Skip to content

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

17.9 KB
Raw Blame
package create

import (
	"bytes"
	"context"
	"encoding/binary"
	"encoding/json"
	"fmt"
	"io"
	"math"
	"path/filepath"
	"slices"
	"sort"
	"testing"

	st "github.com/ollama/ollama/fs/safetensors"
	"github.com/ollama/ollama/mlx/mlxtest"
)

type captureStore struct{ blobs map[string][]byte }

func newCaptureStore() *captureStore { return &captureStore{blobs: make(map[string][]byte)} }

func (c *captureStore) WriteBlob(r io.Reader, mediaType, name string) (LayerInfo, error) {
	data, err := io.ReadAll(r)
	if err != nil {
		return LayerInfo{}, err
	}
	c.blobs[name] = data
	return LayerInfo{Name: name, MediaType: mediaType, Digest: "sha256:" + name, Size: int64(len(data))}, nil
}

func (c *captureStore) names() []string {
	out := make([]string, 0, len(c.blobs))
	for k := range c.blobs {
		out = append(out, k)
	}
	sort.Strings(out)
	return out
}

type headerEntry struct {
	Dtype string  `json:"dtype"`
	Shape []int32 `json:"shape"`
}

func blobRawHeader(t *testing.T, data []byte) map[string]json.RawMessage {
	t.Helper()
	if len(data) < 8 {
		t.Fatalf("blob too small: %d bytes", len(data))
	}
	n := binary.LittleEndian.Uint64(data[:8])
	var raw map[string]json.RawMessage
	if err := json.Unmarshal(data[8:8+n], &raw); err != nil {
		t.Fatalf("parse header: %v", err)
	}
	return raw
}

func blobHeader(t *testing.T, data []byte) map[string]headerEntry {
	t.Helper()
	out := make(map[string]headerEntry)
	for k, v := range blobRawHeader(t, data) {
		if k == "__metadata__" {
			continue
		}
		var e headerEntry
		if err := json.Unmarshal(v, &e); err != nil {
			t.Fatalf("parse header entry %q: %v", k, err)
		}
		out[k] = e
	}
	return out
}

func blobMetadata(t *testing.T, data []byte) map[string]string {
	t.Helper()
	raw, ok := blobRawHeader(t, data)["__metadata__"]
	if !ok {
		return nil
	}
	var metadata map[string]string
	if err := json.Unmarshal(raw, &metadata); err != nil {
		t.Fatalf("parse metadata: %v", err)
	}
	return metadata
}

func f32le(v float32) []byte {
	b := make([]byte, 4)
	binary.LittleEndian.PutUint32(b, math.Float32bits(v))
	return b
}

func TestWriteBlobsCompressedNVFP4(t *testing.T) {
	dir := t.TempDir()
	writeConfigJSON(t, dir, `{"architectures":["TestModel"],"compression_config":{"format":"nvfp4-pack-quantized"}}`)
	createTestSafetensors(t, filepath.Join(dir, "model.safetensors"), []*st.TensorData{
		st.NewTensorDataFromBytes("linear.weight_packed", "U8", []int32{16, 8}, make([]byte, 16*8)),
		st.NewTensorDataFromBytes("linear.weight_scale", "F8_E4M3", []int32{16, 1}, make([]byte, 16)),
		st.NewTensorDataFromBytes("linear.weight_global_scale", "F32", []int32{}, f32le(4.0)),
		st.NewTensorDataFromBytes("linear.input_global_scale", "F32", []int32{}, f32le(8.0)),
		st.NewTensorDataFromBytes("linear.input_scale", "F32", []int32{}, f32le(2.0)),
		st.NewTensorDataFromBytes("norm.weight", "BF16", []int32{16}, make([]byte, 32)),
	})

	inv, err := ReadInventory(dir)
	if err != nil {
		t.Fatalf("ReadInventory() error = %v", err)
	}
	specs, err := Plan(inv, Classification{Kind: SourcePrequantized}, defaultQuantPolicy{})
	if err != nil {
		t.Fatalf("Plan() error = %v", err)
	}

	store := newCaptureStore()
	if _, err := WriteBlobs(context.Background(), specs, dir, store); err != nil {
		t.Fatalf("WriteBlobs() error = %v", err)
	}

	fused, ok := store.blobs["linear.weight"]
	if !ok {
		t.Fatalf("missing fused blob; got %v", store.names())
	}
	hdr := blobHeader(t, fused)

	if w := hdr["linear.weight"]; w.Dtype != "U32" || !slices.Equal(w.Shape, []int32{16, 2}) {
		t.Errorf("fused weight = %+v, want U32 [16 2] (repacked)", w)
	}
	if s := hdr["linear.weight.scale"]; s.Dtype != "U8" {
		t.Errorf("fused scale dtype = %q, want U8 (relabeled from F8_E4M3)", s.Dtype)
	}
	if g, ok := hdr["linear.weight.global_scale"]; !ok || g.Dtype != "F32" {
		t.Errorf("fused global_scale = %+v ok=%v, want F32", g, ok)
	}
	// compressed-tensors stores the global scale inverted.
	gs := readPackedTensorRaw(t, fused, "linear.weight.global_scale")
	if got := math.Float32frombits(binary.LittleEndian.Uint32(gs)); got != 0.25 {
		t.Errorf("global_scale = %v, want 0.25 (reciprocal of 4.0)", got)
	}
	if len(hdr) != 3 {
		t.Errorf("fused tensors = %v, want only weight, block scale and global scale", hdr)
	}
	if got := store.names(); !slices.Equal(got, []string{"linear.weight", "norm.weight"}) {
		t.Errorf("blob names = %v, want only fused weight and norm", got)
	}

	// the scale companion is folded in, not its own blob.
	if _, leaked := store.blobs["linear.weight_scale"]; leaked {
		t.Error("scale companion leaked as its own blob")
	}

	// the norm passes through unchanged as its own blob.
	norm, ok := store.blobs["norm.weight"]
	if !ok {
		t.Fatalf("missing norm blob; got %v", store.names())
	}
	if nh := blobHeader(t, norm)["norm.weight"]; nh.Dtype != "BF16" || !slices.Equal(nh.Shape, []int32{16}) {
		t.Errorf("norm = %+v, want BF16 [16]", nh)
	}
}

func TestWriteBlobsModelOptSkipsActivationScale(t *testing.T) {
	dir := t.TempDir()
	writeConfigJSON(t, dir, `{"architectures":["TestModel"]}`)
	createTestSafetensors(t, filepath.Join(dir, "model.safetensors"), []*st.TensorData{
		st.NewTensorDataFromBytes("linear.weight", "U8", []int32{16, 8}, make([]byte, 16*8)),
		st.NewTensorDataFromBytes("linear.weight_scale", "F8_E4M3", []int32{16, 1}, make([]byte, 16)),
		st.NewTensorDataFromBytes("linear.weight_scale_2", "F32", []int32{}, f32le(4.0)),
		st.NewTensorDataFromBytes("linear.input_scale", "BF16", []int32{}, make([]byte, 2)),
		st.NewTensorDataFromBytes("attention.k_scale", "F32", []int32{}, f32le(0.25)),
		st.NewTensorDataFromBytes("attention.v_scale", "F32", []int32{}, f32le(0.5)),
	})

	inv, err := ReadInventory(dir)
	if err != nil {
		t.Fatalf("ReadInventory() error = %v", err)
	}
	specs, err := Plan(inv, Classification{Kind: SourcePrequantized}, defaultQuantPolicy{})
	if err != nil {
		t.Fatalf("Plan() error = %v", err)
	}
	store := newCaptureStore()
	if _, err := WriteBlobs(context.Background(), specs, dir, store); err != nil {
		t.Fatalf("WriteBlobs() error = %v", err)
	}

	fused := store.blobs["linear.weight"]
	hdr := blobHeader(t, fused)
	if len(hdr) != 3 {
		t.Errorf("fused tensors = %v, want only weight, block scale and global scale", hdr)
	}
	if _, ok := store.blobs["linear.input_scale"]; ok {
		t.Error("unused activation scale should not be emitted")
	}
	for name, want := range map[string]float32{
		"attention.k_scale": 0.25,
		"attention.v_scale": 0.5,
	} {
		raw := readPackedTensorRaw(t, store.blobs[name], name)
		if got := math.Float32frombits(binary.LittleEndian.Uint32(raw)); got != want {
			t.Errorf("%s = %v, want %v (stored as-is)", name, got, want)
		}
	}
}

func TestWriteBlobsNVFP4ExpertScaleBanks(t *testing.T) {
	for _, tc := range []struct {
		name, weight, global, input string
		wantGlobal                  []float32
	}{
		{"modelopt", ".weight", ".weight_scale_2", ".input_scale", []float32{2, 4}},
		{"compressed-tensors", ".weight_packed", ".weight_global_scale", ".input_global_scale", []float32{0.5, 0.25}},
	} {
		t.Run(tc.name, func(t *testing.T) {
			dir := t.TempDir()
			writeConfigJSON(t, dir, `{"architectures":["TestModel"]}`)
			createTestSafetensors(t, filepath.Join(dir, "model.safetensors"), []*st.TensorData{
				st.NewTensorDataFromBytes("experts"+tc.weight, "U32", []int32{2, 16, 2}, make([]byte, 2*16*2*4)),
				st.NewTensorDataFromBytes("experts.weight_scale", "U8", []int32{2, 16, 1}, make([]byte, 2*16)),
				st.NewTensorDataFromBytes("experts"+tc.global, "F32", []int32{2}, append(f32le(2.0), f32le(4.0)...)),
				st.NewTensorDataFromBytes("experts"+tc.input, "F32", []int32{2}, append(f32le(8.0), f32le(16.0)...)),
			})

			inv, err := ReadInventory(dir)
			if err != nil {
				t.Fatalf("ReadInventory() error = %v", err)
			}
			specs, err := Plan(inv, Classification{Kind: SourcePrequantized}, defaultQuantPolicy{})
			if err != nil {
				t.Fatalf("Plan() error = %v", err)
			}
			store := newCaptureStore()
			if _, err := WriteBlobs(context.Background(), specs, dir, store); err != nil {
				t.Fatalf("WriteBlobs() error = %v", err)
			}

			hdr := blobHeader(t, store.blobs["experts.weight"])
			name := "experts.weight.global_scale"
			if scale, ok := hdr[name]; !ok || scale.Dtype != "F32" || !slices.Equal(scale.Shape, []int32{2}) {
				t.Errorf("%s = %+v ok=%v, want F32 [2]", name, scale, ok)
			}
			if got, want := readPackedTensorRaw(t, store.blobs["experts.weight"], name), encodeFloat32s(tc.wantGlobal...); !bytes.Equal(got, want) {
				t.Errorf("%s bytes = %v, want %v", name, got, want)
			}
			if len(hdr) != 3 || len(store.blobs) != 1 {
				t.Errorf("tensors = %v, blobs = %v, want only fused weight, block scale and global scale", hdr, store.names())
			}
		})
	}
}

func TestWriteBlobsModelOptMXFP8(t *testing.T) {
	dir := t.TempDir()
	writeConfigJSON(t, dir, `{"architectures":["TestModel"],"dtype":"bfloat16"}`)
	weight := make([]byte, 16*128)
	for i := range weight {
		weight[i] = byte(i)
	}
	createTestSafetensors(t, filepath.Join(dir, "model.safetensors"), []*st.TensorData{
		st.NewTensorDataFromBytes("linear.weight", "F8_E4M3", []int32{16, 128}, weight),
		st.NewTensorDataFromBytes("linear.weight_scale", "U8", []int32{16, 4}, make([]byte, 16*4)),
	})

	inv, err := ReadInventory(dir)
	if err != nil {
		t.Fatal(err)
	}
	specs, err := Plan(inv, Classification{Kind: SourcePrequantized}, defaultQuantPolicy{})
	if err != nil {
		t.Fatal(err)
	}
	store := newCaptureStore()
	if _, err := WriteBlobs(context.Background(), specs, dir, store); err != nil {
		t.Fatal(err)
	}
	if len(store.blobs) != 1 {
		t.Fatalf("got blobs %v, want only fused weight", store.names())
	}
	blob := store.blobs["linear.weight"]
	h := blobHeader(t, blob)
	if w := h["linear.weight"]; w.Dtype != "U32" || !slices.Equal(w.Shape, []int32{16, 32}) {
		t.Errorf("weight header = %+v, want U32 [16 32]", w)
	}
	if s := h["linear.weight.scale"]; s.Dtype != "U8" || !slices.Equal(s.Shape, []int32{16, 4}) {
		t.Errorf("scale header = %+v, want U8 [16 4]", s)
	}
	if got := readPackedTensorRaw(t, blob, "linear.weight"); !slices.Equal(got, weight) {
		t.Error("FP8 weight bytes changed during packing")
	}
	if md := blobMetadata(t, blob); md["quant_type"] != "mxfp8" || md["group_size"] != "32" {
		t.Errorf("metadata = %v, want mxfp8 group 32", md)
	}
}

func TestWriteBlobsFlintModelOptExpertBank(t *testing.T) {
	for _, tc := range []struct {
		weightDtype string
		weightShape []int32
		scaleDtype  string
	}{
		{"U32", []int32{2, 2}, "U8"},
		{"U8", []int32{2, 8}, "U8"},
		{"U32", []int32{2, 2}, "F8_E4M3"},
		{"U8", []int32{2, 8}, "F8_E4M3"},
	} {
		t.Run(tc.weightDtype+"/"+tc.scaleDtype, func(t *testing.T) {
			dir := t.TempDir()
			writeConfigJSON(t, dir, `{"architectures":["FlintForCausalLM"]}`)
			const prefix = "model.layers.2.mlp.experts"
			var tensors []*st.TensorData
			for expert := range 2 {
				stem := fmt.Sprintf("%s.%d.gate_proj", prefix, expert)
				tensors = append(tensors,
					st.NewTensorDataFromBytes(stem+".weight", tc.weightDtype, tc.weightShape, bytes.Repeat([]byte{byte(expert + 1)}, 16)),
					st.NewTensorDataFromBytes(stem+".weight_scale", tc.scaleDtype, []int32{2, 1}, []byte{byte(expert + 3), byte(expert + 4)}),
					st.NewTensorDataFromBytes(stem+".weight_scale_2", "F32", []int32{1}, f32le(float32(expert+1)/4)),
					st.NewTensorDataFromBytes(stem+".input_scale", "F32", []int32{1}, f32le(8)),
				)
			}
			createTestSafetensors(t, filepath.Join(dir, "model.safetensors"), tensors)
			inv, err := ReadInventory(dir)
			if err != nil {
				t.Fatal(err)
			}
			class, err := Classify(inv, "")
			if err != nil {
				t.Fatal(err)
			}
			specs, err := Plan(inv, class, defaultQuantPolicy{})
			if err != nil {
				t.Fatal(err)
			}
			store := newCaptureStore()
			if _, err := WriteBlobs(context.Background(), specs, dir, store); err != nil {
				t.Fatal(err)
			}
			if names := store.names(); !slices.Equal(names, []string{prefix}) {
				t.Fatalf("blob names = %v, want one expert bank", names)
			}
			blob := store.blobs[prefix]
			name := prefix + ".gate_proj.weight"
			header := blobHeader(t, blob)
			for _, tt := range []struct {
				name  string
				dtype string
				shape []int32
			}{
				{name, "U32", []int32{2, 2, 2}},
				{name + ".scale", "U8", []int32{2, 2, 1}},
				{name + ".global_scale", "F32", []int32{2}},
			} {
				got, ok := header[tt.name]
				if !ok || got.Dtype != tt.dtype || !slices.Equal(got.Shape, tt.shape) {
					t.Errorf("%s header = %+v, want %s %v", tt.name, got, tt.dtype, tt.shape)
				}
			}
			if got := readPackedTensorRaw(t, blob, name); !bytes.Equal(got, append(bytes.Repeat([]byte{1}, 16), bytes.Repeat([]byte{2}, 16)...)) {
				t.Error("packed expert words changed or were stacked out of order")
			}
			if got := readPackedTensorRaw(t, blob, name+".scale"); !bytes.Equal(got, []byte{3, 4, 4, 5}) {
				t.Error("expert scale bytes changed or were stacked out of order")
			}
			if got := readPackedTensorRaw(t, blob, name+".global_scale"); !bytes.Equal(got, append(f32le(0.25), f32le(0.5)...)) {
				t.Error("ModelOpt global scales changed or were stacked out of order")
			}
		})
	}
}

func TestWriteBlobsQuantizeFloat(t *testing.T) {
	mlxtest.SkipIfUnavailable(t)
	for _, quantize := range []string{"int4", "nvfp4", "mxfp8"} {
		t.Run(quantize, func(t *testing.T) {
			dir := t.TempDir()
			writeConfigJSON(t, dir, `{"architectures":["TestModel"]}`)
			createTestSafetensors(t, filepath.Join(dir, "model.safetensors"), []*st.TensorData{
				st.NewTensorDataFromBytes("model.layers.0.self_attn.q_proj.weight", "BF16", []int32{128, 128}, make([]byte, 128*128*2)),
				st.NewTensorDataFromBytes("model.norm.weight", "BF16", []int32{128}, make([]byte, 128*2)),
			})

			inv, err := ReadInventory(dir)
			if err != nil {
				t.Fatalf("ReadInventory() error = %v", err)
			}
			specs, err := Plan(inv, Classification{Kind: SourceFloat, Quantize: quantize}, defaultQuantPolicy{})
			if err != nil {
				t.Fatalf("Plan() error = %v", err)
			}
			store := newCaptureStore()
			if _, err := WriteBlobs(context.Background(), specs, dir, store); err != nil {
				t.Fatalf("WriteBlobs() error = %v", err)
			}

			q, ok := store.blobs["model.layers.0.self_attn.q_proj.weight"]
			if !ok {
				t.Fatalf("missing q_proj blob; got %v", store.names())
			}
			hdr := blobHeader(t, q)
			if w := hdr["model.layers.0.self_attn.q_proj.weight"]; w.Dtype != "U32" {
				t.Errorf("quantized weight dtype = %q, want U32 (packed %s)", w.Dtype, quantize)
			}
			if _, ok := hdr["model.layers.0.self_attn.q_proj.weight.scale"]; !ok {
				t.Error("quantized blob missing scale")
			}

			norm, ok := store.blobs["model.norm.weight"]
			if !ok {
				t.Fatalf("missing norm blob; got %v", store.names())
			}
			if nh := blobHeader(t, norm)["model.norm.weight"]; nh.Dtype != "BF16" {
				t.Errorf("norm dtype = %q, want BF16 (kept, not quantized)", nh.Dtype)
			}
		})
	}
}

func TestWriteBlobsQuantizeGemma4PublisherNames(t *testing.T) {
	mlxtest.SkipIfUnavailable(t)
	const (
		gateUpName = "model.language_model.layers.0.experts.gate_up_proj"
		downName   = "model.language_model.layers.0.experts.down_proj"
	)
	dir := t.TempDir()
	writeConfigJSON(t, dir, `{"architectures":["Gemma4ForCausalLM"],"num_hidden_layers":30,"num_experts":128}`)
	createTestSafetensors(t, filepath.Join(dir, "model.safetensors"), []*st.TensorData{
		st.NewTensorDataFromBytes(gateUpName, "BF16", []int32{8, 256, 128}, make([]byte, 8*256*128*2)),
		st.NewTensorDataFromBytes(downName, "BF16", []int32{8, 128, 128}, make([]byte, 8*128*128*2)),
	})

	inv, err := ReadInventory(dir)
	if err != nil {
		t.Fatalf("ReadInventory() error = %v", err)
	}
	policy, err := newTensorImportTransform(inv)
	if err != nil {
		t.Fatalf("newTensorImportTransform() error = %v", err)
	}
	specs, err := Plan(inv, Classification{Kind: SourceFloat, Quantize: "int4"}, policy)
	if err != nil {
		t.Fatalf("Plan() error = %v", err)
	}
	store := newCaptureStore()
	if _, err := WriteBlobs(context.Background(), specs, dir, store); err != nil {
		t.Fatalf("WriteBlobs() error = %v", err)
	}

	for name, wantQuantize := range map[string]string{gateUpName: "int4", downName: "int8"} {
		blob, ok := store.blobs[name]
		if !ok {
			t.Fatalf("missing publisher-named blob %q; got %v", name, store.names())
		}
		header := blobHeader(t, blob)
		if len(header) != 3 {
			t.Errorf("blob %q has tensors %v, want weight, scale, and bias", name, header)
		}
		for _, tensorName := range []string{name, name + ".scale", name + ".bias"} {
			if _, ok := header[tensorName]; !ok {
				t.Errorf("blob %q missing exact tensor name %q", name, tensorName)
			}
		}
		metadata := blobMetadata(t, blob)
		if metadata["quant_type"] != wantQuantize || metadata["group_size"] != "64" {
			t.Errorf("blob %q metadata = %v, want quant_type=%s group_size=64", name, metadata, wantQuantize)
		}
	}
}

func TestWriteBlobsBlockFP8Decode(t *testing.T) {
	mlxtest.SkipIfUnavailable(t)
	dir := t.TempDir()
	writeConfigJSON(t, dir, `{"architectures":["TestModel"]}`)
	createTestSafetensors(t, filepath.Join(dir, "model.safetensors"), []*st.TensorData{
		st.NewTensorDataFromBytes("model.layers.0.mlp.down_proj.weight", "F8_E4M3", []int32{128, 128}, make([]byte, 128*128)),
		st.NewTensorDataFromBytes("model.layers.0.mlp.down_proj.weight_scale_inv", "F32", []int32{1, 1}, f32le(1.0)),
	})

	inv, err := ReadInventory(dir)
	if err != nil {
		t.Fatalf("ReadInventory() error = %v", err)
	}
	specs, err := Plan(inv, Classification{Kind: SourceBlockFP8, Quantize: "mxfp8"}, defaultQuantPolicy{})
	if err != nil {
		t.Fatalf("Plan() error = %v", err)
	}
	store := newCaptureStore()
	if _, err := WriteBlobs(context.Background(), specs, dir, store); err != nil {
		t.Fatalf("WriteBlobs() error = %v", err)
	}

	b, ok := store.blobs["model.layers.0.mlp.down_proj.weight"]
	if !ok {
		t.Fatalf("missing decoded blob; got %v", store.names())
	}
	hdr := blobHeader(t, b)
	if w := hdr["model.layers.0.mlp.down_proj.weight"]; w.Dtype != "U32" {
		t.Errorf("decoded+quantized weight dtype = %q, want U32 (packed mxfp8)", w.Dtype)
	}
	if _, ok := hdr["model.layers.0.mlp.down_proj.weight.scale"]; !ok {
		t.Error("mxfp8 blob missing scale")
	}
	if _, leaked := store.blobs["model.layers.0.mlp.down_proj.weight_scale_inv"]; leaked {
		t.Error("fp8 scale companion leaked as its own blob")
	}
}