Many bug/performance fixes - #88

Merged

ndryden merged 31 commits into

mainfrom

review-fixes

Jul 30, 2026

Commits on Jul 30, 2026

Add test infrastructure: pytest config, shared fixtures, MPI/torchrun helpers
ndryden
committed
01176bd
Fix reporting and visualization bugs
ndryden
committed
75ec1a4
Fix U-Net construction validation and activation checkpointing
ndryden
committed
609e53d
Generate restart scripts from the run's actual environment and launch shape
ndryden
committed
4f6b72b
Fix IFS point-generation kernel bugs
ndryden
committed
4ea15e3
Make dataset id ordering deterministic and v1 label mapping split-consistent
ndryden
committed
9b44c7c
Harden checkpointing: atomic writes, corruption fallback, race-free best selection
ndryden
committed
4069b79
Compute validation Dice on hard predictions and sample-weighted val loss
ndryden
committed
a62c7ae
Make datagen RNG deterministic and category-search resume safe
ndryden
committed
5e18bac
Validate generated instances, write them atomically, and fix rasterization
ndryden
committed
2d42d7e
Validate config keys, wire --config overrides, implement benchmark sweeps
ndryden
committed
4287ea2
Make distributed datagen failure paths collective-safe
ndryden
committed
eb09eb1
Fix DDP device pinning, launcher detection, and worker teardown order
ndryden
committed
889001c
Reduce training/validation metrics globally across data-parallel replicas
ndryden
committed
e1d9556
Fix resume flag handling, stats-file headers, and step-counter persistence
ndryden
committed
4bf347b
Handle a missing parallel strategy and CPU-only devices in training
ndryden
committed
48c900a
Only treat .csv files as fractal categories when indexing instances
ndryden
committed
f13cc73
Clean up unused imports in test_reporting
ndryden
committed
56d55fb
Apply ruff formatting to checkpointing
ndryden
committed
22201f0
Rasterize by scattering point indices and store point clouds as float32
ndryden
committed
a41201d
Resolve mask paths once and size category search rounds adaptively
ndryden
committed
fe5af09
Skip the upsample pad when no padding is needed
ndryden
committed
57c1322
Trim per-step overhead in the training and validation compute path
ndryden
committed
6e35a3d
Cut redundant dataset I/O per sample and per startup scan
ndryden
committed
27b497b
Load the resume checkpoint on rank 0 and broadcast it
ndryden
committed
159a4b0
Adapt tests and helpers to the always-distributed design
ndryden
committed
7385cc2
Persist the cumulative optimizer-step total across resumes
ndryden
committed
5819931
Gather the dataset id digest on the backend's device
ndryden
committed
d796633
Preserve composite label rows in the legacy mask-value union
ndryden
committed
61b25da
Gate the torch profiler solely on its own environment variable
ndryden
committed
dcac332
Require Python 3.10+
ndryden
committed
ea6c8e3

Navigation Menu

Uh oh!

Many bug/performance fixes - #88

Many bug/performance fixes#88
ndryden merged 31 commits into
mainfrom
review-fixes

Commits on Jul 30, 2026

Add test infrastructure: pytest config, shared fixtures, MPI/torchrun helpers

Fix reporting and visualization bugs

Fix U-Net construction validation and activation checkpointing

Generate restart scripts from the run's actual environment and launch shape

Fix IFS point-generation kernel bugs

Make dataset id ordering deterministic and v1 label mapping split-consistent

Harden checkpointing: atomic writes, corruption fallback, race-free best selection

Compute validation Dice on hard predictions and sample-weighted val loss

Make datagen RNG deterministic and category-search resume safe

Validate generated instances, write them atomically, and fix rasterization

Validate config keys, wire --config overrides, implement benchmark sweeps

Make distributed datagen failure paths collective-safe

Fix DDP device pinning, launcher detection, and worker teardown order

Reduce training/validation metrics globally across data-parallel replicas

Fix resume flag handling, stats-file headers, and step-counter persistence

Handle a missing parallel strategy and CPU-only devices in training

Only treat .csv files as fractal categories when indexing instances

Clean up unused imports in test_reporting

Apply ruff formatting to checkpointing

Rasterize by scattering point indices and store point clouds as float32

Resolve mask paths once and size category search rounds adaptively

Skip the upsample pad when no padding is needed

Trim per-step overhead in the training and validation compute path

Cut redundant dataset I/O per sample and per startup scan

Load the resume checkpoint on rank 0 and broadcast it

Adapt tests and helpers to the always-distributed design

Persist the cumulative optimizer-step total across resumes

Gather the dataset id digest on the backend's device

Preserve composite label rows in the legacy mask-value union

Gate the torch profiler solely on its own environment variable

Require Python 3.10+