Rework QJIT class into distinct compilation stages #531
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
dime10
force-pushed
the
frontend-refactor-1
branch
from
de6fe29
to
2ce0add
Compare
Also switch to only returning the "canonical" jaxpr output without implicit result values.
This cache can replace the existing caching mechanism for static arguments, and also caches different function definitions based on PyTreeDefs of dynamic arguments. Different function definitions based on the ShapedArrays of dynamic is not (yet) provided, as that would incurr additional overhead for type promotion checks.
The compilation_pipelines module is renamed as the name is outdated. The QJIT class is rewritten into more easily reuseable portions and simplified where possible. The updated class makes use of the new compilation cache to access previous versions of compiled functions.
dime10
force-pushed
the
frontend-refactor-2
branch
from
6139eae
to
1c555ba
Compare
rmoyard
reviewed
Feb 21, 2024
josh146
reviewed
Feb 22, 2024
dime10
force-pushed
the
frontend-refactor-2
branch
from
728ae41
to
3805b29
Compare
dime10
added
the
author:build-wheels
dime10
force-pushed
the
frontend-refactor-2
branch
from
5917cb4
to
42791fe
Compare
dime10
force-pushed
the
frontend-refactor-2
branch
from
42791fe
to
b0a9338
Compare
Codecov ReportAttention: Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## main #531 +/- ##
==========================================
- Coverage 99.55% 99.42% -0.13%
==========================================
Files 52 52
Lines 8457 8510 +53
Branches 559 568 +9
==========================================
+ Hits 8419 8461 +42
- Misses 20 27 +7
- Partials 18 22 +4 ☔ View full report in Codecov by Sentry. |
Also fixes wrong attribute access uncovered by test.
dime10
force-pushed
the
frontend-refactor-2
branch
from
cf11747
to
414de95
Compare
rauletorresc
pushed a commit
that referenced
this pull request
Feb 26, 2024
This is part 2 of a refactor started in #529. The QJIT class is reworked into 5 distinct compilation stages: - pre-compilation (like autograph) - capture (jaxpr generation) - ir-generation (mlir generation) - compilation (llvm and binary code generation - cannot be split up since this happens in the compiler driver) - execution The class is also streamlined by using a new compilation cache to handle previously compiled functions and signature lookups. One point of contention might be the results produced by the split of the `trace_to_mlir` function, which have been simplified and need to be double checked against #520. EDIT: c71c322 should address this concern [sc-57014] closes #268 closes #520
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.
This is part 2 of a refactor started in #529.
The QJIT class is reworked into 5 distinct compilation stages:
The class is also streamlined by using a new compilation cache to handle previously compiled functions and signature lookups.
One point of contention might be the results produced by the split of the
trace_to_mlirfunction, which have been simplified and need to be double checked against #520. EDIT: c71c322 should address this concern[sc-57014]
closes #268
closes #520