MINIMIZATION PROOF
Source vs Kernel
| Metric | tinygrad/device.py (leader) | kernel.py (ours) | Notes |
|---|---|---|---|
| Total LOC | 420 | 10 | wc -l |
| Absorbed LOC | 16 (lines 39–54) | 10 | direct distillation |
| Reduction ratio (total) | — | 42.0× | 420 ÷ 10 |
| Reduction ratio (absorbed) | — | 1.6× | 16 ÷ 10 |
What was kept
| tinygrad concept | Our kernel equivalent |
|---|---|
ALL_DEVICES = [...] |
PATHS = ["CPU","GPU","QUANTIZED","MOE"] |
get_available_devices() iterator |
implicit iteration over PATHS dict |
next(self.get_available_devices()) first-available |
min(costs, key=costs.__getitem__) min-energy |
| No energy model (tinygrad picks first live device) | NINA Butler-Volmer nina(η) supplies energy cost per path |
What was added (not from leader)
- NINA Butler-Volmer lambda (5 constants, 1 math expression) — domain-specific extension for energy-aware dispatch.
random.Random(seed)for deterministic η sampling fromworldsignal inputs.
Verification
$ wc -l kernel.py
10 kernel.py
Every line is either a comment/import, a constant, the NINA formula, or one of the 4 dispatch lines. No padding.