next ollama runner (#7913)

feat: add new Ollama engine using ggml through cgo

This change introduces a new way to run pretrained models. It introduces 3 high level interfaces and a bunch of smaller helper interfaces to facilitate this.

- `model.Model` defines the interface for a model architecture. Models such as `llama` and `mllama`, which are provided as examples, can implement the model's forward propagation in the `Forward` method. This method will be called to generate completions. This interface can be found in `model/model.go`
- `ml.Backend` defines the interface for a backend tensor library, in this case `ggml`. Among other things, a Backend is responsible for loading a pretrained model into hardware (GPU, CPU, etc) and providing an interface for Models to access loaded tensors. This interface can be found in `ml/backend.go`
- `ml.Tensor` defines the interface for a tensor and tensor operations

This is the first implementation of the new engine. Follow up PRs will implement more features:

- non-greedy sampling (#8410)
- integration with Ollama and KV caching (#8301)
- more model support (#9080) with more coming soon

Co-authored-by: Bruce MacDonald <brucewmacdonald@gmail.com>

server/create.go

@@ -21,8 +21,8 @@ import (
 	"github.com/ollama/ollama/convert"
 	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/format"
+	"github.com/ollama/ollama/fs/ggml"
 	"github.com/ollama/ollama/llama"
-	"github.com/ollama/ollama/llm"
 	"github.com/ollama/ollama/template"
 	"github.com/ollama/ollama/types/errtypes"
 	"github.com/ollama/ollama/types/model"
@@ -205,7 +205,7 @@ func detectModelTypeFromFiles(files map[string]string) string {
 				return ""
 			}
 
-			ct := llm.DetectGGMLType(buf)
+			ct := ggml.DetectContentType(buf)
 			if ct == "gguf" {
 				return "gguf"
 			}
@@ -271,11 +271,11 @@ func convertFromSafetensors(files map[string]string, baseLayers []*layerGGML, is
 		return nil, err
 	}
 
-	ggml, _, err := llm.DecodeGGML(bin, 0)
+	f, _, err := ggml.Decode(bin, 0)
 	if err != nil {
 		return nil, err
 	}
-	layers := []*layerGGML{{layer, ggml}}
+	layers := []*layerGGML{{layer, f}}
 
 	if !isAdapter {
 		return detectChatTemplate(layers)
@@ -283,13 +283,13 @@ func convertFromSafetensors(files map[string]string, baseLayers []*layerGGML, is
 	return layers, nil
 }
 
-func kvFromLayers(baseLayers []*layerGGML) (llm.KV, error) {
+func kvFromLayers(baseLayers []*layerGGML) (ggml.KV, error) {
 	for _, l := range baseLayers {
 		if l.GGML != nil {
 			return l.KV(), nil
 		}
 	}
-	return llm.KV{}, fmt.Errorf("no base model was found")
+	return ggml.KV{}, fmt.Errorf("no base model was found")
 }
 
 func createModel(r api.CreateRequest, name model.Name, baseLayers []*layerGGML, fn func(resp api.ProgressResponse)) (err error) {
@@ -306,7 +306,7 @@ func createModel(r api.CreateRequest, name model.Name, baseLayers []*layerGGML,
 		if layer.GGML != nil {
 			quantType := strings.ToUpper(cmp.Or(r.Quantize, r.Quantization))
 			if quantType != "" && layer.GGML.Name() == "gguf" && layer.MediaType == "application/vnd.ollama.image.model" {
-				want, err := llm.ParseFileType(quantType)
+				want, err := ggml.ParseFileType(quantType)
 				if err != nil {
 					return err
 				}
@@ -403,7 +403,7 @@ func quantizeLayer(layer *layerGGML, quantizeType string, fn func(resp api.Progr
 	ft := layer.GGML.KV().FileType()
 	fn(api.ProgressResponse{Status: fmt.Sprintf("quantizing %s model to %s", ft, quantizeType)})
 
-	want, err := llm.ParseFileType(quantizeType)
+	want, err := ggml.ParseFileType(quantizeType)
 	if err != nil {
 		return nil, err
 	}
@@ -433,13 +433,13 @@ func quantizeLayer(layer *layerGGML, quantizeType string, fn func(resp api.Progr
 		return nil, err
 	}
 
-	ggml, _, err := llm.DecodeGGML(temp, 0)
+	f, _, err := ggml.Decode(temp, 0)
 	if err != nil {
 		slog.Error(fmt.Sprintf("error decoding ggml: %s\n", err))
 		return nil, err
 	}
 
-	return &layerGGML{newLayer, ggml}, nil
+	return &layerGGML{newLayer, f}, nil
 }
 
 func ggufLayers(digest string, fn func(resp api.ProgressResponse)) ([]*layerGGML, error) {
@@ -475,7 +475,7 @@ func ggufLayers(digest string, fn func(resp api.ProgressResponse)) ([]*layerGGML
 
 	var offset int64
 	for offset < stat.Size() {
-		ggml, n, err := llm.DecodeGGML(blob, 0)
+		f, n, err := ggml.Decode(blob, 0)
 		if errors.Is(err, io.EOF) {
 			break
 		} else if err != nil {
@@ -483,9 +483,9 @@ func ggufLayers(digest string, fn func(resp api.ProgressResponse)) ([]*layerGGML
 		}
 
 		mediatype := "application/vnd.ollama.image.model"
-		if ggml.KV().Kind() == "adapter" {
+		if f.KV().Kind() == "adapter" {
 			mediatype = "application/vnd.ollama.image.adapter"
-		} else if _, ok := ggml.KV()[fmt.Sprintf("%s.vision.block_count", ggml.KV().Architecture())]; ok || ggml.KV().Kind() == "projector" {
+		} else if _, ok := f.KV()[fmt.Sprintf("%s.vision.block_count", f.KV().Architecture())]; ok || f.KV().Kind() == "projector" {
 			mediatype = "application/vnd.ollama.image.projector"
 		}
 
@@ -506,7 +506,7 @@ func ggufLayers(digest string, fn func(resp api.ProgressResponse)) ([]*layerGGML
 			}
 		}
 
-		layers = append(layers, &layerGGML{layer, ggml})
+		layers = append(layers, &layerGGML{layer, f})
 		offset = n
 	}