mirror of
https://github.com/ollama/ollama.git
synced 2026-09-19 12:24:30 -04:00
Jesse Gross
2e036e7cdf
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.
75 lines
2.6 KiB
Go
75 lines
2.6 KiB
Go
package create
|
|
|
|
import (
|
|
"bytes"
|
|
"fmt"
|
|
"io"
|
|
|
|
"github.com/ollama/ollama/fs/safetensors"
|
|
)
|
|
|
|
// applyByteTransform produces a TensorSpec's output tensor from its resolved
|
|
// source tensors using only byte-level (non-MLX) operations. The MLX transform
|
|
// (decode_fp8) and quantization are handled separately by the MLX writer path.
|
|
func applyByteTransform(ts TensorSpec, sources []*safetensors.TensorData) (*safetensors.TensorData, error) {
|
|
switch ts.Transform {
|
|
case TransformNone:
|
|
if len(sources) != 1 {
|
|
return nil, fmt.Errorf("transform none expects 1 source, got %d", len(sources))
|
|
}
|
|
return sources[0].WithName(ts.Name), nil
|
|
|
|
case TransformRepackFP4, TransformRelabelU8:
|
|
// Both relabel the header (dtype, and for the fp4 repack the last
|
|
// dimension); the bytes are unchanged, so the reader is reused.
|
|
if len(sources) != 1 {
|
|
return nil, fmt.Errorf("transform %s expects 1 source, got %d", ts.Transform, len(sources))
|
|
}
|
|
td := sources[0].WithName(ts.Name)
|
|
if ts.OutDtype != "" {
|
|
td.Dtype = ts.OutDtype
|
|
}
|
|
if ts.OutShape != nil {
|
|
td.Shape = append([]int32(nil), ts.OutShape...)
|
|
}
|
|
return td, nil
|
|
|
|
case TransformScalarF32:
|
|
if len(sources) != 1 {
|
|
return nil, fmt.Errorf("transform scalar_f32 expects 1 source, got %d", len(sources))
|
|
}
|
|
return validateScalarFloat32TensorData(sources[0], ts.Name)
|
|
|
|
case TransformReciprocalF32:
|
|
if len(sources) != 1 {
|
|
return nil, fmt.Errorf("transform reciprocal_f32 expects 1 source, got %d", len(sources))
|
|
}
|
|
return invertScalarFloat32TensorData(sources[0], ts.Name)
|
|
|
|
case TransformStackExperts:
|
|
return stackExpertTensors(ts.Name, ts.OutDtype, ts.OutShape, sources)
|
|
|
|
default:
|
|
return nil, fmt.Errorf("transform %q requires the MLX writer path", ts.Transform)
|
|
}
|
|
}
|
|
|
|
// stackExpertTensors concatenates per-expert tensors (in the given order) into
|
|
// one [experts, ...] tensor. Row-major layout means the stacked bytes are
|
|
// exactly the per-expert byte blocks back to back.
|
|
func stackExpertTensors(name, dtype string, shape []int32, sources []*safetensors.TensorData) (*safetensors.TensorData, error) {
|
|
if len(sources) == 0 {
|
|
return nil, fmt.Errorf("stack_experts expects at least one source")
|
|
}
|
|
var buf bytes.Buffer
|
|
for i, s := range sources {
|
|
if s.Dtype != sources[0].Dtype {
|
|
return nil, fmt.Errorf("stack_experts source %d dtype %s != %s", i, s.Dtype, sources[0].Dtype)
|
|
}
|
|
if _, err := io.Copy(&buf, s.Reader()); err != nil {
|
|
return nil, fmt.Errorf("stack_experts read source %d (%s): %w", i, s.Name, err)
|
|
}
|
|
}
|
|
return safetensors.NewTensorDataFromBytes(name, dtype, append([]int32(nil), shape...), buf.Bytes()), nil
|
|
}
|