Add specialization rewrite for solve with batched b #482
Conversation
This file contains 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
|
The vectorization of shape operations (mostly reshape) from #454 is also critical for the discourse-related performance issue |
ricardoV94
force-pushed
the
solve_opt
branch
2 times, most recently
from
22f01cc
to
2eae804
Compare
Codecov Report
Additional details and impacted files@@ Coverage Diff @@
## main #482 +/- ##
==========================================
+ Coverage 80.73% 80.75% +0.01%
==========================================
Files 160 161 +1
Lines 46007 46058 +51
Branches 11245 11258 +13
==========================================
+ Hits 37146 37195 +49
Misses 6636 6636
- Partials 2225 2227 +2
|
ricardoV94
commented
Nov 8, 2023
| matrix_b_solve = Blockwise(new_core_op) | ||
|
|
||
| # Apply the rewrite | ||
| new_solve = _T(matrix_b_solve(a, _T(b))) |
There was a problem hiding this comment.
Actually we could put any batched b dimension here when there are multiple of them. We may choose the larger one to reduce outer looping
ricardoV94
commented
Nov 8, 2023
| @@ -131,6 +138,52 @@ def generic_solve_to_solve_triangular(fgraph, node): | |||
| ] | |||
|
|
|||
|
|
|||
| @register_stabilize | |||
There was a problem hiding this comment.
I want to make this a specialize only rewrite but for now it's in stabilize because pymc includes those before calling grad, and otherwise we still end up with messy unoptimized blockwise graphs
ricardoV94
force-pushed
the
solve_opt
branch
from
8d254f9
to
4f4379c
Compare
twiecki
approved these changes
Nov 10, 2023
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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Outdated suggestions cannot be applied.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.
Related to pymc-devs/pymc#6993 and https://discourse.pymc.io/t/version-dependant-slowing-down-of-gaussian-mixture-sampling-in-ubuntu-20-04/13219
This PR adds a rewrite that optimizes solve when there is a batched vector b (b_ndim=1) and a single a.
This form was previously used for the logp of MvNormal in PyMC and is potentially much faster when the batched dimensions of b are large. More importantly the grad will look better.
I have a couple of other PRs that improve the blockwise grads for the truly new supported cases, but that it's slower work.
PS: Obviously we should also implement the JAX Ops #430, but this looks like a nice optimization, specially if it solves the regression described in the discourse issue above.