server: chunk quantization writes to reduce create memory usage

Only the last token's processing time is included in prompt processing,
giving an artificially high rate. In addition, the number of tokens
only included the tokens that miss the cache, instead of our historic
total tokens.

api/types.go

@@ -15,6 +15,7 @@ import (
 	"github.com/google/uuid"
 
 	"github.com/ollama/ollama/envconfig"
+	"github.com/ollama/ollama/format"
 	"github.com/ollama/ollama/internal/orderedmap"
 	"github.com/ollama/ollama/types/model"
 )
@@ -569,6 +570,7 @@ type DebugInfo struct {
 
 type Metrics struct {
 	TotalDuration      time.Duration `json:"total_duration,omitempty"`
+	PeakMemory         uint64        `json:"peak_memory,omitempty"`
 	LoadDuration       time.Duration `json:"load_duration,omitempty"`
 	PromptEvalCount    int           `json:"prompt_eval_count,omitempty"`
 	PromptEvalDuration time.Duration `json:"prompt_eval_duration,omitempty"`
@@ -934,6 +936,10 @@ func (m *Metrics) Summary() {
 		fmt.Fprintf(os.Stderr, "total duration:       %v\n", m.TotalDuration)
 	}
 
+	if m.PeakMemory > 0 {
+		fmt.Fprintf(os.Stderr, "peak memory:          %s\n", formatPeakMemory(m.PeakMemory))
+	}
+
 	if m.LoadDuration > 0 {
 		fmt.Fprintf(os.Stderr, "load duration:        %v\n", m.LoadDuration)
 	}
@@ -957,6 +963,14 @@ func (m *Metrics) Summary() {
 	}
 }
 
+func formatPeakMemory(b uint64) string {
+	if b >= format.GibiByte {
+		return fmt.Sprintf("%.3f GiB", float64(b)/float64(format.GibiByte))
+	}
+
+	return format.HumanBytes2(b)
+}
+
 func (opts *Options) FromMap(m map[string]any) error {
 	valueOpts := reflect.ValueOf(opts).Elem() // names of the fields in the options struct
 	typeOpts := reflect.TypeOf(opts).Elem()   // types of the fields in the options struct

llm/server.go

@@ -74,8 +74,7 @@ type LlamaServer interface {
 	Tokenize(ctx context.Context, content string) ([]int, error)
 	Detokenize(ctx context.Context, tokens []int) (string, error)
 	Close() error
-	VRAMSize() uint64 // Total VRAM across all GPUs
-	TotalSize() uint64
+	MemorySize() (total, vram uint64)
 	VRAMByGPU(id ml.DeviceID) uint64
 	Pid() int
 	GetPort() int
@@ -685,8 +684,9 @@ func (s *llamaServer) Load(ctx context.Context, systemInfo ml.SystemInfo, system
 	// Windows CUDA should not use mmap for best performance
 	// Linux  with a model larger than free space, mmap leads to thrashing
 	// For CPU loads we want the memory to be allocated, not FS cache
+	totalSize, _ := s.MemorySize()
 	if (runtime.GOOS == "windows" && len(gpus) > 0 && gpus[0].Library == "CUDA" && s.options.UseMMap == nil) ||
-		(runtime.GOOS == "linux" && systemInfo.FreeMemory < s.TotalSize() && s.options.UseMMap == nil) ||
+		(runtime.GOOS == "linux" && systemInfo.FreeMemory < totalSize && s.options.UseMMap == nil) ||
 		(len(gpus) == 0 && s.options.UseMMap == nil) ||
 		(len(gpus) > 0 && gpus[0].Library == "Vulkan" && s.options.UseMMap == nil) ||
 		(s.options.UseMMap != nil && !*s.options.UseMMap) {
@@ -1518,6 +1518,7 @@ type CompletionResponse struct {
 	PromptEvalDuration time.Duration `json:"prompt_eval_duration"`
 	EvalCount          int           `json:"eval_count"`
 	EvalDuration       time.Duration `json:"eval_duration"`
+	PeakMemory         uint64        `json:"peak_memory,omitempty"`
 
 	// Logprobs contains log probability information if requested
 	Logprobs []Logprob `json:"logprobs,omitempty"`
@@ -1848,17 +1849,17 @@ func (s *llamaServer) GetDeviceInfos(ctx context.Context) []ml.DeviceInfo {
 	return nil
 }
 
-func (s *llmServer) VRAMSize() uint64 {
+func (s *llmServer) MemorySize() (total, vram uint64) {
 	if s.mem == nil {
-		return 0
+		return 0, 0
 	}
 
-	var mem uint64
-
 	for _, g := range s.mem.GPUs {
-		mem += g.Size()
+		vram += g.Size()
 	}
 
+	total = s.mem.InputWeights + s.mem.CPU.Size() + vram
+
 	// Some elements are always on CPU. However, if we have allocated all layers
 	// on the GPU then include the CPU components as well, to represent complete offloading.
 	noCPULayers := true
@@ -1869,25 +1870,11 @@ func (s *llmServer) VRAMSize() uint64 {
 		}
 	}
 	if noCPULayers {
-		mem += s.mem.InputWeights
-		mem += s.mem.CPU.Graph
+		vram += s.mem.InputWeights
+		vram += s.mem.CPU.Graph
 	}
 
-	return mem
-}
-
-func (s *llmServer) TotalSize() uint64 {
-	if s.mem == nil {
-		return 0
-	}
-
-	mem := s.mem.InputWeights
-	mem += s.mem.CPU.Size()
-	for _, g := range s.mem.GPUs {
-		mem += g.Size()
-	}
-
-	return mem
+	return total, vram
 }
 
 func (s *llmServer) VRAMByGPU(id ml.DeviceID) uint64 {

model/models/qwen3next/deltanet.go

@@ -41,8 +41,8 @@ type GatedDeltaNet struct {
 	SSMBeta      *nn.Linear  `gguf:"ssm_beta"`  // -> beta (qwen35)
 	SSMAlpha     *nn.Linear  `gguf:"ssm_alpha"` // -> alpha (qwen35)
 	SSMConv1D    *convKernel `gguf:"ssm_conv1d"`
-	SSMDT        ml.Tensor   `gguf:"ssm_dt"` // alpha bias
-	SSMA         ml.Tensor   `gguf:"ssm_a"`  // -A_log.exp()
+	SSMDT        ml.Tensor   `gguf:"ssm_dt,alt:ssm_dt.bias"` // alpha bias
+	SSMA         ml.Tensor   `gguf:"ssm_a"`                  // -A_log.exp()
 	SSMNorm      *nn.RMSNorm `gguf:"ssm_norm"`
 	SSMOut       *nn.Linear  `gguf:"ssm_out"`
 
@@ -135,6 +135,18 @@ func (gdn *GatedDeltaNet) Forward(ctx ml.Context, hiddenStates, _ ml.Tensor, cac
 	default:
 		return nil, errors.New("qwen3next: missing linear attention beta/alpha projections")
 	}
+	if gdn.SSMDT == nil {
+		return nil, errors.New("qwen3next: missing linear attention ssm_dt tensor")
+	}
+	if gdn.SSMA == nil {
+		return nil, errors.New("qwen3next: missing linear attention ssm_a tensor")
+	}
+	if gdn.SSMConv1D == nil || gdn.SSMConv1D.Weight == nil {
+		return nil, errors.New("qwen3next: missing linear attention ssm_conv1d tensor")
+	}
+	if gdn.SSMNorm == nil || gdn.SSMOut == nil {
+		return nil, errors.New("qwen3next: missing linear attention ssm_norm/ssm_out projections")
+	}
 
 	// Compute gate: softplus(alpha + dt_bias) * -A
 	alphaBiased := alpha.Add(ctx, gdn.SSMDT)

model/models/qwen3next/model.go

@@ -437,6 +437,46 @@ func (m *Model) Forward(ctx ml.Context, batch input.Batch) (ml.Tensor, error) {
 	return m.Output.Forward(ctx, hiddenStates), nil
 }
 
+func (m *Model) Validate() error {
+	if m.Options == nil {
+		return fmt.Errorf("qwen3next: missing model options")
+	}
+	if len(m.Layers) != len(m.Options.isRecurrent) {
+		return fmt.Errorf("qwen3next: layer config mismatch: have %d layers, %d recurrent flags", len(m.Layers), len(m.Options.isRecurrent))
+	}
+
+	for i, layer := range m.Layers {
+		if !m.Options.isRecurrent[i] {
+			continue
+		}
+
+		gdn, ok := layer.Operator.(*GatedDeltaNet)
+		if !ok || gdn == nil {
+			return fmt.Errorf("qwen3next: layer %d expected recurrent operator", i)
+		}
+		if gdn.SSMQKV == nil || gdn.SSMQKVGate == nil {
+			return fmt.Errorf("qwen3next: layer %d missing attn_qkv/attn_gate projections", i)
+		}
+		if gdn.SSMBetaAlpha == nil && (gdn.SSMBeta == nil || gdn.SSMAlpha == nil) {
+			return fmt.Errorf("qwen3next: layer %d missing linear attention beta/alpha projections", i)
+		}
+		if gdn.SSMDT == nil {
+			return fmt.Errorf("qwen3next: layer %d missing ssm_dt tensor", i)
+		}
+		if gdn.SSMA == nil {
+			return fmt.Errorf("qwen3next: layer %d missing ssm_a tensor", i)
+		}
+		if gdn.SSMConv1D == nil || gdn.SSMConv1D.Weight == nil {
+			return fmt.Errorf("qwen3next: layer %d missing ssm_conv1d tensor", i)
+		}
+		if gdn.SSMNorm == nil || gdn.SSMOut == nil {
+			return fmt.Errorf("qwen3next: layer %d missing ssm_norm/ssm_out projections", i)
+		}
+	}
+
+	return nil
+}
+
 func (m *Model) Shift(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error) {
 	m.positionCache = nil
 	if len(m.mropeSections) > 0 {
@@ -450,6 +490,64 @@ var (
 	_ model.MultimodalProcessor = (*Model)(nil)
 )
 
+func defaultVHeadReordered(arch string) bool {
+	return arch == "qwen35" || arch == "qwen35moe"
+}
+
+func inferRecurrentLayers(headCountKV []uint64, numLayers int, fullAttentionInterval uint32) ([]bool, error) {
+	isRecurrent := make([]bool, numLayers)
+
+	hasZero := false
+	hasFull := false
+	for i := range numLayers {
+		if i >= len(headCountKV) {
+			continue
+		}
+
+		if headCountKV[i] == 0 {
+			isRecurrent[i] = true
+			hasZero = true
+		} else {
+			hasFull = true
+		}
+	}
+	if hasZero && hasFull {
+		return isRecurrent, nil
+	}
+	if !hasFull {
+		return nil, fmt.Errorf("qwen3next: attention.head_count_kv must include at least one non-zero value")
+	}
+
+	// Compatibility path: older imports store a scalar KV head count and omit
+	// per-layer recurrent flags. Derive the hybrid layout from the interval.
+	interval := int(fullAttentionInterval)
+	if interval == 0 {
+		interval = min(4, numLayers)
+	}
+	if interval <= 0 {
+		return nil, fmt.Errorf("qwen3next: invalid block_count (%d)", numLayers)
+	}
+	if interval > numLayers {
+		return nil, fmt.Errorf("qwen3next: full_attention_interval (%d) exceeds block_count (%d)", interval, numLayers)
+	}
+
+	hasZero = false
+	hasFull = false
+	for i := range numLayers {
+		isRecurrent[i] = (i+1)%interval != 0
+		if isRecurrent[i] {
+			hasZero = true
+		} else {
+			hasFull = true
+		}
+	}
+	if !hasZero || !hasFull {
+		return nil, fmt.Errorf("qwen3next: full_attention_interval (%d) does not produce a mixed recurrent/full layout", interval)
+	}
+
+	return isRecurrent, nil
+}
+
 func New(c fs.Config) (model.Model, error) {
 	numLayers := int(c.Uint("block_count"))
 	layers := make([]Layer, numLayers)
@@ -460,26 +558,14 @@ func New(c fs.Config) (model.Model, error) {
 		HeadCountKV() []uint64
 	}
 
-	var isRecurrent []bool
 	var headCountKV []uint64
 	if hc, ok := c.(headCounts); ok {
 		headCountKV = hc.HeadCountKV()
 	}
 
-	isRecurrent = make([]bool, numLayers)
-	hasZero := false
-	hasFull := false
-	for i := range numLayers {
-		// If KV head count is 0, it's a recurrent layer
-		if i < len(headCountKV) && headCountKV[i] == 0 {
-			isRecurrent[i] = true
-			hasZero = true
-		} else if i < len(headCountKV) && headCountKV[i] > 0 {
-			hasFull = true
-		}
-	}
-	if !hasZero || !hasFull {
-		return nil, fmt.Errorf("qwen3next: invalid attention.head_count_kv array; expected mix of zero and non-zero values")
+	isRecurrent, err := inferRecurrentLayers(headCountKV, numLayers, c.Uint("full_attention_interval"))
+	if err != nil {
+		return nil, err
 	}
 
 	// Determine if MoE
@@ -543,7 +629,7 @@ func New(c fs.Config) (model.Model, error) {
 		ssmNGroup:             int(c.Uint("ssm.group_count")),
 		ssmDtRank:             int(c.Uint("ssm.time_step_rank")),
 		convKernelSize:        int(c.Uint("ssm.conv_kernel")),
-		vHeadReordered:        c.Bool("ssm.v_head_reordered", false),
+		vHeadReordered:        c.Bool("ssm.v_head_reordered", defaultVHeadReordered(c.Architecture())),
 		isRecurrent:           isRecurrent,
 		mropeSections: slices.Collect(func(yield func(int) bool) {
 			for _, section := range mropeSections {
@@ -555,7 +641,7 @@ func New(c fs.Config) (model.Model, error) {
 		mropeInterleaved: c.Bool("rope.mrope_interleaved", c.Bool("mrope_interleaved", false)),
 	}
 	if opts.numKVHeads == 0 {
-		return nil, fmt.Errorf("qwen3next: attention.head_count_kv array must include at least one non-zero value")
+		return nil, fmt.Errorf("qwen3next: attention.head_count_kv must include at least one non-zero value")
 	}
 
 	// Calculate cache dimensions

model/models/qwen3next/model_new_test.go

@@ -0,0 +1,65 @@
+package qwen3next
+
+import (
+	"slices"
+	"strings"
+	"testing"
+)
+
+func TestInferRecurrentLayersMixedKVArray(t *testing.T) {
+	got, err := inferRecurrentLayers([]uint64{0, 2, 0, 2}, 4, 0)
+	if err != nil {
+		t.Fatalf("inferRecurrentLayers() error = %v", err)
+	}
+
+	want := []bool{true, false, true, false}
+	if !slices.Equal(got, want) {
+		t.Fatalf("inferRecurrentLayers() = %v, want %v", got, want)
+	}
+}
+
+func TestInferRecurrentLayersScalarKVDefaultInterval(t *testing.T) {
+	got, err := inferRecurrentLayers([]uint64{2, 2, 2, 2, 2, 2, 2, 2}, 8, 0)
+	if err != nil {
+		t.Fatalf("inferRecurrentLayers() error = %v", err)
+	}
+
+	want := []bool{true, true, true, false, true, true, true, false}
+	if !slices.Equal(got, want) {
+		t.Fatalf("inferRecurrentLayers() = %v, want %v", got, want)
+	}
+}
+
+func TestInferRecurrentLayersScalarKVConfiguredInterval(t *testing.T) {
+	got, err := inferRecurrentLayers([]uint64{2, 2, 2, 2, 2, 2}, 6, 3)
+	if err != nil {
+		t.Fatalf("inferRecurrentLayers() error = %v", err)
+	}
+
+	want := []bool{true, true, false, true, true, false}
+	if !slices.Equal(got, want) {
+		t.Fatalf("inferRecurrentLayers() = %v, want %v", got, want)
+	}
+}
+
+func TestInferRecurrentLayersAllZeroRejects(t *testing.T) {
+	_, err := inferRecurrentLayers([]uint64{0, 0, 0, 0}, 4, 0)
+	if err == nil {
+		t.Fatal("inferRecurrentLayers() expected error, got nil")
+	}
+	if !strings.Contains(err.Error(), "must include at least one non-zero value") {
+		t.Fatalf("unexpected error = %v", err)
+	}
+}
+
+func TestDefaultVHeadReordered(t *testing.T) {
+	if !defaultVHeadReordered("qwen35") {
+		t.Fatal("defaultVHeadReordered(qwen35) = false, want true")
+	}
+	if !defaultVHeadReordered("qwen35moe") {
+		t.Fatal("defaultVHeadReordered(qwen35moe) = false, want true")
+	}
+	if defaultVHeadReordered("qwen3next") {
+		t.Fatal("defaultVHeadReordered(qwen3next) = true, want false")
+	}
+}

model/models/qwen3next/model_validate_test.go

@@ -0,0 +1,45 @@
+package qwen3next
+
+import (
+	"strings"
+	"testing"
+
+	"github.com/ollama/ollama/ml/nn"
+)
+
+func TestValidateRecurrentLayerRequiresSSMDT(t *testing.T) {
+	m := &Model{
+		Layers: []Layer{{
+			Operator: &GatedDeltaNet{
+				SSMQKV:     &nn.Linear{},
+				SSMQKVGate: &nn.Linear{},
+				SSMBeta:    &nn.Linear{},
+				SSMAlpha:   &nn.Linear{},
+			},
+		}},
+		Options: &Options{
+			isRecurrent: []bool{true},
+		},
+	}
+
+	err := m.Validate()
+	if err == nil {
+		t.Fatal("Validate() expected error, got nil")
+	}
+	if !strings.Contains(err.Error(), "missing ssm_dt") {
+		t.Fatalf("unexpected error = %v", err)
+	}
+}
+
+func TestValidateNonRecurrentSkipsLinearChecks(t *testing.T) {
+	m := &Model{
+		Layers: []Layer{{Operator: &FullAttention{}}},
+		Options: &Options{
+			isRecurrent: []bool{false},
+		},
+	}
+
+	if err := m.Validate(); err != nil {
+		t.Fatalf("Validate() error = %v", err)
+	}
+}

server/images.go

@@ -71,6 +71,10 @@ type Model struct {
 	Template *template.Template
 }
 
+func (m *Model) IsMLX() bool {
+	return m.Config.ModelFormat == "safetensors"
+}
+
 // Capabilities returns the capabilities that the model supports
 func (m *Model) Capabilities() []model.Capability {
 	capabilities := []model.Capability{}

server/prompt.go

@@ -30,42 +30,44 @@ func chatPrompt(ctx context.Context, m *Model, tokenize tokenizeFunc, opts *api.
 	lastMsgIdx := len(msgs) - 1
 	currMsgIdx := 0
 
-	// Start with all messages and remove from the front until it fits in context
-	for i := 0; i <= lastMsgIdx; i++ {
-		// Collect system messages from the portion we're about to skip
-		system = make([]api.Message, 0)
-		for j := range i {
-			if msgs[j].Role == "system" {
-				system = append(system, msgs[j])
+	if truncate {
+		// Start with all messages and remove from the front until it fits in context
+		for i := 0; i <= lastMsgIdx; i++ {
+			// Collect system messages from the portion we're about to skip
+			system = make([]api.Message, 0)
+			for j := range i {
+				if msgs[j].Role == "system" {
+					system = append(system, msgs[j])
+				}
 			}
-		}
 
-		p, err := renderPrompt(m, append(system, msgs[i:]...), tools, think)
-		if err != nil {
-			return "", nil, err
-		}
-
-		s, err := tokenize(ctx, p)
-		if err != nil {
-			return "", nil, err
-		}
-
-		ctxLen := len(s)
-		if m.ProjectorPaths != nil {
-			for _, msg := range msgs[i:] {
-				ctxLen += imageNumTokens * len(msg.Images)
+			p, err := renderPrompt(m, append(system, msgs[i:]...), tools, think)
+			if err != nil {
+				return "", nil, err
 			}
-		}
 
-		if !truncate || ctxLen <= opts.NumCtx {
-			currMsgIdx = i
-			break
-		}
+			s, err := tokenize(ctx, p)
+			if err != nil {
+				return "", nil, err
+			}
 
-		// Must always include at least the last message
-		if i == lastMsgIdx {
-			currMsgIdx = lastMsgIdx
-			break
+			ctxLen := len(s)
+			if m.ProjectorPaths != nil {
+				for _, msg := range msgs[i:] {
+					ctxLen += imageNumTokens * len(msg.Images)
+				}
+			}
+
+			if ctxLen <= opts.NumCtx {
+				currMsgIdx = i
+				break
+			}
+
+			// Must always include at least the last message
+			if i == lastMsgIdx {
+				currMsgIdx = lastMsgIdx
+				break
+			}
 		}
 	}
 

server/quantization.go

@@ -21,33 +21,76 @@ type quantizer struct {
 	progressFn func(n uint64)
 }
 
+const quantizationChunkElements uint64 = 4 * 1024 * 1024
+
 func (q quantizer) WriteTo(w io.Writer) (int64, error) {
 	quantize := q.from.Kind != q.to.Kind
 	sr := io.NewSectionReader(q, int64(q.offset), int64(q.from.Size()))
 	if !quantize {
 		n, err := io.Copy(w, sr)
-		q.progressFn(q.from.Size())
+		if q.progressFn != nil {
+			q.progressFn(q.from.Size())
+		}
 		return n, err
 	}
-	data, err := io.ReadAll(sr)
-	if err != nil {
-		slog.Warn("file read error", "tensor", q.from.Name, "file", q.Name(), "error", err)
-		return 0, fmt.Errorf("unable to read tensor %s from %s: %s", q.from.Name, q.Name(), err)
+
+	if len(q.from.Shape) == 0 || q.from.Shape[0] == 0 {
+		return 0, fmt.Errorf("tensor %s has invalid shape %v", q.from.Name, q.from.Shape)
 	}
-	if uint64(len(data)) < q.from.Size() {
-		return 0, fmt.Errorf("tensor %s data size %d is less than expected %d from shape %v", q.from.Name, len(data), q.from.Size(), q.from.Shape)
+
+	fromType := fsggml.TensorType(q.from.Kind)
+	toType := fsggml.TensorType(q.to.Kind)
+	nPerRow := q.from.Shape[0]
+	totalElements := q.from.Elements()
+	if totalElements%nPerRow != 0 {
+		return 0, fmt.Errorf("tensor %s has non-row-aligned shape %v", q.from.Name, q.from.Shape)
 	}
-	var f32s []float32
-	newType := fsggml.TensorType(q.to.Kind)
-	if fsggml.TensorType(q.from.Kind) == fsggml.TensorTypeF32 {
-		f32s = unsafe.Slice((*float32)(unsafe.Pointer(&data[0])), q.from.Elements())
-	} else {
-		f32s = ggml.ConvertToF32(data, q.from.Kind, q.from.Elements())
+
+	inRowSize := fromType.RowSize(nPerRow)
+	if inRowSize == 0 {
+		return 0, fmt.Errorf("tensor %s has unsupported source type %v", q.from.Name, fromType)
 	}
-	data = ggml.Quantize(newType, f32s, q.from.Shape)
-	n, err := w.Write(data)
-	q.progressFn(q.from.Size())
-	return int64(n), err
+
+	totalRows := totalElements / nPerRow
+	rowsPerChunk := max(quantizationChunkElements/nPerRow, uint64(1))
+	chunkBuf := make([]byte, inRowSize*rowsPerChunk)
+	var written int64
+
+	for row := uint64(0); row < totalRows; {
+		chunkRows := min(rowsPerChunk, totalRows-row)
+		chunkBytes := inRowSize * chunkRows
+		data := chunkBuf[:chunkBytes]
+
+		if _, err := io.ReadFull(sr, data); err != nil {
+			slog.Warn("file read error", "tensor", q.from.Name, "file", q.Name(), "error", err)
+			return written, fmt.Errorf("unable to read tensor %s from %s: %w", q.from.Name, q.Name(), err)
+		}
+
+		var f32s []float32
+		chunkElements := chunkRows * nPerRow
+		if fromType == fsggml.TensorTypeF32 {
+			f32s = unsafe.Slice((*float32)(unsafe.Pointer(&data[0])), chunkElements)
+		} else {
+			f32s = ggml.ConvertToF32(data, q.from.Kind, chunkElements)
+		}
+
+		quantized := ggml.Quantize(toType, f32s, []uint64{nPerRow, chunkRows})
+		n, err := w.Write(quantized)
+		written += int64(n)
+		if err != nil {
+			return written, err
+		}
+		if n != len(quantized) {
+			return written, io.ErrShortWrite
+		}
+
+		if q.progressFn != nil {
+			q.progressFn(chunkBytes)
+		}
+		row += chunkRows
+	}
+
+	return written, nil
 }
 
 type quantizeState struct {

server/routes.go

@@ -484,7 +484,8 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 		// the real chat handler, but doing this as a stopgap to get renderer
 		// support for generate
 		if values.Messages != nil && values.Suffix == "" && req.Template == "" {
-			prompt, images, err = chatPrompt(c.Request.Context(), m, r.Tokenize, opts, values.Messages, []api.Tool{}, req.Think, req.Truncate == nil || *req.Truncate)
+			genTruncate := (req.Truncate == nil || *req.Truncate) && !m.IsMLX()
+			prompt, images, err = chatPrompt(c.Request.Context(), m, r.Tokenize, opts, values.Messages, []api.Tool{}, req.Think, genTruncate)
 			if err != nil {
 				c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
 				return
@@ -557,6 +558,7 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 					PromptEvalDuration: cr.PromptEvalDuration,
 					EvalCount:          cr.EvalCount,
 					EvalDuration:       cr.EvalDuration,
+					PeakMemory:         cr.PeakMemory,
 				},
 				Logprobs: toAPILogprobs(cr.Logprobs),
 			}
@@ -1951,6 +1953,9 @@ func (s *Server) PsHandler(c *gin.Context) {
 		}
 		if v.llama != nil {
 			mr.ContextLength = v.llama.ContextLength()
+			total, vram := v.llama.MemorySize()
+			mr.Size = int64(total)
+			mr.SizeVRAM = int64(vram)
 		}
 		// The scheduler waits to set expiresAt, so if a model is loading it's
 		// possible that it will be set to the unix epoch. For those cases, just
@@ -2213,6 +2218,9 @@ func (s *Server) ChatHandler(c *gin.Context) {
 	}
 
 	truncate := req.Truncate == nil || *req.Truncate
+	if m.IsMLX() {
+		truncate = false
+	}
 	prompt, images, err := chatPrompt(c.Request.Context(), m, r.Tokenize, opts, msgs, processedTools, req.Think, truncate)
 	if err != nil {
 		slog.Error("chat prompt error", "error", err)
@@ -2309,6 +2317,7 @@ func (s *Server) ChatHandler(c *gin.Context) {
 						PromptEvalDuration: r.PromptEvalDuration,
 						EvalCount:          r.EvalCount,
 						EvalDuration:       r.EvalDuration,
+						PeakMemory:         r.PeakMemory,
 					},
 					Logprobs: toAPILogprobs(r.Logprobs),
 				}

server/sched.go

@@ -231,7 +231,7 @@ func (s *Scheduler) processPending(ctx context.Context) {
 					}
 
 					// Check for experimental safetensors LLM models
-					if pending.model.Config.ModelFormat == "safetensors" {
+					if pending.model.IsMLX() {
 						if slices.Contains(pending.model.Config.Capabilities, "completion") {
 							// LLM model with safetensors format - use MLX runner
 							if s.loadMLX(pending) {
@@ -536,6 +536,7 @@ iGPUScan:
 		}
 	}
 
+	totalSize, vramSize := llama.MemorySize()
 	runner := &runnerRef{
 		model:           req.model,
 		modelPath:       req.model.ModelPath,
@@ -545,8 +546,8 @@ iGPUScan:
 		sessionDuration: sessionDuration,
 		gpus:            gpuIDs,
 		discreteGPUs:    discreteGPUs,
-		vramSize:        llama.VRAMSize(),
-		totalSize:       llama.TotalSize(),
+		totalSize:       totalSize,
+		vramSize:        vramSize,
 		loading:         true,
 		pid:             llama.Pid(),
 	}
@@ -619,6 +620,7 @@ func (s *Scheduler) loadMLX(req *LlmRequest) bool {
 		sessionDuration = req.sessionDuration.Duration
 	}
 
+	totalSize, vramSize := server.MemorySize()
 	runner := &runnerRef{
 		model:           req.model,
 		modelPath:       req.model.ModelPath,
@@ -628,8 +630,8 @@ func (s *Scheduler) loadMLX(req *LlmRequest) bool {
 		loading:         false,
 		isImagegen:      isImagegen,
 		sessionDuration: sessionDuration,
-		totalSize:       server.TotalSize(),
-		vramSize:        server.VRAMSize(),
+		totalSize:       totalSize,
+		vramSize:        vramSize,
 	}
 
 	s.loadedMu.Lock()
@@ -762,7 +764,7 @@ func (runner *runnerRef) needsReload(ctx context.Context, req *LlmRequest) bool
 	defer cancel()
 	if !reflect.DeepEqual(runner.model.AdapterPaths, req.model.AdapterPaths) || // have the adapters changed?
 		!reflect.DeepEqual(runner.model.ProjectorPaths, req.model.ProjectorPaths) || // have the projectors changed?
-		!reflect.DeepEqual(optsExisting, optsNew) || // have the runner options changed?
+		(!runner.model.IsMLX() && !reflect.DeepEqual(optsExisting, optsNew)) || // have the runner options changed?
 		runner.llama.Ping(ctx) != nil {
 		return true
 	}

server/sched_test.go

@@ -861,8 +861,7 @@ func (s *mockLlm) Close() error {
 	s.closeCalled = true
 	return s.closeResp
 }
-func (s *mockLlm) VRAMSize() uint64                                   { return s.vramSize }
-func (s *mockLlm) TotalSize() uint64                                  { return s.totalSize }
+func (s *mockLlm) MemorySize() (uint64, uint64)                       { return s.totalSize, s.vramSize }
 func (s *mockLlm) VRAMByGPU(id ml.DeviceID) uint64                    { return s.vramByGPU[id] }
 func (s *mockLlm) Pid() int                                           { return -1 }
 func (s *mockLlm) GetPort() int                                       { return -1 }

x/imagegen/server.go

@@ -374,14 +374,9 @@ func (s *Server) Close() error {
 	return nil
 }
 
-// VRAMSize returns the estimated VRAM usage.
-func (s *Server) VRAMSize() uint64 {
-	return s.vramSize
-}
-
-// TotalSize returns the total memory usage.
-func (s *Server) TotalSize() uint64 {
-	return s.vramSize
+// MemorySize returns the total and VRAM memory usage.
+func (s *Server) MemorySize() (total, vram uint64) {
+	return s.vramSize, s.vramSize
 }
 
 // VRAMByGPU returns VRAM usage for a specific GPU.

x/mlxrunner/cache.go

@@ -78,6 +78,12 @@ func (c *kvCache) findRemaining(tokens []int32) []int32 {
 		prefix++
 	}
 
+	// Always keep at least one token to re-evaluate so the
+	// pipeline can seed token generation from it.
+	if prefix == len(tokens) && prefix > 0 {
+		prefix--
+	}
+
 	if prefix < len(c.tokens) {
 		trim := len(c.tokens) - prefix
 		for _, kv := range c.caches {

x/mlxrunner/client.go

@@ -8,7 +8,6 @@ import (
 	"fmt"
 	"io"
 	"log/slog"
-	"math"
 	"math/rand"
 	"net"
 	"net/http"
@@ -19,25 +18,27 @@ import (
 	"strconv"
 	"strings"
 	"sync"
+	"sync/atomic"
 	"time"
 
+	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/llm"
 	"github.com/ollama/ollama/ml"
 	"github.com/ollama/ollama/x/imagegen"
-	"github.com/ollama/ollama/x/imagegen/manifest"
 )
 
 // Client wraps an MLX runner subprocess to implement llm.LlamaServer for LLM models.
 type Client struct {
-	port        int
-	modelName   string
-	vramSize    uint64
-	done        chan error
-	client      *http.Client
-	lastErr     string
-	lastErrLock sync.Mutex
-	mu          sync.Mutex
-	cmd         *exec.Cmd
+	port          int
+	modelName     string
+	contextLength atomic.Int64
+	memory        atomic.Uint64
+	done          chan error
+	client        *http.Client
+	lastErr       string
+	lastErrLock   sync.Mutex
+	mu            sync.Mutex
+	cmd           *exec.Cmd
 }
 
 // NewClient spawns a new MLX runner subprocess for LLM models and waits until it's ready.
@@ -98,18 +99,9 @@ func NewClient(modelName string) (*Client, error) {
 		slog.Debug("mlx subprocess library path", "LD_LIBRARY_PATH", pathEnvVal)
 	}
 
-	// Estimate VRAM based on tensor size from manifest
-	var vramSize uint64
-	if modelManifest, err := manifest.LoadManifest(modelName); err == nil {
-		vramSize = uint64(modelManifest.TotalTensorSize())
-	} else {
-		vramSize = 8 * 1024 * 1024 * 1024
-	}
-
 	c := &Client{
 		port:      port,
 		modelName: modelName,
-		vramSize:  vramSize,
 		done:      make(chan error, 1),
 		client:    &http.Client{Timeout: 10 * time.Minute},
 		cmd:       cmd,
@@ -201,6 +193,20 @@ type completionOpts struct {
 	NumPredict  int     `json:"num_predict,omitempty"`
 }
 
+type CompletionResponse struct {
+	Content    string
+	Done       bool
+	DoneReason int
+
+	PromptEvalCount    int
+	PromptEvalDuration time.Duration
+	EvalCount          int
+	EvalDuration       time.Duration
+	PeakMemory         uint64
+
+	Error *api.StatusError
+}
+
 // Close terminates the subprocess.
 func (c *Client) Close() error {
 	c.mu.Lock()
@@ -260,28 +266,25 @@ func (c *Client) Completion(ctx context.Context, req llm.CompletionRequest, fn f
 
 	scanner := bufio.NewScanner(resp.Body)
 	for scanner.Scan() {
-		var raw struct {
-			Content            string `json:"content,omitempty"`
-			Done               bool   `json:"done"`
-			DoneReason         int    `json:"done_reason,omitempty"`
-			PromptEvalCount    int    `json:"prompt_eval_count,omitempty"`
-			PromptEvalDuration int    `json:"prompt_eval_duration,omitempty"`
-			EvalCount          int    `json:"eval_count,omitempty"`
-			EvalDuration       int    `json:"eval_duration,omitempty"`
-		}
+		var raw CompletionResponse
 		if err := json.Unmarshal(scanner.Bytes(), &raw); err != nil {
 			slog.Debug("mlx response parse error", "error", err, "line", string(scanner.Bytes()))
 			continue
 		}
 
+		if raw.Error != nil {
+			return *raw.Error
+		}
+
 		cresp := llm.CompletionResponse{
 			Content:            raw.Content,
 			Done:               raw.Done,
 			DoneReason:         llm.DoneReason(raw.DoneReason),
 			PromptEvalCount:    raw.PromptEvalCount,
-			PromptEvalDuration: time.Duration(raw.PromptEvalDuration),
+			PromptEvalDuration: raw.PromptEvalDuration,
 			EvalCount:          raw.EvalCount,
-			EvalDuration:       time.Duration(raw.EvalDuration),
+			EvalDuration:       raw.EvalDuration,
+			PeakMemory:         raw.PeakMemory,
 		}
 
 		fn(cresp)
@@ -294,7 +297,7 @@ func (c *Client) Completion(ctx context.Context, req llm.CompletionRequest, fn f
 }
 
 func (c *Client) ContextLength() int {
-	return math.MaxInt
+	return int(c.contextLength.Load())
 }
 
 // Detokenize implements llm.LlamaServer.
@@ -347,9 +350,16 @@ func (c *Client) Pid() int {
 	return -1
 }
 
+type statusResponse struct {
+	Status        int
+	Progress      int
+	ContextLength int
+	Memory        uint64
+}
+
 // Ping implements llm.LlamaServer.
 func (c *Client) Ping(ctx context.Context) error {
-	reqURL := fmt.Sprintf("http://127.0.0.1:%d/health", c.port)
+	reqURL := fmt.Sprintf("http://127.0.0.1:%d/v1/status", c.port)
 	req, err := http.NewRequestWithContext(ctx, "GET", reqURL, nil)
 	if err != nil {
 		return err
@@ -362,6 +372,15 @@ func (c *Client) Ping(ctx context.Context) error {
 	if resp.StatusCode != http.StatusOK {
 		return fmt.Errorf("health check failed: %d", resp.StatusCode)
 	}
+
+	var status statusResponse
+	if err := json.NewDecoder(resp.Body).Decode(&status); err != nil {
+		return err
+	}
+
+	c.contextLength.Store(int64(status.ContextLength))
+	c.memory.Store(status.Memory)
+
 	return nil
 }
 
@@ -388,19 +407,24 @@ func (c *Client) Tokenize(ctx context.Context, content string) ([]int, error) {
 	return tokens, nil
 }
 
-// TotalSize implements llm.LlamaServer.
-func (c *Client) TotalSize() uint64 {
-	return c.vramSize
+func (c *Client) currentMemory() uint64 {
+	ctx, cancel := context.WithTimeout(context.Background(), time.Second)
+	defer cancel()
+	if err := c.Ping(ctx); err != nil {
+		slog.Warn("failed to get current memory", "error", err)
+	}
+	return c.memory.Load()
+}
+
+// MemorySize implements llm.LlamaServer.
+func (c *Client) MemorySize() (total, vram uint64) {
+	mem := c.currentMemory()
+	return mem, mem
 }
 
 // VRAMByGPU implements llm.LlamaServer.
 func (c *Client) VRAMByGPU(id ml.DeviceID) uint64 {
-	return c.vramSize
-}
-
-// VRAMSize implements llm.LlamaServer.
-func (c *Client) VRAMSize() uint64 {
-	return c.vramSize
+	return c.currentMemory()
 }
 
 // WaitUntilRunning implements llm.LlamaServer.

x/mlxrunner/mlx/memory.go

@@ -64,6 +64,10 @@ func PeakMemory() int {
 	return int(peak)
 }
 
+func ResetPeakMemory() {
+	C.mlx_reset_peak_memory()
+}
+
 type Memory struct{}
 
 func (Memory) LogValue() slog.Value {

x/mlxrunner/model/base/base.go

@@ -20,6 +20,7 @@ type Model interface {
 	Unembed(x *mlx.Array) *mlx.Array
 	NumLayers() int
 	Tokenizer() *tokenizer.Tokenizer
+	MaxContextLength() int
 
 	// LoadWeights receives all tensors loaded from the manifest and assigns
 	// them to model fields. Model-specific logic (MLA absorption, expert

x/mlxrunner/pipeline.go

@@ -6,9 +6,12 @@ import (
 	"bytes"
 	"context"
 	"errors"
+	"fmt"
 	"log/slog"
+	"net/http"
 	"time"
 
+	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/logutil"
 	"github.com/ollama/ollama/x/mlxrunner/mlx"
 )
@@ -44,16 +47,35 @@ func (r *Runner) TextGenerationPipeline(request Request) error {
 	} else {
 		mlx.DisableCompile()
 	}
+	mlx.ResetPeakMemory()
 
 	inputs := r.Tokenizer.Encode(request.Prompt, true)
+	if len(inputs) == 0 {
+		return errors.New("empty prompt")
+	}
+
+	if len(inputs) >= r.contextLength {
+		return api.StatusError{
+			StatusCode:   http.StatusBadRequest,
+			ErrorMessage: fmt.Sprintf("input length (%d tokens) exceeds the model's maximum context length (%d tokens)", len(inputs), r.contextLength),
+		}
+	}
+
+	// Cap generation to stay within the model's context length
+	maxGenerate := r.contextLength - len(inputs)
+	if request.Options.MaxTokens <= 0 {
+		request.Options.MaxTokens = maxGenerate
+	} else {
+		request.Options.MaxTokens = min(request.Options.MaxTokens, maxGenerate)
+	}
+
 	session := r.cache.begin(r.Model, inputs)
 	defer session.close()
-
 	caches := session.caches
 	tokens := session.remaining
 
+	now := time.Now()
 	total, processed := len(tokens), 0
-	slog.Info("Prompt processing progress", "processed", processed, "total", total)
 	for total-processed > 1 {
 		if err := request.Ctx.Err(); err != nil {
 			return err
@@ -93,8 +115,7 @@ func (r *Runner) TextGenerationPipeline(request Request) error {
 
 	var b bytes.Buffer
 
-	now := time.Now()
-	final := Response{Done: true, PromptTokens: total, CompletionTokens: request.Options.MaxTokens, DoneReason: 1}
+	final := CompletionResponse{Done: true, PromptEvalCount: len(inputs), EvalCount: request.Options.MaxTokens, DoneReason: 1}
 	for i := range request.Options.MaxTokens {
 		if err := request.Ctx.Err(); err != nil {
 			return err
@@ -103,9 +124,8 @@ func (r *Runner) TextGenerationPipeline(request Request) error {
 		nextSample, nextLogprobs = step(sample)
 
 		if i == 0 {
-			slog.Info("Prompt processing progress", "processed", total, "total", total)
 			mlx.Eval(sample)
-			final.PromptTokensDuration = time.Since(now)
+			final.PromptEvalDuration = time.Since(now)
 			now = time.Now()
 		}
 
@@ -113,18 +133,16 @@ func (r *Runner) TextGenerationPipeline(request Request) error {
 		session.outputs = append(session.outputs, output)
 
 		if r.Tokenizer.IsEOS(output) {
-			final.Token = int(output)
 			final.DoneReason = 0
-			final.CompletionTokens = i
+			final.EvalCount = i
 			break
 		}
 
 		select {
 		case <-request.Ctx.Done():
 			return request.Ctx.Err()
-		case request.Responses <- Response{
-			Text:  r.Decode(output, &b),
-			Token: int(output),
+		case request.Responses <- CompletionResponse{
+			Content: r.Decode(output, &b),
 		}:
 		}
 
@@ -137,7 +155,8 @@ func (r *Runner) TextGenerationPipeline(request Request) error {
 		}
 	}
 
-	final.CompletionTokensDuration = time.Since(now)
+	final.EvalDuration = time.Since(now)
+	final.PeakMemory = uint64(mlx.PeakMemory())
 	select {
 	case <-request.Ctx.Done():
 		return request.Ctx.Err()

x/mlxrunner/runner.go

@@ -4,14 +4,15 @@ package mlxrunner
 
 import (
 	"context"
+	"errors"
 	"log/slog"
 	"net"
 	"net/http"
 	"strings"
-	"time"
 
 	"golang.org/x/sync/errgroup"
 
+	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/x/mlxrunner/mlx"
 	"github.com/ollama/ollama/x/mlxrunner/model"
 	"github.com/ollama/ollama/x/mlxrunner/model/base"
@@ -21,7 +22,7 @@ import (
 
 type Request struct {
 	TextCompletionsRequest
-	Responses chan Response
+	Responses chan CompletionResponse
 	Pipeline  func(Request) error
 
 	Ctx context.Context
@@ -43,25 +44,12 @@ type TextCompletionsRequest struct {
 	} `json:"options"`
 }
 
-type Response struct {
-	Text       string    `json:"content,omitempty"`
-	Token      int       `json:"token,omitempty"`
-	Logprobs   []float32 `json:"logprobs,omitempty"`
-	Done       bool      `json:"done,omitempty"`
-	DoneReason int       `json:"done_reason,omitempty"`
-
-	PromptTokens             int           `json:"prompt_eval_count,omitempty"`
-	PromptTokensDuration     time.Duration `json:"prompt_eval_duration,omitempty"`
-	CompletionTokens         int           `json:"eval_count,omitempty"`
-	CompletionTokensDuration time.Duration `json:"eval_duration,omitempty"`
-	TotalTokens              int           `json:"total_tokens,omitempty"`
-}
-
 type Runner struct {
-	Model     base.Model
-	Tokenizer *tokenizer.Tokenizer
-	Requests  chan Request
-	cache     kvCache
+	Model         base.Model
+	Tokenizer     *tokenizer.Tokenizer
+	Requests      chan Request
+	cache         kvCache
+	contextLength int
 }
 
 func (r *Runner) Load(modelName string) error {
@@ -90,6 +78,7 @@ func (r *Runner) Load(modelName string) error {
 
 	r.Model = m
 	r.Tokenizer = m.Tokenizer()
+	r.contextLength = m.MaxContextLength()
 	return nil
 }
 
@@ -158,6 +147,17 @@ func (r *Runner) Run(host, port string, mux http.Handler) error {
 			case request := <-r.Requests:
 				if err := request.Pipeline(request); err != nil {
 					slog.Info("Request terminated", "error", err)
+					var statusErr api.StatusError
+					if !errors.As(err, &statusErr) {
+						statusErr = api.StatusError{
+							StatusCode:   http.StatusInternalServerError,
+							ErrorMessage: err.Error(),
+						}
+					}
+					select {
+					case request.Responses <- CompletionResponse{Error: &statusErr}:
+					case <-request.Ctx.Done():
+					}
 				}
 
 				close(request.Responses)

x/mlxrunner/server.go

@@ -50,9 +50,11 @@ func Execute(args []string) error {
 
 	mux := http.NewServeMux()
 	mux.HandleFunc("GET /v1/status", func(w http.ResponseWriter, r *http.Request) {
-		if err := json.NewEncoder(w).Encode(map[string]any{
-			"status":   0,
-			"progress": 100,
+		if err := json.NewEncoder(w).Encode(statusResponse{
+			Status:        0,
+			Progress:      100,
+			ContextLength: runner.contextLength,
+			Memory:        uint64(mlx.ActiveMemory() + mlx.CacheMemory()),
 		}); err != nil {
 			slog.Error("Failed to encode response", "error", err)
 			http.Error(w, "Internal Server Error", http.StatusInternalServerError)
@@ -78,7 +80,7 @@ func Execute(args []string) error {
 	})
 
 	mux.HandleFunc("POST /v1/completions", func(w http.ResponseWriter, r *http.Request) {
-		request := Request{Responses: make(chan Response)}
+		request := Request{Responses: make(chan CompletionResponse)}
 
 		if err := json.NewDecoder(r.Body).Decode(&request.TextCompletionsRequest); err != nil {
 			slog.Error("Failed to decode request", "error", err)
@@ -87,9 +89,6 @@ func Execute(args []string) error {
 		}
 
 		request.Options.MaxTokens = cmp.Or(request.Options.MaxTokens, request.Options.NumPredict)
-		if request.Options.MaxTokens < 1 {
-			request.Options.MaxTokens = 16 << 10
-		}
 
 		request.Pipeline = runner.TextGenerationPipeline
 		request.Sampler = sample.New(

x/models/gemma3/gemma3.go

@@ -430,6 +430,10 @@ func (m *Model) NumLayers() int {
 	return len(m.Layers)
 }
 
+func (m *Model) MaxContextLength() int {
+	return int(m.MaxPositionEmbeddings)
+}
+
 func (m *Model) Tokenizer() *tokenizer.Tokenizer {
 	return m.tok
 }

x/models/glm4_moe_lite/glm4_moe_lite.go

@@ -733,7 +733,7 @@ func (m *Model) Unembed(x *mlx.Array) *mlx.Array {
 func (m *Model) NumLayers() int { return len(m.Layers) }
 
 // MaxContextLength returns the maximum context length
-func (m *Model) MaxContextLength() int32 { return m.MaxPositionEmbeddings }
+func (m *Model) MaxContextLength() int { return int(m.MaxPositionEmbeddings) }
 
 // VocabSize returns the vocabulary size
 func (m *Model) VocabSize() int32 { return m.Config.VocabSize }

x/models/llama/llama.go

@@ -262,6 +262,10 @@ func (m *Model) NumLayers() int {
 	return len(m.Layers)
 }
 
+func (m *Model) MaxContextLength() int {
+	return int(m.MaxPositionEmbeddings)
+}
+
 func (m *Model) Tokenizer() *tokenizer.Tokenizer {
 	return m.tok
 }

x/models/qwen3/qwen3.go

@@ -279,6 +279,10 @@ func (m *Model) NumLayers() int {
 	return len(m.Layers)
 }
 
+func (m *Model) MaxContextLength() int {
+	return int(m.MaxPositionEmbeddings)
+}
+
 func (m *Model) Tokenizer() *tokenizer.Tokenizer {
 	return m.tok
 }