kvcache: Use SetRows to store cache data

We currently copy data into the KV cache in contiguous buffers using
ggml_cpy(). ggml_set_rows() was introduced to allow scatter operation
so that contiguous buffers are no longer required. The direct primary
benefit of this is that we no longer need to perform defragmentation.

However, GGML recently removed an optimization for ggml_cpy() and
we picked it up in 544b673 "ggml update to b6840 (#12791)". This
caused a roughly 40% drop in token generation performance on CUDA
due to CUDA graphs no longer being used. By switching to
ggml_set_rows(), the original optimization is no longer necessary
and CUDA performance is restored.

Fixes #13112

ml/backend.go

@@ -194,6 +194,7 @@ type Tensor interface {
 	Repeat(ctx Context, dim, n int) Tensor
 	Concat(ctx Context, t2 Tensor, dim int) Tensor
 	Rows(ctx Context, t2 Tensor) Tensor
+	SetRows(ctx Context, src Tensor, idxs Tensor) Tensor
 	Copy(ctx Context, t2 Tensor) Tensor
 	Duplicate(ctx Context) Tensor
 

ml/backend/ggml/ggml.go

@@ -1338,6 +1338,13 @@ func (t *Tensor) Rows(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
 	}
 }
 
+func (t *Tensor) SetRows(ctx ml.Context, src ml.Tensor, idxs ml.Tensor) ml.Tensor {
+	return &Tensor{
+		b: t.b,
+		t: C.ggml_set_rows(ctx.(*Context).ctx, t.t, src.(*Tensor).t, idxs.(*Tensor).t),
+	}
+}
+
 func (t *Tensor) Copy(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
 	return &Tensor{
 		b: t.b,