amaye15/timesfm-gguf

timesfm-rs

Pure Rust converter and inference engine for google/timesfm-2.5-200m-pytorch.

Pre-converted GGUF files are available at amaye15/timesfm-gguf. Produces GGUF v3 files and runs native forecasting — no Python required.

Build

cargo build --release

Convert

Downloads the model from HuggingFace and writes a GGUF file:

# F16 (recommended)
./target/release/timesfm-rs convert --model google/timesfm-2.5-200m-pytorch --dtype f16 --output gguf/timesfm-f16.gguf

# Q8_0 (smallest)
./target/release/timesfm-rs convert --dtype q8 --output gguf/timesfm-q8.gguf

# F32 (full precision)
./target/release/timesfm-rs convert --dtype f32 --output gguf/timesfm-f32.gguf

To convert all dtypes at once:

./scripts/convert_all.sh

HuggingFace token (optional for public models):

HF_TOKEN=hf_... ./scripts/convert_all.sh

Inspect tensors

Print all tensor names and shapes from a GGUF file:

./target/release/timesfm-rs inspect-tensors gguf/timesfm-f16.gguf

Infer

Run forecasting from comma-separated context values:

echo '{"context": [1.0, 1.2, 1.5, 1.3, 1.8, 2.0, 1.9, 2.1], "horizon": 64}' \
  | ./target/release/timesfm-rs infer --gguf gguf/timesfm-f16.gguf

Output is JSON in an OpenAI-compatible forecast format with point forecast and all 9 quantile channels (q0.1–q0.9):

{
  "id": "forecast-000001932b7a1234",
  "object": "forecast",
  "created": 1749686400,
  "model": "timesfm",
  "choices": [{
    "index": 0,
    "forecast": {
      "point": [2.1, 2.3, 2.5, "..."],
      "quantiles": {
        "0.10": [1.8, 2.0, 2.2, "..."],
        "0.50": [2.1, 2.3, 2.5, "..."],
        "0.90": [2.4, 2.6, 2.8, "..."]
      }
    },
    "finish_reason": "stop"
  }],
  "usage": {"context_length": 8, "forecast_length": 64}
}

Batch inference — pass multiple series as a nested array to get one Choice per series:

echo '{"context": [[1.0, 1.2, 1.5], [2.0, 2.2, 2.5]], "horizon": 64}' \
  | ./target/release/timesfm-rs infer --gguf gguf/timesfm-f16.gguf

Python bindings

Install with maturin inside a virtual environment:

python -m venv .venv && source .venv/bin/activate
pip install maturin
maturin develop --features python

import timesfm_rs

model = timesfm_rs.TimesFM("gguf/timesfm-f16.gguf")

result = model.forecast([1.0, 1.2, 1.5, 1.3, 1.8, 2.0], horizon=64)
fc     = result["choices"][0]["forecast"]
point  = fc["point"]           # median forecast
q10    = fc["quantiles"]["0.10"]  # 10th-percentile
q90    = fc["quantiles"]["0.90"]  # 90th-percentile

# Batch — one Choice per series
result = model.forecast([[1.0, 1.2, 1.5], [2.0, 2.2, 2.5]], horizon=64)

forecast returns a Python dict in the same OpenAI-compatible format as the CLI.

Architecture notes

TimesFM 2.5 is a decoder-only patch transformer:

• Input: Context is split into 32-value patches; each patch is instance-normalized (RevIN with cumulative running statistics) and concatenated with an observation mask, producing 64-dim tokenizer inputs
• Backbone: 20-layer causal transformer, d_model=1280, 16 heads, head_dim=80, d_ff=1280; 200M parameters
• Attention: Sandwich norms (pre-norm → attention → post-norm + residual); fused QKV projection; per-dimension query scale 1.442695/√80 × softplus(param); QK RMSNorm after RoPE
• AR decoding: KV cache with 4-patch decode stride — O(n) instead of O(n²) re-forward per step
• Output: Each output patch decoded through a ResidualBlock to [n_patches, 128, 10]; index 0 is point forecast, indices 1–9 are quantiles q0.1–q0.9