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>

cache/cache.go

@@ -0,0 +1,63 @@
+package cache
+
+import (
+	"github.com/ollama/ollama/ml"
+)
+
+type Options struct {
+	Position int
+}
+
+type Cache interface {
+	Sub(i int) Cache
+	Put(ctx ml.Context, key, value ml.Tensor, opts Options) (ml.Tensor, ml.Tensor)
+}
+
+type Simple struct {
+	DType    ml.DType
+	Capacity int
+
+	keys, values []ml.Tensor
+}
+
+func (c *Simple) Sub(i int) Cache {
+	if i >= len(c.keys) {
+		c.keys = append(c.keys, make([]ml.Tensor, i-len(c.keys)+1)...)
+		c.values = append(c.values, make([]ml.Tensor, i-len(c.values)+1)...)
+	}
+
+	return &Simple{
+		keys:     c.keys[i : i+1],
+		values:   c.values[i : i+1],
+		Capacity: c.Capacity,
+		DType:    c.DType,
+	}
+}
+
+func (c *Simple) Put(ctx ml.Context, key, value ml.Tensor, opts Options) (ml.Tensor, ml.Tensor) {
+	if c.keys[0] == nil || c.values[0] == nil {
+		c.keys[0] = ctx.Zeros(c.DType, int(key.Dim(0)*key.Dim(1))*c.Capacity)
+		c.values[0] = ctx.Zeros(c.DType, int(value.Dim(0)*value.Dim(1))*c.Capacity)
+	}
+
+	ctx.Forward(key.Copy(ctx, c.keys[0].View(ctx, int(key.Stride(2))*opts.Position, int(key.Dim(0)*key.Dim(1)*key.Dim(2)))))
+	ctx.Forward(value.Copy(ctx, c.values[0].View(ctx, int(value.Stride(2))*opts.Position, int(value.Dim(0)*value.Dim(1)*value.Dim(2)))))
+
+	n := min(c.Capacity, int(key.Dim(2))+opts.Position)
+
+	key = c.keys[0].View(ctx, 0,
+		int(key.Dim(0)), int(key.Stride(1)),
+		int(key.Dim(1)), int(key.Stride(2)),
+		n,
+	)
+
+	value = c.values[0].View(ctx, 0,
+		int(value.Dim(0)), int(value.Stride(1)),
+		int(value.Dim(1)), int(value.Stride(2)),
+		n,
+	)
+
+	// TODO shift context if necessary
+
+	return key, value
+}