feat: default --xla_gpu_autotune_level=0 in jax_wrapper (#131) - #132
Merged
Conversation
GPU autotuning dominates cold JAX compiles (up to ~7min for one fusion, autolens_profiling#74: 17x cold probe, -40% cold full fit off) with no measured eval speed-up on PyAuto likelihoods. Appended to XLA_FLAGS unless the user already sets an autotune level. Stacks with the default compilation cache (#128): worst-case first-fit ~70min -> ~30s. Co-Authored-By: Claude Fable 5 <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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Overview
Rollout of autolens_profiling#74 verdict 2 (closes #131): default
--xla_gpu_autotune_level=0injax_wrapper.py. GPU autotuning dominates cold JAX compile times (up to ~7m30 for a single fusion on the A100 pathological case) while giving no measurable evaluation speed-up on PyAuto likelihoods (steady-eval parity across the measured matrix; the 4800-eval full fit was 40 % faster end-to-end with it off; fixed-input logL bit-identical). Stacked with the default compilation cache (#128): worst-case first-fit UX ~70 min → ~30 s.API Changes
None (public Python API untouched). Environment behavior at import:
XLA_FLAGSnow also gains--xla_gpu_autotune_level=0unless the user's flags already contain an--xla_gpu_autotune_levelsetting (any value — always respected). Same append/env-respecting pattern as feat: JAX persistent compilation cache by default + XLA_FLAGS append fix (#127) #128.Tests
test_autoconf/test_jax_wrapper.py(default applied; preset level respected and default suppressed; composes with user flags); 2 pre-feat: default --xla_gpu_autotune_level=0 in jax_wrapper (env-respecting) #131 exact-equality tests updated for the new default.--xla_gpu_autotune_level=2survives and suppresses the default).Heart
RED at ship time — the same 5 pre-existing unrelated reasons as autolens_profiling#76's ship; human-acknowledged for this PR-open (merge stays human).
🤖 Generated with Claude Code