis check to a joint one over the whole advanced index group, sharing the logic across subtensor.py and subtensor_lift.py. - _index_provably_unique: per-axis uniqueness, now also proving single-signed aranges and views (Reshape/DimShuffle) that preserve the value multiset. - _indices_jointly_unique: distinct joint coordinate tuples via all-axes uniqueness, a single Nonzero (e.g. tril_indices), or jointly-unique constants where no single axis is unique on its own. - _indices_provably_not_larger: bound a gather by the indexed axes' size using static shapes, per-axis bounds, or joint uniqueness.
ricardoV94
force-pushed
the
proper_idx_gating
branch
from
July 7, 2026 09:34
756c14b to
3814ee9
Compare
jessegrabowski
approved these changes
Jul 7, 2026
| if _is_provably_positive(step): # ascending | ||
| if _is_provably_non_negative(start): # all >= 0 | ||
| return True | ||
| if stop_c is not None and stop_c <= 0: # all <= -1 |
Member
There was a problem hiding this comment.
stop_c can't be symbolic? why not _is_provably_negative(stop_c) here
Member
Author
There was a problem hiding this comment.
Cuz we don't have a _is_provably_negative helper :)
| and step_c is not None | ||
| and step_c != 0 | ||
| ): | ||
| n = max(0, -(-(stop_c - start_c) // step_c)) # length, via ceil division |
Member
There was a problem hiding this comment.
push in some of these minus signs
Member
Author
There was a problem hiding this comment.
no bug ever came from simple algebra
Member
Author
There was a problem hiding this comment.
bedside reading: https://discuss.python.org/t/integer-ceiling-divide/91269/13
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
2 participants
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.
We had heuristics that work per-axis, but we should reason about the advanced indices together since they interact. E.g., tril_indices, would fail the per-axis checks, as each index (row, col) is larger than the respective axis and contains duplicates. But jointly they index less than the full array and don't have duplicates, so we want to lift it through Elemwise and the like.
Shows up in pymc-devs/pymc#8297