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>

model/mllama/model.go

@@ -0,0 +1,99 @@
+package mllama
+
+import (
+	"github.com/ollama/ollama/ml"
+	"github.com/ollama/ollama/ml/nn"
+	"github.com/ollama/ollama/model"
+)
+
+type Model struct {
+	model.Base
+	model.BytePairEncoding
+
+	*VisionModel `gguf:"v,vision"`
+	*TextModel
+
+	Projector *nn.Linear `gguf:"mm.0"`
+
+	ImageProcessor
+}
+
+func New(c ml.Config) (model.Model, error) {
+	return &Model{
+		BytePairEncoding: model.NewBytePairEncoding(
+			c.String("tokenizer.ggml.pretokenizer", `(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\r\n\p{L}\p{N}]?\p{L}+|\p{N}{1,3}| ?[^\s\p{L}\p{N}]+[\r\n]*|\s*[\r\n]+|\s+(?!\S)|\s+`),
+			&model.Vocabulary{
+				Values: c.Strings("tokenizer.ggml.tokens"),
+				Types:  c.Uints("tokenizer.ggml.token_type"),
+				Merges: c.Strings("tokenizer.ggml.merges"),
+				BOS:    c.Uint("tokenizer.ggml.bos_token_id"),
+				EOS:    c.Uint("tokenizer.ggml.eos_token_id"),
+			},
+		),
+		ImageProcessor: newImageProcessor(c),
+		VisionModel:    newVisionModel(c),
+		TextModel:      newTextModel(c),
+	}, nil
+}
+
+func (m *Model) Forward(ctx ml.Context, opts model.Options) (ml.Tensor, error) {
+	var crossAttentionStates ml.Tensor
+	if opts.Images != nil {
+		f32s, aspectRatioID, err := m.ImageProcessor.ProcessImage(opts.Images[0])
+		if err != nil {
+			return nil, err
+		}
+
+		pixelValues, err := ctx.FromFloatSlice(f32s,
+			m.ImageProcessor.imageSize,
+			m.ImageProcessor.imageSize,
+			m.ImageProcessor.numChannels,
+			m.ImageProcessor.maxNumTiles,
+		)
+		if err != nil {
+			return nil, err
+		}
+
+		aspectRatio, err := ctx.FromIntSlice([]int32{int32(aspectRatioID)}, 1)
+		if err != nil {
+			return nil, err
+		}
+
+		positions := make([]int32, 1601)
+		for i := range positions {
+			positions[i] = int32(i)
+		}
+
+		positionIDs, err := ctx.FromIntSlice(positions, len(positions))
+		if err != nil {
+			return nil, err
+		}
+
+		crossAttentionStates = m.VisionModel.Forward(ctx, pixelValues, positionIDs, aspectRatio)
+		crossAttentionStates = m.Projector.Forward(ctx, crossAttentionStates)
+	}
+
+	inputs, err := ctx.FromIntSlice(opts.Inputs(), len(opts.Inputs()))
+	if err != nil {
+		return nil, err
+	}
+
+	positions, err := ctx.FromIntSlice(opts.Positions(), len(opts.Positions()))
+	if err != nil {
+		return nil, err
+	}
+
+	// TODO: attention mask, cross attention mask
+	hiddenState := m.TextModel.Forward(ctx, inputs, positions, nil, crossAttentionStates, nil, opts.Cache)
+
+	outputs, err := ctx.FromIntSlice([]int32{int32(len(opts.Positions())) - 1}, 1)
+	if err != nil {
+		return nil, err
+	}
+
+	return hiddenState.Rows(ctx, outputs), nil
+}
+
+func init() {
+	model.Register("mllama", New)
+}