server/internal/chunks: remove chunks package (#9755)

Replace large-chunk blob downloads with parallel small-chunk
verification to solve timeout and performance issues. Registry users
experienced progressively slowing download speeds as large-chunk
transfers aged, often timing out completely.

The previous approach downloaded blobs in a few large chunks but
required a separate, single-threaded pass to read the entire blob back
from disk for verification after download completion.

This change uses the new chunksums API to fetch many smaller
chunk+digest pairs, allowing concurrent downloads and immediate
verification as each chunk arrives. Chunks are written directly to their
final positions, eliminating the entire separate verification pass.

The result is more reliable downloads that maintain speed throughout the
transfer process and significantly faster overall completion, especially
over unstable connections or with large blobs.

README.md

@@ -54,6 +54,10 @@ Here are some example models that can be downloaded:
 
 | Model              | Parameters | Size  | Download                         |
 | ------------------ | ---------- | ----- | -------------------------------- |
+| Gemma 3            | 1B         | 815MB | `ollama run gemma3:1b`           |
+| Gemma 3            | 4B         | 3.3GB | `ollama run gemma3`              |
+| Gemma 3            | 12B        | 8.1GB | `ollama run gemma3:12b`          |
+| Gemma 3            | 27B        | 17GB  | `ollama run gemma3:27b`          |
 | QwQ                | 32B        | 20GB  | `ollama run qwq`                 |
 | DeepSeek-R1        | 7B         | 4.7GB | `ollama run deepseek-r1`         |
 | DeepSeek-R1        | 671B       | 404GB | `ollama run deepseek-r1:671b`    |
@@ -66,9 +70,6 @@ Here are some example models that can be downloaded:
 | Llama 3.1          | 405B       | 231GB | `ollama run llama3.1:405b`       |
 | Phi 4              | 14B        | 9.1GB | `ollama run phi4`                |
 | Phi 4 Mini         | 3.8B       | 2.5GB | `ollama run phi4-mini`           |
-| Gemma 2            | 2B         | 1.6GB | `ollama run gemma2:2b`           |
-| Gemma 2            | 9B         | 5.5GB | `ollama run gemma2`              |
-| Gemma 2            | 27B        | 16GB  | `ollama run gemma2:27b`          |
 | Mistral            | 7B         | 4.1GB | `ollama run mistral`             |
 | Moondream 2        | 1.4B       | 829MB | `ollama run moondream`           |
 | Neural Chat        | 7B         | 4.1GB | `ollama run neural-chat`         |
@@ -571,6 +572,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [llama.cpp](https://github.com/ggerganov/llama.cpp) project founded by Georgi Gerganov.
 
 ### Observability
+- [Opik](https://www.comet.com/docs/opik/cookbook/ollama) is an open-source platform to debug, evaluate, and monitor your LLM applications, RAG systems, and agentic workflows with comprehensive tracing, automated evaluations, and production-ready dashboards. Opik supports native intergration to Ollama.
 - [Lunary](https://lunary.ai/docs/integrations/ollama) is the leading open-source LLM observability platform. It provides a variety of enterprise-grade features such as real-time analytics, prompt templates management, PII masking, and comprehensive agent tracing.
 - [OpenLIT](https://github.com/openlit/openlit) is an OpenTelemetry-native tool for monitoring Ollama Applications & GPUs using traces and metrics.
 - [HoneyHive](https://docs.honeyhive.ai/integrations/ollama) is an AI observability and evaluation platform for AI agents. Use HoneyHive to evaluate agent performance, interrogate failures, and monitor quality in production.

api/types.go

@@ -349,6 +349,7 @@ type ShowResponse struct {
 	Messages      []Message      `json:"messages,omitempty"`
 	ModelInfo     map[string]any `json:"model_info,omitempty"`
 	ProjectorInfo map[string]any `json:"projector_info,omitempty"`
+	Tensors       []Tensor       `json:"tensors,omitempty"`
 	ModifiedAt    time.Time      `json:"modified_at,omitempty"`
 }
 
@@ -467,6 +468,13 @@ type ModelDetails struct {
 	QuantizationLevel string   `json:"quantization_level"`
 }
 
+// Tensor describes the metadata for a given tensor.
+type Tensor struct {
+	Name  string   `json:"name"`
+	Type  string   `json:"type"`
+	Shape []uint64 `json:"shape"`
+}
+
 func (m *Metrics) Summary() {
 	if m.TotalDuration > 0 {
 		fmt.Fprintf(os.Stderr, "total duration:       %v\n", m.TotalDuration)

cmd/cmd.go

@@ -18,6 +18,7 @@ import (
 	"os/signal"
 	"path/filepath"
 	"runtime"
+	"sort"
 	"strconv"
 	"strings"
 	"sync/atomic"
@@ -568,8 +569,9 @@ func ShowHandler(cmd *cobra.Command, args []string) error {
 	parameters, errParams := cmd.Flags().GetBool("parameters")
 	system, errSystem := cmd.Flags().GetBool("system")
 	template, errTemplate := cmd.Flags().GetBool("template")
+	verbose, errVerbose := cmd.Flags().GetBool("verbose")
 
-	for _, boolErr := range []error{errLicense, errModelfile, errParams, errSystem, errTemplate} {
+	for _, boolErr := range []error{errLicense, errModelfile, errParams, errSystem, errTemplate, errVerbose} {
 		if boolErr != nil {
 			return errors.New("error retrieving flags")
 		}
@@ -607,7 +609,7 @@ func ShowHandler(cmd *cobra.Command, args []string) error {
 		return errors.New("only one of '--license', '--modelfile', '--parameters', '--system', or '--template' can be specified")
 	}
 
-	req := api.ShowRequest{Name: args[0]}
+	req := api.ShowRequest{Name: args[0], Verbose: verbose}
 	resp, err := client.Show(cmd.Context(), &req)
 	if err != nil {
 		return err
@@ -630,10 +632,10 @@ func ShowHandler(cmd *cobra.Command, args []string) error {
 		return nil
 	}
 
-	return showInfo(resp, os.Stdout)
+	return showInfo(resp, verbose, os.Stdout)
 }
 
-func showInfo(resp *api.ShowResponse, w io.Writer) error {
+func showInfo(resp *api.ShowResponse, verbose bool, w io.Writer) error {
 	tableRender := func(header string, rows func() [][]string) {
 		fmt.Fprintln(w, " ", header)
 		table := tablewriter.NewWriter(w)
@@ -690,6 +692,45 @@ func showInfo(resp *api.ShowResponse, w io.Writer) error {
 		})
 	}
 
+	if resp.ModelInfo != nil && verbose {
+		tableRender("Metadata", func() (rows [][]string) {
+			keys := make([]string, 0, len(resp.ModelInfo))
+			for k := range resp.ModelInfo {
+				keys = append(keys, k)
+			}
+			sort.Strings(keys)
+
+			for _, k := range keys {
+				var v string
+				switch vData := resp.ModelInfo[k].(type) {
+				case string:
+					v = vData
+				case float64:
+					v = fmt.Sprintf("%g", vData)
+				case []any:
+					n := 3
+					if len(vData) < n {
+						n = len(vData)
+					}
+					v = fmt.Sprintf("%v", vData[:n])
+				default:
+					v = fmt.Sprintf("%T", vData)
+				}
+				rows = append(rows, []string{"", k, v})
+			}
+			return
+		})
+	}
+
+	if len(resp.Tensors) > 0 && verbose {
+		tableRender("Tensors", func() (rows [][]string) {
+			for _, t := range resp.Tensors {
+				rows = append(rows, []string{"", t.Name, t.Type, fmt.Sprint(t.Shape)})
+			}
+			return
+		})
+	}
+
 	head := func(s string, n int) (rows [][]string) {
 		scanner := bufio.NewScanner(strings.NewReader(s))
 		for scanner.Scan() && (len(rows) < n || n < 0) {
@@ -1196,6 +1237,7 @@ func NewCLI() *cobra.Command {
 	showCmd.Flags().Bool("parameters", false, "Show parameters of a model")
 	showCmd.Flags().Bool("template", false, "Show template of a model")
 	showCmd.Flags().Bool("system", false, "Show system message of a model")
+	showCmd.Flags().BoolP("verbose", "v", false, "Show detailed model information")
 
 	runCmd := &cobra.Command{
 		Use:     "run MODEL [PROMPT]",

cmd/cmd_test.go

@@ -27,7 +27,7 @@ func TestShowInfo(t *testing.T) {
 				ParameterSize:     "7B",
 				QuantizationLevel: "FP16",
 			},
-		}, &b); err != nil {
+		}, false, &b); err != nil {
 			t.Fatal(err)
 		}
 
@@ -57,7 +57,7 @@ func TestShowInfo(t *testing.T) {
 				ParameterSize:     "7B",
 				QuantizationLevel: "FP16",
 			},
-		}, &b); err != nil {
+		}, false, &b); err != nil {
 			t.Fatal(err)
 		}
 
@@ -68,6 +68,56 @@ func TestShowInfo(t *testing.T) {
     embedding length    0       
     quantization        FP16    
 
+`
+		if diff := cmp.Diff(expect, b.String()); diff != "" {
+			t.Errorf("unexpected output (-want +got):\n%s", diff)
+		}
+	})
+
+	t.Run("verbose model", func(t *testing.T) {
+		var b bytes.Buffer
+		if err := showInfo(&api.ShowResponse{
+			Details: api.ModelDetails{
+				Family:            "test",
+				ParameterSize:     "8B",
+				QuantizationLevel: "FP16",
+			},
+			Parameters: `
+			stop up`,
+			ModelInfo: map[string]any{
+				"general.architecture":    "test",
+				"general.parameter_count": float64(8_000_000_000),
+				"test.context_length":     float64(1000),
+				"test.embedding_length":   float64(11434),
+			},
+			Tensors: []api.Tensor{
+				{Name: "blk.0.attn_k.weight", Type: "BF16", Shape: []uint64{42, 3117}},
+				{Name: "blk.0.attn_q.weight", Type: "FP16", Shape: []uint64{3117, 42}},
+			},
+		}, true, &b); err != nil {
+			t.Fatal(err)
+		}
+
+		expect := `  Model
+    architecture        test     
+    parameters          8B       
+    context length      1000     
+    embedding length    11434    
+    quantization        FP16     
+
+  Parameters
+    stop    up    
+
+  Metadata
+    general.architecture       test     
+    general.parameter_count    8e+09    
+    test.context_length        1000     
+    test.embedding_length      11434    
+
+  Tensors
+    blk.0.attn_k.weight    BF16    [42 3117]    
+    blk.0.attn_q.weight    FP16    [3117 42]    
+
 `
 		if diff := cmp.Diff(expect, b.String()); diff != "" {
 			t.Errorf("unexpected output (-want +got):\n%s", diff)
@@ -89,7 +139,7 @@ func TestShowInfo(t *testing.T) {
 			stop you
 			stop up
 			temperature 99`,
-		}, &b); err != nil {
+		}, false, &b); err != nil {
 			t.Fatal(err)
 		}
 
@@ -126,7 +176,7 @@ func TestShowInfo(t *testing.T) {
 				"clip.vision.embedding_length": float64(0),
 				"clip.vision.projection_dim":   float64(0),
 			},
-		}, &b); err != nil {
+		}, false, &b); err != nil {
 			t.Fatal(err)
 		}
 
@@ -159,7 +209,7 @@ func TestShowInfo(t *testing.T) {
 Ahoy, matey!
 Weigh anchor!
 			`,
-		}, &b); err != nil {
+		}, false, &b); err != nil {
 			t.Fatal(err)
 		}
 
@@ -188,7 +238,7 @@ Weigh anchor!
 				QuantizationLevel: "FP16",
 			},
 			License: license,
-		}, &b); err != nil {
+		}, false, &b); err != nil {
 			t.Fatal(err)
 		}
 

cmd/interactive.go

@@ -195,6 +195,10 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 			opts.Messages = []api.Message{}
 			fmt.Printf("Loading model '%s'\n", opts.Model)
 			if err := loadOrUnloadModel(cmd, &opts); err != nil {
+				if strings.Contains(err.Error(), "not found") {
+					fmt.Printf("error: %v\n", err)
+					continue
+				}
 				return err
 			}
 			continue
@@ -343,7 +347,7 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 
 				switch args[1] {
 				case "info":
-					_ = showInfo(resp, os.Stderr)
+					_ = showInfo(resp, false, os.Stderr)
 				case "license":
 					if resp.License == "" {
 						fmt.Println("No license was specified for this model.")

convert/convert.go

@@ -13,8 +13,13 @@ import (
 )
 
 type ModelParameters struct {
-	Architectures []string `json:"architectures"`
-	VocabSize     uint32   `json:"vocab_size"`
+	Architectures []string       `json:"architectures"`
+	VocabSize     uint32         `json:"vocab_size"`
+	TextModel     TextParameters `json:"text_config"`
+}
+
+type TextParameters struct {
+	VocabSize uint32 `json:"vocab_size"`
 }
 
 type AdapterParameters struct {
@@ -185,6 +190,8 @@ func ConvertModel(fsys fs.FS, ws io.WriteSeeker) error {
 		conv = &gemmaModel{}
 	case "Gemma2ForCausalLM":
 		conv = &gemma2Model{}
+	case "Gemma3ForCausalLM", "Gemma3ForConditionalGeneration":
+		conv = &gemma3Model{Architecture: p.Architectures[0]}
 	case "Phi3ForCausalLM":
 		conv = &phi3Model{}
 	case "Qwen2ForCausalLM":
@@ -213,7 +220,14 @@ func ConvertModel(fsys fs.FS, ws io.WriteSeeker) error {
 	}
 
 	vocabSize := int(p.VocabSize)
+	if vocabSize == 0 {
+		tVocabSize := int(p.TextModel.VocabSize)
+		vocabSize = tVocabSize
+	}
+
 	switch {
+	case vocabSize == 0:
+		slog.Warn("vocabulary size was not explicitly set by the model", "default size", len(t.Vocabulary.Tokens))
 	case vocabSize > len(t.Vocabulary.Tokens):
 		slog.Warn("vocabulary is smaller than expected, padding with dummy tokens", "expect", vocabSize, "actual", len(t.Vocabulary.Tokens))
 		for i := range vocabSize - len(t.Vocabulary.Tokens) {

convert/convert_gemma.go

@@ -45,7 +45,7 @@ func (p *gemmaModel) KV(t *Tokenizer) ggml.KV {
 func (p *gemmaModel) Tensors(ts []Tensor) []ggml.Tensor {
 	var out []ggml.Tensor
 	for _, t := range ts {
-		if strings.HasSuffix(t.Name(), "_norm.weight") {
+		if !strings.HasPrefix(t.Name(), "v.") && strings.HasSuffix(t.Name(), "_norm.weight") {
 			t.SetRepacker(p.addOne)
 		}
 

convert/convert_gemma3.go

@@ -0,0 +1,142 @@
+package convert
+
+import (
+	"cmp"
+
+	"github.com/ollama/ollama/fs/ggml"
+)
+
+type gemma3Model struct {
+	gemmaModel
+	Architecture string
+	TextModel    struct {
+		HeadDim          uint32 `json:"head_dim"`
+		HiddenSize       uint32 `json:"hidden_size"`
+		HiddenLayers     uint32 `json:"num_hidden_layers"`
+		IntermediateSize uint32 `json:"intermediate_size"`
+		SlidingWindow    uint32 `json:"sliding_window"`
+	} `json:"text_config"`
+	VisionModel struct {
+		NumAttentionHeads uint32  `json:"num_attention_heads"` // attention.head_count 16
+		LayerNormEpsilon  float32 `json:"layer_norm_eps"`      // attention.layer_norm_epsilon 1e-05
+		NumHiddenLayers   uint32  `json:"num_hidden_layers"`   // block_count 32
+		HiddenSize        uint32  `json:"hidden_size"`         // embedding_length 1280
+		IntermediateSize  uint32  `json:"intermediate_size"`   // feed_forward_length 5120
+		ImageSize         uint32  `json:"image_size"`          // image_size 560
+		NumChannels       uint32  `json:"num_channels"`        // num_channels 3
+		PatchSize         uint32  `json:"patch_size"`          // patch_size 14
+	} `json:"vision_config"`
+	MaxPositionEmbeddings    uint32  `json:"max_position_embeddings"`
+	NumAttentionHeads        uint32  `json:"num_attention_heads"`
+	NumKeyValueHeads         uint32  `json:"num_key_value_heads"`
+	RMSNormEPS               float32 `json:"rms_norm_eps"`
+	HeadDim                  uint32  `json:"head_dim"`
+	FinalLogitSoftcap        float32 `json:"final_logit_softcapping"`
+	RopeLocalTheta           float32 `json:"rope_local_base_freq"`
+	RopeGlobalTheta          float32 `json:"rope_global_base_freq"`
+	SlidingWindow            uint32  `json:"sliding_window"`
+	MultiModalTokensPerImage uint32  `json:"mm_tokens_per_image"`
+}
+
+const (
+	gemma4BLayerCount  = 34
+	gemma12BLayerCount = 48
+	gemma27BLayerCount = 62
+)
+
+func (p *gemma3Model) KV(t *Tokenizer) ggml.KV {
+	kv := p.ModelParameters.KV(t)
+	kv["general.architecture"] = "gemma3"
+
+	numBlocks := cmp.Or(p.HiddenLayers, p.TextModel.HiddenLayers)
+	kv["gemma3.block_count"] = numBlocks
+
+	var (
+		numHeads   uint32
+		numKVHeads uint32
+	)
+
+	switch numBlocks {
+	case gemma4BLayerCount:
+		numHeads = 8
+		numKVHeads = 4
+	case gemma12BLayerCount:
+		numHeads = 16
+		numKVHeads = 8
+	case gemma27BLayerCount:
+		numHeads = 32
+		numKVHeads = 16
+	default:
+		numHeads = p.NumAttentionHeads
+		numKVHeads = p.NumKeyValueHeads
+	}
+
+	kv["gemma3.attention.head_count"] = numHeads
+	kv["gemma3.attention.head_count_kv"] = numKVHeads
+
+	switch p.Architecture {
+	case "Gemma3ForCausalLM":
+		kv["gemma3.context_length"] = p.MaxPositionEmbeddings
+		kv["gemma3.attention.layer_norm_rms_epsilon"] = p.RMSNormEPS
+		kv["gemma3.attention.key_length"] = p.HeadDim
+		kv["gemma3.attention.value_length"] = p.HeadDim
+		kv["gemma3.attention.sliding_window"] = p.SlidingWindow
+		kv["gemma3.final_logit_softcapping"] = cmp.Or(p.FinalLogitSoftcap, 30)
+		kv["gemma3.rope.local.freq_base"] = cmp.Or(p.RopeLocalTheta, 10000.0)
+		kv["gemma3.rope.global.freq_base"] = cmp.Or(p.RopeGlobalTheta, 1000000.0)
+		kv["gemma3.embedding_length"] = p.HiddenSize
+		kv["gemma3.feed_forward_length"] = p.IntermediateSize
+	default:
+		kv["gemma3.context_length"] = cmp.Or(p.MaxPositionEmbeddings, 131072)
+		kv["gemma3.embedding_length"] = p.TextModel.HiddenSize
+		kv["gemma3.feed_forward_length"] = p.TextModel.IntermediateSize
+		kv["gemma3.attention.sliding_window"] = p.TextModel.SlidingWindow
+		kv["gemma3.vision.block_count"] = p.VisionModel.NumHiddenLayers
+		kv["gemma3.vision.embedding_length"] = p.VisionModel.HiddenSize
+		kv["gemma3.vision.feed_forward_length"] = p.VisionModel.IntermediateSize
+		kv["gemma3.vision.image_size"] = p.VisionModel.ImageSize
+		kv["gemma3.vision.patch_size"] = p.VisionModel.PatchSize
+		kv["gemma3.vision.num_channels"] = cmp.Or(p.VisionModel.NumChannels, 3)
+		kv["gemma3.vision.attention.head_count"] = p.VisionModel.NumAttentionHeads
+		kv["gemma3.vision.attention.layer_norm_epsilon"] = cmp.Or(p.VisionModel.LayerNormEpsilon, 1e-6)
+		kv["gemma3.attention.key_length"] = cmp.Or(p.TextModel.HeadDim, 256)
+		kv["gemma3.attention.value_length"] = cmp.Or(p.TextModel.HeadDim, 256)
+	}
+
+	if p.MultiModalTokensPerImage > 0 {
+		kv["gemma3.mm.tokens_per_image"] = p.MultiModalTokensPerImage
+	}
+
+	return kv
+}
+
+func (p *gemma3Model) Replacements() []string {
+	return []string{
+		"lm_head", "output",
+		"model.embed_tokens", "token_embd",
+		"model.norm", "output_norm",
+		"vision_tower.vision_model.embeddings", "v",
+		"vision_tower.vision_model", "v",
+		"vision_model.vision_model.embeddings", "v",
+		"vision_model.vision_model", "v",
+		"language_model.", "",
+		"model.layers", "blk",
+		"encoder.layers", "blk",
+		"input_layernorm", "attn_norm",
+		"self_attn.q_proj", "attn_q",
+		"self_attn.q_norm", "attn_q_norm",
+		"self_attn.k_proj", "attn_k",
+		"self_attn.k_norm", "attn_k_norm",
+		"self_attn.v_proj", "attn_v",
+		"self_attn.o_proj", "attn_output",
+		"self_attn.out_proj", "attn_output",
+		"mlp.gate_proj", "ffn_gate",
+		"mlp.down_proj", "ffn_down",
+		"mlp.up_proj", "ffn_up",
+		"post_attention_layernorm", "post_attention_norm",
+		"pre_feedforward_layernorm", "ffn_norm",
+		"post_feedforward_layernorm", "post_ffw_norm",
+		"input_projection_weight", "input_projection.weight",
+		"multi_modal_projector", "mm",
+	}
+}

convert/tokenizer_spm.go

@@ -6,7 +6,9 @@ import (
 	"errors"
 	"fmt"
 	"io/fs"
+	"log/slog"
 	"os"
+	"reflect"
 	"slices"
 
 	"google.golang.org/protobuf/proto"
@@ -15,6 +17,8 @@ import (
 )
 
 func parseSentencePiece(fsys fs.FS) (*Vocabulary, error) {
+	slog.Debug("using spm vocabulary")
+
 	ast, err := parseAdditionalSpecialTokens(fsys)
 	if err != nil {
 		return nil, err
@@ -43,10 +47,19 @@ func parseSentencePiece(fsys fs.FS) (*Vocabulary, error) {
 			v.Types = append(v.Types, int32(t))
 		default:
 			tt := int32(sentencepiece.ModelProto_SentencePiece_NORMAL)
-			if slices.Contains(ast, piece.GetPiece()) {
+
+			// temporary fix to handle gemma3 broken configs
+			if slices.Contains([]string{"<end_of_turn>", "<start_of_turn>"}, piece.GetPiece()) {
 				tt = int32(sentencepiece.ModelProto_SentencePiece_CONTROL)
 			}
 
+			for _, t := range ast {
+				if t.Content == piece.GetPiece() {
+					tt = int32(sentencepiece.ModelProto_SentencePiece_CONTROL)
+					break
+				}
+			}
+
 			v.Types = append(v.Types, tt)
 		}
 	}
@@ -78,10 +91,16 @@ func parseSentencePiece(fsys fs.FS) (*Vocabulary, error) {
 		return cmp.Compare(i.id, j.id)
 	})
 
-	n := len(v.Tokens)
-	for i, t := range ts {
-		if t.id != i+n {
-			return nil, fmt.Errorf("invalid token id: %d", t.id)
+	for _, t := range ts {
+		if t.id < len(v.Tokens) {
+			if v.Tokens[t.id] == t.content {
+				slog.Warn("tokenizer", "duplicate token", t.content, "id", t.id)
+				continue
+			}
+			return nil, fmt.Errorf("token mismatch: %s != %s at pos [%d]", t.content, v.Tokens[t.id], t.id)
+		}
+		if t.id != len(v.Tokens) {
+			return nil, fmt.Errorf("invalid token id: [%d] as pos [%d]", t.id, len(v.Tokens))
 		}
 
 		v.Tokens = append(v.Tokens, t.content)
@@ -92,7 +111,15 @@ func parseSentencePiece(fsys fs.FS) (*Vocabulary, error) {
 	return &v, nil
 }
 
-func parseAdditionalSpecialTokens(fsys fs.FS) ([]string, error) {
+type specialToken struct {
+	Content    string `json:"content"`
+	Lstrip     bool   `json:"lstrip"`
+	Normalized bool   `json:"normalized"`
+	Rstrip     bool   `json:"rstrip"`
+	SingleWord bool   `json:"single_word"`
+}
+
+func parseAdditionalSpecialTokens(fsys fs.FS) ([]specialToken, error) {
 	f, err := fsys.Open("special_tokens_map.json")
 	if errors.Is(err, os.ErrNotExist) {
 		return nil, nil
@@ -102,12 +129,43 @@ func parseAdditionalSpecialTokens(fsys fs.FS) ([]string, error) {
 	defer f.Close()
 
 	var m struct {
-		AdditionalSpecialTokens []string `json:"additional_special_tokens"`
+		AdditionalSpecialTokens any `json:"additional_special_tokens"`
 	}
 
 	if err := json.NewDecoder(f).Decode(&m); err != nil {
 		return nil, err
 	}
 
-	return m.AdditionalSpecialTokens, nil
+	var ast []specialToken
+
+	switch st := m.AdditionalSpecialTokens.(type) {
+	case []string:
+		for _, s := range st {
+			ast = append(ast, specialToken{Content: s})
+		}
+	case []any:
+		for _, s := range st {
+			// marshal and unmarshal the object to get the special token
+			tMap := s.(map[string]any)
+			data, err := json.Marshal(tMap)
+			if err != nil {
+				return nil, err
+			}
+
+			var token specialToken
+			err = json.Unmarshal(data, &token)
+			if err != nil {
+				return nil, err
+			}
+
+			ast = append(ast, token)
+		}
+
+	default:
+		slog.Warn("special token", "unknown token", reflect.TypeOf(st))
+	}
+
+	slog.Debug("spm tokenizer", "additional tokens", ast)
+
+	return ast, nil
 }

docs/faq.md

@@ -187,6 +187,13 @@ cloudflared tunnel --url http://localhost:11434 --http-host-header="localhost:11
 
 Ollama allows cross-origin requests from `127.0.0.1` and `0.0.0.0` by default. Additional origins can be configured with `OLLAMA_ORIGINS`.
 
+For browser extensions, you'll need to explicitly allow the extension's origin pattern. Set `OLLAMA_ORIGINS` to include `chrome-extension://*`, `moz-extension://*`, and `safari-web-extension://*` if you wish to allow all browser extensions access, or specific extensions as needed:
+
+```
+# Allow all Chrome, Firefox, and Safari extensions
+OLLAMA_ORIGINS=chrome-extension://*,moz-extension://*,safari-web-extension://* ollama serve
+```
+
 Refer to the section [above](#how-do-i-configure-ollama-server) for how to set environment variables on your platform.
 
 ## Where are models stored?

docs/linux.md

@@ -75,7 +75,7 @@ RestartSec=3
 Environment="PATH=$PATH"
 
 [Install]
-WantedBy=default.target
+WantedBy=multi-user.target
 ```
 
 Then start the service:

fs/ggml/ggml.go

@@ -124,6 +124,19 @@ func (kv KV) Uints(key string, defaultValue ...[]uint32) []uint32 {
 	return s
 }
 
+func (kv KV) Floats(key string, defaultValue ...[]float32) []float32 {
+	r := keyValue(kv, key, &array{})
+	s := make([]float32, r.size)
+	for i := range r.size {
+		s[i] = float32(r.values[i].(float32))
+	}
+	return s
+}
+
+func (kv KV) OllamaEngineRequired() bool {
+	return kv.Architecture() == "gemma3"
+}
+
 func keyValue[T string | uint32 | uint64 | float32 | *array | bool](kv KV, key string, defaultValue ...T) T {
 	if !strings.HasPrefix(key, "tokenizer.") && !strings.HasPrefix(key, "general.") {
 		key = kv.Architecture() + "." + key
@@ -314,6 +327,10 @@ func (t Tensor) Size() uint64 {
 	return t.parameters() * t.typeSize() / t.blockSize()
 }
 
+func (t Tensor) Type() string {
+	return fileType(t.Kind).String()
+}
+
 type container interface {
 	Name() string
 	Decode(io.ReadSeeker) (model, error)
@@ -476,7 +493,7 @@ func (f GGML) GraphSize(context, batch uint64, kvCacheType string) (kv, partialO
 			// vocab graph
 			4*batch*(embedding+vocab)+embedding*vocab*105/128,
 		)
-	case "gemma", "gemma2":
+	case "gemma", "gemma2", "gemma3":
 		fullOffload = max(
 			4*batch*(embedding+vocab),
 			4*batch*(2+context+context*heads+2*embedding+2*embeddingHeadsK*heads),
@@ -566,39 +583,52 @@ func (f GGML) GraphSize(context, batch uint64, kvCacheType string) (kv, partialO
 }
 
 func (llm GGML) VisionGraphSize() (weights, graphSize uint64) {
+	if llm.KV().Uint("vision.block_count") == 0 {
+		return
+	}
+
+	for name, layer := range llm.Tensors().GroupLayers() {
+		if name == "v" || strings.HasPrefix(name, "v.") {
+			for _, tensor := range layer {
+				weights += tensor.Size()
+			}
+		}
+	}
+
+	imageSize := uint64(llm.KV().Uint("vision.image_size"))
+	patchSize := uint64(llm.KV().Uint("vision.patch_size"))
+	if patchSize == 0 {
+		slog.Warn("unknown patch size for vision model")
+		return
+	}
+
+	numChannels := uint64(llm.KV().Uint("vision.num_channels"))
+
+	numPatches := (imageSize / patchSize) * (imageSize / patchSize)
+	if _, ok := llm.Tensors().GroupLayers()["v"]["class_embd"]; ok {
+		numPatches++
+	}
+
+	headCount := uint64(llm.KV().Uint("vision.attention.head_count"))
+	embeddingLength := uint64(llm.KV().Uint("vision.embedding_length"))
+
 	switch llm.KV().Architecture() {
 	case "mllama":
-		for _, layer := range llm.Tensors().GroupLayers()["v"] {
-			weights += layer.Size()
-		}
-
-		kv := func(n string) uint64 {
-			if v, ok := llm.KV()["mllama.vision."+n].(uint32); ok {
-				return uint64(v)
-			}
-
-			return 0
-		}
-
-		imageSize := kv("image_size")
-
-		maxNumTiles := kv("max_num_tiles")
-		embeddingLength := kv("embedding_length")
-		headCount := kv("attention.head_count")
-
-		numPatches := (imageSize / kv("patch_size")) * (imageSize / kv("patch_size"))
-		if _, ok := llm.Tensors().GroupLayers()["v"]["class_embd"]; ok {
-			numPatches++
-		}
-
 		numPaddedPatches := numPatches + 8 - (numPatches%8)%8
 
+		maxNumTiles := uint64(llm.KV().Uint("vision.max_num_tiles"))
+
 		graphSize = 4 * (8 +
-			imageSize*imageSize*kv("num_channels")*maxNumTiles +
+			imageSize*imageSize*numChannels*maxNumTiles +
 			embeddingLength*numPatches*maxNumTiles +
 			9*embeddingLength*numPaddedPatches*maxNumTiles +
 			numPaddedPatches*maxNumTiles*numPaddedPatches*maxNumTiles*headCount)
+	case "gemma3":
+		graphSize = 4 * (imageSize*imageSize*numChannels +
+			embeddingLength*patchSize +
+			numPatches*numPatches*headCount)
 	}
+
 	return weights, graphSize
 }
 

kvcache/causal.go

@@ -21,9 +21,10 @@ type shiftFn func(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, e
 type Causal struct {
 	DType      ml.DType
 	Capacity   int32
-	causal     bool
 	windowSize int32
 
+	opts CausalOptions
+
 	// config controls mostly backend-specific optimizations
 	config *ml.CacheConfig
 
@@ -79,7 +80,6 @@ type cellRange struct {
 
 func NewCausalCache(shift shiftFn) *Causal {
 	return &Causal{
-		causal:     true,
 		windowSize: math.MaxInt32,
 		shiftFn:    shift,
 		ctxs:       make(map[int]ml.Context),
@@ -90,7 +90,6 @@ func NewCausalCache(shift shiftFn) *Causal {
 
 func NewSWACache(windowSize int32, shift shiftFn) *Causal {
 	return &Causal{
-		causal:     true,
 		windowSize: windowSize,
 		shiftFn:    shift,
 		ctxs:       make(map[int]ml.Context),
@@ -145,6 +144,7 @@ func (c *Causal) StartForward(ctx ml.Context, opts input.Options) error {
 	c.curBatchSize = len(opts.Positions)
 	c.curSequences = opts.Sequences
 	c.curPositions = opts.Positions
+	c.opts.Except = nil
 
 	var err error
 	c.curLoc, err = c.findStartLoc()
@@ -235,9 +235,10 @@ func (c *Causal) buildMask(ctx ml.Context) (ml.Tensor, error) {
 	mask := make([]float32, batchSize*length)
 
 	for i := range c.curBatchSize {
+		enabled := !slices.Contains(c.opts.Except, i)
 		for j := c.curCellRange.min; j <= c.curCellRange.max; j++ {
 			if !slices.Contains(c.cells[j].sequences, c.curSequences[i]) ||
-				(c.causal && c.cells[j].pos > c.curPositions[i]) ||
+				(enabled && c.cells[j].pos > c.curPositions[i]) ||
 				c.cells[j].pos < c.curPositions[i]-c.windowSize {
 				mask[i*length+(j-c.curCellRange.min)] = float32(math.Inf(-1))
 			}
@@ -404,15 +405,16 @@ func (c *Causal) SetLayer(layer int) {
 	c.curLayer = layer
 }
 
-// SetCausal enables or disables causal mask generation for subsequent calls to Get.
-// This state carries over to future forward passes. The default value is true.
-//
-// ctx may be set to nil if this is called from outside of a forward pass, for
-// example, when initializing the cache.
-func (c *Causal) SetCausal(ctx ml.Context, causal bool) {
-	if c.causal != causal {
-		c.causal = causal
+type CausalOptions struct {
+	// Enabled controls whether the causal mask is generated for a particular index in a batch
+	Except []int
+}
 
+// SetCausal disables causal mask generation for a particular range of indicies in
+// the current batch for subsequent calls to Get. The state resets for the next forward pass.
+func (c *Causal) SetCausal(ctx ml.Context, opts CausalOptions) {
+	if !slices.Equal(c.opts.Except, opts.Except) {
+		c.opts = opts
 		if ctx != nil {
 			var err error
 			c.curMask, err = c.buildMask(ctx)

kvcache/causal_test.go

@@ -441,11 +441,19 @@ func (t *testTensor) Scale(ctx ml.Context, s float64) ml.Tensor {
 	panic("not implemented")
 }
 
+func (t *testTensor) AvgPool1D(ctx ml.Context, k, s, p int) ml.Tensor {
+	panic("not implemented")
+}
+
+func (t *testTensor) AvgPool2D(ctx ml.Context, k, s int, p float32) ml.Tensor {
+	panic("not implemented")
+}
+
 func (t *testTensor) Conv2D(ctx ml.Context, weight ml.Tensor, s0, s1, p0, p1, d0, d1 int) ml.Tensor {
 	panic("not implemented")
 }
 
-func (t *testTensor) RoPE(ctx ml.Context, positionIDs, ropeFactors ml.Tensor, dim uint32, base, scale float32) ml.Tensor {
+func (t *testTensor) RoPE(ctx ml.Context, positionIDs, ropeFactors ml.Tensor, dim, ropeType uint32, base, scale float32) ml.Tensor {
 	panic("not implemented")
 }
 
@@ -495,6 +503,10 @@ func (t *testTensor) Contiguous(ctx ml.Context) ml.Tensor {
 	panic("not implemented")
 }
 
+func (t *testTensor) Set(ctx ml.Context, t2 ml.Tensor, offset int, strides ...int) ml.Tensor {
+	panic("not implemented")
+}
+
 func (t *testTensor) Pad(ctx ml.Context, shape ...int) ml.Tensor {
 	panic("not implemented")
 }

llama/patches/0020-ollama-debug-tensor.patch

@@ -0,0 +1,33 @@
+From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
+From: Michael Yang <mxyng@pm.me>
+Date: Sun, 9 Mar 2025 14:44:16 -0700
+Subject: [PATCH] ollama debug tensor
+
+---
+ ggml/src/ggml-cpu/ggml-cpu.c | 6 ++++++
+ 1 file changed, 6 insertions(+)
+
+diff --git a/ggml/src/ggml-cpu/ggml-cpu.c b/ggml/src/ggml-cpu/ggml-cpu.c
+index 2f606d82..ec60e8fc 100644
+--- a/ggml/src/ggml-cpu/ggml-cpu.c
++++ b/ggml/src/ggml-cpu/ggml-cpu.c
+@@ -11,6 +11,8 @@
+ #include "ggml-threading.h"
+ #include "ggml.h"
+ 
++#include "ollama-debug.h"
++
+ #if defined(_MSC_VER) || defined(__MINGW32__)
+ #include <malloc.h> // using malloc.h with MSC/MINGW
+ #elif !defined(__FreeBSD__) && !defined(__NetBSD__) && !defined(__OpenBSD__)
+@@ -14103,6 +14105,10 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
+ 
+         ggml_compute_forward(&params, node);
+ 
++#ifdef OLLAMA_DEBUG
++        ollama_debug(node, true);
++#endif
++
+         if (state->ith == 0 && cplan->abort_callback &&
+                 cplan->abort_callback(cplan->abort_callback_data)) {
+             atomic_store_explicit(&tp->abort, node_n + 1, memory_order_relaxed);

llm/memory.go

@@ -218,8 +218,8 @@ func EstimateGPULayers(gpus []discover.GpuInfo, f *ggml.GGML, projectors []strin
 		if blk, ok := layers[fmt.Sprintf("blk.%d", i)]; ok {
 			layerSize = blk.Size()
 			layerSize += kv / f.KV().BlockCount()
+			memoryWeights += blk.Size()
 		}
-		memoryWeights += layerSize
 
 		if opts.NumGPU >= 0 && layerCount >= opts.NumGPU {
 			// Stop allocating on GPU(s) once we hit the users target NumGPU
@@ -376,7 +376,7 @@ func (m MemoryEstimate) LogValue() slog.Value {
 				// memory of the weights
 				"total", format.HumanBytes2(m.memoryWeights),
 				// memory of repeating layers
-				"repeating", format.HumanBytes2(m.memoryWeights-m.memoryLayerOutput),
+				"repeating", format.HumanBytes2(m.memoryWeights),
 				// memory of non-repeating layers
 				"nonrepeating", format.HumanBytes2(m.memoryLayerOutput),
 			),

llm/server.go

@@ -271,7 +271,7 @@ func NewLlamaServer(gpus discover.GpuInfoList, modelPath string, f *ggml.GGML, a
 
 	var llamaModel *llama.Model
 	var textProcessor model.TextProcessor
-	if envconfig.NewEngine() {
+	if envconfig.NewEngine() || f.KV().OllamaEngineRequired() {
 		textProcessor, err = model.NewTextProcessor(modelPath)
 		if err != nil {
 			// To prepare for opt-out mode, instead of treating this as an error, we fallback to the old runner

ml/backend.go

@@ -5,6 +5,7 @@ import (
 	"encoding/binary"
 	"fmt"
 	"os"
+	"slices"
 	"strconv"
 	"strings"
 )
@@ -18,6 +19,7 @@ type Config interface {
 
 	Strings(string, ...[]string) []string
 	Uints(string, ...[]uint32) []uint32
+	Floats(string, ...[]float32) []float32
 }
 
 type Backend interface {
@@ -133,8 +135,10 @@ type Tensor interface {
 	RMSNorm(ctx Context, weight Tensor, eps float32) Tensor
 	Scale(ctx Context, s float64) Tensor
 
+	AvgPool2D(ctx Context, k, s int, p float32) Tensor
 	Conv2D(ctx Context, weight Tensor, s0, s1, p0, p1, d0, d1 int) Tensor
-	RoPE(ctx Context, positionIDs, ropeFactors Tensor, dim uint32, base, scale float32) Tensor
+
+	RoPE(ctx Context, positionIDs, ropeFactors Tensor, dim, ropeType uint32, base, scale float32) Tensor
 
 	Tanh(ctx Context) Tensor
 	GELU(ctx Context) Tensor
@@ -144,6 +148,7 @@ type Tensor interface {
 	View(ctx Context, offset int, shape ...int) Tensor
 	Permute(ctx Context, shape ...int) Tensor
 	Contiguous(ctx Context) Tensor
+	Set(ctx Context, t2 Tensor, offset int, strides ...int) Tensor
 
 	Pad(ctx Context, shape ...int) Tensor
 	Unpad(ctx Context, shape ...int) Tensor
@@ -241,16 +246,17 @@ func dump[S ~[]E, E number](ctx Context, t Tensor, items int, fn func(E) string)
 	}
 
 	shape := t.Shape()
+	slices.Reverse(shape)
 
 	var sb strings.Builder
 	var f func([]int, int)
 	f = func(dims []int, stride int) {
 		prefix := strings.Repeat(" ", len(shape)-len(dims)+1)
-		fmt.Fprint(&sb, "[")
-		defer func() { fmt.Fprint(&sb, "]") }()
+		sb.WriteString("[")
+		defer func() { sb.WriteString("]") }()
 		for i := 0; i < dims[0]; i++ {
 			if i >= items && i < dims[0]-items {
-				fmt.Fprint(&sb, "..., ")
+				sb.WriteString("..., ")
 				// skip to next printable element
 				skip := dims[0] - 2*items
 				if len(dims) > 1 {
@@ -265,9 +271,14 @@ func dump[S ~[]E, E number](ctx Context, t Tensor, items int, fn func(E) string)
 					fmt.Fprint(&sb, ",", strings.Repeat("\n", len(dims)-1), prefix)
 				}
 			} else {
-				fmt.Fprint(&sb, fn(s[stride+i]))
+				text := fn(s[stride+i])
+				if len(text) > 0 && text[0] != '-' {
+					sb.WriteString(" ")
+				}
+
+				sb.WriteString(text)
 				if i < dims[0]-1 {
-					fmt.Fprint(&sb, ", ")
+					sb.WriteString(", ")
 				}
 			}
 		}

ml/backend/ggml/ggml.go

@@ -240,11 +240,22 @@ func New(r *os.File, params ml.BackendParams) (ml.Backend, error) {
 		switch {
 		case contains(t.Name, "position_embd", "token_embd", "token_norm_embd", "token_types"):
 			createTensor(tensor{source: t}, input.bts)
+			if _, ok := meta.Tensors().GroupLayers()["output"]; !ok && t.Name == "token_embd.weight" {
+				createTensor(tensor{source: t, target: "output.weight"}, output.bts)
+			}
 		case contains(t.Name, "cls", "output", "output_norm"):
 			createTensor(tensor{source: t}, output.bts)
 		case strings.HasPrefix(t.Name, "v.") || strings.HasPrefix(t.Name, "mm."):
 			// TODO: assign vision tensors to the gpu if possible
-			createTensor(tensor{source: t}, input.bts)
+			createTensor(tensor{source: t}, output.bts)
+		case contains(t.Name, "rope_freqs", "rope_factors_long", "rope_factors_short"):
+			// these tensors should be repeated per layer
+			for i, layer := range layers {
+				createTensor(tensor{
+					source: t,
+					target: "blk." + strconv.Itoa(i) + "." + t.Name,
+				}, layer.bts)
+			}
 		default:
 			layerIndex := -1
 			if fields := strings.FieldsFunc(t.Name, func(r rune) bool { return !unicode.IsNumber(r) }); len(fields) > 0 {
@@ -256,14 +267,8 @@ func New(r *os.File, params ml.BackendParams) (ml.Backend, error) {
 			if layerIndex >= 0 {
 				createTensor(tensor{source: t}, layers[layerIndex].bts)
 			} else {
-				// this is a repeating tensor that doesn't explicitly associated with a layer so
-				// duplicate it for each layer
-				for i, layer := range layers {
-					createTensor(tensor{
-						source: t,
-						target: "blk." + strconv.Itoa(i) + "." + t.Name,
-					}, layer.bts)
-				}
+				// load all other tensors on the cpu
+				createTensor(tensor{source: t}, input.bts)
 			}
 		}
 	}
@@ -352,7 +357,7 @@ func New(r *os.File, params ml.BackendParams) (ml.Backend, error) {
 
 		if C.ggml_backend_is_cpu(b) {
 			// set number of threads for cpu backend
-			C.ggml_backend_cpu_set_n_threads(b, C.int(params.NumThreads))
+			C.ggml_backend_cpu_set_n_threads(b, C.int(Threads(params.NumThreads)))
 		}
 	}
 
@@ -893,10 +898,13 @@ func (t *Tensor) View(ctx ml.Context, offset int, shape ...int) ml.Tensor {
 }
 
 const (
-	ropeTypeNorm C.int = iota
+	ropeTypeNorm   C.int = 0
+	ropeTypeNeox   C.int = 2
+	ropeTypeMrope  C.int = 8
+	ropeTypeVision C.int = 24
 )
 
-func (t *Tensor) RoPE(ctx ml.Context, positionIDs, ropeFactors ml.Tensor, ropeDim uint32, ropeBase, ropeScale float32) ml.Tensor {
+func (t *Tensor) RoPE(ctx ml.Context, positionIDs, ropeFactors ml.Tensor, ropeDim, ropeType uint32, ropeBase, ropeScale float32) ml.Tensor {
 	if ropeFactors == nil {
 		ropeFactors = &Tensor{b: t.b}
 	}
@@ -911,8 +919,8 @@ func (t *Tensor) RoPE(ctx ml.Context, positionIDs, ropeFactors ml.Tensor, ropeDi
 		t: C.ggml_rope_ext(
 			ctx.(*Context).ctx, dequant, positionIDs.(*Tensor).t, ropeFactors.(*Tensor).t,
 			C.int(ropeDim),
-			131072,       // YaRN n_ctx_train
-			ropeTypeNorm, // ROPE_TYPE_NORM
+			C.int(ropeType),
+			131072, // YaRN n_ctx_train
 			C.float(ropeBase),
 			C.float(ropeScale),
 			0.,  // YaRN ext_factor
@@ -944,6 +952,27 @@ func (t *Tensor) Conv2D(ctx ml.Context, t2 ml.Tensor, s0, s1, p0, p1, d0, d1 int
 	}
 }
 
+func (t *Tensor) AvgPool2D(ctx ml.Context, k, s int, p float32) ml.Tensor {
+	return &Tensor{
+		b: t.b,
+		t: C.ggml_pool_2d(ctx.(*Context).ctx, t.t, C.GGML_OP_POOL_AVG, C.int(k), C.int(k), C.int(s), C.int(s), C.float(p), C.float(p)),
+	}
+}
+
+func (t *Tensor) Set(ctx ml.Context, t2 ml.Tensor, offset int, strides ...int) ml.Tensor {
+	var tt *C.struct_ggml_tensor
+	switch len(strides) {
+	case 0:
+		tt = C.ggml_set_1d(ctx.(*Context).ctx, t.t, t2.(*Tensor).t, C.size_t(offset))
+	case 1:
+		tt = C.ggml_set_2d(ctx.(*Context).ctx, t.t, t2.(*Tensor).t, C.size_t(offset), C.size_t(strides[0]))
+	default:
+		panic("unsupported number of dimensions")
+	}
+
+	return &Tensor{b: t.b, t: tt}
+}
+
 func (t *Tensor) ScaledDotProductAttention(ctx ml.Context, key, value, mask ml.Tensor, scale float64) ml.Tensor {
 	var kqMask *C.struct_ggml_tensor
 	if mask != nil {

ml/backend/ggml/ggml/include/ollama-debug.h

@@ -0,0 +1,11 @@
+#include "ggml.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void ollama_debug(const struct ggml_tensor *tensor, bool verbose);
+
+#ifdef __cplusplus
+}
+#endif

ml/backend/ggml/ggml/src/ggml-cpu/cpu_debug.go

@@ -0,0 +1,6 @@
+//go:build debug
+
+package cpu
+
+// #cgo CPPFLAGS: -DOLLAMA_DEBUG
+import "C"

ml/backend/ggml/ggml/src/ggml-cpu/ggml-cpu.c

@@ -11,6 +11,8 @@
 #include "ggml-threading.h"
 #include "ggml.h"
 
+#include "ollama-debug.h"
+
 #if defined(_MSC_VER) || defined(__MINGW32__)
 #include <malloc.h> // using malloc.h with MSC/MINGW
 #elif !defined(__FreeBSD__) && !defined(__NetBSD__) && !defined(__OpenBSD__)
@@ -14103,6 +14105,10 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
 
         ggml_compute_forward(&params, node);
 
+#ifdef OLLAMA_DEBUG
+        ollama_debug(node, true);
+#endif
+
         if (state->ith == 0 && cplan->abort_callback &&
                 cplan->abort_callback(cplan->abort_callback_data)) {
             atomic_store_explicit(&tp->abort, node_n + 1, memory_order_relaxed);

ml/backend/ggml/ggml/src/ollama-debug.c

@@ -0,0 +1,116 @@
+#include <string.h>
+#include <inttypes.h>
+
+#include "ollama-debug.h"
+
+static int mul(int64_t *dims, int ndims) {
+    int result = 1;
+    for (int i = 0; i < ndims; i++) {
+        result *= dims[i];
+    }
+
+    return result;
+}
+
+static void repeat(char c, int n) {
+    for (int i = 0; i < n; i++) {
+        fprintf(stderr, "%c", c);
+    }
+}
+
+static void print_tensor(const void *tensor, void (*cb)(const void *, int),
+                         int shape,
+                         int64_t *dims, int ndims, int stride,
+                         int nitems, int pad) {
+    fprintf(stderr, "[");
+    for (int i = 0; i < dims[0]; i++) {
+        if (i >= nitems && i < dims[0] - nitems) {
+            fprintf(stderr, "... (%" PRIi64 " more), ", dims[0] - 2 * nitems);
+            int skip = dims[0] - 2 * nitems;
+            if (ndims > 1) {
+                stride += mul(dims + 1, ndims - 1) * skip;
+                repeat('\n', ndims - 1);
+                repeat(' ', shape - ndims + 1 + pad);
+            }
+            i += skip - 1;
+        } else if (ndims > 1) {
+            print_tensor(tensor, cb, shape, dims + 1, ndims - 1, stride,
+                         nitems, pad);
+            stride += mul(dims + 1, ndims - 1);
+            if (i < dims[0] - 1) {
+                fprintf(stderr, ", ");
+                repeat('\n', ndims - 1);
+                repeat(' ', shape - ndims + 1 + pad);
+            }
+        } else {
+            cb(tensor, stride + i);
+            if (i < dims[0] - 1) {
+                fprintf(stderr, ", ");
+            }
+        }
+    }
+    fprintf(stderr, "]");
+}
+
+static void print_tensor_f16(const void *tensor, int i) {
+    float value = ggml_fp16_to_fp32(((const ggml_fp16_t *)tensor)[i]);
+    fprintf(stderr, "%s%f", value < 0 ? "" : " ", value);
+}
+
+static void print_tensor_f32(const void *tensor, int i) {
+    float value = ((const float *)tensor)[i];
+    fprintf(stderr, "%s%f", value < 0 ? "" : " ", value);
+}
+
+static void print_tensor_i32(const void *tensor, int i) {
+    int32_t value = ((const int32_t *)tensor)[i];
+    fprintf(stderr, "%s%d", value < 0 ? "" : " ", value);
+}
+
+static void ollama_debug_tensor(const struct ggml_tensor *tensor, bool verbose, const char *prefix, int indent) {
+    fprintf(stderr, "%s%s %s (%s): [%" PRIi64 " %" PRIi64 " %" PRIi64 " %" PRIi64 "]\n", prefix, tensor->name,
+            ggml_op_name(tensor->op), ggml_type_name(tensor->type), tensor->ne[0],
+            tensor->ne[1], tensor->ne[2], tensor->ne[3]);
+
+    if (!verbose) {
+        return;
+    }
+
+    for (int i = 0; i < indent; i++) {
+        fprintf(stderr, " ");
+    }
+
+    switch (tensor->type) {
+    case GGML_TYPE_F16:
+        print_tensor(ggml_get_data(tensor), print_tensor_f16, ggml_n_dims(tensor),
+                     (int64_t *)tensor->ne, ggml_n_dims(tensor), 0, 3, indent);
+        break;
+    case GGML_TYPE_F32:
+        print_tensor(ggml_get_data(tensor), print_tensor_f32, ggml_n_dims(tensor),
+                     (int64_t *)tensor->ne, ggml_n_dims(tensor), 0, 3, indent);
+        break;
+    case GGML_TYPE_I32:
+        print_tensor(ggml_get_data(tensor), print_tensor_i32, ggml_n_dims(tensor),
+                     (int64_t *)tensor->ne, ggml_n_dims(tensor), 0, 3, indent);
+        break;
+    default:
+        fprintf(stderr, "<unsupported type>\n");
+        return;
+    }
+
+    fprintf(stderr, "\n");
+}
+
+void ollama_debug(const struct ggml_tensor *tensor, bool verbose) {
+    ollama_debug_tensor(tensor, verbose, ">>> ", 4);
+
+    for (int i = 0; i < GGML_MAX_SRC && tensor->src[i] != NULL; ++i) {
+        char src[8];
+        const int n = snprintf(src, sizeof(src), " src%d ", i);
+        if (n >= sizeof(src)) {
+            src[sizeof(src) - 1] = '\0';
+        }
+
+        ollama_debug_tensor(tensor->src[i], verbose, src, 4);
+    }
+}

ml/backend/ggml/threads.go

@@ -0,0 +1,7 @@
+//go:build !debug
+
+package ggml
+
+func Threads(n int) int {
+	return n
+}

ml/backend/ggml/threads_debug.go

@@ -0,0 +1,7 @@
+//go:build debug
+
+package ggml
+
+func Threads(_ int) int {
+	return 1
+}

model/model.go

@@ -22,6 +22,8 @@ import (
 	"github.com/ollama/ollama/model/input"
 )
 
+var ErrNoVisionModel = errors.New("this model is missing data required for image input")
+
 // Model implements a specific model architecture, defining the forward pass and any model-specific configuration
 type Model interface {
 	Forward(ml.Context, input.Options) (ml.Tensor, error)

model/models/gemma2/model.go

@@ -0,0 +1,220 @@
+package gemma2
+
+import (
+	"math"
+
+	"github.com/ollama/ollama/kvcache"
+	"github.com/ollama/ollama/ml"
+	"github.com/ollama/ollama/ml/nn"
+	"github.com/ollama/ollama/model"
+	"github.com/ollama/ollama/model/input"
+)
+
+type Options struct {
+	hiddenSize, numHeads, numKVHeads int
+	attnKeyLen, attnValLen           int
+	eps, ropeBase, ropeScale         float32
+	attnLogitSoftcap                 float32
+	finalLogitSoftcap                float32
+	largeModelScaling                bool
+}
+
+type Model struct {
+	model.Base
+	model.SentencePieceModel
+
+	TokenEmbedding *nn.Embedding `gguf:"token_embd"`
+	Layers         []Layer       `gguf:"blk"`
+	OutputNorm     *nn.RMSNorm   `gguf:"output_norm"`
+	Output         *nn.Linear    `gguf:"output,alt:token_embd"` // just set to token_embd?
+
+	*Options
+}
+
+const (
+	gemma27BLayerCount = 46
+)
+
+func New(c ml.Config) (model.Model, error) {
+	m := Model{
+		SentencePieceModel: model.NewSentencePieceModel(
+			c.String("tokenizer.ggml.pretokenizer", `(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\r\n\p{L}\p{N}]?\p{L}+|\p{N}{1,3}| ?[^\s\p{L}\p{N}]+[\r\n]*|\s*[\r\n]+|\s+(?!\S)|\s+`),
+			&model.Vocabulary{
+				Values: c.Strings("tokenizer.ggml.tokens"),
+				Scores: c.Floats("tokenizer.ggml.scores"),
+				Types:  c.Uints("tokenizer.ggml.token_type"),
+				BOS:    int32(c.Uint("tokenizer.ggml.bos_token_id")),
+				EOS:    int32(c.Uint("tokenizer.ggml.eos_token_id")),
+			},
+		),
+		Layers: make([]Layer, c.Uint("block_count")),
+		Options: &Options{
+			hiddenSize:        int(c.Uint("embedding_length")),
+			numHeads:          int(c.Uint("attention.head_count")),
+			numKVHeads:        int(c.Uint("attention.head_count_kv")),
+			attnKeyLen:        int(c.Uint("attention.key_length")),
+			attnValLen:        int(c.Uint("attention.value_length")),
+			eps:               c.Float("attention.layer_norm_rms_epsilon"),
+			ropeBase:          c.Float("rope.freq_base", 10000.0),
+			ropeScale:         c.Float("rope.freq_scale", 1.0),
+			attnLogitSoftcap:  c.Float("attn_logit_softcapping"),
+			finalLogitSoftcap: c.Float("final_logit_softcapping"),
+		},
+	}
+
+	slidingWindowLen := int32(c.Uint("attention.sliding_window"))
+	m.Cache = kvcache.NewWrapperCache(kvcache.NewSWACache(slidingWindowLen, m.Shift), kvcache.NewCausalCache(m.Shift))
+	m.Cache.SetConfig(ml.CacheConfig{})
+
+	return &m, nil
+}
+
+type SelfAttention struct {
+	Query  *nn.Linear `gguf:"attn_q"`
+	Key    *nn.Linear `gguf:"attn_k"`
+	Value  *nn.Linear `gguf:"attn_v"`
+	Output *nn.Linear `gguf:"attn_output"`
+}
+
+func (sa *SelfAttention) Forward(ctx ml.Context, hiddenState, positionIDs ml.Tensor, cache kvcache.Cache, opts *Options) ml.Tensor {
+	batchSize := hiddenState.Dim(1)
+	ropeType := uint32(2)
+
+	q := sa.Query.Forward(ctx, hiddenState)
+	q = q.Reshape(ctx, opts.attnKeyLen, opts.numHeads, batchSize)
+	q = q.RoPE(ctx, positionIDs, nil, uint32(opts.attnKeyLen), ropeType, opts.ropeBase, opts.ropeScale)
+
+	if opts.largeModelScaling {
+		q = q.Scale(ctx, 1.0/math.Sqrt(float64(opts.hiddenSize/opts.numHeads)))
+	} else {
+		q = q.Scale(ctx, 1.0/math.Sqrt(float64(opts.attnKeyLen)))
+	}
+
+	k := sa.Key.Forward(ctx, hiddenState)
+	k = k.Reshape(ctx, opts.attnKeyLen, opts.numKVHeads, batchSize)
+	k = k.RoPE(ctx, positionIDs, nil, uint32(opts.attnKeyLen), ropeType, opts.ropeBase, opts.ropeScale)
+
+	v := sa.Value.Forward(ctx, hiddenState)
+	v = v.Reshape(ctx, opts.attnValLen, opts.numKVHeads, batchSize)
+
+	cache.Put(ctx, k, v)
+	k, v, mask := cache.Get(ctx)
+
+	q = q.Permute(ctx, 0, 2, 1, 3)
+	k = k.Permute(ctx, 0, 2, 1, 3)
+	v = v.Permute(ctx, 1, 2, 0, 3).Contiguous(ctx)
+
+	kq := k.Mulmat(ctx, q)
+
+	// logit softcap
+	kq = kq.Scale(ctx, 1.0/float64(opts.attnLogitSoftcap))
+	kq = kq.Tanh(ctx)
+	kq = kq.Scale(ctx, float64(opts.attnLogitSoftcap))
+
+	kq = kq.Add(ctx, mask)
+	kq = kq.Softmax(ctx)
+
+	kqv := v.Mulmat(ctx, kq)
+	kqv = kqv.Permute(ctx, 0, 2, 1, 3).Contiguous(ctx)
+	kqv = kqv.Reshape(ctx, opts.attnValLen*opts.numHeads, batchSize)
+
+	return sa.Output.Forward(ctx, kqv)
+}
+
+func (m *Model) Shift(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error) {
+	return key.RoPE(ctx, shift, nil, uint32(m.Options.attnKeyLen), uint32(2), m.Options.ropeBase, m.Options.ropeScale), nil
+}
+
+type MLP struct {
+	Up   *nn.Linear `gguf:"ffn_up"`
+	Down *nn.Linear `gguf:"ffn_down"`
+	Gate *nn.Linear `gguf:"ffn_gate"`
+}
+
+func (mlp *MLP) Forward(ctx ml.Context, hiddenState ml.Tensor, opts *Options) ml.Tensor {
+	hiddenState = mlp.Gate.Forward(ctx, hiddenState).GELU(ctx).Mul(ctx, mlp.Up.Forward(ctx, hiddenState))
+	return mlp.Down.Forward(ctx, hiddenState)
+}
+
+type Layer struct {
+	AttentionNorm     *nn.RMSNorm `gguf:"attn_norm"`
+	SelfAttention     *SelfAttention
+	PostAttentionNorm *nn.RMSNorm `gguf:"post_attention_norm"`
+	MLPNorm           *nn.RMSNorm `gguf:"ffn_norm"`
+	MLP               *MLP
+	PostMLPNorm       *nn.RMSNorm `gguf:"post_ffw_norm"`
+}
+
+func (l *Layer) Forward(ctx ml.Context, hiddenState, positionIDs, outputs ml.Tensor, cache kvcache.Cache, opts *Options) ml.Tensor {
+	residual := hiddenState
+
+	hiddenState = l.AttentionNorm.Forward(ctx, hiddenState, opts.eps)
+	hiddenState = l.SelfAttention.Forward(ctx, hiddenState, positionIDs, cache, opts)
+	hiddenState = l.PostAttentionNorm.Forward(ctx, hiddenState, opts.eps)
+
+	// In the final layer (outputs != nil), optimize by pruning to just the token positions
+	// we need logits for.
+	if outputs != nil {
+		hiddenState = hiddenState.Rows(ctx, outputs)
+		residual = residual.Rows(ctx, outputs)
+	}
+
+	hiddenState = hiddenState.Add(ctx, residual)
+	residual = hiddenState
+
+	hiddenState = l.MLPNorm.Forward(ctx, hiddenState, opts.eps)
+	hiddenState = l.MLP.Forward(ctx, hiddenState, opts)
+	hiddenState = l.PostMLPNorm.Forward(ctx, hiddenState, opts.eps)
+	return hiddenState.Add(ctx, residual)
+}
+
+func (m *Model) Forward(ctx ml.Context, opts input.Options) (ml.Tensor, error) {
+	inputs, err := ctx.Input().FromIntSlice(opts.Inputs, len(opts.Inputs))
+	if err != nil {
+		return nil, err
+	}
+
+	positions, err := ctx.Input().FromIntSlice(opts.Positions, len(opts.Positions))
+	if err != nil {
+		return nil, err
+	}
+
+	outputs, err := ctx.Output().FromIntSlice(opts.Outputs, len(opts.Outputs))
+	if err != nil {
+		return nil, err
+	}
+
+	hiddenState := m.TokenEmbedding.Forward(ctx, inputs)
+	hiddenState = hiddenState.Scale(ctx, math.Sqrt(float64(m.Options.hiddenSize)))
+
+	if len(m.Layers) == gemma27BLayerCount {
+		m.Options.largeModelScaling = true
+	}
+
+	for i, layer := range m.Layers {
+		cacheType := i % 2
+		m.Cache.SetLayer(i)
+		wc := m.Cache.(*kvcache.WrapperCache)
+		wc.SetLayerType(cacheType)
+
+		var lastLayerOutputs ml.Tensor
+		if i == len(m.Layers)-1 {
+			lastLayerOutputs = outputs
+		}
+
+		hiddenState = layer.Forward(ctx, hiddenState, positions, lastLayerOutputs, m.Cache, m.Options)
+	}
+
+	hiddenState = m.OutputNorm.Forward(ctx, hiddenState, m.eps)
+	hiddenState = m.Output.Forward(ctx, hiddenState)
+
+	// final logit softcap
+	hiddenState = hiddenState.Scale(ctx, 1.0/float64(m.Options.finalLogitSoftcap))
+	hiddenState = hiddenState.Tanh(ctx)
+	hiddenState = hiddenState.Scale(ctx, float64(m.Options.finalLogitSoftcap))
+	return hiddenState.Rows(ctx, outputs), nil
+}
+
+func init() {
+	model.Register("gemma2", New)
+}

model/models/gemma3/model.go

@@ -0,0 +1,177 @@
+package gemma3
+
+import (
+	"bytes"
+	"encoding/binary"
+	"hash/fnv"
+	"image"
+	"math"
+
+	"github.com/ollama/ollama/kvcache"
+	"github.com/ollama/ollama/ml"
+	"github.com/ollama/ollama/ml/nn"
+	"github.com/ollama/ollama/model"
+	"github.com/ollama/ollama/model/input"
+)
+
+type Model struct {
+	model.Base
+	model.SentencePieceModel
+
+	*VisionModel `gguf:"v,vision"`
+	*TextModel
+
+	*MultiModalProjector `gguf:"mm"`
+
+	ImageProcessor
+}
+
+var _ model.MultimodalProcessor = (*Model)(nil)
+
+type MultiModalProjector struct {
+	SoftEmbNorm     *nn.RMSNorm `gguf:"mm_soft_emb_norm"`
+	InputProjection *nn.Linear  `gguf:"mm_input_projection"`
+
+	tokensPerImage int
+}
+
+func (p *MultiModalProjector) Forward(ctx ml.Context, visionOutputs ml.Tensor, imageSize, patchSize int, eps float32) ml.Tensor {
+	l := visionOutputs.Dim(0)
+
+	visionOutputs = visionOutputs.Permute(ctx, 1, 0, 2, 3).Contiguous(ctx)
+	patchesPerImage := imageSize / patchSize
+	visionOutputs = visionOutputs.Reshape(ctx, patchesPerImage, patchesPerImage, l)
+
+	kernelSize := patchesPerImage / int(math.Sqrt(float64(p.tokensPerImage)))
+	visionOutputs = visionOutputs.AvgPool2D(ctx, kernelSize, kernelSize, 0)
+	visionOutputs = visionOutputs.Reshape(ctx, visionOutputs.Dim(0)*visionOutputs.Dim(1), l)
+	visionOutputs = visionOutputs.Permute(ctx, 1, 0, 2, 3).Contiguous(ctx)
+	visionOutputs = p.SoftEmbNorm.Forward(ctx, visionOutputs, eps)
+
+	// TODO: inputProjection must be transposed since they're incompatible with visionOutputs
+	visionOutputs = p.InputProjection.Weight.Permute(ctx, 1, 0, 2, 3).Contiguous(ctx).Mulmat(ctx, visionOutputs)
+	return visionOutputs
+}
+
+func New(c ml.Config) (model.Model, error) {
+	m := Model{
+		SentencePieceModel: model.NewSentencePieceModel(
+			c.String("tokenizer.ggml.pretokenizer", `(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\r\n\p{L}\p{N}]?\p{L}+|\p{N}{1,3}| ?[^\s\p{L}\p{N}]+[\r\n]*|\s*[\r\n]+|\s+(?!\S)|\s+`),
+			&model.Vocabulary{
+				Values: c.Strings("tokenizer.ggml.tokens"),
+				Scores: c.Floats("tokenizer.ggml.scores"),
+				Types:  c.Uints("tokenizer.ggml.token_type"),
+				BOS:    int32(c.Uint("tokenizer.ggml.bos_token_id")),
+				AddBOS: c.Bool("tokenizer.ggml.add_bos_token", true),
+				EOS:    int32(1),
+				AddEOS: c.Bool("tokenizer.ggml.add_eos_token", false),
+				EOT:    int32(106),
+				AddEOT: c.Bool("tokenizer.ggml.add_eot_token", false),
+			},
+		),
+		ImageProcessor: newImageProcessor(c),
+		VisionModel:    newVisionModel(c),
+		TextModel:      newTextModel(c),
+		MultiModalProjector: &MultiModalProjector{
+			tokensPerImage: int(c.Uint("mm_tokens_per_image", 256)),
+		},
+	}
+
+	slidingWindowLen := int32(c.Uint("attention.sliding_window"))
+	m.Cache = kvcache.NewWrapperCache(kvcache.NewSWACache(slidingWindowLen, m.Shift), kvcache.NewCausalCache(m.Shift))
+
+	return &m, nil
+}
+
+func (m *Model) EncodeMultimodal(ctx ml.Context, multimodalData []byte) (any, error) {
+	if len(m.VisionModel.Layers) == 0 {
+		return nil, model.ErrNoVisionModel
+	}
+
+	image, _, err := image.Decode(bytes.NewReader(multimodalData))
+	if err != nil {
+		return nil, err
+	}
+
+	f32s, err := m.ImageProcessor.ProcessImage(image)
+	if err != nil {
+		return nil, err
+	}
+
+	pixelValues, err := ctx.Input().FromFloatSlice(f32s,
+		m.ImageProcessor.imageSize,
+		m.ImageProcessor.imageSize,
+		m.ImageProcessor.numChannels,
+	)
+	if err != nil {
+		return nil, err
+	}
+
+	visionOutputs := m.VisionModel.Forward(ctx, pixelValues)
+	visionOutputs = m.MultiModalProjector.Forward(ctx, visionOutputs, m.imageSize, m.patchSize, m.VisionModel.eps)
+	return visionOutputs, nil
+}
+
+type imageToken struct {
+	embedding ml.Tensor
+	index     int
+}
+
+func (m *Model) PostTokenize(ctx ml.Context, inputs []input.Input) ([]input.Input, error) {
+	var result []input.Input
+	fnvHash := fnv.New64a()
+
+	for _, inp := range inputs {
+		if inp.Multimodal == nil {
+			result = append(result, inp)
+		} else {
+			imageInputs := []input.Input{
+				{Token: 108},    // "\n\n"
+				{Token: 255999}, // "<start_of_image>""
+			}
+			result = append(result, imageInputs...)
+
+			// add image embeddings
+			inputMultimodal := inp.Multimodal.(ml.Tensor)
+
+			for i := range inputMultimodal.Dim(1) {
+				fnvHash.Reset()
+				binary.Write(fnvHash, binary.NativeEndian, inp.MultimodalHash)
+				fnvHash.Write([]byte{byte(i)})
+
+				imageToken := imageToken{embedding: inputMultimodal, index: i}
+				result = append(result, input.Input{Multimodal: imageToken, MultimodalHash: fnvHash.Sum64()})
+			}
+
+			result = append(result,
+				input.Input{Token: 256000}, // <end_of_image>
+				input.Input{Token: 108},    // "\n\n"
+			)
+		}
+	}
+
+	return result, nil
+}
+
+func (m *Model) Forward(ctx ml.Context, opts input.Options) (ml.Tensor, error) {
+	inputs, err := ctx.Input().FromIntSlice(opts.Inputs, len(opts.Inputs))
+	if err != nil {
+		return nil, err
+	}
+
+	positions, err := ctx.Input().FromIntSlice(opts.Positions, len(opts.Positions))
+	if err != nil {
+		return nil, err
+	}
+
+	outputs, err := ctx.Output().FromIntSlice(opts.Outputs, len(opts.Outputs))
+	if err != nil {
+		return nil, err
+	}
+
+	return m.TextModel.Forward(ctx, inputs, positions, outputs, opts, m.Cache), nil
+}
+
+func init() {
+	model.Register("gemma3", New)
+}

model/models/gemma3/model_text.go

@@ -0,0 +1,247 @@
+package gemma3
+
+import (
+	"math"
+
+	"github.com/ollama/ollama/kvcache"
+	"github.com/ollama/ollama/ml"
+	"github.com/ollama/ollama/ml/nn"
+	"github.com/ollama/ollama/model"
+	"github.com/ollama/ollama/model/input"
+)
+
+type TextOptions struct {
+	hiddenSize, numHeads, numKVHeads int
+	attnKeyLen, attnValLen           int
+	eps, ropeScale                   float32
+	ropeLocalBase, ropeGlobalBase    float32
+	largeModelScaling                bool
+}
+
+type TextModel struct {
+	model.Base
+	model.SentencePieceModel
+
+	TokenEmbedding *nn.Embedding `gguf:"token_embd"`
+	Layers         []TextLayer   `gguf:"blk"`
+	OutputNorm     *nn.RMSNorm   `gguf:"output_norm"`
+	Output         *nn.Linear    `gguf:"output,alt:token_embd"`
+
+	*TextOptions
+}
+
+const (
+	gemmaGlobalCacheCount = 6
+	gemma27BLayerCount    = 62
+)
+
+const (
+	cacheTypeSWA = iota
+	cacheTypeCausal
+)
+
+func newTextModel(c ml.Config) *TextModel {
+	numBlocks := int(c.Uint("block_count"))
+
+	m := TextModel{
+		SentencePieceModel: model.NewSentencePieceModel(
+			c.String("tokenizer.ggml.pretokenizer", `(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\r\n\p{L}\p{N}]?\p{L}+|\p{N}{1,3}| ?[^\s\p{L}\p{N}]+[\r\n]*|\s*[\r\n]+|\s+(?!\S)|\s+`),
+			&model.Vocabulary{
+				Values: c.Strings("tokenizer.ggml.tokens"),
+				Scores: c.Floats("tokenizer.ggml.scores"),
+				Types:  c.Uints("tokenizer.ggml.token_type"),
+				BOS:    int32(c.Uint("tokenizer.ggml.bos_token_id")),
+				EOS:    int32(c.Uint("tokenizer.ggml.eos_token_id")),
+			},
+		),
+		Layers: make([]TextLayer, numBlocks),
+		TextOptions: &TextOptions{
+			hiddenSize:     int(c.Uint("embedding_length")),
+			numHeads:       int(c.Uint("attention.head_count")),
+			numKVHeads:     int(c.Uint("attention.head_count_kv")),
+			attnKeyLen:     int(c.Uint("attention.key_length", 256)),
+			attnValLen:     int(c.Uint("attention.value_length", 256)),
+			eps:            c.Float("attention.layer_norm_rms_epsilon", 1e-06),
+			ropeLocalBase:  c.Float("rope.local.freq_base", 10000.0),
+			ropeGlobalBase: c.Float("rope.global.freq_base", 1000000.0),
+			ropeScale:      c.Float("rope.freq_scale", 1.0),
+		},
+	}
+
+	if numBlocks == gemma27BLayerCount {
+		m.largeModelScaling = true
+	}
+
+	return &m
+}
+
+type TextSelfAttention struct {
+	Query     *nn.Linear  `gguf:"attn_q"`
+	QueryNorm *nn.RMSNorm `gguf:"attn_q_norm"`
+	Key       *nn.Linear  `gguf:"attn_k"`
+	KeyNorm   *nn.RMSNorm `gguf:"attn_k_norm"`
+	Value     *nn.Linear  `gguf:"attn_v"`
+	Output    *nn.Linear  `gguf:"attn_output"`
+}
+
+func (sa *TextSelfAttention) Forward(ctx ml.Context, layer int, hiddenState, positionIDs ml.Tensor, cache kvcache.Cache, opts *TextOptions) ml.Tensor {
+	batchSize := hiddenState.Dim(1)
+	ropeType := uint32(2)
+
+	ropeBase := opts.ropeLocalBase
+	if (layer+1)%gemmaGlobalCacheCount == 0 {
+		ropeBase = opts.ropeGlobalBase
+	}
+
+	q := sa.Query.Forward(ctx, hiddenState)
+	q = q.Reshape(ctx, opts.attnKeyLen, opts.numHeads, batchSize)
+	q = sa.QueryNorm.Forward(ctx, q, opts.eps)
+	q = q.RoPE(ctx, positionIDs, nil, uint32(opts.attnKeyLen), ropeType, ropeBase, opts.ropeScale)
+
+	if opts.largeModelScaling {
+		q = q.Scale(ctx, 1.0/math.Sqrt(float64(opts.hiddenSize/opts.numHeads)))
+	} else {
+		q = q.Scale(ctx, 1.0/math.Sqrt(float64(opts.attnKeyLen)))
+	}
+
+	k := sa.Key.Forward(ctx, hiddenState)
+	k = k.Reshape(ctx, opts.attnKeyLen, opts.numKVHeads, batchSize)
+	k = sa.KeyNorm.Forward(ctx, k, opts.eps)
+	k = k.RoPE(ctx, positionIDs, nil, uint32(opts.attnKeyLen), ropeType, ropeBase, opts.ropeScale)
+
+	v := sa.Value.Forward(ctx, hiddenState)
+	v = v.Reshape(ctx, opts.attnValLen, opts.numKVHeads, batchSize)
+
+	scaleFactor := 1.0
+	kqv := nn.Attention(ctx, q, k, v, scaleFactor, cache)
+	kqv = kqv.Reshape(ctx, opts.attnValLen*opts.numHeads, batchSize)
+
+	return sa.Output.Forward(ctx, kqv)
+}
+
+func (m *TextModel) Shift(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error) {
+	ropeBase := m.TextOptions.ropeLocalBase
+	if (layer+1)%gemmaGlobalCacheCount == 0 {
+		ropeBase = m.TextOptions.ropeGlobalBase
+	}
+
+	return key.RoPE(ctx, shift, nil, uint32(m.TextOptions.attnKeyLen), uint32(2), ropeBase, m.TextOptions.ropeScale), nil
+}
+
+type TextMLP struct {
+	Up   *nn.Linear `gguf:"ffn_up"`
+	Down *nn.Linear `gguf:"ffn_down"`
+	Gate *nn.Linear `gguf:"ffn_gate"`
+}
+
+func (mlp *TextMLP) Forward(ctx ml.Context, hiddenState ml.Tensor, opts *TextOptions) ml.Tensor {
+	hiddenState = mlp.Gate.Forward(ctx, hiddenState).GELU(ctx).Mul(ctx, mlp.Up.Forward(ctx, hiddenState))
+	return mlp.Down.Forward(ctx, hiddenState)
+}
+
+type TextLayer struct {
+	AttentionNorm     *nn.RMSNorm `gguf:"attn_norm"`
+	SelfAttention     *TextSelfAttention
+	PostAttentionNorm *nn.RMSNorm `gguf:"post_attention_norm"`
+	MLPNorm           *nn.RMSNorm `gguf:"ffn_norm"`
+	MLP               *TextMLP
+	PostMLPNorm       *nn.RMSNorm `gguf:"post_ffw_norm"`
+}
+
+func (l *TextLayer) Forward(ctx ml.Context, layer int, hiddenState, positionIDs, outputs ml.Tensor, cache kvcache.Cache, opts *TextOptions) ml.Tensor {
+	residual := hiddenState
+
+	hiddenState = l.AttentionNorm.Forward(ctx, hiddenState, opts.eps)
+	hiddenState = l.SelfAttention.Forward(ctx, layer, hiddenState, positionIDs, cache, opts)
+	hiddenState = l.PostAttentionNorm.Forward(ctx, hiddenState, opts.eps)
+
+	// In the final layer (outputs != nil), optimize by pruning to just the token positions
+	// we need logits for.
+	if outputs != nil {
+		hiddenState = hiddenState.Rows(ctx, outputs)
+		residual = residual.Rows(ctx, outputs)
+	}
+
+	hiddenState = hiddenState.Add(ctx, residual)
+	residual = hiddenState
+
+	hiddenState = l.MLPNorm.Forward(ctx, hiddenState, opts.eps)
+	hiddenState = l.MLP.Forward(ctx, hiddenState, opts)
+	hiddenState = l.PostMLPNorm.Forward(ctx, hiddenState, opts.eps)
+	return hiddenState.Add(ctx, residual)
+}
+
+func setImageEmbeddings(ctx ml.Context, hiddenState ml.Tensor, multimodal []input.MultimodalIndex) []int {
+	var embedding ml.Tensor
+	var src, dst, length int
+	var except []int
+
+	for _, image := range multimodal {
+		imageToken := image.Multimodal.(imageToken)
+		imageSrc := imageToken.index
+		imageDst := image.Index
+
+		if embedding == nil {
+			embedding = imageToken.embedding
+			src = imageSrc
+			dst = imageDst
+			length = 1
+		} else if embedding == imageToken.embedding && imageSrc+1 == src && imageDst+1 == dst {
+			src = imageSrc
+			dst = imageDst
+			length++
+		} else if embedding == imageToken.embedding && src+length == imageSrc && dst+length == imageDst {
+			length++
+		} else {
+			visionOutputs := embedding.View(ctx, src*embedding.Stride(1), length*embedding.Dim(0))
+			ctx.Forward(visionOutputs.Copy(ctx, hiddenState.View(ctx, dst*hiddenState.Stride(1), length*hiddenState.Dim(0))))
+
+			embedding = imageToken.embedding
+			src = imageSrc
+			dst = imageDst
+			length = 1
+		}
+
+		except = append(except, imageDst)
+	}
+
+	if embedding != nil {
+		visionOutputs := embedding.View(ctx, src*embedding.Stride(1), length*embedding.Dim(0))
+		ctx.Forward(visionOutputs.Copy(ctx, hiddenState.View(ctx, dst*hiddenState.Stride(1), length*hiddenState.Dim(0))))
+	}
+
+	return except
+}
+
+func (m *TextModel) Forward(ctx ml.Context, inputs, positions, outputs ml.Tensor, opts input.Options, cache kvcache.Cache) ml.Tensor {
+	hiddenState := m.TokenEmbedding.Forward(ctx, inputs)
+	hiddenState = hiddenState.Scale(ctx, math.Sqrt(float64(m.TextOptions.hiddenSize)))
+
+	except := setImageEmbeddings(ctx, hiddenState, opts.Multimodal)
+
+	for i, layer := range m.Layers {
+		// gemma alternates between the sliding window (local) and causal (global)
+		// kv cache every 6 layers
+		cacheType := cacheTypeSWA
+		if (i+1)%gemmaGlobalCacheCount == 0 {
+			cacheType = cacheTypeCausal
+		}
+		cache.SetLayer(i)
+		wc := cache.(*kvcache.WrapperCache)
+		wc.SetLayerType(cacheType)
+
+		if causal, ok := wc.UnderlyingCache().(*kvcache.Causal); ok {
+			causal.SetCausal(ctx, kvcache.CausalOptions{Except: except})
+		}
+
+		var lastLayerOutputs ml.Tensor
+		if i == len(m.Layers)-1 {
+			lastLayerOutputs = outputs
+		}
+
+		hiddenState = layer.Forward(ctx, i, hiddenState, positions, lastLayerOutputs, cache, m.TextOptions)
+	}
+
+	hiddenState = m.OutputNorm.Forward(ctx, hiddenState, m.eps)
+	return m.Output.Forward(ctx, hiddenState)
+}

model/models/gemma3/model_vision.go

@@ -0,0 +1,127 @@
+package gemma3
+
+import (
+	"math"
+
+	"github.com/ollama/ollama/ml"
+	"github.com/ollama/ollama/ml/nn"
+)
+
+var batchSize int = 1
+
+type VisionSelfAttention struct {
+	Query  *nn.Linear `gguf:"attn_q"`
+	Key    *nn.Linear `gguf:"attn_k"`
+	Value  *nn.Linear `gguf:"attn_v"`
+	Output *nn.Linear `gguf:"attn_output"`
+}
+
+func (sa *VisionSelfAttention) Forward(ctx ml.Context, hiddenState ml.Tensor, opts *VisionModelOptions) ml.Tensor {
+	headDim := opts.hiddenSize / opts.numHeads
+
+	query := sa.Query.Forward(ctx, hiddenState)
+	key := sa.Key.Forward(ctx, hiddenState)
+	value := sa.Value.Forward(ctx, hiddenState)
+
+	query = query.Reshape(ctx, headDim, opts.numHeads, query.Dim(1), batchSize)
+	key = key.Reshape(ctx, headDim, opts.numHeads, key.Dim(1), batchSize)
+	value = value.Reshape(ctx, headDim, opts.numHeads, value.Dim(1), batchSize)
+
+	attention := nn.Attention(ctx, query, key, value, 1.0/math.Sqrt(float64(headDim)), nil)
+	attention = attention.Reshape(ctx, opts.hiddenSize, attention.Dim(2), batchSize)
+
+	hiddenState = sa.Output.Forward(ctx, attention)
+	return hiddenState
+}
+
+type VisionMLP struct {
+	FC1 *nn.Linear `gguf:"fc1"`
+	FC2 *nn.Linear `gguf:"fc2"`
+}
+
+func (mlp *VisionMLP) Forward(ctx ml.Context, hiddenState ml.Tensor, opts *VisionModelOptions) ml.Tensor {
+	hiddenState = mlp.FC1.Forward(ctx, hiddenState).GELU(ctx)
+	hiddenState = mlp.FC2.Forward(ctx, hiddenState)
+	return hiddenState
+}
+
+type VisionEncoderLayer struct {
+	LayerNorm1    *nn.LayerNorm `gguf:"layer_norm1"`
+	SelfAttention *VisionSelfAttention
+
+	LayerNorm2 *nn.LayerNorm `gguf:"layer_norm2"`
+	MLP        *VisionMLP    `gguf:"mlp"`
+}
+
+func (e *VisionEncoderLayer) Forward(ctx ml.Context, hiddenState ml.Tensor, opts *VisionModelOptions) ml.Tensor {
+	residual := hiddenState
+
+	// self attention
+	hiddenState = e.LayerNorm1.Forward(ctx, hiddenState, opts.eps)
+	hiddenState = e.SelfAttention.Forward(ctx, hiddenState, opts)
+	hiddenState = hiddenState.Add(ctx, residual)
+	residual = hiddenState
+
+	// feed forward
+	hiddenState = e.LayerNorm2.Forward(ctx, hiddenState, opts.eps)
+	hiddenState = e.MLP.Forward(ctx, hiddenState, opts)
+	return hiddenState.Add(ctx, residual)
+}
+
+type VisionModelOptions struct {
+	hiddenSize, numHeads int
+	imageSize, patchSize int
+	eps                  float32
+}
+
+type VisionModel struct {
+	PatchEmbedding    *nn.Conv2D    `gguf:"patch_embedding"`
+	PositionEmbedding *nn.Embedding `gguf:"position_embedding"`
+	PostLayerNorm     *nn.LayerNorm `gguf:"post_layernorm"`
+
+	Layers []VisionEncoderLayer `gguf:"blk"`
+
+	*VisionModelOptions
+}
+
+func (m *VisionModel) Forward(ctx ml.Context, pixelValues ml.Tensor) ml.Tensor {
+	numPatches := (m.imageSize / m.patchSize) * (m.imageSize / m.patchSize)
+
+	hiddenState := m.PatchEmbedding.Forward(ctx, pixelValues, m.patchSize, m.patchSize, 0, 0, 1, 1)
+	hiddenState = hiddenState.Reshape(ctx, numPatches, m.hiddenSize)
+	hiddenState = hiddenState.Permute(ctx, 1, 0, 2, 3).Contiguous(ctx)
+
+	positions := make([]int32, numPatches)
+	for i := range positions {
+		positions[i] = int32(i)
+	}
+
+	positionIDs, err := ctx.Input().FromIntSlice(positions, len(positions))
+	if err != nil {
+		panic(err)
+	}
+
+	hiddenState = hiddenState.Add(ctx, m.PositionEmbedding.Forward(ctx, positionIDs))
+
+	for _, layer := range m.Layers {
+		hiddenState = layer.Forward(ctx, hiddenState, m.VisionModelOptions)
+	}
+
+	hiddenState = m.PostLayerNorm.Forward(ctx, hiddenState, m.eps)
+	return hiddenState
+}
+
+func newVisionModel(c ml.Config) *VisionModel {
+	return &VisionModel{
+		Layers: make([]VisionEncoderLayer, c.Uint("vision.block_count")),
+		VisionModelOptions: &VisionModelOptions{
+			hiddenSize: int(c.Uint("vision.embedding_length")),
+			numHeads:   int(c.Uint("vision.attention.head_count")),
+
+			imageSize: int(c.Uint("vision.image_size")),
+			patchSize: int(c.Uint("vision.patch_size")),
+
+			eps: c.Float("vision.attention.layer_norm_epsilon"),
+		},
+	}
+}

model/models/gemma3/process_image.go

@@ -0,0 +1,58 @@
+package gemma3
+
+import (
+	"image"
+
+	"github.com/ollama/ollama/ml"
+	"github.com/ollama/ollama/model/imageproc"
+)
+
+type ImageProcessor struct {
+	imageSize, patchSize, numChannels int
+}
+
+func newImageProcessor(c ml.Config) ImageProcessor {
+	return ImageProcessor{
+		imageSize:   int(c.Uint("vision.image_size")),
+		patchSize:   int(c.Uint("vision.patch_size")),
+		numChannels: int(c.Uint("vision.num_channels")),
+	}
+}
+
+func (p *ImageProcessor) pack(img image.Image, mean, std [3]float32) []float32 {
+	var pixelVals, rVals, gVals, bVals []float32
+
+	bounds := img.Bounds()
+	for y := bounds.Min.Y; y < bounds.Max.Y; y++ {
+		for x := bounds.Min.X; x < bounds.Max.X; x++ {
+			c := img.At(x, y)
+			r, g, b, _ := c.RGBA()
+			rVal := float32(r>>8) / 255.0
+			gVal := float32(g>>8) / 255.0
+			bVal := float32(b>>8) / 255.0
+
+			rVal = (rVal - mean[0]) / std[0]
+			gVal = (gVal - mean[1]) / std[1]
+			bVal = (bVal - mean[2]) / std[2]
+
+			rVals = append(rVals, rVal)
+			gVals = append(gVals, gVal)
+			bVals = append(bVals, bVal)
+		}
+	}
+
+	pixelVals = append(pixelVals, rVals...)
+	pixelVals = append(pixelVals, gVals...)
+	pixelVals = append(pixelVals, bVals...)
+
+	return pixelVals
+}
+
+func (p ImageProcessor) ProcessImage(img image.Image) ([]float32, error) {
+	outputSize := image.Point{p.imageSize, p.imageSize}
+	newImage := imageproc.Composite(img)
+	newImage = imageproc.Resize(newImage, outputSize, imageproc.ResizeBilinear)
+
+	data := p.pack(newImage, imageproc.ImageNetStandardMean, imageproc.ImageNetStandardSTD)
+	return data, nil
+}

model/models/llama/model.go

@@ -76,14 +76,15 @@ type SelfAttention struct {
 func (sa *SelfAttention) Forward(ctx ml.Context, hiddenState, positionIDs ml.Tensor, cache kvcache.Cache, opts *Options) ml.Tensor {
 	batchSize := hiddenState.Dim(1)
 	headDim := opts.hiddenSize / opts.numHeads
+	ropeType := uint32(0)
 
 	q := sa.Query.Forward(ctx, hiddenState)
 	q = q.Reshape(ctx, headDim, opts.numHeads, batchSize)
-	q = q.RoPE(ctx, positionIDs, sa.RopeFactors, opts.ropeDim, opts.ropeBase, opts.ropeScale)
+	q = q.RoPE(ctx, positionIDs, sa.RopeFactors, opts.ropeDim, ropeType, opts.ropeBase, opts.ropeScale)
 
 	k := sa.Key.Forward(ctx, hiddenState)
 	k = k.Reshape(ctx, headDim, opts.numKVHeads, batchSize)
-	k = k.RoPE(ctx, positionIDs, sa.RopeFactors, opts.ropeDim, opts.ropeBase, opts.ropeScale)
+	k = k.RoPE(ctx, positionIDs, sa.RopeFactors, opts.ropeDim, ropeType, opts.ropeBase, opts.ropeScale)
 
 	v := sa.Value.Forward(ctx, hiddenState)
 	v = v.Reshape(ctx, headDim, opts.numKVHeads, batchSize)
@@ -96,7 +97,7 @@ func (sa *SelfAttention) Forward(ctx ml.Context, hiddenState, positionIDs ml.Ten
 }
 
 func (m *Model) Shift(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error) {
-	return key.RoPE(ctx, shift, m.Layers[layer].SelfAttention.RopeFactors, m.ropeDim, m.ropeBase, m.ropeScale), nil
+	return key.RoPE(ctx, shift, m.Layers[layer].SelfAttention.RopeFactors, uint32(0), m.ropeDim, m.ropeBase, m.ropeScale), nil
 }
 
 type MLP struct {

model/models/mllama/model.go

@@ -63,6 +63,10 @@ func New(c ml.Config) (model.Model, error) {
 }
 
 func (m *Model) EncodeMultimodal(ctx ml.Context, multimodalData []byte) (any, error) {
+	if len(m.VisionModel.Transformer.Layers) == 0 || len(m.GlobalTransformer.Layers) == 0 {
+		return nil, model.ErrNoVisionModel
+	}
+
 	image, _, err := image.Decode(bytes.NewReader(multimodalData))
 	if err != nil {
 		return nil, err

model/models/mllama/model_text.go

@@ -20,14 +20,15 @@ type TextSelfAttention struct {
 func (sa *TextSelfAttention) Forward(ctx ml.Context, hiddenState, positions, _ ml.Tensor, cache *kvcache.WrapperCache, opts *TextModelOptions) ml.Tensor {
 	batchSize := hiddenState.Dim(1)
 	headDim := opts.hiddenSize / opts.numHeads
+	ropeType := uint32(0)
 
 	query := sa.Query.Forward(ctx, hiddenState)
 	query = query.Reshape(ctx, headDim, opts.numHeads, batchSize)
-	query = query.RoPE(ctx, positions, sa.RopeFactors, opts.ropeDim, opts.ropeBase, opts.ropeScale)
+	query = query.RoPE(ctx, positions, sa.RopeFactors, opts.ropeDim, ropeType, opts.ropeBase, opts.ropeScale)
 
 	key := sa.Key.Forward(ctx, hiddenState)
 	key = key.Reshape(ctx, headDim, opts.numKVHeads, batchSize)
-	key = key.RoPE(ctx, positions, sa.RopeFactors, opts.ropeDim, opts.ropeBase, opts.ropeScale)
+	key = key.RoPE(ctx, positions, sa.RopeFactors, opts.ropeDim, ropeType, opts.ropeBase, opts.ropeScale)
 
 	value := sa.Value.Forward(ctx, hiddenState)
 	value = value.Reshape(ctx, headDim, opts.numKVHeads, batchSize)
@@ -40,8 +41,9 @@ func (sa *TextSelfAttention) Forward(ctx ml.Context, hiddenState, positions, _ m
 }
 
 func (m *TextModel) Shift(ctx ml.Context, layer int, key, shift ml.Tensor) (ml.Tensor, error) {
+	// This will only get called for layers in the cache, which are just the self attention layers
 	if sa, ok := m.Transformer.Layers[layer].(*TextSelfAttentionDecoderLayer); ok {
-		return key.RoPE(ctx, shift, sa.SelfAttention.RopeFactors, m.ropeDim, m.ropeBase, m.ropeScale), nil
+		return key.RoPE(ctx, shift, sa.SelfAttention.RopeFactors, m.ropeDim, uint32(0), m.ropeBase, m.ropeScale), nil
 	}
 
 	return key, nil

model/models/mllama/process_image.go

@@ -144,8 +144,6 @@ func (p *ImageProcessor) splitToTiles(img image.Image, numTilesSize image.Point)
 	return images
 }
 
-// remove the "alpha" channel by drawing over a prefilled image
-//
 // remove the "alpha" channel by drawing over a prefilled image
 //
 //nolint:unused

model/models/models.go

@@ -1,6 +1,8 @@
 package models
 
 import (
+	_ "github.com/ollama/ollama/model/models/gemma2"
+	_ "github.com/ollama/ollama/model/models/gemma3"
 	_ "github.com/ollama/ollama/model/models/llama"
 	_ "github.com/ollama/ollama/model/models/mllama"
 )

model/process_text.go

@@ -4,6 +4,7 @@ import (
 	"cmp"
 	"iter"
 	"log/slog"
+	"slices"
 	"strings"
 	"sync"
 
@@ -18,6 +19,15 @@ const (
 	SpecialEOS
 )
 
+const (
+	TOKEN_TYPE_NORMAL = iota + 1
+	TOKEN_TYPE_UNKNOWN
+	TOKEN_TYPE_CONTROL
+	TOKEN_TYPE_USER_DEFINED
+	TOKEN_TYPE_UNUSED
+	TOKEN_TYPE_BYTE
+)
+
 type TextProcessor interface {
 	Encode(s string, addSpecial bool) ([]int32, error)
 	Decode([]int32) (string, error)
@@ -27,11 +37,11 @@ type TextProcessor interface {
 type Vocabulary struct {
 	Values []string
 	Types  []uint32
-	Scores []uint32
+	Scores []float32
 	Merges []string
 
-	BOS, EOS       int32
-	AddBOS, AddEOS bool
+	BOS, EOS, EOT          int32
+	AddBOS, AddEOS, AddEOT bool
 
 	specialOnce sync.Once
 	special     []string
@@ -48,7 +58,7 @@ func (v *Vocabulary) Is(id int32, special Special) bool {
 	case SpecialBOS:
 		return id == v.BOS
 	case SpecialEOS:
-		return id == v.EOS
+		return id == v.EOS || id == v.EOT
 	default:
 		return false
 	}
@@ -76,7 +86,9 @@ func (v *Vocabulary) Decode(id int32) string {
 func (v *Vocabulary) SpecialVocabulary() []string {
 	v.specialOnce.Do(func() {
 		for i := range v.Values {
-			if v.Types[i] == 3 {
+			if slices.Contains([]int{105, 106}, i) {
+				v.special = append(v.special, v.Values[i])
+			} else if v.Types[i] == TOKEN_TYPE_CONTROL {
 				v.special = append(v.special, v.Values[i])
 			}
 		}

model/process_text_spm.go

@@ -0,0 +1,246 @@
+package model
+
+import (
+	"iter"
+	"log/slog"
+	"strings"
+
+	"github.com/dlclark/regexp2"
+	queue "github.com/emirpasic/gods/v2/queues/priorityqueue"
+)
+
+const spmWhitespaceSep = "▁"
+
+func replaceWhitespaceBySeperator(s string) string {
+	return strings.ReplaceAll(s, " ", spmWhitespaceSep)
+}
+
+type SentencePieceModel struct {
+	maxTokenLen int
+	pre         *regexp2.Regexp
+	vocab       *Vocabulary
+}
+
+var _ TextProcessor = (*SentencePieceModel)(nil)
+
+func NewSentencePieceModel(pre string, vocab *Vocabulary) SentencePieceModel {
+	slog.Debug("Tokens", "num tokens", len(vocab.Values), "vals", vocab.Values[:5], "scores", vocab.Scores[:5], "types", vocab.Types[:5])
+
+	counter := map[int]int{}
+	var maxTokenLen int
+	for cnt := range vocab.Types {
+		switch vocab.Types[cnt] {
+		case TOKEN_TYPE_NORMAL, TOKEN_TYPE_USER_DEFINED, TOKEN_TYPE_UNUSED:
+			maxTokenLen = max(maxTokenLen, len(vocab.Values[cnt]))
+			fallthrough
+		default:
+			counter[int(vocab.Types[cnt])] += 1
+		}
+	}
+
+	slog.Debug("Token counts", "normal", counter[TOKEN_TYPE_NORMAL], "unknown", counter[TOKEN_TYPE_UNKNOWN], "control", counter[TOKEN_TYPE_CONTROL],
+		"user defined", counter[TOKEN_TYPE_USER_DEFINED], "unused", counter[TOKEN_TYPE_UNUSED], "byte", counter[TOKEN_TYPE_BYTE],
+		"max token len", maxTokenLen)
+
+	return SentencePieceModel{
+		maxTokenLen: maxTokenLen,
+		pre:         regexp2.MustCompile(pre, regexp2.Unicode|regexp2.RE2),
+		vocab:       vocab,
+	}
+}
+
+func (spm SentencePieceModel) Is(id int32, special Special) bool {
+	return spm.vocab.Is(id, special)
+}
+
+func (spm *SentencePieceModel) split(s string) iter.Seq[string] {
+	return func(yield func(string) bool) {
+		for m, _ := spm.pre.FindStringMatch(s); m != nil; m, _ = spm.pre.FindNextMatch(m) {
+			if !yield(m.String()) {
+				break
+			}
+		}
+	}
+}
+
+func (spm SentencePieceModel) Encode(s string, addSpecial bool) ([]int32, error) {
+	fragments := []fragment{{value: s}}
+	for _, special := range spm.vocab.SpecialVocabulary() {
+		// TODO: process special tokens concurrently
+		id := spm.vocab.Encode(special)
+		for i := 0; i < len(fragments); i++ {
+			frag := fragments[i]
+			if len(frag.ids) > 0 {
+				continue
+			}
+
+			var middle []fragment
+			switch i := strings.Index(frag.value, special); {
+			case i < 0:
+				middle = append(middle, frag)
+			case i > 0:
+				middle = append(middle, fragment{value: frag.value[:i]})
+				fallthrough
+			default:
+				middle = append(middle, fragment{value: special, ids: []int32{id}})
+				if rest := frag.value[i+len(special):]; rest != "" {
+					middle = append(middle, fragment{value: rest})
+				}
+			}
+
+			fragments = append(fragments[:i], append(middle, fragments[i+1:]...)...)
+		}
+	}
+	slog.Debug("fragments", "frags", fragments)
+
+	var ids []int32
+	for _, frag := range fragments {
+		if len(frag.ids) > 0 {
+			ids = append(ids, frag.ids...)
+			continue
+		}
+
+		for split := range spm.split(frag.value) {
+			split = replaceWhitespaceBySeperator(split)
+
+			var sb strings.Builder
+			sb.Write([]byte(split))
+			if id := spm.vocab.Encode(sb.String()); id >= 0 {
+				ids = append(ids, id)
+				continue
+			}
+
+			runes := []rune(sb.String())
+			pq := queue.NewWith(func(a, b any) int {
+				priA := a.(*candidate)
+				priB := b.(*candidate)
+				if priA.score > priB.score || (priA.score == priB.score && priA.a < priB.a) {
+					return -1
+				}
+				return 1
+			})
+
+			merges := make([]merge, len(runes))
+			for r := range runes {
+				merges[r] = merge{
+					p:     r - 1,
+					n:     r + 1,
+					runes: []rune{runes[r]},
+				}
+			}
+
+			slog.Debug("tokenizer", "merges", merges)
+
+			pairwise := func(a, b int) *candidate {
+				if a < 0 || b >= len(runes) {
+					return nil
+				}
+
+				left, right := string(merges[a].runes), string(merges[b].runes)
+				if id := spm.vocab.Encode(left + right); id >= 0 {
+					return &candidate{
+						a:     a,
+						b:     b,
+						score: spm.vocab.Scores[id],
+					}
+				}
+				return nil
+			}
+
+			for i := range len(runes) - 1 {
+				if pair := pairwise(i, i+1); pair != nil {
+					pq.Enqueue(pair)
+				}
+			}
+
+			pqv := pq.Values()
+			for _, v := range pqv {
+				e := v.(*candidate)
+				slog.Debug("candidate", "candidate", e)
+			}
+
+			for !pq.Empty() {
+				v, _ := pq.Dequeue()
+				pair := v.(*candidate)
+				left, right := merges[pair.a], merges[pair.b]
+
+				slog.Debug("pair", "left", left, "right", right)
+				if len(left.runes) == 0 || len(right.runes) == 0 {
+					continue
+				}
+
+				if id := spm.vocab.Encode(string(left.runes) + string(right.runes)); id < 0 {
+					continue
+				}
+
+				merges[pair.a].runes = append(left.runes, right.runes...)
+				merges[pair.b].runes = nil
+				merges[pair.a].n = right.n
+				if right.n < len(merges) {
+					merges[right.n].p = pair.a
+				}
+
+				if pair := pairwise(merges[pair.a].p, pair.a); pair != nil {
+					pq.Enqueue(pair)
+				}
+
+				if pair := pairwise(pair.a, merges[pair.a].n); pair != nil {
+					pq.Enqueue(pair)
+				}
+			}
+
+			slog.Debug("merges", "merges", merges)
+
+			for _, merge := range merges {
+				if len(merge.runes) > 0 {
+					if id := spm.vocab.Encode(string(merge.runes)); id >= 0 {
+						ids = append(ids, id)
+					} else {
+						slog.Debug("missing token", "token", string(merge.runes))
+					}
+				}
+			}
+		}
+	}
+
+	if addSpecial && len(ids) > 0 {
+		if spm.vocab.AddBOS {
+			if ids[0] == spm.vocab.BOS {
+				slog.Warn("adding bos token to prompt which already has it", "id", spm.vocab.BOS)
+			}
+
+			slog.Debug("adding bos token to prompt", "id", spm.vocab.BOS)
+			ids = append([]int32{spm.vocab.BOS}, ids...)
+		}
+
+		if spm.vocab.AddEOS {
+			if ids[len(ids)-1] == spm.vocab.EOS {
+				slog.Warn("adding eos token to prompt which already has it", "id", spm.vocab.EOS)
+			}
+
+			slog.Debug("adding eos token to prompt", "id", spm.vocab.EOS)
+			ids = append(ids, spm.vocab.EOS)
+		}
+	}
+
+	return ids, nil
+}
+
+type candidate struct {
+	a, b  int
+	score float32
+}
+
+func (spm SentencePieceModel) Decode(ids []int32) (string, error) {
+	var sb strings.Builder
+	for _, id := range ids {
+		data := spm.vocab.Decode(id)
+		data = strings.ReplaceAll(data, spmWhitespaceSep, " ")
+		if _, err := sb.WriteString(data); err != nil {
+			return "", err
+		}
+	}
+
+	slog.Debug("decoded", "ids", ids, "text", sb.String())
+	return sb.String(), nil
+}

model/process_text_spm_test.go

@@ -0,0 +1,118 @@
+package model
+
+import (
+	"log/slog"
+	"os"
+	"path/filepath"
+	"slices"
+	"testing"
+
+	"google.golang.org/protobuf/proto"
+
+	"github.com/ollama/ollama/convert/sentencepiece"
+)
+
+func loadSentencePieceVocab(t *testing.T) SentencePieceModel {
+	t.Helper()
+
+	bts, err := os.ReadFile(filepath.Join("testdata", "gemma2", "tokenizer.model"))
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	var spm sentencepiece.ModelProto
+	if err := proto.Unmarshal(bts, &spm); err != nil {
+		t.Fatal(err)
+	}
+
+	preTokenizer := `(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\r\n\p{L}\p{N}]?\p{L}+|\p{N}{1,3}| ?[^\s\p{L}\p{N}]+[\r\n]*|\s*[\r\n]+|\s+(?!\S)|\s+`
+
+	var v Vocabulary
+
+	for _, piece := range spm.GetPieces() {
+		v.Values = append(v.Values, piece.GetPiece())
+		v.Scores = append(v.Scores, piece.GetScore())
+		switch t := piece.GetType(); t {
+		case sentencepiece.ModelProto_SentencePiece_UNKNOWN,
+			sentencepiece.ModelProto_SentencePiece_CONTROL,
+			sentencepiece.ModelProto_SentencePiece_UNUSED,
+			sentencepiece.ModelProto_SentencePiece_BYTE:
+			v.Types = append(v.Types, uint32(t))
+		default:
+			tt := uint32(sentencepiece.ModelProto_SentencePiece_NORMAL)
+			// todo parse the special tokens file
+			//   - this will roundtrip correctly but the <start_of_turn> and
+			//     <end_of_turn> tokens aren't processed
+			v.Types = append(v.Types, tt)
+		}
+	}
+
+	return NewSentencePieceModel(preTokenizer, &v)
+}
+
+func TestSentencePieceEncode(t *testing.T) {
+	logger := slog.New(slog.NewTextHandler(os.Stdout, &slog.HandlerOptions{Level: slog.LevelDebug}))
+	slog.SetDefault(logger)
+
+	tokenizer := loadSentencePieceVocab(t)
+
+	t.Run("basic roundtrip", func(t *testing.T) {
+		t.Parallel()
+
+		cases := []string{
+			"hello",
+			"hello ",
+			"hello  ",
+			" hello",
+			" hello ",
+			" hello  ",
+			"hello world",
+			"请考试我的软件！12345",
+			"你好",
+			"Hello 你好 world!",
+			"Special characters: !@#$%^&*()_+-=[]{}|;':\",./<>?",
+			"Multilingual: 你好 こんにちは Привет Hola مرحبا",
+			"Numbers and symbols: 123456789 +- */",
+			"Special tokens: <bos> text <eos>",
+			"Code snippets: func main() { fmt.Println(\"Hello World\") }",
+			"Long text: " + "Lorem ipsum dolor sit amet, consectetur adipiscing elit. " +
+				"Sed do eiusmod tempor incididunt ut labore et dolore magna aliqua. " +
+				"Ut enim ad minim veniam, quis nostrud exercitation ullamco laboris.",
+		}
+
+		for _, want := range cases {
+			ids, err := tokenizer.Encode(want, true)
+			if err != nil {
+				t.Fatal(err)
+			}
+
+			if got, err := tokenizer.Decode(ids); err != nil {
+				t.Fatal(err)
+			} else if got != want {
+				t.Errorf("got %q, want %q [%#v]", got, want, ids)
+			}
+		}
+	})
+
+	t.Run("special tokens", func(t *testing.T) {
+		type candidate struct {
+			token string
+			ids   []int32
+		}
+
+		cases := []candidate{
+			{"<bos>", []int32{2}},
+			{"<eos>", []int32{1}},
+		}
+
+		for _, want := range cases {
+			ids, err := tokenizer.Encode(want.token, true)
+			if err != nil {
+				t.Fatal(err)
+			}
+			if !slices.Equal(ids, want.ids) {
+				t.Errorf("got %#v, want %#v", ids, want.ids)
+			}
+		}
+	})
+}

readline/readline.go

@@ -116,19 +116,9 @@ func (i *Instance) Readline() (string, error) {
 
 			switch r {
 			case KeyUp:
-				if i.History.Pos > 0 {
-					if i.History.Pos == i.History.Size() {
-						currentLineBuf = []rune(buf.String())
-					}
-					buf.Replace([]rune(i.History.Prev()))
-				}
+				i.historyPrev(buf, &currentLineBuf)
 			case KeyDown:
-				if i.History.Pos < i.History.Size() {
-					buf.Replace([]rune(i.History.Next()))
-					if i.History.Pos == i.History.Size() {
-						buf.Replace(currentLineBuf)
-					}
-				}
+				i.historyNext(buf, &currentLineBuf)
 			case KeyLeft:
 				buf.MoveLeft()
 			case KeyRight:
@@ -185,6 +175,10 @@ func (i *Instance) Readline() (string, error) {
 			esc = true
 		case CharInterrupt:
 			return "", ErrInterrupt
+		case CharPrev:
+			i.historyPrev(buf, &currentLineBuf)
+		case CharNext:
+			i.historyNext(buf, &currentLineBuf)
 		case CharLineStart:
 			buf.MoveToStart()
 		case CharLineEnd:
@@ -246,6 +240,24 @@ func (i *Instance) HistoryDisable() {
 	i.History.Enabled = false
 }
 
+func (i *Instance) historyPrev(buf *Buffer, currentLineBuf *[]rune) {
+	if i.History.Pos > 0 {
+		if i.History.Pos == i.History.Size() {
+			*currentLineBuf = []rune(buf.String())
+		}
+		buf.Replace([]rune(i.History.Prev()))
+	}
+}
+
+func (i *Instance) historyNext(buf *Buffer, currentLineBuf *[]rune) {
+	if i.History.Pos < i.History.Size() {
+		buf.Replace([]rune(i.History.Next()))
+		if i.History.Pos == i.History.Size() {
+			buf.Replace(*currentLineBuf)
+		}
+	}
+}
+
 func NewTerminal() (*Terminal, error) {
 	fd := os.Stdin.Fd()
 	termios, err := SetRawMode(fd)

runner/ollamarunner/runner.go

@@ -691,65 +691,6 @@ type EmbeddingResponse struct {
 	Embedding []float32 `json:"embedding"`
 }
 
-func (s *Server) embeddings(w http.ResponseWriter, r *http.Request) {
-	var req EmbeddingRequest
-	if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
-		http.Error(w, fmt.Sprintf("bad request: %s", err), http.StatusBadRequest)
-		return
-	}
-
-	w.Header().Set("Content-Type", "application/json")
-
-	slog.Debug("embedding request", "content", req.Content)
-
-	seq, err := s.NewSequence(req.Content, nil, NewSequenceParams{embedding: true})
-	if err != nil {
-		http.Error(w, fmt.Sprintf("Failed to create new sequence: %v", err), http.StatusInternalServerError)
-		return
-	}
-
-	// Ensure there is a place to put the sequence, released when removed from s.seqs
-	if err := s.seqsSem.Acquire(r.Context(), 1); err != nil {
-		if errors.Is(err, context.Canceled) {
-			slog.Info("aborting embeddings request due to client closing the connection")
-		} else {
-			slog.Error("Failed to acquire semaphore", "error", err)
-		}
-		return
-	}
-
-	s.mu.Lock()
-	found := false
-	for i, sq := range s.seqs {
-		if sq == nil {
-			seq.cache, seq.inputs, err = s.cache.LoadCacheSlot(seq.inputs, req.CachePrompt)
-			if err != nil {
-				s.mu.Unlock()
-				http.Error(w, fmt.Sprintf("Failed to load cache: %v", err), http.StatusInternalServerError)
-				return
-			}
-			s.seqs[i] = seq
-			s.cond.Signal()
-			found = true
-			break
-		}
-	}
-	s.mu.Unlock()
-
-	if !found {
-		http.Error(w, "could not find an available sequence", http.StatusInternalServerError)
-		return
-	}
-
-	embedding := <-seq.embedding
-
-	if err := json.NewEncoder(w).Encode(&EmbeddingResponse{
-		Embedding: embedding,
-	}); err != nil {
-		http.Error(w, fmt.Sprintf("failed to encode response: %v", err), http.StatusInternalServerError)
-	}
-}
-
 type HealthResponse struct {
 	Status   string  `json:"status"`
 	Progress float32 `json:"progress"`
@@ -927,9 +868,13 @@ func Execute(args []string) error {
 	defer listener.Close()
 
 	mux := http.NewServeMux()
-	mux.HandleFunc("/embedding", server.embeddings)
-	mux.HandleFunc("/completion", server.completion)
-	mux.HandleFunc("/health", server.health)
+	// TODO: support embeddings
+	mux.HandleFunc("POST /embedding", func(w http.ResponseWriter, r *http.Request) {
+		http.Error(w, "this model does not support embeddings", http.StatusNotImplemented)
+	})
+
+	mux.HandleFunc("POST /completion", server.completion)
+	mux.HandleFunc("GET /health", server.health)
 
 	httpServer := http.Server{
 		Handler: mux,

sample/samplers.go

@@ -84,14 +84,12 @@ func (s *Sampler) sample(tokens []token) (token, error) {
 		return greedy(tokens), nil
 	}
 
-	if s.topK > 0 {
-		tokens = topK(tokens, s.topK)
-	} else {
-		sortLogits(tokens)
-	}
+	// topK also sorts the tokens in descending order of logits
+	tokens = topK(tokens, s.topK)
 
 	tokens = temperature(tokens, s.temperature)
 	tokens = softmax(tokens)
+
 	tokens = topP(tokens, s.topP)
 	tokens = minP(tokens, s.minP)
 

sample/transforms.go

@@ -1,17 +1,58 @@
 package sample
 
 import (
+	"container/heap"
 	"math"
 	"slices"
 )
 
+// tokenHeap implements heap.Interface and holds tokens as a min-heap to track k largest elements
+type tokenHeap []token
+
+func (h tokenHeap) Len() int           { return len(h) }
+func (h tokenHeap) Less(i, j int) bool { return h[i].value < h[j].value }
+func (h tokenHeap) Swap(i, j int)      { h[i], h[j] = h[j], h[i] }
+
+func (h *tokenHeap) Push(x any) {
+	*h = append(*h, x.(token))
+}
+
+func (h *tokenHeap) Pop() any {
+	old := *h
+	n := len(old)
+	x := old[n-1]
+	*h = old[0 : n-1]
+	return x
+}
+
+// temperature applies scaling to the logits
+func temperature(ts []token, temp float32) []token {
+	// Ensure temperature clipping near 0 to avoid numerical instability
+	temp = max(temp, 1e-7)
+	for i := range ts {
+		ts[i].value = ts[i].value / temp
+	}
+	return ts
+}
+
+// softmax applies normalization to the logits
 func softmax(ts []token) []token {
+	// Find max logit for numerical stability
+	maxLogit := float32(math.Inf(-1))
+	for _, t := range ts {
+		if t.value > maxLogit {
+			maxLogit = t.value
+		}
+	}
+
+	// Compute exp(x - max)
 	var sum float32
 	for i, v := range ts {
-		ts[i].value = float32(math.Exp(float64(v.value)))
+		ts[i].value = float32(math.Exp(float64(v.value - maxLogit)))
 		sum += ts[i].value
 	}
 
+	// exp(x - max) / sum(exp(x - max))
 	for i := range ts {
 		ts[i].value /= sum
 	}
@@ -19,83 +60,42 @@ func softmax(ts []token) []token {
 	return ts
 }
 
-func temperature(ti []token, t float32) []token {
-	if t == 1 {
-		return ti
-	}
-
-	temp := max(t, 1e-7)
-	maxLogit := float32(math.Inf(-1))
-	for _, token := range ti {
-		if token.value > maxLogit {
-			maxLogit = token.value
-		}
-	}
-
-	// subtracting max logit to avoid under/overflow
-	for i := range ti {
-		ti[i].value = (ti[i].value - maxLogit) / temp
-	}
-
-	return ti
-}
-
-// siftDown maintains a min-heap property by recursively moving larger elements down the heap.
-//
-// The heap is represented as an array where for any node at index i:
-// - Left child is at index 2i + 1
-// - Right child is at index 2i + 2
-// - Parent is at index (i-1)/2
-//
-// The function compares a node with its children and:
-// 1. Finds the smallest value between the node and its children
-// 2. If the node is not the smallest, swaps it with its smallest child
-// 3. Continues this process down the affected path until the min-heap property is restored
-func siftDown(data []token, start, end int) {
-	root := start
-	for {
-		child := 2*root + 1
-		if child >= end {
-			break
-		}
-		// Find smaller child (we want min heap)
-		if child+1 < end && data[child+1].value < data[child].value {
-			child++
-		}
-		// Exit if root is already smaller than children
-		if data[root].value <= data[child].value {
-			break
-		}
-		// Swap with smaller child and continue
-		data[root], data[child] = data[child], data[root]
-		root = child
-	}
-}
-
 // topK limits the number of tokens considered to the k highest logits
 func topK(ts []token, k int) []token {
-	if k >= len(ts) {
+	if k >= len(ts) || k <= 0 {
+		slices.SortFunc(ts, func(a, b token) int {
+			switch {
+			case a.value < b.value:
+				return 1
+			case a.value > b.value:
+				return -1
+			default:
+				return 0
+			}
+		})
 		return ts
 	}
-	// Heapify + siftDown - O(nlog(k))
-	// Build min-heap of first k elements
-	heap := ts[:k]
-	for i := k/2 - 1; i >= 0; i-- {
-		siftDown(heap, i, k)
-	}
 
-	// Process remaining elements - if larger than heap root, replace root
+	// Initialize min-heap with first k elements
+	h := make(tokenHeap, k)
+	copy(h, ts[:k])
+	heap.Init(&h)
+
+	// Process remaining elements
 	for i := k; i < len(ts); i++ {
-		if ts[i].value > heap[0].value {
-			heap[0] = ts[i]
-			siftDown(heap, 0, k)
+		if ts[i].value > h[0].value {
+			heap.Pop(&h)
+			heap.Push(&h, ts[i])
 		}
 	}
 
-	slices.Reverse(heap)
+	// Convert heap to sorted slice in descending order
+	result := make([]token, len(h))
+	for i := k - 1; i >= 0; i-- {
+		result[i] = heap.Pop(&h).(token)
+	}
 
-	ts = heap
-	return ts
+	return result
 }
 
 // topP limits tokens to those with cumulative probability p
@@ -143,61 +143,3 @@ func minP(ts []token, p float32) []token {
 	ts = validTokens
 	return ts
 }
-
-// TODO(parthsareen): possibly replace with simpler implementation https://github.com/ollama/ollama/issues/9584
-// Conting sort implementation to sort tokens by logits
-func sortLogits(tokens []token) {
-	if len(tokens) <= 1 {
-		return
-	}
-
-	// Find max/min in a single pass
-	minLogit, maxLogit := tokens[0].value, tokens[0].value
-	for _, t := range tokens[1:] {
-		if t.value < minLogit {
-			minLogit = t.value
-		} else if t.value > maxLogit {
-			maxLogit = t.value
-		}
-	}
-
-	// Calculate scaling to map to uint32 range
-	logitRange := maxLogit - minLogit
-	if logitRange < 1e-6 {
-		return // All values effectively equal
-	}
-
-	// Count frequencies directly from tokens
-	const maxInt = (1 << 24) - 1 // Use 24 bits for good granularity
-	var counts [256]int          // For first byte
-
-	// First pass: count frequencies
-	for _, t := range tokens {
-		// Map to [0, maxInt] range
-		score := min(uint32((t.value-minLogit)*float32(maxInt)/logitRange), maxInt)
-		counts[score>>16]++
-	}
-
-	// Calculate offsets
-	var offset int
-	for i := range counts {
-		count := counts[i]
-		counts[i] = offset
-		offset += count
-	}
-
-	// Second pass: place elements in correct position
-	output := make([]token, len(tokens))
-	// Track current positions
-	countsCopy := counts
-
-	for i, t := range tokens {
-		score := min(uint32((t.value-minLogit)*float32(maxInt)/logitRange), maxInt)
-
-		pos := countsCopy[score>>16]
-		countsCopy[score>>16]++
-		output[len(tokens)-1-pos] = tokens[i]
-	}
-
-	copy(tokens, output)
-}

sample/transforms_test.go

@@ -6,86 +6,155 @@ import (
 	"testing"
 )
 
-// Helper to convert float64 slice to logit slice
-func toTokens(values []float64) []token {
+// Helper to convert float32 slice to logit slice
+func toTokens(values []float32) []token {
 	tokens := make([]token, len(values))
 	for i, v := range values {
 		tokens[i] = token{
 			id:    int32(i),
-			value: float32(v),
+			value: v,
 		}
 	}
 	return tokens
 }
 
 // Helper to compare logit slices
-func compareLogits(t *testing.T, name string, want []float64, got []token) {
+func compareLogits(t *testing.T, name string, want []float32, got []token) {
 	t.Helper()
 	if len(want) != len(got) {
 		t.Errorf("%s: length mismatch: want %d, got %d", name, len(want), len(got))
 		return
 	}
 	for i := range want {
-		if math.Abs(float64(got[i].value)-want[i]) > 1e-6 {
+		if math.Abs(float64(got[i].value-want[i])) > 1e-6 {
 			t.Errorf("%s: index %d: want %f, got %f", name, i, want[i], got[i].value)
 		}
 	}
 }
 
 func TestTemperature(t *testing.T) {
-	input := []float64{2, -1, 4, -3, 1, -2, 0}
-	want := []float64{-4, -10, 0, -14, -6, -12, -8} // (logit - max logit) / temp
-
+	input := []float32{1.0, 4.0, -2.0, 0.0}
 	got := temperature(toTokens(input), 0.5)
-	compareLogits(t, "Temperature", want, got)
+	want := []float32{2.0, 8.0, -4.0, 0.0}
+	compareLogits(t, "temperature(0.5)", want, got)
+
+	got = temperature(toTokens(input), 1.0)
+	want = []float32{1.0, 4.0, -2.0, 0.0}
+	compareLogits(t, "temperature(1)", want, got)
+
+	got = temperature(toTokens(input), 0.0)
+	want = []float32{1e7, 4e7, -2e7, 0.0}
+	compareLogits(t, "temperature(0)", want, got)
 }
 
 func TestSoftmax(t *testing.T) {
-	input := []float64{-3, -2, -1, 0, 1, 2, 4}
-	got := softmax(toTokens(input))
-
-	// Check probabilities sum to 1
-	var sum float32
-	for _, token := range got {
-		sum += token.value
-	}
-	if math.Abs(float64(sum)-1.0) > 1e-6 {
-		t.Errorf("probabilities don't sum to 1: got %f", sum)
+	tests := []struct {
+		name     string
+		input    []float32
+		expected []float32
+	}{
+		{
+			name:     "correctness softmax",
+			input:    []float32{1, -2, 3, 0},
+			expected: []float32{0.113550, 0.005653, 0.839024, 0.041773},
+		},
+		{
+			name:  "normal distribution",
+			input: []float32{0.026986899, 0.043722924, 0.036774673, 0.27755088, 0.0046718004, 0.08582123, 0.20409796, 0.00412893, 0.15720603, 0.045046154, 0.0030491839, 0.01681367},
+		},
+		{
+			name:  "single value",
+			input: []float32{1.0},
+		},
+		{
+			name:  "identical values",
+			input: []float32{0.9, 0.9, 0.9},
+		},
+		{
+			name:  "large values",
+			input: []float32{1000.0, 2000.0, 3000.0},
+		},
+		{
+			name:  "small values",
+			input: []float32{1e-6, 2e-6, 3e-6},
+		},
+		{
+			name:  "negative values",
+			input: []float32{-1.0, -2.0, -3.0},
+		},
+		{
+			name:  "mixed values",
+			input: []float32{-100.0, 0.0, 100.0},
+		},
 	}
 
-	// Check relative ordering is preserved
-	for i := 1; i < len(got); i++ {
-		if got[i].value < got[i-1].value {
-			t.Errorf("probability ordering not preserved at index %d", i)
-		}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			got := softmax(toTokens(tt.input))
+
+			if tt.expected != nil {
+				compareLogits(t, tt.name, tt.expected, got)
+				return
+			}
+
+			// Check probabilities sum to 1
+			var sum float32
+			for _, token := range got {
+				sum += token.value
+				if token.value < 0 || token.value > 1 {
+					t.Errorf("probability out of range [0,1]: got %f", token.value)
+				}
+			}
+			if math.Abs(float64(sum-1.0)) > 1e-6 {
+				t.Errorf("probabilities don't sum to 1: got %f", sum)
+			}
+		})
 	}
 }
 
 func TestTopK(t *testing.T) {
-	input := []float64{-3, -2, -1, 0, 1, 2, 4}
+	input := []float32{0.026986899, 0.043722924, 0.036774673, 0.27755088, 0.0046718004, 0.08582123, 0.20409796, 0.00412893, 0.15720603, 0.045046154, 0.0030491839, 0.01681367}
 
-	// Test k=3
-	got := topK(toTokens(input), 3)
-	if len(got) != 3 {
-		t.Errorf("topK(3): wrong length: want 3, got %d", len(got))
+	// Test k=5
+	got := topK(toTokens(input), 5)
+	if len(got) != 5 {
+		t.Errorf("topK(5): wrong length: want 5, got %d", len(got))
 	}
-	// Should keep highest 3 values: 4, 2, 1
-	want := []float64{4, 2, 1}
+	// Should keep highest 3 values in descending order
+	want := []float32{0.27755088, 0.20409796, 0.15720603, 0.08582123, 0.045046154}
 	compareLogits(t, "topK(3)", want, got)
 
-	// Test k > len
-	got = topK(toTokens(input), 10)
-	compareLogits(t, "topK(10)", input, got)
+	got = topK(toTokens(input), 20)
+	if len(got) != len(input) {
+		t.Errorf("topK(20): wrong length: want %d, got %d", len(input), len(got))
+	}
+
+	// Test k=-1
+	input = []float32{0.026986899, 0.043722924, 0.036774673, 0.27755088, 0.0046718004, 0.08582123, 0.20409796, 0.00412893, 0.15720603, 0.045046154, 0.0030491839, 0.01681367}
+	want = []float32{0.27755088, 0.20409796, 0.15720603, 0.08582123, 0.045046154, 0.043722924, 0.036774673, 0.026986899, 0.01681367, 0.0046718004, 0.00412893, 0.0030491839}
+	got = topK(toTokens(input), -1)
+	if len(got) != len(input) {
+		t.Errorf("topK(-1): wrong length: want %d, got %d", len(input), len(got))
+	}
+	compareLogits(t, "topK(-1)", want, got)
+
+	// Test k=0
+	input = []float32{0.026986899, 0.043722924, 0.036774673, 0.27755088, 0.0046718004, 0.08582123, 0.20409796, 0.00412893, 0.15720603, 0.045046154, 0.0030491839, 0.01681367}
+	want = []float32{0.27755088, 0.20409796, 0.15720603, 0.08582123, 0.045046154, 0.043722924, 0.036774673, 0.026986899, 0.01681367, 0.0046718004, 0.00412893, 0.0030491839}
+	got = topK(toTokens(input), 0)
+	if len(got) != len(input) {
+		t.Errorf("topK(-1): wrong length: want %d, got %d", len(input), len(got))
+	}
+	compareLogits(t, "topK(-1)", want, got)
 }
 
 func TestTopP(t *testing.T) {
-	input := []float64{-3, -2, -1, 0, 1, 2, 4}
+	input := []float32{-3, -2, -1, 0, 1, 2, 4}
 	tokens := toTokens(input)
 
 	// First apply temperature and softmax to get probabilities
-	tokens = temperature(tokens, 1)
 	tokens = softmax(tokens)
-	sortLogits(tokens)
+	tokens = topK(tokens, 20)
 
 	// Then apply topP
 	got := topP(tokens, 0.95)
@@ -98,11 +167,10 @@ func TestTopP(t *testing.T) {
 }
 
 func TestMinP(t *testing.T) {
-	input := []float64{-3, -2, -1, 0, 1, 2, 4, 3}
+	input := []float32{-3, -2, -1, 0, 1, 2, 4, 3}
 	tokens := toTokens(input)
 
 	// First apply temperature and softmax
-	tokens = temperature(tokens, 1)
 	tokens = softmax(tokens)
 
 	// Then apply minP
@@ -115,10 +183,10 @@ func TestMinP(t *testing.T) {
 }
 
 func TestSortLogits(t *testing.T) {
-	input := []float64{3, 1, 4, 2, -1, 0, -2}
+	input := []float32{0.026986899, 0.043722924, 0.036774673, 0.27755088, 0.0046718004, 0.08582123, 0.20409796, 0.00412893, 0.15720603, 0.045046154, 0.0030491839, 0.01681367}
 	tokens := toTokens(input)
 
-	sortLogits(tokens)
+	tokens = topK(tokens, 20)
 
 	for i := 1; i < len(tokens); i++ {
 		if tokens[i].value > tokens[i-1].value {
@@ -127,7 +195,7 @@ func TestSortLogits(t *testing.T) {
 		}
 	}
 
-	want := []float64{4, 3, 2, 1, 0, -1, -2}
+	want := []float32{0.27755088, 0.20409796, 0.15720603, 0.08582123, 0.045046154, 0.043722924, 0.036774673, 0.026986899, 0.01681367, 0.0046718004, 0.00412893, 0.0030491839}
 	compareLogits(t, "sortLogits", want, tokens)
 }
 
@@ -151,6 +219,14 @@ func BenchmarkTransforms(b *testing.B) {
 		}
 	})
 
+	b.Run("Softmax", func(b *testing.B) {
+		b.ResetTimer()
+		for b.Loop() {
+			copy(tokensCopy, tokens)
+			softmax(tokensCopy)
+		}
+	})
+
 	b.Run("TopK", func(b *testing.B) {
 		b.ResetTimer()
 		for b.Loop() {
@@ -179,7 +255,7 @@ func BenchmarkTransforms(b *testing.B) {
 		b.ResetTimer()
 		for b.Loop() {
 			copy(tokensCopy, tokens)
-			sortLogits(tokensCopy)
+			topK(tokensCopy, 200000)
 		}
 	})
 }

scripts/build_windows.ps1

@@ -80,13 +80,14 @@ function checkEnv() {
 
 
 function buildOllama() {
+    mkdir -Force -path "${script:DIST_DIR}\"
     if ($script:ARCH -ne "arm64") {
         Remove-Item -ea 0 -recurse -force -path "${script:SRC_DIR}\dist\windows-${script:ARCH}"
         New-Item "${script:SRC_DIR}\dist\windows-${script:ARCH}\lib\ollama\" -ItemType Directory -ea 0
 
         & cmake --fresh --preset CPU --install-prefix $script:DIST_DIR
         if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
-        & cmake --build --preset CPU --parallel $script:JOBS
+        & cmake --build --preset CPU  --config Release --parallel $script:JOBS
         if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
         & cmake --install build --component CPU --strip
         if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
@@ -101,7 +102,7 @@ function buildOllama() {
             # to avoid 2022 (or newer) from being used as the default
             & cmake --fresh --preset "CUDA 11" -G "Visual Studio 16 2019" --install-prefix $script:DIST_DIR
             if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
-            & cmake --build --preset "CUDA 11" --parallel $script:JOBS
+            & cmake --build --preset "CUDA 11"  --config Release --parallel $script:JOBS
             if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
             & cmake --install build --component "CUDA" --strip
             if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
@@ -112,7 +113,7 @@ function buildOllama() {
             write-host "Building CUDA v12 backend libraries"
             & cmake --fresh --preset "CUDA 12" --install-prefix $script:DIST_DIR
             if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
-            & cmake --build --preset "CUDA 12" --parallel $script:JOBS
+            & cmake --build --preset "CUDA 12"  --config Release --parallel $script:JOBS
             if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
             & cmake --install build --component "CUDA" --strip
             if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
@@ -131,7 +132,7 @@ function buildOllama() {
             $env:HIPCXX=""
             $env:HIP_PLATFORM=""
             $env:CMAKE_PREFIX_PATH=""
-            & cmake --build --preset "ROCm" --parallel $script:JOBS
+            & cmake --build --preset "ROCm"  --config Release --parallel $script:JOBS
             if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
             & cmake --install build --component "HIP" --strip
             if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}

server/internal/cache/blob/cache.go

@@ -146,7 +146,7 @@ func debugger(err *error) func(step string) {
 // be in either of the following forms:
 //
 //	@<digest>
-//	<name>
+//	<name>@<digest>
 //	<name>
 //
 // If a digest is provided, it is returned as is and nothing else happens.
@@ -160,8 +160,6 @@ func debugger(err *error) func(step string) {
 // hashed is passed to a PutBytes call to ensure that the manifest is in the
 // blob store. This is done to ensure that future calls to [Get] succeed in
 // these cases.
-//
-// TODO(bmizerany): Move Links/Resolve/etc. out of this package.
 func (c *DiskCache) Resolve(name string) (Digest, error) {
 	name, digest := splitNameDigest(name)
 	if digest != "" {
@@ -279,18 +277,6 @@ func (c *DiskCache) Get(d Digest) (Entry, error) {
 // It returns an error if either the name or digest is invalid, or if link
 // creation encounters any issues.
 func (c *DiskCache) Link(name string, d Digest) error {
-	// TODO(bmizerany): Move link handling from cache to registry.
-	//
-	// We originally placed links in the cache due to its storage
-	// knowledge. However, the registry likely offers better context for
-	// naming concerns, and our API design shouldn't be tightly coupled to
-	// our on-disk format.
-	//
-	// Links work effectively when independent from physical location -
-	// they can reference content with matching SHA regardless of storage
-	// location. In an upcoming change, we plan to shift this
-	// responsibility to the registry where it better aligns with the
-	// system's conceptual model.
 	manifest, err := c.manifestPath(name)
 	if err != nil {
 		return err
@@ -341,7 +327,9 @@ func (c *DiskCache) GetFile(d Digest) string {
 	return absJoin(c.dir, "blobs", filename)
 }
 
-// Links returns a sequence of links in the cache in lexical order.
+// Links returns a sequence of link names. The sequence is in lexical order.
+// Names are converted from their relative path form to their name form but are
+// not guaranteed to be valid. Callers should validate the names before using.
 func (c *DiskCache) Links() iter.Seq2[string, error] {
 	return func(yield func(string, error) bool) {
 		for path, err := range c.links() {
@@ -414,12 +402,14 @@ func (c *DiskCache) links() iter.Seq2[string, error] {
 }
 
 type checkWriter struct {
-	d    Digest
 	size int64
-	n    int64
-	h    hash.Hash
+	d    Digest
 	f    *os.File
-	err  error
+	h    hash.Hash
+
+	w   io.Writer // underlying writer; set by creator
+	n   int64
+	err error
 
 	testHookBeforeFinalWrite func(*os.File)
 }
@@ -435,6 +425,10 @@ func (w *checkWriter) seterr(err error) error {
 // underlying writer is guaranteed to be the last byte of p as verified by the
 // hash.
 func (w *checkWriter) Write(p []byte) (int, error) {
+	if w.err != nil {
+		return 0, w.err
+	}
+
 	_, err := w.h.Write(p)
 	if err != nil {
 		return 0, w.seterr(err)
@@ -453,7 +447,7 @@ func (w *checkWriter) Write(p []byte) (int, error) {
 	if nextSize > w.size {
 		return 0, w.seterr(fmt.Errorf("content exceeds expected size: %d > %d", nextSize, w.size))
 	}
-	n, err := w.f.Write(p)
+	n, err := w.w.Write(p)
 	w.n += int64(n)
 	return n, w.seterr(err)
 }
@@ -493,10 +487,12 @@ func (c *DiskCache) copyNamedFile(name string, file io.Reader, out Digest, size
 
 	// Copy file to f, but also into h to double-check hash.
 	cw := &checkWriter{
-		d:                        out,
-		size:                     size,
-		h:                        sha256.New(),
-		f:                        f,
+		d:    out,
+		size: size,
+		h:    sha256.New(),
+		f:    f,
+		w:    f,
+
 		testHookBeforeFinalWrite: c.testHookBeforeFinalWrite,
 	}
 	n, err := io.Copy(cw, file)
@@ -532,11 +528,6 @@ func splitNameDigest(s string) (name, digest string) {
 var errInvalidName = errors.New("invalid name")
 
 func nameToPath(name string) (_ string, err error) {
-	if strings.Contains(name, "@") {
-		// TODO(bmizerany): HACK: Fix names.Parse to validate.
-		// TODO(bmizerany): merge with default parts (maybe names.Merge(a, b))
-		return "", errInvalidName
-	}
 	n := names.Parse(name)
 	if !n.IsFullyQualified() {
 		return "", errInvalidName
@@ -547,8 +538,7 @@ func nameToPath(name string) (_ string, err error) {
 func absJoin(pp ...string) string {
 	abs, err := filepath.Abs(filepath.Join(pp...))
 	if err != nil {
-		// Likely a bug bug or a bad OS problem. Just panic.
-		panic(err)
+		panic(err) // this should never happen
 	}
 	return abs
 }

server/internal/cache/blob/chunked.go

@@ -0,0 +1,73 @@
+package blob
+
+import (
+	"crypto/sha256"
+	"errors"
+	"io"
+	"os"
+)
+
+// Chunk represents a range of bytes in a blob.
+type Chunk struct {
+	Start int64
+	End   int64
+}
+
+// Size returns end minus start plus one.
+func (c Chunk) Size() int64 {
+	return c.End - c.Start + 1
+}
+
+// Chunker writes to a blob in chunks.
+// Its zero value is invalid. Use [DiskCache.Chunked] to create a new Chunker.
+type Chunker struct {
+	digest Digest
+	size   int64
+	f      *os.File // nil means pre-validated
+}
+
+// Chunked returns a new Chunker, ready for use storing a blob of the given
+// size in chunks.
+//
+// Use [Chunker.Put] to write data to the blob at specific offsets.
+func (c *DiskCache) Chunked(d Digest, size int64) (*Chunker, error) {
+	name := c.GetFile(d)
+	info, err := os.Stat(name)
+	if err == nil && info.Size() == size {
+		return &Chunker{}, nil
+	}
+	f, err := os.OpenFile(name, os.O_CREATE|os.O_WRONLY, 0o666)
+	if err != nil {
+		return nil, err
+	}
+	return &Chunker{digest: d, size: size, f: f}, nil
+}
+
+// Put copies chunk.Size() bytes from r to the blob at the given offset,
+// merging the data with the existing blob. It returns an error if any. As a
+// special case, if r has less than chunk.Size() bytes, Put returns
+// io.ErrUnexpectedEOF.
+func (c *Chunker) Put(chunk Chunk, d Digest, r io.Reader) error {
+	if c.f == nil {
+		return nil
+	}
+
+	cw := &checkWriter{
+		d:    d,
+		size: chunk.Size(),
+		h:    sha256.New(),
+		f:    c.f,
+		w:    io.NewOffsetWriter(c.f, chunk.Start),
+	}
+
+	_, err := io.CopyN(cw, r, chunk.Size())
+	if err != nil && errors.Is(err, io.EOF) {
+		return io.ErrUnexpectedEOF
+	}
+	return err
+}
+
+// Close closes the underlying file.
+func (c *Chunker) Close() error {
+	return c.f.Close()
+}

server/internal/cache/blob/digest.go

@@ -63,6 +63,10 @@ func (d Digest) Short() string {
 	return fmt.Sprintf("%x", d.sum[:4])
 }
 
+func (d Digest) Sum() [32]byte {
+	return d.sum
+}
+
 func (d Digest) Compare(other Digest) int {
 	return slices.Compare(d.sum[:], other.sum[:])
 }

server/internal/chunks/chunks.go

@@ -1,78 +0,0 @@
-package chunks
-
-import (
-	"fmt"
-	"iter"
-	"strconv"
-	"strings"
-)
-
-type Chunk struct {
-	Start, End int64
-}
-
-func New(start, end int64) Chunk {
-	return Chunk{start, end}
-}
-
-// ParseRange parses a string in the form "unit=range" where unit is a string
-// and range is a string in the form "start-end". It returns the unit and the
-// range as a Chunk.
-func ParseRange(s string) (unit string, _ Chunk, _ error) {
-	unit, r, _ := strings.Cut(s, "=")
-	if r == "" {
-		return unit, Chunk{}, nil
-	}
-	c, err := Parse(r)
-	if err != nil {
-		return "", Chunk{}, err
-	}
-	return unit, c, err
-}
-
-// Parse parses a string in the form "start-end" and returns the Chunk.
-func Parse(s string) (Chunk, error) {
-	startStr, endStr, _ := strings.Cut(s, "-")
-	start, err := strconv.ParseInt(startStr, 10, 64)
-	if err != nil {
-		return Chunk{}, fmt.Errorf("invalid start: %v", err)
-	}
-	end, err := strconv.ParseInt(endStr, 10, 64)
-	if err != nil {
-		return Chunk{}, fmt.Errorf("invalid end: %v", err)
-	}
-	if start > end {
-		return Chunk{}, fmt.Errorf("invalid range %d-%d: start > end", start, end)
-	}
-	return Chunk{start, end}, nil
-}
-
-// Of returns a sequence of contiguous Chunks of size chunkSize that cover
-// the range [0, size), in order.
-func Of(size, chunkSize int64) iter.Seq[Chunk] {
-	return func(yield func(Chunk) bool) {
-		for start := int64(0); start < size; start += chunkSize {
-			end := min(start+chunkSize-1, size-1)
-			if !yield(Chunk{start, end}) {
-				break
-			}
-		}
-	}
-}
-
-// Count returns the number of Chunks of size chunkSize needed to cover the
-// range [0, size).
-func Count(size, chunkSize int64) int64 {
-	return (size + chunkSize - 1) / chunkSize
-}
-
-// Size returns end minus start plus one.
-func (c Chunk) Size() int64 {
-	return c.End - c.Start + 1
-}
-
-// String returns the string representation of the Chunk in the form
-// "{start}-{end}".
-func (c Chunk) String() string {
-	return fmt.Sprintf("%d-%d", c.Start, c.End)
-}

server/internal/chunks/chunks_test.go

@@ -1,65 +0,0 @@
-package chunks
-
-import (
-	"slices"
-	"testing"
-)
-
-func TestOf(t *testing.T) {
-	cases := []struct {
-		total     int64
-		chunkSize int64
-		want      []Chunk
-	}{
-		{0, 1, nil},
-		{1, 1, []Chunk{{0, 0}}},
-		{1, 2, []Chunk{{0, 0}}},
-		{2, 1, []Chunk{{0, 0}, {1, 1}}},
-		{10, 9, []Chunk{{0, 8}, {9, 9}}},
-	}
-
-	for _, tt := range cases {
-		got := slices.Collect(Of(tt.total, tt.chunkSize))
-		if !slices.Equal(got, tt.want) {
-			t.Errorf("[%d/%d]: got %v; want %v", tt.total, tt.chunkSize, got, tt.want)
-		}
-	}
-}
-
-func TestSize(t *testing.T) {
-	cases := []struct {
-		c    Chunk
-		want int64
-	}{
-		{Chunk{0, 0}, 1},
-		{Chunk{0, 1}, 2},
-		{Chunk{3, 4}, 2},
-	}
-
-	for _, tt := range cases {
-		got := tt.c.Size()
-		if got != tt.want {
-			t.Errorf("%v: got %d; want %d", tt.c, got, tt.want)
-		}
-	}
-}
-
-func TestCount(t *testing.T) {
-	cases := []struct {
-		total     int64
-		chunkSize int64
-		want      int64
-	}{
-		{0, 1, 0},
-		{1, 1, 1},
-		{1, 2, 1},
-		{2, 1, 2},
-		{10, 9, 2},
-	}
-	for _, tt := range cases {
-		got := Count(tt.total, tt.chunkSize)
-		if got != tt.want {
-			t.Errorf("[%d/%d]: got %d; want %d", tt.total, tt.chunkSize, got, tt.want)
-		}
-	}
-}

server/internal/client/ollama/registry.go

@@ -19,6 +19,7 @@ import (
 	"fmt"
 	"io"
 	"io/fs"
+	"iter"
 	"log/slog"
 	"net/http"
 	"os"
@@ -35,10 +36,8 @@ import (
 	"golang.org/x/sync/errgroup"
 
 	"github.com/ollama/ollama/server/internal/cache/blob"
-	"github.com/ollama/ollama/server/internal/chunks"
 	"github.com/ollama/ollama/server/internal/internal/backoff"
 	"github.com/ollama/ollama/server/internal/internal/names"
-	"github.com/ollama/ollama/server/internal/internal/syncs"
 
 	_ "embed"
 )
@@ -66,12 +65,7 @@ var (
 const (
 	// DefaultChunkingThreshold is the threshold at which a layer should be
 	// split up into chunks when downloading.
-	DefaultChunkingThreshold = 128 << 20
-
-	// DefaultMaxChunkSize is the default maximum size of a chunk to
-	// download. It is configured based on benchmarks and aims to strike a
-	// balance between download speed and memory usage.
-	DefaultMaxChunkSize = 8 << 20
+	DefaultChunkingThreshold = 64 << 20
 )
 
 var defaultCache = sync.OnceValues(func() (*blob.DiskCache, error) {
@@ -211,8 +205,7 @@ type Registry struct {
 	// pushing or pulling models. If zero, the number of streams is
 	// determined by [runtime.GOMAXPROCS].
 	//
-	// Clients that want "unlimited" streams should set this to a large
-	// number.
+	// A negative value means no limit.
 	MaxStreams int
 
 	// ChunkingThreshold is the maximum size of a layer to download in a single
@@ -282,24 +275,13 @@ func DefaultRegistry() (*Registry, error) {
 }
 
 func (r *Registry) maxStreams() int {
-	n := cmp.Or(r.MaxStreams, runtime.GOMAXPROCS(0))
-
-	// Large downloads require a writter stream, so ensure we have at least
-	// two streams to avoid a deadlock.
-	return max(n, 2)
+	return cmp.Or(r.MaxStreams, runtime.GOMAXPROCS(0))
 }
 
 func (r *Registry) maxChunkingThreshold() int64 {
 	return cmp.Or(r.ChunkingThreshold, DefaultChunkingThreshold)
 }
 
-// chunkSizeFor returns the chunk size for a layer of the given size. If the
-// size is less than or equal to the max chunking threshold, the size is
-// returned; otherwise, the max chunk size is returned.
-func (r *Registry) maxChunkSize() int64 {
-	return cmp.Or(r.MaxChunkSize, DefaultMaxChunkSize)
-}
-
 type PushParams struct {
 	// From is an optional destination name for the model. If empty, the
 	// destination name is the same as the source name.
@@ -426,6 +408,21 @@ func canRetry(err error) bool {
 	return re.Status >= 500
 }
 
+// trackingReader is an io.Reader that tracks the number of bytes read and
+// calls the update function with the layer, the number of bytes read.
+//
+// It always calls update with a nil error.
+type trackingReader struct {
+	r io.Reader
+	n *atomic.Int64
+}
+
+func (r *trackingReader) Read(p []byte) (n int, err error) {
+	n, err = r.r.Read(p)
+	r.n.Add(int64(n))
+	return
+}
+
 // Pull pulls the model with the given name from the remote registry into the
 // cache.
 //
@@ -434,11 +431,6 @@ func canRetry(err error) bool {
 // typically slower than splitting the model up across layers, and is mostly
 // utilized for layers of type equal to "application/vnd.ollama.image".
 func (r *Registry) Pull(ctx context.Context, name string) error {
-	scheme, n, _, err := r.parseNameExtended(name)
-	if err != nil {
-		return err
-	}
-
 	m, err := r.Resolve(ctx, name)
 	if err != nil {
 		return err
@@ -457,126 +449,95 @@ func (r *Registry) Pull(ctx context.Context, name string) error {
 		return err == nil && info.Size == l.Size
 	}
 
-	t := traceFromContext(ctx)
-
-	g, ctx := errgroup.WithContext(ctx)
-	g.SetLimit(r.maxStreams())
-
 	layers := m.Layers
 	if m.Config != nil && m.Config.Digest.IsValid() {
 		layers = append(layers, m.Config)
 	}
 
-	for _, l := range layers {
+	// Send initial layer trace events to allow clients to have an
+	// understanding of work to be done before work starts.
+	t := traceFromContext(ctx)
+	skip := make([]bool, len(layers))
+	for i, l := range layers {
+		t.update(l, 0, nil)
 		if exists(l) {
+			skip[i] = true
 			t.update(l, l.Size, ErrCached)
+		}
+	}
+
+	g, ctx := errgroup.WithContext(ctx)
+	g.SetLimit(r.maxStreams())
+	for i, l := range layers {
+		if skip[i] {
 			continue
 		}
 
-		blobURL := fmt.Sprintf("%s://%s/v2/%s/%s/blobs/%s", scheme, n.Host(), n.Namespace(), n.Model(), l.Digest)
-		req, err := r.newRequest(ctx, "GET", blobURL, nil)
+		chunked, err := c.Chunked(l.Digest, l.Size)
 		if err != nil {
 			t.update(l, 0, err)
 			continue
 		}
+		defer chunked.Close()
 
-		t.update(l, 0, nil)
-
-		if l.Size <= r.maxChunkingThreshold() {
-			g.Go(func() error {
-				// TODO(bmizerany): retry/backoff like below in
-				// the chunking case
-				res, err := sendRequest(r.client(), req)
-				if err != nil {
-					return err
-				}
-				defer res.Body.Close()
-				err = c.Put(l.Digest, res.Body, l.Size)
-				if err == nil {
-					t.update(l, l.Size, nil)
-				}
-				return err
-			})
-		} else {
-			q := syncs.NewRelayReader()
+		var progress atomic.Int64
+		for cs, err := range r.chunksums(ctx, name, l) {
+			if err != nil {
+				t.update(l, progress.Load(), err)
+				break
+			}
 
 			g.Go(func() (err error) {
-				defer func() { q.CloseWithError(err) }()
-				return c.Put(l.Digest, q, l.Size)
-			})
+				defer func() { t.update(l, progress.Load(), err) }()
 
-			var progress atomic.Int64
-
-			// We want to avoid extra round trips per chunk due to
-			// redirects from the registry to the blob store, so
-			// fire an initial request to get the final URL and
-			// then use that URL for the chunk requests.
-			req.Header.Set("Range", "bytes=0-0")
-			res, err := sendRequest(r.client(), req)
-			if err != nil {
-				return err
-			}
-			res.Body.Close()
-			req = res.Request.WithContext(req.Context())
-
-			wp := writerPool{size: r.maxChunkSize()}
-
-			for chunk := range chunks.Of(l.Size, r.maxChunkSize()) {
-				if ctx.Err() != nil {
-					break
-				}
-
-				ticket := q.Take()
-				g.Go(func() (err error) {
-					defer func() {
-						if err != nil {
-							q.CloseWithError(err)
-						}
-						ticket.Close()
-						t.update(l, progress.Load(), err)
-					}()
-
-					for _, err := range backoff.Loop(ctx, 3*time.Second) {
-						if err != nil {
-							return err
-						}
-						err := func() error {
-							req := req.Clone(req.Context())
-							req.Header.Set("Range", fmt.Sprintf("bytes=%s", chunk))
-							res, err := sendRequest(r.client(), req)
-							if err != nil {
-								return err
-							}
-							defer res.Body.Close()
-
-							tw := wp.get()
-							tw.Reset(ticket)
-							defer wp.put(tw)
-
-							_, err = io.CopyN(tw, res.Body, chunk.Size())
-							if err != nil {
-								return maybeUnexpectedEOF(err)
-							}
-							if err := tw.Flush(); err != nil {
-								return err
-							}
-
-							total := progress.Add(chunk.Size())
-							if total >= l.Size {
-								q.Close()
-							}
-							return nil
-						}()
-						if !canRetry(err) {
-							return err
-						}
+				for _, err := range backoff.Loop(ctx, 3*time.Second) {
+					if err != nil {
+						return err
 					}
-					return nil
-				})
-			}
+					err := func() error {
+						req, err := http.NewRequestWithContext(ctx, "GET", cs.URL, nil)
+						if err != nil {
+							return err
+						}
+						req.Header.Set("Range", fmt.Sprintf("bytes=%d-%d", cs.Chunk.Start, cs.Chunk.End))
+						res, err := sendRequest(r.client(), req)
+						if err != nil {
+							return err
+						}
+						defer res.Body.Close()
+
+						// Count bytes towards
+						// progress, as they arrive, so
+						// that our bytes piggyback
+						// other chunk updates on
+						// completion.
+						//
+						// This tactic is enough to
+						// show "smooth" progress given
+						// the current CLI client. In
+						// the near future, the server
+						// should report download rate
+						// since it knows better than
+						// a client that is measuring
+						// rate based on wall-clock
+						// time-since-last-update.
+						body := &trackingReader{r: res.Body, n: &progress}
+
+						err = chunked.Put(cs.Chunk, cs.Digest, body)
+						if err != nil {
+							return err
+						}
+
+						return nil
+					}()
+					if !canRetry(err) {
+						return err
+					}
+				}
+				return nil
+			})
 		}
 	}
-
 	if err := g.Wait(); err != nil {
 		return err
 	}
@@ -615,8 +576,6 @@ type Manifest struct {
 	Config *Layer `json:"config"`
 }
 
-var emptyDigest, _ = blob.ParseDigest("sha256:0000000000000000000000000000000000000000000000000000000000000000")
-
 // Layer returns the layer with the given
 // digest, or nil if not found.
 func (m *Manifest) Layer(d blob.Digest) *Layer {
@@ -643,10 +602,9 @@ func (m Manifest) MarshalJSON() ([]byte, error) {
 		// last phase of the commit which expects it, but does nothing
 		// with it. This will be fixed in a future release of
 		// ollama.com.
-		Config *Layer `json:"config"`
+		Config Layer `json:"config"`
 	}{
-		M:      M(m),
-		Config: &Layer{Digest: emptyDigest},
+		M: M(m),
 	}
 	return json.Marshal(v)
 }
@@ -736,6 +694,123 @@ func (r *Registry) Resolve(ctx context.Context, name string) (*Manifest, error)
 	return m, nil
 }
 
+type chunksum struct {
+	URL    string
+	Chunk  blob.Chunk
+	Digest blob.Digest
+}
+
+// chunksums returns a sequence of chunksums for the given layer. If the layer is under the
+// chunking threshold, a single chunksum is returned that covers the entire layer. If the layer
+// is over the chunking threshold, the chunksums are read from the chunksums endpoint.
+func (r *Registry) chunksums(ctx context.Context, name string, l *Layer) iter.Seq2[chunksum, error] {
+	return func(yield func(chunksum, error) bool) {
+		scheme, n, _, err := r.parseNameExtended(name)
+		if err != nil {
+			yield(chunksum{}, err)
+			return
+		}
+
+		if l.Size < r.maxChunkingThreshold() {
+			// any layer under the threshold should be downloaded
+			// in one go.
+			cs := chunksum{
+				URL: fmt.Sprintf("%s://%s/v2/%s/%s/blobs/%s",
+					scheme,
+					n.Host(),
+					n.Namespace(),
+					n.Model(),
+					l.Digest,
+				),
+				Chunk:  blob.Chunk{Start: 0, End: l.Size - 1},
+				Digest: l.Digest,
+			}
+			yield(cs, nil)
+			return
+		}
+
+		// A chunksums response is a sequence of chunksums in a
+		// simple, easy to parse line-oriented format.
+		//
+		// Example:
+		//
+		//     >> GET /v2/<namespace>/<model>/chunksums/<digest>
+		//
+		//     << HTTP/1.1 200 OK
+		//     << Content-Location: <blobURL>
+		//     <<
+		//     << <digest> <start>-<end>
+		//     << ...
+		//
+		// The blobURL is the URL to download the chunks from.
+
+		chunksumsURL := fmt.Sprintf("%s://%s/v2/%s/%s/chunksums/%s",
+			scheme,
+			n.Host(),
+			n.Namespace(),
+			n.Model(),
+			l.Digest,
+		)
+
+		req, err := r.newRequest(ctx, "GET", chunksumsURL, nil)
+		if err != nil {
+			yield(chunksum{}, err)
+			return
+		}
+		res, err := sendRequest(r.client(), req)
+		if err != nil {
+			yield(chunksum{}, err)
+			return
+		}
+		defer res.Body.Close()
+		if res.StatusCode != 200 {
+			err := fmt.Errorf("chunksums: unexpected status code %d", res.StatusCode)
+			yield(chunksum{}, err)
+			return
+		}
+		blobURL := res.Header.Get("Content-Location")
+
+		s := bufio.NewScanner(res.Body)
+		s.Split(bufio.ScanWords)
+		for {
+			if !s.Scan() {
+				if s.Err() != nil {
+					yield(chunksum{}, s.Err())
+				}
+				return
+			}
+			d, err := blob.ParseDigest(s.Bytes())
+			if err != nil {
+				yield(chunksum{}, fmt.Errorf("invalid digest: %q", s.Bytes()))
+				return
+			}
+
+			if !s.Scan() {
+				err := s.Err()
+				if err == nil {
+					err = fmt.Errorf("missing chunk range for digest %s", d)
+				}
+				yield(chunksum{}, err)
+				return
+			}
+			chunk, err := parseChunk(s.Bytes())
+			if err != nil {
+				yield(chunksum{}, fmt.Errorf("invalid chunk range for digest %s: %q", d, s.Bytes()))
+				return
+			}
+
+			cs := chunksum{
+				URL:    blobURL,
+				Chunk:  chunk,
+				Digest: d,
+			}
+			if !yield(cs, nil) {
+				return
+			}
+		}
+	}
+}
+
 func (r *Registry) client() *http.Client {
 	if r.HTTPClient != nil {
 		return r.HTTPClient
@@ -898,13 +973,6 @@ func checkData(url string) string {
 	return fmt.Sprintf("GET,%s,%s", url, zeroSum)
 }
 
-func maybeUnexpectedEOF(err error) error {
-	if errors.Is(err, io.EOF) {
-		return io.ErrUnexpectedEOF
-	}
-	return err
-}
-
 type publicError struct {
 	wrapped error
 	message string
@@ -991,27 +1059,22 @@ func splitExtended(s string) (scheme, name, digest string) {
 	return scheme, s, digest
 }
 
-type writerPool struct {
-	size int64 // set by the caller
-
-	mu sync.Mutex
-	ws []*bufio.Writer
-}
-
-func (p *writerPool) get() *bufio.Writer {
-	p.mu.Lock()
-	defer p.mu.Unlock()
-	if len(p.ws) == 0 {
-		return bufio.NewWriterSize(nil, int(p.size))
+// parseChunk parses a string in the form "start-end" and returns the Chunk.
+func parseChunk[S ~string | ~[]byte](s S) (blob.Chunk, error) {
+	startPart, endPart, found := strings.Cut(string(s), "-")
+	if !found {
+		return blob.Chunk{}, fmt.Errorf("chunks: invalid range %q: missing '-'", s)
 	}
-	w := p.ws[len(p.ws)-1]
-	p.ws = p.ws[:len(p.ws)-1]
-	return w
-}
-
-func (p *writerPool) put(w *bufio.Writer) {
-	p.mu.Lock()
-	defer p.mu.Unlock()
-	w.Reset(nil)
-	p.ws = append(p.ws, w)
+	start, err := strconv.ParseInt(startPart, 10, 64)
+	if err != nil {
+		return blob.Chunk{}, fmt.Errorf("chunks: invalid start to %q: %v", s, err)
+	}
+	end, err := strconv.ParseInt(endPart, 10, 64)
+	if err != nil {
+		return blob.Chunk{}, fmt.Errorf("chunks: invalid end to %q: %v", s, err)
+	}
+	if start > end {
+		return blob.Chunk{}, fmt.Errorf("chunks: invalid range %q: start > end", s)
+	}
+	return blob.Chunk{Start: start, End: end}, nil
 }

server/internal/client/ollama/registry_test.go

@@ -21,7 +21,6 @@ import (
 	"time"
 
 	"github.com/ollama/ollama/server/internal/cache/blob"
-	"github.com/ollama/ollama/server/internal/chunks"
 	"github.com/ollama/ollama/server/internal/testutil"
 )
 
@@ -428,7 +427,7 @@ func TestRegistryPullCached(t *testing.T) {
 	err := rc.Pull(ctx, "single")
 	testutil.Check(t, err)
 
-	want := []int64{6}
+	want := []int64{0, 6}
 	if !errors.Is(errors.Join(errs...), ErrCached) {
 		t.Errorf("errs = %v; want %v", errs, ErrCached)
 	}
@@ -531,54 +530,6 @@ func TestRegistryPullMixedCachedNotCached(t *testing.T) {
 	}
 }
 
-func TestRegistryPullChunking(t *testing.T) {
-	rc, _ := newClient(t, func(w http.ResponseWriter, r *http.Request) {
-		t.Log("request:", r.URL.Host, r.Method, r.URL.Path, r.Header.Get("Range"))
-		if r.URL.Host != "blob.store" {
-			// The production registry redirects to the blob store.
-			http.Redirect(w, r, "http://blob.store"+r.URL.Path, http.StatusFound)
-			return
-		}
-		if strings.Contains(r.URL.Path, "/blobs/") {
-			rng := r.Header.Get("Range")
-			if rng == "" {
-				http.Error(w, "missing range", http.StatusBadRequest)
-				return
-			}
-			_, c, err := chunks.ParseRange(r.Header.Get("Range"))
-			if err != nil {
-				panic(err)
-			}
-			io.WriteString(w, "remote"[c.Start:c.End+1])
-			return
-		}
-		fmt.Fprintf(w, `{"layers":[{"digest":%q,"size":6}]}`, blob.DigestFromBytes("remote"))
-	})
-
-	// Force chunking by setting the threshold to less than the size of the
-	// layer.
-	rc.ChunkingThreshold = 3
-	rc.MaxChunkSize = 3
-
-	var reads []int64
-	ctx := WithTrace(t.Context(), &Trace{
-		Update: func(d *Layer, n int64, err error) {
-			if err != nil {
-				t.Errorf("update %v %d %v", d, n, err)
-			}
-			reads = append(reads, n)
-		},
-	})
-
-	err := rc.Pull(ctx, "remote")
-	testutil.Check(t, err)
-
-	want := []int64{0, 3, 6}
-	if !slices.Equal(reads, want) {
-		t.Errorf("reads = %v; want %v", reads, want)
-	}
-}
-
 func TestRegistryResolveByDigest(t *testing.T) {
 	check := testutil.Checker(t)
 

server/internal/cmd/oppbench/oppbench.go

@@ -1,11 +0,0 @@
-package main
-
-import (
-	"fmt"
-	"os"
-)
-
-func main() {
-	fmt.Println("Run as 'go test -bench=.' to run the benchmarks")
-	os.Exit(1)
-}

server/internal/cmd/oppbench/oppbench_test.go

@@ -1,107 +0,0 @@
-package main
-
-import (
-	"bytes"
-	"context"
-	"fmt"
-	"io"
-	"net/http"
-	"os"
-	"path/filepath"
-	"runtime"
-	"sync/atomic"
-	"testing"
-	"time"
-
-	"github.com/ollama/ollama/server/internal/chunks"
-	"golang.org/x/sync/errgroup"
-)
-
-func BenchmarkDownload(b *testing.B) {
-	run := func(fileSize, chunkSize int64) {
-		name := fmt.Sprintf("size=%d/chunksize=%d", fileSize, chunkSize)
-		b.Run(name, func(b *testing.B) { benchmarkDownload(b, fileSize, chunkSize) })
-	}
-
-	run(100<<20, 8<<20)
-	run(100<<20, 16<<20)
-	run(100<<20, 32<<20)
-	run(100<<20, 64<<20)
-	run(100<<20, 128<<20) // 1 chunk
-}
-
-func run(ctx context.Context, c *http.Client, chunk chunks.Chunk) error {
-	const blobURL = "https://ollama.com/v2/x/x/blobs/sha256-4824460d29f2058aaf6e1118a63a7a197a09bed509f0e7d4e2efb1ee273b447d"
-	req, err := http.NewRequestWithContext(ctx, "GET", blobURL, nil)
-	if err != nil {
-		return err
-	}
-	req.Header.Set("Range", fmt.Sprintf("bytes=%s", chunk))
-	res, err := c.Do(req)
-	if err != nil {
-		return err
-	}
-	defer res.Body.Close()
-
-	_, err = io.CopyN(io.Discard, res.Body, chunk.Size()) // will io.EOF on short read
-	return err
-}
-
-var sleepTime atomic.Int64
-
-func benchmarkDownload(b *testing.B, fileSize, chunkSize int64) {
-	client := &http.Client{
-		Transport: func() http.RoundTripper {
-			tr := http.DefaultTransport.(*http.Transport).Clone()
-			tr.DisableKeepAlives = true
-			return tr
-		}(),
-	}
-	defer client.CloseIdleConnections()
-
-	// warm up the client
-	run(context.Background(), client, chunks.New(0, 1<<20))
-
-	b.SetBytes(fileSize)
-	b.ReportAllocs()
-
-	// Give our CDN a min to breathe between benchmarks.
-	time.Sleep(time.Duration(sleepTime.Swap(3)))
-
-	for b.Loop() {
-		g, ctx := errgroup.WithContext(b.Context())
-		g.SetLimit(runtime.GOMAXPROCS(0))
-		for chunk := range chunks.Of(fileSize, chunkSize) {
-			g.Go(func() error { return run(ctx, client, chunk) })
-		}
-		if err := g.Wait(); err != nil {
-			b.Fatal(err)
-		}
-	}
-}
-
-func BenchmarkWrite(b *testing.B) {
-	b.Run("chunksize=1MiB", func(b *testing.B) { benchmarkWrite(b, 1<<20) })
-}
-
-func benchmarkWrite(b *testing.B, chunkSize int) {
-	b.ReportAllocs()
-
-	dir := b.TempDir()
-	f, err := os.Create(filepath.Join(dir, "write-single"))
-	if err != nil {
-		b.Fatal(err)
-	}
-	defer f.Close()
-
-	data := make([]byte, chunkSize)
-	b.SetBytes(int64(chunkSize))
-	r := bytes.NewReader(data)
-	for b.Loop() {
-		r.Reset(data)
-		_, err := io.Copy(f, r)
-		if err != nil {
-			b.Fatal(err)
-		}
-	}
-}

server/internal/registry/server.go

@@ -1,6 +1,5 @@
-// Package registry provides an http.Handler for handling local Ollama API
-// requests for performing tasks related to the ollama.com model registry and
-// the local disk cache.
+// Package registry implements an http.Handler for handling local Ollama API
+// model management requests. See [Local] for details.
 package registry
 
 import (
@@ -10,6 +9,7 @@ import (
 	"fmt"
 	"io"
 	"log/slog"
+	"maps"
 	"net/http"
 	"sync"
 	"time"
@@ -18,16 +18,11 @@ import (
 	"github.com/ollama/ollama/server/internal/client/ollama"
 )
 
-// Local is an http.Handler for handling local Ollama API requests for
-// performing tasks related to the ollama.com model registry combined with the
-// local disk cache.
+// Local implements an http.Handler for handling local Ollama API model
+// management requests, such as pushing, pulling, and deleting models.
 //
-// It is not concern of Local, or this package, to handle model creation, which
-// proceeds any registry operations for models it produces.
-//
-// NOTE: The package built for dealing with model creation should use
-// [DefaultCache] to access the blob store and not attempt to read or write
-// directly to the blob disk cache.
+// It can be arranged for all unknown requests to be passed through to a
+// fallback handler, if one is provided.
 type Local struct {
 	Client *ollama.Registry // required
 	Logger *slog.Logger     // required
@@ -63,6 +58,7 @@ func (e serverError) Error() string {
 var (
 	errMethodNotAllowed = &serverError{405, "method_not_allowed", "method not allowed"}
 	errNotFound         = &serverError{404, "not_found", "not found"}
+	errModelNotFound    = &serverError{404, "not_found", "model not found"}
 	errInternalError    = &serverError{500, "internal_error", "internal server error"}
 )
 
@@ -175,8 +171,16 @@ func (s *Local) serveHTTP(rec *statusCodeRecorder, r *http.Request) {
 }
 
 type params struct {
-	DeprecatedName string `json:"name"`  // Use [params.model]
-	Model          string `json:"model"` // Use [params.model]
+	// DeprecatedName is the name of the model to push, pull, or delete,
+	// but is deprecated. New clients should use [Model] instead.
+	//
+	// Use [model()] to get the model name for both old and new API requests.
+	DeprecatedName string `json:"name"`
+
+	// Model is the name of the model to push, pull, or delete.
+	//
+	// Use [model()] to get the model name for both old and new API requests.
+	Model string `json:"model"`
 
 	// AllowNonTLS is a flag that indicates a client using HTTP
 	// is doing so, deliberately.
@@ -189,9 +193,18 @@ type params struct {
 	// confusing flags such as this.
 	AllowNonTLS bool `json:"insecure"`
 
-	// ProgressStream is a flag that indicates the client is expecting a stream of
-	// progress updates.
-	ProgressStream bool `json:"stream"`
+	// Stream, if true, will make the server send progress updates in a
+	// streaming of JSON objects. If false, the server will send a single
+	// JSON object with the final status as "success", or an error object
+	// if an error occurred.
+	//
+	// Unfortunately, this API was designed to be a bit awkward. Stream is
+	// defined to default to true if not present, so we need a way to check
+	// if the client decisively it to false. So, we use a pointer to a
+	// bool. Gross.
+	//
+	// Use [stream()] to get the correct value for this field.
+	Stream *bool `json:"stream"`
 }
 
 // model returns the model name for both old and new API requests.
@@ -199,6 +212,13 @@ func (p params) model() string {
 	return cmp.Or(p.Model, p.DeprecatedName)
 }
 
+func (p params) stream() bool {
+	if p.Stream == nil {
+		return true
+	}
+	return *p.Stream
+}
+
 func (s *Local) handleDelete(_ http.ResponseWriter, r *http.Request) error {
 	if r.Method != "DELETE" {
 		return errMethodNotAllowed
@@ -212,16 +232,16 @@ func (s *Local) handleDelete(_ http.ResponseWriter, r *http.Request) error {
 		return err
 	}
 	if !ok {
-		return &serverError{404, "not_found", "model not found"}
+		return errModelNotFound
 	}
-	if s.Prune == nil {
-		return nil
+	if s.Prune != nil {
+		return s.Prune()
 	}
-	return s.Prune()
+	return nil
 }
 
 type progressUpdateJSON struct {
-	Status    string      `json:"status"`
+	Status    string      `json:"status,omitempty,omitzero"`
 	Digest    blob.Digest `json:"digest,omitempty,omitzero"`
 	Total     int64       `json:"total,omitempty,omitzero"`
 	Completed int64       `json:"completed,omitempty,omitzero"`
@@ -237,6 +257,17 @@ func (s *Local) handlePull(w http.ResponseWriter, r *http.Request) error {
 		return err
 	}
 
+	enc := json.NewEncoder(w)
+	if !p.stream() {
+		if err := s.Client.Pull(r.Context(), p.model()); err != nil {
+			if errors.Is(err, ollama.ErrModelNotFound) {
+				return errModelNotFound
+			}
+			return err
+		}
+		return enc.Encode(progressUpdateJSON{Status: "success"})
+	}
+
 	maybeFlush := func() {
 		fl, _ := w.(http.Flusher)
 		if fl != nil {
@@ -246,69 +277,67 @@ func (s *Local) handlePull(w http.ResponseWriter, r *http.Request) error {
 	defer maybeFlush()
 
 	var mu sync.Mutex
-	enc := json.NewEncoder(w)
-	enc.Encode(progressUpdateJSON{Status: "pulling manifest"})
+	progress := make(map[*ollama.Layer]int64)
 
-	ctx := ollama.WithTrace(r.Context(), &ollama.Trace{
-		Update: func(l *ollama.Layer, n int64, err error) {
-			mu.Lock()
-			defer mu.Unlock()
+	progressCopy := make(map[*ollama.Layer]int64, len(progress))
+	pushUpdate := func() {
+		defer maybeFlush()
 
-			// TODO(bmizerany): coalesce these updates; writing per
-			// update is expensive
+		// TODO(bmizerany): This scales poorly with more layers due to
+		// needing to flush out them all in one big update. We _could_
+		// just flush on the changed ones, or just track the whole
+		// download. Needs more thought. This is fine for now.
+		mu.Lock()
+		maps.Copy(progressCopy, progress)
+		mu.Unlock()
+		for l, n := range progress {
 			enc.Encode(progressUpdateJSON{
 				Digest:    l.Digest,
-				Status:    "pulling",
 				Total:     l.Size,
 				Completed: n,
 			})
+		}
+	}
+
+	t := time.NewTicker(time.Hour) // "unstarted" timer
+	start := sync.OnceFunc(func() {
+		pushUpdate()
+		t.Reset(100 * time.Millisecond)
+	})
+	ctx := ollama.WithTrace(r.Context(), &ollama.Trace{
+		Update: func(l *ollama.Layer, n int64, err error) {
+			if n > 0 {
+				start() // flush initial state
+			}
+			mu.Lock()
+			progress[l] = n
+			mu.Unlock()
 		},
 	})
 
 	done := make(chan error, 1)
 	go func() {
-		// TODO(bmizerany): continue to support non-streaming responses
 		done <- s.Client.Pull(ctx, p.model())
 	}()
 
-	func() {
-		t := time.NewTicker(100 * time.Millisecond)
-		defer t.Stop()
-		for {
-			select {
-			case <-t.C:
-				mu.Lock()
-				maybeFlush()
-				mu.Unlock()
-			case err := <-done:
-				if err != nil {
-					var status string
-					if errors.Is(err, ollama.ErrModelNotFound) {
-						status = fmt.Sprintf("error: model %q not found", p.model())
-						enc.Encode(progressUpdateJSON{Status: status})
-					} else {
-						status = fmt.Sprintf("error: %v", err)
-						enc.Encode(progressUpdateJSON{Status: status})
-					}
-					return
+	for {
+		select {
+		case <-t.C:
+			pushUpdate()
+		case err := <-done:
+			pushUpdate()
+			if err != nil {
+				var status string
+				if errors.Is(err, ollama.ErrModelNotFound) {
+					status = fmt.Sprintf("error: model %q not found", p.model())
+				} else {
+					status = fmt.Sprintf("error: %v", err)
 				}
-
-				// These final updates are not strictly necessary, because they have
-				// already happened at this point. Our pull handler code used to do
-				// these steps after, not during, the pull, and they were slow, so we
-				// wanted to provide feedback to users what was happening. For now, we
-				// keep them to not jar users who are used to seeing them. We can phase
-				// them out with a new and nicer UX later. One without progress bars
-				// and digests that no one cares about.
-				enc.Encode(progressUpdateJSON{Status: "verifying layers"})
-				enc.Encode(progressUpdateJSON{Status: "writing manifest"})
-				enc.Encode(progressUpdateJSON{Status: "success"})
-				return
+				enc.Encode(progressUpdateJSON{Status: status})
 			}
+			return nil
 		}
-	}()
-
-	return nil
+	}
 }
 
 func decodeUserJSON[T any](r io.Reader) (T, error) {

server/internal/registry/server_test.go

@@ -4,7 +4,6 @@ import (
 	"bytes"
 	"context"
 	"encoding/json"
-	"fmt"
 	"io"
 	"io/fs"
 	"net"
@@ -160,7 +159,6 @@ var registryFS = sync.OnceValue(func() fs.FS {
 	// to \n when parsing the txtar on Windows.
 	data := bytes.ReplaceAll(registryTXT, []byte("\r\n"), []byte("\n"))
 	a := txtar.Parse(data)
-	fmt.Printf("%q\n", a.Comment)
 	fsys, err := txtar.FS(a)
 	if err != nil {
 		panic(err)
@@ -179,7 +177,7 @@ func TestServerPull(t *testing.T) {
 			w.WriteHeader(404)
 			io.WriteString(w, `{"errors": [{"code": "MANIFEST_UNKNOWN", "message": "manifest unknown"}]}`)
 		default:
-			t.Logf("serving file: %s", r.URL.Path)
+			t.Logf("serving blob: %s", r.URL.Path)
 			modelsHandler.ServeHTTP(w, r)
 		}
 	})
@@ -188,7 +186,7 @@ func TestServerPull(t *testing.T) {
 		t.Helper()
 
 		if got.Code != 200 {
-			t.Fatalf("Code = %d; want 200", got.Code)
+			t.Errorf("Code = %d; want 200", got.Code)
 		}
 		gotlines := got.Body.String()
 		t.Logf("got:\n%s", gotlines)
@@ -197,35 +195,29 @@ func TestServerPull(t *testing.T) {
 			want, unwanted := strings.CutPrefix(want, "!")
 			want = strings.TrimSpace(want)
 			if !unwanted && !strings.Contains(gotlines, want) {
-				t.Fatalf("! missing %q in body", want)
+				t.Errorf("! missing %q in body", want)
 			}
 			if unwanted && strings.Contains(gotlines, want) {
-				t.Fatalf("! unexpected %q in body", want)
+				t.Errorf("! unexpected %q in body", want)
 			}
 		}
 	}
 
 	got := s.send(t, "POST", "/api/pull", `{"model": "BOOM"}`)
 	checkResponse(got, `
-		{"status":"pulling manifest"}
 		{"status":"error: request error https://example.com/v2/library/BOOM/manifests/latest: registry responded with status 999: boom"}
 	`)
 
 	got = s.send(t, "POST", "/api/pull", `{"model": "smol"}`)
 	checkResponse(got, `
-		{"status":"pulling manifest"}
-		{"status":"pulling","digest":"sha256:68e0ec597aee59d35f8dc44942d7b17d471ade10d3aca07a5bb7177713950312","total":5}
-		{"status":"pulling","digest":"sha256:ca3d163bab055381827226140568f3bef7eaac187cebd76878e0b63e9e442356","total":3}
-		{"status":"pulling","digest":"sha256:68e0ec597aee59d35f8dc44942d7b17d471ade10d3aca07a5bb7177713950312","total":5,"completed":5}
-		{"status":"pulling","digest":"sha256:ca3d163bab055381827226140568f3bef7eaac187cebd76878e0b63e9e442356","total":3,"completed":3}
-		{"status":"verifying layers"}
-		{"status":"writing manifest"}
-		{"status":"success"}
+		{"digest":"sha256:68e0ec597aee59d35f8dc44942d7b17d471ade10d3aca07a5bb7177713950312","total":5}
+		{"digest":"sha256:ca3d163bab055381827226140568f3bef7eaac187cebd76878e0b63e9e442356","total":3}
+		{"digest":"sha256:68e0ec597aee59d35f8dc44942d7b17d471ade10d3aca07a5bb7177713950312","total":5,"completed":5}
+		{"digest":"sha256:ca3d163bab055381827226140568f3bef7eaac187cebd76878e0b63e9e442356","total":3,"completed":3}
 	`)
 
 	got = s.send(t, "POST", "/api/pull", `{"model": "unknown"}`)
 	checkResponse(got, `
-		{"status":"pulling manifest"}
 		{"status":"error: model \"unknown\" not found"}
 	`)
 
@@ -240,19 +232,39 @@ func TestServerPull(t *testing.T) {
 
 	got = s.send(t, "POST", "/api/pull", `{"model": "://"}`)
 	checkResponse(got, `
-		{"status":"pulling manifest"}
 		{"status":"error: invalid or missing name: \"\""}
-
-		!verifying
-		!writing
-		!success
 	`)
+
+	// Non-streaming pulls
+	got = s.send(t, "POST", "/api/pull", `{"model": "://", "stream": false}`)
+	checkErrorResponse(t, got, 400, "bad_request", "invalid or missing name")
+	got = s.send(t, "POST", "/api/pull", `{"model": "smol", "stream": false}`)
+	checkResponse(got, `
+		{"status":"success"}
+		!digest
+		!total
+		!completed
+	`)
+	got = s.send(t, "POST", "/api/pull", `{"model": "unknown", "stream": false}`)
+	checkErrorResponse(t, got, 404, "not_found", "model not found")
 }
 
 func TestServerUnknownPath(t *testing.T) {
 	s := newTestServer(t, nil)
 	got := s.send(t, "DELETE", "/api/unknown", `{}`)
 	checkErrorResponse(t, got, 404, "not_found", "not found")
+
+	var fellback bool
+	s.Fallback = http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		fellback = true
+	})
+	got = s.send(t, "DELETE", "/api/unknown", `{}`)
+	if !fellback {
+		t.Fatal("expected Fallback to be called")
+	}
+	if got.Code != 200 {
+		t.Fatalf("Code = %d; want 200", got.Code)
+	}
 }
 
 func checkErrorResponse(t *testing.T, got *httptest.ResponseRecorder, status int, code, msg string) {

server/prompt.go

@@ -26,6 +26,7 @@ func chatPrompt(ctx context.Context, m *Model, tokenize tokenizeFunc, opts *api.
 	var system []api.Message
 
 	isMllama := checkMllamaModelFamily(m)
+	isGemma3 := checkGemma3ModelFamily(m)
 
 	var imageNumTokens int
 	// TODO: Ideally we would compute this from the projector metadata but some pieces are implementation dependent
@@ -40,7 +41,7 @@ func chatPrompt(ctx context.Context, m *Model, tokenize tokenizeFunc, opts *api.
 	n := len(msgs) - 1
 	// in reverse, find all messages that fit into context window
 	for i := n; i >= 0; i-- {
-		if isMllama && len(msgs[i].Images) > 1 {
+		if (isMllama || isGemma3) && len(msgs[i].Images) > 1 {
 			return "", nil, errTooManyImages
 		}
 
@@ -157,3 +158,12 @@ func checkMllamaModelFamily(m *Model) bool {
 	}
 	return false
 }
+
+func checkGemma3ModelFamily(m *Model) bool {
+	for _, arch := range m.Config.ModelFamilies {
+		if arch == "gemma3" {
+			return true
+		}
+	}
+	return false
+}

server/routes.go

@@ -435,7 +435,7 @@ func (s *Server) EmbedHandler(c *gin.Context) {
 		return
 	}
 
-	kvData, err := getKVData(m.ModelPath, false)
+	kvData, _, err := getModelData(m.ModelPath, false)
 	if err != nil {
 		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
 		return
@@ -483,8 +483,7 @@ func (s *Server) EmbedHandler(c *gin.Context) {
 	}
 
 	if err := g.Wait(); err != nil {
-		slog.Error("embedding generation failed", "error", err)
-		c.JSON(http.StatusInternalServerError, gin.H{"error": fmt.Errorf("failed to generate embeddings: %v", err)})
+		c.AbortWithStatusJSON(http.StatusInternalServerError, gin.H{"error": strings.TrimSpace(err.Error())})
 		return
 	}
 
@@ -545,8 +544,7 @@ func (s *Server) EmbeddingsHandler(c *gin.Context) {
 
 	embedding, err := r.Embedding(c.Request.Context(), req.Prompt)
 	if err != nil {
-		slog.Info(fmt.Sprintf("embedding generation failed: %v", err))
-		c.JSON(http.StatusInternalServerError, gin.H{"error": fmt.Errorf("failed to generate embedding: %v", err)})
+		c.AbortWithStatusJSON(http.StatusInternalServerError, gin.H{"error": strings.TrimSpace(err.Error())})
 		return
 	}
 
@@ -850,16 +848,23 @@ func GetModelInfo(req api.ShowRequest) (*api.ShowResponse, error) {
 	fmt.Fprint(&sb, m.String())
 	resp.Modelfile = sb.String()
 
-	kvData, err := getKVData(m.ModelPath, req.Verbose)
+	kvData, tensors, err := getModelData(m.ModelPath, req.Verbose)
 	if err != nil {
 		return nil, err
 	}
+
 	delete(kvData, "general.name")
 	delete(kvData, "tokenizer.chat_template")
 	resp.ModelInfo = kvData
 
+	tensorData := make([]api.Tensor, len(tensors.Items()))
+	for cnt, t := range tensors.Items() {
+		tensorData[cnt] = api.Tensor{Name: t.Name, Type: t.Type(), Shape: t.Shape}
+	}
+	resp.Tensors = tensorData
+
 	if len(m.ProjectorPaths) > 0 {
-		projectorData, err := getKVData(m.ProjectorPaths[0], req.Verbose)
+		projectorData, _, err := getModelData(m.ProjectorPaths[0], req.Verbose)
 		if err != nil {
 			return nil, err
 		}
@@ -869,17 +874,17 @@ func GetModelInfo(req api.ShowRequest) (*api.ShowResponse, error) {
 	return resp, nil
 }
 
-func getKVData(digest string, verbose bool) (ggml.KV, error) {
+func getModelData(digest string, verbose bool) (ggml.KV, ggml.Tensors, error) {
 	maxArraySize := 0
 	if verbose {
 		maxArraySize = -1
 	}
-	kvData, err := llm.LoadModel(digest, maxArraySize)
+	data, err := llm.LoadModel(digest, maxArraySize)
 	if err != nil {
-		return nil, err
+		return nil, ggml.Tensors{}, err
 	}
 
-	kv := kvData.KV()
+	kv := data.KV()
 
 	if !verbose {
 		for k := range kv {
@@ -889,7 +894,7 @@ func getKVData(digest string, verbose bool) (ggml.KV, error) {
 		}
 	}
 
-	return kv, nil
+	return kv, data.Tensors(), nil
 }
 
 func (s *Server) ListHandler(c *gin.Context) {