Kumo-Forecast — GGUF

GGUF conversion of NVIDIA's Kumo-Forecast multivariate time-series forecasting 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/kumoforecast-gguf", "kumoforecast-f32.gguf")
model = zsfm.KumoForecastModel(gguf_path)

# channels: number of variates (1–7)
# history_flat: flat [channels × 96] history values, row-major
# mask: optional flat [channels × 96], 1=observed, 0=missing
point = model.forecast(history_flat, channels=7, mask=None)
# -> List[float], 24 denormalized point forecasts for channel 0

Rust / CLI (cargo install zsfm)

cargo install zsfm --locked

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

echo '{"channels": 7, "sequence": [[1,2,3,4,5,6,7,8], [10,20,30,40,50,60,70,80]], "mask": [[1,1,1,1,1,1,1,1], [1,1,1,0,1,1,1,1]]}' \
  | zsfm kumoforecast infer --gguf gguf/kumoforecast-f32.gguf

Response format

{
  "forecast": [12.3, 12.5, 12.6, 12.7, ...]
}

24 denormalized point forecasts for channel 0.

Architecture

Kumo-Forecast is built on the MOMENT backbone with a trainable forecasting head:

• Input: multivariate time series with up to 7 channels, 96-step context.
• Encoder: RevIN → patchify → value embed + sinusoidal position → per-channel T5 encoder (relative + observation-aware bias, gated gelu_new FFN) → cross-channel attention → flat head on channel 0 → RevIN denormalize.
• Output: 24-step point forecast for channel 0.
• Supports missing-value masks: observed steps marked 1, missing 0 are replaced by the model's mask embedding.

License

Conversion code: MIT (amaye15/zsfm-rs). Weights: Apache-2.0, per NVIDIA's release — unrestricted, including commercial use.