Add experimental MLX backend and engine with imagegen support (#13648)

* WIP - MLX backend with gemma3

* MLX: add cmake and go tag build toggles

To build the new MLX backend code:
  cmake --preset MLX
  cmake --build --preset MLX --parallel
  cmake --install build --component MLX
  go build -tags mlx .

Note: the main.go entrypoint for the MLX engine will change in a follow up commit.

* add experimental image generation runtime

* add experimental image generation runtime

* MLX: wire up cuda build for linux

* MLX: get dependencies correct and dedup

This is still too large for a unified github artifact, but is now "correct" for the mlx_cuda_v13
directory.

* fix relative link bug in dedup

* Add darwin build and readme

* add go build tag for mlx dependent code and wire up build_darwin.sh

* lint cleanup

* macos: build mlx for x86

This will be CPU only.

* cuda build instructions and fix drift from mlx bump

* stale comment

* Delete agent helper doc

* Clean up readme.md

* Revise README for tokenizer clarity and details

Updated README to clarify tokenizer functionality and removed correctness section.

---------
Co-authored-by: jmorganca <jmorganca@gmail.com>

x/imagegen/models/qwen_image/scheduler_test.go

+//go:build mlx
+
+package qwen_image
+
+import (
+	"math"
+	"testing"
+)
+
+// TestSchedulerSetTimesteps verifies scheduler sigmas match Python diffusers reference.
+// Golden values generated via:
+//
+//	python3 -c "
+//	from diffusers.schedulers import FlowMatchEulerDiscreteScheduler
+//	import numpy as np
+//	s = FlowMatchEulerDiscreteScheduler(num_train_timesteps=1000, base_shift=0.5, max_shift=0.9,
+//	    base_image_seq_len=256, max_image_seq_len=8192, shift_terminal=0.02, use_dynamic_shifting=True)
+//	mu = 4096 * (0.9-0.5)/(8192-256) + 0.5 - (0.9-0.5)/(8192-256)*256
+//	sigmas = np.linspace(1.0, 1.0/30, 30)
+//	s.set_timesteps(sigmas=sigmas, mu=mu)
+//	print(s.sigmas.numpy())"
+func TestSchedulerSetTimesteps(t *testing.T) {
+	cfg := DefaultSchedulerConfig()
+	scheduler := NewFlowMatchScheduler(cfg)
+	scheduler.SetTimesteps(30, 4096)
+
+	// Golden values from Python diffusers (first 3, last 3 before terminal)
+	wantFirst := []float32{1.000000, 0.982251, 0.963889}
+	wantLast := []float32{0.142924, 0.083384, 0.020000}
+
+	// Check first 3
+	for i, want := range wantFirst {
+		got := scheduler.Sigmas[i]
+		if abs32(got-want) > 1e-4 {
+			t.Errorf("sigma[%d]: got %v, want %v", i, got, want)
+		}
+	}
+
+	// Check last 3 (indices 27, 28, 29)
+	for i, want := range wantLast {
+		idx := 27 + i
+		got := scheduler.Sigmas[idx]
+		if abs32(got-want) > 1e-4 {
+			t.Errorf("sigma[%d]: got %v, want %v", idx, got, want)
+		}
+	}
+
+	// Check terminal is 0
+	if scheduler.Sigmas[30] != 0.0 {
+		t.Errorf("terminal sigma: got %v, want 0", scheduler.Sigmas[30])
+	}
+
+	// Check length
+	if len(scheduler.Sigmas) != 31 {
+		t.Errorf("sigmas length: got %d, want 31", len(scheduler.Sigmas))
+	}
+}
+
+// TestSchedulerProperties tests mathematical invariants of the scheduler.
+func TestSchedulerProperties(t *testing.T) {
+	cfg := DefaultSchedulerConfig()
+	scheduler := NewFlowMatchScheduler(cfg)
+	scheduler.SetTimesteps(30, 4096)
+
+	// Property: sigmas monotonically decreasing
+	for i := 1; i < len(scheduler.Sigmas); i++ {
+		if scheduler.Sigmas[i] > scheduler.Sigmas[i-1] {
+			t.Errorf("sigmas not monotonically decreasing at %d: %v > %v",
+				i, scheduler.Sigmas[i], scheduler.Sigmas[i-1])
+		}
+	}
+
+	// Property: first sigma should be ~1.0 (with time shift)
+	if scheduler.Sigmas[0] < 0.9 || scheduler.Sigmas[0] > 1.01 {
+		t.Errorf("first sigma out of expected range [0.9, 1.01]: %v", scheduler.Sigmas[0])
+	}
+
+	// Property: terminal sigma should be exactly 0
+	if scheduler.Sigmas[len(scheduler.Sigmas)-1] != 0.0 {
+		t.Errorf("terminal sigma should be 0, got %v", scheduler.Sigmas[len(scheduler.Sigmas)-1])
+	}
+
+	// Property: last non-terminal sigma should be shift_terminal (0.02)
+	lastNonTerminal := scheduler.Sigmas[len(scheduler.Sigmas)-2]
+	if abs32(lastNonTerminal-0.02) > 1e-5 {
+		t.Errorf("last non-terminal sigma should be 0.02, got %v", lastNonTerminal)
+	}
+
+	// Property: length = steps + 1
+	if len(scheduler.Sigmas) != scheduler.NumSteps+1 {
+		t.Errorf("sigmas length should be steps+1: got %d, want %d",
+			len(scheduler.Sigmas), scheduler.NumSteps+1)
+	}
+}
+
+// TestCalculateShift verifies the mu calculation against Python reference.
+// Golden values from: mu = img_seq_len * m + b where m = (max_shift - base_shift) / (max_seq_len - base_seq_len)
+func TestCalculateShift(t *testing.T) {
+	cases := []struct {
+		imgSeqLen int32
+		want      float32
+	}{
+		{256, 0.5},     // base case
+		{8192, 0.9},    // max case
+		{4096, 0.6935}, // middle case (rounded)
+	}
+
+	for _, c := range cases {
+		got := CalculateShift(c.imgSeqLen, 256, 8192, 0.5, 0.9)
+		if abs32(got-c.want) > 0.001 {
+			t.Errorf("CalculateShift(%d): got %v, want %v", c.imgSeqLen, got, c.want)
+		}
+	}
+}
+
+// TestSchedulerStep verifies the Euler step formula.
+func TestSchedulerStep(t *testing.T) {
+	cfg := DefaultSchedulerConfig()
+	scheduler := NewFlowMatchScheduler(cfg)
+	scheduler.SetTimesteps(30, 4096)
+
+	// Verify dt calculation for first step
+	sigma0 := scheduler.Sigmas[0]
+	sigma1 := scheduler.Sigmas[1]
+	expectedDt := sigma1 - sigma0
+
+	// dt should be negative (sigmas decrease)
+	if expectedDt >= 0 {
+		t.Errorf("expected negative dt, got %v (sigma0=%v, sigma1=%v)", expectedDt, sigma0, sigma1)
+	}
+}
+
+func abs32(x float32) float32 {
+	return float32(math.Abs(float64(x)))
+}

x/imagegen/models/qwen_image/text_encoder_test.go

+//go:build mlx
+
+package qwen_image
+
+import (
+	"encoding/json"
+	"math"
+	"os"
+	"path/filepath"
+	"slices"
+	"testing"
+
+	"github.com/ollama/ollama/x/imagegen/mlx"
+	"github.com/ollama/ollama/x/imagegen/safetensors"
+)
+
+// TinyTextEncoderConfig holds config for the tiny test text encoder
+type TinyTextEncoderConfig struct {
+	HiddenSize        int32   `json:"hidden_size"`
+	NumHiddenLayers   int32   `json:"num_hidden_layers"`
+	IntermediateSize  int32   `json:"intermediate_size"`
+	NumAttentionHeads int32   `json:"num_attention_heads"`
+	NumKeyValueHeads  int32   `json:"num_key_value_heads"`
+	VocabSize         int32   `json:"vocab_size"`
+	RMSNormEps        float32 `json:"rms_norm_eps"`
+	RopeTheta         float32 `json:"rope_theta"`
+	HeadDim           int32   `json:"head_dim"`
+	MRoPESection      []int32 `json:"mrope_section"`
+}
+
+// loadTinyTextEncoder loads the tiny text encoder from testdata
+func loadTinyTextEncoder(t *testing.T) (*Qwen25VL, *TinyTextEncoderConfig) {
+	t.Helper()
+
+	testdataDir := filepath.Join("testdata", "tiny_text_encoder")
+
+	// Load config
+	configData, err := os.ReadFile(filepath.Join(testdataDir, "config.json"))
+	if err != nil {
+		t.Skipf("Skipping: tiny weights not found. Regenerate with Python (see models/CLAUDE.md)")
+	}
+
+	var tinyCfg TinyTextEncoderConfig
+	if err := json.Unmarshal(configData, &tinyCfg); err != nil {
+		t.Fatalf("Failed to parse config: %v", err)
+	}
+
+	// Create encoder config (using Qwen25VLConfig)
+	cfg := &Qwen25VLConfig{
+		HiddenSize:        tinyCfg.HiddenSize,
+		NumHiddenLayers:   tinyCfg.NumHiddenLayers,
+		IntermediateSize:  tinyCfg.IntermediateSize,
+		NumAttentionHeads: tinyCfg.NumAttentionHeads,
+		NumKeyValueHeads:  tinyCfg.NumKeyValueHeads,
+		VocabSize:         tinyCfg.VocabSize,
+		RMSNormEps:        tinyCfg.RMSNormEps,
+		RopeTheta:         tinyCfg.RopeTheta,
+		HeadDim:           tinyCfg.HeadDim,
+		MRoPESection:      tinyCfg.MRoPESection,
+	}
+
+	// Load weights
+	weights, err := safetensors.LoadModelWeights(testdataDir)
+	if err != nil {
+		t.Fatalf("Failed to load weights: %v", err)
+	}
+
+	if err := weights.Load(mlx.DtypeBFloat16); err != nil {
+		t.Fatalf("Failed to bulk load weights: %v", err)
+	}
+
+	// Build encoder
+	embedding, err := weights.Get("model.embed_tokens.weight")
+	if err != nil {
+		t.Fatalf("Failed to get embedding: %v", err)
+	}
+
+	blocks := make([]*VLTextBlock, cfg.NumHiddenLayers)
+	for i := int32(0); i < cfg.NumHiddenLayers; i++ {
+		block, err := newVLTextBlock(weights, int(i), cfg)
+		if err != nil {
+			t.Fatalf("Failed to load block %d: %v", i, err)
+		}
+		blocks[i] = block
+	}
+
+	finalNorm, err := weights.Get("model.norm.weight")
+	if err != nil {
+		t.Fatalf("Failed to get final norm: %v", err)
+	}
+
+	encoder := &Qwen25VL{
+		Config:    cfg,
+		Embedding: embedding,
+		Blocks:    blocks,
+		FinalNorm: finalNorm,
+		HasVision: false, // Text-only mode
+	}
+
+	return encoder, &tinyCfg
+}
+
+// TestTextEncoderForward verifies the text encoder forward pass with tiny weights.
+func TestTextEncoderForward(t *testing.T) {
+	encoder, cfg := loadTinyTextEncoder(t)
+
+	// Create test tokens (within vocab range)
+	tokens := []int32{1, 2, 3, 4, 5}
+
+	// Forward pass using EncodeTextOnly
+	out := encoder.EncodeTextOnly(tokens)
+	mlx.Eval(out)
+
+	// Verify output shape: [batch, seq_len, hidden_size]
+	wantShape := []int32{1, 5, cfg.HiddenSize}
+	if !slices.Equal(out.Shape(), wantShape) {
+		t.Errorf("output shape: got %v, want %v", out.Shape(), wantShape)
+	}
+
+	// Verify output is finite (not NaN or Inf)
+	data := out.Data()
+	for i, v := range data {
+		if math.IsNaN(float64(v)) || math.IsInf(float64(v), 0) {
+			t.Errorf("output[%d] is not finite: %v", i, v)
+			break
+		}
+	}
+}
+
+// TestTextEncoderBatch tests batch processing.
+func TestTextEncoderBatch(t *testing.T) {
+	encoder, cfg := loadTinyTextEncoder(t)
+
+	// For batch test, we'll use EncodeTextOnly with a single sequence
+	// (EncodeTextOnly doesn't support batch, but we can verify single sequence works)
+	tokens := []int32{1, 2, 3}
+
+	out := encoder.EncodeTextOnly(tokens)
+	mlx.Eval(out)
+
+	wantShape := []int32{1, 3, cfg.HiddenSize}
+	if !slices.Equal(out.Shape(), wantShape) {
+		t.Errorf("shape: got %v, want %v", out.Shape(), wantShape)
+	}
+}
+
+// TestMRoPEComputation verifies M-RoPE frequency computation produces valid values.
+func TestMRoPEComputation(t *testing.T) {
+	encoder, cfg := loadTinyTextEncoder(t)
+
+	cossin := encoder.computeTextRoPE(10, 1)
+	mlx.Eval(cossin[0], cossin[1])
+
+	// Verify shapes: [3, B, L, head_dim]
+	wantShape := []int32{3, 1, 10, cfg.HeadDim}
+	if !slices.Equal(cossin[0].Shape(), wantShape) {
+		t.Errorf("cos shape: got %v, want %v", cossin[0].Shape(), wantShape)
+	}
+	if !slices.Equal(cossin[1].Shape(), wantShape) {
+		t.Errorf("sin shape: got %v, want %v", cossin[1].Shape(), wantShape)
+	}
+
+	// Verify cos/sin values are in valid range [-1, 1]
+	cosData := cossin[0].Data()
+	sinData := cossin[1].Data()
+	for i := 0; i < min(100, len(cosData)); i++ {
+		if cosData[i] < -1.01 || cosData[i] > 1.01 {
+			t.Errorf("cos[%d] out of range: %v", i, cosData[i])
+		}
+		if sinData[i] < -1.01 || sinData[i] > 1.01 {
+			t.Errorf("sin[%d] out of range: %v", i, sinData[i])
+		}
+	}
+}

x/imagegen/models/qwen_image/transformer.go

+//go:build mlx
+
+package qwen_image
+
+import (
+	"fmt"
+	"math"
+	"path/filepath"
+
+	"github.com/ollama/ollama/x/imagegen/cache"
+	"github.com/ollama/ollama/x/imagegen/mlx"
+	"github.com/ollama/ollama/x/imagegen/safetensors"
+)
+
+// TransformerConfig holds Qwen-Image transformer configuration
+type TransformerConfig struct {
+	HiddenDim         int32   `json:"hidden_dim"`          // 3072 (24 * 128)
+	NHeads            int32   `json:"num_attention_heads"` // 24
+	HeadDim           int32   `json:"attention_head_dim"`  // 128
+	NLayers           int32   `json:"num_layers"`          // 60
+	InChannels        int32   `json:"in_channels"`         // 64
+	OutChannels       int32   `json:"out_channels"`        // 16
+	PatchSize         int32   `json:"patch_size"`          // 2
+	JointAttentionDim int32   `json:"joint_attention_dim"` // 3584 (text encoder dim)
+	NormEps           float32 `json:"norm_eps"`            // 1e-6
+	AxesDimsRope      []int32 `json:"axes_dims_rope"`      // [16, 56, 56]
+	GuidanceEmbeds    bool    `json:"guidance_embeds"`     // false
+}
+
+// defaultTransformerConfig returns config for Qwen-Image transformer
+func defaultTransformerConfig() *TransformerConfig {
+	return &TransformerConfig{
+		HiddenDim:         3072, // 24 * 128
+		NHeads:            24,
+		HeadDim:           128,
+		NLayers:           60,
+		InChannels:        64,
+		OutChannels:       16,
+		PatchSize:         2,
+		JointAttentionDim: 3584,
+		NormEps:           1e-6,
+		AxesDimsRope:      []int32{16, 56, 56},
+		GuidanceEmbeds:    false,
+	}
+}
+
+// TimestepEmbedder creates timestep embeddings
+type TimestepEmbedder struct {
+	Linear1Weight *mlx.Array // [256, hidden_dim]
+	Linear1Bias   *mlx.Array
+	Linear2Weight *mlx.Array // [hidden_dim, hidden_dim]
+	Linear2Bias   *mlx.Array
+}
+
+// newTimestepEmbedder creates a timestep embedder from weights
+func newTimestepEmbedder(weights *safetensors.ModelWeights) (*TimestepEmbedder, error) {
+	linear1Weight, err := weights.Get("time_text_embed.timestep_embedder.linear_1.weight")
+	if err != nil {
+		return nil, err
+	}
+	linear1Bias, err := weights.Get("time_text_embed.timestep_embedder.linear_1.bias")
+	if err != nil {
+		return nil, err
+	}
+	linear2Weight, err := weights.Get("time_text_embed.timestep_embedder.linear_2.weight")
+	if err != nil {
+		return nil, err
+	}
+	linear2Bias, err := weights.Get("time_text_embed.timestep_embedder.linear_2.bias")
+	if err != nil {
+		return nil, err
+	}
+
+	return &TimestepEmbedder{
+		Linear1Weight: mlx.Transpose(linear1Weight, 1, 0),
+		Linear1Bias:   linear1Bias,
+		Linear2Weight: mlx.Transpose(linear2Weight, 1, 0),
+		Linear2Bias:   linear2Bias,
+	}, nil
+}
+
+// Forward computes timestep embeddings
+// t: [B] timesteps (normalized 0-1, will be scaled by 1000 internally)
+func (te *TimestepEmbedder) Forward(t *mlx.Array) *mlx.Array {
+	half := int32(128) // embedding_dim / 2
+
+	// Sinusoidal embedding with flip_sin_to_cos=True, scale=1000
+	freqs := make([]float32, half)
+	for i := int32(0); i < half; i++ {
+		freqs[i] = float32(math.Exp(-math.Log(10000.0) * float64(i) / float64(half)))
+	}
+	freqsArr := mlx.NewArray(freqs, []int32{1, half})
+
+	tExpanded := mlx.ExpandDims(t, 1)
+	args := mlx.Mul(tExpanded, freqsArr)
+	args = mlx.MulScalar(args, 1000.0) // scale
+
+	// [cos, sin] (flip_sin_to_cos=True)
+	sinArgs := mlx.Sin(args)
+	cosArgs := mlx.Cos(args)
+	embedding := mlx.Concatenate([]*mlx.Array{cosArgs, sinArgs}, 1) // [B, 256]
+
+	// MLP: linear1 -> silu -> linear2
+	h := mlx.Linear(embedding, te.Linear1Weight)
+	h = mlx.Add(h, te.Linear1Bias)
+	h = mlx.SiLU(h)
+	h = mlx.Linear(h, te.Linear2Weight)
+	h = mlx.Add(h, te.Linear2Bias)
+
+	return h
+}
+
+// JointAttention implements dual-stream joint attention
+type JointAttention struct {
+	// Image projections
+	ToQ    *mlx.Array
+	ToQB   *mlx.Array
+	ToK    *mlx.Array
+	ToKB   *mlx.Array
+	ToV    *mlx.Array
+	ToVB   *mlx.Array
+	ToOut  *mlx.Array
+	ToOutB *mlx.Array
+	NormQ  *mlx.Array
+	NormK  *mlx.Array
+
+	// Text (added) projections
+	AddQProj  *mlx.Array
+	AddQProjB *mlx.Array
+	AddKProj  *mlx.Array
+	AddKProjB *mlx.Array
+	AddVProj  *mlx.Array
+	AddVProjB *mlx.Array
+	ToAddOut  *mlx.Array
+	ToAddOutB *mlx.Array
+	NormAddQ  *mlx.Array
+	NormAddK  *mlx.Array
+
+	NHeads  int32
+	HeadDim int32
+	Scale   float32
+}
+
+// newJointAttention creates a joint attention layer
+func newJointAttention(weights *safetensors.ModelWeights, prefix string, cfg *TransformerConfig) (*JointAttention, error) {
+	toQ, _ := weights.Get(prefix + ".attn.to_q.weight")
+	toQB, _ := weights.Get(prefix + ".attn.to_q.bias")
+	toK, _ := weights.Get(prefix + ".attn.to_k.weight")
+	toKB, _ := weights.Get(prefix + ".attn.to_k.bias")
+	toV, _ := weights.Get(prefix + ".attn.to_v.weight")
+	toVB, _ := weights.Get(prefix + ".attn.to_v.bias")
+	toOut, _ := weights.Get(prefix + ".attn.to_out.0.weight")
+	toOutB, _ := weights.Get(prefix + ".attn.to_out.0.bias")
+	normQ, _ := weights.Get(prefix + ".attn.norm_q.weight")
+	normK, _ := weights.Get(prefix + ".attn.norm_k.weight")
+
+	addQProj, _ := weights.Get(prefix + ".attn.add_q_proj.weight")
+	addQProjB, _ := weights.Get(prefix + ".attn.add_q_proj.bias")
+	addKProj, _ := weights.Get(prefix + ".attn.add_k_proj.weight")
+	addKProjB, _ := weights.Get(prefix + ".attn.add_k_proj.bias")
+	addVProj, _ := weights.Get(prefix + ".attn.add_v_proj.weight")
+	addVProjB, _ := weights.Get(prefix + ".attn.add_v_proj.bias")
+	toAddOut, _ := weights.Get(prefix + ".attn.to_add_out.weight")
+	toAddOutB, _ := weights.Get(prefix + ".attn.to_add_out.bias")
+	normAddQ, _ := weights.Get(prefix + ".attn.norm_added_q.weight")
+	normAddK, _ := weights.Get(prefix + ".attn.norm_added_k.weight")
+
+	return &JointAttention{
+		ToQ:       mlx.Transpose(toQ, 1, 0),
+		ToQB:      toQB,
+		ToK:       mlx.Transpose(toK, 1, 0),
+		ToKB:      toKB,
+		ToV:       mlx.Transpose(toV, 1, 0),
+		ToVB:      toVB,
+		ToOut:     mlx.Transpose(toOut, 1, 0),
+		ToOutB:    toOutB,
+		NormQ:     normQ,
+		NormK:     normK,
+		AddQProj:  mlx.Transpose(addQProj, 1, 0),
+		AddQProjB: addQProjB,
+		AddKProj:  mlx.Transpose(addKProj, 1, 0),
+		AddKProjB: addKProjB,
+		AddVProj:  mlx.Transpose(addVProj, 1, 0),
+		AddVProjB: addVProjB,
+		ToAddOut:  mlx.Transpose(toAddOut, 1, 0),
+		ToAddOutB: toAddOutB,
+		NormAddQ:  normAddQ,
+		NormAddK:  normAddK,
+		NHeads:    cfg.NHeads,
+		HeadDim:   cfg.HeadDim,
+		Scale:     float32(1.0 / math.Sqrt(float64(cfg.HeadDim))),
+	}, nil
+}
+
+// Forward computes joint attention
+// img: [B, L_img, D], txt: [B, L_txt, D]
+// imgFreqs, txtFreqs: complex RoPE frequencies [L, head_dim/2] as interleaved real/imag
+func (attn *JointAttention) Forward(img, txt *mlx.Array, imgFreqs, txtFreqs *mlx.Array) (*mlx.Array, *mlx.Array) {
+	imgShape := img.Shape()
+	B := imgShape[0]
+	Limg := imgShape[1]
+	D := imgShape[2]
+
+	txtShape := txt.Shape()
+	Ltxt := txtShape[1]
+
+	// === Image Q/K/V ===
+	imgFlat := mlx.Reshape(img, B*Limg, D)
+	qImg := mlx.Add(mlx.Linear(imgFlat, attn.ToQ), attn.ToQB)
+	kImg := mlx.Add(mlx.Linear(imgFlat, attn.ToK), attn.ToKB)
+	vImg := mlx.Add(mlx.Linear(imgFlat, attn.ToV), attn.ToVB)
+
+	qImg = mlx.Reshape(qImg, B, Limg, attn.NHeads, attn.HeadDim)
+	kImg = mlx.Reshape(kImg, B, Limg, attn.NHeads, attn.HeadDim)
+	vImg = mlx.Reshape(vImg, B, Limg, attn.NHeads, attn.HeadDim)
+
+	// QK norm (RMSNorm per head)
+	qImg = mlx.RMSNorm(qImg, attn.NormQ, 1e-6)
+	kImg = mlx.RMSNorm(kImg, attn.NormK, 1e-6)
+
+	// Apply RoPE
+	if imgFreqs != nil {
+		qImg = applyRoPE(qImg, imgFreqs)
+		kImg = applyRoPE(kImg, imgFreqs)
+	}
+
+	// === Text Q/K/V ===
+	txtFlat := mlx.Reshape(txt, B*Ltxt, D)
+	qTxt := mlx.Add(mlx.Linear(txtFlat, attn.AddQProj), attn.AddQProjB)
+	kTxt := mlx.Add(mlx.Linear(txtFlat, attn.AddKProj), attn.AddKProjB)
+	vTxt := mlx.Add(mlx.Linear(txtFlat, attn.AddVProj), attn.AddVProjB)
+
+	qTxt = mlx.Reshape(qTxt, B, Ltxt, attn.NHeads, attn.HeadDim)
+	kTxt = mlx.Reshape(kTxt, B, Ltxt, attn.NHeads, attn.HeadDim)
+	vTxt = mlx.Reshape(vTxt, B, Ltxt, attn.NHeads, attn.HeadDim)
+
+	qTxt = mlx.RMSNorm(qTxt, attn.NormAddQ, 1e-6)
+	kTxt = mlx.RMSNorm(kTxt, attn.NormAddK, 1e-6)
+
+	if txtFreqs != nil {
+		qTxt = applyRoPE(qTxt, txtFreqs)
+		kTxt = applyRoPE(kTxt, txtFreqs)
+	}
+
+	// Concatenate for joint attention: [txt, img] order
+	qJoint := mlx.Concatenate([]*mlx.Array{qTxt, qImg}, 1)
+	kJoint := mlx.Concatenate([]*mlx.Array{kTxt, kImg}, 1)
+	vJoint := mlx.Concatenate([]*mlx.Array{vTxt, vImg}, 1)
+
+	// Transpose to [B, nheads, L, head_dim]
+	qJoint = mlx.Transpose(qJoint, 0, 2, 1, 3)
+	kJoint = mlx.Transpose(kJoint, 0, 2, 1, 3)
+	vJoint = mlx.Transpose(vJoint, 0, 2, 1, 3)
+
+	// SDPA
+	outJoint := mlx.ScaledDotProductAttention(qJoint, kJoint, vJoint, attn.Scale, false)
+
+	// Transpose back and split
+	outJoint = mlx.Transpose(outJoint, 0, 2, 1, 3) // [B, L, nheads, head_dim]
+	outJoint = mlx.Reshape(outJoint, B, Ltxt+Limg, D)
+
+	outTxt := mlx.Slice(outJoint, []int32{0, 0, 0}, []int32{B, Ltxt, D})
+	outImg := mlx.Slice(outJoint, []int32{0, Ltxt, 0}, []int32{B, Ltxt + Limg, D})
+
+	// Output projections
+	outImg = mlx.Reshape(outImg, B*Limg, D)
+	outImg = mlx.Add(mlx.Linear(outImg, attn.ToOut), attn.ToOutB)
+	outImg = mlx.Reshape(outImg, B, Limg, D)
+
+	outTxt = mlx.Reshape(outTxt, B*Ltxt, D)
+	outTxt = mlx.Add(mlx.Linear(outTxt, attn.ToAddOut), attn.ToAddOutB)
+	outTxt = mlx.Reshape(outTxt, B, Ltxt, D)
+
+	return outImg, outTxt
+}
+
+// applyRoPE applies rotary embeddings using complex multiplication
+// x: [B, L, nheads, head_dim]
+// freqs: [L, head_dim] as complex (interleaved real/imag pairs)
+func applyRoPE(x *mlx.Array, freqs *mlx.Array) *mlx.Array {
+	shape := x.Shape()
+	B := shape[0]
+	L := shape[1]
+	nheads := shape[2]
+	headDim := shape[3]
+	halfDim := headDim / 2
+
+	// Reshape x to pairs: [B, L, nheads, half, 2]
+	xPairs := mlx.Reshape(x, B, L, nheads, halfDim, 2)
+
+	// freqs: [L, head_dim] -> [1, L, 1, half, 2]
+	freqsExp := mlx.Reshape(freqs, 1, L, 1, halfDim, 2)
+
+	// Extract real/imag parts
+	xReal := mlx.SliceStride(xPairs, []int32{0, 0, 0, 0, 0}, []int32{B, L, nheads, halfDim, 1}, []int32{1, 1, 1, 1, 1})
+	xImag := mlx.SliceStride(xPairs, []int32{0, 0, 0, 0, 1}, []int32{B, L, nheads, halfDim, 2}, []int32{1, 1, 1, 1, 1})
+	xReal = mlx.Squeeze(xReal, 4)
+	xImag = mlx.Squeeze(xImag, 4)
+
+	freqReal := mlx.SliceStride(freqsExp, []int32{0, 0, 0, 0, 0}, []int32{1, L, 1, halfDim, 1}, []int32{1, 1, 1, 1, 1})
+	freqImag := mlx.SliceStride(freqsExp, []int32{0, 0, 0, 0, 1}, []int32{1, L, 1, halfDim, 2}, []int32{1, 1, 1, 1, 1})
+	freqReal = mlx.Squeeze(freqReal, 4)
+	freqImag = mlx.Squeeze(freqImag, 4)
+
+	// Complex multiplication: (a + bi) * (c + di) = (ac - bd) + (ad + bc)i
+	outReal := mlx.Sub(mlx.Mul(xReal, freqReal), mlx.Mul(xImag, freqImag))
+	outImag := mlx.Add(mlx.Mul(xReal, freqImag), mlx.Mul(xImag, freqReal))
+
+	// Interleave back
+	outReal = mlx.ExpandDims(outReal, 4)
+	outImag = mlx.ExpandDims(outImag, 4)
+	out := mlx.Concatenate([]*mlx.Array{outReal, outImag}, 4)
+
+	return mlx.Reshape(out, B, L, nheads, headDim)
+}
+
+// MLP implements GELU MLP (not GEGLU)
+type MLP struct {
+	ProjWeight *mlx.Array
+	ProjBias   *mlx.Array
+	OutWeight  *mlx.Array
+	OutBias    *mlx.Array
+}
+
+// newMLP creates a GELU MLP
+func newMLP(weights *safetensors.ModelWeights, prefix string) (*MLP, error) {
+	projWeight, _ := weights.Get(prefix + ".net.0.proj.weight")
+	projBias, _ := weights.Get(prefix + ".net.0.proj.bias")
+	outWeight, _ := weights.Get(prefix + ".net.2.weight")
+	outBias, _ := weights.Get(prefix + ".net.2.bias")
+
+	return &MLP{
+		ProjWeight: mlx.Transpose(projWeight, 1, 0),
+		ProjBias:   projBias,
+		OutWeight:  mlx.Transpose(outWeight, 1, 0),
+		OutBias:    outBias,
+	}, nil
+}
+
+// Forward applies GELU MLP
+func (m *MLP) Forward(x *mlx.Array) *mlx.Array {
+	shape := x.Shape()
+	B := shape[0]
+	L := shape[1]
+	D := shape[2]
+
+	xFlat := mlx.Reshape(x, B*L, D)
+	h := mlx.Add(mlx.Linear(xFlat, m.ProjWeight), m.ProjBias)
+	h = geluApprox(h)
+	h = mlx.Add(mlx.Linear(h, m.OutWeight), m.OutBias)
+	return mlx.Reshape(h, B, L, m.OutBias.Dim(0))
+}
+
+// geluApprox implements approximate GELU
+func geluApprox(x *mlx.Array) *mlx.Array {
+	sqrt2OverPi := float32(math.Sqrt(2.0 / math.Pi))
+	x3 := mlx.Mul(mlx.Mul(x, x), x)
+	inner := mlx.Add(x, mlx.MulScalar(x3, 0.044715))
+	inner = mlx.MulScalar(inner, sqrt2OverPi)
+	return mlx.Mul(mlx.MulScalar(x, 0.5), mlx.AddScalar(mlx.Tanh(inner), 1.0))
+}
+
+// TransformerBlock is a single dual-stream transformer block
+type TransformerBlock struct {
+	Attention *JointAttention
+	ImgMLP    *MLP
+	TxtMLP    *MLP
+
+	ImgModWeight *mlx.Array
+	ImgModBias   *mlx.Array
+	TxtModWeight *mlx.Array
+	TxtModBias   *mlx.Array
+
+	HiddenDim int32
+	NormEps   float32
+}
+
+// newTransformerBlock creates a transformer block
+func newTransformerBlock(weights *safetensors.ModelWeights, prefix string, cfg *TransformerConfig) (*TransformerBlock, error) {
+	attn, err := newJointAttention(weights, prefix, cfg)
+	if err != nil {
+		return nil, err
+	}
+
+	imgMLP, _ := newMLP(weights, prefix+".img_mlp")
+	txtMLP, _ := newMLP(weights, prefix+".txt_mlp")
+
+	imgModWeight, _ := weights.Get(prefix + ".img_mod.1.weight")
+	imgModBias, _ := weights.Get(prefix + ".img_mod.1.bias")
+	txtModWeight, _ := weights.Get(prefix + ".txt_mod.1.weight")
+	txtModBias, _ := weights.Get(prefix + ".txt_mod.1.bias")
+
+	return &TransformerBlock{
+		Attention:    attn,
+		ImgMLP:       imgMLP,
+		TxtMLP:       txtMLP,
+		ImgModWeight: mlx.Transpose(imgModWeight, 1, 0),
+		ImgModBias:   imgModBias,
+		TxtModWeight: mlx.Transpose(txtModWeight, 1, 0),
+		TxtModBias:   txtModBias,
+		HiddenDim:    cfg.HiddenDim,
+		NormEps:      cfg.NormEps,
+	}, nil
+}
+
+// Forward applies the transformer block
+func (tb *TransformerBlock) Forward(img, txt, temb *mlx.Array, imgFreqs, txtFreqs *mlx.Array) (*mlx.Array, *mlx.Array) {
+	// Compute modulation: silu(temb) -> linear -> [B, 6*D]
+	siluT := mlx.SiLU(temb)
+	imgMod := mlx.Add(mlx.Linear(siluT, tb.ImgModWeight), tb.ImgModBias)
+	txtMod := mlx.Add(mlx.Linear(siluT, tb.TxtModWeight), tb.TxtModBias)
+
+	// Split into 6 parts: shift1, scale1, gate1, shift2, scale2, gate2
+	imgModParts := splitMod6(imgMod, tb.HiddenDim)
+	txtModParts := splitMod6(txtMod, tb.HiddenDim)
+
+	// Pre-attention: norm + modulate
+	imgNorm := layerNormNoAffine(img, tb.NormEps)
+	imgNorm = mlx.Add(mlx.Mul(imgNorm, mlx.AddScalar(imgModParts[1], 1.0)), imgModParts[0])
+
+	txtNorm := layerNormNoAffine(txt, tb.NormEps)
+	txtNorm = mlx.Add(mlx.Mul(txtNorm, mlx.AddScalar(txtModParts[1], 1.0)), txtModParts[0])
+
+	// Joint attention
+	attnImg, attnTxt := tb.Attention.Forward(imgNorm, txtNorm, imgFreqs, txtFreqs)
+
+	// Residual with gate
+	img = mlx.Add(img, mlx.Mul(imgModParts[2], attnImg))
+	txt = mlx.Add(txt, mlx.Mul(txtModParts[2], attnTxt))
+
+	// Pre-MLP: norm + modulate
+	imgNorm2 := layerNormNoAffine(img, tb.NormEps)
+	imgNorm2 = mlx.Add(mlx.Mul(imgNorm2, mlx.AddScalar(imgModParts[4], 1.0)), imgModParts[3])
+
+	txtNorm2 := layerNormNoAffine(txt, tb.NormEps)
+	txtNorm2 = mlx.Add(mlx.Mul(txtNorm2, mlx.AddScalar(txtModParts[4], 1.0)), txtModParts[3])
+
+	// MLP
+	mlpImg := tb.ImgMLP.Forward(imgNorm2)
+	mlpTxt := tb.TxtMLP.Forward(txtNorm2)
+
+	// Residual with gate
+	img = mlx.Add(img, mlx.Mul(imgModParts[5], mlpImg))
+	txt = mlx.Add(txt, mlx.Mul(txtModParts[5], mlpTxt))
+
+	return img, txt
+}
+
+// splitMod6 splits modulation into 6 parts each [B, 1, D]
+func splitMod6(mod *mlx.Array, hiddenDim int32) []*mlx.Array {
+	shape := mod.Shape()
+	B := shape[0]
+	parts := make([]*mlx.Array, 6)
+	for i := int32(0); i < 6; i++ {
+		part := mlx.Slice(mod, []int32{0, i * hiddenDim}, []int32{B, (i + 1) * hiddenDim})
+		parts[i] = mlx.ExpandDims(part, 1)
+	}
+	return parts
+}
+
+// layerNormNoAffine applies layer norm without learnable parameters
+func layerNormNoAffine(x *mlx.Array, eps float32) *mlx.Array {
+	ndim := x.Ndim()
+	lastAxis := ndim - 1
+	mean := mlx.Mean(x, lastAxis, true)
+	xCentered := mlx.Sub(x, mean)
+	variance := mlx.Mean(mlx.Square(xCentered), lastAxis, true)
+	return mlx.Div(xCentered, mlx.Sqrt(mlx.AddScalar(variance, eps)))
+}
+
+// Transformer is the full Qwen-Image transformer model
+type Transformer struct {
+	Config *TransformerConfig
+
+	ImgIn     *mlx.Array
+	ImgInBias *mlx.Array
+	TxtIn     *mlx.Array
+	TxtInBias *mlx.Array
+	TxtNorm   *mlx.Array
+
+	TEmbed *TimestepEmbedder
+	Layers []*TransformerBlock
+
+	NormOutWeight *mlx.Array
+	NormOutBias   *mlx.Array
+	ProjOut       *mlx.Array
+	ProjOutBias   *mlx.Array
+}
+
+// Load loads the transformer from a directory
+func (m *Transformer) Load(path string) error {
+	fmt.Println("Loading Qwen-Image transformer...")
+
+	cfg := defaultTransformerConfig()
+	m.Config = cfg
+
+	weights, err := safetensors.LoadModelWeights(path)
+	if err != nil {
+		return fmt.Errorf("weights: %w", err)
+	}
+
+	// Bulk load all weights as bf16
+	fmt.Print("  Loading weights as bf16... ")
+	if err := weights.Load(mlx.DtypeBFloat16); err != nil {
+		return fmt.Errorf("load weights: %w", err)
+	}
+	fmt.Printf("✓ (%.1f GB)\n", float64(mlx.MetalGetActiveMemory())/(1024*1024*1024))
+
+	fmt.Print("  Loading input projections... ")
+	imgIn, _ := weights.Get("img_in.weight")
+	imgInBias, _ := weights.Get("img_in.bias")
+	txtIn, _ := weights.Get("txt_in.weight")
+	txtInBias, _ := weights.Get("txt_in.bias")
+	txtNorm, _ := weights.Get("txt_norm.weight")
+	m.ImgIn = mlx.Transpose(imgIn, 1, 0)
+	m.ImgInBias = imgInBias
+	m.TxtIn = mlx.Transpose(txtIn, 1, 0)
+	m.TxtInBias = txtInBias
+	m.TxtNorm = txtNorm
+	fmt.Println("✓")
+
+	fmt.Print("  Loading timestep embedder... ")
+	m.TEmbed, err = newTimestepEmbedder(weights)
+	if err != nil {
+		return fmt.Errorf("timestep embedder: %w", err)
+	}
+	fmt.Println("✓")
+
+	m.Layers = make([]*TransformerBlock, cfg.NLayers)
+	for i := int32(0); i < cfg.NLayers; i++ {
+		fmt.Printf("\r  Loading transformer layers... %d/%d", i+1, cfg.NLayers)
+		prefix := fmt.Sprintf("transformer_blocks.%d", i)
+		m.Layers[i], err = newTransformerBlock(weights, prefix, cfg)
+		if err != nil {
+			return fmt.Errorf("layer %d: %w", i, err)
+		}
+	}
+	fmt.Printf("\r  Loading transformer layers... ✓ [%d blocks]          \n", cfg.NLayers)
+
+	fmt.Print("  Loading output layers... ")
+	normOutWeight, _ := weights.Get("norm_out.linear.weight")
+	normOutBias, _ := weights.Get("norm_out.linear.bias")
+	projOut, _ := weights.Get("proj_out.weight")
+	projOutBias, _ := weights.Get("proj_out.bias")
+	m.NormOutWeight = mlx.Transpose(normOutWeight, 1, 0)
+	m.NormOutBias = normOutBias
+	m.ProjOut = mlx.Transpose(projOut, 1, 0)
+	m.ProjOutBias = projOutBias
+	fmt.Println("✓")
+
+	weights.ReleaseAll()
+	return nil
+}
+
+// LoadFromPath is a convenience function to load transformer from path
+func LoadTransformerFromPath(path string) (*Transformer, error) {
+	m := &Transformer{}
+	if err := m.Load(filepath.Join(path, "transformer")); err != nil {
+		return nil, err
+	}
+	return m, nil
+}
+
+// Forward runs the transformer
+// img: [B, L_img, in_channels] patchified latents
+// txt: [B, L_txt, joint_attention_dim] text embeddings
+// t: [B] timesteps (0-1)
+// imgFreqs, txtFreqs: RoPE frequencies
+func (tr *Transformer) Forward(img, txt, t *mlx.Array, imgFreqs, txtFreqs *mlx.Array) *mlx.Array {
+	imgShape := img.Shape()
+	B := imgShape[0]
+	Limg := imgShape[1]
+
+	txtShape := txt.Shape()
+	Ltxt := txtShape[1]
+
+	// Timestep embedding
+	temb := tr.TEmbed.Forward(t)
+
+	// Project image: [B, L, in_channels] -> [B, L, hidden_dim]
+	imgFlat := mlx.Reshape(img, B*Limg, tr.Config.InChannels)
+	imgH := mlx.Add(mlx.Linear(imgFlat, tr.ImgIn), tr.ImgInBias)
+	imgH = mlx.Reshape(imgH, B, Limg, tr.Config.HiddenDim)
+
+	// Project text: RMSNorm then linear
+	txtFlat := mlx.Reshape(txt, B*Ltxt, tr.Config.JointAttentionDim)
+	txtNormed := mlx.RMSNorm(txtFlat, tr.TxtNorm, 1e-6)
+	txtH := mlx.Add(mlx.Linear(txtNormed, tr.TxtIn), tr.TxtInBias)
+	txtH = mlx.Reshape(txtH, B, Ltxt, tr.Config.HiddenDim)
+
+	for _, layer := range tr.Layers {
+		imgH, txtH = layer.Forward(imgH, txtH, temb, imgFreqs, txtFreqs)
+	}
+
+	// Final norm with modulation (AdaLayerNormContinuous)
+	// Python: scale, shift = torch.chunk(emb, 2, dim=1)
+	finalMod := mlx.Add(mlx.Linear(mlx.SiLU(temb), tr.NormOutWeight), tr.NormOutBias)
+	modShape := finalMod.Shape()
+	halfDim := modShape[1] / 2
+	scale := mlx.ExpandDims(mlx.Slice(finalMod, []int32{0, 0}, []int32{B, halfDim}), 1)
+	shift := mlx.ExpandDims(mlx.Slice(finalMod, []int32{0, halfDim}, []int32{B, modShape[1]}), 1)
+
+	imgH = layerNormNoAffine(imgH, tr.Config.NormEps)
+	imgH = mlx.Add(mlx.Mul(imgH, mlx.AddScalar(scale, 1.0)), shift)
+
+	// Final projection: [B, L, hidden_dim] -> [B, L, patch_size^2 * out_channels]
+	imgFlat = mlx.Reshape(imgH, B*Limg, tr.Config.HiddenDim)
+	out := mlx.Add(mlx.Linear(imgFlat, tr.ProjOut), tr.ProjOutBias)
+
+	outChannels := tr.Config.PatchSize * tr.Config.PatchSize * tr.Config.OutChannels
+	return mlx.Reshape(out, B, Limg, outChannels)
+}
+
+// ForwardWithCache runs the transformer with layer caching for speedup.
+// Based on DeepCache (CVPR 2024) / Learning-to-Cache (NeurIPS 2024):
+// shallow layers change little between denoising steps, so we cache their
+// outputs and reuse them on non-refresh steps.
+//
+// stepCache: cache for layer outputs (use cache.NewStepCache(cacheLayers))
+// step: current denoising step (0-indexed)
+// cacheInterval: refresh cache every N steps (e.g., 3)
+// cacheLayers: number of shallow layers to cache (e.g., 15)
+func (tr *Transformer) ForwardWithCache(
+	img, txt, t *mlx.Array,
+	imgFreqs, txtFreqs *mlx.Array,
+	stepCache *cache.StepCache,
+	step, cacheInterval, cacheLayers int,
+) *mlx.Array {
+	imgShape := img.Shape()
+	B := imgShape[0]
+	Limg := imgShape[1]
+
+	txtShape := txt.Shape()
+	Ltxt := txtShape[1]
+
+	// Timestep embedding
+	temb := tr.TEmbed.Forward(t)
+
+	// Project image: [B, L, in_channels] -> [B, L, hidden_dim]
+	imgFlat := mlx.Reshape(img, B*Limg, tr.Config.InChannels)
+	imgH := mlx.Add(mlx.Linear(imgFlat, tr.ImgIn), tr.ImgInBias)
+	imgH = mlx.Reshape(imgH, B, Limg, tr.Config.HiddenDim)
+
+	// Project text: RMSNorm then linear
+	txtFlat := mlx.Reshape(txt, B*Ltxt, tr.Config.JointAttentionDim)
+	txtNormed := mlx.RMSNorm(txtFlat, tr.TxtNorm, 1e-6)
+	txtH := mlx.Add(mlx.Linear(txtNormed, tr.TxtIn), tr.TxtInBias)
+	txtH = mlx.Reshape(txtH, B, Ltxt, tr.Config.HiddenDim)
+
+	// Check if we should refresh the cache
+	refreshCache := stepCache.ShouldRefresh(step, cacheInterval)
+
+	for i, layer := range tr.Layers {
+		if i < cacheLayers && !refreshCache && stepCache.Get(i) != nil {
+			// Use cached outputs for shallow layers
+			imgH = stepCache.Get(i)
+			txtH = stepCache.Get2(i)
+		} else {
+			// Compute layer
+			imgH, txtH = layer.Forward(imgH, txtH, temb, imgFreqs, txtFreqs)
+			// Cache shallow layers on refresh steps
+			if i < cacheLayers && refreshCache {
+				stepCache.Set(i, imgH)
+				stepCache.Set2(i, txtH)
+			}
+		}
+	}
+
+	// Final norm with modulation (AdaLayerNormContinuous)
+	finalMod := mlx.Add(mlx.Linear(mlx.SiLU(temb), tr.NormOutWeight), tr.NormOutBias)
+	modShape := finalMod.Shape()
+	halfDim := modShape[1] / 2
+	scale := mlx.ExpandDims(mlx.Slice(finalMod, []int32{0, 0}, []int32{B, halfDim}), 1)
+	shift := mlx.ExpandDims(mlx.Slice(finalMod, []int32{0, halfDim}, []int32{B, modShape[1]}), 1)
+
+	imgH = layerNormNoAffine(imgH, tr.Config.NormEps)
+	imgH = mlx.Add(mlx.Mul(imgH, mlx.AddScalar(scale, 1.0)), shift)
+
+	// Final projection: [B, L, hidden_dim] -> [B, L, patch_size^2 * out_channels]
+	imgFlat = mlx.Reshape(imgH, B*Limg, tr.Config.HiddenDim)
+	out := mlx.Add(mlx.Linear(imgFlat, tr.ProjOut), tr.ProjOutBias)
+
+	outChannels := tr.Config.PatchSize * tr.Config.PatchSize * tr.Config.OutChannels
+	return mlx.Reshape(out, B, Limg, outChannels)
+}
+
+// RoPECache holds precomputed RoPE frequencies
+type RoPECache struct {
+	ImgFreqs *mlx.Array // [L_img, head_dim]
+	TxtFreqs *mlx.Array // [L_txt, head_dim]
+}
+
+// PrepareRoPE computes RoPE for image and text sequences
+// This matches Python's QwenEmbedRope with scale_rope=True
+func PrepareRoPE(imgH, imgW int32, txtLen int32, axesDims []int32) *RoPECache {
+	theta := float64(10000)
+	maxIdx := int32(4096)
+
+	// Compute base frequencies for each axis dimension
+	freqsT := ComputeAxisFreqs(axesDims[0], theta)
+	freqsH := ComputeAxisFreqs(axesDims[1], theta)
+	freqsW := ComputeAxisFreqs(axesDims[2], theta)
+
+	// Build frequency lookup tables
+	posFreqsT := MakeFreqTable(maxIdx, freqsT, false)
+	posFreqsH := MakeFreqTable(maxIdx, freqsH, false)
+	posFreqsW := MakeFreqTable(maxIdx, freqsW, false)
+	negFreqsH := MakeFreqTable(maxIdx, freqsH, true)
+	negFreqsW := MakeFreqTable(maxIdx, freqsW, true)
+
+	// Image frequencies with scale_rope=True
+	imgLen := imgH * imgW
+	headDim := int32(len(freqsT)+len(freqsH)+len(freqsW)) * 2
+	imgFreqsData := make([]float32, imgLen*headDim)
+
+	hHalf := imgH / 2
+	wHalf := imgW / 2
+
+	idx := int32(0)
+	for y := int32(0); y < imgH; y++ {
+		for x := int32(0); x < imgW; x++ {
+			// Frame = 0
+			for i := 0; i < len(freqsT)*2; i++ {
+				imgFreqsData[idx+int32(i)] = posFreqsT[0][i]
+			}
+			idx += int32(len(freqsT) * 2)
+
+			// Height: scale_rope pattern
+			hNegCount := imgH - hHalf
+			if y < hNegCount {
+				negTableIdx := maxIdx - hNegCount + y
+				for i := 0; i < len(freqsH)*2; i++ {
+					imgFreqsData[idx+int32(i)] = negFreqsH[negTableIdx][i]
+				}
+			} else {
+				posIdx := y - hNegCount
+				for i := 0; i < len(freqsH)*2; i++ {
+					imgFreqsData[idx+int32(i)] = posFreqsH[posIdx][i]
+				}
+			}
+			idx += int32(len(freqsH) * 2)
+
+			// Width: scale_rope pattern
+			wNegCount := imgW - wHalf
+			if x < wNegCount {
+				negTableIdx := maxIdx - wNegCount + x
+				for i := 0; i < len(freqsW)*2; i++ {
+					imgFreqsData[idx+int32(i)] = negFreqsW[negTableIdx][i]
+				}
+			} else {
+				posIdx := x - wNegCount
+				for i := 0; i < len(freqsW)*2; i++ {
+					imgFreqsData[idx+int32(i)] = posFreqsW[posIdx][i]
+				}
+			}
+			idx += int32(len(freqsW) * 2)
+		}
+	}
+
+	imgFreqs := mlx.NewArray(imgFreqsData, []int32{imgLen, headDim})
+	imgFreqs = mlx.ToBFloat16(imgFreqs)
+
+	// Text frequencies
+	maxVidIdx := max(hHalf, wHalf)
+	txtFreqsData := make([]float32, txtLen*headDim)
+
+	idx = 0
+	for t := int32(0); t < txtLen; t++ {
+		pos := maxVidIdx + t
+		for i := 0; i < len(freqsT)*2; i++ {
+			txtFreqsData[idx+int32(i)] = posFreqsT[pos][i]
+		}
+		idx += int32(len(freqsT) * 2)
+		for i := 0; i < len(freqsH)*2; i++ {
+			txtFreqsData[idx+int32(i)] = posFreqsH[pos][i]
+		}
+		idx += int32(len(freqsH) * 2)
+		for i := 0; i < len(freqsW)*2; i++ {
+			txtFreqsData[idx+int32(i)] = posFreqsW[pos][i]
+		}
+		idx += int32(len(freqsW) * 2)
+	}
+
+	txtFreqs := mlx.NewArray(txtFreqsData, []int32{txtLen, headDim})
+	txtFreqs = mlx.ToBFloat16(txtFreqs)
+
+	return &RoPECache{
+		ImgFreqs: imgFreqs,
+		TxtFreqs: txtFreqs,
+	}
+}
+
+// ComputeAxisFreqs computes RoPE base frequencies for a given dimension.
+func ComputeAxisFreqs(dim int32, theta float64) []float64 {
+	halfDim := dim / 2
+	freqs := make([]float64, halfDim)
+	for i := int32(0); i < halfDim; i++ {
+		freqs[i] = 1.0 / math.Pow(theta, float64(i)/float64(halfDim))
+	}
+	return freqs
+}
+
+// MakeFreqTable builds a table of cos/sin values for RoPE positions.
+func MakeFreqTable(maxIdx int32, baseFreqs []float64, negative bool) [][]float32 {
+	table := make([][]float32, maxIdx)
+	for idx := int32(0); idx < maxIdx; idx++ {
+		var pos float64
+		if negative {
+			pos = float64(-maxIdx + int32(idx))
+		} else {
+			pos = float64(idx)
+		}
+
+		row := make([]float32, len(baseFreqs)*2)
+		for i, f := range baseFreqs {
+			angle := pos * f
+			row[i*2] = float32(math.Cos(angle))
+			row[i*2+1] = float32(math.Sin(angle))
+		}
+		table[idx] = row
+	}
+	return table
+}
+
+func max(a, b int32) int32 {
+	if a > b {
+		return a
+	}
+	return b
+}
+
+// PackLatents converts [B, C, H, W] to [B, L, C*4] patches
+func PackLatents(latents *mlx.Array, patchSize int32) *mlx.Array {
+	shape := latents.Shape()
+	B := shape[0]
+	C := shape[1]
+	H := shape[2]
+	W := shape[3]
+
+	pH := H / patchSize
+	pW := W / patchSize
+
+	// [B, C, H, W] -> [B, C, pH, 2, pW, 2]
+	x := mlx.Reshape(latents, B, C, pH, patchSize, pW, patchSize)
+	// -> [B, pH, pW, C, 2, 2]
+	x = mlx.Transpose(x, 0, 2, 4, 1, 3, 5)
+	// -> [B, pH*pW, C*4]
+	return mlx.Reshape(x, B, pH*pW, C*patchSize*patchSize)
+}
+
+// UnpackLatents converts [B, L, C*4] back to [B, C, 1, H, W] (5D for VAE)
+func UnpackLatents(patches *mlx.Array, H, W, patchSize int32) *mlx.Array {
+	shape := patches.Shape()
+	B := shape[0]
+	channels := shape[2] / (patchSize * patchSize)
+
+	pH := H / patchSize
+	pW := W / patchSize
+
+	// [B, L, C*4] -> [B, pH, pW, C, 2, 2]
+	x := mlx.Reshape(patches, B, pH, pW, channels, patchSize, patchSize)
+	// -> [B, C, pH, 2, pW, 2]
+	x = mlx.Transpose(x, 0, 3, 1, 4, 2, 5)
+	// -> [B, C, H, W]
+	x = mlx.Reshape(x, B, channels, pH*patchSize, pW*patchSize)
+	// Add temporal dimension for VAE: [B, C, 1, H, W]
+	return mlx.ExpandDims(x, 2)
+}

x/imagegen/models/qwen_image/transformer_test.go

+//go:build mlx
+
+package qwen_image
+
+import (
+	"math"
+	"os"
+	"testing"
+
+	"github.com/ollama/ollama/x/imagegen/mlx"
+)
+
+// TestTransformerConfig tests configuration invariants.
+func TestTransformerConfig(t *testing.T) {
+	cfg := defaultTransformerConfig()
+
+	// Property: hidden_dim = n_heads * head_dim
+	if cfg.HiddenDim != cfg.NHeads*cfg.HeadDim {
+		t.Errorf("hidden_dim != n_heads * head_dim: %d != %d * %d",
+			cfg.HiddenDim, cfg.NHeads, cfg.HeadDim)
+	}
+
+	// Property: axes_dims_rope sums to head_dim
+	var ropeSum int32
+	for _, d := range cfg.AxesDimsRope {
+		ropeSum += d
+	}
+	if ropeSum != cfg.HeadDim {
+		t.Errorf("axes_dims_rope sum != head_dim: %d != %d", ropeSum, cfg.HeadDim)
+	}
+
+	// Property: in_channels = out_channels * patch_size^2
+	expectedIn := cfg.OutChannels * cfg.PatchSize * cfg.PatchSize
+	if cfg.InChannels != expectedIn {
+		t.Errorf("in_channels != out_channels * patch_size^2: %d != %d", cfg.InChannels, expectedIn)
+	}
+}
+
+// TestTransformerRoPE tests RoPE frequency computation produces valid values.
+func TestTransformerRoPE(t *testing.T) {
+	cfg := defaultTransformerConfig()
+
+	// Test with small image dimensions
+	imgH, imgW := int32(4), int32(4) // 4x4 latent = 16 patches
+	txtLen := int32(5)
+
+	ropeCache := PrepareRoPE(imgH, imgW, txtLen, cfg.AxesDimsRope)
+	mlx.Eval(ropeCache.ImgFreqs, ropeCache.TxtFreqs)
+
+	// Verify shapes: [seq_len, head_dim]
+	imgSeqLen := imgH * imgW
+	if ropeCache.ImgFreqs.Shape()[0] != imgSeqLen {
+		t.Errorf("ImgFreqs seq_len: got %d, want %d", ropeCache.ImgFreqs.Shape()[0], imgSeqLen)
+	}
+	if ropeCache.ImgFreqs.Shape()[1] != cfg.HeadDim {
+		t.Errorf("ImgFreqs head_dim: got %d, want %d", ropeCache.ImgFreqs.Shape()[1], cfg.HeadDim)
+	}
+
+	if ropeCache.TxtFreqs.Shape()[0] != txtLen {
+		t.Errorf("TxtFreqs seq_len: got %d, want %d", ropeCache.TxtFreqs.Shape()[0], txtLen)
+	}
+
+	// Verify values are finite
+	imgData := ropeCache.ImgFreqs.Data()
+	for i := 0; i < min(100, len(imgData)); i++ {
+		if math.IsNaN(float64(imgData[i])) || math.IsInf(float64(imgData[i]), 0) {
+			t.Errorf("ImgFreqs[%d] not finite: %v", i, imgData[i])
+			break
+		}
+	}
+}
+
+// TestTransformerForward tests full forward pass (integration test).
+// Skips if model weights are not available.
+func TestTransformerForward(t *testing.T) {
+	weightsPath := "../../../weights/Qwen-Image-2512/transformer"
+	if _, err := os.Stat(weightsPath); os.IsNotExist(err) {
+		t.Skip("Skipping: model weights not found at " + weightsPath)
+	}
+
+	transformer := &Transformer{}
+	if err := transformer.Load(weightsPath); err != nil {
+		t.Fatalf("Failed to load transformer: %v", err)
+	}
+	mlx.Keep(mlx.Collect(transformer)...)
+	cfg := transformer.Config
+
+	// Small test inputs
+	batchSize := int32(1)
+	imgH, imgW := int32(4), int32(4)
+	imgSeqLen := imgH * imgW
+	txtSeqLen := int32(5)
+
+	hiddenStates := mlx.RandomNormal([]int32{batchSize, imgSeqLen, cfg.InChannels}, 0)
+	encoderHiddenStates := mlx.RandomNormal([]int32{batchSize, txtSeqLen, cfg.JointAttentionDim}, 0)
+	timestep := mlx.NewArray([]float32{0.5}, []int32{batchSize})
+
+	ropeCache := PrepareRoPE(imgH, imgW, txtSeqLen, cfg.AxesDimsRope)
+
+	// Forward pass
+	out := transformer.Forward(hiddenStates, encoderHiddenStates, timestep, ropeCache.ImgFreqs, ropeCache.TxtFreqs)
+	mlx.Eval(out)
+
+	// Verify output shape: [batch, img_seq_len, in_channels]
+	wantShape := []int32{batchSize, imgSeqLen, cfg.InChannels}
+	gotShape := out.Shape()
+	if gotShape[0] != wantShape[0] || gotShape[1] != wantShape[1] || gotShape[2] != wantShape[2] {
+		t.Errorf("output shape: got %v, want %v", gotShape, wantShape)
+	}
+
+	// Verify output is finite
+	outData := out.Data()
+	for i := 0; i < min(100, len(outData)); i++ {
+		if math.IsNaN(float64(outData[i])) || math.IsInf(float64(outData[i]), 0) {
+			t.Errorf("output[%d] not finite: %v", i, outData[i])
+			break
+		}
+	}
+}

x/imagegen/models/qwen_image/vae.go

+//go:build mlx
+
+package qwen_image
+
+import (
+	"fmt"
+	"math"
+	"path/filepath"
+
+	"github.com/ollama/ollama/x/imagegen/mlx"
+	"github.com/ollama/ollama/x/imagegen/safetensors"
+)
+
+// VAEConfig holds Qwen-Image VAE configuration
+type VAEConfig struct {
+	ZDim               int32     `json:"z_dim"`               // 16
+	BaseDim            int32     `json:"base_dim"`            // 96
+	DimMult            []int32   `json:"dim_mult"`            // [1, 2, 4, 4]
+	NumResBlocks       int32     `json:"num_res_blocks"`      // 2
+	LatentsMean        []float32 `json:"latents_mean"`        // 16 values
+	LatentsStd         []float32 `json:"latents_std"`         // 16 values
+	TemperalDownsample []bool    `json:"temperal_downsample"` // [false, true, true]
+}
+
+// defaultVAEConfig returns config for Qwen-Image VAE
+func defaultVAEConfig() *VAEConfig {
+	return &VAEConfig{
+		ZDim:         16,
+		BaseDim:      96,
+		DimMult:      []int32{1, 2, 4, 4},
+		NumResBlocks: 2,
+		LatentsMean: []float32{
+			-0.7571, -0.7089, -0.9113, 0.1075,
+			-0.1745, 0.9653, -0.1517, 1.5508,
+			0.4134, -0.0715, 0.5517, -0.3632,
+			-0.1922, -0.9497, 0.2503, -0.2921,
+		},
+		LatentsStd: []float32{
+			2.8184, 1.4541, 2.3275, 2.6558,
+			1.2196, 1.7708, 2.6052, 2.0743,
+			3.2687, 2.1526, 2.8652, 1.5579,
+			1.6382, 1.1253, 2.8251, 1.916,
+		},
+		TemperalDownsample: []bool{false, true, true},
+	}
+}
+
+// CausalConv3d is a causal 3D convolution (for temporal causality)
+type CausalConv3d struct {
+	Weight       *mlx.Array
+	Bias         *mlx.Array
+	BiasReshaped *mlx.Array // [1, C, 1, 1, 1]
+	KernelT      int32
+}
+
+// newCausalConv3d creates a 3D causal conv
+func newCausalConv3d(weights *safetensors.ModelWeights, prefix string) (*CausalConv3d, error) {
+	weight, err := weights.Get(prefix + ".weight")
+	if err != nil {
+		return nil, fmt.Errorf("weight not found: %s", prefix)
+	}
+	bias, _ := weights.Get(prefix + ".bias")
+
+	kernelT := weight.Shape()[2]
+	outC := weight.Shape()[0]
+
+	var biasReshaped *mlx.Array
+	if bias != nil {
+		biasReshaped = mlx.Reshape(bias, 1, outC, 1, 1, 1)
+	}
+
+	return &CausalConv3d{
+		Weight:       weight,
+		Bias:         bias,
+		BiasReshaped: biasReshaped,
+		KernelT:      kernelT,
+	}, nil
+}
+
+// Forward applies causal 3D convolution
+// x: [B, T, H, W, C] (channels-last, MLX format)
+func (c *CausalConv3d) Forward(x *mlx.Array) *mlx.Array {
+	shape := c.Weight.Shape() // PyTorch format: [O, I, kT, kH, kW]
+	kernelT := shape[2]
+	kernelH := shape[3]
+	kernelW := shape[4]
+
+	// Causal temporal padding, same spatial padding
+	// Input is channels-last: [B, T, H, W, C]
+	padT := kernelT - 1
+	padH := kernelH / 2
+	padW := kernelW / 2
+
+	// Stage 1: Pad
+	{
+			x = pad3DChannelsLast(x, padT, 0, padH, padH, padW, padW)
+		mlx.Eval(x)
+	}
+
+	// Stage 2: Conv + bias
+	var out *mlx.Array
+	{
+			prev := x
+		weight := mlx.Transpose(c.Weight, 0, 2, 3, 4, 1)
+		out = mlx.Conv3d(x, weight, 1, 1, 1, 0, 0, 0)
+		if c.Bias != nil {
+			bias := mlx.Reshape(c.Bias, 1, 1, 1, 1, c.Bias.Dim(0))
+			out = mlx.Add(out, bias)
+		}
+		prev.Free()
+		mlx.Eval(out)
+	}
+
+	return out
+}
+
+// RMSNorm3D applies RMS normalization over channels
+// Works with channels-last [B, T, H, W, C] format
+type RMSNorm3D struct {
+	Gamma *mlx.Array // [1, 1, 1, 1, C] for broadcasting
+}
+
+// newRMSNorm3D creates an RMS norm
+func newRMSNorm3D(weights *safetensors.ModelWeights, prefix string, dim int32) (*RMSNorm3D, error) {
+	gamma, err := weights.Get(prefix + ".gamma")
+	if err != nil {
+		return nil, err
+	}
+	// Reshape for channels-last broadcasting: [1, 1, 1, 1, C]
+	gamma = mlx.Reshape(gamma, 1, 1, 1, 1, gamma.Dim(0))
+	return &RMSNorm3D{Gamma: gamma}, nil
+}
+
+// Forward applies RMS norm to channels-last input [B, T, H, W, C]
+func (n *RMSNorm3D) Forward(x *mlx.Array) *mlx.Array {
+	// RMSNorm: x * rsqrt(mean(x^2) + eps) * gamma
+	normalized := mlx.RMSNormNoWeight(x, 1e-6)
+	return mlx.Mul(normalized, n.Gamma)
+}
+
+// ResBlock is a residual block with RMS norm and causal convs
+type ResBlock struct {
+	Norm1    *RMSNorm3D
+	Conv1    *CausalConv3d
+	Norm2    *RMSNorm3D
+	Conv2    *CausalConv3d
+	Shortcut *CausalConv3d
+}
+
+// newResBlock creates a residual block
+func newResBlock(weights *safetensors.ModelWeights, prefix string, inDim, outDim int32) (*ResBlock, error) {
+	norm1, err := newRMSNorm3D(weights, prefix+".norm1", inDim)
+	if err != nil {
+		return nil, err
+	}
+	conv1, err := newCausalConv3d(weights, prefix+".conv1")
+	if err != nil {
+		return nil, err
+	}
+	norm2, err := newRMSNorm3D(weights, prefix+".norm2", outDim)
+	if err != nil {
+		return nil, err
+	}
+	conv2, err := newCausalConv3d(weights, prefix+".conv2")
+	if err != nil {
+		return nil, err
+	}
+
+	var shortcut *CausalConv3d
+	if inDim != outDim {
+		shortcut, err = newCausalConv3d(weights, prefix+".conv_shortcut")
+		if err != nil {
+			return nil, err
+		}
+	}
+
+	return &ResBlock{
+		Norm1:    norm1,
+		Conv1:    conv1,
+		Norm2:    norm2,
+		Conv2:    conv2,
+		Shortcut: shortcut,
+	}, nil
+}
+
+// Forward applies the residual block
+func (r *ResBlock) Forward(x *mlx.Array) *mlx.Array {
+	// Use h as working variable, keep x intact for residual (caller will free x)
+	// Conv handles its own pools, so we just need pools for non-conv operations
+	var h *mlx.Array
+
+	// Keep x so it survives Eval() cleanup - needed for residual connection
+	mlx.Keep(x)
+
+	// Stage 1: norm1 + silu
+	{
+			h = r.Norm1.Forward(x)
+		h = silu3D(h)
+		mlx.Eval(h)
+	}
+
+	// Stage 2: conv1 (handles its own pools)
+	{
+		prev := h
+		h = r.Conv1.Forward(h)
+		prev.Free()
+	}
+
+	// Stage 3: norm2 + silu
+	{
+			prev := h
+		h = r.Norm2.Forward(h)
+		h = silu3D(h)
+		prev.Free()
+		mlx.Eval(h)
+	}
+
+	// Stage 4: conv2 (handles its own pools)
+	{
+		prev := h
+		h = r.Conv2.Forward(h)
+		prev.Free()
+	}
+
+	// Residual connection (shortcut handles its own pools if present)
+	if r.Shortcut != nil {
+		shortcut := r.Shortcut.Forward(x)
+			h = mlx.Add(h, shortcut)
+		mlx.Eval(h)
+	} else {
+			h = mlx.Add(h, x)
+		mlx.Eval(h)
+	}
+
+	return h
+}
+
+// AttentionBlock is a 2D attention block
+type AttentionBlock struct {
+	Norm      *RMSNorm3D
+	ToQKV     *mlx.Array
+	ToQKVBias *mlx.Array
+	Proj      *mlx.Array
+	ProjBias  *mlx.Array
+	Dim       int32
+}
+
+// newAttentionBlock creates an attention block
+func newAttentionBlock(weights *safetensors.ModelWeights, prefix string, dim int32) (*AttentionBlock, error) {
+	norm, err := newRMSNorm3D(weights, prefix+".norm", dim)
+	if err != nil {
+		return nil, err
+	}
+	toQKV, _ := weights.Get(prefix + ".to_qkv.weight")
+	toQKVBias, _ := weights.Get(prefix + ".to_qkv.bias")
+	proj, _ := weights.Get(prefix + ".proj.weight")
+	projBias, _ := weights.Get(prefix + ".proj.bias")
+
+	return &AttentionBlock{
+		Norm:      norm,
+		ToQKV:     toQKV,
+		ToQKVBias: toQKVBias,
+		Proj:      proj,
+		ProjBias:  projBias,
+		Dim:       dim,
+	}, nil
+}
+
+// Forward applies 2D attention
+// Input: [B, T, H, W, C] (channels-last)
+func (a *AttentionBlock) Forward(x *mlx.Array) *mlx.Array {
+	shape := x.Shape()
+	B := shape[0]
+	T := shape[1]
+	H := shape[2]
+	W := shape[3]
+	C := shape[4]
+
+	identity := x
+
+	// Flatten to [B*T, 1, H, W, C] for norm
+	x = mlx.Reshape(x, B*T, 1, H, W, C)
+	x = a.Norm.Forward(x)
+	x = mlx.Reshape(x, B*T, H, W, C)
+
+	// Flatten spatial to [B*T, H*W, C]
+	x = mlx.Reshape(x, B*T, H*W, C)
+
+	// Linear to get Q, K, V: [B*T, H*W, 3*C]
+	// Weight is [outC, inC] or [outC, inC, 1, 1]
+	wShape := a.ToQKV.Shape()
+	var w *mlx.Array
+	if len(wShape) == 4 {
+		w = mlx.Reshape(a.ToQKV, wShape[0], wShape[1])
+	} else {
+		w = a.ToQKV
+	}
+	w = mlx.Transpose(w, 1, 0) // [inC, outC]
+
+	qkv := mlx.Linear(x, w) // [B*T, H*W, 3*C]
+	if a.ToQKVBias != nil {
+		qkv = mlx.Add(qkv, a.ToQKVBias)
+	}
+	qkv = mlx.Reshape(qkv, B*T, 1, H*W, 3*C)
+
+	q := mlx.Slice(qkv, []int32{0, 0, 0, 0}, []int32{B * T, 1, H * W, C})
+	k := mlx.Slice(qkv, []int32{0, 0, 0, C}, []int32{B * T, 1, H * W, 2 * C})
+	v := mlx.Slice(qkv, []int32{0, 0, 0, 2 * C}, []int32{B * T, 1, H * W, 3 * C})
+
+	scale := float32(1.0 / math.Sqrt(float64(C)))
+	out := mlx.ScaledDotProductAttention(q, k, v, scale, false)
+
+	// out: [B*T, 1, H*W, C]
+	out = mlx.Reshape(out, B*T, H*W, C)
+
+	// Project back
+	pShape := a.Proj.Shape()
+	var p *mlx.Array
+	if len(pShape) == 4 {
+		p = mlx.Reshape(a.Proj, pShape[0], pShape[1])
+	} else {
+		p = a.Proj
+	}
+	p = mlx.Transpose(p, 1, 0) // [inC, outC]
+	out = mlx.Linear(out, p) // [B*T, H*W, C]
+	if a.ProjBias != nil {
+		out = mlx.Add(out, a.ProjBias)
+	}
+
+	out = mlx.Reshape(out, B, T, H, W, C)
+	return mlx.Add(out, identity)
+}
+
+// UpBlock handles upsampling in decoder
+type UpBlock struct {
+	ResBlocks []*ResBlock
+	Upsampler *Upsample
+}
+
+// newUpBlock creates an up block
+func newUpBlock(weights *safetensors.ModelWeights, prefix string, inDim, outDim int32, numBlocks int32, upsampleMode string) (*UpBlock, error) {
+	resBlocks := make([]*ResBlock, numBlocks+1)
+
+	currentDim := inDim
+	for i := int32(0); i <= numBlocks; i++ {
+		resPrefix := fmt.Sprintf("%s.resnets.%d", prefix, i)
+		block, err := newResBlock(weights, resPrefix, currentDim, outDim)
+		if err != nil {
+			return nil, err
+		}
+		resBlocks[i] = block
+		currentDim = outDim
+	}
+
+	var upsampler *Upsample
+	if upsampleMode != "" {
+		upsampler = newUpsample(weights, prefix+".upsamplers.0", outDim, upsampleMode)
+	}
+
+	return &UpBlock{
+		ResBlocks: resBlocks,
+		Upsampler: upsampler,
+	}, nil
+}
+
+// Forward applies up block with staged memory management
+func (u *UpBlock) Forward(x *mlx.Array) *mlx.Array {
+	// ResBlocks handle their own pools
+	for _, block := range u.ResBlocks {
+		prev := x
+		x = block.Forward(x)
+		prev.Free()
+	}
+
+	// Upsampler handles its own pools
+	if u.Upsampler != nil {
+		prev := x
+		x = u.Upsampler.Forward(x)
+		prev.Free()
+	}
+	return x
+}
+
+// Upsample handles spatial upsampling
+type Upsample struct {
+	Conv *mlx.Array
+	Bias *mlx.Array
+	Mode string
+}
+
+// newUpsample creates an upsampler
+func newUpsample(weights *safetensors.ModelWeights, prefix string, dim int32, mode string) *Upsample {
+	conv, _ := weights.Get(prefix + ".resample.1.weight")
+	bias, _ := weights.Get(prefix + ".resample.1.bias")
+	return &Upsample{
+		Conv: conv,
+		Bias: bias,
+		Mode: mode,
+	}
+}
+
+// Forward applies upsampling to channels-last input [B, T, H, W, C]
+// Uses staged pools to reduce peak memory during 2x upsampling
+func (u *Upsample) Forward(x *mlx.Array) *mlx.Array {
+	shape := x.Shape()
+	B := shape[0]
+	T := shape[1]
+	H := shape[2]
+	W := shape[3]
+	C := shape[4]
+	outC := u.Conv.Shape()[0]
+
+	// Stage 1: 2x nearest neighbor upsample
+	{
+			x = mlx.Reshape(x, B*T, H, W, C)
+		x = upsample2xChannelsLast(x)
+		mlx.Eval(x)
+	}
+
+	// Stage 2: Conv + bias
+	{
+			prev := x
+		weight := mlx.Transpose(u.Conv, 0, 2, 3, 1)
+		x = conv2D3x3PaddedChannelsLast(x, weight)
+		if u.Bias != nil {
+			bias := mlx.Reshape(u.Bias, 1, 1, 1, outC)
+			x = mlx.Add(x, bias)
+		}
+		x = mlx.Reshape(x, B, T, H*2, W*2, outC)
+		prev.Free()
+		mlx.Eval(x)
+	}
+
+	return x
+}
+
+// MidBlock is the middle block of decoder
+type MidBlock struct {
+	ResBlock1 *ResBlock
+	Attention *AttentionBlock
+	ResBlock2 *ResBlock
+}
+
+// newMidBlock creates a mid block
+func newMidBlock(weights *safetensors.ModelWeights, prefix string, dim int32) (*MidBlock, error) {
+	res1, err := newResBlock(weights, prefix+".resnets.0", dim, dim)
+	if err != nil {
+		return nil, err
+	}
+	attn, err := newAttentionBlock(weights, prefix+".attentions.0", dim)
+	if err != nil {
+		return nil, err
+	}
+	res2, err := newResBlock(weights, prefix+".resnets.1", dim, dim)
+	if err != nil {
+		return nil, err
+	}
+
+	return &MidBlock{
+		ResBlock1: res1,
+		Attention: attn,
+		ResBlock2: res2,
+	}, nil
+}
+
+// Forward applies mid block
+func (m *MidBlock) Forward(x *mlx.Array) *mlx.Array {
+	// Each component handles its own pools; we just free inputs
+	prev := x
+	x = m.ResBlock1.Forward(x)
+	prev.Free()
+
+	prev = x
+	x = m.Attention.Forward(x)
+	prev.Free()
+
+	prev = x
+	x = m.ResBlock2.Forward(x)
+	prev.Free()
+
+	return x
+}
+
+// VAEDecoder is the full VAE decoder
+type VAEDecoder struct {
+	Config *VAEConfig
+
+	PostQuantConv *CausalConv3d
+	ConvIn        *CausalConv3d
+	MidBlock      *MidBlock
+	UpBlocks      []*UpBlock
+	NormOut       *RMSNorm3D
+	ConvOut       *CausalConv3d
+}
+
+// Load loads the VAE decoder from a directory
+func (m *VAEDecoder) Load(path string) error {
+	fmt.Println("Loading Qwen-Image VAE decoder...")
+
+	cfg := defaultVAEConfig()
+	m.Config = cfg
+
+	weights, err := safetensors.LoadModelWeights(path)
+	if err != nil {
+		return fmt.Errorf("weights: %w", err)
+	}
+
+	// Bulk load all weights as bf16
+	fmt.Print("  Loading weights as bf16... ")
+	if err := weights.Load(mlx.DtypeBFloat16); err != nil {
+		return fmt.Errorf("failed to load weights: %w", err)
+	}
+	fmt.Printf("✓ (%.1f GB)\n", float64(mlx.MetalGetActiveMemory())/(1024*1024*1024))
+
+	fmt.Print("  Loading post_quant_conv... ")
+	postQuantConv, err := newCausalConv3d(weights, "post_quant_conv")
+	if err != nil {
+		return err
+	}
+	m.PostQuantConv = postQuantConv
+	fmt.Println("✓")
+
+	fmt.Print("  Loading conv_in... ")
+	convIn, err := newCausalConv3d(weights, "decoder.conv_in")
+	if err != nil {
+		return err
+	}
+	m.ConvIn = convIn
+	fmt.Println("✓")
+
+	// Mid block (dim = base_dim * dim_mult[-1] = 96 * 4 = 384)
+	fmt.Print("  Loading mid_block... ")
+	midDim := cfg.BaseDim * cfg.DimMult[len(cfg.DimMult)-1]
+	midBlock, err := newMidBlock(weights, "decoder.mid_block", midDim)
+	if err != nil {
+		return err
+	}
+	m.MidBlock = midBlock
+	fmt.Println("✓")
+
+	// Up blocks (reversed dim_mult)
+	fmt.Print("  Loading up_blocks... ")
+	numUpBlocks := len(cfg.DimMult)
+	m.UpBlocks = make([]*UpBlock, numUpBlocks)
+
+	dimsMult := make([]int32, numUpBlocks+1)
+	dimsMult[0] = cfg.DimMult[numUpBlocks-1]
+	for i := 0; i < numUpBlocks; i++ {
+		dimsMult[i+1] = cfg.DimMult[numUpBlocks-1-i]
+	}
+
+	temporalUpsample := make([]bool, len(cfg.TemperalDownsample))
+	for i := range cfg.TemperalDownsample {
+		temporalUpsample[i] = cfg.TemperalDownsample[len(cfg.TemperalDownsample)-1-i]
+	}
+
+	for i := 0; i < numUpBlocks; i++ {
+		inDim := cfg.BaseDim * dimsMult[i]
+		outDim := cfg.BaseDim * dimsMult[i+1]
+
+		if i > 0 {
+			inDim = inDim / 2
+		}
+
+		upsampleMode := ""
+		if i < numUpBlocks-1 {
+			if temporalUpsample[i] {
+				upsampleMode = "upsample3d"
+			} else {
+				upsampleMode = "upsample2d"
+			}
+		}
+
+		prefix := fmt.Sprintf("decoder.up_blocks.%d", i)
+		upBlock, err := newUpBlock(weights, prefix, inDim, outDim, cfg.NumResBlocks, upsampleMode)
+		if err != nil {
+			return err
+		}
+		m.UpBlocks[i] = upBlock
+	}
+	fmt.Printf("✓ [%d blocks]\n", numUpBlocks)
+
+	fmt.Print("  Loading output layers... ")
+	normOut, err := newRMSNorm3D(weights, "decoder.norm_out", cfg.BaseDim)
+	if err != nil {
+		return err
+	}
+	m.NormOut = normOut
+	convOut, err := newCausalConv3d(weights, "decoder.conv_out")
+	if err != nil {
+		return err
+	}
+	m.ConvOut = convOut
+	fmt.Println("✓")
+
+	weights.ReleaseAll()
+	return nil
+}
+
+// LoadVAEDecoderFromPath is a convenience function to load VAE from path
+func LoadVAEDecoderFromPath(path string) (*VAEDecoder, error) {
+	m := &VAEDecoder{}
+	if err := m.Load(filepath.Join(path, "vae")); err != nil {
+		return nil, err
+	}
+	return m, nil
+}
+
+// Decode converts latents to image
+// z: [B, C, T, H, W] normalized latents
+// Uses staged pools to free intermediate arrays and reduce peak memory.
+func (vae *VAEDecoder) Decode(z *mlx.Array) *mlx.Array {
+	var x *mlx.Array
+
+	// Stage 1a: Denormalize and transpose
+	{
+			z = vae.Denormalize(z)
+		// Convert from channels-first [N, C, T, H, W] to channels-last [N, T, H, W, C]
+		z = mlx.Contiguous(mlx.Transpose(z, 0, 2, 3, 4, 1))
+		mlx.Eval(z)
+	}
+
+	// Stage 1b: PostQuantConv (handles its own pools)
+	x = vae.PostQuantConv.Forward(z)
+	z.Free()
+
+	// Stage 1c: ConvIn (handles its own pools)
+	{
+		prev := x
+		x = vae.ConvIn.Forward(x)
+		prev.Free()
+	}
+
+	// Stage 2: Mid block (handles its own pools)
+	x = vae.MidBlock.Forward(x)
+
+	// Stage 3: Up blocks (each handles its own pools)
+	for _, upBlock := range vae.UpBlocks {
+		x = upBlock.Forward(x)
+	}
+
+	// Stage 4a: NormOut + silu
+	{
+			prev := x
+		x = vae.NormOut.Forward(x)
+		x = silu3D(x)
+		prev.Free()
+		mlx.Eval(x)
+	}
+
+	// Stage 4b: ConvOut (handles its own pools)
+	{
+		prev := x
+		x = vae.ConvOut.Forward(x)
+		prev.Free()
+	}
+
+	// Stage 4c: Post-processing
+	{
+			prev := x
+		// Clamp to [-1, 1]
+		x = mlx.ClipScalar(x, -1.0, 1.0, true, true)
+		// Convert back from channels-last to channels-first
+		x = mlx.Contiguous(mlx.Transpose(x, 0, 4, 1, 2, 3))
+		prev.Free()
+		mlx.Eval(x)
+	}
+
+	return x
+}
+
+// Denormalize reverses the normalization applied during encoding
+func (vae *VAEDecoder) Denormalize(z *mlx.Array) *mlx.Array {
+	shape := z.Shape()
+	C := shape[1]
+
+	mean := mlx.NewArray(vae.Config.LatentsMean[:C], []int32{1, C, 1, 1, 1})
+	std := mlx.NewArray(vae.Config.LatentsStd[:C], []int32{1, C, 1, 1, 1})
+
+	mean = mlx.ToBFloat16(mean)
+	std = mlx.ToBFloat16(std)
+
+	return mlx.Add(mlx.Mul(z, std), mean)
+}
+
+// Helper functions
+
+func silu3D(x *mlx.Array) *mlx.Array {
+	return mlx.Mul(x, mlx.Sigmoid(x))
+}
+
+// pad3DChannelsLast pads a channels-last [B, T, H, W, C] tensor
+func pad3DChannelsLast(x *mlx.Array, tBefore, tAfter, hBefore, hAfter, wBefore, wAfter int32) *mlx.Array {
+	if tBefore == 0 && tAfter == 0 && hBefore == 0 && hAfter == 0 && wBefore == 0 && wAfter == 0 {
+		return x
+	}
+	// Pad dims: [B before, B after, T before, T after, H before, H after, W before, W after, C before, C after]
+	return mlx.Pad(x, []int32{0, 0, tBefore, tAfter, hBefore, hAfter, wBefore, wAfter, 0, 0})
+}
+
+func pad2D(x *mlx.Array, hBefore, hAfter, wBefore, wAfter int32) *mlx.Array {
+	if hBefore == 0 && hAfter == 0 && wBefore == 0 && wAfter == 0 {
+		return x
+	}
+	return mlx.Pad(x, []int32{0, 0, 0, 0, hBefore, hAfter, wBefore, wAfter})
+}
+
+func conv2D1x1(x, weight *mlx.Array) *mlx.Array {
+	shape := x.Shape()
+	B := shape[0]
+	H := shape[2]
+	W := shape[3]
+
+	x = mlx.Transpose(x, 0, 2, 3, 1)
+	x = mlx.Reshape(x, B*H*W, shape[1])
+
+	wShape := weight.Shape()
+	var w *mlx.Array
+	if len(wShape) == 4 {
+		w = mlx.Reshape(weight, wShape[0], wShape[1])
+	} else {
+		w = weight
+	}
+	w = mlx.Transpose(w, 1, 0)
+
+	out := mlx.Linear(x, w)
+	outC := w.Dim(1)
+	out = mlx.Reshape(out, B, H, W, outC)
+	return mlx.Transpose(out, 0, 3, 1, 2)
+}
+
+func conv2D3x3Padded(x, weight *mlx.Array) *mlx.Array {
+	x = pad2D(x, 1, 1, 1, 1)
+	return conv2D(x, weight, 1, 1)
+}
+
+func conv2D(x, w *mlx.Array, strideH, strideW int32) *mlx.Array {
+	x = mlx.Transpose(x, 0, 2, 3, 1)
+	w = mlx.Transpose(w, 0, 2, 3, 1)
+
+	shape := x.Shape()
+	B := shape[0]
+	H := shape[1]
+	W := shape[2]
+
+	wShape := w.Shape()
+	Cout := wShape[0]
+	kH := wShape[1]
+	kW := wShape[2]
+
+	outH := (H-kH)/strideH + 1
+	outW := (W-kW)/strideW + 1
+
+	patches := extractPatches2D(x, kH, kW, strideH, strideW)
+	wFlat := mlx.Reshape(w, Cout, -1)
+	patches = mlx.Reshape(patches, B*outH*outW, -1)
+	out := mlx.Linear(patches, mlx.Transpose(wFlat, 1, 0))
+	out = mlx.Reshape(out, B, outH, outW, Cout)
+	return mlx.Transpose(out, 0, 3, 1, 2)
+}
+
+func extractPatches2D(x *mlx.Array, kH, kW, strideH, strideW int32) *mlx.Array {
+	shape := x.Shape()
+	B := shape[0]
+	H := shape[1]
+	W := shape[2]
+	C := shape[3]
+
+	outH := (H-kH)/strideH + 1
+	outW := (W-kW)/strideW + 1
+
+	patches := make([]*mlx.Array, outH*outW)
+	idx := 0
+	for i := int32(0); i < outH; i++ {
+		for j := int32(0); j < outW; j++ {
+			startH := i * strideH
+			startW := j * strideW
+			patch := mlx.Slice(x, []int32{0, startH, startW, 0}, []int32{B, startH + kH, startW + kW, C})
+			patch = mlx.Reshape(patch, B, kH*kW*C)
+			patches[idx] = patch
+			idx++
+		}
+	}
+
+	for i := range patches {
+		patches[i] = mlx.ExpandDims(patches[i], 1)
+	}
+	stacked := mlx.Concatenate(patches, 1)
+	return mlx.Reshape(stacked, B, outH, outW, kH*kW*C)
+}
+
+func upsample2x(x *mlx.Array) *mlx.Array {
+	shape := x.Shape()
+	H := shape[2]
+	W := shape[3]
+
+	rowIdxData := make([]int32, H*2)
+	for i := int32(0); i < H; i++ {
+		rowIdxData[i*2] = i
+		rowIdxData[i*2+1] = i
+	}
+	rowIdx := mlx.NewArrayInt32(rowIdxData, []int32{H * 2})
+
+	colIdxData := make([]int32, W*2)
+	for i := int32(0); i < W; i++ {
+		colIdxData[i*2] = i
+		colIdxData[i*2+1] = i
+	}
+	colIdx := mlx.NewArrayInt32(colIdxData, []int32{W * 2})
+
+	x = mlx.Take(x, rowIdx, 2)
+	x = mlx.Take(x, colIdx, 3)
+
+	return x
+}
+
+// upsample2xChannelsLast upsamples channels-last input [B, H, W, C] by 2x
+func upsample2xChannelsLast(x *mlx.Array) *mlx.Array {
+	shape := x.Shape()
+	H := shape[1]
+	W := shape[2]
+
+	// Create repeat indices for rows
+	rowIdxData := make([]int32, H*2)
+	for i := int32(0); i < H; i++ {
+		rowIdxData[i*2] = i
+		rowIdxData[i*2+1] = i
+	}
+	rowIdx := mlx.NewArrayInt32(rowIdxData, []int32{H * 2})
+
+	// Create repeat indices for columns
+	colIdxData := make([]int32, W*2)
+	for i := int32(0); i < W; i++ {
+		colIdxData[i*2] = i
+		colIdxData[i*2+1] = i
+	}
+	colIdx := mlx.NewArrayInt32(colIdxData, []int32{W * 2})
+
+	// Take along H (axis 1) then W (axis 2)
+	x = mlx.Take(x, rowIdx, 1)
+	x = mlx.Take(x, colIdx, 2)
+
+	return x
+}
+
+// conv2D3x3PaddedChannelsLast applies 3x3 conv with padding to channels-last input [B, H, W, C]
+// weight: [outC, kH, kW, inC] (MLX channels-last format)
+func conv2D3x3PaddedChannelsLast(x, weight *mlx.Array) *mlx.Array {
+	// Pad spatial dims: [B, H, W, C] -> pad H and W by 1 each side
+	x = mlx.Pad(x, []int32{0, 0, 1, 1, 1, 1, 0, 0})
+	// Conv2d expects: input [B, H, W, inC], weight [outC, kH, kW, inC]
+	// stride=1, padding=0 (we already padded manually)
+	return mlx.Conv2d(x, weight, 1, 0)
+}

x/imagegen/models/qwen_image/vae_test.go

+//go:build mlx
+
+package qwen_image
+
+import (
+	"math"
+	"os"
+	"testing"
+
+	"github.com/ollama/ollama/x/imagegen/mlx"
+)
+
+// TestVAEConfig tests configuration invariants.
+func TestVAEConfig(t *testing.T) {
+	cfg := defaultVAEConfig()
+
+	// Property: latents_mean and latents_std have z_dim elements
+	if int32(len(cfg.LatentsMean)) != cfg.ZDim {
+		t.Errorf("latents_mean length != z_dim: %d != %d", len(cfg.LatentsMean), cfg.ZDim)
+	}
+	if int32(len(cfg.LatentsStd)) != cfg.ZDim {
+		t.Errorf("latents_std length != z_dim: %d != %d", len(cfg.LatentsStd), cfg.ZDim)
+	}
+
+	// Property: dim_mult defines 4 stages
+	if len(cfg.DimMult) != 4 {
+		t.Errorf("dim_mult should have 4 stages: got %d", len(cfg.DimMult))
+	}
+
+	// Property: temperal_downsample has 3 elements (for 3 transitions)
+	if len(cfg.TemperalDownsample) != 3 {
+		t.Errorf("temperal_downsample should have 3 elements: got %d", len(cfg.TemperalDownsample))
+	}
+}
+
+// TestVAELatentsNormalization tests the latent denormalization values.
+func TestVAELatentsNormalization(t *testing.T) {
+	cfg := defaultVAEConfig()
+
+	// Verify latents_std values are all positive
+	for i, std := range cfg.LatentsStd {
+		if std <= 0 {
+			t.Errorf("latents_std[%d] should be positive: %v", i, std)
+		}
+	}
+
+	// Verify values are in reasonable range (from actual model)
+	for i, mean := range cfg.LatentsMean {
+		if math.Abs(float64(mean)) > 5 {
+			t.Errorf("latents_mean[%d] seems too large: %v", i, mean)
+		}
+	}
+	for i, std := range cfg.LatentsStd {
+		if std > 10 {
+			t.Errorf("latents_std[%d] seems too large: %v", i, std)
+		}
+	}
+}
+
+// TestVAEDecoderForward tests full forward pass (integration test).
+// Skips if model weights are not available.
+func TestVAEDecoderForward(t *testing.T) {
+	weightsPath := "../../../weights/Qwen-Image-2512/vae"
+	if _, err := os.Stat(weightsPath); os.IsNotExist(err) {
+		t.Skip("Skipping: model weights not found at " + weightsPath)
+	}
+
+	vae := &VAEDecoder{}
+	if err := vae.Load(weightsPath); err != nil {
+		t.Fatalf("Failed to load VAE decoder: %v", err)
+	}
+	mlx.Keep(mlx.Collect(vae)...)
+
+	// Small test input: [B, C, T, H, W]
+	// After 4 upsampling stages (2x each), H/W multiply by 16
+	batchSize := int32(1)
+	channels := int32(16)
+	frames := int32(1)
+	latentH := int32(4)
+	latentW := int32(4)
+
+	latents := mlx.RandomNormal([]int32{batchSize, channels, frames, latentH, latentW}, 0)
+
+	// Decode
+	out := vae.Decode(latents)
+	mlx.Eval(out)
+
+	// Verify output shape: [B, 3, T, H*16, W*16]
+	outShape := out.Shape()
+	if outShape[0] != batchSize {
+		t.Errorf("batch size: got %d, want %d", outShape[0], batchSize)
+	}
+	if outShape[1] != 3 {
+		t.Errorf("channels: got %d, want 3", outShape[1])
+	}
+	if outShape[2] != frames {
+		t.Errorf("frames: got %d, want %d", outShape[2], frames)
+	}
+	expectedH := latentH * 16 // 4 stages of 2x upsampling
+	expectedW := latentW * 16
+	if outShape[3] != expectedH || outShape[4] != expectedW {
+		t.Errorf("spatial dims: got [%d, %d], want [%d, %d]",
+			outShape[3], outShape[4], expectedH, expectedW)
+	}
+
+	// Verify output is in valid range (should be clamped to [0, 1] by decode)
+	outData := out.Data()
+	for i := 0; i < min(100, len(outData)); i++ {
+		if math.IsNaN(float64(outData[i])) || math.IsInf(float64(outData[i]), 0) {
+			t.Errorf("output[%d] not finite: %v", i, outData[i])
+			break
+		}
+	}
+}

x/imagegen/models/qwen_image_edit/layers.go

+//go:build mlx
+
+package qwen_image_edit
+
+import (
+	"fmt"
+	"math"
+
+	"github.com/ollama/ollama/x/imagegen/mlx"
+	"github.com/ollama/ollama/x/imagegen/safetensors"
+)
+
+// CausalConv3d is a causal 3D convolution (for temporal causality)
+type CausalConv3d struct {
+	Weight       *mlx.Array
+	Bias         *mlx.Array
+	BiasReshaped *mlx.Array // [1, C, 1, 1, 1]
+	KernelT      int32
+}
+
+// newCausalConv3d creates a 3D causal conv
+func newCausalConv3d(weights *safetensors.ModelWeights, prefix string) (*CausalConv3d, error) {
+	weight, err := weights.Get(prefix + ".weight")
+	if err != nil {
+		return nil, fmt.Errorf("weight not found: %s", prefix)
+	}
+	bias, _ := weights.Get(prefix + ".bias")
+
+	kernelT := weight.Shape()[2]
+	outC := weight.Shape()[0]
+
+	var biasReshaped *mlx.Array
+	if bias != nil {
+		biasReshaped = mlx.Reshape(bias, 1, outC, 1, 1, 1)
+	}
+
+	return &CausalConv3d{
+		Weight:       weight,
+		Bias:         bias,
+		BiasReshaped: biasReshaped,
+		KernelT:      kernelT,
+	}, nil
+}
+
+// Forward applies causal 3D convolution (or 2D if weight is 4D)
+// x: [B, T, H, W, C] (channels-last, MLX format)
+func (c *CausalConv3d) Forward(x *mlx.Array) *mlx.Array {
+	shape := c.Weight.Shape()
+
+	// Handle both 5D (3D conv) and 4D (2D conv) weights
+	if len(shape) == 4 {
+		// 2D conv: [O, I, kH, kW] - need to apply per-frame
+		return c.forward2D(x)
+	}
+
+	// 3D conv: [O, I, kT, kH, kW]
+	kernelT := shape[2]
+	kernelH := shape[3]
+	kernelW := shape[4]
+
+	// Causal temporal padding, same spatial padding
+	padT := kernelT - 1
+	padH := kernelH / 2
+	padW := kernelW / 2
+
+	// Stage 1: Pad
+	{
+		x = pad3DChannelsLast(x, padT, 0, padH, padH, padW, padW)
+		mlx.Eval(x)
+	}
+
+	// Stage 2: Conv + bias
+	var out *mlx.Array
+	{
+		prev := x
+		weight := mlx.Transpose(c.Weight, 0, 2, 3, 4, 1)
+		out = mlx.Conv3d(x, weight, 1, 1, 1, 0, 0, 0)
+		if c.Bias != nil {
+			bias := mlx.Reshape(c.Bias, 1, 1, 1, 1, c.Bias.Dim(0))
+			out = mlx.Add(out, bias)
+		}
+		prev.Free()
+		mlx.Eval(out)
+	}
+
+	return out
+}
+
+// forward2D applies 2D conv per-frame for [B, T, H, W, C] input
+func (c *CausalConv3d) forward2D(x *mlx.Array) *mlx.Array {
+	xShape := x.Shape()
+	B := xShape[0]
+	T := xShape[1]
+	H := xShape[2]
+	W := xShape[3]
+	C := xShape[4]
+
+	wShape := c.Weight.Shape() // [O, I, kH, kW]
+	kernelH := wShape[2]
+	kernelW := wShape[3]
+	outC := wShape[0]
+
+	padH := kernelH / 2
+	padW := kernelW / 2
+
+	// Reshape to [B*T, H, W, C] for 2D conv
+	x = mlx.Reshape(x, B*T, H, W, C)
+
+	// Pad spatially
+	x = mlx.Pad(x, []int32{0, 0, padH, padH, padW, padW, 0, 0})
+
+	// Apply 2D conv
+	weight := mlx.Transpose(c.Weight, 0, 2, 3, 1) // [O, I, kH, kW] -> [O, kH, kW, I]
+	x = mlx.Conv2d(x, weight, 1, 0)
+
+	if c.Bias != nil {
+		bias := mlx.Reshape(c.Bias, 1, 1, 1, outC)
+		x = mlx.Add(x, bias)
+	}
+
+	// Get output spatial dims
+	outH := H
+	outW := W
+
+	// Reshape back to [B, T, H, W, C]
+	x = mlx.Reshape(x, B, T, outH, outW, outC)
+	mlx.Eval(x)
+
+	return x
+}
+
+// RMSNorm3D applies RMS normalization over channels
+type RMSNorm3D struct {
+	Gamma *mlx.Array // [1, 1, 1, 1, C] for broadcasting
+}
+
+// newRMSNorm3D creates an RMS norm
+func newRMSNorm3D(weights *safetensors.ModelWeights, prefix string, dim int32) (*RMSNorm3D, error) {
+	gamma, err := weights.Get(prefix + ".gamma")
+	if err != nil {
+		return nil, err
+	}
+	gamma = mlx.Reshape(gamma, 1, 1, 1, 1, gamma.Dim(0))
+	return &RMSNorm3D{Gamma: gamma}, nil
+}
+
+// Forward applies RMS norm to channels-last input [B, T, H, W, C]
+func (n *RMSNorm3D) Forward(x *mlx.Array) *mlx.Array {
+	normalized := mlx.RMSNormNoWeight(x, 1e-6)
+	return mlx.Mul(normalized, n.Gamma)
+}
+
+// ResBlock is a residual block with RMS norm and causal convs
+type ResBlock struct {
+	Norm1    *RMSNorm3D
+	Conv1    *CausalConv3d
+	Norm2    *RMSNorm3D
+	Conv2    *CausalConv3d
+	Shortcut *CausalConv3d
+}
+
+// newResBlock creates a residual block
+func newResBlock(weights *safetensors.ModelWeights, prefix string, inDim, outDim int32) (*ResBlock, error) {
+	norm1, err := newRMSNorm3D(weights, prefix+".norm1", inDim)
+	if err != nil {
+		return nil, err
+	}
+	conv1, err := newCausalConv3d(weights, prefix+".conv1")
+	if err != nil {
+		return nil, err
+	}
+	norm2, err := newRMSNorm3D(weights, prefix+".norm2", outDim)
+	if err != nil {
+		return nil, err
+	}
+	conv2, err := newCausalConv3d(weights, prefix+".conv2")
+	if err != nil {
+		return nil, err
+	}
+
+	var shortcut *CausalConv3d
+	if inDim != outDim {
+		shortcut, err = newCausalConv3d(weights, prefix+".conv_shortcut")
+		if err != nil {
+			return nil, err
+		}
+	}
+
+	return &ResBlock{
+		Norm1:    norm1,
+		Conv1:    conv1,
+		Norm2:    norm2,
+		Conv2:    conv2,
+		Shortcut: shortcut,
+	}, nil
+}
+
+// Forward applies the residual block
+func (r *ResBlock) Forward(x *mlx.Array) *mlx.Array {
+	var h *mlx.Array
+
+	mlx.Keep(x)
+
+	// Stage 1: norm1 + silu
+	{
+		h = r.Norm1.Forward(x)
+		h = silu3D(h)
+		mlx.Eval(h)
+	}
+
+	// Stage 2: conv1
+	{
+		prev := h
+		h = r.Conv1.Forward(h)
+		prev.Free()
+	}
+
+	// Stage 3: norm2 + silu
+	{
+		prev := h
+		h = r.Norm2.Forward(h)
+		h = silu3D(h)
+		prev.Free()
+		mlx.Eval(h)
+	}
+
+	// Stage 4: conv2
+	{
+		prev := h
+		h = r.Conv2.Forward(h)
+		prev.Free()
+	}
+
+	// Residual connection
+	if r.Shortcut != nil {
+		shortcut := r.Shortcut.Forward(x)
+		h = mlx.Add(h, shortcut)
+		mlx.Eval(h)
+	} else {
+		h = mlx.Add(h, x)
+		mlx.Eval(h)
+	}
+
+	return h
+}
+
+// AttentionBlock is a 2D attention block
+type AttentionBlock struct {
+	Norm      *RMSNorm3D
+	ToQKV     *mlx.Array
+	ToQKVBias *mlx.Array
+	Proj      *mlx.Array
+	ProjBias  *mlx.Array
+	Dim       int32
+}
+
+// newAttentionBlock creates an attention block
+func newAttentionBlock(weights *safetensors.ModelWeights, prefix string, dim int32) (*AttentionBlock, error) {
+	norm, err := newRMSNorm3D(weights, prefix+".norm", dim)
+	if err != nil {
+		return nil, err
+	}
+	toQKV, _ := weights.Get(prefix + ".to_qkv.weight")
+	toQKVBias, _ := weights.Get(prefix + ".to_qkv.bias")
+	proj, _ := weights.Get(prefix + ".proj.weight")
+	projBias, _ := weights.Get(prefix + ".proj.bias")
+
+	return &AttentionBlock{
+		Norm:      norm,
+		ToQKV:     toQKV,
+		ToQKVBias: toQKVBias,
+		Proj:      proj,
+		ProjBias:  projBias,
+		Dim:       dim,
+	}, nil
+}
+
+// Forward applies 2D attention
+// Input: [B, T, H, W, C] (channels-last)
+func (a *AttentionBlock) Forward(x *mlx.Array) *mlx.Array {
+	shape := x.Shape()
+	B := shape[0]
+	T := shape[1]
+	H := shape[2]
+	W := shape[3]
+	C := shape[4]
+
+	identity := x
+
+	// Flatten to [B*T, 1, H, W, C] for norm
+	x = mlx.Reshape(x, B*T, 1, H, W, C)
+	x = a.Norm.Forward(x)
+	x = mlx.Reshape(x, B*T, H, W, C)
+
+	// Flatten spatial to [B*T, H*W, C]
+	x = mlx.Reshape(x, B*T, H*W, C)
+
+	// Linear to get Q, K, V
+	wShape := a.ToQKV.Shape()
+	var w *mlx.Array
+	if len(wShape) == 4 {
+		w = mlx.Reshape(a.ToQKV, wShape[0], wShape[1])
+	} else {
+		w = a.ToQKV
+	}
+	w = mlx.Transpose(w, 1, 0)
+
+	qkv := mlx.Linear(x, w)
+	if a.ToQKVBias != nil {
+		qkv = mlx.Add(qkv, a.ToQKVBias)
+	}
+	qkv = mlx.Reshape(qkv, B*T, 1, H*W, 3*C)
+
+	q := mlx.Slice(qkv, []int32{0, 0, 0, 0}, []int32{B * T, 1, H * W, C})
+	k := mlx.Slice(qkv, []int32{0, 0, 0, C}, []int32{B * T, 1, H * W, 2 * C})
+	v := mlx.Slice(qkv, []int32{0, 0, 0, 2 * C}, []int32{B * T, 1, H * W, 3 * C})
+
+	scale := float32(1.0 / math.Sqrt(float64(C)))
+	out := mlx.ScaledDotProductAttention(q, k, v, scale, false)
+
+	out = mlx.Reshape(out, B*T, H*W, C)
+
+	// Project back
+	pShape := a.Proj.Shape()
+	var p *mlx.Array
+	if len(pShape) == 4 {
+		p = mlx.Reshape(a.Proj, pShape[0], pShape[1])
+	} else {
+		p = a.Proj
+	}
+	p = mlx.Transpose(p, 1, 0)
+	out = mlx.Linear(out, p)
+	if a.ProjBias != nil {
+		out = mlx.Add(out, a.ProjBias)
+	}
+
+	out = mlx.Reshape(out, B, T, H, W, C)
+	return mlx.Add(out, identity)
+}
+
+// UpBlock handles upsampling in decoder
+type UpBlock struct {
+	ResBlocks []*ResBlock
+	Upsampler *Upsample
+}
+
+// newUpBlock creates an up block
+func newUpBlock(weights *safetensors.ModelWeights, prefix string, inDim, outDim int32, numBlocks int32, upsampleMode string) (*UpBlock, error) {
+	resBlocks := make([]*ResBlock, numBlocks+1)
+
+	currentDim := inDim
+	for i := int32(0); i <= numBlocks; i++ {
+		resPrefix := fmt.Sprintf("%s.resnets.%d", prefix, i)
+		block, err := newResBlock(weights, resPrefix, currentDim, outDim)
+		if err != nil {
+			return nil, err
+		}
+		resBlocks[i] = block
+		currentDim = outDim
+	}
+
+	var upsampler *Upsample
+	if upsampleMode != "" {
+		upsampler = newUpsample(weights, prefix+".upsamplers.0", outDim, upsampleMode)
+	}
+
+	return &UpBlock{
+		ResBlocks: resBlocks,
+		Upsampler: upsampler,
+	}, nil
+}
+
+// Forward applies up block
+func (u *UpBlock) Forward(x *mlx.Array) *mlx.Array {
+	for _, block := range u.ResBlocks {
+		prev := x
+		x = block.Forward(x)
+		prev.Free()
+	}
+
+	if u.Upsampler != nil {
+		prev := x
+		x = u.Upsampler.Forward(x)
+		prev.Free()
+	}
+	return x
+}
+
+// Upsample handles spatial upsampling
+type Upsample struct {
+	Conv *mlx.Array
+	Bias *mlx.Array
+	Mode string
+}
+
+// newUpsample creates an upsampler
+func newUpsample(weights *safetensors.ModelWeights, prefix string, dim int32, mode string) *Upsample {
+	conv, _ := weights.Get(prefix + ".resample.1.weight")
+	bias, _ := weights.Get(prefix + ".resample.1.bias")
+	return &Upsample{
+		Conv: conv,
+		Bias: bias,
+		Mode: mode,
+	}
+}
+
+// Forward applies upsampling to channels-last input [B, T, H, W, C]
+func (u *Upsample) Forward(x *mlx.Array) *mlx.Array {
+	shape := x.Shape()
+	B := shape[0]
+	T := shape[1]
+	H := shape[2]
+	W := shape[3]
+	C := shape[4]
+	outC := u.Conv.Shape()[0]
+
+	// Stage 1: 2x nearest neighbor upsample
+	{
+		x = mlx.Reshape(x, B*T, H, W, C)
+		x = upsample2xChannelsLast(x)
+		mlx.Eval(x)
+	}
+
+	// Stage 2: Conv + bias
+	{
+		prev := x
+		weight := mlx.Transpose(u.Conv, 0, 2, 3, 1)
+		x = conv2D3x3PaddedChannelsLast(x, weight)
+		if u.Bias != nil {
+			bias := mlx.Reshape(u.Bias, 1, 1, 1, outC)
+			x = mlx.Add(x, bias)
+		}
+		x = mlx.Reshape(x, B, T, H*2, W*2, outC)
+		prev.Free()
+		mlx.Eval(x)
+	}
+
+	return x
+}
+
+// MidBlock is the middle block
+type MidBlock struct {
+	ResBlock1 *ResBlock
+	Attention *AttentionBlock
+	ResBlock2 *ResBlock
+}
+
+// newMidBlock creates a mid block
+func newMidBlock(weights *safetensors.ModelWeights, prefix string, dim int32) (*MidBlock, error) {
+	res1, err := newResBlock(weights, prefix+".resnets.0", dim, dim)
+	if err != nil {
+		return nil, err
+	}
+	attn, err := newAttentionBlock(weights, prefix+".attentions.0", dim)
+	if err != nil {
+		return nil, err
+	}
+	res2, err := newResBlock(weights, prefix+".resnets.1", dim, dim)
+	if err != nil {
+		return nil, err
+	}
+
+	return &MidBlock{
+		ResBlock1: res1,
+		Attention: attn,
+		ResBlock2: res2,
+	}, nil
+}
+
+// Forward applies mid block
+func (m *MidBlock) Forward(x *mlx.Array) *mlx.Array {
+	prev := x
+	x = m.ResBlock1.Forward(x)
+	prev.Free()
+
+	prev = x
+	x = m.Attention.Forward(x)
+	prev.Free()
+
+	prev = x
+	x = m.ResBlock2.Forward(x)
+	prev.Free()
+
+	return x
+}
+
+// Helper functions
+
+func silu3D(x *mlx.Array) *mlx.Array {
+	return mlx.Mul(x, mlx.Sigmoid(x))
+}
+
+// pad3DChannelsLast pads a channels-last [B, T, H, W, C] tensor
+func pad3DChannelsLast(x *mlx.Array, tBefore, tAfter, hBefore, hAfter, wBefore, wAfter int32) *mlx.Array {
+	if tBefore == 0 && tAfter == 0 && hBefore == 0 && hAfter == 0 && wBefore == 0 && wAfter == 0 {
+		return x
+	}
+	return mlx.Pad(x, []int32{0, 0, tBefore, tAfter, hBefore, hAfter, wBefore, wAfter, 0, 0})
+}
+
+// upsample2xChannelsLast upsamples channels-last input [B, H, W, C] by 2x
+func upsample2xChannelsLast(x *mlx.Array) *mlx.Array {
+	shape := x.Shape()
+	H := shape[1]
+	W := shape[2]
+
+	rowIdxData := make([]int32, H*2)
+	for i := int32(0); i < H; i++ {
+		rowIdxData[i*2] = i
+		rowIdxData[i*2+1] = i
+	}
+	rowIdx := mlx.NewArrayInt32(rowIdxData, []int32{H * 2})
+
+	colIdxData := make([]int32, W*2)
+	for i := int32(0); i < W; i++ {
+		colIdxData[i*2] = i
+		colIdxData[i*2+1] = i
+	}
+	colIdx := mlx.NewArrayInt32(colIdxData, []int32{W * 2})
+
+	x = mlx.Take(x, rowIdx, 1)
+	x = mlx.Take(x, colIdx, 2)
+
+	return x
+}
+
+// conv2D3x3PaddedChannelsLast applies 3x3 conv with padding to channels-last input [B, H, W, C]
+func conv2D3x3PaddedChannelsLast(x, weight *mlx.Array) *mlx.Array {
+	x = mlx.Pad(x, []int32{0, 0, 1, 1, 1, 1, 0, 0})
+	return mlx.Conv2d(x, weight, 1, 0)
+}
+
+// conv2DStrided applies conv with stride > 1 using manual patch extraction
+// x: [B, H, W, C] (channels-last), weight: [O, kH, kW, I]
+func conv2DStrided(x, weight *mlx.Array, stride int32) *mlx.Array {
+	shape := x.Shape()
+	B := shape[0]
+	H := shape[1]
+	W := shape[2]
+
+	wShape := weight.Shape()
+	Cout := wShape[0]
+	kH := wShape[1]
+	kW := wShape[2]
+
+	outH := (H - kH) / stride + 1
+	outW := (W - kW) / stride + 1
+
+	patches := extractPatches2DStrided(x, kH, kW, stride)
+	wFlat := mlx.Reshape(weight, Cout, -1)
+	patches = mlx.Reshape(patches, B*outH*outW, -1)
+	out := mlx.Linear(patches, mlx.Transpose(wFlat, 1, 0))
+	return mlx.Reshape(out, B, outH, outW, Cout)
+}
+
+// conv3DStrided applies 3D conv with strides using manual patch extraction
+// x: [B, T, H, W, C] (channels-last), weight: [O, I, kT, kH, kW] (PyTorch format)
+// strideT, strideH, strideW are the strides for each dimension
+// Patches are extracted in [C, T, H, W] order to match Python's preprocessing
+func conv3DStrided(x, weight *mlx.Array, strideT, strideH, strideW int32) *mlx.Array {
+	shape := x.Shape()
+	B := shape[0]
+	T := shape[1]
+	H := shape[2]
+	W := shape[3]
+	C := shape[4]
+
+	wShape := weight.Shape()
+	Cout := wShape[0]
+	// I := wShape[1]
+	kT := wShape[2]
+	kH := wShape[3]
+	kW := wShape[4]
+
+	// For temporal: if T < kT, we need to repeat frames temporally
+	// For single image with T=1 and kT=2, we duplicate the frame to T=kT
+	// Python Qwen2.5-VL duplicates the frame, not zero-pads
+	if T < kT {
+		// Tile along T dimension: [B, T, H, W, C] -> [B, kT, H, W, C]
+		x = mlx.Tile(x, []int32{1, kT, 1, 1, 1})
+		T = kT
+	}
+
+	outT := (T - kT) / strideT + 1
+	outH := (H - kH) / strideH + 1
+	outW := (W - kW) / strideW + 1
+
+	// Extract 3D patches in [C, T, H, W] order to match Python
+	patches := extractPatches3DStrided(x, kT, kH, kW, strideT, strideH, strideW)
+	// patches shape: [B, outT, outH, outW, C*kT*kH*kW]
+
+	// Weight is [O, I, kT, kH, kW] - flatten to [O, I*kT*kH*kW] to match patch order [C, T, H, W]
+	wFlat := mlx.Reshape(weight, Cout, -1) // [Cout, I*kT*kH*kW]
+	patches = mlx.Reshape(patches, B*outT*outH*outW, C*kT*kH*kW)
+	out := mlx.Linear(patches, mlx.Transpose(wFlat, 1, 0))
+	return mlx.Reshape(out, B, outT, outH, outW, Cout)
+}
+
+// extractPatches3DStrided extracts 3D patches with given strides
+// Returns patches with values in [C, T, H, W] order to match Python's preprocessing
+func extractPatches3DStrided(x *mlx.Array, kT, kH, kW, strideT, strideH, strideW int32) *mlx.Array {
+	shape := x.Shape()
+	B := shape[0]
+	T := shape[1]
+	H := shape[2]
+	W := shape[3]
+	C := shape[4]
+
+	outT := (T - kT) / strideT + 1
+	outH := (H - kH) / strideH + 1
+	outW := (W - kW) / strideW + 1
+
+	numPatches := outT * outH * outW
+	patches := make([]*mlx.Array, numPatches)
+	idx := 0
+	for t := int32(0); t < outT; t++ {
+		for i := int32(0); i < outH; i++ {
+			for j := int32(0); j < outW; j++ {
+				startT := t * strideT
+				startH := i * strideH
+				startW := j * strideW
+				// Extract patch: [B, kT, kH, kW, C]
+				patch := mlx.Slice(x,
+					[]int32{0, startT, startH, startW, 0},
+					[]int32{B, startT + kT, startH + kH, startW + kW, C})
+				// Transpose from [B, T, H, W, C] to [B, C, T, H, W] to match Python's order
+				patch = mlx.Transpose(patch, 0, 4, 1, 2, 3)
+				// Flatten to [B, C*T*H*W]
+				patch = mlx.Reshape(patch, B, C*kT*kH*kW)
+				patches[idx] = patch
+				idx++
+			}
+		}
+	}
+
+	for i := range patches {
+		patches[i] = mlx.ExpandDims(patches[i], 1)
+	}
+	stacked := mlx.Concatenate(patches, 1)
+	return mlx.Reshape(stacked, B, outT, outH, outW, C*kT*kH*kW)
+}
+
+// extractPatches2DStrided extracts patches with given stride
+func extractPatches2DStrided(x *mlx.Array, kH, kW, stride int32) *mlx.Array {
+	shape := x.Shape()
+	B := shape[0]
+	H := shape[1]
+	W := shape[2]
+	C := shape[3]
+
+	outH := (H - kH) / stride + 1
+	outW := (W - kW) / stride + 1
+
+	patches := make([]*mlx.Array, outH*outW)
+	idx := 0
+	for i := int32(0); i < outH; i++ {
+		for j := int32(0); j < outW; j++ {
+			startH := i * stride
+			startW := j * stride
+			patch := mlx.Slice(x, []int32{0, startH, startW, 0}, []int32{B, startH + kH, startW + kW, C})
+			patch = mlx.Reshape(patch, B, kH*kW*C)
+			patches[idx] = patch
+			idx++
+		}
+	}
+
+	for i := range patches {
+		patches[i] = mlx.ExpandDims(patches[i], 1)
+	}
+	stacked := mlx.Concatenate(patches, 1)
+	return mlx.Reshape(stacked, B, outH, outW, kH*kW*C)
+}
+
+// layerNormNoAffine applies layer norm without learnable parameters
+func layerNormNoAffine(x *mlx.Array, eps float32) *mlx.Array {
+	ndim := x.Ndim()
+	lastAxis := ndim - 1
+	mean := mlx.Mean(x, lastAxis, true)
+	xCentered := mlx.Sub(x, mean)
+	variance := mlx.Mean(mlx.Square(xCentered), lastAxis, true)
+	return mlx.Div(xCentered, mlx.Sqrt(mlx.AddScalar(variance, eps)))
+}

x/imagegen/models/qwen_image_edit/processor.go

+//go:build mlx
+
+package qwen_image_edit
+
+import (
+	"fmt"
+	"image"
+	"image/color"
+	_ "image/jpeg"
+	_ "image/png"
+	"math"
+	"os"
+
+	"github.com/ollama/ollama/x/imagegen/mlx"
+	"golang.org/x/image/draw"
+	_ "golang.org/x/image/webp"
+)
+
+// loadImageFile loads an image from disk
+func loadImageFile(path string) (image.Image, error) {
+	f, err := os.Open(path)
+	if err != nil {
+		return nil, fmt.Errorf("open image: %w", err)
+	}
+	defer f.Close()
+
+	img, _, err := image.Decode(f)
+	if err != nil {
+		return nil, fmt.Errorf("decode image: %w", err)
+	}
+	return img, nil
+}
+
+// imageToFloat32Pixels converts an image to a float32 pixel array [H, W, C] in [0, 1] range
+func imageToFloat32Pixels(img image.Image, width, height int) []float32 {
+	pixels := make([]float32, width*height*3)
+	idx := 0
+	for y := 0; y < height; y++ {
+		for x := 0; x < width; x++ {
+			r, g, b, _ := img.At(x, y).RGBA()
+			pixels[idx] = float32(r) / 65535.0
+			pixels[idx+1] = float32(g) / 65535.0
+			pixels[idx+2] = float32(b) / 65535.0
+			idx += 3
+		}
+	}
+	return pixels
+}
+
+// normalizeImageNet applies ImageNet normalization to an image tensor
+func (p *Processor) normalizeImageNet(arr *mlx.Array) *mlx.Array {
+	mean := mlx.NewArray(p.Config.ImageMean, []int32{1, 1, 3})
+	std := mlx.NewArray(p.Config.ImageStd, []int32{1, 1, 3})
+	return mlx.Div(mlx.Sub(arr, mean), std)
+}
+
+// prepareImageTensor transforms [H, W, C] to [B, C, H, W] and converts to bf16
+func prepareImageTensor(arr *mlx.Array) *mlx.Array {
+	// Transpose to [C, H, W] and make contiguous
+	arr = mlx.Contiguous(mlx.Transpose(arr, 2, 0, 1))
+	// Add batch dimension [1, C, H, W]
+	arr = mlx.ExpandDims(arr, 0)
+	// Convert to bf16
+	arr = mlx.ToBFloat16(arr)
+	mlx.Eval(arr)
+	return arr
+}
+
+// clampFloat clamps a value to [0, 255] and returns uint8
+func clampFloat(v, weightSum float64) uint8 {
+	v /= weightSum
+	if v < 0 {
+		v = 0
+	}
+	if v > 255 {
+		v = 255
+	}
+	return uint8(math.Round(v))
+}
+
+// ImageDims holds dimensions for a preprocessed image
+type ImageDims struct {
+	// Original image dimensions
+	OrigW, OrigH int32
+	// Condition image dimensions (for vision encoder)
+	CondW, CondH int32
+	// VAE image dimensions
+	VaeW, VaeH int32
+	// Latent dimensions (VAE dims / vae_scale_factor)
+	LatentW, LatentH int32
+	// Patch dimensions (latent dims / patch_size)
+	PatchW, PatchH int32
+}
+
+// ProcessorConfig holds image processor configuration
+type ProcessorConfig struct {
+	// Condition image size (target pixel area for vision encoder input)
+	// Python: CONDITION_IMAGE_SIZE = 384 * 384 = 147456
+	// Pipeline resizes image to this area before passing to encode_prompt
+	ConditionImageSize int32
+
+	// VAE image size (target pixel area)
+	// Python: VAE_IMAGE_SIZE = 1024 * 1024 = 1048576
+	VAEImageSize int32
+
+	// Image normalization (ImageNet stats for vision encoder)
+	ImageMean []float32
+	ImageStd  []float32
+}
+
+// defaultProcessorConfig returns default processor config
+func defaultProcessorConfig() *ProcessorConfig {
+	return &ProcessorConfig{
+		ConditionImageSize: 384 * 384,   // 147456 - matches Python CONDITION_IMAGE_SIZE
+		VAEImageSize:       1024 * 1024, // 1048576 - matches Python VAE_IMAGE_SIZE
+		ImageMean:          []float32{0.48145466, 0.4578275, 0.40821073},
+		ImageStd:           []float32{0.26862954, 0.26130258, 0.27577711},
+	}
+}
+
+// Processor handles image preprocessing for Qwen-Image-Edit
+type Processor struct {
+	Config *ProcessorConfig
+}
+
+// Load loads the processor config
+func (p *Processor) Load(path string) error {
+	p.Config = defaultProcessorConfig()
+	return nil
+}
+
+// LoadAndPreprocess loads an image and preprocesses it for both paths
+// Returns: condImage (for vision encoder), vaeImage (for VAE encoding)
+func (p *Processor) LoadAndPreprocess(imagePath string) (*mlx.Array, *mlx.Array, error) {
+	img, err := loadImageFile(imagePath)
+	if err != nil {
+		return nil, nil, err
+	}
+
+	bounds := img.Bounds()
+	origW := bounds.Dx()
+	origH := bounds.Dy()
+	ratio := float64(origW) / float64(origH)
+
+	// Calculate dimensions for condition image (vision encoder)
+	// Python pipeline does TWO resizes:
+	// 1. VaeImageProcessor.resize with Lanczos to CONDITION_IMAGE_SIZE (384x384 area)
+	// 2. Qwen2VLProcessor's smart_resize with Bicubic to multiple of 28
+	intermediateW, intermediateH := calculateDimensions(p.Config.ConditionImageSize, ratio, 32)
+	finalH, finalW := smartResize(intermediateH, intermediateW, 28, 56*56, 28*28*1280)
+
+	// Calculate dimensions for VAE image (1024x1024 area)
+	// Use multiple of 32 (vae_scale_factor * patch_size * 2 = 8 * 2 * 2 = 32)
+	vaeW, vaeH := calculateDimensions(p.Config.VAEImageSize, ratio, 32)
+
+	// Preprocess for condition (vision encoder) - two-step resize
+	condImage := p.preprocessImageTwoStep(img, intermediateW, intermediateH, finalW, finalH)
+
+	// Preprocess for VAE ([-1, 1] range, 5D tensor)
+	vaeImage := p.preprocessImageForVAE(img, vaeW, vaeH)
+
+	return condImage, vaeImage, nil
+}
+
+// preprocessImageLanczos does single-step Lanczos resize for vision encoder
+// Matches Python VaeImageProcessor.resize with resample='lanczos' (the default)
+// Used by edit_plus pipeline for multi-image input
+// Returns: [B, C, H, W] normalized tensor
+func (p *Processor) preprocessImageLanczos(img image.Image, width, height int32) *mlx.Array {
+	resized := resizeImageLanczos(img, int(width), int(height))
+	pixels := imageToFloat32Pixels(resized, int(width), int(height))
+	arr := mlx.NewArray(pixels, []int32{height, width, 3})
+	arr = p.normalizeImageNet(arr)
+	return prepareImageTensor(arr)
+}
+
+// preprocessImageTwoStep does two-step resize for vision encoder to match Python pipeline
+// Step 1: Lanczos resize from original to intermediate size (VaeImageProcessor.resize)
+// Step 2: Bicubic resize from intermediate to final size (Qwen2VLProcessor smart_resize)
+// Returns: [B, C, H, W] normalized tensor
+func (p *Processor) preprocessImageTwoStep(img image.Image, intermediateW, intermediateH, finalW, finalH int32) *mlx.Array {
+	intermediate := resizeImageLanczos(img, int(intermediateW), int(intermediateH))
+	resized := resizeImageBicubic(intermediate, int(finalW), int(finalH))
+	pixels := imageToFloat32Pixels(resized, int(finalW), int(finalH))
+	arr := mlx.NewArray(pixels, []int32{finalH, finalW, 3})
+	arr = p.normalizeImageNet(arr)
+	return prepareImageTensor(arr)
+}
+
+// preprocessImage converts image to tensor for vision encoder
+// Returns: [B, C, H, W] normalized tensor
+func (p *Processor) preprocessImage(img image.Image, width, height int32, normalize bool) *mlx.Array {
+	resized := resizeImageBicubic(img, int(width), int(height))
+	pixels := imageToFloat32Pixels(resized, int(width), int(height))
+	arr := mlx.NewArray(pixels, []int32{height, width, 3})
+	if normalize {
+		arr = p.normalizeImageNet(arr)
+	}
+	return prepareImageTensor(arr)
+}
+
+// preprocessImageForVAE converts image to tensor for VAE encoding
+// Returns: [B, C, T, H, W] tensor in [-1, 1] range
+func (p *Processor) preprocessImageForVAE(img image.Image, width, height int32) *mlx.Array {
+	resized := resizeImageLanczos(img, int(width), int(height))
+	pixels := imageToFloat32Pixels(resized, int(width), int(height))
+	arr := mlx.NewArray(pixels, []int32{height, width, 3})
+
+	// Scale to [-1, 1]: arr * 2 - 1
+	arr = mlx.MulScalar(arr, 2.0)
+	arr = mlx.AddScalar(arr, -1.0)
+
+	// Transpose to [C, H, W] and make contiguous
+	arr = mlx.Contiguous(mlx.Transpose(arr, 2, 0, 1))
+
+	// Add batch and temporal dimensions [1, C, 1, H, W]
+	arr = mlx.ExpandDims(arr, 0) // [1, C, H, W]
+	arr = mlx.ExpandDims(arr, 2) // [1, C, 1, H, W]
+
+	arr = mlx.ToBFloat16(arr)
+	mlx.Eval(arr)
+	return arr
+}
+
+// smartResize implements Python Qwen2VL processor's smart_resize logic
+// Returns (resizedHeight, resizedWidth) that fit within min/max pixel constraints
+func smartResize(height, width, factor, minPixels, maxPixels int32) (int32, int32) {
+	// Round to factor
+	hBar := int32(math.Round(float64(height)/float64(factor))) * factor
+	wBar := int32(math.Round(float64(width)/float64(factor))) * factor
+
+	// Ensure minimum factor size
+	if hBar < factor {
+		hBar = factor
+	}
+	if wBar < factor {
+		wBar = factor
+	}
+
+	// Check pixel constraints
+	total := hBar * wBar
+	if total > maxPixels {
+		// Scale down
+		beta := math.Sqrt(float64(maxPixels) / float64(total))
+		hBar = int32(math.Floor(float64(height)*beta/float64(factor))) * factor
+		wBar = int32(math.Floor(float64(width)*beta/float64(factor))) * factor
+	} else if total < minPixels {
+		// Scale up
+		beta := math.Sqrt(float64(minPixels) / float64(total))
+		hBar = int32(math.Ceil(float64(height)*beta/float64(factor))) * factor
+		wBar = int32(math.Ceil(float64(width)*beta/float64(factor))) * factor
+	}
+
+	return hBar, wBar
+}
+
+// calculateDimensions calculates width and height for a target area while maintaining ratio
+// multiple: the value to round dimensions to (e.g., 28 for vision encoder with patch 14 and 2x2 merge)
+func calculateDimensions(targetArea int32, ratio float64, multiple int32) (int32, int32) {
+	width := math.Sqrt(float64(targetArea) * ratio)
+	height := width / ratio
+
+	m := float64(multiple)
+	width = math.Round(width/m) * m
+	height = math.Round(height/m) * m
+
+	// Ensure minimum dimensions
+	if width < m {
+		width = m
+	}
+	if height < m {
+		height = m
+	}
+
+	return int32(width), int32(height)
+}
+
+// resizeImageLanczos resizes an image using Lanczos3 interpolation (matches PIL.LANCZOS)
+func resizeImageLanczos(img image.Image, width, height int) image.Image {
+	bounds := img.Bounds()
+	dst := image.NewRGBA(image.Rect(0, 0, width, height))
+
+	// Lanczos3 kernel (a=3) to match PIL.LANCZOS
+	lanczos3 := &draw.Kernel{
+		Support: 3.0,
+		At: func(t float64) float64 {
+			if t == 0 {
+				return 1.0
+			}
+			if t < 0 {
+				t = -t
+			}
+			if t >= 3.0 {
+				return 0.0
+			}
+			// sinc(t) * sinc(t/3)
+			piT := math.Pi * t
+			return (math.Sin(piT) / piT) * (math.Sin(piT/3) / (piT / 3))
+		},
+	}
+	lanczos3.Scale(dst, dst.Bounds(), img, bounds, draw.Over, nil)
+
+	return dst
+}
+
+// resizeImageBicubic resizes an image using bicubic interpolation (matches PIL.BICUBIC)
+// Uses separable interpolation with PIL's coordinate mapping for exact match
+func resizeImageBicubic(img image.Image, width, height int) image.Image {
+	bounds := img.Bounds()
+	srcW := bounds.Dx()
+	srcH := bounds.Dy()
+
+	// Convert to RGBA if needed
+	var src *image.RGBA
+	if rgba, ok := img.(*image.RGBA); ok {
+		src = rgba
+	} else {
+		src = image.NewRGBA(bounds)
+		for y := bounds.Min.Y; y < bounds.Max.Y; y++ {
+			for x := bounds.Min.X; x < bounds.Max.X; x++ {
+				src.Set(x, y, img.At(x, y))
+			}
+		}
+	}
+
+	// Keys cubic with a=-0.5 (PIL BICUBIC)
+	cubic := func(x float64) float64 {
+		if x < 0 {
+			x = -x
+		}
+		if x < 1 {
+			return 1.5*x*x*x - 2.5*x*x + 1
+		}
+		if x < 2 {
+			return -0.5*x*x*x + 2.5*x*x - 4*x + 2
+		}
+		return 0
+	}
+
+	// Horizontal pass: srcW -> width, keep srcH rows
+	temp := image.NewRGBA(image.Rect(0, 0, width, srcH))
+	for y := 0; y < srcH; y++ {
+		for dstX := 0; dstX < width; dstX++ {
+			// PIL coordinate mapping: center-to-center
+			srcXf := (float64(dstX)+0.5)*(float64(srcW)/float64(width)) - 0.5
+			baseX := int(math.Floor(srcXf))
+
+			var sumR, sumG, sumB, sumA, weightSum float64
+			for i := -1; i <= 2; i++ {
+				sx := baseX + i
+				if sx < 0 {
+					sx = 0
+				}
+				if sx >= srcW {
+					sx = srcW - 1
+				}
+
+				w := cubic(math.Abs(srcXf - float64(baseX+i)))
+				c := src.RGBAAt(sx, y)
+				sumR += float64(c.R) * w
+				sumG += float64(c.G) * w
+				sumB += float64(c.B) * w
+				sumA += float64(c.A) * w
+				weightSum += w
+			}
+
+			temp.SetRGBA(dstX, y, color.RGBA{
+				clampFloat(sumR, weightSum),
+				clampFloat(sumG, weightSum),
+				clampFloat(sumB, weightSum),
+				clampFloat(sumA, weightSum),
+			})
+		}
+	}
+
+	// Vertical pass: srcH -> height
+	dst := image.NewRGBA(image.Rect(0, 0, width, height))
+	for x := 0; x < width; x++ {
+		for dstY := 0; dstY < height; dstY++ {
+			srcYf := (float64(dstY)+0.5)*(float64(srcH)/float64(height)) - 0.5
+			baseY := int(math.Floor(srcYf))
+
+			var sumR, sumG, sumB, sumA, weightSum float64
+			for j := -1; j <= 2; j++ {
+				sy := baseY + j
+				if sy < 0 {
+					sy = 0
+				}
+				if sy >= srcH {
+					sy = srcH - 1
+				}
+
+				w := cubic(math.Abs(srcYf - float64(baseY+j)))
+				c := temp.RGBAAt(x, sy)
+				sumR += float64(c.R) * w
+				sumG += float64(c.G) * w
+				sumB += float64(c.B) * w
+				sumA += float64(c.A) * w
+				weightSum += w
+			}
+
+			dst.SetRGBA(x, dstY, color.RGBA{
+				clampFloat(sumR, weightSum),
+				clampFloat(sumG, weightSum),
+				clampFloat(sumB, weightSum),
+				clampFloat(sumA, weightSum),
+			})
+		}
+	}
+
+	return dst
+}
+
+// LoadAndPreprocessMultiple loads multiple images and preprocesses them
+// Returns: condImages (for vision encoder), vaeImages (for VAE encoding), dims (per-image dimensions)
+func (p *Processor) LoadAndPreprocessMultiple(imagePaths []string) ([]*mlx.Array, []*mlx.Array, []ImageDims, error) {
+	const vaeScaleFactor int32 = 8
+	const patchSize int32 = 2
+
+	condImages := make([]*mlx.Array, len(imagePaths))
+	vaeImages := make([]*mlx.Array, len(imagePaths))
+	dims := make([]ImageDims, len(imagePaths))
+
+	for i, imagePath := range imagePaths {
+		img, err := loadImageFile(imagePath)
+		if err != nil {
+			return nil, nil, nil, fmt.Errorf("image %d: %w", i, err)
+		}
+
+		bounds := img.Bounds()
+		origW := int32(bounds.Dx())
+		origH := int32(bounds.Dy())
+		ratio := float64(origW) / float64(origH)
+
+		// Calculate dimensions for condition image (vision encoder)
+		// Python pipeline does TWO resizes:
+		// 1. VaeImageProcessor.resize with Lanczos to CONDITION_IMAGE_SIZE (384x384 area)
+		// 2. Qwen2VLProcessor's smart_resize with Bicubic to multiple of 28
+		intermediateW, intermediateH := calculateDimensions(p.Config.ConditionImageSize, ratio, 32)
+		condH, condW := smartResize(intermediateH, intermediateW, 28, 56*56, 28*28*1280)
+
+		// Calculate dimensions for VAE image (1024x1024 area)
+		vaeW, vaeH := calculateDimensions(p.Config.VAEImageSize, ratio, 32)
+
+		// Calculate derived dimensions
+		latentW := vaeW / vaeScaleFactor
+		latentH := vaeH / vaeScaleFactor
+		patchW := latentW / patchSize
+		patchH := latentH / patchSize
+
+		dims[i] = ImageDims{
+			OrigW:   origW,
+			OrigH:   origH,
+			CondW:   condW,
+			CondH:   condH,
+			VaeW:    vaeW,
+			VaeH:    vaeH,
+			LatentW: latentW,
+			LatentH: latentH,
+			PatchW:  patchW,
+			PatchH:  patchH,
+		}
+
+		fmt.Printf("  Image %d: orig=%dx%d, cond=%dx%d, vae=%dx%d, latent=%dx%d, patch=%dx%d\n",
+			i+1, origW, origH, condW, condH, vaeW, vaeH, latentW, latentH, patchW, patchH)
+
+		// Preprocess for condition (vision encoder) - two-step resize to match Python pipeline
+		condImages[i] = p.preprocessImageTwoStep(img, intermediateW, intermediateH, condW, condH)
+
+		// Preprocess for VAE ([-1, 1] range, 5D tensor)
+		vaeImages[i] = p.preprocessImageForVAE(img, vaeW, vaeH)
+	}
+
+	return condImages, vaeImages, dims, nil
+}

x/imagegen/models/qwen_image_edit/qwen_image_edit.go

+//go:build mlx
+
+// Package qwen_image_edit implements the Qwen-Image-Edit diffusion model for image editing.
+// It reuses components from qwen_image where possible.
+package qwen_image_edit
+
+import (
+	"context"
+	"fmt"
+	"path/filepath"
+	"time"
+
+	"github.com/ollama/ollama/x/imagegen/mlx"
+	"github.com/ollama/ollama/x/imagegen/models/qwen_image"
+	"github.com/ollama/ollama/x/imagegen/tokenizer"
+)
+
+// GenerateConfig holds all options for image editing.
+type GenerateConfig struct {
+	Prompt         string
+	NegativePrompt string       // Unconditional prompt for CFG (empty string "" is valid)
+	CFGScale       float32      // CFG enabled when > 1.0 (default: 4.0)
+	Width          int32        // Output width (default: from input image)
+	Height         int32        // Output height (default: from input image)
+	Steps          int          // Denoising steps (default: 50)
+	Seed           int64        // Random seed
+	Progress       ProgressFunc // Optional progress callback
+}
+
+// ProgressFunc is called during generation with step progress.
+type ProgressFunc func(step, totalSteps int)
+
+// Model represents a Qwen-Image-Edit diffusion model.
+type Model struct {
+	ModelPath     string
+	Tokenizer     *tokenizer.Tokenizer
+	Processor     *Processor                // Image processor for vision encoder
+	TextEncoder   *qwen_image.Qwen25VL      // Qwen2.5-VL vision-language encoder (from qwen_image)
+	Transformer   *qwen_image.Transformer   // Reuse qwen_image transformer
+	VAE           *VAE                      // Combined encoder + decoder
+}
+
+// Load loads the Qwen-Image-Edit model from a directory.
+func (m *Model) Load(modelPath string) error {
+	fmt.Println("Loading Qwen-Image-Edit model...")
+	start := time.Now()
+
+	if mlx.GPUIsAvailable() {
+		mlx.SetDefaultDeviceGPU()
+		mlx.EnableCompile()
+	}
+
+	m.ModelPath = modelPath
+
+	// Load tokenizer from processor directory
+	fmt.Print("  Loading tokenizer... ")
+	processorPath := filepath.Join(modelPath, "processor")
+	tok, err := tokenizer.Load(processorPath)
+	if err != nil {
+		// Fallback to tokenizer directory
+		tokenizerPath := filepath.Join(modelPath, "tokenizer")
+		tok, err = tokenizer.Load(tokenizerPath)
+		if err != nil {
+			return fmt.Errorf("tokenizer: %w", err)
+		}
+	}
+	m.Tokenizer = tok
+	fmt.Println("✓")
+
+	// Load processor (image preprocessing config)
+	fmt.Print("  Loading processor... ")
+	m.Processor = &Processor{}
+	if err := m.Processor.Load(processorPath); err != nil {
+		return fmt.Errorf("processor: %w", err)
+	}
+	fmt.Println("✓")
+
+	// Load vision-language text encoder (Qwen2.5-VL from qwen_image package)
+	m.TextEncoder = &qwen_image.Qwen25VL{}
+	if err := m.TextEncoder.Load(filepath.Join(modelPath, "text_encoder")); err != nil {
+		return fmt.Errorf("text encoder: %w", err)
+	}
+	mlx.Eval(mlx.Collect(m.TextEncoder)...)
+	fmt.Printf("  (%.1f GB, peak %.1f GB)\n",
+		float64(mlx.MetalGetActiveMemory())/(1024*1024*1024),
+		float64(mlx.MetalGetPeakMemory())/(1024*1024*1024))
+
+	// Load transformer (reuse qwen_image)
+	m.Transformer = &qwen_image.Transformer{}
+	if err := m.Transformer.Load(filepath.Join(modelPath, "transformer")); err != nil {
+		return fmt.Errorf("transformer: %w", err)
+	}
+	mlx.Eval(mlx.Collect(m.Transformer)...)
+	fmt.Printf("  (%.1f GB, peak %.1f GB)\n",
+		float64(mlx.MetalGetActiveMemory())/(1024*1024*1024),
+		float64(mlx.MetalGetPeakMemory())/(1024*1024*1024))
+
+	// Load VAE (encoder + decoder)
+	m.VAE = &VAE{}
+	if err := m.VAE.Load(filepath.Join(modelPath, "vae")); err != nil {
+		return fmt.Errorf("VAE: %w", err)
+	}
+	mlx.Eval(mlx.Collect(m.VAE)...)
+	fmt.Printf("  (%.1f GB, peak %.1f GB)\n",
+		float64(mlx.MetalGetActiveMemory())/(1024*1024*1024),
+		float64(mlx.MetalGetPeakMemory())/(1024*1024*1024))
+
+	mem := mlx.MetalGetActiveMemory()
+	peak := mlx.MetalGetPeakMemory()
+	fmt.Printf("  Loaded in %.2fs (%.1f GB active, %.1f GB peak)\n",
+		time.Since(start).Seconds(),
+		float64(mem)/(1024*1024*1024),
+		float64(peak)/(1024*1024*1024))
+
+	return nil
+}
+
+// Edit edits an image based on a text prompt.
+// inputImagePath: path to input image
+// prompt: text description of desired edit
+func (m *Model) Edit(inputImagePath string, prompt string, width, height int32, steps int, seed int64) (*mlx.Array, error) {
+	return m.EditFromConfig([]string{inputImagePath}, &GenerateConfig{
+		Prompt: prompt,
+		Width:  width,
+		Height: height,
+		Steps:  steps,
+		Seed:   seed,
+	})
+}
+
+// EditFromConfig edits images using the unified config struct.
+// Accepts one or more input images.
+func (m *Model) EditFromConfig(inputImagePaths []string, cfg *GenerateConfig) (*mlx.Array, error) {
+	if len(inputImagePaths) == 0 {
+		return nil, fmt.Errorf("no input images provided")
+	}
+
+	start := time.Now()
+	result, err := m.edit(inputImagePaths, cfg)
+	if err != nil {
+		return nil, err
+	}
+
+	if cfg.NegativePrompt != "" {
+		fmt.Printf("Edited %d image(s) with CFG (scale=%.1f) in %.2fs (%d steps)\n",
+			len(inputImagePaths), cfg.CFGScale, time.Since(start).Seconds(), cfg.Steps)
+	} else {
+		fmt.Printf("Edited %d image(s) in %.2fs (%d steps)\n",
+			len(inputImagePaths), time.Since(start).Seconds(), cfg.Steps)
+	}
+	return result, nil
+}
+
+// EditImage implements model.ImageEditModel interface.
+func (m *Model) EditImage(ctx context.Context, inputImagePath, prompt string, width, height int32, steps int, seed int64) (*mlx.Array, error) {
+	return m.Edit(inputImagePath, prompt, width, height, steps, seed)
+}
+
+// EditMultiImage edits using multiple source images.
+// This matches diffusers' QwenImageEditPlusPipeline behavior.
+func (m *Model) EditMultiImage(inputImagePaths []string, cfg *GenerateConfig) (*mlx.Array, error) {
+	return m.EditFromConfig(inputImagePaths, cfg)
+}
+
+// edit is the internal editing pipeline that handles one or more images.
+func (m *Model) edit(inputImagePaths []string, cfg *GenerateConfig) (*mlx.Array, error) {
+	// Apply defaults
+	if cfg.Steps <= 0 {
+		cfg.Steps = 50
+	}
+	if cfg.CFGScale <= 0 {
+		cfg.CFGScale = 4.0
+	}
+
+	// Load and preprocess all input images
+	fmt.Printf("Loading %d image(s)...\n", len(inputImagePaths))
+	condImages, vaeImages, inputDims, err := m.Processor.LoadAndPreprocessMultiple(inputImagePaths)
+	if err != nil {
+		return nil, fmt.Errorf("preprocess images: %w", err)
+	}
+	for _, img := range condImages {
+		mlx.Keep(img)
+	}
+	for _, img := range vaeImages {
+		mlx.Keep(img)
+	}
+	mlx.Eval(append(condImages, vaeImages...)...)
+
+	useCFG := cfg.NegativePrompt != ""
+	tcfg := m.Transformer.Config
+	vaeScaleFactor := int32(8)
+
+	// Output dimensions - if not specified, use first input image dimensions
+	if cfg.Width <= 0 {
+		cfg.Width = inputDims[0].VaeW
+	}
+	if cfg.Height <= 0 {
+		cfg.Height = inputDims[0].VaeH
+	}
+
+	// Output (noise) latent dimensions
+	outLatentH := cfg.Height / vaeScaleFactor
+	outLatentW := cfg.Width / vaeScaleFactor
+	outPH := outLatentH / tcfg.PatchSize
+	outPW := outLatentW / tcfg.PatchSize
+	noiseSeqLen := outPH * outPW
+	imgSeqLen := noiseSeqLen
+
+	// Encode prompt with all images for conditioning
+	posEmb, _, _, err := m.TextEncoder.EncodePromptWithImages(m.Tokenizer, cfg.Prompt, condImages)
+	if err != nil {
+		return nil, fmt.Errorf("encoding prompt: %w", err)
+	}
+	mlx.Keep(posEmb)
+	mlx.Eval(posEmb)
+
+	var negEmb *mlx.Array
+	if useCFG {
+		negEmb, _, _, err = m.TextEncoder.EncodePromptWithImages(m.Tokenizer, cfg.NegativePrompt, condImages)
+		if err != nil {
+			return nil, fmt.Errorf("encoding negative prompt: %w", err)
+		}
+		mlx.Keep(negEmb)
+		mlx.Eval(negEmb)
+	}
+
+	// Pad sequences to same length for CFG
+	txtLen := posEmb.Shape()[1]
+	if useCFG {
+		negLen := negEmb.Shape()[1]
+		if negLen > txtLen {
+			txtLen = negLen
+		}
+		if posEmb.Shape()[1] < txtLen {
+			posEmb = padSequence(posEmb, txtLen)
+		}
+		if negEmb.Shape()[1] < txtLen {
+			negEmb = padSequence(negEmb, txtLen)
+		}
+		mlx.Keep(posEmb, negEmb)
+		mlx.Eval(posEmb, negEmb)
+	}
+
+	// Encode all input images to latents and concatenate
+	fmt.Println("Encoding images to latents...")
+	allImageLatentsPacked := make([]*mlx.Array, len(vaeImages))
+	for i, vaeImage := range vaeImages {
+		imageLatents := m.VAE.Encode(vaeImage)
+		imageLatents = m.VAE.Normalize(imageLatents)
+		imageLatents2D := mlx.Squeeze(imageLatents, 2)
+		packed := qwen_image.PackLatents(imageLatents2D, tcfg.PatchSize)
+		mlx.Keep(packed)
+		mlx.Eval(packed)
+		allImageLatentsPacked[i] = packed
+	}
+
+	imageLatentsPacked := mlx.Concatenate(allImageLatentsPacked, 1)
+	mlx.Keep(imageLatentsPacked)
+	mlx.Eval(imageLatentsPacked)
+
+	// Scheduler
+	scheduler := qwen_image.NewFlowMatchScheduler(qwen_image.DefaultSchedulerConfig())
+	scheduler.SetTimesteps(cfg.Steps, noiseSeqLen)
+
+	// Init noise latents in packed format
+	packedChannels := tcfg.OutChannels * tcfg.PatchSize * tcfg.PatchSize
+	packedNoise := scheduler.InitNoisePacked(1, noiseSeqLen, packedChannels, cfg.Seed)
+	latents := qwen_image.UnpackLatents(packedNoise, outLatentH, outLatentW, tcfg.PatchSize)
+	mlx.Eval(latents)
+
+	// RoPE cache
+	ropeCache := PrepareRoPEMultiImage(outPH, outPW, inputDims, txtLen, tcfg.AxesDimsRope)
+	mlx.Keep(ropeCache.ImgFreqs, ropeCache.TxtFreqs)
+	mlx.Eval(ropeCache.ImgFreqs, ropeCache.TxtFreqs)
+
+	// Denoising loop
+	fmt.Printf("Running denoising (%d steps)...\n", cfg.Steps)
+	for i := 0; i < cfg.Steps; i++ {
+		stepStart := time.Now()
+		if cfg.Progress != nil {
+			cfg.Progress(i+1, cfg.Steps)
+		}
+
+		t := scheduler.Timesteps[i]
+		timestep := mlx.ToBFloat16(mlx.NewArray([]float32{t}, []int32{1}))
+		mlx.Eval(timestep)
+
+		latents2D := mlx.Squeeze(latents, 2)
+		patches := qwen_image.PackLatents(latents2D, tcfg.PatchSize)
+		latentInput := mlx.Concatenate([]*mlx.Array{patches, imageLatentsPacked}, 1)
+
+		var output *mlx.Array
+		if useCFG {
+			posOutput := m.Transformer.Forward(latentInput, posEmb, timestep, ropeCache.ImgFreqs, ropeCache.TxtFreqs)
+			negOutput := m.Transformer.Forward(latentInput, negEmb, timestep, ropeCache.ImgFreqs, ropeCache.TxtFreqs)
+
+			posOutput = mlx.Slice(posOutput, []int32{0, 0, 0}, []int32{1, imgSeqLen, posOutput.Shape()[2]})
+			negOutput = mlx.Slice(negOutput, []int32{0, 0, 0}, []int32{1, imgSeqLen, negOutput.Shape()[2]})
+
+			output = applyCFGWithNormRescale(posOutput, negOutput, cfg.CFGScale)
+		} else {
+			output = m.Transformer.Forward(latentInput, posEmb, timestep, ropeCache.ImgFreqs, ropeCache.TxtFreqs)
+			output = mlx.Slice(output, []int32{0, 0, 0}, []int32{1, imgSeqLen, output.Shape()[2]})
+		}
+
+		noisePred := qwen_image.UnpackLatents(output, outLatentH, outLatentW, tcfg.PatchSize)
+		oldLatents := latents
+		latents = scheduler.Step(noisePred, latents, i)
+		mlx.Eval(latents)
+		oldLatents.Free()
+
+		fmt.Printf("  Step %d/%d: t=%.4f (%.2fs)\n", i+1, cfg.Steps, t, time.Since(stepStart).Seconds())
+	}
+
+	// Free denoising temporaries
+	posEmb.Free()
+	if negEmb != nil {
+		negEmb.Free()
+	}
+	ropeCache.ImgFreqs.Free()
+	ropeCache.TxtFreqs.Free()
+	imageLatentsPacked.Free()
+
+	// Decode latents
+	decoded := m.decodeAndPostprocess(latents)
+	latents.Free()
+
+	fmt.Printf("  Peak memory: %.2f GB\n", float64(mlx.MetalGetPeakMemory())/(1024*1024*1024))
+	return decoded, nil
+}
+
+// applyCFGWithNormRescale applies classifier-free guidance with norm rescaling.
+// This prevents CFG from inflating magnitude too much.
+func applyCFGWithNormRescale(posOutput, negOutput *mlx.Array, scale float32) *mlx.Array {
+	// Upcast to float32 for precision
+	posF32 := mlx.AsType(posOutput, mlx.DtypeFloat32)
+	negF32 := mlx.AsType(negOutput, mlx.DtypeFloat32)
+
+	// CFG: pred = neg + scale * (pos - neg)
+	diff := mlx.Sub(posF32, negF32)
+	scaledDiff := mlx.MulScalar(diff, scale)
+	combPred := mlx.Add(negF32, scaledDiff)
+
+	// Norm rescaling: rescale combined prediction to match conditional norm
+	condNorm := mlx.Sqrt(mlx.Sum(mlx.Square(posF32), -1, true))
+	combNorm := mlx.Sqrt(mlx.Sum(mlx.Square(combPred), -1, true))
+	output := mlx.Mul(combPred, mlx.Div(condNorm, combNorm))
+
+	mlx.Eval(output)
+	return mlx.ToBFloat16(output)
+}
+
+// decodeAndPostprocess denormalizes latents, decodes through VAE, and scales to [0,1].
+func (m *Model) decodeAndPostprocess(latents *mlx.Array) *mlx.Array {
+	latents = m.VAE.Denormalize(latents)
+	decoded := m.VAE.Decode(latents)
+
+	// Post-process: squeeze temporal dim and rescale to [0, 1]
+	decoded = mlx.Squeeze(decoded, 2)
+	decoded = mlx.AddScalar(decoded, 1.0)
+	decoded = mlx.DivScalar(decoded, 2.0)
+	decoded = mlx.ClipScalar(decoded, 0.0, 1.0, true, true)
+	mlx.Eval(decoded)
+	return decoded
+}
+
+// padSequence pads a sequence tensor to the target length with zeros
+func padSequence(x *mlx.Array, targetLen int32) *mlx.Array {
+	shape := x.Shape()
+	currentLen := shape[1]
+	if currentLen >= targetLen {
+		return x
+	}
+	padLen := targetLen - currentLen
+	// Pad on sequence dimension (axis 1)
+	return mlx.Pad(x, []int32{0, 0, 0, padLen, 0, 0})
+}
+
+// LoadPersistent is an alias for backward compatibility.
+func LoadPersistent(modelPath string) (*Model, error) {
+	m := &Model{}
+	if err := m.Load(modelPath); err != nil {
+		return nil, err
+	}
+	return m, nil
+}
+
+// PrepareRoPEMultiImage computes RoPE with interpolation for image editing.
+// Handles single or multiple input images with different resolutions.
+//
+// Parameters:
+//   - outPH, outPW: output patch dimensions (noise latent resolution)
+//   - inputDims: patch dimensions for each input image [(pH1, pW1), (pH2, pW2), ...]
+//   - txtLen: text sequence length
+//   - axesDims: RoPE axis dimensions [16, 56, 56]
+//
+// Returns RoPE cache where:
+//   - ImgFreqs has (outPH*outPW + sum(inPH*inPW for each image)) positions
+//   - First outPH*outPW positions are for noise latents (standard RoPE at output res)
+//   - Following positions are for each input image (interpolated from output res)
+func PrepareRoPEMultiImage(outPH, outPW int32, inputDims []ImageDims, txtLen int32, axesDims []int32) *qwen_image.RoPECache {
+	theta := float64(10000)
+	maxIdx := int32(4096)
+
+	// Compute base frequencies for each axis dimension
+	freqsT := qwen_image.ComputeAxisFreqs(axesDims[0], theta)
+	freqsH := qwen_image.ComputeAxisFreqs(axesDims[1], theta)
+	freqsW := qwen_image.ComputeAxisFreqs(axesDims[2], theta)
+
+	// Build frequency lookup tables
+	posFreqsT := qwen_image.MakeFreqTable(maxIdx, freqsT, false)
+	posFreqsH := qwen_image.MakeFreqTable(maxIdx, freqsH, false)
+	posFreqsW := qwen_image.MakeFreqTable(maxIdx, freqsW, false)
+	negFreqsT := qwen_image.MakeFreqTable(maxIdx, freqsT, true) // For frame -1 on last condition image
+	negFreqsH := qwen_image.MakeFreqTable(maxIdx, freqsH, true)
+	negFreqsW := qwen_image.MakeFreqTable(maxIdx, freqsW, true)
+
+	headDim := int32(len(freqsT)+len(freqsH)+len(freqsW)) * 2
+
+	// Helper to compute RoPE for a single position at output resolution with scale_rope
+	computePosFreqs := func(framePos, y, x int32) []float32 {
+		row := make([]float32, headDim)
+		idx := 0
+
+		// Frame position
+		for i := 0; i < len(freqsT)*2; i++ {
+			row[idx+i] = posFreqsT[framePos][i]
+		}
+		idx += len(freqsT) * 2
+
+		// Height with scale_rope centering (using OUTPUT dimensions)
+		outHHalf := outPH / 2
+		hNegCount := outPH - outHHalf
+		if y < hNegCount {
+			negTableIdx := maxIdx - hNegCount + y
+			for i := 0; i < len(freqsH)*2; i++ {
+				row[idx+i] = negFreqsH[negTableIdx][i]
+			}
+		} else {
+			posIdx := y - hNegCount
+			for i := 0; i < len(freqsH)*2; i++ {
+				row[idx+i] = posFreqsH[posIdx][i]
+			}
+		}
+		idx += len(freqsH) * 2
+
+		// Width with scale_rope centering (using OUTPUT dimensions)
+		outWHalf := outPW / 2
+		wNegCount := outPW - outWHalf
+		if x < wNegCount {
+			negTableIdx := maxIdx - wNegCount + x
+			for i := 0; i < len(freqsW)*2; i++ {
+				row[idx+i] = negFreqsW[negTableIdx][i]
+			}
+		} else {
+			posIdx := x - wNegCount
+			for i := 0; i < len(freqsW)*2; i++ {
+				row[idx+i] = posFreqsW[posIdx][i]
+			}
+		}
+
+		return row
+	}
+
+	// Helper to compute RoPE for frame -1 (used for last condition image)
+	// This matches Python's _compute_condition_freqs which uses freqs_neg[0][-1:]
+	computeNegFrameFreqs := func(y, x int32) []float32 {
+		row := make([]float32, headDim)
+		idx := 0
+
+		// Frame -1: use last row of negative frame frequencies
+		negFrameIdx := maxIdx - 1
+		for i := 0; i < len(freqsT)*2; i++ {
+			row[idx+i] = negFreqsT[negFrameIdx][i]
+		}
+		idx += len(freqsT) * 2
+
+		// Height with scale_rope centering (using OUTPUT dimensions)
+		outHHalf := outPH / 2
+		hNegCount := outPH - outHHalf
+		if y < hNegCount {
+			negTableIdx := maxIdx - hNegCount + y
+			for i := 0; i < len(freqsH)*2; i++ {
+				row[idx+i] = negFreqsH[negTableIdx][i]
+			}
+		} else {
+			posIdx := y - hNegCount
+			for i := 0; i < len(freqsH)*2; i++ {
+				row[idx+i] = posFreqsH[posIdx][i]
+			}
+		}
+		idx += len(freqsH) * 2
+
+		// Width with scale_rope centering (using OUTPUT dimensions)
+		outWHalf := outPW / 2
+		wNegCount := outPW - outWHalf
+		if x < wNegCount {
+			negTableIdx := maxIdx - wNegCount + x
+			for i := 0; i < len(freqsW)*2; i++ {
+				row[idx+i] = negFreqsW[negTableIdx][i]
+			}
+		} else {
+			posIdx := x - wNegCount
+			for i := 0; i < len(freqsW)*2; i++ {
+				row[idx+i] = posFreqsW[posIdx][i]
+			}
+		}
+
+		return row
+	}
+
+	// Total image sequence length: noise + all input images
+	noiseSeqLen := outPH * outPW
+	totalImgLen := noiseSeqLen
+	for _, dims := range inputDims {
+		totalImgLen += dims.PatchH * dims.PatchW
+	}
+
+	imgFreqsData := make([]float32, totalImgLen*headDim)
+	idx := int32(0)
+
+	// Segment 0: Noise latents - standard RoPE at output resolution (frame 0)
+	for y := int32(0); y < outPH; y++ {
+		for x := int32(0); x < outPW; x++ {
+			row := computePosFreqs(0, y, x)
+			copy(imgFreqsData[idx:], row)
+			idx += headDim
+		}
+	}
+
+	// Segments 1..N: Edit image latents - INTERPOLATED RoPE
+	// For single image: use frame 1 (matches original PrepareRoPEInterpolated)
+	// For multiple images: Python uses frame -1 for the LAST condition image
+	// (_compute_condition_freqs), positive indices for others.
+	numImages := len(inputDims)
+	lastImgIdx := numImages - 1
+	for imgIdx, dims := range inputDims {
+		inPH := dims.PatchH
+		inPW := dims.PatchW
+
+		// Determine frame index for this image
+		// Single image case: use frame 1 (like original PrepareRoPEInterpolated)
+		// Multi-image case: last image uses frame -1, others use frame 1, 2, etc.
+		useNegFrame := numImages > 1 && imgIdx == lastImgIdx
+
+		// Map each input position to an output position using linear interpolation
+		for y := int32(0); y < inPH; y++ {
+			for x := int32(0); x < inPW; x++ {
+				// Interpolate: map input (y, x) to output grid position
+				// This is the key fix from DiffSynth's forward_sampling
+				var yOut, xOut int32
+				if inPH == 1 {
+					yOut = 0
+				} else {
+					// Linear interpolation: y_out = y * (outPH - 1) / (inPH - 1)
+					yOut = y * (outPH - 1) / (inPH - 1)
+				}
+				if inPW == 1 {
+					xOut = 0
+				} else {
+					xOut = x * (outPW - 1) / (inPW - 1)
+				}
+
+				var row []float32
+				if useNegFrame {
+					// Last image in multi-image uses frame -1
+					row = computeNegFrameFreqs(yOut, xOut)
+				} else {
+					// Single image uses frame 1, multi-image uses frame 1, 2, etc.
+					frameIdx := int32(imgIdx + 1)
+					row = computePosFreqs(frameIdx, yOut, xOut)
+				}
+				copy(imgFreqsData[idx:], row)
+				idx += headDim
+			}
+		}
+	}
+
+	imgFreqs := mlx.NewArray(imgFreqsData, []int32{totalImgLen, headDim})
+	imgFreqs = mlx.ToBFloat16(imgFreqs)
+
+	// Text frequencies - start after max video index
+	maxVidIdx := max(outPH/2, outPW/2)
+
+	txtFreqsData := make([]float32, txtLen*headDim)
+	idx = 0
+	for t := int32(0); t < txtLen; t++ {
+		pos := maxVidIdx + t
+		for i := 0; i < len(freqsT)*2; i++ {
+			txtFreqsData[idx+int32(i)] = posFreqsT[pos][i]
+		}
+		idx += int32(len(freqsT) * 2)
+		for i := 0; i < len(freqsH)*2; i++ {
+			txtFreqsData[idx+int32(i)] = posFreqsH[pos][i]
+		}
+		idx += int32(len(freqsH) * 2)
+		for i := 0; i < len(freqsW)*2; i++ {
+			txtFreqsData[idx+int32(i)] = posFreqsW[pos][i]
+		}
+		idx += int32(len(freqsW) * 2)
+	}
+
+	txtFreqs := mlx.NewArray(txtFreqsData, []int32{txtLen, headDim})
+	txtFreqs = mlx.ToBFloat16(txtFreqs)
+
+	return &qwen_image.RoPECache{
+		ImgFreqs: imgFreqs,
+		TxtFreqs: txtFreqs,
+	}
+}

x/imagegen/models/qwen_image_edit/rope_test.go

+//go:build mlx
+
+package qwen_image_edit
+
+import (
+	"math"
+	"testing"
+
+	"github.com/ollama/ollama/x/imagegen/mlx"
+	"github.com/ollama/ollama/x/imagegen/models/qwen_image"
+)
+
+// TestComputeAxisFreqs verifies frequency computation matches Python reference
+func TestComputeAxisFreqs(t *testing.T) {
+	theta := float64(10000)
+
+	// Expected values from Python:
+	// freqs = 1.0 / (theta ** (np.arange(0, half_dim) / half_dim))
+	expectedFreqsT := []float64{
+		1.000000000000000, 0.316227766016838, 0.100000000000000, 0.031622776601684,
+		0.010000000000000, 0.003162277660168, 0.001000000000000, 0.000316227766017,
+	}
+
+	expectedFreqsH_first4 := []float64{
+		1.000000000000000, 0.719685673001152, 0.517947467923121, 0.372759372031494,
+	}
+
+	expectedFreqsH_last4 := []float64{
+		0.000372759372031, 0.000268269579528, 0.000193069772888, 0.000138949549437,
+	}
+
+	// Test temporal frequencies (dim=16)
+	freqsT := qwen_image.ComputeAxisFreqs(16, theta)
+	if len(freqsT) != 8 {
+		t.Fatalf("expected 8 temporal frequencies, got %d", len(freqsT))
+	}
+	for i, expected := range expectedFreqsT {
+		if diff := math.Abs(freqsT[i] - expected); diff > 1e-10 {
+			t.Errorf("freqsT[%d]: expected %.15f, got %.15f, diff %.2e", i, expected, freqsT[i], diff)
+		}
+	}
+
+	// Test height/width frequencies (dim=56)
+	freqsH := qwen_image.ComputeAxisFreqs(56, theta)
+	if len(freqsH) != 28 {
+		t.Fatalf("expected 28 height frequencies, got %d", len(freqsH))
+	}
+	for i, expected := range expectedFreqsH_first4 {
+		if diff := math.Abs(freqsH[i] - expected); diff > 1e-10 {
+			t.Errorf("freqsH[%d]: expected %.15f, got %.15f, diff %.2e", i, expected, freqsH[i], diff)
+		}
+	}
+	for i, expected := range expectedFreqsH_last4 {
+		idx := 24 + i // last 4 of 28
+		if diff := math.Abs(freqsH[idx] - expected); diff > 1e-10 {
+			t.Errorf("freqsH[%d]: expected %.15f, got %.15f, diff %.2e", idx, expected, freqsH[idx], diff)
+		}
+	}
+}
+
+// TestMakeFreqTable verifies the frequency lookup table for both positive and negative positions
+func TestMakeFreqTable(t *testing.T) {
+	theta := float64(10000)
+	freqsT := qwen_image.ComputeAxisFreqs(16, theta)
+	maxIdx := int32(4096)
+
+	// Test positive table
+	posTable := qwen_image.MakeFreqTable(maxIdx, freqsT, false)
+
+	// Position 0 should give cos=1, sin=0 for all frequencies
+	for i := 0; i < len(freqsT)*2; i += 2 {
+		if posTable[0][i] != 1.0 {
+			t.Errorf("posTable[0][%d] (cos): expected 1.0, got %f", i, posTable[0][i])
+		}
+		if posTable[0][i+1] != 0.0 {
+			t.Errorf("posTable[0][%d] (sin): expected 0.0, got %f", i+1, posTable[0][i+1])
+		}
+	}
+
+	// Position 1, first frequency (1.0): angle = 1*1 = 1
+	// cos(1) = 0.5403, sin(1) = 0.8415
+	if diff := math.Abs(float64(posTable[1][0]) - 0.5403023058681398); diff > 1e-6 {
+		t.Errorf("posTable[1][0] (cos): expected 0.5403, got %f", posTable[1][0])
+	}
+	if diff := math.Abs(float64(posTable[1][1]) - 0.8414709848078965); diff > 1e-6 {
+		t.Errorf("posTable[1][1] (sin): expected 0.8415, got %f", posTable[1][1])
+	}
+
+	// Test negative table
+	negTable := qwen_image.MakeFreqTable(maxIdx, freqsT, true)
+
+	// negTable[4095] corresponds to position -1
+	// cos(-1) = cos(1), sin(-1) = -sin(1)
+	if diff := math.Abs(float64(negTable[4095][0]) - 0.5403023058681398); diff > 1e-6 {
+		t.Errorf("negTable[4095][0] (cos(-1)): expected 0.5403, got %f", negTable[4095][0])
+	}
+	if diff := math.Abs(float64(negTable[4095][1]) - (-0.8414709848078965)); diff > 1e-6 {
+		t.Errorf("negTable[4095][1] (sin(-1)): expected -0.8415, got %f", negTable[4095][1])
+	}
+
+	// negTable[4094] corresponds to position -2
+	// cos(-2) = cos(2), sin(-2) = -sin(2)
+	cos2 := math.Cos(2.0)
+	sin2 := math.Sin(2.0)
+	if diff := math.Abs(float64(negTable[4094][0]) - cos2); diff > 1e-6 {
+		t.Errorf("negTable[4094][0] (cos(-2)): expected %f, got %f", cos2, negTable[4094][0])
+	}
+	if diff := math.Abs(float64(negTable[4094][1]) - (-sin2)); diff > 1e-6 {
+		t.Errorf("negTable[4094][1] (sin(-2)): expected %f, got %f", -sin2, negTable[4094][1])
+	}
+}
+
+// TestPrepareRoPE_QwenImage verifies qwen_image.PrepareRoPE for single-segment case
+func TestPrepareRoPE_QwenImage(t *testing.T) {
+	if !mlx.GPUIsAvailable() {
+		t.Skip("GPU not available")
+	}
+
+	mlx.SetDefaultDeviceCPU()
+
+	// 4x4 patch grid, single image
+	imgH, imgW := int32(4), int32(4)
+	txtLen := int32(5)
+	axesDims := []int32{16, 56, 56}
+
+	cache := qwen_image.PrepareRoPE(imgH, imgW, txtLen, axesDims)
+	mlx.Eval(cache.ImgFreqs, cache.TxtFreqs)
+
+	// Check shapes
+	imgShape := cache.ImgFreqs.Shape()
+	if imgShape[0] != 16 { // 4*4 patches
+		t.Errorf("ImgFreqs seq len: expected 16, got %d", imgShape[0])
+	}
+
+	// For single image (frame=0), all temporal values should be cos=1, sin=0
+	imgFreqsCPU := mlx.AsType(cache.ImgFreqs, mlx.DtypeFloat32)
+	mlx.Eval(imgFreqsCPU)
+	imgData := imgFreqsCPU.Data()
+
+	// Check first 16 values of patch 0 (temporal cos/sin pairs)
+	for i := 0; i < 16; i += 2 {
+		cosVal := imgData[i]
+		sinVal := imgData[i+1]
+		if diff := math.Abs(float64(cosVal - 1.0)); diff > 1e-5 {
+			t.Errorf("ImgFreqs[0][%d] (cos): expected 1.0, got %f", i, cosVal)
+		}
+		if diff := math.Abs(float64(sinVal - 0.0)); diff > 1e-5 {
+			t.Errorf("ImgFreqs[0][%d] (sin): expected 0.0, got %f", i+1, sinVal)
+		}
+	}
+
+	cache.ImgFreqs.Free()
+	cache.TxtFreqs.Free()
+}
+
+// TestScaleRopePositions verifies the centered position calculation for scale_rope=True
+func TestScaleRopePositions(t *testing.T) {
+	// For a 4x4 grid with scale_rope=True:
+	// hHalf = 2, wHalf = 2
+	// hNegCount = 4 - 2 = 2 (positions 0,1 are negative)
+	// wNegCount = 4 - 2 = 2 (positions 0,1 are negative)
+	//
+	// Height positions:
+	//   y=0: -(4-2) + 0 = -2
+	//   y=1: -(4-2) + 1 = -1
+	//   y=2: 2 - 2 = 0
+	//   y=3: 3 - 2 = 1
+	//
+	// Same for width
+
+	pH, pW := int32(4), int32(4)
+	hHalf := pH / 2
+	wHalf := pW / 2
+	hNegCount := pH - hHalf
+	wNegCount := pW - wHalf
+
+	expectedH := []int32{-2, -1, 0, 1}
+	expectedW := []int32{-2, -1, 0, 1}
+
+	for y := int32(0); y < pH; y++ {
+		var hPos int32
+		if y < hNegCount {
+			hPos = -(pH - hHalf) + y
+		} else {
+			hPos = y - hNegCount
+		}
+		if hPos != expectedH[y] {
+			t.Errorf("y=%d: expected h_pos=%d, got %d", y, expectedH[y], hPos)
+		}
+	}
+
+	for x := int32(0); x < pW; x++ {
+		var wPos int32
+		if x < wNegCount {
+			wPos = -(pW - wHalf) + x
+		} else {
+			wPos = x - wNegCount
+		}
+		if wPos != expectedW[x] {
+			t.Errorf("x=%d: expected w_pos=%d, got %d", x, expectedW[x], wPos)
+		}
+	}
+}
+
+// TestRoPEHeadDimensions verifies the head dimension breakdown
+func TestRoPEHeadDimensions(t *testing.T) {
+	// axes_dims_rope = [16, 56, 56]
+	// Each dimension uses half the values for frequencies
+	// So we get: 8 + 28 + 28 = 64 frequency values
+	// Each frequency produces cos + sin, so: 64 * 2 = 128 total values per position
+
+	axesDims := []int32{16, 56, 56}
+	expectedFreqs := (axesDims[0]/2 + axesDims[1]/2 + axesDims[2]/2)
+	expectedHeadDim := expectedFreqs * 2
+
+	if expectedFreqs != 64 {
+		t.Errorf("expected 64 frequency values, got %d", expectedFreqs)
+	}
+	if expectedHeadDim != 128 {
+		t.Errorf("expected head_dim=128, got %d", expectedHeadDim)
+	}
+
+	// This should match the transformer's attention head dimension
+	// hidden_size = 3072, num_heads = 24
+	// head_dim = 3072 / 24 = 128
+}
+

x/imagegen/models/qwen_image_edit/vae.go

+//go:build mlx
+
+package qwen_image_edit
+
+import (
+	"fmt"
+
+	"github.com/ollama/ollama/x/imagegen/mlx"
+	"github.com/ollama/ollama/x/imagegen/safetensors"
+)
+
+// VAEConfig holds Qwen-Image VAE configuration
+type VAEConfig struct {
+	ZDim               int32     `json:"z_dim"`               // 16
+	BaseDim            int32     `json:"base_dim"`            // 96
+	DimMult            []int32   `json:"dim_mult"`            // [1, 2, 4, 4]
+	NumResBlocks       int32     `json:"num_res_blocks"`      // 2
+	LatentsMean        []float32 `json:"latents_mean"`        // 16 values
+	LatentsStd         []float32 `json:"latents_std"`         // 16 values
+	TemperalDownsample []bool    `json:"temperal_downsample"` // [false, true, true]
+}
+
+// defaultVAEConfig returns config for Qwen-Image VAE
+func defaultVAEConfig() *VAEConfig {
+	return &VAEConfig{
+		ZDim:         16,
+		BaseDim:      96,
+		DimMult:      []int32{1, 2, 4, 4},
+		NumResBlocks: 2,
+		LatentsMean: []float32{
+			-0.7571, -0.7089, -0.9113, 0.1075,
+			-0.1745, 0.9653, -0.1517, 1.5508,
+			0.4134, -0.0715, 0.5517, -0.3632,
+			-0.1922, -0.9497, 0.2503, -0.2921,
+		},
+		LatentsStd: []float32{
+			2.8184, 1.4541, 2.3275, 2.6558,
+			1.2196, 1.7708, 2.6052, 2.0743,
+			3.2687, 2.1526, 2.8652, 1.5579,
+			1.6382, 1.1253, 2.8251, 1.916,
+		},
+		TemperalDownsample: []bool{false, true, true},
+	}
+}
+
+// VAE is the full VAE with encoder and decoder
+type VAE struct {
+	Config  *VAEConfig
+	Encoder *VAEEncoder
+	Decoder *VAEDecoder
+}
+
+// Load loads the VAE from a directory
+func (m *VAE) Load(path string) error {
+	fmt.Println("Loading Qwen-Image-Edit VAE (encoder + decoder)...")
+
+	cfg := defaultVAEConfig()
+	m.Config = cfg
+
+	weights, err := safetensors.LoadModelWeights(path)
+	if err != nil {
+		return fmt.Errorf("weights: %w", err)
+	}
+
+	// Load weights as f32 for quality (matches Python default behavior)
+	// VAE decoder precision is critical for final image quality
+	fmt.Print("  Loading weights as f32... ")
+	if err := weights.Load(mlx.DtypeFloat32); err != nil {
+		return fmt.Errorf("failed to load weights: %w", err)
+	}
+	fmt.Printf("✓ (%.1f GB)\n", float64(mlx.MetalGetActiveMemory())/(1024*1024*1024))
+
+	// Load encoder
+	fmt.Print("  Loading encoder... ")
+	m.Encoder = &VAEEncoder{}
+	if err := m.Encoder.loadFromWeights(weights, cfg); err != nil {
+		return fmt.Errorf("encoder: %w", err)
+	}
+	fmt.Println("✓")
+
+	// Load decoder
+	fmt.Print("  Loading decoder... ")
+	m.Decoder = &VAEDecoder{}
+	if err := m.Decoder.loadFromWeights(weights, cfg); err != nil {
+		return fmt.Errorf("decoder: %w", err)
+	}
+	fmt.Println("✓")
+
+	weights.ReleaseAll()
+	return nil
+}
+
+// Encode encodes an image to latents
+// x: [B, C, T, H, W] image tensor in [-1, 1] range
+// Returns: [B, C, T, H/8, W/8] latents (unnormalized)
+func (m *VAE) Encode(x *mlx.Array) *mlx.Array {
+	return m.Encoder.Encode(x)
+}
+
+// Decode decodes latents to image
+// z: [B, C, T, H, W] latents (denormalized)
+// Returns: [B, C, T, H*8, W*8] image in [-1, 1]
+func (m *VAE) Decode(z *mlx.Array) *mlx.Array {
+	return m.Decoder.Decode(z)
+}
+
+// Normalize applies latent normalization
+// Input z should be f32 (from VAE encoder), output is f32 for transformer
+func (m *VAE) Normalize(z *mlx.Array) *mlx.Array {
+	shape := z.Shape()
+	C := shape[1]
+
+	mean := mlx.NewArray(m.Config.LatentsMean[:C], []int32{1, C, 1, 1, 1})
+	std := mlx.NewArray(m.Config.LatentsStd[:C], []int32{1, C, 1, 1, 1})
+
+	// Mean/std are f32, will match z dtype through broadcasting
+	return mlx.Div(mlx.Sub(z, mean), std)
+}
+
+// Denormalize reverses latent normalization
+// Input z is bf16 (from transformer), output converted to f32 for VAE decoder
+func (m *VAE) Denormalize(z *mlx.Array) *mlx.Array {
+	shape := z.Shape()
+	C := shape[1]
+
+	// Convert latents to f32 for VAE decoder quality
+	z = mlx.AsType(z, mlx.DtypeFloat32)
+
+	mean := mlx.NewArray(m.Config.LatentsMean[:C], []int32{1, C, 1, 1, 1})
+	std := mlx.NewArray(m.Config.LatentsStd[:C], []int32{1, C, 1, 1, 1})
+
+	return mlx.Add(mlx.Mul(z, std), mean)
+}
+
+// VAEEncoder is the encoder part of the VAE
+// The encoder uses a flat structure where down_blocks contains a mix of ResBlocks and Downsamplers:
+// - Blocks 0,1: ResBlocks (base_dim)
+// - Block 2: Downsample
+// - Blocks 3,4: ResBlocks (base_dim*2)
+// - Block 5: Downsample + temporal
+// - Blocks 6,7: ResBlocks (base_dim*4)
+// - Block 8: Downsample + temporal
+// - Blocks 9,10: ResBlocks (base_dim*4)
+type VAEEncoder struct {
+	Config *VAEConfig
+
+	ConvIn     *CausalConv3d
+	Blocks     []EncoderBlock // Flat list of ResBlocks and Downsamplers
+	MidBlock   *MidBlock
+	NormOut    *RMSNorm3D
+	ConvOut    *CausalConv3d
+	QuantConv  *CausalConv3d
+}
+
+// EncoderBlock is either a ResBlock or a Downsample
+type EncoderBlock interface {
+	Forward(x *mlx.Array) *mlx.Array
+	IsDownsample() bool
+}
+
+// EncoderResBlock wraps ResBlock
+type EncoderResBlock struct {
+	*ResBlock
+}
+
+func (b *EncoderResBlock) IsDownsample() bool { return false }
+
+// EncoderDownsample is a downsample layer
+type EncoderDownsample struct {
+	Resample *CausalConv3d
+	TimeConv *CausalConv3d // Optional temporal downsample
+}
+
+func (d *EncoderDownsample) IsDownsample() bool { return true }
+
+func (d *EncoderDownsample) Forward(x *mlx.Array) *mlx.Array {
+	// Spatial downsample with stride 2
+	// WAN VAE uses: ZeroPad2d(0,1,0,1) + Conv2d(3x3, stride=2)
+	x = d.forwardSpatialDownsample(x)
+
+	// NOTE: In WAN VAE, time_conv is ONLY used in streaming/chunked mode
+	// with feat_cache. For single-frame encoding (T=1), time_conv is skipped.
+	// The Python forward checks: if feat_cache is not None ... then use time_conv
+	// Since we don't support streaming, we skip time_conv entirely.
+	return x
+}
+
+// forwardSpatialDownsample applies 2D conv with stride 2 for spatial downsampling
+func (d *EncoderDownsample) forwardSpatialDownsample(x *mlx.Array) *mlx.Array {
+	xShape := x.Shape()
+	B := xShape[0]
+	T := xShape[1]
+	H := xShape[2]
+	W := xShape[3]
+	C := xShape[4]
+
+	wShape := d.Resample.Weight.Shape()
+	outC := wShape[0]
+
+	// Reshape to [B*T, H, W, C] for 2D conv
+	x = mlx.Reshape(x, B*T, H, W, C)
+
+	// Asymmetric padding: pad right and bottom by 1 (WAN VAE style)
+	// ZeroPad2d(0, 1, 0, 1) means (left=0, right=1, top=0, bottom=1)
+	x = mlx.Pad(x, []int32{0, 0, 0, 1, 0, 1, 0, 0}) // [B, H, W, C] -> pad H and W
+
+	// Apply 2D conv with stride 2
+	weight := mlx.Transpose(d.Resample.Weight, 0, 2, 3, 1) // [O, I, kH, kW] -> [O, kH, kW, I]
+	x = conv2DStrided(x, weight, 2)
+
+	if d.Resample.Bias != nil {
+		bias := mlx.Reshape(d.Resample.Bias, 1, 1, 1, outC)
+		x = mlx.Add(x, bias)
+	}
+
+	// Output dims after stride 2: (H+1)/2, (W+1)/2
+	outH := (H + 1) / 2
+	outW := (W + 1) / 2
+
+	// Reshape back to [B, T, H', W', C]
+	x = mlx.Reshape(x, B, T, outH, outW, outC)
+	mlx.Eval(x)
+
+	return x
+}
+
+// loadFromWeights loads the encoder from pre-loaded weights
+func (e *VAEEncoder) loadFromWeights(weights *safetensors.ModelWeights, cfg *VAEConfig) error {
+	e.Config = cfg
+
+	// Conv in
+	convIn, err := newCausalConv3d(weights, "encoder.conv_in")
+	if err != nil {
+		return err
+	}
+	e.ConvIn = convIn
+
+	// Encoder uses flat block structure:
+	// dim_mult = [1, 2, 4, 4], num_res_blocks = 2, temporal_downsample = [false, true, true]
+	// Block layout: res,res,down, res,res,down+t, res,res,down+t, res,res
+	// That's 11 blocks: 0,1=res, 2=down, 3,4=res, 5=down+t, 6,7=res, 8=down+t, 9,10=res
+	e.Blocks = make([]EncoderBlock, 0, 11)
+
+	// Track dimensions
+	dims := []int32{cfg.BaseDim, cfg.BaseDim * 2, cfg.BaseDim * 4, cfg.BaseDim * 4}
+	blockIdx := 0
+
+	for stage := 0; stage < len(cfg.DimMult); stage++ {
+		inDim := cfg.BaseDim
+		if stage > 0 {
+			inDim = dims[stage-1]
+		}
+		outDim := dims[stage]
+
+		// ResBlocks for this stage (num_res_blocks per stage)
+		for r := int32(0); r < cfg.NumResBlocks; r++ {
+			prefix := fmt.Sprintf("encoder.down_blocks.%d", blockIdx)
+			currentInDim := inDim
+			if r > 0 {
+				currentInDim = outDim
+			}
+			block, err := newEncoderResBlock(weights, prefix, currentInDim, outDim)
+			if err != nil {
+				return fmt.Errorf("encoder res block %d: %w", blockIdx, err)
+			}
+			e.Blocks = append(e.Blocks, block)
+			blockIdx++
+		}
+
+		// Downsample after each stage except the last
+		if stage < len(cfg.DimMult)-1 {
+			prefix := fmt.Sprintf("encoder.down_blocks.%d", blockIdx)
+			down, err := newEncoderDownsample(weights, prefix, cfg.TemperalDownsample[stage])
+			if err != nil {
+				return fmt.Errorf("encoder downsample %d: %w", blockIdx, err)
+			}
+			e.Blocks = append(e.Blocks, down)
+			blockIdx++
+		}
+	}
+
+	// Mid block
+	midDim := cfg.BaseDim * cfg.DimMult[len(cfg.DimMult)-1]
+	midBlock, err := newMidBlock(weights, "encoder.mid_block", midDim)
+	if err != nil {
+		return err
+	}
+	e.MidBlock = midBlock
+
+	// Norm out
+	normOut, err := newRMSNorm3D(weights, "encoder.norm_out", midDim)
+	if err != nil {
+		return err
+	}
+	e.NormOut = normOut
+
+	// Conv out
+	convOut, err := newCausalConv3d(weights, "encoder.conv_out")
+	if err != nil {
+		return err
+	}
+	e.ConvOut = convOut
+
+	// Quant conv
+	quantConv, err := newCausalConv3d(weights, "quant_conv")
+	if err != nil {
+		return err
+	}
+	e.QuantConv = quantConv
+
+	return nil
+}
+
+// newEncoderResBlock creates a ResBlock for the encoder (flat structure)
+func newEncoderResBlock(weights *safetensors.ModelWeights, prefix string, inDim, outDim int32) (*EncoderResBlock, error) {
+	block, err := newResBlock(weights, prefix, inDim, outDim)
+	if err != nil {
+		return nil, err
+	}
+	return &EncoderResBlock{block}, nil
+}
+
+// newEncoderDownsample creates a downsample layer for the encoder
+func newEncoderDownsample(weights *safetensors.ModelWeights, prefix string, temporal bool) (*EncoderDownsample, error) {
+	resample, err := newCausalConv3d(weights, prefix+".resample.1")
+	if err != nil {
+		return nil, err
+	}
+
+	var timeConv *CausalConv3d
+	if temporal {
+		timeConv, _ = newCausalConv3d(weights, prefix+".time_conv")
+	}
+
+	return &EncoderDownsample{
+		Resample: resample,
+		TimeConv: timeConv,
+	}, nil
+}
+
+// Encode encodes an image to latents
+// x: [B, C, T, H, W] image tensor (channels-first)
+// Returns: [B, latent_C, T, H/8, W/8] latent distribution mode
+func (e *VAEEncoder) Encode(x *mlx.Array) *mlx.Array {
+	// Convert from channels-first [N, C, T, H, W] to channels-last [N, T, H, W, C]
+	x = mlx.Contiguous(mlx.Transpose(x, 0, 2, 3, 4, 1))
+	mlx.Eval(x)
+
+	// Conv in
+	x = e.ConvIn.Forward(x)
+
+	// Encoder blocks (mix of ResBlocks and Downsamplers)
+	for _, block := range e.Blocks {
+		prev := x
+		x = block.Forward(x)
+		prev.Free()
+	}
+
+	// Mid block
+	x = e.MidBlock.Forward(x)
+
+	// Norm + silu
+	{
+		prev := x
+		x = e.NormOut.Forward(x)
+		x = silu3D(x)
+		prev.Free()
+		mlx.Eval(x)
+	}
+
+	// Conv out
+	{
+		prev := x
+		x = e.ConvOut.Forward(x)
+		prev.Free()
+	}
+
+	// Quant conv
+	{
+		prev := x
+		x = e.QuantConv.Forward(x)
+		prev.Free()
+	}
+
+	// Get mode from distribution (first half of channels = mean)
+	// Output is [B, T, H, W, 2*latent_C], we take first latent_C channels
+	shape := x.Shape()
+	latentC := shape[4] / 2
+	x = mlx.Slice(x, []int32{0, 0, 0, 0, 0}, []int32{shape[0], shape[1], shape[2], shape[3], latentC})
+
+	// Convert back to channels-first [N, C, T, H, W]
+	x = mlx.Contiguous(mlx.Transpose(x, 0, 4, 1, 2, 3))
+	mlx.Eval(x)
+
+	return x
+}
+
+// VAEDecoder is the decoder part of the VAE
+type VAEDecoder struct {
+	Config *VAEConfig
+
+	PostQuantConv *CausalConv3d
+	ConvIn        *CausalConv3d
+	MidBlock      *MidBlock
+	UpBlocks      []*UpBlock
+	NormOut       *RMSNorm3D
+	ConvOut       *CausalConv3d
+}
+
+// loadFromWeights loads the decoder from pre-loaded weights
+func (d *VAEDecoder) loadFromWeights(weights *safetensors.ModelWeights, cfg *VAEConfig) error {
+	d.Config = cfg
+
+	postQuantConv, err := newCausalConv3d(weights, "post_quant_conv")
+	if err != nil {
+		return err
+	}
+	d.PostQuantConv = postQuantConv
+
+	convIn, err := newCausalConv3d(weights, "decoder.conv_in")
+	if err != nil {
+		return err
+	}
+	d.ConvIn = convIn
+
+	// Mid block
+	midDim := cfg.BaseDim * cfg.DimMult[len(cfg.DimMult)-1]
+	midBlock, err := newMidBlock(weights, "decoder.mid_block", midDim)
+	if err != nil {
+		return err
+	}
+	d.MidBlock = midBlock
+
+	// Up blocks (reversed dim_mult)
+	numUpBlocks := len(cfg.DimMult)
+	d.UpBlocks = make([]*UpBlock, numUpBlocks)
+
+	dimsMult := make([]int32, numUpBlocks+1)
+	dimsMult[0] = cfg.DimMult[numUpBlocks-1]
+	for i := 0; i < numUpBlocks; i++ {
+		dimsMult[i+1] = cfg.DimMult[numUpBlocks-1-i]
+	}
+
+	temporalUpsample := make([]bool, len(cfg.TemperalDownsample))
+	for i := range cfg.TemperalDownsample {
+		temporalUpsample[i] = cfg.TemperalDownsample[len(cfg.TemperalDownsample)-1-i]
+	}
+
+	for i := 0; i < numUpBlocks; i++ {
+		inDim := cfg.BaseDim * dimsMult[i]
+		outDim := cfg.BaseDim * dimsMult[i+1]
+
+		if i > 0 {
+			inDim = inDim / 2
+		}
+
+		upsampleMode := ""
+		if i < numUpBlocks-1 {
+			if temporalUpsample[i] {
+				upsampleMode = "upsample3d"
+			} else {
+				upsampleMode = "upsample2d"
+			}
+		}
+
+		prefix := fmt.Sprintf("decoder.up_blocks.%d", i)
+		upBlock, err := newUpBlock(weights, prefix, inDim, outDim, cfg.NumResBlocks, upsampleMode)
+		if err != nil {
+			return err
+		}
+		d.UpBlocks[i] = upBlock
+	}
+
+	normOut, err := newRMSNorm3D(weights, "decoder.norm_out", cfg.BaseDim)
+	if err != nil {
+		return err
+	}
+	d.NormOut = normOut
+
+	convOut, err := newCausalConv3d(weights, "decoder.conv_out")
+	if err != nil {
+		return err
+	}
+	d.ConvOut = convOut
+
+	return nil
+}
+
+// Decode converts latents to image
+// z: [B, C, T, H, W] denormalized latents
+func (d *VAEDecoder) Decode(z *mlx.Array) *mlx.Array {
+	var x *mlx.Array
+
+	// Convert from channels-first to channels-last
+	{
+		z = mlx.Contiguous(mlx.Transpose(z, 0, 2, 3, 4, 1))
+		mlx.Eval(z)
+	}
+
+	// PostQuantConv
+	x = d.PostQuantConv.Forward(z)
+	z.Free()
+
+	// ConvIn
+	{
+		prev := x
+		x = d.ConvIn.Forward(x)
+		prev.Free()
+	}
+
+	// Mid block
+	x = d.MidBlock.Forward(x)
+
+	// Up blocks
+	for _, upBlock := range d.UpBlocks {
+		x = upBlock.Forward(x)
+	}
+
+	// NormOut + silu
+	{
+		prev := x
+		x = d.NormOut.Forward(x)
+		x = silu3D(x)
+		prev.Free()
+		mlx.Eval(x)
+	}
+
+	// ConvOut
+	{
+		prev := x
+		x = d.ConvOut.Forward(x)
+		prev.Free()
+	}
+
+	// Post-processing: clamp and convert back to channels-first
+	{
+		prev := x
+		x = mlx.ClipScalar(x, -1.0, 1.0, true, true)
+		x = mlx.Contiguous(mlx.Transpose(x, 0, 4, 1, 2, 3))
+		prev.Free()
+		mlx.Eval(x)
+	}
+
+	return x
+}
+
+// DownBlock handles downsampling in encoder
+type DownBlock struct {
+	ResBlocks   []*ResBlock
+	Downsampler *Downsample
+}
+
+// newDownBlock creates a down block
+func newDownBlock(weights *safetensors.ModelWeights, prefix string, inDim, outDim int32, numBlocks int32, downsampleMode string) (*DownBlock, error) {
+	resBlocks := make([]*ResBlock, numBlocks+1)
+
+	currentDim := inDim
+	for i := int32(0); i <= numBlocks; i++ {
+		resPrefix := fmt.Sprintf("%s.resnets.%d", prefix, i)
+		block, err := newResBlock(weights, resPrefix, currentDim, outDim)
+		if err != nil {
+			return nil, err
+		}
+		resBlocks[i] = block
+		currentDim = outDim
+	}
+
+	var downsampler *Downsample
+	if downsampleMode != "" {
+		downsampler = newDownsample(weights, prefix+".downsamplers.0", outDim, downsampleMode)
+	}
+
+	return &DownBlock{
+		ResBlocks:   resBlocks,
+		Downsampler: downsampler,
+	}, nil
+}
+
+// Forward applies down block
+func (d *DownBlock) Forward(x *mlx.Array) *mlx.Array {
+	for _, block := range d.ResBlocks {
+		prev := x
+		x = block.Forward(x)
+		prev.Free()
+	}
+
+	if d.Downsampler != nil {
+		prev := x
+		x = d.Downsampler.Forward(x)
+		prev.Free()
+	}
+	return x
+}
+
+// Downsample handles spatial downsampling
+type Downsample struct {
+	Conv *mlx.Array
+	Bias *mlx.Array
+	Mode string
+}
+
+// newDownsample creates a downsampler
+func newDownsample(weights *safetensors.ModelWeights, prefix string, dim int32, mode string) *Downsample {
+	conv, _ := weights.Get(prefix + ".resample.1.weight")
+	bias, _ := weights.Get(prefix + ".resample.1.bias")
+	return &Downsample{
+		Conv: conv,
+		Bias: bias,
+		Mode: mode,
+	}
+}
+
+// Forward applies downsampling to channels-last input [B, T, H, W, C]
+func (d *Downsample) Forward(x *mlx.Array) *mlx.Array {
+	shape := x.Shape()
+	B := shape[0]
+	T := shape[1]
+	H := shape[2]
+	W := shape[3]
+	C := shape[4]
+	outC := d.Conv.Shape()[0]
+
+	// Reshape to [B*T, H, W, C] for 2D conv
+	x = mlx.Reshape(x, B*T, H, W, C)
+
+	// Pad for stride-2 conv: need (3-1)/2 = 1 on each side, but for stride 2 we need specific padding
+	// For 3x3 stride 2: pad 1 on all sides
+	x = mlx.Pad(x, []int32{0, 0, 1, 1, 1, 1, 0, 0})
+
+	// Conv with stride 2 using manual strided patching
+	weight := mlx.Transpose(d.Conv, 0, 2, 3, 1)
+	x = conv2DStrided(x, weight, 2)
+	if d.Bias != nil {
+		bias := mlx.Reshape(d.Bias, 1, 1, 1, outC)
+		x = mlx.Add(x, bias)
+	}
+
+	x = mlx.Reshape(x, B, T, H/2, W/2, outC)
+	mlx.Eval(x)
+
+	return x
+}

x/imagegen/models/zimage/scheduler.go

+//go:build mlx
+
+package zimage
+
+import (
+	"math"
+
+	"github.com/ollama/ollama/x/imagegen/mlx"
+)
+
+// FlowMatchSchedulerConfig holds scheduler configuration
+type FlowMatchSchedulerConfig struct {
+	NumTrainTimesteps  int32   `json:"num_train_timesteps"`  // 1000
+	Shift              float32 `json:"shift"`                // 3.0
+	UseDynamicShifting bool    `json:"use_dynamic_shifting"` // false
+}
+
+// DefaultFlowMatchSchedulerConfig returns default config
+func DefaultFlowMatchSchedulerConfig() *FlowMatchSchedulerConfig {
+	return &FlowMatchSchedulerConfig{
+		NumTrainTimesteps:  1000,
+		Shift:              3.0,
+		UseDynamicShifting: true, // Z-Image-Turbo uses dynamic shifting
+	}
+}
+
+// FlowMatchEulerScheduler implements the Flow Match Euler discrete scheduler
+// This is used in Z-Image-Turbo for fast sampling
+type FlowMatchEulerScheduler struct {
+	Config    *FlowMatchSchedulerConfig
+	Timesteps []float32 // Discretized timesteps
+	Sigmas    []float32 // Noise levels at each timestep
+	NumSteps  int       // Number of inference steps
+}
+
+// NewFlowMatchEulerScheduler creates a new scheduler
+func NewFlowMatchEulerScheduler(cfg *FlowMatchSchedulerConfig) *FlowMatchEulerScheduler {
+	return &FlowMatchEulerScheduler{
+		Config: cfg,
+	}
+}
+
+// SetTimesteps sets up the scheduler for the given number of inference steps
+func (s *FlowMatchEulerScheduler) SetTimesteps(numSteps int) {
+	s.SetTimestepsWithMu(numSteps, 0)
+}
+
+// SetTimestepsWithMu sets up the scheduler with dynamic mu shift
+func (s *FlowMatchEulerScheduler) SetTimestepsWithMu(numSteps int, mu float32) {
+	s.NumSteps = numSteps
+
+	// Create evenly spaced timesteps from 1.0 to 0.0 (flow matching goes t=1 to t=0)
+	// Match Python: np.linspace(1.0, 0.0, num_inference_steps + 1)
+	s.Timesteps = make([]float32, numSteps+1)
+	s.Sigmas = make([]float32, numSteps+1)
+
+	for i := 0; i <= numSteps; i++ {
+		t := 1.0 - float32(i)/float32(numSteps)
+
+		// Apply time shift if using dynamic shifting
+		if s.Config.UseDynamicShifting && mu != 0 {
+			t = s.timeShift(mu, t)
+		}
+
+		s.Timesteps[i] = t
+		s.Sigmas[i] = t
+	}
+}
+
+// timeShift applies the dynamic time shift (match Python)
+func (s *FlowMatchEulerScheduler) timeShift(mu float32, t float32) float32 {
+	if t <= 0 {
+		return 0
+	}
+	// exp(mu) / (exp(mu) + (1/t - 1))
+	expMu := float32(math.Exp(float64(mu)))
+	return expMu / (expMu + (1.0/t - 1.0))
+}
+
+// Step performs one denoising step
+// modelOutput: predicted velocity/noise from the model
+// timestepIdx: current timestep index
+// sample: current noisy sample
+// Returns: denoised sample for next step
+func (s *FlowMatchEulerScheduler) Step(modelOutput, sample *mlx.Array, timestepIdx int) *mlx.Array {
+	// Get current and next sigma
+	sigma := s.Sigmas[timestepIdx]
+	sigmaNext := s.Sigmas[timestepIdx+1]
+
+	// Euler step: x_{t-dt} = x_t + (sigma_next - sigma) * v_t
+	// where v_t is the velocity predicted by the model
+	dt := sigmaNext - sigma // This is negative (going from noise to clean)
+
+	// x_next = x + dt * velocity
+	scaledOutput := mlx.MulScalar(modelOutput, dt)
+	return mlx.Add(sample, scaledOutput)
+}
+
+// ScaleSample scales the sample for model input (identity for flow matching)
+func (s *FlowMatchEulerScheduler) ScaleSample(sample *mlx.Array, timestepIdx int) *mlx.Array {
+	// Flow matching doesn't need scaling
+	return sample
+}
+
+// GetTimestep returns the timestep value at the given index
+func (s *FlowMatchEulerScheduler) GetTimestep(idx int) float32 {
+	if idx < len(s.Timesteps) {
+		return s.Timesteps[idx]
+	}
+	return 0.0
+}
+
+// GetTimesteps returns all timesteps (implements Scheduler interface)
+func (s *FlowMatchEulerScheduler) GetTimesteps() []float32 {
+	return s.Timesteps
+}
+
+// AddNoise adds noise to clean samples for a given timestep
+// Used for img2img or inpainting
+func (s *FlowMatchEulerScheduler) AddNoise(cleanSample, noise *mlx.Array, timestepIdx int) *mlx.Array {
+	// In flow matching: x_t = (1-t) * x_0 + t * noise
+	t := s.Timesteps[timestepIdx]
+	oneMinusT := 1.0 - t
+
+	scaledClean := mlx.MulScalar(cleanSample, oneMinusT)
+	scaledNoise := mlx.MulScalar(noise, t)
+
+	return mlx.Add(scaledClean, scaledNoise)
+}
+
+// InitNoise creates initial noise for sampling
+func (s *FlowMatchEulerScheduler) InitNoise(shape []int32, seed int64) *mlx.Array {
+	return RandomNormal(shape, seed)
+}
+
+// RandomNormal creates a random normal tensor using MLX
+func RandomNormal(shape []int32, seed int64) *mlx.Array {
+	return mlx.RandomNormal(shape, uint64(seed))
+}
+
+// GetLatentShape returns the latent shape for a given image size
+func GetLatentShape(batchSize, height, width, latentChannels int32, patchSize int32) []int32 {
+	// Latent is 8x smaller than image (VAE downscale)
+	latentH := height / 8
+	latentW := width / 8
+
+	return []int32{batchSize, latentChannels, latentH, latentW}
+}

x/imagegen/models/zimage/text_encoder.go

+//go:build mlx
+
+package zimage
+
+import (
+	"encoding/json"
+	"fmt"
+	"math"
+	"os"
+	"path/filepath"
+
+	"github.com/ollama/ollama/x/imagegen/mlx"
+	"github.com/ollama/ollama/x/imagegen/nn"
+	"github.com/ollama/ollama/x/imagegen/safetensors"
+	"github.com/ollama/ollama/x/imagegen/tokenizer"
+)
+
+// Qwen3Config holds Qwen3 text encoder configuration
+type Qwen3Config struct {
+	HiddenSize        int32   `json:"hidden_size"`
+	NumHiddenLayers   int32   `json:"num_hidden_layers"`
+	IntermediateSize  int32   `json:"intermediate_size"`
+	NumAttentionHeads int32   `json:"num_attention_heads"`
+	NumKeyValueHeads  int32   `json:"num_key_value_heads"`
+	VocabSize         int32   `json:"vocab_size"`
+	RMSNormEps        float32 `json:"rms_norm_eps"`
+	RopeTheta         float32 `json:"rope_theta"`
+	HeadDim           int32   `json:"head_dim"`
+}
+
+// loadQwen3Config loads text encoder config from a JSON file
+func loadQwen3Config(path string) (*Qwen3Config, error) {
+	data, err := os.ReadFile(path)
+	if err != nil {
+		return nil, fmt.Errorf("read config: %w", err)
+	}
+	var cfg Qwen3Config
+	if err := json.Unmarshal(data, &cfg); err != nil {
+		return nil, fmt.Errorf("parse config: %w", err)
+	}
+	return &cfg, nil
+}
+
+// Qwen3Attention implements Qwen3 attention with QK norms
+type Qwen3Attention struct {
+	QProj *nn.Linear  `weight:"q_proj"`
+	KProj *nn.Linear  `weight:"k_proj"`
+	VProj *nn.Linear  `weight:"v_proj"`
+	OProj *nn.Linear  `weight:"o_proj"`
+	QNorm *nn.RMSNorm `weight:"q_norm"`
+	KNorm *nn.RMSNorm `weight:"k_norm"`
+	// Computed fields
+	NHeads    int32
+	NKVHeads  int32
+	HeadDim   int32
+	Scale     float32
+	RopeTheta float32
+}
+
+// applyRoPEQwen3 applies the custom RoPE for Qwen3 text encoder
+func applyRoPEQwen3(x *mlx.Array, seqLen int32, theta float32) *mlx.Array {
+	shape := x.Shape()
+	B := shape[0]
+	L := shape[1]
+	H := shape[2]
+	D := shape[3]
+	half := D / 2
+
+	freqsArr := make([]float32, half)
+	logTheta := float32(math.Log(float64(theta)))
+	for i := int32(0); i < half; i++ {
+		freqsArr[i] = float32(math.Exp(float64(-logTheta * float32(i) / float32(half))))
+	}
+	freqs := mlx.NewArray(freqsArr, []int32{half})
+
+	posArr := make([]float32, seqLen)
+	for i := int32(0); i < seqLen; i++ {
+		posArr[i] = float32(i)
+	}
+	pos := mlx.NewArray(posArr, []int32{seqLen})
+
+	posExpanded := mlx.Reshape(pos, seqLen, 1)
+	freqsExpanded := mlx.Reshape(freqs, 1, half)
+	args := mlx.Mul(posExpanded, freqsExpanded)
+
+	cosVals := mlx.Cos(args)
+	sinVals := mlx.Sin(args)
+	cosVals = mlx.Reshape(cosVals, seqLen, 1, half)
+	sinVals = mlx.Reshape(sinVals, seqLen, 1, half)
+
+	x1 := mlx.Slice(x, []int32{0, 0, 0, 0}, []int32{B, L, H, half})
+	x2 := mlx.Slice(x, []int32{0, 0, 0, half}, []int32{B, L, H, D})
+
+	part1 := mlx.Sub(mlx.Mul(x1, cosVals), mlx.Mul(x2, sinVals))
+	part2 := mlx.Add(mlx.Mul(x1, sinVals), mlx.Mul(x2, cosVals))
+
+	return mlx.Concatenate([]*mlx.Array{part1, part2}, 3)
+}
+
+// Forward computes attention with causal masking
+func (attn *Qwen3Attention) Forward(x *mlx.Array) *mlx.Array {
+	shape := x.Shape()
+	B := shape[0]
+	L := shape[1]
+
+	q := attn.QProj.Forward(x)
+	k := attn.KProj.Forward(x)
+	v := attn.VProj.Forward(x)
+
+	q = mlx.Reshape(q, B, L, attn.NHeads, attn.HeadDim)
+	k = mlx.Reshape(k, B, L, attn.NKVHeads, attn.HeadDim)
+	v = mlx.Reshape(v, B, L, attn.NKVHeads, attn.HeadDim)
+
+	// QK norm uses 1e-6 hardcoded (Qwen3 specific)
+	q = attn.QNorm.Forward(q, 1e-6)
+	k = attn.KNorm.Forward(k, 1e-6)
+
+	q = applyRoPEQwen3(q, L, attn.RopeTheta)
+	k = applyRoPEQwen3(k, L, attn.RopeTheta)
+
+	q = mlx.Transpose(q, 0, 2, 1, 3)
+	k = mlx.Transpose(k, 0, 2, 1, 3)
+	v = mlx.Transpose(v, 0, 2, 1, 3)
+
+	if attn.NKVHeads < attn.NHeads {
+		repeats := attn.NHeads / attn.NKVHeads
+		k = repeatKV(k, repeats)
+		v = repeatKV(v, repeats)
+	}
+
+	out := mlx.ScaledDotProductAttention(q, k, v, attn.Scale, true)
+
+	out = mlx.Transpose(out, 0, 2, 1, 3)
+	out = mlx.Reshape(out, B, L, attn.NHeads*attn.HeadDim)
+
+	out = attn.OProj.Forward(out)
+
+	return out
+}
+
+// repeatKV repeats key/value heads for GQA
+func repeatKV(x *mlx.Array, repeats int32) *mlx.Array {
+	if repeats == 1 {
+		return x
+	}
+	shape := x.Shape()
+	x = mlx.ExpandDims(x, 2)
+	x = mlx.Tile(x, []int32{1, 1, repeats, 1, 1})
+	return mlx.Reshape(x, shape[0], shape[1]*repeats, shape[2], shape[3])
+}
+
+// Qwen3MLP implements Qwen3 SwiGLU MLP
+type Qwen3MLP struct {
+	GateProj *nn.Linear `weight:"gate_proj"`
+	UpProj   *nn.Linear `weight:"up_proj"`
+	DownProj *nn.Linear `weight:"down_proj"`
+}
+
+// Forward applies the MLP
+func (m *Qwen3MLP) Forward(x *mlx.Array) *mlx.Array {
+	gate := m.GateProj.Forward(x)
+	gate = mlx.SiLU(gate)
+	up := m.UpProj.Forward(x)
+	h := mlx.Mul(gate, up)
+	return m.DownProj.Forward(h)
+}
+
+// Qwen3Block represents a single Qwen3 transformer block
+type Qwen3Block struct {
+	Attention         *Qwen3Attention `weight:"self_attn"`
+	MLP               *Qwen3MLP       `weight:"mlp"`
+	InputLayerNorm    *nn.RMSNorm     `weight:"input_layernorm"`
+	PostAttnLayerNorm *nn.RMSNorm     `weight:"post_attention_layernorm"`
+}
+
+// Forward applies the Qwen3 block
+func (qb *Qwen3Block) Forward(x *mlx.Array, eps float32) *mlx.Array {
+	h := qb.InputLayerNorm.Forward(x, eps)
+	attnOut := qb.Attention.Forward(h)
+	x = mlx.Add(x, attnOut)
+
+	h = qb.PostAttnLayerNorm.Forward(x, eps)
+	mlpOut := qb.MLP.Forward(h)
+	x = mlx.Add(x, mlpOut)
+
+	return x
+}
+
+// Qwen3TextEncoder is the full Qwen3 encoder for Z-Image
+type Qwen3TextEncoder struct {
+	EmbedTokens *nn.Embedding   `weight:"model.embed_tokens"`
+	Layers      []*Qwen3Block   `weight:"model.layers"`
+	FinalNorm   *nn.RMSNorm     `weight:"model.norm"`
+	*Qwen3Config
+}
+
+// Load loads the Qwen3 text encoder from a directory
+func (m *Qwen3TextEncoder) Load(path string) error {
+	fmt.Println("Loading Qwen3 text encoder...")
+
+	// Load config
+	cfg, err := loadQwen3Config(filepath.Join(path, "config.json"))
+	if err != nil {
+		return fmt.Errorf("config: %w", err)
+	}
+	m.Qwen3Config = cfg
+
+	// Pre-allocate layers slice
+	m.Layers = make([]*Qwen3Block, cfg.NumHiddenLayers)
+
+	// Load weights
+	weights, err := safetensors.LoadModelWeights(path)
+	if err != nil {
+		return fmt.Errorf("weights: %w", err)
+	}
+
+	fmt.Print("  Loading weights via struct tags... ")
+	if err := safetensors.LoadModule(m, weights, ""); err != nil {
+		return fmt.Errorf("load module: %w", err)
+	}
+	fmt.Println("✓")
+
+	// Initialize computed fields
+	m.FinalNorm.Eps = cfg.RMSNormEps
+	for _, block := range m.Layers {
+		// Attention
+		block.Attention.NHeads = cfg.NumAttentionHeads
+		block.Attention.NKVHeads = cfg.NumKeyValueHeads
+		block.Attention.HeadDim = cfg.HeadDim
+		block.Attention.Scale = float32(1.0 / math.Sqrt(float64(cfg.HeadDim)))
+		block.Attention.RopeTheta = cfg.RopeTheta
+		block.Attention.QNorm.Eps = cfg.RMSNormEps
+		block.Attention.KNorm.Eps = cfg.RMSNormEps
+		// Block norms
+		block.InputLayerNorm.Eps = cfg.RMSNormEps
+		block.PostAttnLayerNorm.Eps = cfg.RMSNormEps
+	}
+
+	weights.ReleaseAll()
+	return nil
+}
+
+// Forward encodes text tokens
+func (te *Qwen3TextEncoder) Forward(tokens *mlx.Array) *mlx.Array {
+	h := te.EmbedTokens.Forward(tokens)
+	eps := te.RMSNormEps
+
+	for _, layer := range te.Layers {
+		h = layer.Forward(h, eps)
+	}
+
+	// Apply final RMS norm
+	h = te.FinalNorm.Forward(h, eps)
+
+	return h
+}
+
+// ApplyChatTemplate wraps prompt in Qwen3 chat format
+func ApplyChatTemplate(prompt string) string {
+	return "<|im_start|>user\n" + prompt + "<|im_end|>\n<|im_start|>assistant\n"
+}
+
+// EncodePrompt encodes a text prompt using the tokenizer and encoder
+func (te *Qwen3TextEncoder) EncodePrompt(tok *tokenizer.Tokenizer, prompt string, maxLen int) (*mlx.Array, *mlx.Array) {
+	formattedPrompt := ApplyChatTemplate(prompt)
+
+	tokens := tok.Encode(formattedPrompt, false)
+
+	if len(tokens) > maxLen {
+		tokens = tokens[:maxLen]
+	}
+
+	maskData := make([]float32, maxLen)
+	for i := 0; i < len(tokens); i++ {
+		maskData[i] = 1.0
+	}
+
+	// Get PAD token (different from EOS for Qwen3)
+	padToken := tok.PAD()
+	if padToken < 0 {
+		padToken = tok.EOS() // fallback
+	}
+
+	paddedTokens := make([]int32, maxLen)
+	copy(paddedTokens, tokens)
+	for i := len(tokens); i < maxLen; i++ {
+		paddedTokens[i] = padToken
+	}
+
+	tokensArr := mlx.NewArrayInt32(paddedTokens, []int32{1, int32(maxLen)})
+	maskArr := mlx.NewArray(maskData, []int32{1, int32(maxLen)})
+
+	embeddings := te.Forward(tokensArr)
+
+	return embeddings, maskArr
+}

x/imagegen/models/zimage/transformer.go

+//go:build mlx
+
+// Package zimage implements the Z-Image diffusion transformer model.
+package zimage
+
+import (
+	"encoding/json"
+	"fmt"
+	"math"
+	"os"
+	"path/filepath"
+
+	"github.com/ollama/ollama/x/imagegen/cache"
+	"github.com/ollama/ollama/x/imagegen/mlx"
+	"github.com/ollama/ollama/x/imagegen/nn"
+	"github.com/ollama/ollama/x/imagegen/safetensors"
+)
+
+// TransformerConfig holds Z-Image transformer configuration
+type TransformerConfig struct {
+	Dim            int32   `json:"dim"`
+	NHeads         int32   `json:"n_heads"`
+	NKVHeads       int32   `json:"n_kv_heads"`
+	NLayers        int32   `json:"n_layers"`
+	NRefinerLayers int32   `json:"n_refiner_layers"`
+	InChannels     int32   `json:"in_channels"`
+	PatchSize      int32   `json:"-"` // Computed from AllPatchSize
+	CapFeatDim     int32   `json:"cap_feat_dim"`
+	NormEps        float32 `json:"norm_eps"`
+	RopeTheta      float32 `json:"rope_theta"`
+	TScale         float32 `json:"t_scale"`
+	QKNorm         bool    `json:"qk_norm"`
+	AxesDims       []int32 `json:"axes_dims"`
+	AxesLens       []int32 `json:"axes_lens"`
+	AllPatchSize   []int32 `json:"all_patch_size"` // JSON array, PatchSize = first element
+}
+
+// TimestepEmbedder creates sinusoidal timestep embeddings
+// Output dimension is 256 (fixed), used for AdaLN modulation
+type TimestepEmbedder struct {
+	Linear1       *nn.Linear `weight:"mlp.0"`
+	Linear2       *nn.Linear `weight:"mlp.2"`
+	FreqEmbedSize int32      // 256 (computed)
+}
+
+// Forward computes timestep embeddings -> [B, 256]
+func (te *TimestepEmbedder) Forward(t *mlx.Array) *mlx.Array {
+	// t: [B] timesteps
+
+	// Create sinusoidal embedding
+	half := te.FreqEmbedSize / 2
+
+	// freqs = exp(-log(10000) * arange(half) / half)
+	freqs := make([]float32, half)
+	for i := int32(0); i < half; i++ {
+		freqs[i] = float32(math.Exp(-math.Log(10000.0) * float64(i) / float64(half)))
+	}
+	freqsArr := mlx.NewArray(freqs, []int32{1, half})
+
+	// t[:, None] * freqs[None, :] -> [B, half]
+	tExpanded := mlx.ExpandDims(t, 1) // [B, 1]
+	args := mlx.Mul(tExpanded, freqsArr)
+
+	// embedding = [cos(args), sin(args)] -> [B, 256]
+	cosArgs := mlx.Cos(args)
+	sinArgs := mlx.Sin(args)
+	embedding := mlx.Concatenate([]*mlx.Array{cosArgs, sinArgs}, 1)
+
+	// MLP: linear1 -> silu -> linear2
+	h := te.Linear1.Forward(embedding)
+	h = mlx.SiLU(h)
+	h = te.Linear2.Forward(h)
+
+	return h
+}
+
+// XEmbedder embeds image patches to model dimension
+type XEmbedder struct {
+	Linear *nn.Linear `weight:"2-1"`
+}
+
+// Forward embeds patchified image latents
+func (xe *XEmbedder) Forward(x *mlx.Array) *mlx.Array {
+	// x: [B, L, in_channels * 4] -> [B, L, dim]
+	return xe.Linear.Forward(x)
+}
+
+// CapEmbedder projects caption features to model dimension
+type CapEmbedder struct {
+	Norm     *nn.RMSNorm `weight:"0"`
+	Linear   *nn.Linear  `weight:"1"`
+	PadToken *mlx.Array  // loaded separately at root level
+}
+
+// Forward projects caption embeddings: [B, L, cap_feat_dim] -> [B, L, dim]
+func (ce *CapEmbedder) Forward(capFeats *mlx.Array) *mlx.Array {
+	// RMSNorm on last axis (uses 1e-6)
+	h := ce.Norm.Forward(capFeats, 1e-6)
+	// Linear projection
+	return ce.Linear.Forward(h)
+}
+
+// FeedForward implements SwiGLU FFN
+type FeedForward struct {
+	W1     *nn.Linear `weight:"w1"` // gate projection
+	W2     *nn.Linear `weight:"w2"` // down projection
+	W3     *nn.Linear `weight:"w3"` // up projection
+	OutDim int32      // computed from W2
+}
+
+// Forward applies SwiGLU: silu(W1(x)) * W3(x), then W2
+func (ff *FeedForward) Forward(x *mlx.Array) *mlx.Array {
+	shape := x.Shape()
+	B := shape[0]
+	L := shape[1]
+	D := shape[2]
+
+	// Reshape for matmul
+	x = mlx.Reshape(x, B*L, D)
+	gate := ff.W1.Forward(x)
+	gate = mlx.SiLU(gate)
+	up := ff.W3.Forward(x)
+	h := mlx.Mul(gate, up)
+	out := ff.W2.Forward(h)
+
+	return mlx.Reshape(out, B, L, ff.OutDim)
+}
+
+// Attention implements multi-head attention with QK norm
+type Attention struct {
+	ToQ   *nn.Linear `weight:"to_q"`
+	ToK   *nn.Linear `weight:"to_k"`
+	ToV   *nn.Linear `weight:"to_v"`
+	ToOut *nn.Linear `weight:"to_out.0"`
+	NormQ *mlx.Array `weight:"norm_q.weight"` // [head_dim] for per-head RMSNorm
+	NormK *mlx.Array `weight:"norm_k.weight"`
+	// Computed fields
+	NHeads  int32
+	HeadDim int32
+	Dim     int32
+	Scale   float32
+}
+
+// Forward computes attention
+func (attn *Attention) Forward(x *mlx.Array, cos, sin *mlx.Array) *mlx.Array {
+	shape := x.Shape()
+	B := shape[0]
+	L := shape[1]
+	D := shape[2]
+
+	// Project Q, K, V
+	xFlat := mlx.Reshape(x, B*L, D)
+	q := attn.ToQ.Forward(xFlat)
+	k := attn.ToK.Forward(xFlat)
+	v := attn.ToV.Forward(xFlat)
+
+	// Reshape to [B, L, nheads, head_dim]
+	q = mlx.Reshape(q, B, L, attn.NHeads, attn.HeadDim)
+	k = mlx.Reshape(k, B, L, attn.NHeads, attn.HeadDim)
+	v = mlx.Reshape(v, B, L, attn.NHeads, attn.HeadDim)
+
+	// QK norm
+	q = mlx.RMSNorm(q, attn.NormQ, 1e-5)
+	k = mlx.RMSNorm(k, attn.NormK, 1e-5)
+
+	// Apply RoPE if provided
+	if cos != nil && sin != nil {
+		q = applyRoPE3D(q, cos, sin)
+		k = applyRoPE3D(k, cos, sin)
+	}
+
+	// Transpose to [B, nheads, L, head_dim]
+	q = mlx.Transpose(q, 0, 2, 1, 3)
+	k = mlx.Transpose(k, 0, 2, 1, 3)
+	v = mlx.Transpose(v, 0, 2, 1, 3)
+
+	// SDPA
+	out := mlx.ScaledDotProductAttention(q, k, v, attn.Scale, false)
+
+	// Transpose back and reshape
+	out = mlx.Transpose(out, 0, 2, 1, 3)
+	out = mlx.Reshape(out, B*L, attn.Dim)
+	out = attn.ToOut.Forward(out)
+
+	return mlx.Reshape(out, B, L, attn.Dim)
+}
+
+// applyRoPE3D applies 3-axis rotary position embeddings
+// x: [B, L, nheads, head_dim]
+// cos, sin: [B, L, 1, head_dim/2]
+func applyRoPE3D(x *mlx.Array, cos, sin *mlx.Array) *mlx.Array {
+	shape := x.Shape()
+	B := shape[0]
+	L := shape[1]
+	nheads := shape[2]
+	headDim := shape[3]
+	half := headDim / 2
+
+	// Create even/odd index arrays
+	evenIdx := make([]int32, half)
+	oddIdx := make([]int32, half)
+	for i := int32(0); i < half; i++ {
+		evenIdx[i] = i * 2
+		oddIdx[i] = i*2 + 1
+	}
+	evenIndices := mlx.NewArrayInt32(evenIdx, []int32{half})
+	oddIndices := mlx.NewArrayInt32(oddIdx, []int32{half})
+
+	// Extract x1 (even indices) and x2 (odd indices) along last axis
+	x1 := mlx.Take(x, evenIndices, 3) // [B, L, nheads, half]
+	x2 := mlx.Take(x, oddIndices, 3)  // [B, L, nheads, half]
+
+	// Apply rotation: [x1*cos - x2*sin, x1*sin + x2*cos]
+	r1 := mlx.Sub(mlx.Mul(x1, cos), mlx.Mul(x2, sin))
+	r2 := mlx.Add(mlx.Mul(x1, sin), mlx.Mul(x2, cos))
+
+	// Stack and reshape to interleave: [r1_0, r2_0, r1_1, r2_1, ...]
+	r1 = mlx.ExpandDims(r1, 4)                          // [B, L, nheads, half, 1]
+	r2 = mlx.ExpandDims(r2, 4)                          // [B, L, nheads, half, 1]
+	stacked := mlx.Concatenate([]*mlx.Array{r1, r2}, 4) // [B, L, nheads, half, 2]
+	return mlx.Reshape(stacked, B, L, nheads, headDim)
+}
+
+// TransformerBlock is a single transformer block with optional AdaLN modulation
+type TransformerBlock struct {
+	Attention      *Attention   `weight:"attention"`
+	FeedForward    *FeedForward `weight:"feed_forward"`
+	AttentionNorm1 *nn.RMSNorm  `weight:"attention_norm1"`
+	AttentionNorm2 *nn.RMSNorm  `weight:"attention_norm2"`
+	FFNNorm1       *nn.RMSNorm  `weight:"ffn_norm1"`
+	FFNNorm2       *nn.RMSNorm  `weight:"ffn_norm2"`
+	AdaLN          *nn.Linear   `weight:"adaLN_modulation.0,optional"` // only if modulation
+	// Computed fields
+	HasModulation bool
+	Dim           int32
+}
+
+// Forward applies the transformer block
+func (tb *TransformerBlock) Forward(x *mlx.Array, adaln *mlx.Array, cos, sin *mlx.Array, eps float32) *mlx.Array {
+	if tb.AdaLN != nil && adaln != nil {
+		// Compute modulation: [B, 256] -> [B, 4*dim]
+		chunks := tb.AdaLN.Forward(adaln)
+
+		// Split into 4 parts: scale_msa, gate_msa, scale_mlp, gate_mlp
+		chunkShape := chunks.Shape()
+		chunkDim := chunkShape[1] / 4
+
+		scaleMSA := mlx.Slice(chunks, []int32{0, 0}, []int32{chunkShape[0], chunkDim})
+		gateMSA := mlx.Slice(chunks, []int32{0, chunkDim}, []int32{chunkShape[0], chunkDim * 2})
+		scaleMLP := mlx.Slice(chunks, []int32{0, chunkDim * 2}, []int32{chunkShape[0], chunkDim * 3})
+		gateMLP := mlx.Slice(chunks, []int32{0, chunkDim * 3}, []int32{chunkShape[0], chunkDim * 4})
+
+		// Expand for broadcasting: [B, 1, dim]
+		scaleMSA = mlx.ExpandDims(scaleMSA, 1)
+		gateMSA = mlx.ExpandDims(gateMSA, 1)
+		scaleMLP = mlx.ExpandDims(scaleMLP, 1)
+		gateMLP = mlx.ExpandDims(gateMLP, 1)
+
+		// Attention with modulation
+		normX := tb.AttentionNorm1.Forward(x, eps)
+		normX = mlx.Mul(normX, mlx.AddScalar(scaleMSA, 1.0))
+		attnOut := tb.Attention.Forward(normX, cos, sin)
+		attnOut = tb.AttentionNorm2.Forward(attnOut, eps)
+		x = mlx.Add(x, mlx.Mul(mlx.Tanh(gateMSA), attnOut))
+
+		// FFN with modulation
+		normFFN := tb.FFNNorm1.Forward(x, eps)
+		normFFN = mlx.Mul(normFFN, mlx.AddScalar(scaleMLP, 1.0))
+		ffnOut := tb.FeedForward.Forward(normFFN)
+		ffnOut = tb.FFNNorm2.Forward(ffnOut, eps)
+		x = mlx.Add(x, mlx.Mul(mlx.Tanh(gateMLP), ffnOut))
+	} else {
+		// No modulation (context refiner)
+		attnOut := tb.Attention.Forward(tb.AttentionNorm1.Forward(x, eps), cos, sin)
+		x = mlx.Add(x, tb.AttentionNorm2.Forward(attnOut, eps))
+
+		ffnOut := tb.FeedForward.Forward(tb.FFNNorm1.Forward(x, eps))
+		x = mlx.Add(x, tb.FFNNorm2.Forward(ffnOut, eps))
+	}
+
+	return x
+}
+
+// FinalLayer outputs the denoised patches
+type FinalLayer struct {
+	AdaLN  *nn.Linear `weight:"adaLN_modulation.1"` // [256] -> [dim]
+	Output *nn.Linear `weight:"linear"`             // [dim] -> [out_channels]
+	OutDim int32      // computed from Output
+}
+
+// Forward computes final output
+func (fl *FinalLayer) Forward(x *mlx.Array, c *mlx.Array) *mlx.Array {
+	// c: [B, 256] -> scale: [B, dim]
+	scale := mlx.SiLU(c)
+	scale = fl.AdaLN.Forward(scale)
+	scale = mlx.ExpandDims(scale, 1) // [B, 1, dim]
+
+	// LayerNorm (affine=False) then scale
+	x = layerNormNoAffine(x, 1e-6)
+	x = mlx.Mul(x, mlx.AddScalar(scale, 1.0))
+
+	// Output projection
+	shape := x.Shape()
+	B := shape[0]
+	L := shape[1]
+	D := shape[2]
+	x = mlx.Reshape(x, B*L, D)
+	x = fl.Output.Forward(x)
+
+	return mlx.Reshape(x, B, L, fl.OutDim)
+}
+
+// layerNormNoAffine applies layer norm without learnable parameters
+func layerNormNoAffine(x *mlx.Array, eps float32) *mlx.Array {
+	ndim := x.Ndim()
+	lastAxis := ndim - 1
+
+	mean := mlx.Mean(x, lastAxis, true)
+	xCentered := mlx.Sub(x, mean)
+	variance := mlx.Mean(mlx.Square(xCentered), lastAxis, true)
+	return mlx.Div(xCentered, mlx.Sqrt(mlx.AddScalar(variance, eps)))
+}
+
+// Transformer is the full Z-Image DiT model
+type Transformer struct {
+	TEmbed          *TimestepEmbedder   `weight:"t_embedder"`
+	XEmbed          *XEmbedder          `weight:"all_x_embedder"`
+	CapEmbed        *CapEmbedder        `weight:"cap_embedder"`
+	NoiseRefiners   []*TransformerBlock `weight:"noise_refiner"`
+	ContextRefiners []*TransformerBlock `weight:"context_refiner"`
+	Layers          []*TransformerBlock `weight:"layers"`
+	FinalLayer      *FinalLayer         `weight:"all_final_layer.2-1"`
+	XPadToken       *mlx.Array          `weight:"x_pad_token"`
+	CapPadToken     *mlx.Array          `weight:"cap_pad_token"`
+	*TransformerConfig
+}
+
+// Load loads the Z-Image transformer from a directory
+func (m *Transformer) Load(path string) error {
+	fmt.Println("Loading Z-Image transformer...")
+
+	// Load config
+	cfg, err := loadTransformerConfig(filepath.Join(path, "config.json"))
+	if err != nil {
+		return fmt.Errorf("config: %w", err)
+	}
+	m.TransformerConfig = cfg
+
+	// Pre-allocate slices for loader
+	m.NoiseRefiners = make([]*TransformerBlock, cfg.NRefinerLayers)
+	m.ContextRefiners = make([]*TransformerBlock, cfg.NRefinerLayers)
+	m.Layers = make([]*TransformerBlock, cfg.NLayers)
+
+	// Load weights
+	weights, err := safetensors.LoadModelWeights(path)
+	if err != nil {
+		return fmt.Errorf("weights: %w", err)
+	}
+
+	fmt.Print("  Loading weights as bf16... ")
+	if err := weights.Load(mlx.DtypeBFloat16); err != nil {
+		return fmt.Errorf("load weights: %w", err)
+	}
+	fmt.Printf("✓ (%.1f GB)\n", float64(mlx.MetalGetActiveMemory())/(1024*1024*1024))
+
+	fmt.Print("  Loading weights via struct tags... ")
+	if err := safetensors.LoadModule(m, weights, ""); err != nil {
+		return fmt.Errorf("load module: %w", err)
+	}
+	fmt.Println("✓")
+
+	// Initialize computed fields
+	m.TEmbed.FreqEmbedSize = 256
+	m.FinalLayer.OutDim = m.FinalLayer.Output.Weight.Shape()[0]
+	m.CapEmbed.Norm.Eps = 1e-6
+
+	for _, block := range m.NoiseRefiners {
+		initTransformerBlock(block, cfg)
+	}
+	for _, block := range m.ContextRefiners {
+		initTransformerBlock(block, cfg)
+	}
+	for _, block := range m.Layers {
+		initTransformerBlock(block, cfg)
+	}
+
+	weights.ReleaseAll()
+	return nil
+}
+
+// loadTransformerConfig loads transformer config from a JSON file
+func loadTransformerConfig(path string) (*TransformerConfig, error) {
+	data, err := os.ReadFile(path)
+	if err != nil {
+		return nil, fmt.Errorf("read config: %w", err)
+	}
+	var cfg TransformerConfig
+	if err := json.Unmarshal(data, &cfg); err != nil {
+		return nil, fmt.Errorf("parse config: %w", err)
+	}
+	// Extract PatchSize from array
+	if len(cfg.AllPatchSize) > 0 {
+		cfg.PatchSize = cfg.AllPatchSize[0]
+	}
+	return &cfg, nil
+}
+
+// initTransformerBlock sets computed fields on a transformer block
+func initTransformerBlock(block *TransformerBlock, cfg *TransformerConfig) {
+	block.Dim = cfg.Dim
+	block.HasModulation = block.AdaLN != nil
+
+	// Init attention computed fields
+	attn := block.Attention
+	attn.NHeads = cfg.NHeads
+	attn.HeadDim = cfg.Dim / cfg.NHeads
+	attn.Dim = cfg.Dim
+	attn.Scale = float32(1.0 / math.Sqrt(float64(attn.HeadDim)))
+
+	// Init feedforward OutDim
+	block.FeedForward.OutDim = block.FeedForward.W2.Weight.Shape()[0]
+
+	// Set eps on all RMSNorm layers
+	block.AttentionNorm1.Eps = cfg.NormEps
+	block.AttentionNorm2.Eps = cfg.NormEps
+	block.FFNNorm1.Eps = cfg.NormEps
+	block.FFNNorm2.Eps = cfg.NormEps
+}
+
+// RoPECache holds precomputed RoPE values
+type RoPECache struct {
+	ImgCos     *mlx.Array
+	ImgSin     *mlx.Array
+	CapCos     *mlx.Array
+	CapSin     *mlx.Array
+	UnifiedCos *mlx.Array
+	UnifiedSin *mlx.Array
+	ImgLen     int32
+	CapLen     int32
+}
+
+// PrepareRoPECache precomputes RoPE values for the given image and caption lengths.
+// hTok and wTok are the number of tokens in each dimension (latentH/patchSize, latentW/patchSize).
+func (m *Transformer) PrepareRoPECache(hTok, wTok, capLen int32) *RoPECache {
+	imgLen := hTok * wTok
+
+	// Image positions: grid over (1, H, W) starting at (capLen+1, 0, 0)
+	imgPos := createCoordinateGrid(1, hTok, wTok, capLen+1, 0, 0)
+	imgPos = mlx.ToBFloat16(imgPos)
+	// Caption positions: grid over (capLen, 1, 1) starting at (1, 0, 0)
+	capPos := createCoordinateGrid(capLen, 1, 1, 1, 0, 0)
+	capPos = mlx.ToBFloat16(capPos)
+
+	// Compute RoPE from UNIFIED positions
+	unifiedPos := mlx.Concatenate([]*mlx.Array{imgPos, capPos}, 1)
+	unifiedCos, unifiedSin := prepareRoPE3D(unifiedPos, m.TransformerConfig.AxesDims)
+
+	// Slice RoPE for image and caption parts
+	imgCos := mlx.Slice(unifiedCos, []int32{0, 0, 0, 0}, []int32{1, imgLen, 1, 64})
+	imgSin := mlx.Slice(unifiedSin, []int32{0, 0, 0, 0}, []int32{1, imgLen, 1, 64})
+	capCos := mlx.Slice(unifiedCos, []int32{0, imgLen, 0, 0}, []int32{1, imgLen + capLen, 1, 64})
+	capSin := mlx.Slice(unifiedSin, []int32{0, imgLen, 0, 0}, []int32{1, imgLen + capLen, 1, 64})
+
+	return &RoPECache{
+		ImgCos:     imgCos,
+		ImgSin:     imgSin,
+		CapCos:     capCos,
+		CapSin:     capSin,
+		UnifiedCos: unifiedCos,
+		UnifiedSin: unifiedSin,
+		ImgLen:     imgLen,
+		CapLen:     capLen,
+	}
+}
+
+// Forward runs the Z-Image transformer with precomputed RoPE
+func (m *Transformer) Forward(x *mlx.Array, t *mlx.Array, capFeats *mlx.Array, rope *RoPECache) *mlx.Array {
+	imgLen := rope.ImgLen
+
+	// Timestep embedding -> [B, 256]
+	temb := m.TEmbed.Forward(mlx.MulScalar(t, m.TransformerConfig.TScale))
+
+	// Embed image patches -> [B, L_img, dim]
+	x = m.XEmbed.Forward(x)
+
+	// Embed caption features -> [B, L_cap, dim]
+	capEmb := m.CapEmbed.Forward(capFeats)
+
+	eps := m.NormEps
+
+	// Noise refiner: refine image patches with modulation
+	for _, refiner := range m.NoiseRefiners {
+		x = refiner.Forward(x, temb, rope.ImgCos, rope.ImgSin, eps)
+	}
+
+	// Context refiner: refine caption (no modulation)
+	for _, refiner := range m.ContextRefiners {
+		capEmb = refiner.Forward(capEmb, nil, rope.CapCos, rope.CapSin, eps)
+	}
+
+	// Concatenate image and caption for joint attention
+	unified := mlx.Concatenate([]*mlx.Array{x, capEmb}, 1)
+
+	// Main transformer layers use full unified RoPE
+	for _, layer := range m.Layers {
+		unified = layer.Forward(unified, temb, rope.UnifiedCos, rope.UnifiedSin, eps)
+	}
+
+	// Extract image tokens only
+	unifiedShape := unified.Shape()
+	B := unifiedShape[0]
+	imgOut := mlx.Slice(unified, []int32{0, 0, 0}, []int32{B, imgLen, unifiedShape[2]})
+
+	// Final layer
+	return m.FinalLayer.Forward(imgOut, temb)
+}
+
+// ForwardWithCache runs the transformer with layer caching for faster inference.
+// On refresh steps (step % cacheInterval == 0), all layers are computed and cached.
+// On other steps, shallow layers (0 to cacheLayers-1) reuse cached outputs.
+func (m *Transformer) ForwardWithCache(
+	x *mlx.Array,
+	t *mlx.Array,
+	capFeats *mlx.Array,
+	rope *RoPECache,
+	stepCache *cache.StepCache,
+	step int,
+	cacheInterval int,
+) *mlx.Array {
+	imgLen := rope.ImgLen
+	cacheLayers := stepCache.NumLayers()
+	eps := m.NormEps
+
+	// Timestep embedding -> [B, 256]
+	temb := m.TEmbed.Forward(mlx.MulScalar(t, m.TransformerConfig.TScale))
+
+	// Embed image patches -> [B, L_img, dim]
+	x = m.XEmbed.Forward(x)
+
+	// Context refiners: compute once on step 0, reuse forever
+	// (caption embedding doesn't depend on timestep or latents)
+	var capEmb *mlx.Array
+	if stepCache.GetConstant() != nil {
+		capEmb = stepCache.GetConstant()
+	} else {
+		capEmb = m.CapEmbed.Forward(capFeats)
+		for _, refiner := range m.ContextRefiners {
+			capEmb = refiner.Forward(capEmb, nil, rope.CapCos, rope.CapSin, eps)
+		}
+		stepCache.SetConstant(capEmb)
+	}
+
+	// Noise refiners: always compute (depend on x which changes each step)
+	for _, refiner := range m.NoiseRefiners {
+		x = refiner.Forward(x, temb, rope.ImgCos, rope.ImgSin, eps)
+	}
+
+	// Concatenate image and caption for joint attention
+	unified := mlx.Concatenate([]*mlx.Array{x, capEmb}, 1)
+
+	// Determine if this is a cache refresh step
+	refreshCache := stepCache.ShouldRefresh(step, cacheInterval)
+
+	// Main transformer layers with caching
+	for i, layer := range m.Layers {
+		if i < cacheLayers && !refreshCache && stepCache.Get(i) != nil {
+			// Use cached output for shallow layers
+			unified = stepCache.Get(i)
+		} else {
+			// Compute layer
+			unified = layer.Forward(unified, temb, rope.UnifiedCos, rope.UnifiedSin, eps)
+			// Cache shallow layer outputs on refresh steps
+			if i < cacheLayers && refreshCache {
+				stepCache.Set(i, unified)
+			}
+		}
+	}
+
+	// Extract image tokens only
+	unifiedShape := unified.Shape()
+	B := unifiedShape[0]
+	imgOut := mlx.Slice(unified, []int32{0, 0, 0}, []int32{B, imgLen, unifiedShape[2]})
+
+	// Final layer
+	return m.FinalLayer.Forward(imgOut, temb)
+}
+
+// createCoordinateGrid creates 3D position grid [1, d0*d1*d2, 3]
+func createCoordinateGrid(d0, d1, d2, s0, s1, s2 int32) *mlx.Array {
+	// Create meshgrid and stack
+	total := d0 * d1 * d2
+	coords := make([]float32, total*3)
+
+	idx := 0
+	for i := int32(0); i < d0; i++ {
+		for j := int32(0); j < d1; j++ {
+			for k := int32(0); k < d2; k++ {
+				coords[idx*3+0] = float32(s0 + i)
+				coords[idx*3+1] = float32(s1 + j)
+				coords[idx*3+2] = float32(s2 + k)
+				idx++
+			}
+		}
+	}
+
+	return mlx.NewArray(coords, []int32{1, total, 3})
+}
+
+// prepareRoPE3D computes cos/sin for 3-axis RoPE
+// positions: [B, L, 3] with (h, w, t) coordinates
+// axesDims: [32, 48, 48] - dimensions for each axis
+// Returns: cos, sin each [B, L, 1, head_dim/2]
+func prepareRoPE3D(positions *mlx.Array, axesDims []int32) (*mlx.Array, *mlx.Array) {
+	// Compute frequencies for each axis
+	// dims = [32, 48, 48], so halves = [16, 24, 24]
+	ropeTheta := float32(256.0)
+
+	freqs := make([]*mlx.Array, 3)
+	for axis := 0; axis < 3; axis++ {
+		half := axesDims[axis] / 2
+		f := make([]float32, half)
+		for i := int32(0); i < half; i++ {
+			f[i] = float32(math.Exp(-math.Log(float64(ropeTheta)) * float64(i) / float64(half)))
+		}
+		freqs[axis] = mlx.NewArray(f, []int32{1, 1, 1, half})
+	}
+
+	// Extract position coordinates
+	shape := positions.Shape()
+	B := shape[0]
+	L := shape[1]
+
+	// positions[:, :, 0] -> h positions
+	posH := mlx.Slice(positions, []int32{0, 0, 0}, []int32{B, L, 1})
+	posW := mlx.Slice(positions, []int32{0, 0, 1}, []int32{B, L, 2})
+	posT := mlx.Slice(positions, []int32{0, 0, 2}, []int32{B, L, 3})
+
+	// Compute args: pos * freqs for each axis
+	posH = mlx.ExpandDims(posH, 3) // [B, L, 1, 1]
+	posW = mlx.ExpandDims(posW, 3)
+	posT = mlx.ExpandDims(posT, 3)
+
+	argsH := mlx.Mul(posH, freqs[0]) // [B, L, 1, 16]
+	argsW := mlx.Mul(posW, freqs[1]) // [B, L, 1, 24]
+	argsT := mlx.Mul(posT, freqs[2]) // [B, L, 1, 24]
+
+	// Concatenate: [B, L, 1, 16+24+24=64]
+	args := mlx.Concatenate([]*mlx.Array{argsH, argsW, argsT}, 3)
+
+	// Compute cos and sin
+	return mlx.Cos(args), mlx.Sin(args)
+}
+
+// PatchifyLatents converts latents [B, C, H, W] to patches [B, L, C*patch^2]
+// Matches Python: x.reshape(C, 1, 1, H_tok, 2, W_tok, 2).transpose(1,2,3,5,4,6,0).reshape(1,-1,C*4)
+func PatchifyLatents(latents *mlx.Array, patchSize int32) *mlx.Array {
+	shape := latents.Shape()
+	C := shape[1]
+	H := shape[2]
+	W := shape[3]
+
+	pH := H / patchSize // H_tok
+	pW := W / patchSize // W_tok
+
+	// Match Python exactly: reshape treating B=1 as part of contiguous data
+	// [1, C, H, W] -> [C, 1, 1, pH, 2, pW, 2]
+	x := mlx.Reshape(latents, C, 1, 1, pH, patchSize, pW, patchSize)
+
+	// Python: transpose(1, 2, 3, 5, 4, 6, 0)
+	// [C, 1, 1, pH, 2, pW, 2] -> [1, 1, pH, pW, 2, 2, C]
+	x = mlx.Transpose(x, 1, 2, 3, 5, 4, 6, 0)
+
+	// [1, 1, pH, pW, 2, 2, C] -> [1, pH*pW, C*4]
+	return mlx.Reshape(x, 1, pH*pW, C*patchSize*patchSize)
+}
+
+// UnpatchifyLatents converts patches [B, L, C*patch^2] back to [B, C, H, W]
+// Matches Python: out.reshape(1,1,H_tok,W_tok,2,2,C).transpose(6,0,1,2,4,3,5).reshape(1,C,H,W)
+func UnpatchifyLatents(patches *mlx.Array, patchSize, H, W, C int32) *mlx.Array {
+	pH := H / patchSize
+	pW := W / patchSize
+
+	// [1, L, C*4] -> [1, 1, pH, pW, 2, 2, C]
+	x := mlx.Reshape(patches, 1, 1, pH, pW, patchSize, patchSize, C)
+
+	// Python: transpose(6, 0, 1, 2, 4, 3, 5)
+	// [1, 1, pH, pW, 2, 2, C] -> [C, 1, 1, pH, 2, pW, 2]
+	x = mlx.Transpose(x, 6, 0, 1, 2, 4, 3, 5)
+
+	// [C, 1, 1, pH, 2, pW, 2] -> [1, C, H, W]
+	return mlx.Reshape(x, 1, C, H, W)
+}

x/imagegen/models/zimage/vae.go

+//go:build mlx
+
+package zimage
+
+import (
+	"encoding/json"
+	"fmt"
+	"math"
+	"os"
+	"path/filepath"
+
+	"github.com/ollama/ollama/x/imagegen/mlx"
+	"github.com/ollama/ollama/x/imagegen/safetensors"
+)
+
+// VAEConfig holds VAE decoder configuration
+type VAEConfig struct {
+	InChannels       int32   `json:"in_channels"`
+	OutChannels      int32   `json:"out_channels"`
+	LatentChannels   int32   `json:"latent_channels"`
+	BlockOutChannels []int32 `json:"block_out_channels"`
+	LayersPerBlock   int32   `json:"layers_per_block"`
+	NormNumGroups    int32   `json:"norm_num_groups"`
+	ScalingFactor    float32 `json:"scaling_factor"`
+	ShiftFactor      float32 `json:"shift_factor"`
+}
+
+// loadVAEConfig loads VAE config from a JSON file
+func loadVAEConfig(path string) (*VAEConfig, error) {
+	data, err := os.ReadFile(path)
+	if err != nil {
+		return nil, fmt.Errorf("read config: %w", err)
+	}
+	var cfg VAEConfig
+	if err := json.Unmarshal(data, &cfg); err != nil {
+		return nil, fmt.Errorf("parse config: %w", err)
+	}
+	return &cfg, nil
+}
+
+// GroupNormLayer implements group normalization
+type GroupNormLayer struct {
+	Weight    *mlx.Array
+	Bias      *mlx.Array
+	NumGroups int32
+	Eps       float32
+}
+
+// NewGroupNorm creates a group norm layer
+func NewGroupNorm(weight, bias *mlx.Array, numGroups int32) *GroupNormLayer {
+	return &GroupNormLayer{
+		Weight:    weight,
+		Bias:      bias,
+		NumGroups: numGroups,
+		Eps:       1e-5,
+	}
+}
+
+// Forward applies group normalization
+func (gn *GroupNormLayer) Forward(x *mlx.Array) *mlx.Array {
+	// x: [B, C, H, W]
+	shape := x.Shape()
+	B := shape[0]
+	C := shape[1]
+	H := shape[2]
+	W := shape[3]
+
+	// Reshape to [B, groups, C/groups, H, W]
+	groupSize := C / gn.NumGroups
+	x = mlx.Reshape(x, B, gn.NumGroups, groupSize, H, W)
+
+	// Compute mean and variance per group
+	mean := mlx.Mean(x, 2, true)
+	mean = mlx.Mean(mean, 3, true)
+	mean = mlx.Mean(mean, 4, true)
+
+	xCentered := mlx.Sub(x, mean)
+	variance := mlx.Mean(mlx.Square(xCentered), 2, true)
+	variance = mlx.Mean(variance, 3, true)
+	variance = mlx.Mean(variance, 4, true)
+
+	// Normalize
+	xNorm := mlx.Div(xCentered, mlx.Sqrt(mlx.AddScalar(variance, gn.Eps)))
+
+	// Reshape back to [B, C, H, W]
+	xNorm = mlx.Reshape(xNorm, B, C, H, W)
+
+	// Scale and shift (weight and bias are [C])
+	if gn.Weight != nil {
+		weight := mlx.Reshape(gn.Weight, 1, C, 1, 1)
+		xNorm = mlx.Mul(xNorm, weight)
+	}
+	if gn.Bias != nil {
+		bias := mlx.Reshape(gn.Bias, 1, C, 1, 1)
+		xNorm = mlx.Add(xNorm, bias)
+	}
+
+	return xNorm
+}
+
+// Conv2D represents a 2D convolution layer
+// MLX uses NHWC format, but we store weights in OHWI format for MLX conv
+type Conv2D struct {
+	Weight  *mlx.Array // [out_channels, kH, kW, in_channels] (OHWI for MLX)
+	Bias    *mlx.Array // [out_channels]
+	Stride  int32
+	Padding int32
+}
+
+// NewConv2D creates a Conv2D layer
+// weight comes in as [out_channels, in_channels, kH, kW] (OIHW from PyTorch)
+// we transpose to [out_channels, kH, kW, in_channels] (OHWI for MLX)
+func NewConv2D(weight, bias *mlx.Array, stride, padding int32) *Conv2D {
+	// Transpose weight from OIHW to OHWI
+	// [O, I, H, W] -> [O, H, W, I]
+	weightOHWI := mlx.Transpose(weight, 0, 2, 3, 1)
+	return &Conv2D{
+		Weight:  weightOHWI,
+		Bias:    bias,
+		Stride:  stride,
+		Padding: padding,
+	}
+}
+
+// Forward applies convolution
+// Input x is in NCHW format, we convert to NHWC for MLX, then back to NCHW
+func (conv *Conv2D) Forward(x *mlx.Array) *mlx.Array {
+	// x: [N, C, H, W] -> [N, H, W, C]
+	xNHWC := mlx.Transpose(x, 0, 2, 3, 1)
+
+	// Conv in NHWC format
+	outNHWC := mlx.Conv2d(xNHWC, conv.Weight, conv.Stride, conv.Padding)
+
+	// Convert back to NCHW: [N, H, W, C] -> [N, C, H, W]
+	out := mlx.Transpose(outNHWC, 0, 3, 1, 2)
+
+	if conv.Bias != nil {
+		bias := mlx.Reshape(conv.Bias, 1, conv.Bias.Dim(0), 1, 1)
+		out = mlx.Add(out, bias)
+	}
+	return out
+}
+
+// ResnetBlock2D implements a ResNet block for VAE
+type ResnetBlock2D struct {
+	Norm1        *GroupNormLayer
+	Conv1        *Conv2D
+	Norm2        *GroupNormLayer
+	Conv2        *Conv2D
+	ConvShortcut *Conv2D // nil if in_channels == out_channels
+}
+
+// NewResnetBlock2D creates a ResNet block
+func NewResnetBlock2D(weights *safetensors.ModelWeights, prefix string, numGroups int32) (*ResnetBlock2D, error) {
+	norm1Weight, err := weights.GetTensor(prefix + ".norm1.weight")
+	if err != nil {
+		return nil, err
+	}
+	norm1Bias, err := weights.GetTensor(prefix + ".norm1.bias")
+	if err != nil {
+		return nil, err
+	}
+
+	conv1Weight, err := weights.GetTensor(prefix + ".conv1.weight")
+	if err != nil {
+		return nil, err
+	}
+	conv1Bias, err := weights.GetTensor(prefix + ".conv1.bias")
+	if err != nil {
+		return nil, err
+	}
+
+	norm2Weight, err := weights.GetTensor(prefix + ".norm2.weight")
+	if err != nil {
+		return nil, err
+	}
+	norm2Bias, err := weights.GetTensor(prefix + ".norm2.bias")
+	if err != nil {
+		return nil, err
+	}
+
+	conv2Weight, err := weights.GetTensor(prefix + ".conv2.weight")
+	if err != nil {
+		return nil, err
+	}
+	conv2Bias, err := weights.GetTensor(prefix + ".conv2.bias")
+	if err != nil {
+		return nil, err
+	}
+
+	block := &ResnetBlock2D{
+		Norm1: NewGroupNorm(norm1Weight, norm1Bias, numGroups),
+		Conv1: NewConv2D(conv1Weight, conv1Bias, 1, 1),
+		Norm2: NewGroupNorm(norm2Weight, norm2Bias, numGroups),
+		Conv2: NewConv2D(conv2Weight, conv2Bias, 1, 1),
+	}
+
+	if weights.HasTensor(prefix + ".conv_shortcut.weight") {
+		shortcutWeight, err := weights.GetTensor(prefix + ".conv_shortcut.weight")
+		if err != nil {
+			return nil, err
+		}
+		shortcutBias, err := weights.GetTensor(prefix + ".conv_shortcut.bias")
+		if err != nil {
+			return nil, err
+		}
+		block.ConvShortcut = NewConv2D(shortcutWeight, shortcutBias, 1, 0)
+	}
+
+	return block, nil
+}
+
+// Forward applies the ResNet block with staged evaluation
+func (rb *ResnetBlock2D) Forward(x *mlx.Array) *mlx.Array {
+	var h *mlx.Array
+
+	// Stage 1: norm1
+	{
+			h = rb.Norm1.Forward(x)
+		mlx.Eval(h)
+	}
+
+	// Stage 2: silu + conv1
+	{
+			prev := h
+		h = mlx.SiLU(h)
+		h = rb.Conv1.Forward(h)
+		prev.Free()
+		mlx.Eval(h)
+	}
+
+	// Stage 3: norm2
+	{
+			prev := h
+		h = rb.Norm2.Forward(h)
+		prev.Free()
+		mlx.Eval(h)
+	}
+
+	// Stage 4: silu + conv2
+	{
+			prev := h
+		h = mlx.SiLU(h)
+		h = rb.Conv2.Forward(h)
+		prev.Free()
+		mlx.Eval(h)
+	}
+
+	// Residual connection
+	{
+			prev := h
+		if rb.ConvShortcut != nil {
+			shortcut := rb.ConvShortcut.Forward(x)
+			h = mlx.Add(h, shortcut)
+		} else {
+			h = mlx.Add(h, x)
+		}
+		prev.Free()
+		mlx.Eval(h)
+	}
+
+	return h
+}
+
+// VAEAttentionBlock implements self-attention for VAE
+type VAEAttentionBlock struct {
+	GroupNorm   *GroupNormLayer
+	ToQWeight   *mlx.Array
+	ToQBias     *mlx.Array
+	ToKWeight   *mlx.Array
+	ToKBias     *mlx.Array
+	ToVWeight   *mlx.Array
+	ToVBias     *mlx.Array
+	ToOutWeight *mlx.Array
+	ToOutBias   *mlx.Array
+	NumHeads    int32
+}
+
+// NewVAEAttentionBlock creates an attention block
+func NewVAEAttentionBlock(weights *safetensors.ModelWeights, prefix string, numGroups int32) (*VAEAttentionBlock, error) {
+	normWeight, err := weights.GetTensor(prefix + ".group_norm.weight")
+	if err != nil {
+		return nil, err
+	}
+	normBias, err := weights.GetTensor(prefix + ".group_norm.bias")
+	if err != nil {
+		return nil, err
+	}
+
+	toQWeight, err := weights.GetTensor(prefix + ".to_q.weight")
+	if err != nil {
+		return nil, err
+	}
+	toQBias, err := weights.GetTensor(prefix + ".to_q.bias")
+	if err != nil {
+		return nil, err
+	}
+
+	toKWeight, err := weights.GetTensor(prefix + ".to_k.weight")
+	if err != nil {
+		return nil, err
+	}
+	toKBias, err := weights.GetTensor(prefix + ".to_k.bias")
+	if err != nil {
+		return nil, err
+	}
+
+	toVWeight, err := weights.GetTensor(prefix + ".to_v.weight")
+	if err != nil {
+		return nil, err
+	}
+	toVBias, err := weights.GetTensor(prefix + ".to_v.bias")
+	if err != nil {
+		return nil, err
+	}
+
+	toOutWeight, err := weights.GetTensor(prefix + ".to_out.0.weight")
+	if err != nil {
+		return nil, err
+	}
+	toOutBias, err := weights.GetTensor(prefix + ".to_out.0.bias")
+	if err != nil {
+		return nil, err
+	}
+
+	return &VAEAttentionBlock{
+		GroupNorm:   NewGroupNorm(normWeight, normBias, numGroups),
+		ToQWeight:   mlx.Transpose(toQWeight, 1, 0),
+		ToQBias:     toQBias,
+		ToKWeight:   mlx.Transpose(toKWeight, 1, 0),
+		ToKBias:     toKBias,
+		ToVWeight:   mlx.Transpose(toVWeight, 1, 0),
+		ToVBias:     toVBias,
+		ToOutWeight: mlx.Transpose(toOutWeight, 1, 0),
+		ToOutBias:   toOutBias,
+		NumHeads:    1,
+	}, nil
+}
+
+// Forward applies attention with staged evaluation
+func (ab *VAEAttentionBlock) Forward(x *mlx.Array) *mlx.Array {
+	residual := x
+	shape := x.Shape()
+	B := shape[0]
+	C := shape[1]
+	H := shape[2]
+	W := shape[3]
+
+	var h *mlx.Array
+
+	// Stage 1: GroupNorm + reshape
+	{
+			h = ab.GroupNorm.Forward(x)
+		h = mlx.Transpose(h, 0, 2, 3, 1)
+		h = mlx.Reshape(h, B, H*W, C)
+		mlx.Eval(h)
+	}
+
+	var out *mlx.Array
+
+	// Stage 2: Q, K, V projections + attention
+	{
+			q := mlx.Linear(h, ab.ToQWeight)
+		q = mlx.Add(q, ab.ToQBias)
+		k := mlx.Linear(h, ab.ToKWeight)
+		k = mlx.Add(k, ab.ToKBias)
+		v := mlx.Linear(h, ab.ToVWeight)
+		v = mlx.Add(v, ab.ToVBias)
+		h.Free()
+
+		q = mlx.ExpandDims(q, 1)
+		k = mlx.ExpandDims(k, 1)
+		v = mlx.ExpandDims(v, 1)
+
+		scale := float32(1.0 / math.Sqrt(float64(C)))
+		out = mlx.ScaledDotProductAttention(q, k, v, scale, false)
+		out = mlx.Squeeze(out, 1)
+		mlx.Eval(out)
+	}
+
+	// Stage 3: Output projection + reshape + residual
+	{
+			prev := out
+		out = mlx.Linear(out, ab.ToOutWeight)
+		out = mlx.Add(out, ab.ToOutBias)
+		out = mlx.Reshape(out, B, H, W, C)
+		out = mlx.Transpose(out, 0, 3, 1, 2)
+		out = mlx.Add(out, residual)
+		prev.Free()
+		mlx.Eval(out)
+	}
+
+	return out
+}
+
+// UpDecoderBlock2D implements an upsampling decoder block
+type UpDecoderBlock2D struct {
+	ResnetBlocks []*ResnetBlock2D
+	Upsample     *Conv2D
+}
+
+// NewUpDecoderBlock2D creates an up decoder block
+func NewUpDecoderBlock2D(weights *safetensors.ModelWeights, prefix string, numLayers, numGroups int32, hasUpsample bool) (*UpDecoderBlock2D, error) {
+	resnets := make([]*ResnetBlock2D, numLayers)
+	for i := int32(0); i < numLayers; i++ {
+		resPrefix := fmt.Sprintf("%s.resnets.%d", prefix, i)
+		resnet, err := NewResnetBlock2D(weights, resPrefix, numGroups)
+		if err != nil {
+			return nil, err
+		}
+		resnets[i] = resnet
+	}
+
+	var upsample *Conv2D
+	if hasUpsample {
+		upWeight, err := weights.GetTensor(prefix + ".upsamplers.0.conv.weight")
+		if err != nil {
+			return nil, err
+		}
+		upBias, err := weights.GetTensor(prefix + ".upsamplers.0.conv.bias")
+		if err != nil {
+			return nil, err
+		}
+		upsample = NewConv2D(upWeight, upBias, 1, 1)
+	}
+
+	return &UpDecoderBlock2D{
+		ResnetBlocks: resnets,
+		Upsample:     upsample,
+	}, nil
+}
+
+// Forward applies the up decoder block with staged evaluation to reduce peak memory
+func (ub *UpDecoderBlock2D) Forward(x *mlx.Array) *mlx.Array {
+	for _, resnet := range ub.ResnetBlocks {
+		prev := x
+		x = resnet.Forward(x) // ResNet handles its own pools
+		prev.Free()
+	}
+
+	if ub.Upsample != nil {
+		// Stage 1: Upsample2x (nearest neighbor)
+		{
+					prev := x
+			x = Upsample2x(x)
+			prev.Free()
+			mlx.Eval(x)
+		}
+
+		// Stage 2: Upsample conv
+		{
+					prev := x
+			x = ub.Upsample.Forward(x)
+			prev.Free()
+			mlx.Eval(x)
+		}
+	}
+
+	return x
+}
+
+// VAEMidBlock is the middle block with attention
+type VAEMidBlock struct {
+	Resnet1   *ResnetBlock2D
+	Attention *VAEAttentionBlock
+	Resnet2   *ResnetBlock2D
+}
+
+// NewVAEMidBlock creates the mid block
+func NewVAEMidBlock(weights *safetensors.ModelWeights, prefix string, numGroups int32) (*VAEMidBlock, error) {
+	resnet1, err := NewResnetBlock2D(weights, prefix+".resnets.0", numGroups)
+	if err != nil {
+		return nil, err
+	}
+
+	attention, err := NewVAEAttentionBlock(weights, prefix+".attentions.0", numGroups)
+	if err != nil {
+		return nil, err
+	}
+
+	resnet2, err := NewResnetBlock2D(weights, prefix+".resnets.1", numGroups)
+	if err != nil {
+		return nil, err
+	}
+
+	return &VAEMidBlock{
+		Resnet1:   resnet1,
+		Attention: attention,
+		Resnet2:   resnet2,
+	}, nil
+}
+
+// Forward applies the mid block with staged evaluation
+func (mb *VAEMidBlock) Forward(x *mlx.Array) *mlx.Array {
+	prev := x
+	x = mb.Resnet1.Forward(x) // ResNet handles its own pools
+	prev.Free()
+
+	// Attention handles its own pools
+	prev = x
+	x = mb.Attention.Forward(x)
+	prev.Free()
+
+	prev = x
+	x = mb.Resnet2.Forward(x) // ResNet handles its own pools
+	prev.Free()
+
+	return x
+}
+
+// VAEDecoder is the full VAE decoder
+type VAEDecoder struct {
+	Config      *VAEConfig
+	ConvIn      *Conv2D
+	MidBlock    *VAEMidBlock
+	UpBlocks    []*UpDecoderBlock2D
+	ConvNormOut *GroupNormLayer
+	ConvOut     *Conv2D
+}
+
+// Load loads the VAE decoder from a directory
+func (m *VAEDecoder) Load(path string) error {
+	fmt.Println("Loading VAE decoder...")
+
+	// Load config
+	cfg, err := loadVAEConfig(filepath.Join(path, "config.json"))
+	if err != nil {
+		return fmt.Errorf("config: %w", err)
+	}
+	m.Config = cfg
+
+	// Load weights
+	weights, err := safetensors.LoadModelWeights(path)
+	if err != nil {
+		return fmt.Errorf("weights: %w", err)
+	}
+
+	// Load conv_in
+	fmt.Print("  Loading conv_in... ")
+	convInWeight, err := weights.GetTensor("decoder.conv_in.weight")
+	if err != nil {
+		return err
+	}
+	convInBias, err := weights.GetTensor("decoder.conv_in.bias")
+	if err != nil {
+		return err
+	}
+	m.ConvIn = NewConv2D(convInWeight, convInBias, 1, 1)
+	fmt.Println("✓")
+
+	// Load mid block
+	fmt.Print("  Loading mid block... ")
+	m.MidBlock, err = NewVAEMidBlock(weights, "decoder.mid_block", cfg.NormNumGroups)
+	if err != nil {
+		return err
+	}
+	fmt.Println("✓")
+
+	// Load up blocks
+	fmt.Print("  Loading up blocks... ")
+	numBlocks := len(cfg.BlockOutChannels)
+	m.UpBlocks = make([]*UpDecoderBlock2D, numBlocks)
+	for i := 0; i < numBlocks; i++ {
+		prefix := fmt.Sprintf("decoder.up_blocks.%d", i)
+		hasUpsample := i < numBlocks-1
+		m.UpBlocks[i], err = NewUpDecoderBlock2D(weights, prefix, cfg.LayersPerBlock+1, cfg.NormNumGroups, hasUpsample)
+		if err != nil {
+			return err
+		}
+	}
+	fmt.Printf("✓ [%d blocks]\n", numBlocks)
+
+	// Load conv_norm_out
+	fmt.Print("  Loading conv_norm_out... ")
+	normWeight, err := weights.GetTensor("decoder.conv_norm_out.weight")
+	if err != nil {
+		return err
+	}
+	normBias, err := weights.GetTensor("decoder.conv_norm_out.bias")
+	if err != nil {
+		return err
+	}
+	m.ConvNormOut = NewGroupNorm(normWeight, normBias, cfg.NormNumGroups)
+	fmt.Println("✓")
+
+	// Load conv_out
+	fmt.Print("  Loading conv_out... ")
+	convOutWeight, err := weights.GetTensor("decoder.conv_out.weight")
+	if err != nil {
+		return err
+	}
+	convOutBias, err := weights.GetTensor("decoder.conv_out.bias")
+	if err != nil {
+		return err
+	}
+	m.ConvOut = NewConv2D(convOutWeight, convOutBias, 1, 1)
+	fmt.Println("✓")
+
+	weights.ReleaseAll()
+	return nil
+}
+
+// Decode decodes latents to images.
+// Uses staged pools to free intermediate arrays and reduce peak memory.
+func (vae *VAEDecoder) Decode(latents *mlx.Array) *mlx.Array {
+	var h *mlx.Array
+	{
+		z := mlx.DivScalar(latents, vae.Config.ScalingFactor)
+		z = mlx.AddScalar(z, vae.Config.ShiftFactor)
+		h = vae.ConvIn.Forward(z)
+		mlx.Eval(h)
+	}
+
+	h = vae.MidBlock.Forward(h)
+
+	for _, upBlock := range vae.UpBlocks {
+		h = upBlock.Forward(h)
+	}
+
+	{
+			prev := h
+		h = vae.ConvNormOut.Forward(h)
+		h = mlx.SiLU(h)
+		h = vae.ConvOut.Forward(h)
+		// VAE outputs [-1, 1], convert to [0, 1]
+		h = mlx.AddScalar(mlx.MulScalar(h, 0.5), 0.5)
+		h = mlx.ClipScalar(h, 0.0, 1.0, true, true)
+		prev.Free()
+		mlx.Eval(h)
+	}
+
+	return h
+}
+
+// Upsample2x performs 2x nearest neighbor upsampling using broadcast.
+// x: [B, C, H, W] -> [B, C, H*2, W*2]
+func Upsample2x(x *mlx.Array) *mlx.Array {
+	shape := x.Shape()
+	B := shape[0]
+	C := shape[1]
+	H := shape[2]
+	W := shape[3]
+
+	// [B, C, H, W] -> [B, C, H, 1, W, 1]
+	x = mlx.Reshape(x, B, C, H, 1, W, 1)
+	// Broadcast to [B, C, H, 2, W, 2]
+	x = mlx.BroadcastTo(x, []int32{B, C, H, 2, W, 2})
+	// Reshape to [B, C, H*2, W*2]
+	x = mlx.Reshape(x, B, C, H*2, W*2)
+
+	return x
+}

x/imagegen/models/zimage/zimage.go

+//go:build mlx
+
+// Package zimage implements the Z-Image diffusion transformer model.
+package zimage
+
+import (
+	"context"
+	"fmt"
+	"path/filepath"
+	"time"
+
+	"github.com/ollama/ollama/x/imagegen/cache"
+	"github.com/ollama/ollama/x/imagegen/mlx"
+	"github.com/ollama/ollama/x/imagegen/tokenizer"
+)
+
+// GenerateConfig holds all options for image generation.
+type GenerateConfig struct {
+	Prompt         string
+	NegativePrompt string       // Empty = no CFG
+	CFGScale       float32      // Only used if NegativePrompt is set (default: 4.0)
+	Width          int32        // Image width (default: 1024)
+	Height         int32        // Image height (default: 1024)
+	Steps          int          // Denoising steps (default: 9 for turbo)
+	Seed           int64        // Random seed
+	Progress       ProgressFunc // Optional progress callback
+	CapturePath    string       // GPU capture path (debug)
+
+	// Layer caching options (speedup via shallow layer reuse)
+	LayerCache    bool // Enable layer caching (default: false)
+	CacheInterval int  // Refresh cache every N steps (default: 3)
+	CacheLayers   int  // Number of shallow layers to cache (default: 15)
+}
+
+// ProgressFunc is called during generation with step progress.
+type ProgressFunc func(step, totalSteps int)
+
+// Model represents a Z-Image diffusion model.
+type Model struct {
+	ModelPath   string
+	Tokenizer   *tokenizer.Tokenizer
+	TextEncoder *Qwen3TextEncoder
+	Transformer *Transformer
+	VAEDecoder  *VAEDecoder
+}
+
+// Load loads the Z-Image model from a directory.
+func (m *Model) Load(modelPath string) error {
+	fmt.Println("Loading Z-Image model...")
+	start := time.Now()
+
+	if mlx.GPUIsAvailable() {
+		mlx.SetDefaultDeviceGPU()
+		mlx.EnableCompile()
+	}
+
+	m.ModelPath = modelPath
+
+	// Load tokenizer
+	fmt.Print("  Loading tokenizer... ")
+	tokenizerPath := filepath.Join(modelPath, "tokenizer", "tokenizer.json")
+	tok, err := tokenizer.Load(tokenizerPath)
+	if err != nil {
+		return fmt.Errorf("tokenizer: %w", err)
+	}
+	m.Tokenizer = tok
+	fmt.Println("✓")
+
+	// Load text encoder
+	m.TextEncoder = &Qwen3TextEncoder{}
+	if err := m.TextEncoder.Load(filepath.Join(modelPath, "text_encoder")); err != nil {
+		return fmt.Errorf("text encoder: %w", err)
+	}
+	mlx.Eval(mlx.Collect(m.TextEncoder)...)
+	fmt.Printf("  (%.1f GB, peak %.1f GB)\n",
+		float64(mlx.MetalGetActiveMemory())/(1024*1024*1024),
+		float64(mlx.MetalGetPeakMemory())/(1024*1024*1024))
+
+	// Load transformer
+	m.Transformer = &Transformer{}
+	if err := m.Transformer.Load(filepath.Join(modelPath, "transformer")); err != nil {
+		return fmt.Errorf("transformer: %w", err)
+	}
+	mlx.Eval(mlx.Collect(m.Transformer)...)
+	fmt.Printf("  (%.1f GB, peak %.1f GB)\n",
+		float64(mlx.MetalGetActiveMemory())/(1024*1024*1024),
+		float64(mlx.MetalGetPeakMemory())/(1024*1024*1024))
+
+	// Load VAE decoder
+	m.VAEDecoder = &VAEDecoder{}
+	if err := m.VAEDecoder.Load(filepath.Join(modelPath, "vae")); err != nil {
+		return fmt.Errorf("VAE decoder: %w", err)
+	}
+	mlx.Eval(mlx.Collect(m.VAEDecoder)...)
+	fmt.Printf("  (%.1f GB, peak %.1f GB)\n",
+		float64(mlx.MetalGetActiveMemory())/(1024*1024*1024),
+		float64(mlx.MetalGetPeakMemory())/(1024*1024*1024))
+
+	mem := mlx.MetalGetActiveMemory()
+	fmt.Printf("  Loaded in %.2fs (%.1f GB VRAM)\n", time.Since(start).Seconds(), float64(mem)/(1024*1024*1024))
+
+	return nil
+}
+
+// Generate creates an image from a prompt.
+func (m *Model) Generate(prompt string, width, height int32, steps int, seed int64) (*mlx.Array, error) {
+	return m.GenerateFromConfig(&GenerateConfig{
+		Prompt: prompt,
+		Width:  width,
+		Height: height,
+		Steps:  steps,
+		Seed:   seed,
+	})
+}
+
+// GenerateWithProgress creates an image with progress callback.
+func (m *Model) GenerateWithProgress(prompt string, width, height int32, steps int, seed int64, progress ProgressFunc) (*mlx.Array, error) {
+	return m.GenerateFromConfig(&GenerateConfig{
+		Prompt:   prompt,
+		Width:    width,
+		Height:   height,
+		Steps:    steps,
+		Seed:     seed,
+		Progress: progress,
+	})
+}
+
+// GenerateWithCFG creates an image with classifier-free guidance.
+func (m *Model) GenerateWithCFG(prompt, negativePrompt string, width, height int32, steps int, seed int64, cfgScale float32, progress ProgressFunc) (*mlx.Array, error) {
+	return m.GenerateFromConfig(&GenerateConfig{
+		Prompt:         prompt,
+		NegativePrompt: negativePrompt,
+		CFGScale:       cfgScale,
+		Width:          width,
+		Height:         height,
+		Steps:          steps,
+		Seed:           seed,
+		Progress:       progress,
+	})
+}
+
+// GenerateFromConfig generates an image using the unified config struct.
+func (m *Model) GenerateFromConfig(cfg *GenerateConfig) (*mlx.Array, error) {
+	start := time.Now()
+	result, err := m.generate(cfg)
+	if err != nil {
+		return nil, err
+	}
+	if cfg.NegativePrompt != "" {
+		fmt.Printf("Generated with CFG (scale=%.1f) in %.2fs (%d steps)\n", cfg.CFGScale, time.Since(start).Seconds(), cfg.Steps)
+	} else {
+		fmt.Printf("Generated in %.2fs (%d steps)\n", time.Since(start).Seconds(), cfg.Steps)
+	}
+	return result, nil
+}
+
+// GenerateImage implements model.ImageModel interface.
+func (m *Model) GenerateImage(ctx context.Context, prompt string, width, height int32, steps int, seed int64) (*mlx.Array, error) {
+	return m.Generate(prompt, width, height, steps, seed)
+}
+
+// generate is the internal denoising pipeline.
+func (m *Model) generate(cfg *GenerateConfig) (*mlx.Array, error) {
+	// Apply defaults
+	if cfg.Width <= 0 {
+		cfg.Width = 1024
+	}
+	if cfg.Height <= 0 {
+		cfg.Height = 1024
+	}
+	if cfg.Steps <= 0 {
+		cfg.Steps = 9 // Turbo default
+	}
+	if cfg.CFGScale <= 0 {
+		cfg.CFGScale = 4.0
+	}
+	if cfg.LayerCache {
+		if cfg.CacheInterval <= 0 {
+			cfg.CacheInterval = 3
+		}
+		if cfg.CacheLayers <= 0 {
+			cfg.CacheLayers = 15 // Half of 30 layers
+		}
+	}
+
+	useCFG := cfg.NegativePrompt != ""
+	tcfg := m.Transformer.TransformerConfig
+	latentH := cfg.Height / 8
+	latentW := cfg.Width / 8
+	hTok := latentH / tcfg.PatchSize
+	wTok := latentW / tcfg.PatchSize
+
+	// Text encoding with padding to multiple of 32
+	var posEmb, negEmb *mlx.Array
+	{
+		posEmb, _ = m.TextEncoder.EncodePrompt(m.Tokenizer, cfg.Prompt, 512)
+		if useCFG {
+			negEmb, _ = m.TextEncoder.EncodePrompt(m.Tokenizer, cfg.NegativePrompt, 512)
+		}
+
+		// Pad both to same length (multiple of 32)
+		maxLen := posEmb.Shape()[1]
+		if useCFG && negEmb.Shape()[1] > maxLen {
+			maxLen = negEmb.Shape()[1]
+		}
+		if pad := (32 - (maxLen % 32)) % 32; pad > 0 {
+			maxLen += pad
+		}
+
+		posEmb = padToLength(posEmb, maxLen)
+		if useCFG {
+			negEmb = padToLength(negEmb, maxLen)
+			mlx.Keep(posEmb, negEmb)
+			mlx.Eval(posEmb, negEmb)
+		} else {
+			mlx.Keep(posEmb)
+			mlx.Eval(posEmb)
+		}
+	}
+
+	// Scheduler
+	scheduler := NewFlowMatchEulerScheduler(DefaultFlowMatchSchedulerConfig())
+	scheduler.SetTimestepsWithMu(cfg.Steps, CalculateShift(hTok*wTok))
+
+	// Init latents [B, C, H, W]
+	var latents *mlx.Array
+	{
+		latents = scheduler.InitNoise([]int32{1, tcfg.InChannels, latentH, latentW}, cfg.Seed)
+		mlx.Eval(latents)
+	}
+
+	// RoPE cache
+	var ropeCache *RoPECache
+	{
+		ropeCache = m.Transformer.PrepareRoPECache(hTok, wTok, posEmb.Shape()[1])
+		mlx.Keep(ropeCache.ImgCos, ropeCache.ImgSin, ropeCache.CapCos, ropeCache.CapSin,
+			ropeCache.UnifiedCos, ropeCache.UnifiedSin)
+		mlx.Eval(ropeCache.UnifiedCos)
+	}
+
+	// Step cache for shallow layer reuse (DeepCache/Learning-to-Cache style)
+	var stepCache *cache.StepCache
+	if cfg.LayerCache {
+		stepCache = cache.NewStepCache(cfg.CacheLayers)
+		fmt.Printf("  Layer caching enabled: %d layers, refresh every %d steps\n",
+			cfg.CacheLayers, cfg.CacheInterval)
+	}
+
+	// Denoising loop
+	for i := 0; i < cfg.Steps; i++ {
+		stepStart := time.Now()
+		if cfg.Progress != nil {
+			cfg.Progress(i+1, cfg.Steps)
+		}
+
+		// GPU capture on step 2 if requested
+		if cfg.CapturePath != "" && i == 1 {
+			mlx.MetalStartCapture(cfg.CapturePath)
+		}
+
+		tCurr := scheduler.Timesteps[i]
+		timestep := mlx.ToBFloat16(mlx.NewArray([]float32{1.0 - tCurr}, []int32{1}))
+
+		patches := PatchifyLatents(latents, tcfg.PatchSize)
+
+		var output *mlx.Array
+		if stepCache != nil {
+			// Use layer caching for faster inference
+			if useCFG {
+				posOutput := m.Transformer.ForwardWithCache(patches, timestep, posEmb, ropeCache,
+					stepCache, i, cfg.CacheInterval)
+				// Note: CFG with layer cache shares the cache between pos/neg
+				// This is approximate but fast - neg prompt uses same cached shallow layers
+				negOutput := m.Transformer.ForwardWithCache(patches, timestep, negEmb, ropeCache,
+					stepCache, i, cfg.CacheInterval)
+				diff := mlx.Sub(posOutput, negOutput)
+				scaledDiff := mlx.MulScalar(diff, cfg.CFGScale)
+				output = mlx.Add(negOutput, scaledDiff)
+			} else {
+				output = m.Transformer.ForwardWithCache(patches, timestep, posEmb, ropeCache,
+					stepCache, i, cfg.CacheInterval)
+			}
+		} else {
+			// Standard forward without caching
+			if useCFG {
+				posOutput := m.Transformer.Forward(patches, timestep, posEmb, ropeCache)
+				negOutput := m.Transformer.Forward(patches, timestep, negEmb, ropeCache)
+				diff := mlx.Sub(posOutput, negOutput)
+				scaledDiff := mlx.MulScalar(diff, cfg.CFGScale)
+				output = mlx.Add(negOutput, scaledDiff)
+			} else {
+				output = m.Transformer.Forward(patches, timestep, posEmb, ropeCache)
+			}
+		}
+
+		noisePred := UnpatchifyLatents(output, tcfg.PatchSize, latentH, latentW, tcfg.InChannels)
+		noisePred = mlx.Neg(noisePred)
+		oldLatents := latents
+		latents = scheduler.Step(noisePred, latents, i)
+
+		// Keep latents and any cached arrays
+		if stepCache != nil {
+			mlx.Keep(stepCache.Arrays()...)
+		}
+		mlx.Eval(latents)
+		oldLatents.Free()
+
+		if cfg.CapturePath != "" && i == 1 {
+			mlx.MetalStopCapture()
+		}
+
+		activeMem := float64(mlx.MetalGetActiveMemory()) / (1024 * 1024 * 1024)
+		peakMem := float64(mlx.MetalGetPeakMemory()) / (1024 * 1024 * 1024)
+		fmt.Printf("  Step %d/%d: t=%.4f (%.2fs) [%.1f GB active, %.1f GB peak]\n",
+			i+1, cfg.Steps, tCurr, time.Since(stepStart).Seconds(), activeMem, peakMem)
+	}
+
+	// Free denoising temporaries before VAE decode
+	posEmb.Free()
+	if negEmb != nil {
+		negEmb.Free()
+	}
+	ropeCache.ImgCos.Free()
+	ropeCache.ImgSin.Free()
+	ropeCache.CapCos.Free()
+	ropeCache.CapSin.Free()
+	ropeCache.UnifiedCos.Free()
+	ropeCache.UnifiedSin.Free()
+	if stepCache != nil {
+		stepCache.Free()
+	}
+
+	// VAE decode
+	decoded := m.VAEDecoder.Decode(latents)
+	latents.Free()
+
+	return decoded, nil
+}
+
+// padToLength pads a sequence tensor to the target length by repeating the last token.
+func padToLength(x *mlx.Array, targetLen int32) *mlx.Array {
+	shape := x.Shape()
+	currentLen := shape[1]
+	if currentLen >= targetLen {
+		return x
+	}
+	padLen := targetLen - currentLen
+	lastToken := mlx.Slice(x, []int32{0, currentLen - 1, 0}, []int32{shape[0], currentLen, shape[2]})
+	padding := mlx.Tile(lastToken, []int32{1, padLen, 1})
+	return mlx.Concatenate([]*mlx.Array{x, padding}, 1)
+}
+
+// CalculateShift computes the mu shift value for dynamic scheduling
+func CalculateShift(imgSeqLen int32) float32 {
+	baseSeqLen := float32(256)
+	maxSeqLen := float32(4096)
+	baseShift := float32(0.5)
+	maxShift := float32(1.15)
+
+	m := (maxShift - baseShift) / (maxSeqLen - baseSeqLen)
+	b := baseShift - m*baseSeqLen
+	return float32(imgSeqLen)*m + b
+}

x/imagegen/nn/nn.go

+//go:build mlx
+
+// Package nn provides neural network layer types.
+package nn
+
+import "github.com/ollama/ollama/x/imagegen/mlx"
+
+// Layer is the interface for neural network layers with a Forward method.
+type Layer interface {
+	Forward(x *mlx.Array) *mlx.Array
+}
+
+// Linear applies an affine transformation: y = x @ W.T + b
+// Weight is stored as [out_features, in_features], matching PyTorch/MLX convention.
+type Linear struct {
+	Weight *mlx.Array `weight:"weight"`          // [out_features, in_features]
+	Bias   *mlx.Array `weight:"bias,optional"`   // [out_features] or nil
+}
+
+// NewLinear creates a linear layer.
+// Weight should be [out_features, in_features].
+func NewLinear(weight *mlx.Array, bias *mlx.Array) *Linear {
+	return &Linear{Weight: weight, Bias: bias}
+}
+
+// NewQuantizedLinear creates a quantized linear layer directly from bf16 weights.
+// Quantizes the weight immediately and evaluates to break lazy dependencies.
+func NewQuantizedLinear(weight *mlx.Array, bias *mlx.Array, groupSize, bits int, mode string) *QuantizedLinear {
+	qw, scales, qbiases := mlx.Quantize(weight, groupSize, bits, mode)
+	// Eval immediately so bf16 weight can be freed
+	mlx.Eval(qw, scales, qbiases)
+	return &QuantizedLinear{
+		Weight:    qw,
+		Scales:    scales,
+		QBiases:   qbiases,
+		Bias:      bias,
+		GroupSize: groupSize,
+		Bits:      bits,
+		Mode:      mode,
+	}
+}
+
+// Forward applies the linear transformation: x @ W.T + bias
+func (l *Linear) Forward(x *mlx.Array) *mlx.Array {
+	w := mlx.Transpose(l.Weight, 1, 0)
+	if l.Bias != nil {
+		return mlx.AddMM(l.Bias, x, w, 1.0, 1.0)
+	}
+	return mlx.Linear(x, w)
+}
+
+// ToQuantized converts this Linear to a QuantizedLinear.
+func (l *Linear) ToQuantized(groupSize, bits int, mode string) *QuantizedLinear {
+	qw, scales, qbiases := mlx.Quantize(l.Weight, groupSize, bits, mode)
+	return &QuantizedLinear{
+		Weight:    qw,
+		Scales:    scales,
+		QBiases:   qbiases,
+		Bias:      l.Bias,
+		GroupSize: groupSize,
+		Bits:      bits,
+		Mode:      mode,
+	}
+}
+
+// QuantizedLinear applies an affine transformation using quantized weights.
+// Equivalent to mlx.nn.QuantizedLinear.
+type QuantizedLinear struct {
+	Weight    *mlx.Array // Quantized weight data
+	Scales    *mlx.Array // Scale factors for dequantization
+	QBiases   *mlx.Array // Quantization biases (NOT layer bias)
+	Bias      *mlx.Array // Layer bias [output_dims] or nil
+	GroupSize int
+	Bits      int
+	Mode      string
+}
+
+// Forward applies the quantized linear transformation.
+func (ql *QuantizedLinear) Forward(x *mlx.Array) *mlx.Array {
+	out := mlx.QuantizedMatmul(x, ql.Weight, ql.Scales, ql.QBiases, true, ql.GroupSize, ql.Bits, ql.Mode)
+	if ql.Bias != nil {
+		out = mlx.Add(out, ql.Bias)
+	}
+	return out
+}
+
+// RMSNorm represents an RMS normalization layer.
+type RMSNorm struct {
+	Weight *mlx.Array `weight:"weight"`
+	Eps    float32    // optional: used if Forward called with eps=0
+}
+
+// NewRMSNorm creates an RMSNorm layer (for models not using weight loader).
+func NewRMSNorm(weight *mlx.Array, eps float32) *RMSNorm {
+	return &RMSNorm{Weight: weight, Eps: eps}
+}
+
+// Forward applies RMS normalization. If eps=0, uses stored Eps.
+func (rn *RMSNorm) Forward(x *mlx.Array, eps float32) *mlx.Array {
+	if eps == 0 {
+		eps = rn.Eps
+	}
+	return mlx.RMSNorm(x, rn.Weight, eps)
+}
+
+// Embedding represents an embedding layer.
+type Embedding struct {
+	Weight *mlx.Array `weight:"weight"`
+}
+
+// NewEmbedding creates an embedding layer.
+func NewEmbedding(weight *mlx.Array) *Embedding {
+	return &Embedding{Weight: weight}
+}
+
+// Forward looks up embeddings by indices.
+func (e *Embedding) Forward(indices *mlx.Array) *mlx.Array {
+	return mlx.Take(e.Weight, indices, 0)
+}
+
+// RepeatKV repeats K/V tensors for grouped query attention
+// x: [B, num_kv_heads, S, head_dim] -> [B, num_heads, S, head_dim]
+func RepeatKV(x *mlx.Array, repeatFactor int32) *mlx.Array {
+	if repeatFactor == 1 {
+		return x
+	}
+	shape := x.Shape()
+	// [B, num_kv_heads, S, head_dim] -> [B, num_kv_heads, 1, S, head_dim]
+	x = mlx.ExpandDims(x, 2)
+	// Repeat along the new axis
+	reps := []int32{1, 1, repeatFactor, 1, 1}
+	x = mlx.Tile(x, reps)
+	// Reshape: [B, num_kv_heads, repeat, S, head_dim] -> [B, num_kv_heads * repeat, S, head_dim]
+	return mlx.Reshape(x, shape[0], shape[1]*repeatFactor, shape[2], shape[3])
+}
+
+// ApplyCausalMask applies causal (lower triangular) mask to attention scores
+func ApplyCausalMask(scores *mlx.Array) *mlx.Array {
+	// scores: [B, num_heads, S, S]
+	shape := scores.Shape()
+	seqLen := shape[2]
+
+	// Create causal mask: 1 for positions to keep, 0 for positions to mask
+	mask := mlx.Tri(seqLen, seqLen, 0)
+
+	// Where mask is 0, set score to -inf
+	negInf := mlx.NewScalarArray(float32(-1e9))
+
+	// Broadcast mask to match scores shape
+	mask = mlx.ExpandDims(mlx.ExpandDims(mask, 0), 0) // [1, 1, S, S]
+
+	// Use where: if mask > 0, keep scores, else -inf
+	return mlx.Where(mask, scores, negInf)
+}
+
+// ApplyCausalMaskWithOffset applies causal mask for cached attention
+// scores: [B, num_heads, queryLen, keyLen] where keyLen = cacheLen + queryLen
+// offset: the starting position of the new queries (i.e., cache length)
+func ApplyCausalMaskWithOffset(scores *mlx.Array, offset int32) *mlx.Array {
+	if offset == 0 {
+		return ApplyCausalMask(scores)
+	}
+
+	shape := scores.Shape()
+	queryLen := shape[2]
+	keyLen := shape[3]
+
+	// For cached attention, new queries can attend to all cached keys plus
+	// new keys up to and including their position.
+	mask := mlx.Tri(queryLen, keyLen, int(offset))
+
+	negInf := mlx.NewScalarArray(float32(-1e9))
+	mask = mlx.ExpandDims(mlx.ExpandDims(mask, 0), 0) // [1, 1, queryLen, keyLen]
+
+	return mlx.Where(mask, scores, negInf)
+}
+
+// LayerNorm represents a standard layer normalization layer (with bias).
+type LayerNorm struct {
+	Weight *mlx.Array `weight:"weight"`
+	Bias   *mlx.Array `weight:"bias"`
+	Eps    float32
+}
+
+// Forward applies layer normalization: (x - mean) / sqrt(var + eps) * weight + bias
+func (ln *LayerNorm) Forward(x *mlx.Array) *mlx.Array {
+	eps := ln.Eps
+	if eps == 0 {
+		eps = 1e-5
+	}
+	// Compute mean and variance along last dimension
+	mean := mlx.Mean(x, -1, true)
+	centered := mlx.Sub(x, mean)
+	variance := mlx.Mean(mlx.Mul(centered, centered), -1, true)
+	normalized := mlx.Mul(centered, mlx.RSqrt(mlx.AddScalar(variance, eps)))
+
+	// Scale and shift
+	out := mlx.Mul(normalized, ln.Weight)
+	if ln.Bias != nil {
+		out = mlx.Add(out, ln.Bias)
+	}
+	return out
+}

x/imagegen/nn/nn_test.go

+//go:build mlx
+
+package nn
+
+import (
+	"math"
+	"testing"
+
+	"github.com/ollama/ollama/x/imagegen/mlx"
+)
+
+// TestLinearNoBias verifies Linear without bias computes x @ w.T correctly.
+func TestLinearNoBias(t *testing.T) {
+	// Weight: [out=2, in=3] -> transposed at forward time
+	weight := mlx.NewArrayFloat32([]float32{
+		1, 2, 3, // row 0
+		4, 5, 6, // row 1
+	}, []int32{2, 3})
+	mlx.Eval(weight)
+
+	linear := NewLinear(weight, nil)
+
+	// Input: [1, 3]
+	x := mlx.NewArrayFloat32([]float32{1, 1, 1}, []int32{1, 3})
+	mlx.Eval(x)
+
+	out := linear.Forward(x)
+	mlx.Eval(out)
+
+	// Expected: [1,1,1] @ [[1,4],[2,5],[3,6]] = [6, 15]
+	data := out.Data()
+	if len(data) != 2 || data[0] != 6 || data[1] != 15 {
+		t.Errorf("expected [6, 15], got %v", data)
+	}
+}
+
+// TestLinearWithBias verifies Linear with bias computes x @ w.T + b correctly.
+func TestLinearWithBias(t *testing.T) {
+	weight := mlx.NewArrayFloat32([]float32{
+		1, 2, 3,
+		4, 5, 6,
+	}, []int32{2, 3})
+	bias := mlx.NewArrayFloat32([]float32{10, 20}, []int32{2})
+	mlx.Eval(weight, bias)
+
+	linear := NewLinear(weight, bias)
+
+	x := mlx.NewArrayFloat32([]float32{1, 1, 1}, []int32{1, 3})
+	mlx.Eval(x)
+
+	out := linear.Forward(x)
+	mlx.Eval(out)
+
+	// Expected: [6, 15] + [10, 20] = [16, 35]
+	data := out.Data()
+	if len(data) != 2 || data[0] != 16 || data[1] != 35 {
+		t.Errorf("expected [16, 35], got %v", data)
+	}
+}
+
+// TestLinearBatched verifies Linear works with batched input.
+func TestLinearBatched(t *testing.T) {
+	weight := mlx.NewArrayFloat32([]float32{
+		1, 0,
+		0, 1,
+	}, []int32{2, 2}) // Identity
+	mlx.Eval(weight)
+
+	linear := NewLinear(weight, nil)
+
+	// Batch of 3 inputs
+	x := mlx.NewArrayFloat32([]float32{
+		1, 2,
+		3, 4,
+		5, 6,
+	}, []int32{3, 2})
+	mlx.Eval(x)
+
+	out := linear.Forward(x)
+	mlx.Eval(out)
+
+	// Identity should return same values
+	data := out.Data()
+	expected := []float32{1, 2, 3, 4, 5, 6}
+	for i, v := range expected {
+		if data[i] != v {
+			t.Errorf("at %d: expected %f, got %f", i, v, data[i])
+		}
+	}
+}
+
+// TestRMSNorm verifies RMSNorm computation.
+func TestRMSNorm(t *testing.T) {
+	weight := mlx.NewArrayFloat32([]float32{1, 1, 1, 1}, []int32{4})
+	mlx.Eval(weight)
+
+	norm := NewRMSNorm(weight, 1e-5)
+
+	// Input with known RMS
+	x := mlx.NewArrayFloat32([]float32{2, 2, 2, 2}, []int32{1, 4})
+	mlx.Eval(x)
+
+	out := norm.Forward(x, 0) // eps=0 uses stored Eps
+	mlx.Eval(out)
+
+	// RMS of [2,2,2,2] = 2, so normalized = [1,1,1,1]
+	data := out.Data()
+	for i, v := range data {
+		if math.Abs(float64(v-1.0)) > 1e-4 {
+			t.Errorf("at %d: expected ~1.0, got %f", i, v)
+		}
+	}
+}
+
+// TestRMSNormWithScale verifies RMSNorm applies weight scaling.
+func TestRMSNormWithScale(t *testing.T) {
+	weight := mlx.NewArrayFloat32([]float32{2, 2, 2, 2}, []int32{4})
+	mlx.Eval(weight)
+
+	norm := NewRMSNorm(weight, 1e-5)
+
+	x := mlx.NewArrayFloat32([]float32{2, 2, 2, 2}, []int32{1, 4})
+	mlx.Eval(x)
+
+	out := norm.Forward(x, 0) // eps=0 uses stored Eps
+	mlx.Eval(out)
+
+	// Normalized [1,1,1,1] * weight [2,2,2,2] = [2,2,2,2]
+	data := out.Data()
+	for i, v := range data {
+		if math.Abs(float64(v-2.0)) > 1e-4 {
+			t.Errorf("at %d: expected ~2.0, got %f", i, v)
+		}
+	}
+}
+
+// TestEmbedding verifies embedding lookup.
+func TestEmbedding(t *testing.T) {
+	// Embedding table: 4 tokens, dim 3
+	weight := mlx.NewArrayFloat32([]float32{
+		0, 0, 0, // token 0
+		1, 1, 1, // token 1
+		2, 2, 2, // token 2
+		3, 3, 3, // token 3
+	}, []int32{4, 3})
+	mlx.Eval(weight)
+
+	emb := NewEmbedding(weight)
+
+	// Look up tokens [1, 3, 0]
+	indices := mlx.NewArrayInt32([]int32{1, 3, 0}, []int32{3})
+	mlx.Eval(indices)
+
+	out := emb.Forward(indices)
+	mlx.Eval(out)
+
+	data := out.Data()
+	expected := []float32{1, 1, 1, 3, 3, 3, 0, 0, 0}
+	for i, v := range expected {
+		if data[i] != v {
+			t.Errorf("at %d: expected %f, got %f", i, v, data[i])
+		}
+	}
+}
+
+// TestRepeatKV verifies K/V repetition for GQA.
+func TestRepeatKV(t *testing.T) {
+	// [B=1, num_kv_heads=2, S=2, head_dim=2]
+	x := mlx.NewArrayFloat32([]float32{
+		// head 0
+		1, 2, // pos 0
+		3, 4, // pos 1
+		// head 1
+		5, 6, // pos 0
+		7, 8, // pos 1
+	}, []int32{1, 2, 2, 2})
+	mlx.Eval(x)
+
+	// Repeat factor 2: 2 kv heads -> 4 heads
+	out := RepeatKV(x, 2)
+	mlx.Eval(out)
+
+	shape := out.Shape()
+	if shape[0] != 1 || shape[1] != 4 || shape[2] != 2 || shape[3] != 2 {
+		t.Errorf("expected shape [1,4,2,2], got %v", shape)
+	}
+
+	data := out.Data()
+	// After repeat: head0, head0, head1, head1
+	expected := []float32{
+		1, 2, 3, 4, // head 0 (original)
+		1, 2, 3, 4, // head 0 (repeat)
+		5, 6, 7, 8, // head 1 (original)
+		5, 6, 7, 8, // head 1 (repeat)
+	}
+	for i, v := range expected {
+		if data[i] != v {
+			t.Errorf("at %d: expected %f, got %f", i, v, data[i])
+		}
+	}
+}
+
+// TestRepeatKVNoOp verifies RepeatKV with factor 1 returns input unchanged.
+func TestRepeatKVNoOp(t *testing.T) {
+	x := mlx.NewArrayFloat32([]float32{1, 2, 3, 4}, []int32{1, 1, 2, 2})
+	mlx.Eval(x)
+
+	out := RepeatKV(x, 1)
+	// Should return same pointer
+	if out != x {
+		t.Error("RepeatKV with factor 1 should return input unchanged")
+	}
+}
+
+// TestApplyCausalMask verifies causal masking.
+func TestApplyCausalMask(t *testing.T) {
+	// [B=1, heads=1, S=3, S=3] - all ones
+	scores := mlx.Ones(1, 1, 3, 3)
+	mlx.Eval(scores)
+
+	out := ApplyCausalMask(scores)
+	mlx.Eval(out)
+
+	data := out.Data()
+	// Lower triangular should be 1, upper should be -1e9
+	// Row 0: [1, -inf, -inf]
+	// Row 1: [1, 1, -inf]
+	// Row 2: [1, 1, 1]
+	if data[0] != 1 || data[1] >= 0 || data[2] >= 0 {
+		t.Errorf("row 0 wrong: %v", data[0:3])
+	}
+	if data[3] != 1 || data[4] != 1 || data[5] >= 0 {
+		t.Errorf("row 1 wrong: %v", data[3:6])
+	}
+	if data[6] != 1 || data[7] != 1 || data[8] != 1 {
+		t.Errorf("row 2 wrong: %v", data[6:9])
+	}
+}
+
+// TestApplyCausalMaskWithOffset verifies causal masking with cache offset.
+func TestApplyCausalMaskWithOffset(t *testing.T) {
+	// Simulating: cache has 2 tokens, adding 1 new query
+	// scores: [B=1, heads=1, queryLen=1, keyLen=3]
+	scores := mlx.Ones(1, 1, 1, 3)
+	mlx.Eval(scores)
+
+	out := ApplyCausalMaskWithOffset(scores, 2)
+	mlx.Eval(out)
+
+	data := out.Data()
+	// With offset=2, query at position 2 can attend to all 3 positions
+	if data[0] != 1 || data[1] != 1 || data[2] != 1 {
+		t.Errorf("expected [1, 1, 1], got %v", data)
+	}
+}
+
+// TestApplyCausalMaskWithOffsetZero verifies offset=0 falls back to regular causal.
+func TestApplyCausalMaskWithOffsetZero(t *testing.T) {
+	scores := mlx.Ones(1, 1, 2, 2)
+	mlx.Eval(scores)
+
+	out := ApplyCausalMaskWithOffset(scores, 0)
+	mlx.Eval(out)
+
+	data := out.Data()
+	// Standard causal: [1, -inf], [1, 1]
+	if data[0] != 1 || data[1] >= 0 {
+		t.Errorf("row 0 wrong: %v", data[0:2])
+	}
+	if data[2] != 1 || data[3] != 1 {
+		t.Errorf("row 1 wrong: %v", data[2:4])
+	}
+}
+
+// BenchmarkLinearSmall benchmarks small Linear forward pass.
+func BenchmarkLinearSmall(b *testing.B) {
+	weight := mlx.RandomNormal([]int32{256, 256}, 42)
+	mlx.Eval(weight)
+
+	linear := NewLinear(weight, nil)
+
+	x := mlx.RandomNormal([]int32{1, 256}, 43)
+	mlx.Eval(x)
+
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		out := linear.Forward(x)
+		mlx.Eval(out)
+	}
+}
+
+// BenchmarkLinearLarge benchmarks larger Linear forward pass.
+func BenchmarkLinearLarge(b *testing.B) {
+	weight := mlx.RandomNormal([]int32{4096, 4096}, 42)
+	mlx.Eval(weight)
+
+	linear := NewLinear(weight, nil)
+
+	x := mlx.RandomNormal([]int32{1, 4096}, 43)
+	mlx.Eval(x)
+
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		out := linear.Forward(x)
+		mlx.Eval(out)
+	}
+}
+
+// BenchmarkRMSNorm benchmarks RMSNorm forward pass.
+func BenchmarkRMSNorm(b *testing.B) {
+	weight := mlx.Ones(4096)
+	mlx.Eval(weight)
+
+	norm := NewRMSNorm(weight, 1e-5)
+
+	x := mlx.RandomNormal([]int32{1, 4096}, 42)
+	mlx.Eval(x)
+
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		out := norm.Forward(x, 0)
+		mlx.Eval(out)
+	}
+}
+
+// BenchmarkEmbedding benchmarks embedding lookup.
+func BenchmarkEmbedding(b *testing.B) {
+	// Typical vocab size
+	weight := mlx.RandomNormal([]int32{32000, 4096}, 42)
+	mlx.Eval(weight)
+
+	emb := NewEmbedding(weight)
+
+	// Single token lookup
+	indices := mlx.NewArrayInt32([]int32{1000}, []int32{1})
+	mlx.Eval(indices)
+
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		out := emb.Forward(indices)
+		mlx.Eval(out)
+	}
+}
+
+// BenchmarkRepeatKV benchmarks K/V repetition.
+func BenchmarkRepeatKV(b *testing.B) {
+	// Typical GQA setup: 8 kv heads -> 32 heads
+	x := mlx.RandomNormal([]int32{1, 8, 512, 128}, 42)
+	mlx.Eval(x)
+
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		out := RepeatKV(x, 4)
+		mlx.Eval(out)
+	}
+}

x/imagegen/safetensors/loader.go

+//go:build mlx
+
+package safetensors
+
+import (
+	"fmt"
+	"reflect"
+	"strings"
+
+	"github.com/ollama/ollama/x/imagegen/mlx"
+)
+
+// LoadModule loads weights into a struct using reflection and struct tags.
+//
+// Struct tags use the format: `weight:"path[,optional]"`
+//   - path: the weight name suffix (appended to prefix)
+//   - optional: if present, missing weights don't cause errors
+//   - "-": skip this field entirely
+//   - no tag on struct pointer: recurse with current prefix
+//   - no tag on *mlx.Array: skip (computed fields don't need loading)
+//
+// For slices of struct pointers, the loader iterates with .0, .1, .2... suffixes.
+// The slice must be pre-allocated to the correct length.
+//
+// Example:
+//
+//	type Attention struct {
+//	    QProj *nn.Linear  `weight:"self_attn.q_proj"`
+//	    KProj *nn.Linear  `weight:"self_attn.k_proj"`
+//	    Cache *mlx.Array  // no tag = skipped (computed field)
+//	}
+//
+//	err := LoadModule(&attn, weights, "model.layers.0")
+func LoadModule(dst any, weights *ModelWeights, prefix string) error {
+	v := reflect.ValueOf(dst)
+	if v.Kind() != reflect.Ptr || v.IsNil() {
+		return fmt.Errorf("LoadModule: dst must be a non-nil pointer")
+	}
+	v = v.Elem()
+	if v.Kind() != reflect.Struct {
+		return fmt.Errorf("LoadModule: dst must be a pointer to struct, got %v", v.Kind())
+	}
+
+	var errs []string
+	loadStruct(v, weights, prefix, &errs, false)
+
+	if len(errs) > 0 {
+		return fmt.Errorf("LoadModule: missing weights:\n  %s", strings.Join(errs, "\n  "))
+	}
+	return nil
+}
+
+// loadStruct recursively loads weights into a struct value.
+func loadStruct(v reflect.Value, weights *ModelWeights, prefix string, errs *[]string, parentOptional bool) {
+	t := v.Type()
+
+	for i := 0; i < t.NumField(); i++ {
+		field := t.Field(i)
+		fieldVal := v.Field(i)
+
+		// Skip unexported fields
+		if !fieldVal.CanSet() {
+			continue
+		}
+
+		// Parse tag
+		tag, hasTag := field.Tag.Lookup("weight")
+		if tag == "-" {
+			continue
+		}
+
+		// Parse tag options
+		optional := parentOptional
+		weightPath := tag
+		if idx := strings.Index(tag, ","); idx != -1 {
+			weightPath = tag[:idx]
+			if strings.Contains(tag[idx+1:], "optional") {
+				optional = true
+			}
+		}
+
+		// Build full path
+		fullPath := joinPath(prefix, weightPath)
+
+		// For struct pointers without a tag, recurse with current prefix
+		if !hasTag && fieldVal.Kind() == reflect.Ptr {
+			elemType := fieldVal.Type().Elem()
+			if elemType.Kind() == reflect.Struct && elemType != reflect.TypeOf(mlx.Array{}) {
+				if fieldVal.IsNil() {
+					fieldVal.Set(reflect.New(elemType))
+				}
+				loadStruct(fieldVal.Elem(), weights, prefix, errs, optional)
+				continue
+			}
+		}
+
+		// Handle by kind
+		switch fieldVal.Kind() {
+		case reflect.Ptr:
+			elemType := fieldVal.Type().Elem()
+
+			// *mlx.Array - load directly (but skip if no tag - computed fields)
+			if fieldVal.Type() == reflect.TypeOf((*mlx.Array)(nil)) {
+				if !hasTag {
+					continue // no tag on *mlx.Array = computed field, skip
+				}
+				arr, err := weights.GetTensor(fullPath)
+				if err != nil {
+					if !optional {
+						*errs = append(*errs, fullPath)
+					}
+					continue
+				}
+				fieldVal.Set(reflect.ValueOf(arr))
+				continue
+			}
+
+			// Pointer to struct - allocate and recurse
+			if elemType.Kind() == reflect.Struct {
+				if optional && !hasWeightsWithPrefix(weights, fullPath) {
+					continue
+				}
+				if fieldVal.IsNil() {
+					fieldVal.Set(reflect.New(elemType))
+				}
+				loadStruct(fieldVal.Elem(), weights, fullPath, errs, optional)
+			}
+
+		case reflect.Slice:
+			elemType := fieldVal.Type().Elem()
+			if elemType.Kind() == reflect.Ptr && elemType.Elem().Kind() == reflect.Struct {
+				loadSlice(fieldVal, weights, fullPath, errs)
+			}
+		}
+	}
+}
+
+// hasWeightsWithPrefix checks if any weights exist with the given prefix.
+func hasWeightsWithPrefix(weights *ModelWeights, prefix string) bool {
+	for _, name := range weights.ListTensors() {
+		if strings.HasPrefix(name, prefix+".") || name == prefix {
+			return true
+		}
+	}
+	return false
+}
+
+// loadSlice loads weights into each element of a slice of struct pointers.
+func loadSlice(v reflect.Value, weights *ModelWeights, prefix string, errs *[]string) {
+	elemStructType := v.Type().Elem().Elem()
+
+	for i := 0; i < v.Len(); i++ {
+		elem := v.Index(i)
+		if elem.IsNil() {
+			elem.Set(reflect.New(elemStructType))
+		}
+		loadStruct(elem.Elem(), weights, fmt.Sprintf("%s.%d", prefix, i), errs, false)
+	}
+}
+
+// joinPath joins path segments with dots, handling empty segments.
+func joinPath(prefix, suffix string) string {
+	if prefix == "" {
+		return suffix
+	}
+	if suffix == "" {
+		return prefix
+	}
+	return prefix + "." + suffix
+}

x/imagegen/safetensors/safetensors.go

+//go:build mlx
+
+package safetensors
+
+import (
+	"encoding/binary"
+	"encoding/json"
+	"fmt"
+	"os"
+	"path/filepath"
+	"sort"
+	"strings"
+
+	"github.com/ollama/ollama/x/imagegen/mlx"
+)
+
+// SafetensorHeader represents the JSON header of a safetensors file
+type SafetensorHeader map[string]TensorInfo
+
+// TensorInfo contains metadata about a tensor
+type TensorInfo struct {
+	Dtype       string  `json:"dtype"`
+	Shape       []int32 `json:"shape"`
+	DataOffsets [2]int  `json:"data_offsets"`
+}
+
+// parseSafetensorHeader reads only the JSON header from a safetensors file.
+func parseSafetensorHeader(path string) (SafetensorHeader, error) {
+	f, err := os.Open(path)
+	if err != nil {
+		return nil, fmt.Errorf("failed to open file: %w", err)
+	}
+	defer f.Close()
+
+	var headerSize uint64
+	if err := binary.Read(f, binary.LittleEndian, &headerSize); err != nil {
+		return nil, fmt.Errorf("failed to read header size: %w", err)
+	}
+
+	headerBytes := make([]byte, headerSize)
+	if _, err := f.Read(headerBytes); err != nil {
+		return nil, fmt.Errorf("failed to read header: %w", err)
+	}
+
+	var header SafetensorHeader
+	if err := json.Unmarshal(headerBytes, &header); err != nil {
+		return nil, fmt.Errorf("failed to parse header: %w", err)
+	}
+
+	delete(header, "__metadata__")
+	return header, nil
+}
+
+// dtypeFromString converts safetensors dtype string to mlx.Dtype
+func dtypeFromString(s string) mlx.Dtype {
+	switch strings.ToUpper(s) {
+	case "F32", "FLOAT32":
+		return mlx.DtypeFloat32
+	case "F16", "FLOAT16":
+		return mlx.DtypeFloat16
+	case "BF16", "BFLOAT16":
+		return mlx.DtypeBFloat16
+	case "I32", "INT32":
+		return mlx.DtypeInt32
+	case "I64", "INT64":
+		return mlx.DtypeInt64
+	case "U8", "UINT8":
+		return mlx.DtypeUint8
+	default:
+		return mlx.DtypeFloat32
+	}
+}
+
+// ModelWeights manages weights from multiple safetensor files.
+type ModelWeights struct {
+	dir         string                          // Model directory
+	tensorFiles map[string]string               // tensor name -> file path
+	tensorInfo  map[string]TensorInfo           // tensor name -> metadata
+	nativeCache map[string]*mlx.SafetensorsFile // file path -> loaded native handle
+	cache       map[string]*mlx.Array           // tensor name -> array (after Load)
+}
+
+// LoadModelWeights scans safetensor files and builds a tensor index.
+// This only reads JSON headers, not tensor data.
+func LoadModelWeights(dir string) (*ModelWeights, error) {
+	mw := &ModelWeights{
+		dir:         dir,
+		tensorFiles: make(map[string]string),
+		tensorInfo:  make(map[string]TensorInfo),
+		nativeCache: make(map[string]*mlx.SafetensorsFile),
+	}
+
+	entries, err := os.ReadDir(dir)
+	if err != nil {
+		return nil, fmt.Errorf("failed to read directory: %w", err)
+	}
+
+	for _, entry := range entries {
+		if strings.HasSuffix(entry.Name(), ".safetensors") {
+			path := filepath.Join(dir, entry.Name())
+
+			header, err := parseSafetensorHeader(path)
+			if err != nil {
+				return nil, fmt.Errorf("failed to parse %s: %w", entry.Name(), err)
+			}
+
+			for name, info := range header {
+				mw.tensorFiles[name] = path
+				mw.tensorInfo[name] = info
+			}
+		}
+	}
+
+	if len(mw.tensorFiles) == 0 {
+		return nil, fmt.Errorf("no safetensor files found in %s", dir)
+	}
+
+	return mw, nil
+}
+
+// Load loads all tensors into cache with the specified dtype.
+// If dtype is 0, tensors are loaded in their original dtype.
+// Automatically uses streaming (memory-efficient) when dtype conversion is needed,
+// or native loading when tensors are already in the target dtype.
+func (mw *ModelWeights) Load(dtype mlx.Dtype) error {
+	if dtype == 0 {
+		return mw.loadNative()
+	}
+
+	// Check if any tensor needs conversion
+	needsConversion := false
+	for name := range mw.tensorFiles {
+		info := mw.tensorInfo[name]
+		if dtypeFromString(info.Dtype) != dtype {
+			needsConversion = true
+			break
+		}
+	}
+
+	if needsConversion {
+		return mw.loadStreaming(dtype)
+	}
+	return mw.loadNative()
+}
+
+// loadNative loads all tensors using the native memory-mapped loader.
+func (mw *ModelWeights) loadNative() error {
+	mw.cache = make(map[string]*mlx.Array)
+
+	fileToTensors := make(map[string][]string)
+	for name, path := range mw.tensorFiles {
+		fileToTensors[path] = append(fileToTensors[path], name)
+	}
+
+	for path, names := range fileToTensors {
+		native, err := mlx.LoadSafetensorsNative(path)
+		if err != nil {
+			return fmt.Errorf("failed to load %s: %w", path, err)
+		}
+
+		for _, name := range names {
+			arr := native.Get(name)
+			if arr == nil {
+				native.Free()
+				return fmt.Errorf("tensor %q not found in %s", name, path)
+			}
+			mw.cache[name] = arr
+		}
+
+		mw.nativeCache[path] = native
+	}
+
+	return nil
+}
+
+// loadStreaming loads tensors with dtype conversion.
+// Uses the same pattern as Python: replace each entry in the map after conversion,
+// so the original tensor loses its reference and can be freed.
+func (mw *ModelWeights) loadStreaming(dtype mlx.Dtype) error {
+	mw.cache = make(map[string]*mlx.Array)
+
+	fileToTensors := make(map[string][]string)
+	for name, path := range mw.tensorFiles {
+		fileToTensors[path] = append(fileToTensors[path], name)
+	}
+
+	for path, names := range fileToTensors {
+		native, err := mlx.LoadSafetensorsNative(path)
+		if err != nil {
+			return fmt.Errorf("failed to load %s: %w", path, err)
+		}
+
+		for _, name := range names {
+			src := native.Get(name)
+			if src == nil {
+				native.Free()
+				return fmt.Errorf("tensor %q not found in %s", name, path)
+			}
+
+			dst := mlx.AsType(src, dtype)
+			mlx.Eval(dst)
+			native.Set(name, dst)
+			mw.cache[name] = dst
+		}
+
+		native.Free()
+	}
+
+	return nil
+}
+
+// Get returns a tensor from cache. Call Load() first.
+func (mw *ModelWeights) Get(name string) (*mlx.Array, error) {
+	if mw.cache == nil {
+		return nil, fmt.Errorf("cache not initialized: call Load() first")
+	}
+	arr, ok := mw.cache[name]
+	if !ok {
+		return nil, fmt.Errorf("tensor %q not found in cache", name)
+	}
+	return arr, nil
+}
+
+// GetTensor loads a tensor using the native loader without caching.
+// For bulk loading, use Load() + Get() instead.
+func (mw *ModelWeights) GetTensor(name string) (*mlx.Array, error) {
+	if mw.cache != nil {
+		if arr, ok := mw.cache[name]; ok {
+			return arr, nil
+		}
+	}
+
+	path, ok := mw.tensorFiles[name]
+	if !ok {
+		return nil, fmt.Errorf("tensor %q not found", name)
+	}
+
+	native, ok := mw.nativeCache[path]
+	if !ok {
+		var err error
+		native, err = mlx.LoadSafetensorsNative(path)
+		if err != nil {
+			return nil, fmt.Errorf("failed to load %s: %w", path, err)
+		}
+		mw.nativeCache[path] = native
+	}
+
+	return native.Get(name), nil
+}
+
+// GetTensorInfo returns metadata about a tensor without loading it.
+func (mw *ModelWeights) GetTensorInfo(name string) (TensorInfo, bool) {
+	info, ok := mw.tensorInfo[name]
+	return info, ok
+}
+
+// ListTensors returns all tensor names.
+func (mw *ModelWeights) ListTensors() []string {
+	names := make([]string, 0, len(mw.tensorFiles))
+	for name := range mw.tensorFiles {
+		names = append(names, name)
+	}
+	sort.Strings(names)
+	return names
+}
+
+// HasTensor checks if a tensor exists.
+func (mw *ModelWeights) HasTensor(name string) bool {
+	_, ok := mw.tensorFiles[name]
+	return ok
+}
+
+// ReleaseAll releases all cached native file handles.
+func (mw *ModelWeights) ReleaseAll() {
+	for path, native := range mw.nativeCache {
+		native.Free()
+		delete(mw.nativeCache, path)
+	}
+}
+