mirror of
https://github.com/ollama/ollama.git
synced 2026-09-21 05:28:00 -05:00
The MLX runner is the only Go inference runner left and is no longer experimental, so its packages leave x/. The bindings become a top-level mlx package beside the carried patches in mlx/compat, mirroring how llama/ holds the llama.cpp integration, and the runner becomes mlxrunner with the architectures nested under the package they implement. Subpackages move with their parent unless listed. x/mlxrunner/mlx mlx x/internal/mlxthread mlx/mlxthread x/internal/mlxthreadtest mlx/mlxthread/mlxthreadtest x/internal/mlxtest mlx/mlxtest x/quant mlx/quant mlx/compat/*.patch mlx/compat/mlx-c (MLX patches go in mlx/compat/mlx) x/mlxrunner mlxrunner x/models/nn mlxrunner/nn x/models/<arch> mlxrunner/model/<arch> x/mlxrunner/imports.go mlxrunner/model/architectures (new package) x/create create x/safetensors fs/safetensors x/tokenizer mlxrunner/tokenizer Every package keeps its name, so the Go changes are the import path rewrites the moves force, and the CMake, Dockerfile, CI cache keys, drift check and Darwin payload script follow the new paths. Four edits are not paths: the runner's blank architecture imports become the package mlxrunner/model/architectures, so the list to extend for a new model sits beside the architecture directories; a depguard rule keeps the two test harnesses out of non-test code, as the x/internal placement used to; the CI change filter's two entries for the long-deleted x/imagegen/mlx now name the bindings' CMake project and the carried patches, so a change to either builds the payload; and the tokenizer parity test reads its fixtures from its own testdata instead of walking out of x/. x/server and x/imagegen/manifest stay for the next two commits.
178 lines
4.9 KiB
Go
178 lines
4.9 KiB
Go
package mlx
|
|
|
|
// #include <stdlib.h>
|
|
// #include "generated.h"
|
|
//
|
|
// extern int closureCallback(mlx_vector_array* res, mlx_vector_array input, void* payload);
|
|
// extern void closureDestructor(void* payload);
|
|
import "C"
|
|
|
|
import (
|
|
"log/slog"
|
|
"runtime/cgo"
|
|
"sync"
|
|
"unsafe"
|
|
)
|
|
|
|
// CompileFunc is the signature of a function that can be compiled.
|
|
type CompileFunc func(inputs ...*Array) []*Array
|
|
|
|
// CompileOption configures Compile behavior.
|
|
type CompileOption func(*compileConfig)
|
|
|
|
type compileConfig struct {
|
|
shapeless bool
|
|
}
|
|
|
|
// Shapeless traces the function once against symbolic shapes so the compiled
|
|
// graph accepts any input shape afterwards. Without this option, MLX re-traces
|
|
// on each new (shape, dtype) combination and caches each specialization.
|
|
func Shapeless() CompileOption {
|
|
return func(c *compileConfig) { c.shapeless = true }
|
|
}
|
|
|
|
// Compile returns a compiled version of fn. When called during another
|
|
// compile's trace, fn is inlined directly so outer compiles can fuse through
|
|
// inner ones.
|
|
//
|
|
// Compiled functions must not have side effects outside of the function. Do
|
|
// not access data other than the arguments passed in (either Go data or MLX
|
|
// arrays) unless it is a constant.
|
|
func Compile(name string, fn CompileFunc, opts ...CompileOption) CompileFunc {
|
|
var cfg compileConfig
|
|
for _, o := range opts {
|
|
o(&cfg)
|
|
}
|
|
|
|
var closure C.mlx_closure
|
|
var once sync.Once
|
|
|
|
return func(inputs ...*Array) []*Array {
|
|
if tracing {
|
|
return fn(inputs...)
|
|
}
|
|
|
|
once.Do(func() {
|
|
payload := (*cgo.Handle)(C.malloc(C.size_t(unsafe.Sizeof(cgo.Handle(0)))))
|
|
*payload = cgo.NewHandle(fn)
|
|
src := mlxCheck(C.mlx_closure_new_func_payload(
|
|
(*[0]byte)(C.closureCallback),
|
|
unsafe.Pointer(payload),
|
|
(*[0]byte)(C.closureDestructor),
|
|
))
|
|
defer freeClosure(src)
|
|
|
|
closure = mlxCheck(C.mlx_closure_new())
|
|
mlxCheck(C.mlx_compile(&closure, src, C.bool(cfg.shapeless)))
|
|
})
|
|
|
|
inVec := mlxCheck(C.mlx_vector_array_new())
|
|
defer freeVectorArray(inVec)
|
|
for _, in := range inputs {
|
|
mlxCheck(C.mlx_vector_array_append_value(inVec, in.ctx))
|
|
}
|
|
|
|
outVec := mlxCheck(C.mlx_vector_array_new())
|
|
defer freeVectorArray(outVec)
|
|
mlxCheck(C.mlx_closure_apply(&outVec, closure, inVec))
|
|
|
|
n := int(mlxCheck(C.mlx_vector_array_size(outVec)))
|
|
outputs := make([]*Array, n)
|
|
for i := range n {
|
|
outputs[i] = New(name)
|
|
mlxCheck(C.mlx_vector_array_get(&outputs[i].ctx, outVec, C.size_t(i)))
|
|
}
|
|
return outputs
|
|
}
|
|
}
|
|
|
|
// Compile1 compiles a unary function. See Compile.
|
|
func Compile1(name string, fn func(*Array) *Array, opts ...CompileOption) func(*Array) *Array {
|
|
cf := Compile(name, func(in ...*Array) []*Array {
|
|
return []*Array{fn(in[0])}
|
|
}, opts...)
|
|
return func(a *Array) *Array {
|
|
return cf(a)[0]
|
|
}
|
|
}
|
|
|
|
// Compile2 compiles a binary function. See Compile.
|
|
func Compile2(name string, fn func(*Array, *Array) *Array, opts ...CompileOption) func(*Array, *Array) *Array {
|
|
cf := Compile(name, func(in ...*Array) []*Array {
|
|
return []*Array{fn(in[0], in[1])}
|
|
}, opts...)
|
|
return func(a, b *Array) *Array {
|
|
return cf(a, b)[0]
|
|
}
|
|
}
|
|
|
|
// Compile3 compiles a ternary function. See Compile.
|
|
func Compile3(name string, fn func(*Array, *Array, *Array) *Array, opts ...CompileOption) func(*Array, *Array, *Array) *Array {
|
|
cf := Compile(name, func(in ...*Array) []*Array {
|
|
return []*Array{fn(in[0], in[1], in[2])}
|
|
}, opts...)
|
|
return func(a, b, c *Array) *Array {
|
|
return cf(a, b, c)[0]
|
|
}
|
|
}
|
|
|
|
// tracing is true while a compile callback is running. Since MLX is
|
|
// single-threaded at this level a plain Go bool suffices.
|
|
var tracing bool
|
|
|
|
//export closureCallback
|
|
func closureCallback(res *C.mlx_vector_array, input C.mlx_vector_array, payload unsafe.Pointer) (rc C.int) {
|
|
defer func() {
|
|
if r := recover(); r != nil {
|
|
slog.Error("mlx closure callback panicked", "panic", r)
|
|
rc = 1
|
|
}
|
|
}()
|
|
|
|
handle := *(*cgo.Handle)(payload)
|
|
fn := handle.Value().(CompileFunc)
|
|
|
|
// The trace runs in its own scope so its intermediates are freed as a
|
|
// group when the callback returns. This gives the effect of a single op:
|
|
// inputs are owned by the original caller (via the MLX layer) and outputs
|
|
// are transferred back to MLX to create a new Go side tensor.
|
|
if tracing {
|
|
panic("mlx: nested compile trace")
|
|
}
|
|
tracing = true
|
|
s := enterScope()
|
|
s.noEscape = true
|
|
defer func() {
|
|
tracing = false
|
|
exitScope(s)
|
|
}()
|
|
|
|
n := int(mlxCheck(C.mlx_vector_array_size(input)))
|
|
inputs := make([]*Array, n)
|
|
for i := range n {
|
|
a := New("")
|
|
mlxCheck(C.mlx_vector_array_get(&a.ctx, input, C.size_t(i)))
|
|
inputs[i] = a
|
|
}
|
|
|
|
outputs := fn(inputs...)
|
|
|
|
var arrPtr *C.mlx_array
|
|
if len(outputs) > 0 {
|
|
handles := make([]C.mlx_array, len(outputs))
|
|
for i, out := range outputs {
|
|
handles[i] = out.ctx
|
|
}
|
|
arrPtr = &handles[0]
|
|
}
|
|
mlxCheck(C.mlx_vector_array_set_data(res, arrPtr, C.size_t(len(outputs))))
|
|
return 0
|
|
}
|
|
|
|
//export closureDestructor
|
|
func closureDestructor(payload unsafe.Pointer) {
|
|
handle := *(*cgo.Handle)(payload)
|
|
handle.Delete()
|
|
C.free(payload)
|
|
}
|