Chapter 63
Validation report — full suite
github.com/openfluke/w2a✅ 137k cells
Why it exists
Claims are cheap; a green matrix is not. This is the actual output of one full w2a run so the book's ✅ marks are backed by numbers you can reproduce, not asserted.
What it is
Every timed layer sweeps its dtype × format × backend matrix; every suite runs its case checks. The run below is zero-gap and zero-fail across the whole board.
Full suite: PASS137,039 matrix cells — OK 137,039 · GAP 0 · FAIL 0. 326 suite cases — PASS 326 · FAIL 0. Elapsed 24m09s on amd64 with an NVIDIA GTX 1650 SUPER.
Coverage by layer (timed matrices)
| Layer | Cells | OK | Cases |
|---|---|---|---|
| dense | 2,802 | 2,802 | 18 |
| mha | 2,847 | 2,847 | 18 |
| softmax | 2,847 | 2,847 | 19 |
| cnn1 · cnn2 · cnn3 | 2,847 ea | 2,847 ea | 14 ea |
| embedding · layernorm · lstm | 2,847 ea | 2,847 ea | 14 ea |
| residual · rmsnorm · rnn | 2,847 ea | 2,847 ea | 14 ea |
| sequential · swiglu | 2,847 ea | 2,847 ea | 14 ea |
| dna | 16,159 | 16,159 | 6 |
| evolution | 16,152 | 16,152 | 6 |
| step | 32,458 | 32,458 | 11 |
| tween | 32,457 | 32,457 | 10 |
| Total | 137,039 | 137,039 | 326 |
Dense timed matrix — highlights
All 34 dtypes run forward and backward on CPU-tiled, Plan 9 SIMD, and WebGPU with zero gaps. Fastest forward paths on SIMD: int8 45µs, float32 57µs, int4 87µs; WebGPU stays in the ~165–490µs band across every dtype.
Go example
examples/63-validation/main.go
Run:
cd welvet/examples/63-validation && source ../env.sh && go run .package main
import "fmt"
func main() {
cells, cases := 137039, 326
fmt.Printf("w2a suite: %d/%d matrix cells OK, %d/%d cases PASS\n", cells, cells, cases, cases)
fmt.Println("gaps: 0 fails: 0 result: PASS")
}
Output
Validate (harness)
cd w2a && go run . # full timed suite (~24m); writes logs/suite.txt