eugenehp/glucofm-encoder

GlucoFM encoder weights

Frozen encoder-only checkpoints for GlucoFM (dual-stream CGM foundation model), converted for the Rust glucofm / RLX stack.

Not medical advice. Research reimplementation. Official Google GlucoFM weights are not public. This package redistributes / converts a community OpenCGM reconstruction.

Files

File	Format	Notes
glucofm_encoder.safetensors	F32 safetensors	Preferred for glucofm load / train warm-start
glucofm_encoder.f16.safetensors	F16 safetensors	~2× smaller disk
glucofm_encoder.rlxp	RLX package F32	Flat mmap weight-only pack (rlxp / infer feature)
glucofm_encoder.f16.rlxp	RLX package F16	~2× smaller RLXP
glucofm_encoder.f32.gguf	GGUF F32	Lossless GGUF dump
glucofm_encoder.f16.gguf	GGUF F16	Smaller disk; dequantized at load
glucofm_encoder.q8.gguf	GGUF Q8_0	~4× smaller than F32; dequantized at load
glucofm_encoder.bin	Native GLFM v1	Compact glucofm-native dump
glucofm_encoder.onnx	ONNX opset 17	Upstream OpenCGM export (raw grid → 128-d embedding)
glucofm_encoder.onnx.meta.json	JSON	Upstream export metadata
config.json	JSON	Geometry + param counts for this package
SHA256SUMS	text	Integrity checks

Encoder params in weight files: 435,505 (OpenCGM ONNX tensors).
online/mask_token (128) is omitted upstream; glucofm inference defaults it to zeros.
Full encoder including mask_token + online/gaussian.rho: 435,633.

Paper full trainable target (encoder + JEPA heads): 732,593.

Geometry

Symbol	Value
Grid L	288 (5 min)
Patches P	24
Patch len K	12
Hidden D	128
Stream S	64

Provenance

• Upstream community encoder: sfourdrinier/opencgm-stateevent
• OpenCGM project: sfourdrinier/opencgm
• Paper: arXiv:2605.30865
• Conversion via glucofm convert_weights --all (ONNX → safetensors / GGUF / RLXP / .bin)

Upstream ONNX meta (epoch 40 seed 17, raw stats, circadian, learnable σ) is kept in glucofm_encoder.onnx.meta.json.

Use with glucofm

cargo run --release --features infer --example infer -- \
  --weights /path/to/glucofm_encoder.safetensors

# RLXP (same feature set)
cargo run --release --features infer --example infer -- \
  --weights /path/to/glucofm_encoder.rlxp

cargo run --release --features hub --example train_from_hub -- \
  --weights /path/to/glucofm_encoder.safetensors --steps 50

use glucofm::prelude::*;

let mut model = GlucoFm::for_inference(GlucoFmConfig::default(), InferOptions::default(), &mut rng)?;
model.load_weights_path("glucofm_encoder.rlxp")?; // or .safetensors / .gguf / .bin
let emb = model.encode_embedding_day(&day, &mut rng)?; // [128]

Formats

• Prefer safetensors or RLXP F32 for numeric fidelity with RLX Param graphs.
• GGUF / RLXP F16 are for disk savings; runtime still uses dense F32/F16 params after load (no in-graph Q8 matmul).

License

Weights and packaging in this repository: MIT (see LICENSE).

Upstream OpenCGM / paper licenses may impose additional terms on the original checkpoints — review those projects before commercial use.