RFC-003: Neural ODE via TorchScript

This RFC proposes adding a PyTorch-based Neural ODE prediction backend. Models are trained in Python (PyTorch) or Rust (tch-rs), exported as TorchScript (.pt), and loaded for inference. Gated behind an optional torch-backend feature flag.

Summary Table

Phase	Scope	Effort	Status
1	Feature flags & dependencies (tch 0.23)	Low	✅ Done
2	Configuration (model_path in TOML)	Low	✅ Done
3	TorchScript loader + inference (torch_ode.rs)	Medium	✅ Done
4	Query engine integration (fallback to linear)	Medium	✅ Done
5	Python training script (PyTorch)	Low	✅ Done
6	Rust training API (tch-rs)	Medium	✅ Done
7	Python bindings update (PyO3)	Low	✅ Done

Why tch-rs over burn

Factor	tch-rs (libtorch)	burn
Ecosystem	Full PyTorch compat	Rust-native
Pre-trained models	Any .pt from HuggingFace	Requires re-impl
Training	Python + Rust	Rust only
Community	Millions of users	Growing

Architecture

Training (Python/Rust)          Inference (Rust)
─────────────────────           ────────────────────
Train NeuralODEPredictor  →     TorchOdeModel::load()
Export TorchScript .pt    →     model.predict(traj, t)
                                Fallback → linear_extrapolate()

Model I/O Contract

Tensor	Shape	Description
Input trajectory	[1, T, D+1]	normalized_time + vector per step
Input target_t	[1, 1]	normalized target timestamp
Output predicted	[1, D]	predicted vector

Configuration

[analytics]
neural_ode = true
model_path = "./models/neural_ode_d128.pt"

Error Handling

Every failure falls back to linear_extrapolate():

Scenario	Behavior
torch-backend not compiled	Linear fallback (zero overhead)
Model file missing	Warning + linear fallback
Inference fails (NaN, bad dims)	Warning + linear fallback per request
CUDA unavailable	Automatic CPU fallback

Pre-trained Models

No standard model exists for high-dim trajectory prediction. Strategy:

1. Now: Ship without default model
2. Future: Publish manucouto1/cvx-neural-ode-{dim} on HuggingFace

References

1. Chen et al. “Neural ODEs.” NeurIPS 2018.
2. Rubanova et al. “Latent ODEs for Irregularly-Sampled Time Series.” NeurIPS 2019.
3. De Brouwer et al. “GRU-ODE-Bayes.” NeurIPS 2019.
4. Kidger. “On Neural Differential Equations.” PhD Thesis, Oxford, 2022.
5. tch-rs: https://github.com/LaurentMazare/tch-rs

See full RFC in design/CVX_RFC_003_NeuralODE_TorchBackend.md.