Releases: plasma-umass/scalene
Releases · plasma-umass/scalene
Increased profiling accuracy, Python 3.11 support
Changes in this release:
- increased accuracy of time attribution to specific lines for CPU & GPU profiling (also reduces memory consumption)
- increased accuracy of memory attribution to specific liens
- added per-process GPU accounting for NVIDIA, which can dramatically increase accuracy when profiling on shared GPUs
- added support for Python 3.11
- documented the command-line option to force Scalene to ignore options after that point (
---
)
Full Changelog: v1.5.8...v1.5.9.1
Increased profiling accuracy, Python 3.11 support
Changes in this release:
- increased accuracy of time attribution to specific lines for CPU & GPU profiling (also reduces memory consumption)
- increased accuracy of memory attribution to specific liens
- added per-process GPU accounting for NVIDIA, which can dramatically increase accuracy when profiling on shared GPUs
- added support for Python 3.11
- documented the command-line option to force Scalene to ignore options after that point (
---
)
Full Changelog: v1.5.8...v1.5.9
Fixed GUI display issues
Changes in this release:
- fixed missing GUI files from Linux wheels;
- fixed some issues launching browser to display GUI results;
Minor UI update: memory activity now shown as pies instead of numbers
What's Changed
UI improvements:
- Memory activity now shown as pies instead of numbers
Compatibility:
- Working towards conda builds.
Full Changelog: v1.5.6...v1.5.7
Bug fixes for memory attribution and multiprocessing
What's Changed
Improved functionality and accuracy:
- Fixed Python memory attribution for large requests.
- Fixed an issue with the
multiprocessing
library.
UI improvements:
- Fixed reporting of the Python fraction of memory allocated.
Compatibility:
- Removed
nvidia-ml-py
dependency, which was causing a reported issue with Dask (#378).
Full Changelog: v1.5.5...v1.5.6
Bug fixes, improved CPU and memory accounting
What's Changed
Improved functionality and accuracy:
- Fixed occasional segfaults caused by unaligned memory allocations.
- Corrected an issue with attribution of CPU time with threads.
- Leak detection enabled by default.
UI improvements:
- Hovering over memory timelines now shows amount of memory consumed, and when.
- Memory timelines are compressed, reducing the size of profiles and reducing the memory consumption of the UI.
- Suspected leaks are now highlighted.
Compatibility:
- Moved to Python 3.8.
Full Changelog: v1.5.4...v1.5.5
Improved UI, especially for GPU usage; accuracy and compatibility improvements
What's Changed
Improved functionality and accuracy:
- Fixed memcpy attribution (now on specific lines, just like allocations).
- GPU profiling now enabled on Apple systems
UI improvements:
- Memory timelines are now zoomable.
- Fatter bars, zoomable timelines, fixed sorting for code line separators
- Added explanations when hovering over column headers.
- Omit function summaries if no functions.
- Added GPU memory profiling.
- Report peak instead of average GPU memory.
- GPU utilization now pies.
Compatibility:
Full Changelog: v1.5.3...v1.5.4
Maintenance release: fixes virtualized GPU issue, Windows loading, output issue
What's Changed
- Adds exception handling to workaround a virtualized GPU issue (#323).
- Added average memory consumption calculation to function summaries.
- Fixes a missing argument issue in output (#344)
- Fixes an issue with Jupyter notebooks when they don't have access to a web browser.
Full Changelog: v1.5.2...v1.5.3
Jupyter support for new GUI; auto-exit of server
What's Changed
- Scalene's web-based GUI is now integrated into Jupyter notebooks
- When using
--cpu-only
or profiling in Jupyter, columns for memory profiling (which would all be empty) are now hidden - The local webserver now exits after 5 seconds.
Full Changelog: v1.5.1...v1.5.2
Local web-based GUI
What's Changed
- Scalene now launches its web-based GUI locally by default. After profiling, It opens a browser tab to a local webserver and automatically brings up the most recent profile. (The old behavior is still available by using
--cli
on the command line.)
Full Changelog: v1.5.0...v1.5.1