plasma-umass/scalene v2.4.0

Free-threaded wheels, M4 fix, safer API keys, and formal verification

on GitHub

What's Changed

Headlines

• Free-threaded Python wheels on PyPI: cp313t and cp314t wheels for macOS and Windows, and cp314t for Linux. Release builds now check that every wheel's native modules import, and that on free-threaded builds they don't re-enable the GIL.
• Apple Silicon M4+ fix: on chips with SME, profiling numpy/Accelerate code could silently corrupt numerical results (#1056). CPU sampling on those chips no longer uses signals.
• Safer AI-provider keys: API keys are never embedded in generated HTML. The new opt-in scalene view --api-keys-from-env serves them from memory over localhost instead. The GUI is now served from a private directory rather than the whole temp dir.
• Correctness: fixes found by new formal models (TLA+ and Lean) of Scalene's sampling and attribution, including several divide-by-zero and unbounded-sampling bugs.

New feature work

• Publish free-threaded (cp313t/cp314t) wheels by @emeryberger in #1104
• Signal-free CPU sampling on Apple Silicon with SME (M4+) (#1056) by @emeryberger in #1058
• Add opt-in scalene view --api-keys-from-env by @emeryberger in #1097
• Bound --stacks CPU cost on processes with many Python threads (#1056) by @emeryberger in #1057
• Declare free-threaded safety for the _scalene_unwind module by @emeryberger in #1065

Bug fixes

• Build on Windows for free-threaded Python by @rbo-corematic in #1102
• Fix free-threaded builds: limited-API module name, pydantic-core pin, mypy 2.4 by @emeryberger in #1103
• Preserve ReplacementSemLock identity across spawned processes by @HassanRady in #1099
• Prevent API credentials from being embedded in generated HTML by @HassanRady in #1095
• Serve the GUI from a private directory, not the whole temp dir by @emeryberger in #1100
• Fix profiling scripts with a UTF-8 BOM by @mameikagou in #1092
• Fix UnicodeEncodeError for non-ASCII file paths by @emeryberger in #1087
• Fix #1063: non-ASCII --profile-only no longer crashes with SIGSEGV by @emeryberger in #1064
• Retry EINTR on poll/select so profiling signals don't break native blocking calls (#1060) by @emeryberger in #1062
• Fix "dictionary changed size during iteration" race in stack output by @emeryberger in #1067
• Fix divide-by-zero and unbounded-sampling bugs found by formalization by @emeryberger in #1077 and #1078
• Sync requirements.txt with pyproject.toml runtime dependencies by @emeryberger in #1085
• Lowercase scalene view option help to match the rest of the CLI by @emeryberger in #1101

Formal verification

• Add formal models (TLA+ and Lean) of Scalene's concurrency and correctness, and prove the profiler-correctness desideratum (unbiased, consistent attribution), the Poisson sampler, memory-sampler correctness and the Python/native classifier, in #1068, #1070, #1072, #1075, #1076, #1080, #1082, #1083 and #1084

Testing and CI

• CI fails when native modules don't import or tests don't collect by @emeryberger in #1105
• Free-threaded CI jobs run their tests again (#1089, #1103), and a test-suite sys.executable leak no longer skips the free-threaded memory tests (#1066)
• Deflaking: #1069, #1071, #1074, #1096

Note for free-threaded Python 3.13 users

pydantic-core no longer ships cp313t wheels and its newest releases can't be built from source there. Until that changes, install with:

pip install scalene "pydantic-core<=2.46.5"

New Contributors

• @mameikagou made their first contribution in #1092
• @HassanRady made their first contribution in #1095
• @rbo-corematic made their first contribution in #1102

Full Changelog: v2.3.0...v2.4.0