Repository navigation
Releases
Releases · NVIDIA/Kumo-TS
Kumo-TS v1.0.1
Kumo-TS v1.0.1 adds anomaly explainability and historical attribution, missingness-aware forecasting, and inference and fine-tuning reliability improvements.
Changes since v1.0.0.
Added
- Anomaly explanation metrics, per-feature contributions, companion CSV exports, and consolidated PDF reports. (#68, #72)
- YAML-configured anomaly explanation settings and optional reconstruction diagnosis. (#75, #77)
- Historical attribution for anomaly explanations, including event validation and report timestamp preservation. (#93)
- Configurable shared feature embeddings for anomaly detection, with a positional compatibility mode. (#80)
- Missingness-aware forecasting in the forecasting SDK:
ForecastingConfig(handle_missingness=True)routes inputs with
NaN gaps to a missingness-aware model (impute/, Backbone-LF+ CRS) that reads per-channel/per-timestep observation
masks instead of zero-filling. Options:impute_ckpt,impute_normalization(history/provided/checkpoint),
impute_history_rows,impute_scaler_stats,impute_source_dataset,impute_channel_alignment(name/positional),
plusfit_impute_scaler_stats. - The released missingness checkpoint is loaded from
hf://nvidia/Kumo-Forecast/kumo-forecast-1.2.0(any
hf://org/repo[@revision][/subfolder]reference or local folder is accepted). examples/missingness_quickstart.pyandsdk/tests/test_imputation.py(real-checkpoint tests opt-in via
KUMO_IMPUTE_CKPT/KUMO_IMPUTE_HF/KUMO_IMPUTE_CSV).
Changed
- Renamed the repository to Kumo-TS, forecasting to Kumo-Forecast, and AD Diffusion to Kumo-Anomaly; updated package names, directories, artifact prefixes, model references, and documentation. (#83, #84, #85, #86)
- Added published Kumo blogs and research papers to the repository README. (#91)
- The NULL zero-fill warning now points to
handle_missingness=True;clear_model_cache()also clears cached
missingness models.
Fixed
- Replayed saved fine-tuning feature order and preprocessing during anomaly inference rather than fitting transforms to the scoring batch. (#32)
- Corrected
preprocess_simpletarget-dimension forwarding. (#44) - Made forecasting RevIN statistics request-local to prevent interference between concurrent calls. (#70)
- Removed GPU-count-dependent diffusion random-number behavior. (#78)
- Corrected DPM-Solver++ timesteps, clipped schedules, and cumulative unconditional noise; anomaly inference now uses DPM-Solver++ exclusively. (#89)
- Restored fine-tuning run-config support for validation data, labels, dropped columns, pretrained models, and model configuration. (#90)
- Aligned anomaly inference window, split, and scale settings with saved fine-tuned checkpoints, including preprocessing overrides and multi-device paths. (#92)
Notes
- Repository release version:
v1.0.1; package versions:kumo-forecast1.0.1 andkumo-anomaly1.0.1. - Missingness-aware forecasting is introduced by #87. Its
kumo-forecast-1.2.0model asset version is independent of SDK release 1.0.1. - Anomaly output remains int64 (
0= normal,1= anomaly). Explanation and historical attribution are opt-in. - Missingness-aware forecasting is inference-only in this release and cannot be combined with forecasting interpretability.
Known issues
- Pretrained anomaly inference identifies and excludes timestamp/ground-truth columns only when reporting or historical attribution is enabled. Otherwise numeric metadata (including datetime values converted to epoch numbers) can enter the model and alter detections. Pass signal-only inputs consistently, or retain identical reporting/attribution metadata settings when selecting and explaining events. Fine-tuned checkpoints with saved feature schemas select their recorded features. See the input metadata guidance. This behavior predates 1.0.1; this release documents the workaround without changing detection behavior.
Full changelog: v1.0.0...v1.0.1
NV-Tesseract v1.0.0
Changes since v0.1.0.
Added
- Forecasting output for all input channels through
ForecastingConfig(return_all_channels=True). (#26) - Hugging Face Hub model wrappers for both forecasting and AD Diffusion, with
from_pretrained()loading. (#48) - TAO AutoML run-config support and scalar
metrics.jsonoutputs in both fine-tuning examples; per-epoch results are available inepoch_metrics.json. (#54, #59) - Automatic multi-GPU distributed fine-tuning in both examples, with
--num-gpusto limit GPU use. (#61) - Forecasting feature-axis attribution, embedding Integrated Gradients, and embedding stability reporting, with JSON, CSV, and PDF artifacts. (#56, #62)
- Multi-GPU execution for DARR context embedding and channel-flow Interpretability analysis. (#67)
- Automatic MPS device detection for AD inference on Apple Silicon. (#55)
Changed
perform_forecasting()now acceptsconfig=ForecastingConfig(...)or a YAML path in place of flat inference arguments. Keepdfand optionalcontext_dfas direct inputs; move settings such astarget_column,forecast_horizon, andckptinto the config. See the forecasting SDK guide and YAML template. (#74)- AD analysis renames
config_pathtomodel_config_path. Reporting options belong insdk_config=ADDiffusionConfig(...)or a YAML file. See the AD Diffusion guide. (#76) - Simplified forecasting interpretability around batched Jacobian analysis and embedding Integrated Gradients; removed Shapley, coupling, and legacy sharded paths, and corrected semantic-flow report layout. (#60)
- Contributions target
dev;mainis reserved for release-ready code. (#65)
Fixed
- Isolated temporary forecasting CSVs per call to prevent collisions between concurrent requests. (#34)
- Aligned AD inference normalization with training and returned reconstructions in the same scale as targets. (#41)
- Restored AD preprocessing artifacts with the reusable
AdaptiveNormalizer, preserving fitted transforms through JSON save/load; added round-trip coverage for normalizers, scalers, transformers, and PCA. (#71) - Kept shared memory open until AD worker inference completes. (#47)
- Aligned DARR input and context channel ordering when the same columns arrive in different orders. (#52)
- Corrected base/cross-channel checkpoint selection and inferred cross-channel mode from checkpoint weights. (#73)
- Excluded SDK-added padding from AD MAE, L2, and thresholding, while retaining full-width target and reconstruction outputs. Low-level results include
valid_feature_maskandscore_feature_count. Scores may differ for padded inputs. (#79) - Verify downloaded model weight files before reporting a successful download in both SDKs. (#82)
- Clarified that AD
Anomalyoutput uses int64 values (0= normal,1= anomaly); detection behavior is unchanged. - Synchronized forecasting lockfile metadata with its dependency declarations without changing locked package versions.
Notes
- Repository release version:
v1.0.0. - Package metadata versions:
tesseract_forecasting1.0.0 andad-diffusion-oss1.0.0.
NV-Tesseract v0.1.0
First public release of NV-Tesseract.
Added
- Forecasting package with DataFrame-first inference, DARR context-enhanced forecasting, cross-channel forecasting support, and Hugging Face weight loading.
- Forecasting interpretability framework with input attributions, semantic-flow diagnostics, forecast-vs-history ratios, trajectory stability metrics, and optional PDF/JSON artifacts.
- AD Diffusion package for multivariate anomaly detection with SCS and MACS adaptive thresholding, DPM-Solver inference, and Hugging Face weight loading.
- Fine-tuning examples for both forecasting and AD Diffusion.
- Lightweight CI covering linting, SPDX checks, forecasting tests, AD Diffusion tests, and example tests.
- Public documentation for installation, examples, dataset expectations, model assets, contribution flow, security reporting, and third-party notices.
Changed
- Raised the PyTorch dependency floor to
torch>=2.7.0for Blackwell GPU support. - Removed unused forecasting audio/vision PyTorch dependencies and the legacy
mac-mpsextra. - Switched user-facing progress output from direct
printcalls to logging where appropriate. - Updated documentation to reference public Hugging Face model repositories.
Fixed
- Fixed DARR retrieval behavior when forecast horizons exceed the model horizon.
- Fixed AD Diffusion complementary mask aggregation so each target mask selects reconstructions from its own strategy.
- Fixed AD thresholding, packaging, and README examples.
- Fixed pandas frequency deprecation warnings in README examples.
Notes
- Repository release version:
v0.1.0. - Package metadata versions at this release:
forecastingis0.1.0;ad-diffusion-ossis1.0.0.