rasptorch
Operate rasptorch (NumPy deep learning library) via its agent-native CLI, consuming JSON output. Supports tensors, model creation (mlp, cnn, cv, jev, gru, transformer, linear), training, save/load, and UI/chat modes.
rasptorch skill
Use this skill to operate rasptorch from any agent harness by calling its agent-native CLI and consuming structured JSON output.
Interface contract
| Field | Value |
|---|---|
| Transport | CLI (stdout/stderr) |
| Output format | JSON (pass --json on every call) |
| Stateful | Yes — model registry persists within a session |
| Long-running | rasptorch ui (Streamlit), rasptorch chat (REPL) |
How to invoke
Prefer the installed console script:
rasptorch --json <command> [args...]
If the console script is unavailable, use module invocation from this repo:
python -m rasptorch.CLI.cli --json <command> [args...]
Backend selection
Always prefer --backend auto unless the user asks to pin a specific backend:
--backend auto(default): choose best available backend--backend vulkan: prefer Vulkan GPU backend--backend opencl: use OpenCL backend--backend cuda: use CUDA backend (CuPy or PyTorch CUDA provider)--backend numpy: CPU backend (internal key:cpu)
Read-only checks
Environment + GPU status
rasptorch --json info
Key fields: vulkan_available, vulkan_using_real_gpu, vulkan_status.
Tensor workflows
Create tensors
rasptorch --json --backend numpy tensor random --shape 2,3,4 --device cpu --dtype float32
rasptorch --json tensor zeros --shape 3,4 --device auto
rasptorch --json --backend vulkan tensor ones --shape 5,10 --device gpu
Notes:
- Keep shapes small unless the user explicitly requests large workloads.
- Prefer
--backend autounless the user asks to pin a specific backend.
Model workflows (stateful)
Model commands keep a session-like registry of created models.
Create models
# Linear (single layer)
rasptorch --json model linear --input-size 10 --hidden-sizes 32,16 --output-size 2
# MLP (multi-layer perceptron)
rasptorch --json model mlp --layers 64,32,16,2
# CNN (convolutional network)
rasptorch --json model cnn --in-channels 3 --out-channels 16,32
# CV (computer vision: Conv2d blocks -> Flatten -> Linear)
rasptorch --json model cv --in-channels 3 --out-channels 32,64 --num-classes 10 --input-hw 32
# JEV (decision: MLP outputting a decision vector)
rasptorch --json model jev --input-size 10 --hidden-sizes 32,16 --num-decisions 2
# GRU (gated recurrent unit)
rasptorch --json model gru --input-size 8 --hidden-size 32 --num-layers 1
# Transformer
rasptorch --json model transformer --vocab-size 32000 --d-model 256 --num-heads 8 --num-layers 4
List / remove models
rasptorch --json model list
rasptorch --json model remove --model-id <model-id>
Save / load (writes files)
Only do this when the user requests it (or approves writing files):
rasptorch --json model save --model-id <model-id> --path ./model.pkl
rasptorch --json model load --path ./model.pkl
Train (writes compute + may take time)
Only run training after user approval:
rasptorch --json model train --model-id <model-id> --epochs 5 --lr 0.001 --batch-size 32 --device auto --optimizer Adam
UI / chat modes
rasptorch uistarts a Streamlit server (long-running). Only launch with user approval.rasptorch chatstarts an interactive REPL. Avoid in automated tool flows.
Error handling
- Non-zero exit code indicates failure; inspect stderr for details.
- JSON output is emitted on stdout; errors may be returned as JSON with an
errorfield. - Always check the exit code before parsing stdout as JSON.
Capability discovery
Run rasptorch --json info to discover available backends, GPU status, and supported commands at runtime.
Related skills
- Downloads
- 909All time
- Security audit
- Pass
- Last updated
- 1w ago
- Current version
- v2.1.0
- License
- MIT-0
