Kumo-Anomaly — GGUF

GGUF conversion of NVIDIA's Kumo-Anomaly time-series anomaly localization model, converted and run with zsfm, a Rust workspace that ports zero-shot forecasting and tabular foundation models to GGUF + Burn, verified bit-exact against the original PyTorch checkpoint. No PyTorch, no Python runtime required to run inference.

GGUF embeds its own architecture metadata — no separate config.json is needed.

Use it

Python (pip install zsfm)

pip install zsfm huggingface_hub

import zsfm
from huggingface_hub import hf_hub_download

gguf_path = hf_hub_download("amaye15/kumoanomaly-gguf", "kumoanomaly-f32.gguf")
model = zsfm.KumoAnomalyModel(gguf_path)

# data: flat rows × ncols in row-major order (12 values = 4 rows × 3 cols)
data = [1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10.0, 11.0, 12.0]
mask = [1.0, 1.0, 1.0, 1.0, 0.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0]  # optional

# rows=4, ncols=3 — each row is one time step with 3 variates
res = model.localize(data, rows=4, ncols=3, mask=mask, nsample=30)
# res.residual: per (row, column) mean-abs reconstruction residuals (length = rows × ncols)
# res.recon:    reconstruction values
# res.target:   target values
# Large residuals indicate anomalous steps.

Rust / CLI (cargo install zsfm)

cargo install zsfm --locked

# downloads the original NVIDIA weights and converts to GGUF locally
zsfm kumoanomaly convert

echo '{"data": [[1,2,3],[4,5,6],[7,8,9]]}' \
  | zsfm kumoanomaly infer --gguf gguf/kumoanomaly-f32.gguf --nsample 30

Response format

{
  "residual": [0.01, 0.02, ...],
  "recon": [1.0, 2.0, ...],
  "target": [1.0, 2.0, ...],
  "score_feature_count": 3,
  "valid_feature_mask": [1.0, 1.0, 1.0],
  "ncols": 3
}

residual holds per-step mean-abs reconstruction residuals — large values indicate anomalous time steps. Length = rows × columns.

Architecture

Kumo-Anomaly is a diffusion-based anomaly localization model:

• Input: multivariate time series, normalized, with optional missing-value mask.
• Encoder: transformer-based diffusion encoder that reconstructs the input.
• Output: per-cell reconstruction residuals that highlight anomalous time steps.
• Supports up to 40 columns (variates) per time step.

License

Conversion code: MIT (amaye15/zsfm-rs). Weights: OpenMDW-1.1, per NVIDIA's release — commercial use permitted.