[Show & Tell] SpiralTorch — Rust-first ML stack (WebGPU/WASM/MPS/CUDA), trains in Z-space

Show and Tell
1 
# SpiralTorch — Rust-first ML stack that trains in Z-space
**TL;DR:** Pure Rust training framework with WGPU/WebGPU/MPS/CUDA/HIP backends.  
No NumPy/Torch shims. Roundtable planner (A/B/C) self-rewrites SpiralK heuristics.

- Model card: https://huggingface.co/RyoSpiralArchitect/SpiralTorch
- Repo: https://github.com/RyoSpiralArchitect/SpiralTorch

## What it is
- **Rust-first training**: `st-nn` modules (Linear/Conv1d/WaveRnn/Relu/Sequential), datasets, trainer.
- **Z-space hypergrad**: tensors can absorb text/complex waves; gradients flow on an open topos.
- **Unified planner**: SpiralK DSL (hard), SoftLogic (A), Tuner table (C); Wilson-backed consensus.
- **WGPU/WASM**: same heuristics in browser and native (zero shims, zero tracebacks).
- **Python wheel**: thin veneer over the same Rust logic.

## Hello SpiralSession (quickstart)
```python
from spiraltorch import SpiralSession, Tensor

# hello session: barycenter + hypergrad alignment + one training epoch
session = SpiralSession(device="auto", curvature=-0.95)
print(session)  # SpiralSession(device=wgpu, curvature=-0.95, ...)

input  = Tensor(1, 4, [0.1, -0.2, 0.3, -0.4])
target = Tensor(1, 2, [0.0, 1.0])

stats = session.train_epoch(input, target)
print(f"loss={stats.average_loss:.6f}, steps={stats.steps}")

Benchmarks (forward)

On my local M-series CPU vs WGPU (WebGPU path):

Input   CPU (ms)   WGPU (ms)   Speedup
128     0.46       0.14        ×3.3
256     0.73       0.22        ×3.3
512     1.28       0.38        ×3.4
1024    2.45       0.74        ×3.3

(We’d love external numbers on 4090/CUDA / RDNA/ROCm / iGPU via WebGPU.)

Why Rust / WebGPU?

  • Single source of truth: planners, ops, losses — all in Rust.
  • Native WGPU/WASM means same heuristics in browser and native.
  • No C++/Python/JS split, no glue layers, no tracebacks.

WASM demo (browser)

cd crates/st-tensor
wasm-pack build --target web --release
python3 -m http.server 8080
# open http://localhost:8080/wasm_bench/index.html

You’ll see a live Z-space spiral evolving on with WebGPU.

Call for feedback

  • API surface (Rust & Python), esp. SpiralSession ergonomics
  • Planner heuristics & SpiralK snippets you’d want to override
  • Perf reports on MPS/CUDA/HIP/WGPU (browser & native)

License: AGPL-3.0-or-later
Contact: kishkavsesvit@icloud.com (research & integration)
1 Like