31. runtime/training

Part: IV · Runtime
Package: github.com/openfluke/welvet/runtime/training
Status: ok — ✅

When

Walking a Grid/Stack with the shared runtime/training path.

Where

import "github.com/openfluke/welvet/runtime/training"

cd 31-training && source ../env.sh && go run .

Why

Suites and small nets need MSE+SGD and tween hooks without inventing an external trainer or a retained float32 master beside storage.

What

MSE/MSEGrad, SGD, Step, ApplyTween/StepTween, StepMesh. Layer-agnostic ApplyGradSGD dispatch (Dense…Mamba/GDN/…). FormatNone: in-dtype ApplySGD; packed: unpack→update→re-Pack. No QAT dual path — storage dtype/format is truth after every step. Sandwich credit (Split / FastProxy / Sparse / …) lives in layers/parallel TrainMode — see §67.

Sample output (captured)

0 <nil>

Live capture from go run ./cmd/runall on local ../../welvet (exit 0).

Source

Copied from the Welvet feature book examples (openfluke.github.io/welvet/examples/31-training).

main.go

Download source ↓

package main

import (
	"fmt"

	"github.com/openfluke/welvet/architecture"
	"github.com/openfluke/welvet/core"
	"github.com/openfluke/welvet/layers/dense"
	"github.com/openfluke/welvet/runtime/forward"
	"github.com/openfluke/welvet/runtime/training"
)

func main() {
	g := architecture.NewGrid(1, 1, 1, 1)
	l, _ := dense.New(4, 4, core.ActivationLinear, core.DTypeFloat32)
	_ = dense.Place(g, 0, 0, 0, 0, l)
	fwd, _ := forward.Forward(g, core.NewTensor[float32](1, 4))
	loss, err := training.Step(fwd, core.NewTensor[float32](1, 4), 1e-2)
	fmt.Println(loss, err)
}