amaye15/lag-llama-gguf

Lag-Llama — GGUF

GGUF conversion of Lag-Llama — a LLaMA-style autoregressive decoder for time series. Converted and run with zsfm, a Rust workspace that ports zero-shot forecasting and tabular foundation models to GGUF + candle. No PyTorch, no Python runtime required to run inference.

F32	F16	Q8_0
lag_llama-f32.gguf	lag_llama-f16.gguf	lag_llama-q8.gguf

F16 is generally the best size/accuracy trade-off; Q8_0 is smallest. This repo's default recommendation matches the upstream conversion default: F32.

Context must be at least 32 timesteps (Lag-Llama has no hard minimum, but 32 gives a meaningful example) — a shorter context fails with context too short. The examples below use a 32-value context.

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/lag-llama-gguf", "lag_llama-f32.gguf")

model = zsfm.LagLlamaModel(gguf_path)

context = [0.85, 0.93, 1.01, 1.09, 1.17, 1.25, 1.33, 1.06, 1.14, 1.22, 1.3, 1.38, 1.46, 1.54, 1.27, 1.35, 1.43, 1.51, 1.59, 1.67, 1.75, 1.48, 1.56, 1.64, 1.72, 1.8, 1.88, 1.96, 1.69, 1.77, 1.85, 1.93]
point = model.forecast(context, horizon=64)
# -> List[float], the point forecast

Rust / CLI (cargo install zsfm)

cargo install zsfm --locked

# downloads the original weights and converts to GGUF locally
# (produces the same bytes as lag_llama-f32.gguf in this repo):
zsfm lag-llama convert --dtype f32 -o gguf/lag_llama-f32.gguf
echo '{"context": [0.85, 0.93, 1.01, 1.09, 1.17, 1.25, 1.33, 1.06, 1.14, 1.22, 1.3, 1.38, 1.46, 1.54, 1.27, 1.35, 1.43, 1.51, 1.59, 1.67, 1.75, 1.48, 1.56, 1.64, 1.72, 1.8, 1.88, 1.96, 1.69, 1.77, 1.85, 1.93], "horizon": 64}' \
  | zsfm lag-llama infer --gguf gguf/lag_llama-f32.gguf

-m/--model takes the full HuggingFace repo id (default time-series-foundation-models/Lag-Llama) — there's only one published checkpoint for this architecture, so you normally don't need to change it. -o/--output defaults to gguf/lag_llama-f32.gguf regardless of --dtype, so always pass -o explicitly (as above) — otherwise repeated runs overwrite the same file under a name that may not even match the dtype you chose:

zsfm lag-llama convert --dtype f32 -o gguf/lag_llama-f32.gguf
zsfm lag-llama convert --dtype q8  -o gguf/lag_llama-q8.gguf

To skip conversion and run a file already published here:

huggingface-cli download amaye15/lag-llama-gguf lag_llama-f32.gguf --local-dir .
echo '{"context": [0.85, 0.93, 1.01, 1.09, 1.17, 1.25, 1.33, 1.06, 1.14, 1.22, 1.3, 1.38, 1.46, 1.54, 1.27, 1.35, 1.43, 1.51, 1.59, 1.67, 1.75, 1.48, 1.56, 1.64, 1.72, 1.8, 1.88, 1.96, 1.69, 1.77, 1.85, 1.93], "horizon": 64}' \
  | zsfm lag-llama infer --gguf lag_llama-f32.gguf

Source, the other 9 time-series forecasters + 5 tabular models, and full docs: amaye15/zsfm-rs.

Response format

{
  "id": "forecast-000001932b7a1234",
  "object": "forecast",
  "created": 1736290000,
  "model": "lag-llama",
  "choices": [{
    "index": 0,
    "forecast": {
      "point": [2.1, 2.3, 2.5],
      "quantiles": {}
    },
    "finish_reason": "stop"
  }],
  "usage": {"context_length": 32, "forecast_length": 64}
}

This architecture has no quantile head, so quantiles is always empty and point is the only forecast.

Pass a batch of series ("context": [[...], [...]]) for one choice per series.

Architecture

Lag-Llama is a LLaMA-style autoregressive decoder for time series:

• Input: Lagged values of the target series are appended as covariates alongside the most-recent context, giving the model a multi-scale view of the history
• Backbone: Standard LLaMA decoder (RoPE, causal self-attention, SwiGLU FFN)
• Output: Point forecast for each future timestep via autoregressive decoding
• Checkpoint: Reads .ckpt (PyTorch Lightning pickle) directly — no Python required for conversion

License

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