Skip to content

Get up and running with Kimi, GLM, MiniMax, DeepSeek, gpt-oss, Qwen, Gemma and other models.

4.1 KB
Raw Blame
package model

import (
	"testing"

	"github.com/ollama/ollama/mlx"
	"github.com/ollama/ollama/mlx/mlxtest"
	"github.com/ollama/ollama/mlxrunner/nn"
)

func TestMakeEmbeddingLayerDense(t *testing.T) {
	mlxtest.Run(t, func(t *mlxtest.T) {
		weight := mlx.FromValues([]float32{
			1, 2, 3, 4,
			5, 6, 7, 8,
		}, 2, 4).AsType(mlx.DTypeBFloat16)

		emb := MakeEmbeddingLayer(map[string]*mlx.Array{
			"model.embed_tokens.weight": weight,
		}, "model.embed_tokens", 0, 0, "", nil)

		dense, ok := emb.(*nn.Embedding)
		if !ok {
			t.Fatalf("embedding type = %T, want *nn.Embedding", emb)
		}
		if dense.Weight.DType() != mlx.DTypeBFloat16 {
			t.Fatalf("embedding dtype = %v, want %v", dense.Weight.DType(), mlx.DTypeBFloat16)
		}
		if _, ok := emb.AsLinear().(*nn.Linear); !ok {
			t.Fatalf("AsLinear type = %T, want *nn.Linear", emb.AsLinear())
		}
	})
}

func TestMakeEmbeddingLayerQuantized(t *testing.T) {
	mlxtest.Run(t, func(t *mlxtest.T) {
		denseWeight := mlx.FromValues(func() []float32 {
			out := make([]float32, 2*64)
			for i := range out {
				out[i] = float32(i%17) / 8
			}
			return out
		}(), 2, 64).AsType(mlx.DTypeBFloat16)

		qw, scales, qbiases := mlx.Quantize(denseWeight, 64, 4, "affine")
		mlx.Eval(qw, scales, qbiases)

		emb := MakeEmbeddingLayer(map[string]*mlx.Array{
			"model.embed_tokens.weight":       qw,
			"model.embed_tokens.weight_scale": scales,
			"model.embed_tokens.weight_qbias": qbiases,
		}, "model.embed_tokens", 64, 4, "affine", nil)

		qemb, ok := emb.(*nn.QuantizedEmbedding)
		if !ok {
			t.Fatalf("embedding type = %T, want *nn.QuantizedEmbedding", emb)
		}
		if qemb.GroupSize != 64 || qemb.Bits != 4 || qemb.Mode != "affine" {
			t.Fatalf("quant params = (%d, %d, %q), want (64, 4, %q)", qemb.GroupSize, qemb.Bits, qemb.Mode, "affine")
		}

		indices := mlx.FromValues([]int32{1, 0}, 2)
		out := emb.Forward(indices)
		mlx.Eval(out)
		if dims := out.Dims(); len(dims) != 2 || dims[0] != 2 || dims[1] != 64 {
			t.Fatalf("embedding output dims = %v, want [2 64]", dims)
		}
		if _, ok := emb.AsLinear().(*nn.QuantizedLinear); !ok {
			t.Fatalf("AsLinear type = %T, want *nn.QuantizedLinear", emb.AsLinear())
		}
	})
}

// A checkpoint scale is converted to MLX's representation at load, and the
// same converted scale reaches the linear view.
func TestMakeEmbeddingLayerQuantizedGlobalScale(t *testing.T) {
	mlxtest.Run(t, func(t *mlxtest.T) {
		const checkpoint = 0.375
		emb := MakeEmbeddingLayer(map[string]*mlx.Array{
			"model.embed_tokens.weight":              {},
			"model.embed_tokens.weight_scale":        {},
			"model.embed_tokens.weight.global_scale": mlx.FromValues([]float32{checkpoint}, 1),
		}, "model.embed_tokens", 16, 4, "nvfp4", nil)

		qemb, ok := emb.(*nn.QuantizedEmbedding)
		if !ok {
			t.Fatalf("embedding type = %T, want *nn.QuantizedEmbedding", emb)
		}
		mlx.Eval(qemb.GlobalScale)
		want := float32(checkpoint * mlx.Nvfp4MaxProduct)
		if got := qemb.GlobalScale.Floats(); len(got) != 1 || got[0] != want {
			t.Fatalf("GlobalScale = %v, want [%v]", got, want)
		}

		linear, ok := qemb.AsLinear().(*nn.QuantizedLinear)
		if !ok {
			t.Fatalf("AsLinear type = %T, want *nn.QuantizedLinear", qemb.AsLinear())
		}
		if linear.GlobalScale != qemb.GlobalScale {
			t.Fatalf("AsLinear GlobalScale = %p, want %p", linear.GlobalScale, qemb.GlobalScale)
		}
	})
}

// ModelOpt's own "weight_scale_2" name is honoured, and converted the same way.
func TestMakeEmbeddingLayerQuantizedGlobalScaleFallback(t *testing.T) {
	mlxtest.Run(t, func(t *mlxtest.T) {
		const checkpoint = 0.375
		emb := MakeEmbeddingLayer(map[string]*mlx.Array{
			"model.embed_tokens.weight":         {},
			"model.embed_tokens.weight_scale":   {},
			"model.embed_tokens.weight_scale_2": mlx.FromValues([]float32{checkpoint}, 1),
		}, "model.embed_tokens", 16, 4, "nvfp4", nil)

		qemb, ok := emb.(*nn.QuantizedEmbedding)
		if !ok {
			t.Fatalf("embedding type = %T, want *nn.QuantizedEmbedding", emb)
		}
		mlx.Eval(qemb.GlobalScale)
		want := float32(checkpoint * mlx.Nvfp4MaxProduct)
		if got := qemb.GlobalScale.Floats(); len(got) != 1 || got[0] != want {
			t.Fatalf("GlobalScale = %v, want [%v]", got, want)
		}
	})
}