mirror of
https://github.com/ollama/ollama.git
synced 2026-04-17 15:53:27 +02:00
* prefer rocm v6 on windows Avoid building with v7 - more changes are needed * MLX: add header vendoring and remove go build tag This switches to using a vendoring approach for the mlx-c headers so that Go can build without requiring a cmake first. This enables building the new MLX based code by default. Every time cmake runs, the headers are refreshed, so we can easily keep them in sync when we bump mlx versions. Basic Windows and Linux support are verified. * ci: harden for flaky choco repo servers CI sometimes fails due to choco not actually installing cache. Since it just speeds up the build, we can proceed without. * review comments
160 lines
3.7 KiB
Go
160 lines
3.7 KiB
Go
package cache
|
|
|
|
import "github.com/ollama/ollama/x/mlxrunner/mlx"
|
|
|
|
// RecurrentCache stores state for linear-recurrent layers.
|
|
//
|
|
// Conv state shape: [B, convTail, convDim]
|
|
// Delta state shape: [B, numVHeads, headVDim, headKDim]
|
|
type RecurrentCache struct {
|
|
convState *mlx.Array
|
|
deltaState *mlx.Array
|
|
offset int
|
|
|
|
convTail int
|
|
convDim int
|
|
numVHeads int
|
|
headVDim int
|
|
headKDim int
|
|
}
|
|
|
|
func (c *RecurrentCache) setStateRaw(old, v *mlx.Array) *mlx.Array {
|
|
if v == nil || !v.Valid() {
|
|
return old
|
|
}
|
|
if old == v {
|
|
return old
|
|
}
|
|
|
|
mlx.Pin(v)
|
|
if old != nil && old != v {
|
|
mlx.Unpin(old)
|
|
}
|
|
|
|
return v
|
|
}
|
|
|
|
func (c *RecurrentCache) setStateDetached(old, v *mlx.Array, ensureContiguous bool) *mlx.Array {
|
|
if v == nil || !v.Valid() {
|
|
return old
|
|
}
|
|
if old == v {
|
|
return old
|
|
}
|
|
|
|
root := v
|
|
if ensureContiguous {
|
|
root = mlx.Contiguous(v, false)
|
|
}
|
|
detached := root.Clone()
|
|
|
|
mlx.Pin(detached)
|
|
if old != nil && old != detached {
|
|
mlx.Unpin(old)
|
|
}
|
|
|
|
return detached
|
|
}
|
|
|
|
func snapshotPinned(a *mlx.Array) *mlx.Array {
|
|
if a == nil || !a.Valid() {
|
|
return nil
|
|
}
|
|
snap := mlx.Copy(a)
|
|
mlx.Eval(snap)
|
|
mlx.Pin(snap)
|
|
return snap
|
|
}
|
|
|
|
func NewRecurrentCache(convTail, convDim, numVHeads, headVDim, headKDim int32) *RecurrentCache {
|
|
return &RecurrentCache{
|
|
convTail: int(convTail),
|
|
convDim: int(convDim),
|
|
numVHeads: int(numVHeads),
|
|
headVDim: int(headVDim),
|
|
headKDim: int(headKDim),
|
|
}
|
|
}
|
|
|
|
func (c *RecurrentCache) ensure(batch int, dtype mlx.DType) {
|
|
if batch <= 0 {
|
|
batch = 1
|
|
}
|
|
|
|
needConv := c.convState == nil || !c.convState.Valid() || c.convState.DType() != dtype ||
|
|
c.convState.Dim(0) != batch || c.convState.Dim(1) != c.convTail || c.convState.Dim(2) != c.convDim
|
|
needDelta := c.deltaState == nil || !c.deltaState.Valid() || c.deltaState.DType() != dtype ||
|
|
c.deltaState.Dim(0) != batch || c.deltaState.Dim(1) != c.numVHeads || c.deltaState.Dim(2) != c.headVDim || c.deltaState.Dim(3) != c.headKDim
|
|
if !needConv && !needDelta {
|
|
return
|
|
}
|
|
|
|
if needConv {
|
|
c.convState = c.setStateRaw(c.convState, mlx.Zeros(dtype, batch, c.convTail, c.convDim))
|
|
}
|
|
if needDelta {
|
|
c.deltaState = c.setStateRaw(c.deltaState, mlx.Zeros(dtype, batch, c.numVHeads, c.headVDim, c.headKDim))
|
|
}
|
|
}
|
|
|
|
func (c *RecurrentCache) ConvState(batch int, dtype mlx.DType) *mlx.Array {
|
|
c.ensure(batch, dtype)
|
|
return c.convState
|
|
}
|
|
|
|
func (c *RecurrentCache) SetConvState(v *mlx.Array) {
|
|
c.convState = c.setStateDetached(c.convState, v, true)
|
|
}
|
|
|
|
func (c *RecurrentCache) DeltaState(batch int, dtype mlx.DType) *mlx.Array {
|
|
c.ensure(batch, dtype)
|
|
return c.deltaState
|
|
}
|
|
|
|
func (c *RecurrentCache) SetDeltaState(v *mlx.Array) {
|
|
c.deltaState = c.setStateDetached(c.deltaState, v, false)
|
|
}
|
|
|
|
func (c *RecurrentCache) Advance(n int) {
|
|
c.offset += n
|
|
}
|
|
|
|
func (c *RecurrentCache) Update(keys, values *mlx.Array) (*mlx.Array, *mlx.Array) {
|
|
return keys, values
|
|
}
|
|
|
|
func (c *RecurrentCache) State() (*mlx.Array, *mlx.Array) {
|
|
return c.convState, c.deltaState
|
|
}
|
|
|
|
func (c *RecurrentCache) CanTrim() bool { return false }
|
|
|
|
func (c *RecurrentCache) Trim(n int) int {
|
|
// Recurrent state is not directly trimmable. Divergent prefixes must drop the cache.
|
|
_ = n
|
|
return 0
|
|
}
|
|
|
|
func (c *RecurrentCache) Clone() Cache {
|
|
clone := &RecurrentCache{
|
|
offset: c.offset,
|
|
convTail: c.convTail,
|
|
convDim: c.convDim,
|
|
numVHeads: c.numVHeads,
|
|
headVDim: c.headVDim,
|
|
headKDim: c.headKDim,
|
|
convState: snapshotPinned(c.convState),
|
|
deltaState: snapshotPinned(c.deltaState),
|
|
}
|
|
return clone
|
|
}
|
|
|
|
func (c *RecurrentCache) Free() {
|
|
mlx.Unpin(c.convState, c.deltaState)
|
|
c.convState, c.deltaState = nil, nil
|
|
c.offset = 0
|
|
}
|
|
|
|
func (c *RecurrentCache) Offset() int { return c.offset }
|
|
func (c *RecurrentCache) Len() int { return c.offset }
|