makarna/pkg/backend/cuda/cuda_stub.go

//go:build !cuda

package cuda

import (
	"errors"
	"unsafe"

	"makarna/pkg/tensor"
)

var ErrCUDANotAvailable = errors.New("CUDA support not compiled in - build with -tags=cuda")

// MemoryInfo returns (total, free) bytes for the current CUDA device.
// In non-CUDA builds this always returns ErrCUDANotAvailable.
func MemoryInfo() (total uint64, free uint64, err error) {
	return 0, 0, ErrCUDANotAvailable
}

func MemoryInfoDevice(gpu int) (total uint64, free uint64, err error) {
	return 0, 0, ErrCUDANotAvailable
}

func DeviceCount() (int, error) {
	return 0, ErrCUDANotAvailable
}

// Tensor is a stub when CUDA is not available
type Tensor struct {
	shape tensor.Shape
	dtype tensor.DType
	gpu   int
}

func NewTensor(shape tensor.Shape, dtype tensor.DType, gpu int) (*Tensor, error) {
	return nil, ErrCUDANotAvailable
}

func (t *Tensor) Shape() tensor.Shape       { return nil }
func (t *Tensor) DType() tensor.DType       { return 0 }
func (t *Tensor) Device() tensor.DeviceType { return tensor.CPU }
func (t *Tensor) GPU() int                  { return -1 }
func (t *Tensor) Placement() tensor.DevicePlacement {
	return tensor.DevicePlacement{Type: tensor.CPU, GPU: -1}
}
func (t *Tensor) Data() interface{}                                 { return nil }
func (t *Tensor) Free()                                             {}
func (t *Tensor) Add(other tensor.Tensor) error                     { return ErrCUDANotAvailable }
func (t *Tensor) Mul(other tensor.Tensor) error                     { return ErrCUDANotAvailable }
func (t *Tensor) MatMul(other, out tensor.Tensor) error             { return ErrCUDANotAvailable }
func (t *Tensor) Reshape(shape tensor.Shape) (tensor.Tensor, error) { return nil, ErrCUDANotAvailable }
func (t *Tensor) View(shape tensor.Shape) (tensor.Tensor, error)    { return nil, ErrCUDANotAvailable }
func (t *Tensor) ViewAt(shape tensor.Shape, offsetBytes uintptr) (*Tensor, error) {
	return nil, ErrCUDANotAvailable
}
func (t *Tensor) ToDevice(device tensor.DeviceType) (tensor.Tensor, error) {
	return nil, ErrCUDANotAvailable
}
func (t *Tensor) CopyFrom(data interface{}) error                    { return ErrCUDANotAvailable }
func (t *Tensor) CopyToHost(dst []float32) error                     { return ErrCUDANotAvailable }
func (t *Tensor) CopyPartialFrom(dstOffset int, src []float32) error { return ErrCUDANotAvailable }
func (t *Tensor) CopyPartialFromDevice(dstOffset int, src *Tensor, srcOffset int, length int) error {
	return ErrCUDANotAvailable
}

func MemcpyH2D(dst, src unsafe.Pointer, size uintptr, gpu int) error { return ErrCUDANotAvailable }

func MemcpyD2H(dst, src unsafe.Pointer, size uintptr, gpu int) error { return ErrCUDANotAvailable }

func MemcpyD2D(dst, src unsafe.Pointer, size uintptr, gpu int) error { return ErrCUDANotAvailable }

func CastF32ToF16(srcF32, dstF16 unsafe.Pointer, n int, gpu int) error { return ErrCUDANotAvailable }

func KDACausalShortConv1D(x, state, w unsafe.Pointer, tokens, projSize, kernel int, gpu int) error {
	return ErrCUDANotAvailable
}

func L2NormHeads(q, k unsafe.Pointer, tokens, numHeads, headDim int, eps float32, gpu int) error {
	return ErrCUDANotAvailable
}

func KDAGate(g, aLog, dtBias, out unsafe.Pointer, tokens, numHeads, headDim int, gpu int) error {
	return ErrCUDANotAvailable
}

func KDARecurrent(q, k, v, g, beta, state unsafe.Pointer, tokens, numHeads, headDim int, gpu int) error {
	return ErrCUDANotAvailable
}

func RMSNormGated(out, g, weight unsafe.Pointer, n, headDim int, eps float32, gpu int) error {
	return ErrCUDANotAvailable
}

func Sigmoid(x unsafe.Pointer, n int, gpu int) error {
	return ErrCUDANotAvailable
}

func SoftmaxRows(x unsafe.Pointer, rows, cols int, gpu int) error {
	return ErrCUDANotAvailable
}

func TopKPerRow(scores unsafe.Pointer, indices unsafe.Pointer, values unsafe.Pointer, rows, cols, k int, gpu int) error {
	return ErrCUDANotAvailable
}

func PagedAttention(Q, kBlocksDev, vBlocksDev, out unsafe.Pointer, seqLen, kvLen, numHeads, numKVHeads, headDim, blockSize int, scale float32, startPos int, gpu int) error {
	return ErrCUDANotAvailable
}

func PagedAttentionBatch(Q, kBlocksFlatDev, vBlocksFlatDev, blockOffsetsDev, kvLensDev, queryPosDev, out unsafe.Pointer, numTokens, numHeads, numKVHeads, headDim, blockSize int, scale float32, maxKvLen int, gpu int) error {
	return ErrCUDANotAvailable
}

func PagedAttentionF32F16KV(Q, kBlocksDev, vBlocksDev, out unsafe.Pointer, seqLen, kvLen, numHeads, numKVHeads, headDim, blockSize int, scale float32, startPos int, gpu int) error {
	return ErrCUDANotAvailable
}

func PagedAttentionBatchF32F16KV(Q, kBlocksFlatDev, vBlocksFlatDev, blockOffsetsDev, kvLensDev, queryPosDev, out unsafe.Pointer, numTokens, numHeads, numKVHeads, headDim, blockSize int, scale float32, maxKvLen int, gpu int) error {
	return ErrCUDANotAvailable
}

func PagedAttentionRoPEF32F16KV(Q, kBlocksDev, vBlocksDev, out unsafe.Pointer, seqLen, kvLen, numHeads, numKVHeads, headDim, blockSize int, scale float32, startPos int, theta float32, gpu int) error {
	return ErrCUDANotAvailable
}

func PagedAttentionBatchRoPEF32F16KV(Q, kBlocksFlatDev, vBlocksFlatDev, blockOffsetsDev, kvLensDev, queryPosDev, out unsafe.Pointer, numTokens, numHeads, numKVHeads, headDim, blockSize int, scale float32, maxKvLen int, theta float32, gpu int) error {
	return ErrCUDANotAvailable
}

func AllocAndCopyInt32(data []int32, gpu int) (unsafe.Pointer, error) {
	return nil, ErrCUDANotAvailable
}

func TopKLogitsF32(logits unsafe.Pointer, vocab int, repIDs []int32, repPenalty float32, k int, gpu int) ([]int32, []float32, int, error) {
	return nil, nil, 0, ErrCUDANotAvailable
}

func DequantQ8K(blocks unsafe.Pointer, out unsafe.Pointer, numBlocks int, gpu int) error {
	return ErrCUDANotAvailable
}

func DequantQ4K(blocks unsafe.Pointer, out unsafe.Pointer, numBlocks int, gpu int) error {
	return ErrCUDANotAvailable
}

func DequantQ5K(blocks unsafe.Pointer, out unsafe.Pointer, numBlocks int, gpu int) error {
	return ErrCUDANotAvailable
}

func DequantQ6K(blocks unsafe.Pointer, out unsafe.Pointer, numBlocks int, gpu int) error {
	return ErrCUDANotAvailable
}

func DequantQ3K(blocks unsafe.Pointer, out unsafe.Pointer, numBlocks int, gpu int) error {
	return ErrCUDANotAvailable
}

func DequantQ2K(blocks unsafe.Pointer, out unsafe.Pointer, numBlocks int, gpu int) error {
	return ErrCUDANotAvailable
}

func MatMulQ2K(aPtr, bPtr, cPtr unsafe.Pointer, m, k, n int, gpu int) error {
	return ErrCUDANotAvailable
}

func MatMulQ4K(aPtr, bPtr, cPtr unsafe.Pointer, m, k, n int, gpu int) error {
	return ErrCUDANotAvailable
}

func MatMulQ5K(aPtr, bPtr, cPtr unsafe.Pointer, m, k, n int, gpu int) error {
	return ErrCUDANotAvailable
}

func MatMulQ3K(aPtr, bPtr, cPtr unsafe.Pointer, m, k, n int, gpu int) error {
	return ErrCUDANotAvailable
}

func MatMulQ6K(aPtr, bPtr, cPtr unsafe.Pointer, m, k, n int, gpu int) error {
	return ErrCUDANotAvailable
}

func MatMulQ8K(aPtr, bPtr, cPtr unsafe.Pointer, m, k, n int, gpu int) error {
	return ErrCUDANotAvailable
}

func MatMulF16Q8K(aPtr, bPtr, cPtr unsafe.Pointer, m, k, n int, gpu int) error {
	return ErrCUDANotAvailable
}

func MatMulF16Q4K(aPtr, bPtr, cPtr unsafe.Pointer, m, k, n int, gpu int) error {
	return ErrCUDANotAvailable
}

func MatMulF16Q5K(aPtr, bPtr, cPtr unsafe.Pointer, m, k, n int, gpu int) error {
	return ErrCUDANotAvailable
}

func MatMulF16Q2K(aPtr, bPtr, cPtr unsafe.Pointer, m, k, n int, gpu int) error {
	return ErrCUDANotAvailable
}

func MatMulF16Q3K(aPtr, bPtr, cPtr unsafe.Pointer, m, k, n int, gpu int) error {
	return ErrCUDANotAvailable
}

func MatMulF16Q6K(aPtr, bPtr, cPtr unsafe.Pointer, m, k, n int, gpu int) error {
	return ErrCUDANotAvailable
}

func FreeDevicePtr(ptr unsafe.Pointer) {}

func Free(ptr unsafe.Pointer) {}

func AllocAndCopyPtrTable(ptrs []uintptr, gpu int) (unsafe.Pointer, error) {
	return nil, ErrCUDANotAvailable
}

// Available returns whether CUDA is available
func Available() bool {
	return false
}