3. core — types & backends

Part: II · Foundation
Package: github.com/openfluke/welvet/core
Status: ok — ✅

When

You need the core foundation package.

Where

import "github.com/openfluke/welvet/core"

cd 03-core && source ../env.sh && go run .

Why

Every polymorphic path needs one place for DType, LayerType, Activation, Backend, Tensor[T], and slim Layer metadata — without QAT morph defaults.

What

34 storage dtypes, Numeric generics, Tensor[T], ExecConfig (BackendCPUTiled | BackendSIMD | BackendWebGPU), Activate/ActivateDeriv, and converters for f16/bf16/fp8/fp4.

Sample output (captured)

simd bfloat16 8

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/03-core).

package main

import (
	"fmt"

	"github.com/openfluke/welvet/core"
)

func main() {
	t := core.NewTensor[float32](2, 4)
	t.Data[0] = -1
	t.Data[0] = core.Activate(t.Data[0], core.ActivationReLU) // 0

	cfg := core.ExecConfig{
		Backend:   core.BackendSIMD,
		MultiCore: true,
		TileSize:  32,
	}
	fmt.Println(cfg.Backend.String(), core.ParseDType("bfloat16"), t.Len())
}