add docs.json

README.md

@@ -411,8 +411,6 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [ollama launcher](https://github.com/NGC13009/ollama-launcher) (A launcher for Ollama, aiming to provide users with convenient functions such as ollama server launching, management, or configuration.)
 - [ai-hub](https://github.com/Aj-Seven/ai-hub) (AI Hub supports multiple models via API keys and Chat support via Ollama API.)
 - [Mayan EDMS](https://gitlab.com/mayan-edms/mayan-edms) (Open source document management system to organize, tag, search, and automate your files with powerful Ollama driven workflows.)
-- [Serene Pub](https://github.com/doolijb/serene-pub) (Beginner friendly, open source AI Roleplaying App for Windows, Mac OS and Linux. Search, download and use models with Ollama all inside the app.)
-- [Andes](https://github.com/aqerd/andes) (A Visual Studio Code extension that provides a local UI interface for Ollama models)
 
 ### Cloud
 
@@ -539,9 +537,6 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Nichey](https://github.com/goodreasonai/nichey) is a Python package for generating custom wikis for your research topic
 - [Ollama for D](https://github.com/kassane/ollama-d)
 - [OllamaPlusPlus](https://github.com/HardCodeDev777/OllamaPlusPlus) (Very simple C++ library for Ollama)
-- [any-llm](https://github.com/mozilla-ai/any-llm) (A single interface to use different llm providers by [mozilla.ai](https://www.mozilla.ai/))
-- [any-agent](https://github.com/mozilla-ai/any-agent) (A single interface to use and evaluate different agent frameworks by [mozilla.ai](https://www.mozilla.ai/))
-- [Neuro SAN](https://github.com/cognizant-ai-lab/neuro-san-studio) (Data-driven multi-agent orchestration framework) with [example](https://github.com/cognizant-ai-lab/neuro-san-studio/blob/main/docs/user_guide.md#ollama)
 
 ### Mobile
 
@@ -602,7 +597,6 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [UnityCodeLama](https://github.com/HardCodeDev777/UnityCodeLama) (Unity Edtior tool to analyze scripts via Ollama)
 - [NativeMind](https://github.com/NativeMindBrowser/NativeMindExtension) (Private, on-device AI Assistant, no cloud dependencies)
 - [GMAI - Gradle Managed AI](https://gmai.premex.se/) (Gradle plugin for automated Ollama lifecycle management during build phases)
-- [NOMYO Router](https://github.com/nomyo-ai/nomyo-router) (A transparent Ollama proxy with model deployment aware routing which auto-manages multiple Ollama instances in a given network)
 
 ### Supported backends
 

api/types.go

@@ -90,10 +90,6 @@ type GenerateRequest struct {
 	// (request that thinking _not_ be used) and unset (use the old behavior
 	// before this option was introduced)
 	Think *ThinkValue `json:"think,omitempty"`
-
-	// DebugRenderOnly is a debug option that, when set to true, returns the rendered
-	// template instead of calling the model.
-	DebugRenderOnly bool `json:"_debug_render_only,omitempty"`
 }
 
 // ChatRequest describes a request sent by [Client.Chat].
@@ -124,10 +120,6 @@ type ChatRequest struct {
 	// responding. Can be a boolean (true/false) or a string ("high", "medium", "low")
 	// for supported models.
 	Think *ThinkValue `json:"think,omitempty"`
-
-	// DebugRenderOnly is a debug option that, when set to true, returns the rendered
-	// template instead of calling the model.
-	DebugRenderOnly bool `json:"_debug_render_only,omitempty"`
 }
 
 type Tools []Tool
@@ -286,23 +278,16 @@ func mapToTypeScriptType(jsonType string) string {
 	}
 }
 
-type ToolFunctionParameters struct {
-	Type       string                  `json:"type"`
-	Defs       any                     `json:"$defs,omitempty"`
-	Items      any                     `json:"items,omitempty"`
-	Required   []string                `json:"required"`
-	Properties map[string]ToolProperty `json:"properties"`
-}
-
-func (t *ToolFunctionParameters) String() string {
-	bts, _ := json.Marshal(t)
-	return string(bts)
-}
-
 type ToolFunction struct {
-	Name        string                 `json:"name"`
-	Description string                 `json:"description"`
-	Parameters  ToolFunctionParameters `json:"parameters"`
+	Name        string `json:"name"`
+	Description string `json:"description"`
+	Parameters  struct {
+		Type       string                  `json:"type"`
+		Defs       any                     `json:"$defs,omitempty"`
+		Items      any                     `json:"items,omitempty"`
+		Required   []string                `json:"required"`
+		Properties map[string]ToolProperty `json:"properties"`
+	} `json:"parameters"`
 }
 
 func (t *ToolFunction) String() string {
@@ -323,19 +308,6 @@ type ChatResponse struct {
 	Metrics
 }
 
-// DebugInfo contains debug information for template rendering
-type DebugInfo struct {
-	RenderedTemplate string `json:"rendered_template"`
-	ImageCount       int    `json:"image_count,omitempty"`
-}
-
-// DebugTemplateResponse is returned when _debug_render_only is set to true
-type DebugTemplateResponse struct {
-	Model     string    `json:"model"`
-	CreatedAt time.Time `json:"created_at"`
-	DebugInfo DebugInfo `json:"_debug_info"`
-}
-
 type Metrics struct {
 	TotalDuration      time.Duration `json:"total_duration,omitempty"`
 	LoadDuration       time.Duration `json:"load_duration,omitempty"`
@@ -888,7 +860,7 @@ func (d *Duration) UnmarshalJSON(b []byte) (err error) {
 		if t < 0 {
 			d.Duration = time.Duration(math.MaxInt64)
 		} else {
-			d.Duration = time.Duration(t * float64(time.Second))
+			d.Duration = time.Duration(int(t) * int(time.Second))
 		}
 	case string:
 		d.Duration, err = time.ParseDuration(t)

api/types_test.go

@@ -17,11 +17,6 @@ func TestKeepAliveParsingFromJSON(t *testing.T) {
 		req  string
 		exp  *Duration
 	}{
-		{
-			name: "Unset",
-			req:  `{ }`,
-			exp:  nil,
-		},
 		{
 			name: "Positive Integer",
 			req:  `{ "keep_alive": 42 }`,
@@ -30,7 +25,7 @@ func TestKeepAliveParsingFromJSON(t *testing.T) {
 		{
 			name: "Positive Float",
 			req:  `{ "keep_alive": 42.5 }`,
-			exp:  &Duration{42500 * time.Millisecond},
+			exp:  &Duration{42 * time.Second},
 		},
 		{
 			name: "Positive Integer String",
@@ -441,50 +436,3 @@ func TestThinking_UnmarshalJSON(t *testing.T) {
 		})
 	}
 }
-
-func TestToolFunctionParameters_String(t *testing.T) {
-	tests := []struct {
-		name     string
-		params   ToolFunctionParameters
-		expected string
-	}{
-		{
-			name: "simple object with string property",
-			params: ToolFunctionParameters{
-				Type:     "object",
-				Required: []string{"name"},
-				Properties: map[string]ToolProperty{
-					"name": {
-						Type:        PropertyType{"string"},
-						Description: "The name of the person",
-					},
-				},
-			},
-			expected: `{"type":"object","required":["name"],"properties":{"name":{"type":"string","description":"The name of the person"}}}`,
-		},
-		{
-			name: "marshal failure returns empty string",
-			params: ToolFunctionParameters{
-				Type: "object",
-				Defs: func() any {
-					// Create a cycle that will cause json.Marshal to fail
-					type selfRef struct {
-						Self *selfRef
-					}
-					s := &selfRef{}
-					s.Self = s
-					return s
-				}(),
-				Properties: map[string]ToolProperty{},
-			},
-			expected: "",
-		},
-	}
-
-	for _, test := range tests {
-		t.Run(test.name, func(t *testing.T) {
-			result := test.params.String()
-			assert.Equal(t, test.expected, result)
-		})
-	}
-}

convert/convert_gptoss.go

@@ -15,24 +15,19 @@ import (
 
 type gptossModel struct {
 	ModelParameters
-	HiddenLayers          uint32  `json:"num_hidden_layers"`
-	MaxPositionEmbeddings uint32  `json:"max_position_embeddings"`
-	HiddenSize            uint32  `json:"hidden_size"`
-	IntermediateSize      uint32  `json:"intermediate_size"`
-	AttentionHeads        uint32  `json:"num_attention_heads"`
-	KeyValueHeads         uint32  `json:"num_key_value_heads"`
-	HeadDim               uint32  `json:"head_dim"`
-	Experts               uint32  `json:"num_experts"`
-	LocalExperts          uint32  `json:"num_local_experts"`
-	ExpertsPerToken       uint32  `json:"experts_per_token"`
-	RMSNormEpsilon        float32 `json:"rms_norm_eps"`
-	InitialContextLength  uint32  `json:"initial_context_length"`
-	RopeTheta             float32 `json:"rope_theta"`
-	RopeScalingFactor     float32 `json:"rope_scaling_factor"`
-	RopeScaling           struct {
-		Factor float32 `json:"factor"`
-	} `json:"rope_scaling"`
-	SlidingWindow uint32 `json:"sliding_window"`
+	HiddenLayers         uint32  `json:"num_hidden_layers"`
+	HiddenSize           uint32  `json:"hidden_size"`
+	IntermediateSize     uint32  `json:"intermediate_size"`
+	AttentionHeads       uint32  `json:"num_attention_heads"`
+	KeyValueHeads        uint32  `json:"num_key_value_heads"`
+	HeadDim              uint32  `json:"head_dim"`
+	Experts              uint32  `json:"num_experts"`
+	ExpertsPerToken      uint32  `json:"experts_per_token"`
+	RMSNormEpsilon       float32 `json:"rms_norm_eps"`
+	InitialContextLength uint32  `json:"initial_context_length"`
+	RopeTheta            float32 `json:"rope_theta"`
+	RopeScalingFactor    float32 `json:"rope_scaling_factor"`
+	SlidingWindow        uint32  `json:"sliding_window"`
 }
 
 var _ ModelConverter = (*gptossModel)(nil)
@@ -41,11 +36,11 @@ func (m *gptossModel) KV(t *Tokenizer) ggml.KV {
 	kv := m.ModelParameters.KV(t)
 	kv["general.architecture"] = "gptoss"
 	kv["general.file_type"] = uint32(4)
-	kv["gptoss.context_length"] = cmp.Or(m.MaxPositionEmbeddings, uint32(m.RopeScalingFactor*float32(m.InitialContextLength)))
+	kv["gptoss.context_length"] = uint32(m.RopeScalingFactor * float32(m.InitialContextLength))
 	kv["gptoss.block_count"] = m.HiddenLayers
 	kv["gptoss.embedding_length"] = m.HiddenSize
 	kv["gptoss.feed_forward_length"] = m.IntermediateSize
-	kv["gptoss.expert_count"] = cmp.Or(m.Experts, m.LocalExperts)
+	kv["gptoss.expert_count"] = m.Experts
 	kv["gptoss.expert_used_count"] = m.ExpertsPerToken
 	kv["gptoss.attention.head_count"] = m.AttentionHeads
 	kv["gptoss.attention.head_count_kv"] = m.KeyValueHeads
@@ -54,7 +49,7 @@ func (m *gptossModel) KV(t *Tokenizer) ggml.KV {
 	kv["gptoss.attention.layer_norm_rms_epsilon"] = cmp.Or(m.RMSNormEpsilon, 1e-5)
 	kv["gptoss.attention.sliding_window"] = m.SlidingWindow
 	kv["gptoss.rope.freq_base"] = m.RopeTheta
-	kv["gptoss.rope.scaling.factor"] = cmp.Or(m.RopeScalingFactor, m.RopeScaling.Factor)
+	kv["gptoss.rope.scaling.factor"] = m.RopeScalingFactor
 	kv["gptoss.rope.scaling.original_context_length"] = m.InitialContextLength
 	kv["tokenizer.ggml.bos_token_id"] = uint32(199998) // <|startoftext|>
 	kv["tokenizer.ggml.add_bos_token"] = false
@@ -97,11 +92,6 @@ func (m *gptossModel) Tensors(ts []Tensor) []*ggml.Tensor {
 
 	for name, mxfp4 := range mxfp4s {
 		dims := mxfp4.blocks.Shape()
-
-		if !strings.HasSuffix(name, ".weight") {
-			name += ".weight"
-		}
-
 		out = append(out, &ggml.Tensor{
 			Name:     name,
 			Kind:     uint32(ggml.TensorTypeMXFP4),
@@ -114,47 +104,25 @@ func (m *gptossModel) Tensors(ts []Tensor) []*ggml.Tensor {
 }
 
 func (m *gptossModel) Replacements() []string {
-	var replacements []string
-	if m.MaxPositionEmbeddings > 0 {
-		// hf flavored model
-		replacements = []string{
-			"lm_head", "output",
-			"model.embed_tokens", "token_embd",
-			"model.layers", "blk",
-			"input_layernorm", "attn_norm",
-			"self_attn.q_proj", "attn_q",
-			"self_attn.k_proj", "attn_k",
-			"self_attn.v_proj", "attn_v",
-			"self_attn.o_proj", "attn_out",
-			"self_attn.sinks", "attn_sinks",
-			"post_attention_layernorm", "ffn_norm",
-			"mlp.router", "ffn_gate_inp",
-			"mlp.experts.gate_up_proj_", "ffn_gate_up_exps.",
-			"mlp.experts.down_proj_", "ffn_down_exps.",
-			"model.norm", "output_norm",
-		}
-	} else {
-		replacements = []string{
-			// noop replacements so other replacements will not be applied
-			".blocks", ".blocks",
-			".scales", ".scales",
-			// real replacements
-			"block", "blk",
-			"attn.norm", "attn_norm",
-			"attn.qkv", "attn_qkv",
-			"attn.sinks", "attn_sinks",
-			"attn.out", "attn_out",
-			"mlp.norm", "ffn_norm",
-			"mlp.gate", "ffn_gate_inp",
-			"mlp.mlp1_", "ffn_gate_up_exps.",
-			"mlp.mlp2_", "ffn_down_exps.",
-			"embedding", "token_embd",
-			"norm", "output_norm",
-			"unembedding", "output",
-			"scale", "weight",
-		}
+	return []string{
+		// noop replacements so other replacements will not be applied
+		".blocks", ".blocks",
+		".scales", ".scales",
+		// real replacements
+		"block", "blk",
+		"attn.norm", "attn_norm",
+		"attn.qkv", "attn_qkv",
+		"attn.sinks", "attn_sinks",
+		"attn.out", "attn_out",
+		"mlp.norm", "ffn_norm",
+		"mlp.gate", "ffn_gate_inp",
+		"mlp.mlp1_", "ffn_gate_up_exps.",
+		"mlp.mlp2_", "ffn_down_exps.",
+		"embedding", "token_embd",
+		"norm", "output_norm",
+		"unembedding", "output",
+		"scale", "weight",
 	}
-	return replacements
 }
 
 type mxfp4 struct {
@@ -172,20 +140,7 @@ func (m *mxfp4) WriteTo(w io.Writer) (int64, error) {
 		blocksDims[i] = int(d)
 	}
 
-	bts := b.Bytes()
-	var tmp [16]byte
-	for i := 0; i < b.Len(); i += 16 {
-		for j := range 8 {
-			// transform a1b2c3 ... x7y8z9 -> 71xa82yb93zc
-			a, b := bts[i+j], bts[i+j+8]
-			tmp[2*j+0] = (a & 0x0F) | (b << 4)
-			tmp[2*j+1] = (a >> 4) | (b & 0xF0)
-		}
-
-		copy(bts[i:i+16], tmp[:])
-	}
-
-	var blocks tensor.Tensor = tensor.New(tensor.WithShape(blocksDims...), tensor.WithBacking(bts))
+	var blocks tensor.Tensor = tensor.New(tensor.WithShape(blocksDims...), tensor.WithBacking(b.Bytes()))
 
 	var s bytes.Buffer
 	if _, err := m.scales.WriteTo(&s); err != nil {
@@ -219,5 +174,5 @@ func (m *mxfp4) WriteTo(w io.Writer) (int64, error) {
 		return 0, err
 	}
 
-	return int64(len(u8s)), nil
+	return 0, nil
 }

convert/reader.go

@@ -33,8 +33,8 @@ func (t tensorBase) Shape() []uint64 {
 const (
 	tensorKindFP32 uint32 = iota
 	tensorKindFP16
+	tensorKindMXFP4 = 4
 	tensorKindBF16  = 30
-	tensorKindMXFP4 = 39
 )
 
 func (t tensorBase) Kind() uint32 {

convert/reader_safetensors.go

@@ -188,17 +188,17 @@ func (st safetensor) WriteTo(w io.Writer) (int64, error) {
 
 	switch st.Kind() {
 	case tensorKindFP32:
-		return int64(len(f32s) * 4), binary.Write(w, binary.LittleEndian, f32s)
+		return 0, binary.Write(w, binary.LittleEndian, f32s)
 	case tensorKindFP16:
 		f16s := make([]uint16, len(f32s))
 		for i := range f32s {
 			f16s[i] = float16.Fromfloat32(f32s[i]).Bits()
 		}
 
-		return int64(len(f16s) * 2), binary.Write(w, binary.LittleEndian, f16s)
+		return 0, binary.Write(w, binary.LittleEndian, f16s)
 	case tensorKindBF16:
 		u8s := bfloat16.EncodeFloat32(f32s)
-		return int64(len(u8s)), binary.Write(w, binary.LittleEndian, u8s)
+		return 0, binary.Write(w, binary.LittleEndian, u8s)
 	default:
 		return 0, fmt.Errorf("unknown storage type: %d", st.Kind())
 	}

discover/amd_linux.go

@@ -277,7 +277,6 @@ func AMDGetGPUInfo() ([]RocmGPUInfo, error) {
 					FreeMemory:  (totalMemory - usedMemory),
 				},
 				ID:            ID,
-				filterID:      gpuOrdinalID,
 				Name:          name,
 				Compute:       fmt.Sprintf("gfx%d%x%x", major, minor, patch),
 				MinimumMemory: rocmMinimumMemory,
@@ -395,7 +394,7 @@ func AMDGetGPUInfo() ([]RocmGPUInfo, error) {
 
 		// Check for env var workarounds
 		if name == "1002:687f" { // Vega RX 56
-			gpuInfo.EnvWorkarounds = append(gpuInfo.EnvWorkarounds, "HSA_ENABLE_SDMA=0")
+			gpuInfo.EnvWorkarounds = append(gpuInfo.EnvWorkarounds, [2]string{"HSA_ENABLE_SDMA", "0"})
 		}
 
 		// The GPU has passed all the verification steps and is supported
@@ -524,26 +523,19 @@ func verifyKFDDriverAccess() error {
 	return nil
 }
 
-func rocmGetVisibleDevicesEnv(gpuInfo []GpuInfo) string {
+func rocmGetVisibleDevicesEnv(gpuInfo []GpuInfo) (string, string) {
 	ids := []string{}
 	for _, info := range gpuInfo {
 		if info.Library != "rocm" {
+			// TODO shouldn't happen if things are wired correctly...
+			slog.Debug("rocmGetVisibleDevicesEnv skipping over non-rocm device", "library", info.Library)
 			continue
 		}
-		// If the devices requires a numeric ID, for filtering purposes, we use the unfiltered ID number
-		if _, err := strconv.Atoi(info.ID); err == nil {
-			ids = append(ids, fmt.Sprintf("%d", info.filterID))
-		} else {
-			ids = append(ids, info.ID)
-		}
+		ids = append(ids, info.ID)
 	}
-	if len(ids) == 0 {
-		return ""
-	}
-
 	// There are 3 potential env vars to use to select GPUs.
 	// ROCR_VISIBLE_DEVICES supports UUID or numeric so is our preferred on linux
 	// GPU_DEVICE_ORDINAL supports numeric IDs only
 	// HIP_VISIBLE_DEVICES supports numeric IDs only
-	return "ROCR_VISIBLE_DEVICES=" + strings.Join(ids, ",")
+	return "ROCR_VISIBLE_DEVICES", strings.Join(ids, ",")
 }

discover/amd_windows.go

@@ -111,7 +111,6 @@ func AMDGetGPUInfo() ([]RocmGPUInfo, error) {
 				UnreliableFreeMemory: true,
 
 				ID:             strconv.Itoa(i), // TODO this is probably wrong if we specify visible devices
-				filterID:       i,
 				DependencyPath: []string{libDir},
 				MinimumMemory:  rocmMinimumMemory,
 				Name:           name,
@@ -201,26 +200,19 @@ func (gpus RocmGPUInfoList) RefreshFreeMemory() error {
 	return nil
 }
 
-func rocmGetVisibleDevicesEnv(gpuInfo []GpuInfo) string {
+func rocmGetVisibleDevicesEnv(gpuInfo []GpuInfo) (string, string) {
 	ids := []string{}
 	for _, info := range gpuInfo {
 		if info.Library != "rocm" {
+			// TODO shouldn't happen if things are wired correctly...
+			slog.Debug("rocmGetVisibleDevicesEnv skipping over non-rocm device", "library", info.Library)
 			continue
 		}
-		// If the devices requires a numeric ID, for filtering purposes, we use the unfiltered ID number
-		if _, err := strconv.Atoi(info.ID); err == nil {
-			ids = append(ids, fmt.Sprintf("%d", info.filterID))
-		} else {
-			ids = append(ids, info.ID)
-		}
+		ids = append(ids, info.ID)
 	}
-	if len(ids) == 0 {
-		return ""
-	}
-
 	// There are 3 potential env vars to use to select GPUs.
 	// ROCR_VISIBLE_DEVICES supports UUID or numeric but does not work on Windows
 	// HIP_VISIBLE_DEVICES supports numeric IDs only
 	// GPU_DEVICE_ORDINAL supports numeric IDs only
-	return "HIP_VISIBLE_DEVICES=" + strings.Join(ids, ",")
+	return "HIP_VISIBLE_DEVICES", strings.Join(ids, ",")
 }

discover/cuda_common.go

@@ -16,6 +16,19 @@ import (
 // Included to drive logic for reducing Ollama-allocated overhead on L4T/Jetson devices.
 var CudaTegra string = os.Getenv("JETSON_JETPACK")
 
+func cudaGetVisibleDevicesEnv(gpuInfo []GpuInfo) (string, string) {
+	ids := []string{}
+	for _, info := range gpuInfo {
+		if info.Library != "cuda" {
+			// TODO shouldn't happen if things are wired correctly...
+			slog.Debug("cudaGetVisibleDevicesEnv skipping over non-cuda device", "library", info.Library)
+			continue
+		}
+		ids = append(ids, info.ID)
+	}
+	return "CUDA_VISIBLE_DEVICES", strings.Join(ids, ",")
+}
+
 func cudaVariant(gpuInfo CudaGPUInfo) string {
 	if runtime.GOARCH == "arm64" && runtime.GOOS == "linux" {
 		if CudaTegra != "" {

discover/gpu.go

@@ -371,15 +371,6 @@ func GetGPUInfo() GpuInfoList {
 		}
 
 		rocmGPUs, err = AMDGetGPUInfo()
-
-		// The ID field is used in context of the filtered set of GPUS
-		// so we have to replace any of these numeric IDs with their
-		// placement in this set of GPUs
-		for i := range rocmGPUs {
-			if _, err := strconv.Atoi(rocmGPUs[i].ID); err == nil {
-				rocmGPUs[i].ID = strconv.Itoa(i)
-			}
-		}
 		if err != nil {
 			bootstrapErrors = append(bootstrapErrors, err)
 		}
@@ -689,16 +680,23 @@ func getVerboseState() C.uint16_t {
 
 // Given the list of GPUs this instantiation is targeted for,
 // figure out the visible devices environment variable
-func (l GpuInfoList) GetVisibleDevicesEnv() []string {
+//
+// If different libraries are detected, the first one is what we use
+func (l GpuInfoList) GetVisibleDevicesEnv() (string, string) {
 	if len(l) == 0 {
-		return nil
+		return "", ""
 	}
-	vd := []string{}
-	// Only filter the AMD GPUs at this level, let all NVIDIA devices through
-	if tmp := rocmGetVisibleDevicesEnv(l); tmp != "" {
-		vd = append(vd, tmp)
+	switch l[0].Library {
+	case "cuda":
+		return cudaGetVisibleDevicesEnv(l)
+	case "rocm":
+		return rocmGetVisibleDevicesEnv(l)
+	case "oneapi":
+		return oneapiGetVisibleDevicesEnv(l)
+	default:
+		slog.Debug("no filter required for library " + l[0].Library)
+		return "", ""
 	}
-	return vd
 }
 
 func GetSystemInfo() SystemInfo {

discover/gpu_darwin.go

@@ -62,9 +62,9 @@ func GetCPUMem() (memInfo, error) {
 	}, nil
 }
 
-func (l GpuInfoList) GetVisibleDevicesEnv() []string {
+func (l GpuInfoList) GetVisibleDevicesEnv() (string, string) {
 	// No-op on darwin
-	return nil
+	return "", ""
 }
 
 func GetSystemInfo() SystemInfo {

discover/gpu_oneapi.go

@@ -0,0 +1,21 @@
+//go:build linux || windows
+
+package discover
+
+import (
+	"log/slog"
+	"strings"
+)
+
+func oneapiGetVisibleDevicesEnv(gpuInfo []GpuInfo) (string, string) {
+	ids := []string{}
+	for _, info := range gpuInfo {
+		if info.Library != "oneapi" {
+			// TODO shouldn't happen if things are wired correctly...
+			slog.Debug("oneapiGetVisibleDevicesEnv skipping over non-sycl device", "library", info.Library)
+			continue
+		}
+		ids = append(ids, info.ID)
+	}
+	return "ONEAPI_DEVICE_SELECTOR", "level_zero:" + strings.Join(ids, ",")
+}

discover/types.go

@@ -27,8 +27,8 @@ type GpuInfo struct { // TODO better name maybe "InferenceProcessor"?
 	// Any extra PATH/LD_LIBRARY_PATH dependencies required for the Library to operate properly
 	DependencyPath []string `json:"lib_path,omitempty"`
 
-	// Extra environment variables specific to the GPU as list of [key=value]
-	EnvWorkarounds []string `json:"envs,omitempty"`
+	// Extra environment variables specific to the GPU as list of [key,value]
+	EnvWorkarounds [][2]string `json:"envs,omitempty"`
 
 	// Set to true if we can NOT reliably discover FreeMemory.  A value of true indicates
 	// the FreeMemory is best effort, and may over or under report actual memory usage
@@ -36,10 +36,9 @@ type GpuInfo struct { // TODO better name maybe "InferenceProcessor"?
 	UnreliableFreeMemory bool
 
 	// GPU information
-	ID       string `json:"gpu_id"` // string to use for selection of this specific GPU
-	filterID int    //nolint:unused,nolintlint // AMD Workaround: The numeric ID of the device used to filter out other devices
-	Name     string `json:"name"`    // user friendly name if available
-	Compute  string `json:"compute"` // Compute Capability or gfx
+	ID      string `json:"gpu_id"`  // string to use for selection of this specific GPU
+	Name    string `json:"name"`    // user friendly name if available
+	Compute string `json:"compute"` // Compute Capability or gfx
 
 	// Driver Information - TODO no need to put this on each GPU
 	DriverMajor int `json:"driver_major,omitempty"`

docs/docs.json

@@ -0,0 +1,75 @@
+{
+  "$schema": "https://mintlify.com/docs.json",
+  "theme": "mint",
+  "background": {
+    "color": {
+      "light": "#ffffff",
+      "dark": "#000000"
+    }
+  },
+  "appearance": {
+    "default": "light"
+  },
+  "styling": {
+    "codeblocks": "system"
+  },
+  "contextual": {
+    "options": ["copy", "chatgpt", "claude", "view"]
+  },
+  "fonts": {
+    "heading": {
+      "family": "Inter"
+    },
+    "body": {
+      "family": "Inter"
+    }
+  },
+  "name": "Ollama",
+  "colors": {
+    "primary": "#000",
+    "light": "#b5b5b5",
+    "dark": "#fff"
+  },
+  "favicon": "/ollama.png",
+  "logo": {
+    "light": "/ollama.png",
+    "dark": "/favicon.svg"
+  },
+  "navigation": {
+    "tabs": [
+      {
+        "tab": "Documentation",
+        "groups": [
+          {
+            "group": "Home",
+            "pages": ["index", "quickstart", "faq", "troubleshooting"]
+          },
+          {
+            "group": "Platforms",
+            "pages": ["linux", "windows", "docker"]
+          },
+          {
+            "group": "Features",
+            "pages": [
+              "modelfile",
+              "apis",
+              "openai",
+              "import",
+              "gpu",
+              "benchmark"
+            ]
+          }
+        ]
+      },
+      {
+        "tab": "Development",
+        "groups": [
+          {
+            "group": " ",
+            "pages": ["development", "examples", "template"]
+          }
+        ]
+      }
+    ]
+  }
+}

fs/ggml/ggml.go

@@ -7,11 +7,9 @@ import (
 	"fmt"
 	"io"
 	"log/slog"
-	"math"
 	"slices"
 	"strings"
 
-	"github.com/ollama/ollama/format"
 	"github.com/ollama/ollama/fs/util/bufioutil"
 )
 
@@ -277,7 +275,7 @@ type Tensor struct {
 
 func (t Tensor) block() (n int) {
 	if _, err := fmt.Sscanf(t.Name, "blk.%d.", &n); err != nil {
-		return math.MaxInt
+		return -1
 	}
 
 	return
@@ -290,24 +288,24 @@ func (t Tensor) blockSize() uint64 {
 func (t TensorType) BlockSize() uint64 {
 	switch t {
 	case
-		TensorTypeF32,
-		TensorTypeF16,
-		TensorTypeI8,
-		TensorTypeI16,
-		TensorTypeI32,
-		TensorTypeI64,
-		TensorTypeF64,
-		TensorTypeBF16:
+		0,  // F32
+		1,  // F16
+		24, // I8
+		25, // I16
+		26, // I32
+		27, // I64
+		28, // F64
+		30: // BF16
 		return 1
 	case
-		TensorTypeQ4_0,
-		TensorTypeQ4_1,
-		TensorTypeQ5_0,
-		TensorTypeQ5_1,
-		TensorTypeQ8_0,
-		TensorTypeQ8_1,
-		tensorTypeIQ4_NL,
-		4, TensorTypeMXFP4:
+		2,  // Q4_0
+		3,  // Q4_1
+		4,  // MXFP4
+		6,  // Q5_0
+		7,  // Q5_1
+		8,  // Q8_0
+		9,  // Q8_1
+		20: // IQ4_NL
 		return 32
 	default:
 		return 256
@@ -330,6 +328,8 @@ func (t TensorType) TypeSize() uint64 {
 		return 2 + blockSize/2
 	case TensorTypeQ4_1:
 		return 2 + 2 + blockSize/2
+	case TensorTypeMXFP4, 39:
+		return 1 + blockSize/2
 	case TensorTypeQ5_0:
 		return 2 + 4 + blockSize/2
 	case TensorTypeQ5_1:
@@ -380,8 +380,6 @@ func (t TensorType) TypeSize() uint64 {
 		return blockSize/8 + blockSize/16 + blockSize/32
 	case TensorTypeBF16:
 		return 2
-	case 4, TensorTypeMXFP4:
-		return 1 + blockSize/2
 	default:
 		return 0
 	}
@@ -481,7 +479,7 @@ func Decode(rs io.ReadSeeker, maxArraySize int) (*GGML, error) {
 	}, nil
 }
 
-func (f GGML) GraphSize(context, batch uint64, numParallel int, kvCacheType string, useFlashAttention bool) (kv []uint64, partialOffload, fullOffload uint64) {
+func (f GGML) GraphSize(context, batch uint64, numParallel int, kvCacheType string) (kv []uint64, partialOffload, fullOffload uint64) {
 	context *= uint64(numParallel)
 
 	embedding := f.KV().EmbeddingLength()
@@ -679,12 +677,7 @@ func (f GGML) GraphSize(context, batch uint64, numParallel int, kvCacheType stri
 				kv[i] *= context
 			}
 		}
-
 		partialOffload = 2 * f.KV().HeadCountMax() / cmp.Or(f.KV().HeadCountKVMin(), 1) * kvTotal / 6
-		if useFlashAttention {
-			// rough estimate of graph size with flash attention on
-			partialOffload = (4*uint64(numParallel) + context>>10 + 110) * format.MebiByte
-		}
 	}
 
 	return
@@ -780,13 +773,6 @@ func (f GGML) SupportsFlashAttention() bool {
 	return headCountK != 0 && headCountV != 0 && headCountK == headCountV
 }
 
-// FlashAttention checks if the model should enable flash attention
-func (f GGML) FlashAttention() bool {
-	return slices.Contains([]string{
-		"gptoss", "gpt-oss",
-	}, f.KV().String("general.architecture"))
-}
-
 // kvCacheBytesPerElement returns the number of bytes per element for a given KV cache type
 func kvCacheBytesPerElement(cacheType string) float64 {
 	switch cacheType {

fs/ggml/gguf.go

@@ -533,15 +533,12 @@ func WriteGGUF(f *os.File, kv KV, ts []*Tensor) error {
 		}
 	}
 
-	slices.SortStableFunc(
-		ts,
-		func(a, b *Tensor) int {
-			return cmp.Or(
-				cmp.Compare(a.block(), b.block()),
-				cmp.Compare(a.Name, b.Name),
-			)
-		},
-	)
+	slices.SortStableFunc(ts, func(a, b *Tensor) int {
+		if i, j := a.block(), b.block(); i > 0 && j > 0 {
+			return cmp.Compare(i, j)
+		}
+		return cmp.Compare(a.Name, b.Name)
+	})
 
 	var s uint64
 	for i := range ts {

fs/ggml/gguf_test.go

@@ -11,24 +11,24 @@ import (
 )
 
 func TestWriteGGUF(t *testing.T) {
-	b := bytes.NewBuffer(make([]byte, 2*3))
+	r := rand.New(rand.NewPCG(0, 0))
 	for range 8 {
 		t.Run("shuffle", func(t *testing.T) {
 			t.Parallel()
 
 			ts := []*Tensor{
-				{Name: "token_embd.weight", Shape: []uint64{2, 3}, WriterTo: b},
-				{Name: "blk.0.ffn_norm.weight", Shape: []uint64{2, 3}, WriterTo: b},
-				{Name: "blk.0.attn_norm.weight", Shape: []uint64{2, 3}, WriterTo: b},
-				{Name: "blk.1.ffn_up.weight", Shape: []uint64{2, 3}, WriterTo: b},
-				{Name: "blk.2.ffn_norm.weight", Shape: []uint64{2, 3}, WriterTo: b},
-				{Name: "blk.1.ffn_down.weight", Shape: []uint64{2, 3}, WriterTo: b},
-				{Name: "blk.0.attn_k.weight", Shape: []uint64{2, 3}, WriterTo: b},
-				{Name: "output_norm.weight", Shape: []uint64{3, 2}, WriterTo: b},
-				{Name: "output.weight", Shape: []uint64{3, 2}, WriterTo: b},
+				{Name: "token_embd.weight", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(make([]byte, 2*3))},
+				{Name: "blk.0.attn_norm.weight", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(make([]byte, 2*3))},
+				{Name: "blk.1.attn_norm.weight", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(make([]byte, 2*3))},
+				{Name: "blk.2.attn_norm.weight", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(make([]byte, 2*3))},
+				{Name: "blk.3.attn_norm.weight", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(make([]byte, 2*3))},
+				{Name: "blk.4.attn_norm.weight", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(make([]byte, 2*3))},
+				{Name: "blk.5.attn_norm.weight", Shape: []uint64{2, 3}, WriterTo: bytes.NewBuffer(make([]byte, 2*3))},
+				{Name: "output_norm.weight", Shape: []uint64{3, 2}, WriterTo: bytes.NewBuffer(make([]byte, 3*2))},
+				{Name: "output.weight", Shape: []uint64{3, 2}, WriterTo: bytes.NewBuffer(make([]byte, 3*2))},
 			}
 
-			rand.Shuffle(len(ts), func(i, j int) {
+			r.Shuffle(len(ts), func(i, j int) {
 				ts[i], ts[j] = ts[j], ts[i]
 			})
 
@@ -63,14 +63,14 @@ func TestWriteGGUF(t *testing.T) {
 			}
 
 			if diff := cmp.Diff(Tensors{
-				Offset: 592,
+				Offset: 608,
 				items: []*Tensor{
-					{Name: "blk.0.attn_k.weight", Offset: 0, Shape: []uint64{2, 3}},
-					{Name: "blk.0.attn_norm.weight", Offset: 32, Shape: []uint64{2, 3}},
-					{Name: "blk.0.ffn_norm.weight", Offset: 64, Shape: []uint64{2, 3}},
-					{Name: "blk.1.ffn_down.weight", Offset: 96, Shape: []uint64{2, 3}},
-					{Name: "blk.1.ffn_up.weight", Offset: 128, Shape: []uint64{2, 3}},
-					{Name: "blk.2.ffn_norm.weight", Offset: 160, Shape: []uint64{2, 3}},
+					{Name: "blk.0.attn_norm.weight", Offset: 0, Shape: []uint64{2, 3}},
+					{Name: "blk.1.attn_norm.weight", Offset: 32, Shape: []uint64{2, 3}},
+					{Name: "blk.2.attn_norm.weight", Offset: 64, Shape: []uint64{2, 3}},
+					{Name: "blk.3.attn_norm.weight", Offset: 96, Shape: []uint64{2, 3}},
+					{Name: "blk.4.attn_norm.weight", Offset: 128, Shape: []uint64{2, 3}},
+					{Name: "blk.5.attn_norm.weight", Offset: 160, Shape: []uint64{2, 3}},
 					{Name: "output.weight", Offset: 192, Shape: []uint64{3, 2}},
 					{Name: "output_norm.weight", Offset: 224, Shape: []uint64{3, 2}},
 					{Name: "token_embd.weight", Offset: 256, Shape: []uint64{2, 3}},

fs/ggml/type.go

@@ -146,6 +146,8 @@ func (ftype FileType) ToTensorType() TensorType {
 		return TensorTypeQ4_0
 	case fileTypeQ4_1:
 		return TensorTypeQ4_1
+	case fileTypeMXFP4:
+		return TensorTypeMXFP4 // Formerly unused tensorTypeQ4_2
 	case FileTypeQ8_0:
 		return TensorTypeQ8_0
 	case fileTypeQ5_0:
@@ -174,8 +176,6 @@ func (ftype FileType) ToTensorType() TensorType {
 		return TensorTypeQ2_K
 	case FileTypeBF16:
 		return TensorTypeBF16
-	case fileTypeMXFP4:
-		return TensorTypeMXFP4
 	default:
 		slog.Warn("unsupported file type", "type", ftype)
 		return 0 // F32
@@ -191,8 +191,8 @@ const (
 	TensorTypeF16
 	TensorTypeQ4_0
 	TensorTypeQ4_1
-	tensorTypeQ4_2
-	tensorTypeQ4_3 // unused by GGML
+	TensorTypeMXFP4 // Formerly unused tensorTypeQ4_2
+	tensorTypeQ4_3  // unused by GGML
 	TensorTypeQ5_0
 	TensorTypeQ5_1
 	TensorTypeQ8_0
@@ -226,7 +226,6 @@ const (
 	tensorTypeIQ4_NL_4_4 // unused by GGML
 	tensorTypeIQ4_NL_4_8 // unused by GGML
 	tensorTypeIQ4_NL_8_8 // unused by GGML
-	TensorTypeMXFP4
 )
 
 // ParseFileType parses the provided GGUF file type
@@ -319,7 +318,7 @@ func (t TensorType) String() string {
 		return "F64"
 	case TensorTypeBF16:
 		return "BF16"
-	case 4, TensorTypeMXFP4:
+	case TensorTypeMXFP4:
 		return "MXFP4"
 	default:
 		return "unknown"

integration/README.md

@@ -2,13 +2,10 @@
 
 This directory contains integration tests to exercise Ollama end-to-end to verify behavior
 
-By default, these tests are disabled so `go test ./...` will exercise only unit tests.  To run integration tests you must pass the integration tag.  `go test -tags=integration ./...` Some tests require additional tags to enable to allow scoped testing to keep the duration reasonable.  For example, testing a broad set of models requires `-tags=integration,models` and a longer timeout (~60m or more depending on the speed of your GPU.). To view the current set of tag combinations use `find integration -type f | xargs grep "go:build"`
+By default, these tests are disabled so `go test ./...` will exercise only unit tests.  To run integration tests you must pass the integration tag.  `go test -tags=integration ./...`
 
 
 The integration tests have 2 modes of operating.
 
 1. By default, they will start the server on a random port, run the tests, and then shutdown the server.
-2. If `OLLAMA_TEST_EXISTING` is set to a non-empty string, the tests will run against an existing running server, which can be remote based on your `OLLAMA_HOST` environment variable
-
-> [!IMPORTANT]
-> Before running the tests locally without the "test existing" setting, compile ollama from the top of the source tree  `go build .` in addition to GPU support with cmake if applicable on your platform.  The integration tests expect to find an ollama binary at the top of the tree.
+2. If `OLLAMA_TEST_EXISTING` is set to a non-empty string, the tests will run against an existing running server, which can be remote

integration/api_test.go

@@ -390,7 +390,7 @@ func TestAPIEmbeddings(t *testing.T) {
 	client, _, cleanup := InitServerConnection(ctx, t)
 	defer cleanup()
 	req := api.EmbeddingRequest{
-		Model:  libraryEmbedModels[0],
+		Model:  "orca-mini",
 		Prompt: "why is the sky blue?",
 		Options: map[string]interface{}{
 			"temperature": 0,

integration/basic_test.go

@@ -11,6 +11,7 @@ import (
 	"time"
 
 	"github.com/ollama/ollama/api"
+	"github.com/stretchr/testify/require"
 )
 
 func TestBlueSky(t *testing.T) {
@@ -36,8 +37,8 @@ func TestUnicode(t *testing.T) {
 	// Set up the test data
 	req := api.GenerateRequest{
 		// DeepSeek has a Unicode tokenizer regex, making it a unicode torture test
-		Model:  "deepseek-coder-v2:16b-lite-instruct-q2_K", // TODO is there an ollama-engine model we can switch to and keep the coverage?
-		Prompt: "天空为什么是蓝色的?",                               // Why is the sky blue?
+		Model:  "deepseek-coder-v2:16b-lite-instruct-q2_K",
+		Prompt: "天空为什么是蓝色的?",
 		Stream: &stream,
 		Options: map[string]any{
 			"temperature": 0,
@@ -49,20 +50,8 @@ func TestUnicode(t *testing.T) {
 	}
 	client, _, cleanup := InitServerConnection(ctx, t)
 	defer cleanup()
-	if err := PullIfMissing(ctx, client, req.Model); err != nil {
-		t.Fatal(err)
-	}
-	slog.Info("loading", "model", req.Model)
-	err := client.Generate(ctx, &api.GenerateRequest{Model: req.Model}, func(response api.GenerateResponse) error { return nil })
-	if err != nil {
-		t.Fatalf("failed to load model %s: %s", req.Model, err)
-	}
-	skipIfNotGPULoaded(ctx, t, client, req.Model, 100)
-
-	DoGenerate(ctx, t, client, req, []string{
-		"散射", // scattering
-		"频率", // frequency
-	}, 120*time.Second, 120*time.Second)
+	require.NoError(t, PullIfMissing(ctx, client, req.Model))
+	DoGenerate(ctx, t, client, req, []string{"散射", "频率"}, 120*time.Second, 120*time.Second)
 }
 
 func TestExtendedUnicodeOutput(t *testing.T) {
@@ -80,9 +69,7 @@ func TestExtendedUnicodeOutput(t *testing.T) {
 	}
 	client, _, cleanup := InitServerConnection(ctx, t)
 	defer cleanup()
-	if err := PullIfMissing(ctx, client, req.Model); err != nil {
-		t.Fatal(err)
-	}
+	require.NoError(t, PullIfMissing(ctx, client, req.Model))
 	DoGenerate(ctx, t, client, req, []string{"😀", "😊", "😁", "😂", "😄", "😃"}, 120*time.Second, 120*time.Second)
 }
 
@@ -97,9 +84,7 @@ func TestUnicodeModelDir(t *testing.T) {
 	}
 
 	modelDir, err := os.MkdirTemp("", "ollama_埃")
-	if err != nil {
-		t.Fatal(err)
-	}
+	require.NoError(t, err)
 	defer os.RemoveAll(modelDir)
 	slog.Info("unicode", "OLLAMA_MODELS", modelDir)
 

integration/concurrency_test.go

@@ -14,6 +14,8 @@ import (
 	"testing"
 	"time"
 
+	"github.com/stretchr/testify/require"
+
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/format"
@@ -77,21 +79,21 @@ func TestMultiModelStress(t *testing.T) {
 		t.Fatal(err)
 	}
 
-	// All models compatible with ollama-engine
 	smallModels := []string{
 		"llama3.2:1b",
 		"qwen3:0.6b",
-		"gemma2:2b",
-		"deepseek-r1:1.5b", // qwen2 arch
-		"gemma3:270m",
+		"gemma:2b",
+		"deepseek-r1:1.5b",
+		"starcoder2:3b",
 	}
 	mediumModels := []string{
-		"llama3.2:3b",    // ~3.4G
-		"qwen3:8b",       // ~6.6G
-		"gpt-oss:20b",    // ~15G
-		"deepseek-r1:7b", // ~5.6G
-		"gemma3:4b",      // ~5.8G
-		"gemma2:9b",      // ~8.1G
+		"qwen3:8b",
+		"llama2",
+		"deepseek-r1:7b",
+		"mistral",
+		"dolphin-mistral",
+		"gemma:7b",
+		"codellama:7b",
 	}
 
 	var chosenModels []string
@@ -112,9 +114,7 @@ func TestMultiModelStress(t *testing.T) {
 
 	// Make sure all the models are pulled before we get started
 	for _, model := range chosenModels {
-		if err := PullIfMissing(ctx, client, model); err != nil {
-			t.Fatal(err)
-		}
+		require.NoError(t, PullIfMissing(ctx, client, model))
 	}
 
 	// Determine how many models we can load in parallel before we exceed VRAM

integration/context_test.go

@@ -22,7 +22,7 @@ func TestLongInputContext(t *testing.T) {
 	defer cancel()
 	// Set up the test data
 	req := api.GenerateRequest{
-		Model:  smol,
+		Model:  "llama2",
 		Prompt: "Oh, don’t speak to me of Austria. Perhaps I don’t understand things, but Austria never has wished, and does not wish, for war. She is betraying us! Russia alone must save Europe. Our gracious sovereign recognizes his high vocation and will be true to it. That is the one thing I have faith in! Our good and wonderful sovereign has to perform the noblest role on earth, and he is so virtuous and noble that God will not forsake him. He will fulfill his vocation and crush the hydra of revolution, which has become more terrible than ever in the person of this murderer and villain! We alone must avenge the blood of the just one.... Whom, I ask you, can we rely on?... England with her commercial spirit will not and cannot understand the Emperor Alexander’s loftiness of soul. She has refused to evacuate Malta. She wanted to find, and still seeks, some secret motive in our actions. What answer did Novosíltsev get? None. The English have not understood and cannot understand the self-abnegation of our Emperor who wants nothing for himself, but only desires the good of mankind. And what have they promised? Nothing! And what little they have promised they will not perform! Prussia has always declared that Buonaparte is invincible, and that all Europe is powerless before him.... And I don’t believe a word that Hardenburg says, or Haugwitz either. This famous Prussian neutrality is just a trap. I have faith only in God and the lofty destiny of our adored monarch. He will save Europe! What country is this referring to?",
 		Stream: &stream,
 		Options: map[string]any{
@@ -36,7 +36,7 @@ func TestLongInputContext(t *testing.T) {
 	if err := PullIfMissing(ctx, client, req.Model); err != nil {
 		t.Fatalf("PullIfMissing failed: %v", err)
 	}
-	DoGenerate(ctx, t, client, req, []string{"russia", "germany", "france", "england", "austria", "prussia", "individuals", "coalition", "conflict"}, 120*time.Second, 10*time.Second)
+	DoGenerate(ctx, t, client, req, []string{"russia", "germany", "france", "england", "austria", "prussia"}, 120*time.Second, 10*time.Second)
 }
 
 func TestContextExhaustion(t *testing.T) {
@@ -49,7 +49,7 @@ func TestContextExhaustion(t *testing.T) {
 	defer cancel()
 	// Set up the test data
 	req := api.GenerateRequest{
-		Model:  smol,
+		Model:  "llama2",
 		Prompt: "Write me a story with a ton of emojis?",
 		Stream: &stream,
 		Options: map[string]any{
@@ -63,10 +63,10 @@ func TestContextExhaustion(t *testing.T) {
 	if err := PullIfMissing(ctx, client, req.Model); err != nil {
 		t.Fatalf("PullIfMissing failed: %v", err)
 	}
-	DoGenerate(ctx, t, client, req, []string{"once", "upon", "lived", "sunny", "cloudy", "clear", "water"}, 120*time.Second, 10*time.Second)
+	DoGenerate(ctx, t, client, req, []string{"once", "upon", "lived"}, 120*time.Second, 10*time.Second)
 }
 
-// Send multiple generate requests with prior context and ensure the response is coherant and expected
+// Send multiple requests with prior context and ensure the response is coherant and expected
 func TestGenerateWithHistory(t *testing.T) {
 	modelOverride := ollamaEngineChatModels[0] // Most recent ollama engine model
 	req, resp := GenerateRequests()
@@ -111,56 +111,5 @@ func TestGenerateWithHistory(t *testing.T) {
 		}(i)
 	}
 	wg.Wait()
-}
-
-// Send multiple chat requests with prior context and ensure the response is coherant and expected
-func TestChatWithHistory(t *testing.T) {
-	modelOverride := ollamaEngineChatModels[0] // Most recent ollama engine model
-	req, resp := ChatRequests()
-	numParallel := 2
-	iterLimit := 2
-
-	softTimeout, hardTimeout := getTimeouts(t)
-	ctx, cancel := context.WithTimeout(context.Background(), hardTimeout)
-	defer cancel()
-	client, _, cleanup := InitServerConnection(ctx, t)
-	defer cleanup()
-
-	// Get the server running (if applicable) warm the model up with a single initial empty request
-	slog.Info("loading", "model", modelOverride)
-	err := client.Generate(ctx,
-		&api.GenerateRequest{Model: modelOverride, KeepAlive: &api.Duration{Duration: 10 * time.Second}},
-		func(response api.GenerateResponse) error { return nil },
-	)
-	if err != nil {
-		t.Fatalf("failed to load model %s: %s", modelOverride, err)
-	}
-
-	var wg sync.WaitGroup
-	wg.Add(numParallel)
-	for i := range numParallel {
-		go func(i int) {
-			defer wg.Done()
-			k := i % len(req)
-			req[k].Model = modelOverride
-			for j := 0; j < iterLimit; j++ {
-				if time.Now().Sub(started) > softTimeout {
-					slog.Info("exceeded soft timeout, winding down test")
-					return
-				}
-				slog.Info("Starting", "thread", i, "iter", j)
-				// On slower GPUs it can take a while to process the concurrent requests
-				// so we allow a much longer initial timeout
-				assistant := DoChat(ctx, t, client, req[k], resp[k], 120*time.Second, 20*time.Second)
-				if assistant == nil {
-					t.Fatalf("didn't get an assistant response for context")
-				}
-				req[k].Messages = append(req[k].Messages,
-					*assistant,
-					api.Message{Role: "user", Content: "tell me more!"},
-				)
-			}
-		}(i)
-	}
-	wg.Wait()
+
 }

integration/llm_image_test.go

@@ -9,6 +9,7 @@ import (
 	"time"
 
 	"github.com/ollama/ollama/api"
+	"github.com/stretchr/testify/require"
 )
 
 func TestVisionModels(t *testing.T) {
@@ -31,9 +32,7 @@ func TestVisionModels(t *testing.T) {
 	for _, v := range testCases {
 		t.Run(v.model, func(t *testing.T) {
 			image, err := base64.StdEncoding.DecodeString(imageEncoding)
-			if err != nil {
-				t.Fatal(err)
-			}
+			require.NoError(t, err)
 			req := api.GenerateRequest{
 				Model:  v.model,
 				Prompt: "what does the text in this image say?",
@@ -53,9 +52,7 @@ func TestVisionModels(t *testing.T) {
 			// Note: sometimes it returns "the ollamas" sometimes "the ollams"
 			resp := "the ollam"
 			defer cleanup()
-			if err := PullIfMissing(ctx, client, req.Model); err != nil {
-				t.Fatal(err)
-			}
+			require.NoError(t, PullIfMissing(ctx, client, req.Model))
 			// llava models on CPU can be quite slow to start
 			DoGenerate(ctx, t, client, req, []string{resp}, 240*time.Second, 30*time.Second)
 		})
@@ -65,9 +62,7 @@ func TestVisionModels(t *testing.T) {
 func TestIntegrationSplitBatch(t *testing.T) {
 	skipUnderMinVRAM(t, 6)
 	image, err := base64.StdEncoding.DecodeString(imageEncoding)
-	if err != nil {
-		t.Fatal(err)
-	}
+	require.NoError(t, err)
 	req := api.GenerateRequest{
 		Model: "gemma3:4b",
 		// Fill up a chunk of the batch so the image will partially spill over into the next one
@@ -89,9 +84,7 @@ func TestIntegrationSplitBatch(t *testing.T) {
 	defer cancel()
 	client, _, cleanup := InitServerConnection(ctx, t)
 	defer cleanup()
-	if err := PullIfMissing(ctx, client, req.Model); err != nil {
-		t.Fatal(err)
-	}
+	require.NoError(t, PullIfMissing(ctx, client, req.Model))
 	// llava models on CPU can be quite slow to start,
 	DoGenerate(ctx, t, client, req, []string{resp}, 120*time.Second, 30*time.Second)
 }

integration/llm_test.go

@@ -0,0 +1,47 @@
+//go:build integration
+
+package integration
+
+import (
+	"context"
+	"testing"
+	"time"
+
+	"github.com/ollama/ollama/api"
+)
+
+// TODO - this would ideally be in the llm package, but that would require some refactoring of interfaces in the server
+//        package to avoid circular dependencies
+
+var (
+	stream = false
+	req    = [2]api.GenerateRequest{
+		{
+			Model:  smol,
+			Prompt: "why is the ocean blue?",
+			Stream: &stream,
+			Options: map[string]any{
+				"seed":        42,
+				"temperature": 0.0,
+			},
+		}, {
+			Model:  smol,
+			Prompt: "what is the origin of the us thanksgiving holiday?",
+			Stream: &stream,
+			Options: map[string]any{
+				"seed":        42,
+				"temperature": 0.0,
+			},
+		},
+	}
+	resp = [2][]string{
+		{"sunlight", "scattering", "interact"},
+		{"england", "english", "massachusetts", "pilgrims"},
+	}
+)
+
+func TestIntegrationSimple(t *testing.T) {
+	ctx, cancel := context.WithTimeout(context.Background(), time.Second*120)
+	defer cancel()
+	GenerateTestHelper(ctx, t, req[0], resp[0])
+}

integration/max_queue_test.go

@@ -13,12 +13,12 @@ import (
 	"testing"
 	"time"
 
+	"github.com/stretchr/testify/require"
+
 	"github.com/ollama/ollama/api"
 )
 
 func TestMaxQueue(t *testing.T) {
-	t.Skip("this test needs to be re-evaluated to use a proper embedding model")
-
 	if os.Getenv("OLLAMA_TEST_EXISTING") != "" {
 		t.Skip("Max Queue test requires spawning a local server so we can adjust the queue size")
 		return
@@ -45,9 +45,7 @@ func TestMaxQueue(t *testing.T) {
 	client, _, cleanup := InitServerConnection(ctx, t)
 	defer cleanup()
 
-	if err := PullIfMissing(ctx, client, req.Model); err != nil {
-		t.Fatal(err)
-	}
+	require.NoError(t, PullIfMissing(ctx, client, req.Model))
 
 	// Context for the worker threads so we can shut them down
 	// embedCtx, embedCancel := context.WithCancel(ctx)
@@ -91,9 +89,7 @@ func TestMaxQueue(t *testing.T) {
 			switch {
 			case genErr == nil:
 				successCount++
-				if len(resp.Embedding) < 5 { // somewhat arbitrary, but sufficient to be reasonable
-					t.Fatalf("embeddings shorter than expected: %d", len(resp.Embedding))
-				}
+				require.Greater(t, len(resp.Embedding), 5) // somewhat arbitrary, but sufficient to be reasonable
 			case errors.Is(genErr, context.Canceled):
 				canceledCount++
 			case strings.Contains(genErr.Error(), "busy"):
@@ -101,9 +97,7 @@ func TestMaxQueue(t *testing.T) {
 			case strings.Contains(genErr.Error(), "connection reset by peer"):
 				resetByPeerCount++
 			default:
-				if genErr != nil {
-					t.Fatalf("%d request failed", i)
-				}
+				require.NoError(t, genErr, "%d request failed", i)
 			}
 
 			slog.Info("embed finished", "id", i)
@@ -114,13 +108,8 @@ func TestMaxQueue(t *testing.T) {
 	embedwg.Wait()
 
 	slog.Info("embeds completed", "success", successCount, "busy", busyCount, "reset", resetByPeerCount, "canceled", canceledCount)
-	if resetByPeerCount != 0 {
-		t.Fatalf("Connections reset by peer, have you updated your fd and socket limits? %d", resetByPeerCount)
-	}
-	if busyCount == 0 {
-		t.Fatalf("no requests hit busy error but some should have")
-	}
-	if canceledCount > 0 {
-		t.Fatalf("no requests should have been canceled due to timeout %d", canceledCount)
-	}
+	require.Equal(t, resetByPeerCount, 0, "Connections reset by peer, have you updated your fd and socket limits?")
+	require.True(t, busyCount > 0, "no requests hit busy error but some should have")
+	require.True(t, canceledCount == 0, "no requests should have been canceled due to timeout")
+
 }

integration/utils_test.go

@@ -9,7 +9,6 @@ import (
 	"fmt"
 	"io"
 	"log/slog"
-	"math"
 	"math/rand"
 	"net"
 	"net/http"
@@ -26,11 +25,11 @@ import (
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/app/lifecycle"
 	"github.com/ollama/ollama/format"
+	"github.com/stretchr/testify/require"
 )
 
 var (
-	smol   = "llama3.2:1b"
-	stream = false
+	smol = "llama3.2:1b"
 )
 
 var (
@@ -436,9 +435,7 @@ func InitServerConnection(ctx context.Context, t *testing.T) (*api.Client, strin
 		}
 		lifecycle.ServerLogFile = fp.Name()
 		fp.Close()
-		if err := startServer(t, ctx, testEndpoint); err != nil {
-			t.Fatal(err)
-		}
+		require.NoError(t, startServer(t, ctx, testEndpoint))
 	}
 
 	return client, testEndpoint, func() {
@@ -471,9 +468,7 @@ func InitServerConnection(ctx context.Context, t *testing.T) (*api.Client, strin
 func GenerateTestHelper(ctx context.Context, t *testing.T, genReq api.GenerateRequest, anyResp []string) {
 	client, _, cleanup := InitServerConnection(ctx, t)
 	defer cleanup()
-	if err := PullIfMissing(ctx, client, genReq.Model); err != nil {
-		t.Fatal(err)
-	}
+	require.NoError(t, PullIfMissing(ctx, client, genReq.Model))
 	DoGenerate(ctx, t, client, genReq, anyResp, 30*time.Second, 10*time.Second)
 }
 
@@ -514,9 +509,7 @@ func DoGenerate(ctx context.Context, t *testing.T, client *api.Client, genReq ap
 			slog.Warn("model is too large for the target test system", "model", genReq.Model, "error", genErr)
 			return context
 		}
-		if genErr != nil {
-			t.Fatalf("%s failed with %s request prompt %s", genErr, genReq.Model, genReq.Prompt)
-		}
+		require.NoError(t, genErr, "failed with %s request prompt %s ", genReq.Model, genReq.Prompt)
 		// Verify the response contains the expected data
 		response := buf.String()
 		atLeastOne := false
@@ -526,9 +519,7 @@ func DoGenerate(ctx context.Context, t *testing.T, client *api.Client, genReq ap
 				break
 			}
 		}
-		if !atLeastOne {
-			t.Fatalf("%s: none of %v found in %s", genReq.Model, anyResp, response)
-		}
+		require.True(t, atLeastOne, "%s: none of %v found in %s", genReq.Model, anyResp, response)
 		slog.Info("test pass", "model", genReq.Model, "prompt", genReq.Prompt, "contains", anyResp, "response", response)
 	case <-ctx.Done():
 		t.Error("outer test context done while waiting for generate")
@@ -570,97 +561,17 @@ func GenerateRequests() ([]api.GenerateRequest, [][]string) {
 		[][]string{
 			{"sunlight", "scattering", "interact", "color", "surface", "depth", "red", "orange", "yellow", "absorbs", "wavelength"},
 			{"soil", "organic", "earth", "black", "tan", "chemical", "processes", "pigments", "particles", "iron oxide", "rust", "air", "water", "mixture", "mixing"},
-			{"england", "english", "massachusetts", "pilgrims", "colonists", "independence", "british", "feast", "family", "gatherings", "traditions", "turkey", "colonial", "period", "harvest", "agricultural", "european settlers", "american revolution", "civil war", "16th century", "17th century", "native american", "united states", "cultural", "hardship", "autumn", "festival"},
+			{"england", "english", "massachusetts", "pilgrims", "colonists", "independence", "british", "feast", "family", "gatherings", "traditions", "turkey", "colonial", "period", "harvest", "agricultural", "european settlers", "american revolution", "civil war", "16th century", "17th century", "native american", "united states"},
 			{"fourth", "july", "declaration", "independence"},
 			{"nitrogen", "oxygen", "carbon", "dioxide"},
 		}
 }
 
-func DoChat(ctx context.Context, t *testing.T, client *api.Client, req api.ChatRequest, anyResp []string, initialTimeout, streamTimeout time.Duration) *api.Message {
-	stallTimer := time.NewTimer(initialTimeout)
-	var buf bytes.Buffer
-	role := "assistant"
-	fn := func(response api.ChatResponse) error {
-		// fmt.Print(".")
-		role = response.Message.Role
-		buf.Write([]byte(response.Message.Content))
-		if !stallTimer.Reset(streamTimeout) {
-			return errors.New("stall was detected while streaming response, aborting")
-		}
-		return nil
-	}
-
-	stream := true
-	req.Stream = &stream
-	done := make(chan int)
-	var genErr error
-	go func() {
-		genErr = client.Chat(ctx, &req, fn)
-		done <- 0
-	}()
-
-	select {
-	case <-stallTimer.C:
-		if buf.Len() == 0 {
-			t.Errorf("generate never started.  Timed out after :%s", initialTimeout.String())
-		} else {
-			t.Errorf("generate stalled.  Response so far:%s", buf.String())
-		}
-	case <-done:
-		if genErr != nil && strings.Contains(genErr.Error(), "model requires more system memory") {
-			slog.Warn("model is too large for the target test system", "model", req.Model, "error", genErr)
-			return nil
-		}
-		if genErr != nil {
-			t.Fatalf("%s failed with %s request prompt %v", genErr, req.Model, req.Messages)
-		}
-
-		// Verify the response contains the expected data
-		response := buf.String()
-		atLeastOne := false
-		for _, resp := range anyResp {
-			if strings.Contains(strings.ToLower(response), resp) {
-				atLeastOne = true
-				break
-			}
-		}
-		if !atLeastOne {
-			t.Fatalf("%s: none of %v found in \"%s\" -- request was:%v", req.Model, anyResp, response, req.Messages)
-		}
-
-		slog.Info("test pass", "model", req.Model, "messages", req.Messages, "contains", anyResp, "response", response)
-	case <-ctx.Done():
-		t.Error("outer test context done while waiting for generate")
-	}
-	return &api.Message{Role: role, Content: buf.String()}
-}
-
-func ChatRequests() ([]api.ChatRequest, [][]string) {
-	genReqs, results := GenerateRequests()
-	reqs := make([]api.ChatRequest, len(genReqs))
-	// think := api.ThinkValue{Value: "low"}
-	for i := range reqs {
-		reqs[i].Model = genReqs[i].Model
-		reqs[i].Stream = genReqs[i].Stream
-		reqs[i].KeepAlive = genReqs[i].KeepAlive
-		// reqs[i].Think = &think
-		reqs[i].Messages = []api.Message{
-			{
-				Role:    "user",
-				Content: genReqs[i].Prompt,
-			},
-		}
-	}
-	return reqs, results
-}
-
 func skipUnderMinVRAM(t *testing.T, gb uint64) {
 	// TODO use info API in the future
 	if s := os.Getenv("OLLAMA_MAX_VRAM"); s != "" {
 		maxVram, err := strconv.ParseUint(s, 10, 64)
-		if err != nil {
-			t.Fatal(err)
-		}
+		require.NoError(t, err)
 		// Don't hammer on small VRAM cards...
 		if maxVram < gb*format.GibiByte {
 			t.Skip("skipping with small VRAM to avoid timeouts")
@@ -668,39 +579,6 @@ func skipUnderMinVRAM(t *testing.T, gb uint64) {
 	}
 }
 
-// Skip if the target model isn't X% GPU loaded to avoid excessive runtime
-func skipIfNotGPULoaded(ctx context.Context, t *testing.T, client *api.Client, model string, minPercent int) {
-	models, err := client.ListRunning(ctx)
-	if err != nil {
-		t.Fatalf("failed to list running models: %s", err)
-	}
-	loaded := []string{}
-	for _, m := range models.Models {
-		loaded = append(loaded, m.Name)
-		if m.Name != model {
-			continue
-		}
-		gpuPercent := 0
-		switch {
-		case m.SizeVRAM == 0:
-			gpuPercent = 0
-		case m.SizeVRAM == m.Size:
-			gpuPercent = 100
-		case m.SizeVRAM > m.Size || m.Size == 0:
-			t.Logf("unexpected size detected: %d", m.SizeVRAM)
-		default:
-			sizeCPU := m.Size - m.SizeVRAM
-			cpuPercent := math.Round(float64(sizeCPU) / float64(m.Size) * 110)
-			gpuPercent = int(100 - cpuPercent)
-		}
-		if gpuPercent < minPercent {
-			t.Skip(fmt.Sprintf("test requires minimum %d%% GPU load, but model %s only has %d%%", minPercent, model, gpuPercent))
-		}
-		return
-	}
-	t.Skip(fmt.Sprintf("model %s not loaded - actually loaded: %v", model, loaded))
-}
-
 func getTimeouts(t *testing.T) (soft time.Duration, hard time.Duration) {
 	deadline, hasDeadline := t.Deadline()
 	if !hasDeadline {

kvcache/causal.go

@@ -378,7 +378,9 @@ func (c *Causal) buildMask(ctx ml.Context) ml.Tensor {
 	maskTensor := ctx.Input().FromFloatSlice(mask, length, batchSize)
 
 	if c.config.MaskDType != ml.DTypeF32 {
-		maskTensor = maskTensor.Cast(ctx, c.config.MaskDType)
+		out := ctx.Input().Empty(c.config.MaskDType, maskTensor.Shape()...)
+		ctx.Forward(maskTensor.Copy(ctx, out))
+		maskTensor = out
 	}
 
 	return maskTensor

llama/llama.cpp/src/llama-context.cpp

@@ -962,7 +962,8 @@ int llama_context::decode(const llama_batch & batch_inp) {
     const int64_t n_vocab = vocab.n_tokens();
     const int64_t n_embd  = hparams.n_embd;
 
-    const bool output_all = false;
+    // when computing embeddings, all tokens are output
+    const bool output_all = cparams.embeddings;
 
     if (!balloc->init(batch_inp, vocab, memory.get(), n_embd, cparams.kv_unified ? LLAMA_MAX_SEQ : cparams.n_seq_max, output_all)) {
         LLAMA_LOG_ERROR("%s: failed to initialize batch\n", __func__);

llama/patches/0019-Enable-CUDA-Graphs-for-gemma3n.patch

@@ -13,7 +13,7 @@ checks.
  1 file changed, 18 insertions(+)
 
 diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
-index 57eae461..c7f9dc3a 100644
+index 57eae461..9db0c8b5 100644
 --- a/ggml/src/ggml-cuda/ggml-cuda.cu
 +++ b/ggml/src/ggml-cuda/ggml-cuda.cu
 @@ -2671,12 +2671,24 @@ static bool check_node_graph_compatibility_and_refresh_copy_ops(ggml_backend_cud

llama/patches/0023-decode-disable-output_all.patch

@@ -1,23 +0,0 @@
-From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
-From: Michael Yang <git@mxy.ng>
-Date: Mon, 18 Aug 2025 16:58:39 -0700
-Subject: [PATCH] decode: disable output_all
-
----
- src/llama-context.cpp | 3 +--
- 1 file changed, 1 insertion(+), 2 deletions(-)
-
-diff --git a/src/llama-context.cpp b/src/llama-context.cpp
-index 26a5cf9c..6ece5263 100644
---- a/src/llama-context.cpp
-+++ b/src/llama-context.cpp
-@@ -962,8 +962,7 @@ int llama_context::decode(const llama_batch & batch_inp) {
-     const int64_t n_vocab = vocab.n_tokens();
-     const int64_t n_embd  = hparams.n_embd;
- 
--    // when computing embeddings, all tokens are output
--    const bool output_all = cparams.embeddings;
-+    const bool output_all = false;
- 
-     if (!balloc->init(batch_inp, vocab, memory.get(), n_embd, cparams.kv_unified ? LLAMA_MAX_SEQ : cparams.n_seq_max, output_all)) {
-         LLAMA_LOG_ERROR("%s: failed to initialize batch\n", __func__);

llama/patches/0024-ggml-Enable-resetting-backend-devices.patch

@@ -1,130 +0,0 @@
-From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
-From: Jesse Gross <jesse@ollama.com>
-Date: Wed, 27 Aug 2025 14:39:48 -0700
-Subject: [PATCH] ggml: Enable resetting backend devices
-
-Touching a CUDA device causes the allocation of a primary context
-with CUDA data structures (~300 MB of VRAM). If a device is
-unused then it can be reset to free these data structures.
----
- ggml/include/ggml-backend.h      |  1 +
- ggml/src/ggml-backend-impl.h     |  4 ++++
- ggml/src/ggml-backend.cpp        |  8 ++++++++
- ggml/src/ggml-cuda/ggml-cuda.cu  | 17 +++++++++++++++--
- ggml/src/ggml-cuda/vendors/hip.h |  1 +
- 5 files changed, 29 insertions(+), 2 deletions(-)
-
-diff --git a/ggml/include/ggml-backend.h b/ggml/include/ggml-backend.h
-index b602a7c78..fda5ceb24 100644
---- a/ggml/include/ggml-backend.h
-+++ b/ggml/include/ggml-backend.h
-@@ -167,6 +167,7 @@ extern "C" {
-     GGML_API void                          ggml_backend_dev_get_props(ggml_backend_dev_t device, struct ggml_backend_dev_props * props);
-     GGML_API ggml_backend_reg_t            ggml_backend_dev_backend_reg(ggml_backend_dev_t device);
-     GGML_API ggml_backend_t                ggml_backend_dev_init(ggml_backend_dev_t device, const char * params);
-+    GGML_API void                          ggml_backend_dev_reset(ggml_backend_dev_t device);
-     GGML_API ggml_backend_buffer_type_t    ggml_backend_dev_buffer_type(ggml_backend_dev_t device);
-     GGML_API ggml_backend_buffer_type_t    ggml_backend_dev_host_buffer_type(ggml_backend_dev_t device);
-     GGML_API ggml_backend_buffer_t         ggml_backend_dev_buffer_from_host_ptr(ggml_backend_dev_t device, void * ptr, size_t size, size_t max_tensor_size);
-diff --git a/ggml/src/ggml-backend-impl.h b/ggml/src/ggml-backend-impl.h
-index 81749a5a3..6f10c353b 100644
---- a/ggml/src/ggml-backend-impl.h
-+++ b/ggml/src/ggml-backend-impl.h
-@@ -178,6 +178,10 @@ extern "C" {
-         ggml_backend_event_t (*event_new)         (ggml_backend_dev_t dev);
-         void                 (*event_free)        (ggml_backend_dev_t dev, ggml_backend_event_t event);
-         void                 (*event_synchronize) (ggml_backend_dev_t dev, ggml_backend_event_t event);
-+
-+        // (optional) reset device, clearing existing allocations and context
-+        // the caller must ensure that there are no outstanding buffers, as these will become invalid
-+        void (*reset)(ggml_backend_dev_t dev);
-     };
- 
-     struct ggml_backend_device {
-diff --git a/ggml/src/ggml-backend.cpp b/ggml/src/ggml-backend.cpp
-index 05a842ed5..6556943b0 100644
---- a/ggml/src/ggml-backend.cpp
-+++ b/ggml/src/ggml-backend.cpp
-@@ -477,6 +477,14 @@ ggml_backend_t ggml_backend_dev_init(ggml_backend_dev_t device, const char * par
-     return device->iface.init_backend(device, params);
- }
- 
-+void ggml_backend_dev_reset(ggml_backend_dev_t device) {
-+    if (device->iface.reset == NULL) {
-+        return;
-+    }
-+
-+    device->iface.reset(device);
-+}
-+
- ggml_backend_buffer_type_t ggml_backend_dev_buffer_type(ggml_backend_dev_t device) {
-     return device->iface.get_buffer_type(device);
- }
-diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
-index c7f9dc3a5..e43fde523 100644
---- a/ggml/src/ggml-cuda/ggml-cuda.cu
-+++ b/ggml/src/ggml-cuda/ggml-cuda.cu
-@@ -103,6 +103,11 @@ int ggml_cuda_get_device() {
-     return id;
- }
- 
-+void ggml_cuda_reset_device(int device) {
-+    ggml_cuda_set_device(device);
-+    CUDA_CHECK(cudaDeviceReset());
-+}
-+
- static cudaError_t ggml_cuda_device_malloc(void ** ptr, size_t size, int device) {
-     ggml_cuda_set_device(device);
-     cudaError_t err;
-@@ -3243,7 +3248,10 @@ static void ggml_backend_cuda_device_get_props(ggml_backend_dev_t dev, ggml_back
-     props->description = ggml_backend_cuda_device_get_description(dev);
-     props->id          = ggml_backend_cuda_device_get_id(dev);
-     props->type        = ggml_backend_cuda_device_get_type(dev);
--    ggml_backend_cuda_device_get_memory(dev, &props->memory_free, &props->memory_total);
-+
-+    // Memory reporting is disabled to avoid allocation of a CUDA primary context (~300 MB per device).
-+    // If you need the memory data, call ggml_backend_dev_memory() explicitly.
-+    props->memory_total = props->memory_free = 0;
- 
-     bool host_buffer = getenv("GGML_CUDA_NO_PINNED") == nullptr;
- #ifdef GGML_CUDA_NO_PEER_COPY
-@@ -3700,6 +3708,11 @@ static void ggml_backend_cuda_device_event_synchronize(ggml_backend_dev_t dev, g
-     CUDA_CHECK(cudaEventSynchronize((cudaEvent_t)event->context));
- }
- 
-+static void ggml_backend_cuda_device_reset(ggml_backend_dev_t dev) {
-+    ggml_backend_cuda_device_context * ctx = (ggml_backend_cuda_device_context *)dev->context;
-+    ggml_cuda_reset_device(ctx->device);
-+}
-+
- static const ggml_backend_device_i ggml_backend_cuda_device_interface = {
-     /* .get_name                = */ ggml_backend_cuda_device_get_name,
-     /* .get_description         = */ ggml_backend_cuda_device_get_description,
-@@ -3716,6 +3729,7 @@ static const ggml_backend_device_i ggml_backend_cuda_device_interface = {
-     /* .event_new               = */ ggml_backend_cuda_device_event_new,
-     /* .event_free              = */ ggml_backend_cuda_device_event_free,
-     /* .event_synchronize       = */ ggml_backend_cuda_device_event_synchronize,
-+    /* .reset                   = */ ggml_backend_cuda_device_reset,
- };
- 
- // backend reg
-@@ -3835,7 +3849,6 @@ ggml_backend_reg_t ggml_backend_cuda_reg() {
-                 dev_ctx->device = i;
-                 dev_ctx->name = GGML_CUDA_NAME + std::to_string(i);
- 
--                ggml_cuda_set_device(i);
-                 cudaDeviceProp prop;
-                 CUDA_CHECK(cudaGetDeviceProperties(&prop, i));
-                 dev_ctx->description = prop.name;
-diff --git a/ggml/src/ggml-cuda/vendors/hip.h b/ggml/src/ggml-cuda/vendors/hip.h
-index c31f31923..cf22e60d2 100644
---- a/ggml/src/ggml-cuda/vendors/hip.h
-+++ b/ggml/src/ggml-cuda/vendors/hip.h
-@@ -40,6 +40,7 @@
- #define cudaDeviceDisablePeerAccess hipDeviceDisablePeerAccess
- #define cudaDeviceEnablePeerAccess hipDeviceEnablePeerAccess
- #define cudaDeviceProp hipDeviceProp_t
-+#define cudaDeviceReset hipDeviceReset
- #define cudaDeviceSynchronize hipDeviceSynchronize
- #define cudaError_t hipError_t
- #define cudaErrorPeerAccessAlreadyEnabled hipErrorPeerAccessAlreadyEnabled

llama/patches/0025-harden-uncaught-exception-registration.patch

@@ -1,28 +0,0 @@
-From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
-From: Daniel Hiltgen <daniel@ollama.com>
-Date: Fri, 29 Aug 2025 16:53:08 -0700
-Subject: [PATCH] harden uncaught exception registration
-
----
- ggml/src/ggml.cpp | 8 ++++++--
- 1 file changed, 6 insertions(+), 2 deletions(-)
-
-diff --git a/ggml/src/ggml.cpp b/ggml/src/ggml.cpp
-index 0d388d45..f5bcb446 100644
---- a/ggml/src/ggml.cpp
-+++ b/ggml/src/ggml.cpp
-@@ -19,8 +19,12 @@ static bool ggml_uncaught_exception_init = []{
-         return false;
-     }
-     const auto prev{std::get_terminate()};
--    GGML_ASSERT(prev != ggml_uncaught_exception);
--    previous_terminate_handler = prev;
-+    // GGML_ASSERT(prev != ggml_uncaught_exception);
-+    if (prev != ggml_uncaught_exception) {
-+        previous_terminate_handler = prev;
-+    } else {
-+        GGML_LOG_WARN("%s double registration of ggml_uncaught_exception\n", __func__);
-+    }
-     std::set_terminate(ggml_uncaught_exception);
-     return true;
- }();

llm/memory.go

@@ -30,7 +30,7 @@ func pickBestFullFitByLibrary(f *ggml.GGML, modelPath string, projectors []strin
 			// Try to pack into as few GPUs as possible, starting from 1 GPU
 			for numGPUs := 1; numGPUs <= len(sgl); numGPUs++ {
 				gpuSubset := sgl[:numGPUs]
-				ok, estimatedVRAM := predictServerFit(gpuSubset, f, adapters, projectors, opts, numParallel)
+				ok, estimatedVRAM := PredictServerFit(gpuSubset, f, adapters, projectors, opts, numParallel)
 
 				if ok {
 					slog.Info("new model will fit in available VRAM across minimum required GPUs, loading",
@@ -48,7 +48,7 @@ func pickBestFullFitByLibrary(f *ggml.GGML, modelPath string, projectors []strin
 			// - try subsets of GPUs instead of just falling back to 1 or all in a family
 
 			// Now try all the GPUS (OLLAMA_SCHED_SPREAD is set)
-			if ok, estimatedVRAM := predictServerFit(sgl, f, adapters, projectors, opts, numParallel); ok {
+			if ok, estimatedVRAM := PredictServerFit(sgl, f, adapters, projectors, opts, numParallel); ok {
 				slog.Info("new model will fit in available VRAM, loading",
 					"model", modelPath,
 					"library", sgl[0].Library,
@@ -71,7 +71,7 @@ func pickBestPartialFitByLibrary(f *ggml.GGML, projectors []string, adapters []s
 	var bestEstimate uint64
 	var bestFit int
 	for i, gl := range byLibrary {
-		_, estimatedVRAM := predictServerFit(gl, f, adapters, projectors, opts, numParallel)
+		_, estimatedVRAM := PredictServerFit(gl, f, adapters, projectors, opts, numParallel)
 		if estimatedVRAM > bestEstimate {
 			bestEstimate = estimatedVRAM
 			bestFit = i
@@ -81,7 +81,7 @@ func pickBestPartialFitByLibrary(f *ggml.GGML, projectors []string, adapters []s
 }
 
 // This algorithm looks for a complete fit to determine if we need to unload other models
-func predictServerFit(allGpus discover.GpuInfoList, f *ggml.GGML, adapters, projectors []string, opts api.Options, numParallel int) (bool, uint64) {
+func PredictServerFit(allGpus discover.GpuInfoList, f *ggml.GGML, adapters, projectors []string, opts api.Options, numParallel int) (bool, uint64) {
 	// Split up the GPUs by type and try them
 	var estimatedVRAM uint64
 	for _, gpus := range allGpus.ByLibrary() {
@@ -97,10 +97,6 @@ func predictServerFit(allGpus discover.GpuInfoList, f *ggml.GGML, adapters, proj
 				return true, estimatedVRAM
 			}
 		}
-
-		if len(gpus) == 1 && gpus[0].Library == "cpu" && estimate.TotalSize <= gpus[0].FreeMemory {
-			return true, estimatedVRAM
-		}
 	}
 	return false, estimatedVRAM
 }
@@ -195,19 +191,17 @@ func estimateGPULayers(gpus []discover.GpuInfo, f *ggml.GGML, projectors []strin
 		slog.Warn("model missing blk.0 layer size")
 	}
 
-	useFlashAttention := (envconfig.FlashAttention() || f.FlashAttention()) &&
-		discover.GetGPUInfo().FlashAttentionSupported() &&
-		f.SupportsFlashAttention()
-
 	var kvct string
-	if useFlashAttention {
+	if envconfig.FlashAttention() &&
+		discover.GetGPUInfo().FlashAttentionSupported() &&
+		f.SupportsFlashAttention() {
 		requested := strings.ToLower(envconfig.KvCacheType())
 		if requested != "" && f.SupportsKVCacheType(requested) {
 			kvct = requested
 		}
 	}
 
-	kv, graphPartialOffload, graphFullOffload := f.GraphSize(uint64(opts.NumCtx), uint64(min(opts.NumCtx, opts.NumBatch)), numParallel, kvct, useFlashAttention)
+	kv, graphPartialOffload, graphFullOffload := f.GraphSize(uint64(opts.NumCtx), uint64(min(opts.NumCtx, opts.NumBatch)), numParallel, kvct)
 
 	if len(kv) > 0 {
 		layerSize += kv[0]

llm/server.go

@@ -195,11 +195,6 @@ func NewLlamaServer(gpus discover.GpuInfoList, modelPath string, f *ggml.GGML, a
 	// This will disable flash attention unless all GPUs on the system support it, even if we end up selecting a subset
 	// that can handle it.
 	fa := envconfig.FlashAttention()
-	if f.FlashAttention() {
-		slog.Info("model wants flash attention")
-		fa = true
-	}
-
 	if fa && !gpus.FlashAttentionSupported() {
 		slog.Warn("flash attention enabled but not supported by gpu")
 		fa = false
@@ -360,28 +355,23 @@ func NewLlamaServer(gpus discover.GpuInfoList, modelPath string, f *ggml.GGML, a
 
 		s.cmd.Env = append(s.cmd.Env, "OLLAMA_LIBRARY_PATH="+strings.Join(ggmlPaths, string(filepath.ListSeparator)))
 
-		envWorkarounds := []string{}
+		envWorkarounds := [][2]string{}
 		for _, gpu := range gpus {
 			envWorkarounds = append(envWorkarounds, gpu.EnvWorkarounds...)
 		}
-		// Always filter down the set of GPUs in case there are any unsupported devices that might crash
-		envWorkarounds = append(envWorkarounds, gpus.GetVisibleDevicesEnv()...)
 		pathEnvVal := strings.Join(libraryPaths, string(filepath.ListSeparator))
 
 		// Update or add the path variable with our adjusted version
 		pathNeeded := true
-		envWorkaroundDone := make([]bool, len(envWorkarounds))
 		for i := range s.cmd.Env {
 			cmp := strings.SplitN(s.cmd.Env[i], "=", 2)
 			if strings.EqualFold(cmp[0], pathEnv) {
 				s.cmd.Env[i] = pathEnv + "=" + pathEnvVal
 				pathNeeded = false
 			} else if len(envWorkarounds) != 0 {
-				for j, kv := range envWorkarounds {
-					tmp := strings.SplitN(kv, "=", 2)
-					if strings.EqualFold(cmp[0], tmp[0]) {
-						s.cmd.Env[i] = kv
-						envWorkaroundDone[j] = true
+				for _, kv := range envWorkarounds {
+					if strings.EqualFold(cmp[0], kv[0]) {
+						s.cmd.Env[i] = kv[0] + "=" + kv[1]
 					}
 				}
 			}
@@ -389,11 +379,6 @@ func NewLlamaServer(gpus discover.GpuInfoList, modelPath string, f *ggml.GGML, a
 		if pathNeeded {
 			s.cmd.Env = append(s.cmd.Env, pathEnv+"="+pathEnvVal)
 		}
-		for i, done := range envWorkaroundDone {
-			if !done {
-				s.cmd.Env = append(s.cmd.Env, envWorkarounds[i])
-			}
-		}
 
 		slog.Info("starting runner", "cmd", s.cmd)
 		slog.Debug("subprocess", "", filteredEnv(s.cmd.Env))
@@ -507,7 +492,6 @@ func (s *llamaServer) Load(ctx context.Context, gpus discover.GpuInfoList, requi
 		if !requireFull {
 			g = pickBestPartialFitByLibrary(s.ggml, []string{s.loadRequest.ProjectorPath}, s.loadRequest.LoraPath, s.options, gpus, s.numParallel)
 		} else {
-			slog.Info("model requires more memory than is currently available, evicting a model to make space", "estimate", s.estimate)
 			return ErrLoadRequiredFull
 		}
 	}
@@ -540,6 +524,10 @@ func (s *llamaServer) Load(ctx context.Context, gpus discover.GpuInfoList, requi
 		}
 	}
 
+	if requireFull && len(gpus) == 1 && gpus[0].Library == "cpu" && s.estimate.TotalSize > gpus[0].FreeMemory {
+		return ErrLoadRequiredFull
+	}
+
 	slog.Info("offload", "", s.estimate)
 
 	s.gpus = gpus
@@ -663,9 +651,7 @@ func (s *ollamaServer) Load(ctx context.Context, gpus discover.GpuInfoList, requ
 		if !success {
 			s.initModel(ctx, LoadRequest{}, LoadOperationClose)
 		}
-		if s.mem != nil {
-			s.mem.Log(slog.LevelInfo)
-		}
+		s.mem.Log(slog.LevelInfo)
 	}()
 
 	slog.Info("loading model", "model layers", s.totalLayers, "requested", s.options.NumGPU)
@@ -678,12 +664,8 @@ func (s *ollamaServer) Load(ctx context.Context, gpus discover.GpuInfoList, requ
 
 	if !(len(gpus) == 1 && gpus[0].Library == "cpu") {
 		for _, gpu := range gpus {
-			available := gpu.FreeMemory - envconfig.GpuOverhead() - gpu.MinimumMemory
-			if gpu.FreeMemory < envconfig.GpuOverhead()+gpu.MinimumMemory {
-				available = 0
-			}
 			slog.Info("gpu memory", "id", gpu.ID,
-				"available", format.HumanBytes2(available),
+				"available", format.HumanBytes2(gpu.FreeMemory-envconfig.GpuOverhead()-gpu.MinimumMemory),
 				"free", format.HumanBytes2(gpu.FreeMemory),
 				"minimum", format.HumanBytes2(gpu.MinimumMemory),
 				"overhead", format.HumanBytes2(envconfig.GpuOverhead()))
@@ -865,7 +847,7 @@ func (s *ollamaServer) createLayout(systemInfo discover.SystemInfo, systemGPUs d
 		}
 		layers[i] += memory.CPU.Weights[i].Size
 		layers[i] += memory.CPU.Cache[i].Size
-		logutil.Trace("layer to assign", "layer", i, "size", format.HumanBytes2(layers[i]))
+		slog.Log(context.TODO(), logutil.LevelTrace, "layer to assign", "layer", i, "size", format.HumanBytes2(layers[i]))
 	}
 
 	gpuLayers := ml.GPULayersList{}

logutil/logutil.go

@@ -1,7 +1,6 @@
 package logutil
 
 import (
-	"context"
 	"io"
 	"log/slog"
 	"path/filepath"
@@ -28,11 +27,3 @@ func NewLogger(w io.Writer, level slog.Level) *slog.Logger {
 		},
 	}))
 }
-
-func Trace(msg string, args ...any) {
-	slog.Log(context.TODO(), LevelTrace, msg, args...)
-}
-
-func TraceContext(ctx context.Context, msg string, args ...any) {
-	slog.Log(ctx, LevelTrace, msg, args...)
-}

ml/backend.go

@@ -266,7 +266,7 @@ func (m DeviceMemory) LogValue() slog.Value {
 // allocation is guaranteed to be provided so that if it failed, the caller can
 // accommodate that to make forward progress.
 type BackendMemory struct {
-	// InputWeights are always located on the CPU and cannot be moved
+	// InputsWeights are always located on the CPU and cannot be moved
 	InputWeights Memory
 
 	// CPU model components are located in system memory. This does not
@@ -372,7 +372,6 @@ type Context interface {
 
 	Forward(...Tensor) Context
 	Compute(...Tensor)
-	ComputeWithNotify(func(), ...Tensor) // notify callback once compute has begun
 
 	// Reserve is analogous to Compute but rather than executing a
 	// graph, simply preallocates memory. Typically called with a
@@ -397,13 +396,10 @@ type Tensor interface {
 
 	Shape() []int
 	DType() DType
-	Cast(ctx Context, dtype DType) Tensor
 
 	Bytes() []byte
 	Floats() []float32
 
-	SetValueFromIntSlice(s []int32)
-
 	Neg(ctx Context) Tensor
 	Add(ctx Context, t2 Tensor) Tensor
 	Sub(ctx Context, t2 Tensor) Tensor

ml/backend/ggml/ggml.go

@@ -82,7 +82,6 @@ type Backend struct {
 	// to the name that is used by the model definition
 	tensorLoadTargets map[string][]string
 
-	schedMu       sync.Mutex // Only one Compute can run at a time
 	sched         C.ggml_backend_sched_t
 	schedBackends []C.ggml_backend_t
 	schedBufts    []C.ggml_backend_buffer_type_t
@@ -271,7 +270,7 @@ func New(modelPath string, params ml.BackendParams) (ml.Backend, error) {
 			tt := C.ggml_new_tensor(ctxs[bt], kind, C.int(len(t.source.Shape)), (*C.int64_t)(unsafe.Pointer(&t.source.Shape[0])))
 			C.ggml_set_name(tt, cname)
 
-			logutil.Trace("created tensor", "name", name, "shape", t.source.Shape, "dtype", t.source.Kind, "buffer_type", C.GoString(C.ggml_backend_buft_name(bt)))
+			slog.Log(context.TODO(), logutil.LevelTrace, "created tensor", "name", name, "shape", t.source.Shape, "dtype", t.source.Kind, "buffer_type", C.GoString(C.ggml_backend_buft_name(bt)))
 
 			size := pad(C.ggml_backend_buft_get_alloc_size(bt, tt), C.ggml_backend_buft_get_alignment(bt))
 			if layer == -1 {
@@ -378,7 +377,7 @@ func New(modelPath string, params ml.BackendParams) (ml.Backend, error) {
 	}
 
 	for bs := range maps.Values(bbs) {
-		logutil.Trace("model weights", "buffer", C.GoString(C.ggml_backend_buffer_name(bs)),
+		slog.Log(context.TODO(), logutil.LevelTrace, "model weights", "buffer", C.GoString(C.ggml_backend_buffer_name(bs)),
 			"size", format.HumanBytes2(uint64(C.ggml_backend_buffer_get_size(bs))))
 	}
 
@@ -536,7 +535,6 @@ func (b *Backend) Load(ctx context.Context, progress func(float32)) error {
 				const BS = 17                             // MXFP4 block size
 				bts := make([]byte, 8*BS*format.KibiByte) // ~128k block aligned
 				var s uint64
-				var tmp [16]byte
 				for s < t.Size() {
 					// Stop if either the parent context has been canceled or if any of the other tensors returned an error
 					if err := ctx.Err(); err != nil {
@@ -548,13 +546,37 @@ func (b *Backend) Load(ctx context.Context, progress func(float32)) error {
 						return err
 					}
 					for j := range n / BS {
-						for i := 1; i < 9; i++ {
-							// transform a1b2c3 ... x7y8z9 -> 71xa82yb93zc
-							a, b := bts[j*BS+i], bts[j*BS+i+8]
-							tmp[2*(i-1)] = (a & 0x0F) | (b << 4)
-							tmp[2*(i-1)+1] = (a >> 4) | (b & 0xF0)
+						for i := 1; i < BS; i++ {
+							// swap nibbles
+							t_lo := bts[j*BS+i] & 0x0F
+							t_hi := bts[j*BS+i] & 0xF0
+							bts[j*BS+i] = (t_lo << 4) | (t_hi >> 4)
+						}
+						// transform aaaa...bbbb... to abababab...
+						oi := 0
+						tmp := [16]byte{}
+						for i := 1; i < 9; i++ {
+							blk_a0 := bts[j*BS+i] & 0xF0
+							blk_a1 := bts[j*BS+i] << 4
+							blk_b0 := bts[j*BS+i+8] >> 4
+							blk_b1 := bts[j*BS+i+8] & 0x0F
+							// swap once more
+							out0 := blk_a0 | blk_b0
+							out1 := blk_a1 | blk_b1
+							out_h0 := out0 & 0xF0
+							out_l0 := out0 & 0x0F
+							out_h1 := out1 & 0xF0
+							out_l1 := out1 & 0x0F
+							out0 = (out_h0 >> 4) | (out_l0 << 4)
+							out1 = (out_h1 >> 4) | (out_l1 << 4)
+							tmp[oi] = out0
+							oi++
+							tmp[oi] = out1
+							oi++
+						}
+						for i := range tmp {
+							bts[j*BS+i+1] = tmp[i]
 						}
-						copy(bts[j*BS+1:j*BS+17], tmp[:])
 					}
 
 					for _, tt := range tts {
@@ -630,18 +652,6 @@ func (b *Backend) Load(ctx context.Context, progress func(float32)) error {
 		})
 	}
 
-	// Cleanup any backend state from devices that we didn't end up using
-nextDevice:
-	for _, d := range append(gpus, append(accels, cpus...)...) {
-		for _, backend := range b.schedBackends {
-			if d == C.ggml_backend_get_device(backend) {
-				continue nextDevice
-			}
-		}
-
-		C.ggml_backend_dev_reset(d)
-	}
-
 	if err := g.Wait(); err != nil {
 		return err
 	}
@@ -759,15 +769,6 @@ func (c *Context) Forward(tensors ...ml.Tensor) ml.Context {
 }
 
 func (c *Context) Compute(tensors ...ml.Tensor) {
-	c.ComputeWithNotify(nil, tensors...)
-}
-
-func (c *Context) ComputeWithNotify(cb func(), tensors ...ml.Tensor) {
-	c.b.schedMu.Lock()
-	defer c.b.schedMu.Unlock()
-	if cb != nil {
-		go cb()
-	}
 	if status := C.ggml_backend_sched_graph_compute_async(c.b.sched, c.graph); status != C.GGML_STATUS_SUCCESS {
 		panic(fmt.Errorf("error computing ggml graph: %v", status))
 	}
@@ -811,7 +812,7 @@ func (c *Context) Reserve() {
 			}
 		}
 
-		logutil.Trace("compute graph", "backend", C.GoString(C.ggml_backend_name(c.b.schedBackends[i])),
+		slog.Log(context.TODO(), logutil.LevelTrace, "compute graph", "backend", C.GoString(C.ggml_backend_name(c.b.schedBackends[i])),
 			"buffer_type", C.GoString(C.ggml_backend_buft_name(c.b.schedBufts[i])), "size", format.HumanBytes2(uint64(bufferStatus.size)))
 	}
 
@@ -842,7 +843,23 @@ func (c *Context) newTensor(dtype ml.DType, shape []int) ml.Tensor {
 		panic("set Input or Layer before creating tensors")
 	}
 
-	cdtype := ggmlDType(dtype)
+	var cdtype uint32
+	switch dtype {
+	case ml.DTypeF32:
+		cdtype = C.GGML_TYPE_F32
+	case ml.DTypeF16:
+		cdtype = C.GGML_TYPE_F16
+	case ml.DTypeQ80:
+		cdtype = C.GGML_TYPE_Q8_0
+	case ml.DTypeQ40:
+		cdtype = C.GGML_TYPE_Q4_0
+	case ml.DTypeI32:
+		cdtype = C.GGML_TYPE_I32
+	case ml.DTypeMXFP4:
+		cdtype = C.GGML_TYPE_MXFP4
+	default:
+		panic("unsupported dtype")
+	}
 
 	if len(shape) < 1 || shape[0] == 0 {
 		var shape C.int64_t = 0
@@ -1020,12 +1037,6 @@ func (t *Tensor) Floats() (data []float32) {
 	return
 }
 
-func (t *Tensor) SetValueFromIntSlice(s []int32) {
-	if len(s) > 0 {
-		C.ggml_backend_tensor_set(t.t, unsafe.Pointer(&s[0]), 0, C.ggml_nbytes(t.t))
-	}
-}
-
 func (t *Tensor) DType() ml.DType {
 	switch t.t._type {
 	case C.GGML_TYPE_F32:
@@ -1045,32 +1056,6 @@ func (t *Tensor) DType() ml.DType {
 	}
 }
 
-func ggmlDType(dtype ml.DType) uint32 {
-	switch dtype {
-	case ml.DTypeF32:
-		return C.GGML_TYPE_F32
-	case ml.DTypeF16:
-		return C.GGML_TYPE_F16
-	case ml.DTypeQ80:
-		return C.GGML_TYPE_Q8_0
-	case ml.DTypeQ40:
-		return C.GGML_TYPE_Q4_0
-	case ml.DTypeI32:
-		return C.GGML_TYPE_I32
-	case ml.DTypeMXFP4:
-		return C.GGML_TYPE_MXFP4
-	default:
-		panic("unsupported dtype")
-	}
-}
-
-func (t *Tensor) Cast(ctx ml.Context, dtype ml.DType) ml.Tensor {
-	return &Tensor{
-		b: t.b,
-		t: C.ggml_cast(ctx.(*Context).ctx, t.t, ggmlDType(dtype)),
-	}
-}
-
 func (t *Tensor) Neg(ctx ml.Context) ml.Tensor {
 	return &Tensor{
 		b: t.b,

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

@@ -167,7 +167,6 @@ extern "C" {
     GGML_API void                          ggml_backend_dev_get_props(ggml_backend_dev_t device, struct ggml_backend_dev_props * props);
     GGML_API ggml_backend_reg_t            ggml_backend_dev_backend_reg(ggml_backend_dev_t device);
     GGML_API ggml_backend_t                ggml_backend_dev_init(ggml_backend_dev_t device, const char * params);
-    GGML_API void                          ggml_backend_dev_reset(ggml_backend_dev_t device);
     GGML_API ggml_backend_buffer_type_t    ggml_backend_dev_buffer_type(ggml_backend_dev_t device);
     GGML_API ggml_backend_buffer_type_t    ggml_backend_dev_host_buffer_type(ggml_backend_dev_t device);
     GGML_API ggml_backend_buffer_t         ggml_backend_dev_buffer_from_host_ptr(ggml_backend_dev_t device, void * ptr, size_t size, size_t max_tensor_size);

ml/backend/ggml/ggml/src/ggml-backend-impl.h

@@ -178,10 +178,6 @@ extern "C" {
         ggml_backend_event_t (*event_new)         (ggml_backend_dev_t dev);
         void                 (*event_free)        (ggml_backend_dev_t dev, ggml_backend_event_t event);
         void                 (*event_synchronize) (ggml_backend_dev_t dev, ggml_backend_event_t event);
-
-        // (optional) reset device, clearing existing allocations and context
-        // the caller must ensure that there are no outstanding buffers, as these will become invalid
-        void (*reset)(ggml_backend_dev_t dev);
     };
 
     struct ggml_backend_device {

ml/backend/ggml/ggml/src/ggml-backend.cpp

@@ -477,14 +477,6 @@ ggml_backend_t ggml_backend_dev_init(ggml_backend_dev_t device, const char * par
     return device->iface.init_backend(device, params);
 }
 
-void ggml_backend_dev_reset(ggml_backend_dev_t device) {
-    if (device->iface.reset == NULL) {
-        return;
-    }
-
-    device->iface.reset(device);
-}
-
 ggml_backend_buffer_type_t ggml_backend_dev_buffer_type(ggml_backend_dev_t device) {
     return device->iface.get_buffer_type(device);
 }

ml/backend/ggml/ggml/src/ggml-cuda/ggml-cuda.cu

@@ -103,11 +103,6 @@ int ggml_cuda_get_device() {
     return id;
 }
 
-void ggml_cuda_reset_device(int device) {
-    ggml_cuda_set_device(device);
-    CUDA_CHECK(cudaDeviceReset());
-}
-
 static cudaError_t ggml_cuda_device_malloc(void ** ptr, size_t size, int device) {
     ggml_cuda_set_device(device);
     cudaError_t err;
@@ -3248,10 +3243,7 @@ static void ggml_backend_cuda_device_get_props(ggml_backend_dev_t dev, ggml_back
     props->description = ggml_backend_cuda_device_get_description(dev);
     props->id          = ggml_backend_cuda_device_get_id(dev);
     props->type        = ggml_backend_cuda_device_get_type(dev);
-
-    // Memory reporting is disabled to avoid allocation of a CUDA primary context (~300 MB per device).
-    // If you need the memory data, call ggml_backend_dev_memory() explicitly.
-    props->memory_total = props->memory_free = 0;
+    ggml_backend_cuda_device_get_memory(dev, &props->memory_free, &props->memory_total);
 
     bool host_buffer = getenv("GGML_CUDA_NO_PINNED") == nullptr;
 #ifdef GGML_CUDA_NO_PEER_COPY
@@ -3708,11 +3700,6 @@ static void ggml_backend_cuda_device_event_synchronize(ggml_backend_dev_t dev, g
     CUDA_CHECK(cudaEventSynchronize((cudaEvent_t)event->context));
 }
 
-static void ggml_backend_cuda_device_reset(ggml_backend_dev_t dev) {
-    ggml_backend_cuda_device_context * ctx = (ggml_backend_cuda_device_context *)dev->context;
-    ggml_cuda_reset_device(ctx->device);
-}
-
 static const ggml_backend_device_i ggml_backend_cuda_device_interface = {
     /* .get_name                = */ ggml_backend_cuda_device_get_name,
     /* .get_description         = */ ggml_backend_cuda_device_get_description,
@@ -3729,7 +3716,6 @@ static const ggml_backend_device_i ggml_backend_cuda_device_interface = {
     /* .event_new               = */ ggml_backend_cuda_device_event_new,
     /* .event_free              = */ ggml_backend_cuda_device_event_free,
     /* .event_synchronize       = */ ggml_backend_cuda_device_event_synchronize,
-    /* .reset                   = */ ggml_backend_cuda_device_reset,
 };
 
 // backend reg
@@ -3849,6 +3835,7 @@ ggml_backend_reg_t ggml_backend_cuda_reg() {
                 dev_ctx->device = i;
                 dev_ctx->name = GGML_CUDA_NAME + std::to_string(i);
 
+                ggml_cuda_set_device(i);
                 cudaDeviceProp prop;
                 CUDA_CHECK(cudaGetDeviceProperties(&prop, i));
                 dev_ctx->description = prop.name;

ml/backend/ggml/ggml/src/ggml-cuda/vendors/hip.h

@@ -40,7 +40,6 @@
 #define cudaDeviceDisablePeerAccess hipDeviceDisablePeerAccess
 #define cudaDeviceEnablePeerAccess hipDeviceEnablePeerAccess
 #define cudaDeviceProp hipDeviceProp_t
-#define cudaDeviceReset hipDeviceReset
 #define cudaDeviceSynchronize hipDeviceSynchronize
 #define cudaError_t hipError_t
 #define cudaErrorPeerAccessAlreadyEnabled hipErrorPeerAccessAlreadyEnabled

ml/backend/ggml/ggml/src/ggml.cpp

@@ -19,12 +19,8 @@ static bool ggml_uncaught_exception_init = []{
         return false;
     }
     const auto prev{std::get_terminate()};
-    // GGML_ASSERT(prev != ggml_uncaught_exception);
-    if (prev != ggml_uncaught_exception) {
-        previous_terminate_handler = prev;
-    } else {
-        GGML_LOG_WARN("%s double registration of ggml_uncaught_exception\n", __func__);
-    }
+    GGML_ASSERT(prev != ggml_uncaught_exception);
+    previous_terminate_handler = prev;
     std::set_terminate(ggml_uncaught_exception);
     return true;
 }();

model/bytepairencoding.go

@@ -2,6 +2,7 @@ package model
 
 import (
 	"cmp"
+	"context"
 	"fmt"
 	"iter"
 	"log/slog"
@@ -108,7 +109,7 @@ func (bpe BytePairEncoding) Encode(s string, addSpecial bool) ([]int32, error) {
 					r = 0x0143
 				case r <= 0x0020:
 					r = r + 0x0100
-				case r >= 0x007f && r <= 0x00a0:
+				case r >= 0x007e && r <= 0x00a0:
 					r = r + 0x00a2
 				}
 
@@ -201,7 +202,7 @@ func (bpe BytePairEncoding) Encode(s string, addSpecial bool) ([]int32, error) {
 		}
 	}
 
-	logutil.Trace("encoded", "string", s, "ids", ids)
+	slog.Log(context.TODO(), logutil.LevelTrace, "encoded", "string", s, "ids", ids)
 
 	if addSpecial && len(ids) > 0 {
 		ids = bpe.vocab.addSpecials(ids)
@@ -242,6 +243,6 @@ func (bpe BytePairEncoding) Decode(ids []int32) (string, error) {
 		}
 	}
 
-	logutil.Trace("decoded", "string", sb.String(), "from", lazyIdsString{ids: ids})
+	slog.Log(context.TODO(), logutil.LevelTrace, "decoded", "string", sb.String(), "from", lazyIdsString{ids: ids})
 	return sb.String(), nil
 }

model/bytepairencoding_test.go

@@ -207,36 +207,6 @@ func TestLlama(t *testing.T) {
 			}
 		}
 	})
-
-	t.Run("roundtriping 0x00-0xFF", func(t *testing.T) {
-		t.Parallel()
-
-		for b := 0x00; b <= 0xFF; b++ {
-			input := string(rune(b))
-			ids, err := tokenizer.Encode(input, false)
-			if err != nil {
-				t.Errorf("failed to encode rune 0x%02X: %v", b, err)
-				continue
-			}
-
-			decoded, err := tokenizer.Decode(ids)
-			if err != nil {
-				t.Errorf("failed to decode rune 0x%02X: %v", b, err)
-				continue
-			}
-
-			if b == 0x00 {
-				if len(decoded) != 0 {
-					t.Errorf("Decode(Encode(0x00)) should be empty, got %v", ids)
-				}
-				continue
-			}
-
-			if decoded != input {
-				t.Errorf("rune 0x%02X failed roundtrip: got %q, want %q", b, decoded, input)
-			}
-		}
-	})
 }
 
 func BenchmarkBytePairEncoding(b *testing.B) {

model/model.go

@@ -1,10 +1,12 @@
 package model
 
 import (
+	"context"
 	"errors"
 	"fmt"
 	_ "image/jpeg"
 	_ "image/png"
+	"log/slog"
 	"os"
 	"reflect"
 	"strconv"
@@ -62,7 +64,7 @@ type MultimodalProcessor interface {
 	// This function is also responsible for updating MultimodalHash for any Multimodal
 	// that is modified to ensure that there is a unique hash value that accurately
 	// represents the contents.
-	PostTokenize([]*input.Input) ([]*input.Input, error)
+	PostTokenize([]input.Input) ([]input.Input, error)
 }
 
 // Base implements the common fields and methods for all models
@@ -196,7 +198,7 @@ func populateFields(base Base, v reflect.Value, tags ...Tag) reflect.Value {
 				names := fn(tagsCopy)
 				for _, name := range names {
 					if tensor := base.Backend().Get(strings.Join(name, ".")); tensor != nil {
-						logutil.Trace("found tensor", "", tensor)
+						slog.Log(context.TODO(), logutil.LevelTrace, "found tensor", "", tensor)
 						vv.Set(reflect.ValueOf(tensor))
 						break
 					}
@@ -276,7 +278,7 @@ func canNil(t reflect.Type) bool {
 		t.Kind() == reflect.Slice
 }
 
-func Forward(ctx ml.Context, m Model, batch input.Batch) (ml.Tensor, error) {
+func Forward(ctx ml.Context, m Model, inputs []int32, batch input.Batch) (ml.Tensor, error) {
 	if len(batch.Positions) != len(batch.Sequences) {
 		return nil, fmt.Errorf("length of positions (%v) must match length of seqs (%v)", len(batch.Positions), len(batch.Sequences))
 	}
@@ -285,6 +287,8 @@ func Forward(ctx ml.Context, m Model, batch input.Batch) (ml.Tensor, error) {
 		return nil, errors.New("batch size cannot be less than 1")
 	}
 
+	batch.Inputs = ctx.Input().FromIntSlice(inputs, len(inputs))
+
 	cache := m.Config().Cache
 	if cache != nil {
 		err := cache.StartForward(ctx, batch, false)
@@ -298,7 +302,7 @@ func Forward(ctx ml.Context, m Model, batch input.Batch) (ml.Tensor, error) {
 		return nil, err
 	}
 
-	ctx.Forward(t)
+	ctx.Forward(t).Compute(t)
 
 	return t, nil
 }

model/models/gemma3/model.go

@@ -18,7 +18,7 @@ type Model struct {
 	model.Base
 	model.SentencePieceModel
 
-	*VisionModel `gguf:"v"`
+	*VisionModel `gguf:"v,vision"`
 	*TextModel
 
 	*MultiModalProjector `gguf:"mm"`
@@ -112,8 +112,8 @@ func (m *Model) EncodeMultimodal(ctx ml.Context, multimodalData []byte) ([]input
 	return []input.Multimodal{{Tensor: visionOutputs}}, nil
 }
 
-func (m *Model) PostTokenize(inputs []*input.Input) ([]*input.Input, error) {
-	var result []*input.Input
+func (m *Model) PostTokenize(inputs []input.Input) ([]input.Input, error) {
+	var result []input.Input
 
 	for _, inp := range inputs {
 		if len(inp.Multimodal) == 0 {
@@ -122,17 +122,17 @@ func (m *Model) PostTokenize(inputs []*input.Input) ([]*input.Input, error) {
 			inputMultimodal := inp.Multimodal[0].Tensor
 
 			result = append(result,
-				&input.Input{Token: 108, SameBatch: inputMultimodal.Dim(1) + 3}, // "\n\n"
-				&input.Input{Token: 255999},                                     // "<start_of_image>""
-				&input.Input{Multimodal: []input.Multimodal{{Tensor: inputMultimodal}}, MultimodalHash: inp.MultimodalHash}, // image data is on the first placeholder
+				input.Input{Token: 108, SameBatch: inputMultimodal.Dim(1) + 3}, // "\n\n"
+				input.Input{Token: 255999},                                     // "<start_of_image>""
+				input.Input{Multimodal: []input.Multimodal{{Tensor: inputMultimodal}}, MultimodalHash: inp.MultimodalHash}, // image data is on the first placeholder
 			)
 
 			// add image token placeholders
-			result = append(result, slices.Repeat([]*input.Input{{Token: 0}}, inputMultimodal.Dim(1)-1)...)
+			result = append(result, slices.Repeat([]input.Input{{Token: 0}}, inputMultimodal.Dim(1)-1)...)
 
 			result = append(result,
-				&input.Input{Token: 256000}, // <end_of_image>
-				&input.Input{Token: 108},    // "\n\n"
+				input.Input{Token: 256000}, // <end_of_image>
+				input.Input{Token: 108},    // "\n\n"
 			)
 		}
 	}

model/models/llama4/model.go

@@ -18,7 +18,7 @@ type Model struct {
 	model.BytePairEncoding
 	ImageProcessor
 
-	*VisionModel `gguf:"v"`
+	*VisionModel `gguf:"v,vision"`
 	*Projector   `gguf:"mm"`
 	*TextModel
 }
@@ -134,16 +134,16 @@ type separator struct {
 	y bool
 }
 
-func (m *Model) PostTokenize(inputs []*input.Input) ([]*input.Input, error) {
-	var result []*input.Input
+func (m *Model) PostTokenize(inputs []input.Input) ([]input.Input, error) {
+	var result []input.Input
 	for _, inp := range inputs {
 		if len(inp.Multimodal) == 0 {
 			result = append(result, inp)
 			continue
 		}
 
-		var imageInputs []*input.Input
-		imageInputs = append(imageInputs, &input.Input{Token: 200080}) // <|image_start|>
+		var imageInputs []input.Input
+		imageInputs = append(imageInputs, input.Input{Token: 200080}) // <|image_start|>
 
 		for i, mm := range inp.Multimodal {
 			patchesPerChunk := mm.Tensor.Dim(1)
@@ -151,20 +151,20 @@ func (m *Model) PostTokenize(inputs []*input.Input) ([]*input.Input, error) {
 			if i < len(inp.Multimodal)-1 {
 				separator := mm.Data.(*separator)
 
-				imageInputs = append(imageInputs, &input.Input{Token: 200092, Multimodal: []input.Multimodal{{Tensor: mm.Tensor}}, MultimodalHash: inp.MultimodalHash, SameBatch: patchesPerChunk}) // <|patch|>
-				imageInputs = append(imageInputs, slices.Repeat([]*input.Input{{Token: 200092}}, patchesPerChunk-1)...)
+				imageInputs = append(imageInputs, input.Input{Token: 200092, Multimodal: []input.Multimodal{{Tensor: mm.Tensor}}, MultimodalHash: inp.MultimodalHash, SameBatch: patchesPerChunk}) // <|patch|>
+				imageInputs = append(imageInputs, slices.Repeat([]input.Input{{Token: 200092}}, patchesPerChunk-1)...)
 
 				if separator.x {
-					imageInputs = append(imageInputs, &input.Input{Token: 200084}) // <|tile_x_separator|>
+					imageInputs = append(imageInputs, input.Input{Token: 200084}) // <|tile_x_separator|>
 				}
 				if separator.y {
-					imageInputs = append(imageInputs, &input.Input{Token: 200085}) // <|tile_y_separator|>
+					imageInputs = append(imageInputs, input.Input{Token: 200085}) // <|tile_y_separator|>
 				}
 			} else {
-				imageInputs = append(imageInputs, &input.Input{Token: 200090})                                                                                                                      // <|image|>
-				imageInputs = append(imageInputs, &input.Input{Token: 200092, Multimodal: []input.Multimodal{{Tensor: mm.Tensor}}, MultimodalHash: inp.MultimodalHash, SameBatch: patchesPerChunk}) // <|patch|>
-				imageInputs = append(imageInputs, slices.Repeat([]*input.Input{{Token: 200092}}, patchesPerChunk-1)...)
-				imageInputs = append(imageInputs, &input.Input{Token: 200080}) // <|image_end|>
+				imageInputs = append(imageInputs, input.Input{Token: 200090})                                                                                                                      // <|image|>
+				imageInputs = append(imageInputs, input.Input{Token: 200092, Multimodal: []input.Multimodal{{Tensor: mm.Tensor}}, MultimodalHash: inp.MultimodalHash, SameBatch: patchesPerChunk}) // <|patch|>
+				imageInputs = append(imageInputs, slices.Repeat([]input.Input{{Token: 200092}}, patchesPerChunk-1)...)
+				imageInputs = append(imageInputs, input.Input{Token: 200080}) // <|image_end|>
 			}
 		}
 

model/models/mistral3/model.go

@@ -18,7 +18,7 @@ type Model struct {
 	model.BytePairEncoding
 
 	*TextModel
-	*VisionModel         `gguf:"v"`
+	*VisionModel         `gguf:"v,vision"`
 	*MultiModalProjector `gguf:"mm"`
 
 	ImageProcessor
@@ -133,22 +133,22 @@ func (m *Model) EncodeMultimodal(ctx ml.Context, multimodalData []byte) ([]input
 // [IMG]...[IMG][IMG_BREAK][IMG]...[IMG][IMG_BREAK][IMG]...[IMG][IMG_END]
 // Each sequence of [IMG]...[IMG] is a set of patches of vision embeddings
 // that can be processed together.
-func (m *Model) PostTokenize(inputs []*input.Input) ([]*input.Input, error) {
-	var result []*input.Input
+func (m *Model) PostTokenize(inputs []input.Input) ([]input.Input, error) {
+	var result []input.Input
 	for _, inp := range inputs {
 		if len(inp.Multimodal) == 0 {
 			result = append(result, inp)
 		} else {
 			for i, row := range inp.Multimodal {
 				// [IMG]
-				result = append(result, &input.Input{Token: 10, Multimodal: []input.Multimodal{{Tensor: row.Tensor}}, MultimodalHash: inp.MultimodalHash, SameBatch: row.Tensor.Dim(1)})
-				result = append(result, slices.Repeat([]*input.Input{{Token: 10}}, row.Tensor.Dim(1)-1)...)
+				result = append(result, input.Input{Token: 10, Multimodal: []input.Multimodal{{Tensor: row.Tensor}}, MultimodalHash: inp.MultimodalHash, SameBatch: row.Tensor.Dim(1)})
+				result = append(result, slices.Repeat([]input.Input{{Token: 10}}, row.Tensor.Dim(1)-1)...)
 				if i == len(inp.Multimodal)-1 {
 					// [IMG_END]
-					result = append(result, &input.Input{Token: 13})
+					result = append(result, input.Input{Token: 13})
 				} else {
 					// [IMG_BREAK]
-					result = append(result, &input.Input{Token: 12})
+					result = append(result, input.Input{Token: 12})
 				}
 			}
 		}

model/models/mllama/model.go

@@ -17,7 +17,7 @@ type Model struct {
 	model.Base
 	model.BytePairEncoding
 
-	*VisionModel `gguf:"v"`
+	*VisionModel `gguf:"v,vision"`
 	*TextModel
 
 	Projector *nn.Linear `gguf:"mm.0"`
@@ -90,7 +90,7 @@ func (m *Model) EncodeMultimodal(ctx ml.Context, multimodalData []byte) ([]input
 	return []input.Multimodal{{Tensor: projectedOutputs}}, nil
 }
 
-func (m *Model) PostTokenize(inputs []*input.Input) ([]*input.Input, error) {
+func (m *Model) PostTokenize(inputs []input.Input) ([]input.Input, error) {
 	for i := range inputs {
 		if inputs[i].Multimodal != nil {
 			inputs[i].Token = 128256 // <|image|>

model/models/qwen25vl/model.go

@@ -18,7 +18,7 @@ type Model struct {
 	model.BytePairEncoding
 
 	*TextModel
-	*VisionModel `gguf:"v"`
+	*VisionModel `gguf:"v,vision"`
 
 	ImageProcessor
 }
@@ -89,8 +89,8 @@ func (m *Model) EncodeMultimodal(ctx ml.Context, multimodalData []byte) ([]input
 }
 
 // PostTokenize arranges Qwen-2.5-VL's inputs for the forward pass
-func (m *Model) PostTokenize(inputs []*input.Input) ([]*input.Input, error) {
-	var result []*input.Input
+func (m *Model) PostTokenize(inputs []input.Input) ([]input.Input, error) {
+	var result []input.Input
 
 	var (
 		imageToken       int32 = 151655
@@ -112,16 +112,16 @@ func (m *Model) PostTokenize(inputs []*input.Input) ([]*input.Input, error) {
 				return nil, fmt.Errorf("failed to encode image prompt: %w", err)
 			}
 			for i := range pre {
-				result = append(result, &input.Input{Token: pre[i]})
+				result = append(result, input.Input{Token: pre[i]})
 			}
 
 			patchesPerChunk := inp.Multimodal[0].Tensor.Dim(1)
 
 			// First add the vision start token
-			result = append(result, &input.Input{Token: visionStartToken})
+			result = append(result, input.Input{Token: visionStartToken})
 
 			// Add the image token with the multimodal tensor data at the first position
-			result = append(result, &input.Input{
+			result = append(result, input.Input{
 				Token:          imageToken,
 				Multimodal:     inp.Multimodal,
 				MultimodalHash: inp.MultimodalHash,
@@ -129,9 +129,9 @@ func (m *Model) PostTokenize(inputs []*input.Input) ([]*input.Input, error) {
 			})
 
 			// Add the placeholder tokens for the remaining positions (tokensPerGrid-1)
-			result = append(result, slices.Repeat([]*input.Input{{Token: imageToken}}, patchesPerChunk-1)...)
+			result = append(result, slices.Repeat([]input.Input{{Token: imageToken}}, patchesPerChunk-1)...)
 
-			result = append(result, &input.Input{Token: visionEndToken})
+			result = append(result, input.Input{Token: visionEndToken})
 		}
 	}
 

model/sentencepiece.go

@@ -2,6 +2,7 @@ package model
 
 import (
 	"container/heap"
+	"context"
 	"fmt"
 	"log/slog"
 	"strconv"
@@ -24,7 +25,7 @@ func (spm SentencePieceModel) Vocabulary() *Vocabulary {
 }
 
 func NewSentencePieceModel(vocab *Vocabulary) SentencePieceModel {
-	logutil.Trace("Tokens", "num tokens", len(vocab.Values), "vals", vocab.Values[:5], "scores", vocab.Scores[:5], "types", vocab.Types[:5])
+	slog.Log(context.TODO(), logutil.LevelTrace, "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
@@ -38,7 +39,7 @@ func NewSentencePieceModel(vocab *Vocabulary) SentencePieceModel {
 		}
 	}
 
-	logutil.Trace("Token counts", "normal", counter[TOKEN_TYPE_NORMAL], "unknown", counter[TOKEN_TYPE_UNKNOWN], "control", counter[TOKEN_TYPE_CONTROL],
+	slog.Log(context.TODO(), logutil.LevelTrace, "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)
 
@@ -181,7 +182,7 @@ func (spm SentencePieceModel) Encode(s string, addSpecial bool) ([]int32, error)
 		}
 	}
 
-	logutil.Trace("encoded", "string", s, "ids", ids)
+	slog.Log(context.TODO(), logutil.LevelTrace, "encoded", "string", s, "ids", ids)
 
 	if addSpecial && len(ids) > 0 {
 		ids = spm.vocab.addSpecials(ids)
@@ -245,6 +246,6 @@ func (spm SentencePieceModel) Decode(ids []int32) (string, error) {
 		}
 	}
 
-	logutil.Trace("decoded", "ids", ids, "string", sb.String())
+	slog.Log(context.TODO(), logutil.LevelTrace, "decoded", "ids", ids, "string", sb.String())
 	return sb.String(), nil
 }

openai/openai.go

@@ -557,10 +557,12 @@ func fromChatRequest(r ChatCompletionRequest) (*api.ChatRequest, error) {
 
 	var think *api.ThinkValue
 	if r.Reasoning != nil {
+		options["reasoning"] = *r.Reasoning.Effort
 		think = &api.ThinkValue{
 			Value: *r.Reasoning.Effort,
 		}
 	} else if r.ReasoningEffort != nil {
+		options["reasoning"] = *r.ReasoningEffort
 		think = &api.ThinkValue{
 			Value: *r.ReasoningEffort,
 		}

runner/llamarunner/cache.go

@@ -46,7 +46,7 @@ func NewInputCache(lc *llama.Context, kvSize int, numSlots int, multiUserCache b
 }
 
 // Locking: Operations on InputCacheSlot (including finding one
-// through LoadCacheSlot) require a lock to be held that serializes
+// through LoadCacheSlot) require a lock to be be held that serializes
 // these operations with each other and llama.Decode
 
 type InputCacheSlot struct {

runner/ollamarunner/cache.go

@@ -78,7 +78,7 @@ func (c *InputCache) Close() {
 }
 
 // Locking: Operations on InputCacheSlot (including finding one
-// through LoadCacheSlot) require a lock to be held that serializes
+// through LoadCacheSlot) require a lock to be be held that serializes
 // these operations with each other and processBatch
 
 type InputCacheSlot struct {
@@ -86,7 +86,7 @@ type InputCacheSlot struct {
 	Id int
 
 	// Inputs that are stored in the KV cache
-	Inputs []*input.Input
+	Inputs []input.Input
 
 	// is this cache actively being processed as part of a sequence?
 	InUse bool
@@ -95,7 +95,7 @@ type InputCacheSlot struct {
 	lastUsed time.Time
 }
 
-func (c *InputCache) LoadCacheSlot(prompt []*input.Input) (*InputCacheSlot, []*input.Input, error) {
+func (c *InputCache) LoadCacheSlot(prompt []input.Input) (*InputCacheSlot, []input.Input, error) {
 	var slot *InputCacheSlot
 	var numPast int32
 	var err error
@@ -146,7 +146,7 @@ func (c *InputCache) LoadCacheSlot(prompt []*input.Input) (*InputCacheSlot, []*i
 	return slot, prompt, nil
 }
 
-func (c *InputCache) findLongestCacheSlot(prompt []*input.Input) (*InputCacheSlot, int32, error) {
+func (c *InputCache) findLongestCacheSlot(prompt []input.Input) (*InputCacheSlot, int32, error) {
 	longest := int32(-1)
 	var longestSlot *InputCacheSlot
 
@@ -169,7 +169,7 @@ func (c *InputCache) findLongestCacheSlot(prompt []*input.Input) (*InputCacheSlo
 	return longestSlot, longest, nil
 }
 
-func (c *InputCache) findBestCacheSlot(prompt []*input.Input) (*InputCacheSlot, int32, error) {
+func (c *InputCache) findBestCacheSlot(prompt []input.Input) (*InputCacheSlot, int32, error) {
 	oldest := time.Now()
 	var oldestSlot *InputCacheSlot
 
@@ -205,7 +205,7 @@ func (c *InputCache) findBestCacheSlot(prompt []*input.Input) (*InputCacheSlot,
 	if longest > 0 && longestSlot != oldestSlot {
 		slog.Debug("forking cache slot", "src", longestSlot.Id, "dst", oldestSlot.Id, "inputs", longest, "total",
 			len(longestSlot.Inputs))
-		oldestSlot.Inputs = make([]*input.Input, longest)
+		oldestSlot.Inputs = make([]input.Input, longest)
 		copy(oldestSlot.Inputs, longestSlot.Inputs[:longest])
 		if c.cache != nil {
 			c.cache.CopyPrefix(longestSlot.Id, oldestSlot.Id, longest)
@@ -215,7 +215,7 @@ func (c *InputCache) findBestCacheSlot(prompt []*input.Input) (*InputCacheSlot,
 	return oldestSlot, longest, nil
 }
 
-func countCommonPrefix(a []*input.Input, b []*input.Input) int32 {
+func countCommonPrefix(a []input.Input, b []input.Input) int32 {
 	var count int32
 
 	for i := range a {
@@ -250,7 +250,7 @@ func (c *InputCache) ShiftDiscard(inputLen int32, numKeep int32) int32 {
 }
 
 type ErrReprocessInputs struct {
-	Inputs []*input.Input
+	Inputs []input.Input
 }
 
 func (e *ErrReprocessInputs) Error() string {
@@ -283,13 +283,13 @@ func (c *InputCache) ShiftCacheSlot(slot *InputCacheSlot, numKeep int32) error {
 				"id", slot.Id, "error", err)
 
 			// Create new input slice with preserved tokens (numKeep + remaining tokens after discard)
-			newInputs := make([]*input.Input, numKeep+inputLen-(numKeep+discard))
+			newInputs := make([]input.Input, numKeep+inputLen-(numKeep+discard))
 			copy(newInputs[:numKeep], slot.Inputs[:numKeep])
 			copy(newInputs[numKeep:], slot.Inputs[numKeep+discard:])
 
 			// Reset the cache
 			_ = c.cache.Remove(slot.Id, 0, math.MaxInt32)
-			slot.Inputs = []*input.Input{}
+			slot.Inputs = []input.Input{}
 
 			// Return error with inputs that need to be reprocessed
 			return &ErrReprocessInputs{Inputs: newInputs}

runner/ollamarunner/cache_test.go

@@ -13,50 +13,50 @@ import (
 func TestCountCommon(t *testing.T) {
 	tests := []struct {
 		name     string
-		t1       []*input.Input
-		t2       []*input.Input
+		t1       []input.Input
+		t2       []input.Input
 		expected int32
 	}{
 		{
 			name:     "Equal",
-			t1:       []*input.Input{{Token: 1}, {Token: 2}, {Token: 3}},
-			t2:       []*input.Input{{Token: 1}, {Token: 2}, {Token: 3}},
+			t1:       []input.Input{{Token: 1}, {Token: 2}, {Token: 3}},
+			t2:       []input.Input{{Token: 1}, {Token: 2}, {Token: 3}},
 			expected: 3,
 		},
 		{
 			name:     "Prefix",
-			t1:       []*input.Input{{Token: 1}},
-			t2:       []*input.Input{{Token: 1}, {Token: 2}, {Token: 3}},
+			t1:       []input.Input{{Token: 1}},
+			t2:       []input.Input{{Token: 1}, {Token: 2}, {Token: 3}},
 			expected: 1,
 		},
 		{
 			name:     "Image Prefix",
-			t1:       []*input.Input{{MultimodalHash: 1}},
-			t2:       []*input.Input{{MultimodalHash: 1}, {MultimodalHash: 2}, {MultimodalHash: 3}},
+			t1:       []input.Input{{MultimodalHash: 1}},
+			t2:       []input.Input{{MultimodalHash: 1}, {MultimodalHash: 2}, {MultimodalHash: 3}},
 			expected: 1,
 		},
 		{
 			name:     "Mixed",
-			t1:       []*input.Input{{Token: 1}, {MultimodalHash: 1}},
-			t2:       []*input.Input{{Token: 1}, {MultimodalHash: 1}, {Token: 5}},
+			t1:       []input.Input{{Token: 1}, {MultimodalHash: 1}},
+			t2:       []input.Input{{Token: 1}, {MultimodalHash: 1}, {Token: 5}},
 			expected: 2,
 		},
 		{
 			name:     "Mixed, Same Length",
-			t1:       []*input.Input{{Token: 1}, {MultimodalHash: 1}},
-			t2:       []*input.Input{{Token: 1}, {MultimodalHash: 2}},
+			t1:       []input.Input{{Token: 1}, {MultimodalHash: 1}},
+			t2:       []input.Input{{Token: 1}, {MultimodalHash: 2}},
 			expected: 1,
 		},
 		{
 			name:     "Empty",
-			t1:       []*input.Input{},
-			t2:       []*input.Input{{Token: 1}, {Token: 2}, {Token: 3}},
+			t1:       []input.Input{},
+			t2:       []input.Input{{Token: 1}, {Token: 2}, {Token: 3}},
 			expected: 0,
 		},
 		{
 			name:     "Both Empty",
-			t1:       []*input.Input{},
-			t2:       []*input.Input{},
+			t1:       []input.Input{},
+			t2:       []input.Input{},
 			expected: 0,
 		},
 	}
@@ -80,7 +80,7 @@ func TestFindCacheSlot(t *testing.T) {
 	tests := []struct {
 		name    string
 		cache   InputCache
-		prompt  []*input.Input
+		prompt  []input.Input
 		longest expected
 		best    expected
 	}{
@@ -89,18 +89,18 @@ func TestFindCacheSlot(t *testing.T) {
 			cache: InputCache{slots: []InputCacheSlot{
 				{
 					Id:       0,
-					Inputs:   []*input.Input{},
+					Inputs:   []input.Input{},
 					InUse:    false,
 					lastUsed: time.Time{},
 				},
 				{
 					Id:       1,
-					Inputs:   []*input.Input{},
+					Inputs:   []input.Input{},
 					InUse:    false,
 					lastUsed: time.Time{},
 				},
 			}},
-			prompt:  []*input.Input{{Token: 1}},
+			prompt:  []input.Input{{Token: 1}},
 			longest: expected{result: 0, len: 0},
 			best:    expected{result: 0, len: 0},
 		},
@@ -109,18 +109,18 @@ func TestFindCacheSlot(t *testing.T) {
 			cache: InputCache{slots: []InputCacheSlot{
 				{
 					Id:       0,
-					Inputs:   []*input.Input{{Token: 1}},
+					Inputs:   []input.Input{{Token: 1}},
 					InUse:    false,
 					lastUsed: time.Now().Add(-time.Second),
 				},
 				{
 					Id:       1,
-					Inputs:   []*input.Input{{Token: 1}, {Token: 2}},
+					Inputs:   []input.Input{{Token: 1}, {Token: 2}},
 					InUse:    false,
 					lastUsed: time.Now().Add(-2 * time.Second),
 				},
 			}},
-			prompt:  []*input.Input{{Token: 1}, {Token: 2}},
+			prompt:  []input.Input{{Token: 1}, {Token: 2}},
 			longest: expected{result: 1, len: 2},
 			best:    expected{result: 1, len: 2},
 		},
@@ -129,18 +129,18 @@ func TestFindCacheSlot(t *testing.T) {
 			cache: InputCache{slots: []InputCacheSlot{
 				{
 					Id:       0,
-					Inputs:   []*input.Input{{Token: 1}, {Token: 2}},
+					Inputs:   []input.Input{{Token: 1}, {Token: 2}},
 					InUse:    false,
 					lastUsed: time.Now().Add(-time.Second),
 				},
 				{
 					Id:       1,
-					Inputs:   []*input.Input{},
+					Inputs:   []input.Input{},
 					InUse:    false,
 					lastUsed: time.Time{},
 				},
 			}},
-			prompt:  []*input.Input{{Token: 2}},
+			prompt:  []input.Input{{Token: 2}},
 			longest: expected{result: 0, len: 0},
 			best:    expected{result: 1, len: 0},
 		},
@@ -150,19 +150,19 @@ func TestFindCacheSlot(t *testing.T) {
 				slots: []InputCacheSlot{
 					{
 						Id:       0,
-						Inputs:   []*input.Input{{Token: 1}, {Token: 2}},
+						Inputs:   []input.Input{{Token: 1}, {Token: 2}},
 						InUse:    false,
 						lastUsed: time.Now().Add(-time.Second),
 					},
 					{
 						Id:       1,
-						Inputs:   []*input.Input{},
+						Inputs:   []input.Input{},
 						InUse:    false,
 						lastUsed: time.Time{},
 					},
 				},
 			},
-			prompt:  []*input.Input{{Token: 1}},
+			prompt:  []input.Input{{Token: 1}},
 			longest: expected{result: 0, len: 1},
 			best:    expected{result: 1, len: 1},
 		},
@@ -171,18 +171,18 @@ func TestFindCacheSlot(t *testing.T) {
 			cache: InputCache{slots: []InputCacheSlot{
 				{
 					Id:       0,
-					Inputs:   []*input.Input{{Token: 1}},
+					Inputs:   []input.Input{{Token: 1}},
 					InUse:    false,
 					lastUsed: time.Now().Add(-time.Second),
 				},
 				{
 					Id:       1,
-					Inputs:   []*input.Input{{Token: 1}, {Token: 2}},
+					Inputs:   []input.Input{{Token: 1}, {Token: 2}},
 					InUse:    false,
 					lastUsed: time.Now().Add(-2 * time.Second),
 				},
 			}},
-			prompt:  []*input.Input{{Token: 2}, {Token: 3}},
+			prompt:  []input.Input{{Token: 2}, {Token: 3}},
 			longest: expected{result: 0, len: 0},
 			best:    expected{result: 1, len: 0},
 		},
@@ -191,18 +191,18 @@ func TestFindCacheSlot(t *testing.T) {
 			cache: InputCache{slots: []InputCacheSlot{
 				{
 					Id:       0,
-					Inputs:   []*input.Input{{Token: 1}, {Token: 2}},
+					Inputs:   []input.Input{{Token: 1}, {Token: 2}},
 					InUse:    true,
 					lastUsed: time.Now().Add(-time.Second),
 				},
 				{
 					Id:       1,
-					Inputs:   []*input.Input{{Token: 1}},
+					Inputs:   []input.Input{{Token: 1}},
 					InUse:    false,
 					lastUsed: time.Now().Add(-2 * time.Second),
 				},
 			}},
-			prompt:  []*input.Input{{Token: 1}, {Token: 2}},
+			prompt:  []input.Input{{Token: 1}, {Token: 2}},
 			longest: expected{result: 1, len: 1},
 			best:    expected{result: 1, len: 2},
 		},
@@ -300,7 +300,7 @@ func TestLoadCacheSlot(t *testing.T) {
 	tests := []struct {
 		name           string
 		cache          InputCache
-		prompt         []*input.Input
+		prompt         []input.Input
 		wantErr        bool
 		expectedSlotId int
 		expectedPrompt int // expected length of remaining prompt
@@ -312,19 +312,19 @@ func TestLoadCacheSlot(t *testing.T) {
 				slots: []InputCacheSlot{
 					{
 						Id:       0,
-						Inputs:   []*input.Input{{Token: 1}, {Token: 2}},
+						Inputs:   []input.Input{{Token: 1}, {Token: 2}},
 						InUse:    false,
 						lastUsed: time.Now().Add(-time.Second),
 					},
 					{
 						Id:       1,
-						Inputs:   []*input.Input{},
+						Inputs:   []input.Input{},
 						InUse:    false,
 						lastUsed: time.Now().Add(-2 * time.Second),
 					},
 				},
 			},
-			prompt:         []*input.Input{{Token: 1}, {Token: 2}, {Token: 3}},
+			prompt:         []input.Input{{Token: 1}, {Token: 2}, {Token: 3}},
 			wantErr:        false,
 			expectedSlotId: 0,
 			expectedPrompt: 1, // Only token 3 remains
@@ -336,19 +336,19 @@ func TestLoadCacheSlot(t *testing.T) {
 				slots: []InputCacheSlot{
 					{
 						Id:       0,
-						Inputs:   []*input.Input{{Token: 1}, {Token: 2}},
+						Inputs:   []input.Input{{Token: 1}, {Token: 2}},
 						InUse:    false,
 						lastUsed: time.Now().Add(-time.Second),
 					},
 					{
 						Id:       1,
-						Inputs:   []*input.Input{},
+						Inputs:   []input.Input{},
 						InUse:    false,
 						lastUsed: time.Now().Add(-2 * time.Second),
 					},
 				},
 			},
-			prompt:         []*input.Input{{Token: 1}, {Token: 2}, {Token: 3}},
+			prompt:         []input.Input{{Token: 1}, {Token: 2}, {Token: 3}},
 			wantErr:        false,
 			expectedSlotId: 0,
 			expectedPrompt: 1, // Only token 3 remains
@@ -360,13 +360,13 @@ func TestLoadCacheSlot(t *testing.T) {
 				slots: []InputCacheSlot{
 					{
 						Id:       0,
-						Inputs:   []*input.Input{{Token: 1}, {Token: 2}},
+						Inputs:   []input.Input{{Token: 1}, {Token: 2}},
 						InUse:    false,
 						lastUsed: time.Now().Add(-time.Second),
 					},
 				},
 			},
-			prompt:         []*input.Input{{Token: 1}, {Token: 2}},
+			prompt:         []input.Input{{Token: 1}, {Token: 2}},
 			wantErr:        false,
 			expectedSlotId: 0,
 			expectedPrompt: 1, // Should leave 1 token for sampling
@@ -378,13 +378,13 @@ func TestLoadCacheSlot(t *testing.T) {
 				slots: []InputCacheSlot{
 					{
 						Id:       0,
-						Inputs:   []*input.Input{{Token: 1}, {Token: 2}},
+						Inputs:   []input.Input{{Token: 1}, {Token: 2}},
 						InUse:    true,
 						lastUsed: time.Now().Add(-time.Second),
 					},
 				},
 			},
-			prompt:         []*input.Input{{Token: 1}, {Token: 2}, {Token: 3}},
+			prompt:         []input.Input{{Token: 1}, {Token: 2}, {Token: 3}},
 			wantErr:        true,
 			expectedSlotId: -1,
 			expectedPrompt: -1,
@@ -452,7 +452,7 @@ func TestShiftCacheSlot(t *testing.T) {
 	tests := []struct {
 		name          string
 		numCtx        int32
-		inputs        []*input.Input
+		inputs        []input.Input
 		numKeep       int32
 		cacheErr      bool
 		wantErr       any
@@ -461,7 +461,7 @@ func TestShiftCacheSlot(t *testing.T) {
 		{
 			name:          "Normal shift",
 			numCtx:        10,
-			inputs:        []*input.Input{{Token: 1}, {Token: 2}, {Token: 3}, {Token: 4}, {Token: 5}, {Token: 6}, {Token: 7}, {Token: 8}, {Token: 9}, {Token: 10}},
+			inputs:        []input.Input{{Token: 1}, {Token: 2}, {Token: 3}, {Token: 4}, {Token: 5}, {Token: 6}, {Token: 7}, {Token: 8}, {Token: 9}, {Token: 10}},
 			numKeep:       2,
 			cacheErr:      false, // No error
 			wantErr:       nil,
@@ -470,7 +470,7 @@ func TestShiftCacheSlot(t *testing.T) {
 		{
 			name:          "Cache removal fails",
 			numCtx:        10,
-			inputs:        []*input.Input{{Token: 1}, {Token: 2}, {Token: 3}, {Token: 4}, {Token: 5}, {Token: 6}, {Token: 7}, {Token: 8}, {Token: 9}, {Token: 10}},
+			inputs:        []input.Input{{Token: 1}, {Token: 2}, {Token: 3}, {Token: 4}, {Token: 5}, {Token: 6}, {Token: 7}, {Token: 8}, {Token: 9}, {Token: 10}},
 			numKeep:       2,
 			cacheErr:      true,
 			wantErr:       &ErrReprocessInputs{},
@@ -487,7 +487,7 @@ func TestShiftCacheSlot(t *testing.T) {
 			}
 			slot := &InputCacheSlot{
 				Id:     123,
-				Inputs: make([]*input.Input, len(tt.inputs)),
+				Inputs: make([]input.Input, len(tt.inputs)),
 			}
 			copy(slot.Inputs, tt.inputs)
 

runner/ollamarunner/runner.go

@@ -17,7 +17,6 @@ import (
 	"reflect"
 	"regexp"
 	"runtime"
-	"runtime/debug"
 	"strconv"
 	"strings"
 	"sync"
@@ -52,10 +51,10 @@ type Sequence struct {
 	iBatch int
 
 	// prompt inputs left to evaluate
-	inputs []*input.Input
+	inputs []input.Input
 
 	// inputs that have been added to a batch but not yet submitted to Forward
-	pendingInputs []*input.Input
+	pendingInputs []input.Input
 
 	// tokens that have been generated but not returned yet (e.g. for stop sequences)
 	pendingResponses []string
@@ -183,8 +182,8 @@ func (s *Server) NewSequence(prompt string, images []llm.ImageData, params NewSe
 // inputs processes the prompt and images into a list of inputs
 // by splitting the prompt on [img-<n>] tags, tokenizing text and
 // decoding images
-func (s *Server) inputs(prompt string, images []llm.ImageData) ([]*input.Input, []ml.Context, multimodalStore, error) {
-	var inputs []*input.Input
+func (s *Server) inputs(prompt string, images []llm.ImageData) ([]input.Input, []ml.Context, multimodalStore, error) {
+	var inputs []input.Input
 	var ctxs []ml.Context
 	var mmStore multimodalStore
 
@@ -211,7 +210,7 @@ func (s *Server) inputs(prompt string, images []llm.ImageData) ([]*input.Input,
 		}
 
 		for _, t := range tokens {
-			inputs = append(inputs, &input.Input{Token: t})
+			inputs = append(inputs, input.Input{Token: t})
 		}
 
 		// image - decode and store
@@ -244,7 +243,7 @@ func (s *Server) inputs(prompt string, images []llm.ImageData) ([]*input.Input,
 
 			mmStore.addMultimodal(imageEmbeddings)
 
-			inputs = append(inputs, &input.Input{Multimodal: imageEmbeddings, MultimodalHash: imageHash})
+			inputs = append(inputs, input.Input{Multimodal: imageEmbeddings, MultimodalHash: imageHash})
 			postTokenize = true
 		}
 	}
@@ -260,37 +259,6 @@ func (s *Server) inputs(prompt string, images []llm.ImageData) ([]*input.Input,
 	return inputs, ctxs, mmStore, nil
 }
 
-type batchState struct {
-	// id provides a counter for trace logging batches
-	id int
-
-	// ctx holds the backend context used for this batch
-	ctx ml.Context
-
-	// modelOutput holds the outputs from this batch
-	modelOutput ml.Tensor
-
-	// batchInputs holds the input token pointers which may start as
-	// placeholders later filled in before calling ctx.Compute
-	batchInputs []*input.Input
-
-	// batch contains the inputs for a model forward pass
-	batch input.Batch
-
-	// full set of seqs at the time this batch was initiated
-	seqs []*Sequence
-
-	// Signaled when this batches inputs are ready and compute can proceed
-	inputsReadyCh chan struct{}
-
-	// Signaling when Compute is about to begin on this batch, and
-	// seqs have been updated to prepare for the next batch
-	computeStartedCh chan struct{}
-
-	// Signaled when this batches outputs are complete and the next batch can proceed
-	outputsReadyCh chan struct{}
-}
-
 type Server struct {
 	// modelPath is the location of the model to be loaded
 	modelPath string
@@ -322,12 +290,6 @@ type Server struct {
 	// TODO (jmorganca): make this n_batch
 	batchSize int
 
-	// Used to signal a hard failure during async processing which will panic the runner
-	hardErrCh chan error
-
-	// Simple counter used only for trace logging batches
-	batchID int
-
 	// protects access to everything below this line
 	// this is context state needed for decoding
 	mu sync.Mutex
@@ -400,66 +362,33 @@ func (s *Server) removeSequence(seqIndex int, reason llm.DoneReason) {
 	s.seqsSem.Release(1)
 }
 
-// track batch state between forwardBatch, computeBatch and predictForwardBatch
-
 func (s *Server) run(ctx context.Context) {
 	s.ready.Wait()
 
-	var activeBatch batchState
 	for {
 		select {
 		case <-ctx.Done():
 			return
-		case err := <-s.hardErrCh:
-			panic(err)
 		default:
-			var err error
-			activeBatch, err = s.forwardBatch(activeBatch)
+			err := s.processBatch()
 			if err != nil {
 				panic(err)
 			}
-			go s.computeBatch(activeBatch)
 		}
 	}
 }
 
-// forwardBatch will calculate a batch.
-func (s *Server) forwardBatch(pendingBatch batchState) (nextBatch batchState, err error) {
-	// If we have a pending batch still processing, wait until Compute has started
-	// before setting up the next batch so the seqs inputs are ready to receive their
-	// token values and we get the correct input pointers for the batchInputs
-	if pendingBatch.ctx != nil {
-		slog.Log(context.TODO(), logutil.LevelTrace, "forwardBatch waiting for compute to start", "pendingBatch.id", pendingBatch.id)
-		<-pendingBatch.computeStartedCh
-		slog.Log(context.TODO(), logutil.LevelTrace, "forwardBatch compute started, setting up next batch", "pendingBatch.id", pendingBatch.id, "id", s.batchID)
-		nextBatch.inputsReadyCh = pendingBatch.outputsReadyCh // Chain the ouputs from the pending batch to the next inputs batch
-	} else {
-		slog.Log(context.TODO(), logutil.LevelTrace, "forwardBatch no pending batch detected", "batchID", s.batchID)
-		// No pendingBatch, so the inputs will be ready in the seqs immediately
-		nextBatch.inputsReadyCh = make(chan struct{}, 1)
-		nextBatch.inputsReadyCh <- struct{}{}
-	}
-
+func (s *Server) processBatch() error {
 	s.mu.Lock()
 	for s.allNil() {
 		s.cond.Wait() // Wait until an item is added
 	}
 	defer s.mu.Unlock()
 
-	nextBatch.ctx = s.model.Backend().NewContext()
-	defer func() {
-		if err != nil {
-			nextBatch.ctx.Close()
-			nextBatch.ctx = nil
-		}
-	}()
-	nextBatch.id = s.batchID
-	nextBatch.seqs = append([]*Sequence{}, s.seqs...)
-	nextBatch.computeStartedCh = make(chan struct{}, 1)
-	nextBatch.outputsReadyCh = make(chan struct{}, 1)
+	ctx := s.model.Backend().NewContext()
+	defer ctx.Close()
 
-	// Prepare the seqs and batch, but defer the input token values as we may not be ready yet
-	var batchInputs []*input.Input
+	var batchInputs []int32
 	var batch input.Batch
 
 	resumeSeq := -1
@@ -467,6 +396,7 @@ func (s *Server) forwardBatch(pendingBatch batchState) (nextBatch batchState, er
 	for range s.seqs {
 		seqIdx = (seqIdx + 1) % len(s.seqs)
 		seq := s.seqs[seqIdx]
+
 		if seq == nil {
 			continue
 		}
@@ -474,13 +404,12 @@ func (s *Server) forwardBatch(pendingBatch batchState) (nextBatch batchState, er
 		// if past the num predict limit
 		if seq.numPredict > 0 && seq.numPredicted >= seq.numPredict {
 			s.removeSequence(seqIdx, llm.DoneReasonLength)
-			nextBatch.seqs[seqIdx] = nil
 			continue
 		}
 
 		if !s.cache.enabled {
 			seq.inputs = append(seq.cache.Inputs, seq.inputs...)
-			seq.cache.Inputs = []*input.Input{}
+			seq.cache.Inputs = []input.Input{}
 		}
 
 		batchSize := s.batchSize
@@ -513,28 +442,25 @@ func (s *Server) forwardBatch(pendingBatch batchState) (nextBatch batchState, er
 					break
 				}
 
-				err = s.cache.ShiftCacheSlot(seq.cache, seq.numKeep)
+				err := s.cache.ShiftCacheSlot(seq.cache, seq.numKeep)
 				if err != nil {
 					var reprocess *ErrReprocessInputs
 					if errors.As(err, &reprocess) {
 						// Prepend these inputs to the sequence's inputs queue for reprocessing
 						seq.inputs = append(reprocess.Inputs, seq.inputs...)
 						// Skip this sequence but continue processing the rest
-						nextBatch.seqs[seqIdx] = nil // clear this sequence for this batch
-						err = nil
 						continue
 					} else {
-						return
+						return err
 					}
 				}
 			}
 
-			batchInputs = append(batchInputs, seq.inputs[i])
+			batchInputs = append(batchInputs, inp.Token)
 			if inp.Multimodal != nil {
-				var mm []input.Multimodal
-				mm, err = seq.mmStore.getMultimodal(s.model.Backend(), nextBatch.ctx, inp.Multimodal, false)
+				mm, err := seq.mmStore.getMultimodal(s.model.Backend(), ctx, inp.Multimodal, false)
 				if err != nil {
-					return
+					return err
 				}
 				batch.Multimodal = append(batch.Multimodal, input.MultimodalIndex{Index: len(batchInputs) - 1, Multimodal: mm})
 			}
@@ -546,7 +472,6 @@ func (s *Server) forwardBatch(pendingBatch batchState) (nextBatch batchState, er
 			if i+1 == len(seq.inputs) {
 				batch.Outputs = append(batch.Outputs, int32(len(batchInputs)-1))
 			}
-			slog.Log(context.TODO(), logutil.LevelTrace, "forwardBatch iBatch", "batchID", s.batchID, "seqIdx", seqIdx, "seq.iBatch", seq.iBatch, "i+1", i+1, "len(seq.inputs)", len(seq.inputs))
 			seq.pendingInputs = append(seq.pendingInputs, inp)
 		}
 
@@ -560,129 +485,36 @@ func (s *Server) forwardBatch(pendingBatch batchState) (nextBatch batchState, er
 	}
 
 	if len(batchInputs) == 0 {
-		slog.Log(context.TODO(), logutil.LevelTrace, "forwardBatch no batchInputs, going idle", "batchID", s.batchID)
-		nextBatch.ctx.Close()
-		nextBatch.ctx = nil
-		return
+		return nil
 	}
-	s.batchID++
 
-	// Actual batchInputs values will be injected into the batch.Inputs tensor before calling Compute
-	batch.Inputs = nextBatch.ctx.Input().Empty(ml.DTypeI32, len(batchInputs))
-	nextBatch.modelOutput, err = model.Forward(nextBatch.ctx, s.model, batch)
+	modelOutput, err := model.Forward(ctx, s.model, batchInputs, batch)
 	if err != nil {
-		err = fmt.Errorf("failed to build graph: %w", err)
-		return
-	}
-	nextBatch.batchInputs = batchInputs
-	nextBatch.batch = batch
-
-	return
-}
-
-// Async processing of the next batch
-func (s *Server) computeBatch(activeBatch batchState) {
-	if activeBatch.ctx == nil {
-		// Nothing to compute
-		return
-	}
-	defer activeBatch.ctx.Close()
-
-	// Wait until inputs are ready
-	slog.Log(context.TODO(), logutil.LevelTrace, "computeBatch: waiting for inputs to be ready", "batchID", activeBatch.id)
-	<-activeBatch.inputsReadyCh
-	slog.Log(context.TODO(), logutil.LevelTrace, "computeBatch: inputs are ready", "batchID", activeBatch.id)
-
-	// Once we complete, signal the next batch of inputs are ready
-	// This will unblock the next computeBatch, or forwardBatch if new seqs come in
-	defer func() {
-		slog.Log(context.TODO(), logutil.LevelTrace, "computeBatch: outputs are ready", "batchID", activeBatch.id)
-		activeBatch.outputsReadyCh <- struct{}{}
-	}()
-
-	s.mu.Lock()
-
-	// Gather the actual input token values now that they're ready
-	batchInputs := make([]int32, len(activeBatch.batchInputs))
-	for i := range batchInputs {
-		batchInputs[i] = activeBatch.batchInputs[i].Token
+		return fmt.Errorf("failed to decode batch: %w", err)
 	}
 
-	// Now we run part of the decoding algorithm to adjust the seq.inputs with placeholder tokens
-	// so that forwardBatch can build a batchInputs set which will eventually contain the actual
-	// decoded tokens.
-	nextBatchTokens := make([]*input.Input, len(s.seqs))
-	iBatches := make([]int, len(s.seqs)) // Record the iBatch values before releasing the lock
+	logits := modelOutput.Floats()
+
 	for i, seq := range s.seqs {
-		iBatches[i] = -1
 		if seq == nil {
 			continue
 		}
-		// Skip over any newly added or skipped sequences
-		if activeBatch.seqs[i] == nil {
-			continue
-		}
 
-		// Detect if the sequence we're processing has already been completed and replaced
-		// with a new sequence
-		if seq != activeBatch.seqs[i] {
-			slog.Log(context.TODO(), logutil.LevelTrace, "computeBatch: sequence replaced, discarding its results", "batchID", activeBatch.id, "seqIdx", i)
-			continue
-		}
-
-		// Pending inputs will actually be in the cache after we call Compute.
-		// However, we have already resolved any placeholder tokens.
-		//
-		// It's possible for incoming sequences to look at the values that we've
-		// added to the cache here and start relying on them before we've done
-		// the computation. This is OK as long as we ensure that this batch's
-		// computation happens before any future batch's and we never fail
-		// (unless we take down the whole runner).
+		// After calling Forward, pending inputs are now in the cache
 		if len(seq.pendingInputs) > 0 {
 			seq.cache.Inputs = append(seq.cache.Inputs, seq.pendingInputs...)
-			seq.pendingInputs = []*input.Input{}
+			seq.pendingInputs = []input.Input{}
 		}
 
 		// don't sample prompt processing
 		if len(seq.inputs) != 0 {
 			if !s.cache.enabled {
-				s.hardErrCh <- fmt.Errorf("caching disabled but unable to fit entire input in a batch")
-				s.mu.Unlock()
-				return
+				return errors.New("caching disabled but unable to fit entire input in a batch")
 			}
 			continue
 		}
 
 		seq.numPredicted++
-		nextToken := &input.Input{Token: 0} // placeholder we'll fill in after Compute/Floats
-		seq.inputs = []*input.Input{nextToken}
-		nextBatchTokens[i] = nextToken
-		iBatches[i] = seq.iBatch
-	}
-
-	// At this point the seqs are ready for forwardBatch to move forward so unblock
-	s.mu.Unlock()
-
-	activeBatch.batch.Inputs.SetValueFromIntSlice(batchInputs)
-	activeBatch.ctx.ComputeWithNotify(
-		func() {
-			slog.Log(context.TODO(), logutil.LevelTrace, "computeBatch: signaling computeStartedCh", "batchID", activeBatch.id)
-			activeBatch.computeStartedCh <- struct{}{}
-		},
-		activeBatch.modelOutput)
-	logits := activeBatch.modelOutput.Floats()
-
-	slog.Log(context.TODO(), logutil.LevelTrace, "computeBatch: logits ready", "batchID", activeBatch.id)
-
-	s.mu.Lock()
-	defer s.mu.Unlock()
-
-	slog.Log(context.TODO(), logutil.LevelTrace, "computeBatch: decoding", "batchID", activeBatch.id)
-	for i, seq := range s.seqs {
-		if seq == nil || nextBatchTokens[i] == nil {
-			continue
-		}
-
 		if seq.numPredicted == 1 {
 			seq.startGenerationTime = time.Now()
 		}
@@ -690,38 +522,36 @@ func (s *Server) computeBatch(activeBatch batchState) {
 		// if done processing the prompt, generate an embedding and return
 		if seq.embeddingOnly {
 			// TODO(jessegross): Embedding support
-			slog.Warn("generation of embedding outputs not yet supported", "id", activeBatch.id, "seqIdx", i)
+			slog.Warn("generation of embedding outputs not yet supported")
 			s.removeSequence(i, llm.DoneReasonStop)
 			continue
 		}
 
 		// sample a token
-		vocabSize := len(logits) / len(activeBatch.batch.Outputs)
-		slog.Log(context.TODO(), logutil.LevelTrace, "computeBatch: vocab details", "batchID", activeBatch.id, "seqIdx", i, "len(logits)", len(logits), "len(activeBatch.batch.Outputs)", len(activeBatch.batch.Outputs), "vocabSize", vocabSize, "iBatches", iBatches)
-		token, err := seq.sampler.Sample(logits[iBatches[i]*vocabSize : (iBatches[i]+1)*vocabSize])
-		if err != nil {
-			s.hardErrCh <- fmt.Errorf("failed to sample token: %w", err)
-			return
-		}
+		vocabSize := len(logits) / len(batch.Outputs)
 
-		nextBatchTokens[i].Token = token
+		token, err := seq.sampler.Sample(logits[seq.iBatch*vocabSize : (seq.iBatch+1)*vocabSize])
+		if err != nil {
+			return fmt.Errorf("failed to sample token: %w", err)
+		}
 
 		// if it's an end of sequence token, break
 		if s.model.(model.TextProcessor).Is(token, model.SpecialEOS) {
 			// TODO (jmorganca): we should send this back
 			// as it's important for the /api/generate context
 			// seq.responses <- piece
-			slog.Log(context.TODO(), logutil.LevelTrace, "computeBatch: EOS", "batchID", activeBatch.id, "seqIdx", i)
+
 			s.removeSequence(i, llm.DoneReasonStop)
 			continue
 		}
 
 		piece, err := s.model.(model.TextProcessor).Decode([]int32{token})
 		if err != nil {
-			s.hardErrCh <- fmt.Errorf("failed to decode token: %w", err)
-			return
+			return err
 		}
 
+		seq.inputs = []input.Input{{Token: token}}
+
 		seq.pendingResponses = append(seq.pendingResponses, piece)
 		sequence := strings.Join(seq.pendingResponses, "")
 
@@ -745,7 +575,6 @@ func (s *Server) computeBatch(activeBatch batchState) {
 			if tokenTruncated || origLen == newLen {
 				tokenLen--
 			}
-
 			seq.cache.Inputs = seq.cache.Inputs[:tokenLen]
 
 			s.removeSequence(i, llm.DoneReasonStop)
@@ -764,6 +593,8 @@ func (s *Server) computeBatch(activeBatch batchState) {
 			s.removeSequence(i, llm.DoneReasonConnectionClosed)
 		}
 	}
+
+	return nil
 }
 
 func (s *Server) completion(w http.ResponseWriter, r *http.Request) {
@@ -905,10 +736,7 @@ func (s *Server) reserveWorstCaseGraph() error {
 	defer ctx.Close()
 
 	var err error
-	inputs := make([]*input.Input, s.batchSize)
-	for i := range inputs {
-		inputs[i] = &input.Input{}
-	}
+	inputs := make([]input.Input, s.batchSize)
 	mmStore := newMultimodalStore()
 
 	// Multimodal strategy:
@@ -950,11 +778,8 @@ func (s *Server) reserveWorstCaseGraph() error {
 			}
 
 			if len(inputs) < s.batchSize {
-				newInputs := make([]*input.Input, s.batchSize)
+				newInputs := make([]input.Input, s.batchSize)
 				copy(newInputs, inputs)
-				for i := len(inputs); i < s.batchSize; i++ {
-					newInputs[i] = &input.Input{}
-				}
 				inputs = newInputs
 			}
 		}
@@ -1017,7 +842,6 @@ func (s *Server) allocModel(
 	// Convert memory allocation panics to errors
 	defer func() {
 		if r := recover(); r != nil {
-			debug.PrintStack()
 			if err, ok := r.(error); ok {
 				panicErr = err
 			} else {
@@ -1187,7 +1011,6 @@ func Execute(args []string) error {
 	server := &Server{
 		modelPath: *mpath,
 		status:    llm.ServerStatusLaunched,
-		hardErrCh: make(chan error, 1),
 	}
 
 	server.cond = sync.NewCond(&server.mu)

server/harmonyparser.go

@@ -1,8 +1,9 @@
-package harmony
+package server
 
 import (
-	"fmt"
+	"context"
 	"log/slog"
+	"slices"
 	"strings"
 	"unicode"
 
@@ -18,6 +19,18 @@ const (
 	harmonyParserState_ParsingContent
 )
 
+func shouldUseHarmony(model Model) bool {
+	if slices.Contains([]string{"gptoss", "gpt-oss"}, model.Config.ModelFamily) {
+		// heuristic to check whether the template expects to be parsed via harmony:
+		// search for harmony tags that are nearly always used
+		if model.Template.Contains("<|start|>") && model.Template.Contains("<|end|>") {
+			return true
+		}
+	}
+
+	return false
+}
+
 func (s harmonyParserState) String() string {
 	switch s {
 	// we're looking for the message start tag
@@ -262,21 +275,19 @@ const (
 // HarmonyMessageHandler processes harmony events and accumulates content appropriately.
 // This is a higher level interface that maps harmony concepts into ollama concepts
 type HarmonyMessageHandler struct {
-	state           harmonyMessageState
-	HarmonyParser   *HarmonyParser
-	FunctionNameMap *FunctionNameMap
+	state         harmonyMessageState
+	harmonyParser *HarmonyParser
 }
 
 // NewHarmonyMessageHandler creates a new message handler
 func NewHarmonyMessageHandler() *HarmonyMessageHandler {
 	return &HarmonyMessageHandler{
 		state: harmonyMessageState_Normal,
-		HarmonyParser: &HarmonyParser{
+		harmonyParser: &HarmonyParser{
 			MessageStartTag: "<|start|>",
 			MessageEndTag:   "<|end|>",
 			HeaderEndTag:    "<|message|>",
 		},
-		FunctionNameMap: NewFunctionNameMap(),
 	}
 }
 
@@ -287,11 +298,11 @@ func (h *HarmonyMessageHandler) AddContent(content string, toolParser *HarmonyTo
 	thinkingSb := strings.Builder{}
 	toolContentSb := strings.Builder{}
 
-	events := h.HarmonyParser.AddContent(content)
+	events := h.harmonyParser.AddContent(content)
 	for _, event := range events {
 		switch event := event.(type) {
 		case HarmonyEventHeaderComplete:
-			logutil.Trace("harmony event header complete", "header", event.Header)
+			slog.Log(context.TODO(), logutil.LevelTrace, "harmony event header complete", "header", event.Header)
 			switch event.Header.Channel {
 			case "analysis":
 				if event.Header.Recipient != "" {
@@ -314,7 +325,7 @@ func (h *HarmonyMessageHandler) AddContent(content string, toolParser *HarmonyTo
 				h.state = harmonyMessageState_Normal
 			}
 		case HarmonyEventContentEmitted:
-			logutil.Trace("harmony event content", "content", event.Content, "state", h.state)
+			slog.Log(context.TODO(), logutil.LevelTrace, "harmony event content", "content", event.Content, "state", h.state)
 			if h.state == harmonyMessageState_Normal {
 				contentSb.WriteString(event.Content)
 			} else if h.state == harmonyMessageState_Thinking {
@@ -367,97 +378,3 @@ func (a *HarmonyToolCallAccumulator) Drain() (*string, string) {
 func (a *HarmonyToolCallAccumulator) Content() string {
 	return a.acc.String()
 }
-
-// FunctionNameMap maps a user-specified function name to a valid function
-// name for harmony (which look like TypeScript identifiers). This is needed to
-// transform user-specified function names, which might contain characters that
-// are not allowed in TypeScript identifiers
-type FunctionNameMap struct {
-	userToHarmony map[string]string
-	harmonyToUser map[string]string
-}
-
-func NewFunctionNameMap() *FunctionNameMap {
-	return &FunctionNameMap{
-		userToHarmony: make(map[string]string),
-		harmonyToUser: make(map[string]string),
-	}
-}
-
-func (m *FunctionNameMap) ConvertAndAdd(userFunctionName string) string {
-	harmonyFunctionName := m.deriveName(userFunctionName)
-	m.userToHarmony[userFunctionName] = harmonyFunctionName
-	m.harmonyToUser[harmonyFunctionName] = userFunctionName
-	return harmonyFunctionName
-}
-
-// OriginalFromConverted looks up the reverse-mapping of a previously-converted
-// user->harmony function name. To unmap reliably, the mapping must exist, as
-// the conversion process is not reversible without the appropriate state
-func (m *FunctionNameMap) OriginalFromConverted(harmonyFunctionName string) string {
-	if userFunctionName, ok := m.harmonyToUser[harmonyFunctionName]; ok {
-		return userFunctionName
-	}
-	slog.Warn("harmony parser: no reverse mapping found for function name", "harmonyFunctionName", harmonyFunctionName)
-	// fallback to the original function name if we can't find a mapping
-	return harmonyFunctionName
-}
-
-// convertToValidChars converts a user-specified function name to a valid
-// TypeScript identifier.
-//
-// Limitations:
-//
-//   - This doesn't restrict reserved TypeScript keywords.
-//   - We don't perform a real ID_Start/ID_Continue check, and instead use the more
-//     restrictive unicode.IsLetter/unicode.IsDigit check. Unclear what kind of
-//     identifiers these models were trained on, so in the end we might want to
-//     convert unicode-heavy identifiers to their closest ASCII equivalents.
-func (m *FunctionNameMap) convertToValidChars(userFunctionName string) string {
-	mapper := func(r rune) rune {
-		// first, replace certain characters with underscores
-		if r == ' ' || r == '-' || r == '.' {
-			return '_'
-		}
-
-		if unicode.IsLetter(r) || unicode.IsDigit(r) || r == '_' || r == '$' {
-			return r
-		}
-
-		// finally, remove any other characters
-		return -1
-	}
-	candidate := strings.Map(mapper, userFunctionName)
-
-	// set a default name if we end up with nothing left
-	if candidate == "" {
-		return "unnamed"
-	}
-
-	// if the candidate starts with a number, prepend an underscore to make it a
-	// valid identifier
-	if unicode.IsDigit(rune(candidate[0])) {
-		candidate = "_" + candidate
-	}
-
-	return candidate
-}
-
-func (m *FunctionNameMap) deriveName(userFunctionName string) string {
-	originalCandidate := m.convertToValidChars(userFunctionName)
-	candidate := originalCandidate
-
-	// Check for dupes, and if so, add a number to the end.
-	// We start at 2 because if we have dupes and the first is never renamed, it
-	// makes sense for them to be named, say, `f`, `f_2`, `f_3`
-	count := 2
-	for {
-		if _, exists := m.harmonyToUser[candidate]; !exists {
-			break
-		}
-		candidate = fmt.Sprintf("%s_%d", originalCandidate, count)
-		count++
-	}
-
-	return candidate
-}

server/harmonyparser_test.go

@@ -1,4 +1,4 @@
-package harmony
+package server
 
 import (
 	"fmt"
@@ -467,71 +467,3 @@ func TestHarmonyParserStreaming(t *testing.T) {
 		})
 	}
 }
-
-// TestFunctionConvertToValidChars tests only FunctionNameMap.convert(), which doesn't
-// handle any saving (and therefore no dupe handling)
-func TestFunctionConvertToValidChars(t *testing.T) {
-	tests := []struct {
-		name string
-		in   string
-		want string
-	}{
-		{name: "replace spaces with underscores", in: "get weather", want: "get_weather"},
-		{name: "replace hyphens with underscores", in: "get-weather", want: "get_weather"},
-		{name: "replace periods with underscores", in: "get.weather", want: "get_weather"},
-		{name: "disallow non-word characters", in: "get weather!", want: "get_weather"},
-		{name: "strip out invalid non-alphanumeric unicode characters", in: "a🫠bc", want: "abc"},
-		{name: "names that only contain invalid characters", in: "🫠", want: "unnamed"},
-		{name: "leading number", in: "123", want: "_123"},
-		{name: "$ allowed", in: "$", want: "$"},
-		// show that we allow weird unicode letter characters, though we might want
-		// to convert them to their closest ASCII equivalents in the future
-		{name: "allow weird unicode letter characters", in: "𝓸𝓵𝓵𝓪𝓶𝓪", want: "𝓸𝓵𝓵𝓪𝓶𝓪"},
-		// names that look like words but are invalid (i.e., not ID_Start/ID_Continue)
-		{name: "disallow non-word characters that look like words", in: "ⓞⓛⓛⓐⓜⓐ123", want: "_123"},
-	}
-
-	for i, tt := range tests {
-		t.Run(tt.name, func(t *testing.T) {
-			parser := NewFunctionNameMap()
-			got := parser.convertToValidChars(tt.in)
-			if got != tt.want {
-				t.Errorf("case %d: got %q, want %q", i, got, tt.want)
-			}
-		})
-	}
-}
-
-func TestFunctionConvertAndAdd(t *testing.T) {
-	// make a fresh map for each test, but within a test use the same map so we can test for dupe handling
-	tests := []struct {
-		name string
-		in   []string
-		want []string
-	}{
-		{name: "basic dupe handling", in: []string{"get weather", "get weather"}, want: []string{"get_weather", "get_weather_2"}},
-		{name: "dupes from different user-specified names", in: []string{"get weather", "get_weather", "get-weather"}, want: []string{"get_weather", "get_weather_2", "get_weather_3"}},
-		{name: "non dupes after dupes", in: []string{"get weather", "get_weather", "get-weather", "something-different"}, want: []string{"get_weather", "get_weather_2", "get_weather_3", "something_different"}},
-		{name: "multiple sets of dupes", in: []string{"a", "a", "b", "a", "a", "b", "a"}, want: []string{"a", "a_2", "b", "a_3", "a_4", "b_2", "a_5"}},
-	}
-
-	for i, tt := range tests {
-		parser := NewFunctionNameMap()
-		t.Run(tt.name, func(t *testing.T) {
-			for j, in := range tt.in {
-				got := parser.ConvertAndAdd(in)
-				want := tt.want[j]
-				if got != want {
-					t.Errorf("case %d: got %q, want %q", i, got, want)
-				}
-				// check that the maps are correct
-				if parser.userToHarmony[in] != want {
-					t.Errorf("case %d: userToHarmony[%q] = %q, want %q", i, in, parser.userToHarmony[in], want)
-				}
-				if parser.harmonyToUser[want] != in {
-					t.Errorf("case %d: harmonyToUser[%q] = %q, want %q", i, want, parser.harmonyToUser[want], in)
-				}
-			}
-		})
-	}
-}

server/routes.go

@@ -32,7 +32,6 @@ import (
 	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/format"
 	"github.com/ollama/ollama/fs/ggml"
-	"github.com/ollama/ollama/harmony"
 	"github.com/ollama/ollama/llm"
 	"github.com/ollama/ollama/logutil"
 	"github.com/ollama/ollama/openai"
@@ -46,18 +45,6 @@ import (
 	"github.com/ollama/ollama/version"
 )
 
-func shouldUseHarmony(model *Model) bool {
-	if slices.Contains([]string{"gptoss", "gpt-oss"}, model.Config.ModelFamily) {
-		// heuristic to check whether the template expects to be parsed via harmony:
-		// search for harmony tags that are nearly always used
-		if model.Template.Contains("<|start|>") && model.Template.Contains("<|end|>") {
-			return true
-		}
-	}
-
-	return false
-}
-
 func experimentEnabled(name string) bool {
 	return slices.Contains(strings.Split(os.Getenv("OLLAMA_EXPERIMENT"), ","), name)
 }
@@ -189,7 +176,7 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 	}
 
 	// expire the runner
-	if req.Prompt == "" && req.KeepAlive != nil && req.KeepAlive.Duration == 0 {
+	if req.Prompt == "" && req.KeepAlive != nil && int(req.KeepAlive.Seconds()) == 0 {
 		s.sched.expireRunner(m)
 
 		c.JSON(http.StatusOK, api.GenerateResponse{
@@ -207,12 +194,12 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 		return
 	}
 
-	useHarmony := shouldUseHarmony(m) && !req.Raw
-	var harmonyMessageHandler *harmony.HarmonyMessageHandler
-	var harmonyToolParser *harmony.HarmonyToolCallAccumulator
+	useHarmony := shouldUseHarmony(*m) && !req.Raw
+	var harmonyMessageHandler *HarmonyMessageHandler
+	var harmonyToolParser *HarmonyToolCallAccumulator
 	if useHarmony {
-		harmonyMessageHandler = harmony.NewHarmonyMessageHandler()
-		harmonyMessageHandler.HarmonyParser.AddImplicitStart()
+		harmonyMessageHandler = NewHarmonyMessageHandler()
+		harmonyMessageHandler.harmonyParser.AddImplicitStart()
 		harmonyToolParser = harmonyMessageHandler.CreateToolParser()
 	}
 
@@ -327,19 +314,6 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 		prompt = b.String()
 	}
 
-	// If debug mode is enabled, return the rendered template instead of calling the model
-	if req.DebugRenderOnly {
-		c.JSON(http.StatusOK, api.DebugTemplateResponse{
-			Model:     req.Model,
-			CreatedAt: time.Now().UTC(),
-			DebugInfo: api.DebugInfo{
-				RenderedTemplate: prompt,
-				ImageCount:       len(images),
-			},
-		})
-		return
-	}
-
 	var thinkingState *thinking.Parser
 	if !useHarmony {
 		openingTag, closingTag := thinking.InferTags(m.Template.Template)
@@ -1544,7 +1518,7 @@ func (s *Server) ChatHandler(c *gin.Context) {
 	}
 
 	// expire the runner
-	if len(req.Messages) == 0 && req.KeepAlive != nil && req.KeepAlive.Duration == 0 {
+	if len(req.Messages) == 0 && req.KeepAlive != nil && int(req.KeepAlive.Seconds()) == 0 {
 		model, err := GetModel(req.Model)
 		if err != nil {
 			switch {
@@ -1616,49 +1590,14 @@ func (s *Server) ChatHandler(c *gin.Context) {
 	}
 	msgs = filterThinkTags(msgs, m)
 
-	var harmonyMessageHandler *harmony.HarmonyMessageHandler
-	var harmonyToolParser *harmony.HarmonyToolCallAccumulator
-
-	useHarmony := shouldUseHarmony(m)
-
-	processedTools := req.Tools
-	if useHarmony {
-		harmonyMessageHandler = harmony.NewHarmonyMessageHandler()
-		var lastMessage *api.Message
-		if len(msgs) > 0 {
-			lastMessage = &msgs[len(msgs)-1]
-		}
-		harmonyMessageHandler.HarmonyParser.AddImplicitStartOrPrefill(lastMessage)
-		harmonyToolParser = harmonyMessageHandler.CreateToolParser()
-
-		// make a copy of tools to pass to the chat prompt. Function names may be
-		// renamed to be valid Harmony function names.
-		processedTools = make([]api.Tool, len(req.Tools))
-		copy(processedTools, req.Tools)
-		for i, tool := range processedTools {
-			processedTools[i].Function.Name = harmonyMessageHandler.FunctionNameMap.ConvertAndAdd(tool.Function.Name)
-		}
-	}
-
-	prompt, images, err := chatPrompt(c.Request.Context(), m, r.Tokenize, opts, msgs, processedTools, req.Think)
+	prompt, images, err := chatPrompt(c.Request.Context(), m, r.Tokenize, opts, msgs, req.Tools, req.Think)
 	if err != nil {
 		slog.Error("chat prompt error", "error", err)
 		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
 		return
 	}
 
-	// If debug mode is enabled, return the rendered template instead of calling the model
-	if req.DebugRenderOnly {
-		c.JSON(http.StatusOK, api.DebugTemplateResponse{
-			Model:     req.Model,
-			CreatedAt: time.Now().UTC(),
-			DebugInfo: api.DebugInfo{
-				RenderedTemplate: prompt,
-				ImageCount:       len(images),
-			},
-		})
-		return
-	}
+	useHarmony := shouldUseHarmony(*m)
 
 	// Validate Think value: string values currently only allowed for gptoss models
 	if req.Think != nil && req.Think.IsString() && !useHarmony {
@@ -1666,6 +1605,19 @@ func (s *Server) ChatHandler(c *gin.Context) {
 		return
 	}
 
+	var harmonyMessageHandler *HarmonyMessageHandler
+	var harmonyToolParser *HarmonyToolCallAccumulator
+
+	if useHarmony {
+		harmonyMessageHandler = NewHarmonyMessageHandler()
+		var lastMessage *api.Message
+		if len(msgs) > 0 {
+			lastMessage = &msgs[len(msgs)-1]
+		}
+		harmonyMessageHandler.harmonyParser.AddImplicitStartOrPrefill(lastMessage)
+		harmonyToolParser = harmonyMessageHandler.CreateToolParser()
+	}
+
 	var thinkingState *thinking.Parser
 	openingTag, closingTag := thinking.InferTags(m.Template.Template)
 	if req.Think != nil && req.Think.Bool() && openingTag != "" && closingTag != "" {
@@ -1673,10 +1625,6 @@ func (s *Server) ChatHandler(c *gin.Context) {
 			OpeningTag: openingTag,
 			ClosingTag: closingTag,
 		}
-
-		if strings.HasSuffix(strings.TrimSpace(prompt), openingTag) {
-			thinkingState.AddContent(openingTag)
-		}
 	}
 
 	var toolParser *tools.Parser
@@ -1722,7 +1670,6 @@ func (s *Server) ChatHandler(c *gin.Context) {
 					toolName, toolContent := harmonyToolParser.Drain()
 					if toolName != nil {
 						*toolName = strings.TrimPrefix(*toolName, "functions.")
-						*toolName = harmonyMessageHandler.FunctionNameMap.OriginalFromConverted(*toolName)
 						var args api.ToolCallFunctionArguments
 						if err := json.Unmarshal([]byte(toolContent), &args); err != nil {
 							errStr := fmt.Sprintf("error parsing tool call: raw='%s', err=%s", toolContent, err.Error())

server/routes_debug_test.go

@@ -1,413 +0,0 @@
-package server
-
-import (
-	"bytes"
-	"encoding/json"
-	"net/http"
-	"testing"
-	"time"
-
-	"github.com/gin-gonic/gin"
-	"github.com/ollama/ollama/api"
-	"github.com/ollama/ollama/discover"
-	"github.com/ollama/ollama/fs/ggml"
-	"github.com/ollama/ollama/llm"
-)
-
-func TestGenerateDebugRenderOnly(t *testing.T) {
-	gin.SetMode(gin.TestMode)
-
-	mock := mockRunner{
-		CompletionResponse: llm.CompletionResponse{
-			Done:               true,
-			DoneReason:         llm.DoneReasonStop,
-			PromptEvalCount:    1,
-			PromptEvalDuration: 1,
-			EvalCount:          1,
-			EvalDuration:       1,
-		},
-	}
-
-	s := Server{
-		sched: &Scheduler{
-			pendingReqCh:  make(chan *LlmRequest, 1),
-			finishedReqCh: make(chan *LlmRequest, 1),
-			expiredCh:     make(chan *runnerRef, 1),
-			unloadedCh:    make(chan any, 1),
-			loaded:        make(map[string]*runnerRef),
-			newServerFn:   newMockServer(&mock),
-			getGpuFn:      discover.GetGPUInfo,
-			getCpuFn:      discover.GetCPUInfo,
-			reschedDelay:  250 * time.Millisecond,
-			loadFn: func(req *LlmRequest, _ *ggml.GGML, _ discover.GpuInfoList, _ bool) bool {
-				// add small delay to simulate loading
-				time.Sleep(time.Millisecond)
-				req.successCh <- &runnerRef{
-					llama: &mock,
-				}
-				return false
-			},
-		},
-	}
-
-	go s.sched.Run(t.Context())
-
-	// Create a test model
-	stream := false
-	_, digest := createBinFile(t, ggml.KV{
-		"general.architecture":          "llama",
-		"llama.block_count":             uint32(1),
-		"llama.context_length":          uint32(8192),
-		"llama.embedding_length":        uint32(4096),
-		"llama.attention.head_count":    uint32(32),
-		"llama.attention.head_count_kv": uint32(8),
-		"tokenizer.ggml.tokens":         []string{""},
-		"tokenizer.ggml.scores":         []float32{0},
-		"tokenizer.ggml.token_type":     []int32{0},
-	}, []*ggml.Tensor{
-		{Name: "token_embd.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
-		{Name: "blk.0.attn_norm.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
-		{Name: "blk.0.ffn_down.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
-		{Name: "blk.0.ffn_gate.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
-		{Name: "blk.0.ffn_up.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
-		{Name: "blk.0.ffn_norm.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
-		{Name: "blk.0.attn_k.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
-		{Name: "blk.0.attn_output.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
-		{Name: "blk.0.attn_q.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
-		{Name: "blk.0.attn_v.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
-		{Name: "output.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
-	})
-
-	w := createRequest(t, s.CreateHandler, api.CreateRequest{
-		Model:    "test-model",
-		Files:    map[string]string{"file.gguf": digest},
-		Template: "{{ .Prompt }}",
-		Stream:   &stream,
-	})
-
-	if w.Code != http.StatusOK {
-		t.Fatalf("expected status 200, got %d", w.Code)
-	}
-
-	tests := []struct {
-		name            string
-		request         api.GenerateRequest
-		expectDebug     bool
-		expectTemplate  string
-		expectNumImages int
-	}{
-		{
-			name: "debug render only enabled",
-			request: api.GenerateRequest{
-				Model:           "test-model",
-				Prompt:          "Hello, world!",
-				DebugRenderOnly: true,
-			},
-			expectDebug:    true,
-			expectTemplate: "Hello, world!",
-		},
-		{
-			name: "debug render only disabled",
-			request: api.GenerateRequest{
-				Model:           "test-model",
-				Prompt:          "Hello, world!",
-				DebugRenderOnly: false,
-			},
-			expectDebug: false,
-		},
-		{
-			name: "debug render only with system prompt",
-			request: api.GenerateRequest{
-				Model:           "test-model",
-				Prompt:          "User question",
-				System:          "You are a helpful assistant",
-				DebugRenderOnly: true,
-			},
-			expectDebug:    true,
-			expectTemplate: "User question",
-		},
-		{
-			name: "debug render only with template",
-			request: api.GenerateRequest{
-				Model:           "test-model",
-				Prompt:          "Hello",
-				Template:        "PROMPT: {{ .Prompt }}",
-				DebugRenderOnly: true,
-			},
-			expectDebug:    true,
-			expectTemplate: "PROMPT: Hello",
-		},
-		{
-			name: "debug render only with images",
-			request: api.GenerateRequest{
-				Model:           "test-model",
-				Prompt:          "Describe this image",
-				Images:          []api.ImageData{[]byte("fake-image-data")},
-				DebugRenderOnly: true,
-			},
-			expectDebug:     true,
-			expectTemplate:  "[img-0]\n\nDescribe this image",
-			expectNumImages: 1,
-		},
-		{
-			name: "debug render only with raw mode",
-			request: api.GenerateRequest{
-				Model:           "test-model",
-				Prompt:          "Raw prompt text",
-				Raw:             true,
-				DebugRenderOnly: true,
-			},
-			expectDebug:    true,
-			expectTemplate: "Raw prompt text",
-		},
-	}
-
-	for _, tt := range tests {
-		// Test both with and without streaming
-		streamValues := []bool{false, true}
-		for _, stream := range streamValues {
-			streamSuffix := ""
-			if stream {
-				streamSuffix = " (streaming)"
-			}
-			t.Run(tt.name+streamSuffix, func(t *testing.T) {
-				req := tt.request
-				req.Stream = &stream
-				w := createRequest(t, s.GenerateHandler, req)
-
-				if tt.expectDebug {
-					if w.Code != http.StatusOK {
-						t.Errorf("expected status %d, got %d, body: %s", http.StatusOK, w.Code, w.Body.String())
-					}
-
-					var response api.DebugTemplateResponse
-					if err := json.Unmarshal(w.Body.Bytes(), &response); err != nil {
-						t.Fatalf("failed to unmarshal response: %v", err)
-					}
-
-					if response.Model != tt.request.Model {
-						t.Errorf("expected model %s, got %s", tt.request.Model, response.Model)
-					}
-
-					if tt.expectTemplate != "" && response.DebugInfo.RenderedTemplate != tt.expectTemplate {
-						t.Errorf("expected template %q, got %q", tt.expectTemplate, response.DebugInfo.RenderedTemplate)
-					}
-
-					if tt.expectNumImages > 0 && response.DebugInfo.ImageCount != tt.expectNumImages {
-						t.Errorf("expected image count %d, got %d", tt.expectNumImages, response.DebugInfo.ImageCount)
-					}
-				} else {
-					// When debug is disabled, it should attempt normal processing
-					if w.Code != http.StatusOK {
-						t.Errorf("expected status %d, got %d", http.StatusOK, w.Code)
-					}
-				}
-			})
-		}
-	}
-}
-
-func TestChatDebugRenderOnly(t *testing.T) {
-	gin.SetMode(gin.TestMode)
-
-	mock := mockRunner{
-		CompletionResponse: llm.CompletionResponse{
-			Done:               true,
-			DoneReason:         llm.DoneReasonStop,
-			PromptEvalCount:    1,
-			PromptEvalDuration: 1,
-			EvalCount:          1,
-			EvalDuration:       1,
-		},
-	}
-
-	s := Server{
-		sched: &Scheduler{
-			pendingReqCh:  make(chan *LlmRequest, 1),
-			finishedReqCh: make(chan *LlmRequest, 1),
-			expiredCh:     make(chan *runnerRef, 1),
-			unloadedCh:    make(chan any, 1),
-			loaded:        make(map[string]*runnerRef),
-			newServerFn:   newMockServer(&mock),
-			getGpuFn:      discover.GetGPUInfo,
-			getCpuFn:      discover.GetCPUInfo,
-			reschedDelay:  250 * time.Millisecond,
-			loadFn: func(req *LlmRequest, _ *ggml.GGML, _ discover.GpuInfoList, _ bool) bool {
-				// add small delay to simulate loading
-				time.Sleep(time.Millisecond)
-				req.successCh <- &runnerRef{
-					llama: &mock,
-				}
-				return false
-			},
-		},
-	}
-
-	go s.sched.Run(t.Context())
-
-	// Create a test model
-	stream := false
-	_, digest := createBinFile(t, ggml.KV{
-		"general.architecture":          "llama",
-		"llama.block_count":             uint32(1),
-		"llama.context_length":          uint32(8192),
-		"llama.embedding_length":        uint32(4096),
-		"llama.attention.head_count":    uint32(32),
-		"llama.attention.head_count_kv": uint32(8),
-		"tokenizer.ggml.tokens":         []string{""},
-		"tokenizer.ggml.scores":         []float32{0},
-		"tokenizer.ggml.token_type":     []int32{0},
-	}, []*ggml.Tensor{
-		{Name: "token_embd.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
-		{Name: "blk.0.attn_norm.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
-		{Name: "blk.0.ffn_down.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
-		{Name: "blk.0.ffn_gate.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
-		{Name: "blk.0.ffn_up.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
-		{Name: "blk.0.ffn_norm.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
-		{Name: "blk.0.attn_k.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
-		{Name: "blk.0.attn_output.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
-		{Name: "blk.0.attn_q.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
-		{Name: "blk.0.attn_v.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
-		{Name: "output.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
-	})
-
-	w := createRequest(t, s.CreateHandler, api.CreateRequest{
-		Model:    "test-model",
-		Files:    map[string]string{"file.gguf": digest},
-		Template: "{{ if .Tools }}{{ .Tools }}{{ end }}{{ range .Messages }}{{ .Role }}: {{ .Content }}\n{{ end }}",
-		Stream:   &stream,
-	})
-
-	if w.Code != http.StatusOK {
-		t.Fatalf("expected status 200, got %d", w.Code)
-	}
-
-	tests := []struct {
-		name            string
-		request         api.ChatRequest
-		expectDebug     bool
-		expectTemplate  string
-		expectNumImages int
-	}{
-		{
-			name: "chat debug render only enabled",
-			request: api.ChatRequest{
-				Model: "test-model",
-				Messages: []api.Message{
-					{Role: "system", Content: "You are a helpful assistant"},
-					{Role: "user", Content: "Hello"},
-				},
-				DebugRenderOnly: true,
-			},
-			expectDebug:    true,
-			expectTemplate: "system: You are a helpful assistant\nuser: Hello\n",
-		},
-		{
-			name: "chat debug render only disabled",
-			request: api.ChatRequest{
-				Model: "test-model",
-				Messages: []api.Message{
-					{Role: "user", Content: "Hello"},
-				},
-				DebugRenderOnly: false,
-			},
-			expectDebug: false,
-		},
-		{
-			name: "chat debug with assistant message",
-			request: api.ChatRequest{
-				Model: "test-model",
-				Messages: []api.Message{
-					{Role: "user", Content: "Hello"},
-					{Role: "assistant", Content: "Hi there!"},
-					{Role: "user", Content: "How are you?"},
-				},
-				DebugRenderOnly: true,
-			},
-			expectDebug:    true,
-			expectTemplate: "user: Hello\nassistant: Hi there!\nuser: How are you?\n",
-		},
-		{
-			name: "chat debug with images",
-			request: api.ChatRequest{
-				Model: "test-model",
-				Messages: []api.Message{
-					{
-						Role:    "user",
-						Content: "What's in this image?",
-						Images:  []api.ImageData{[]byte("fake-image-data")},
-					},
-				},
-				DebugRenderOnly: true,
-			},
-			expectDebug:     true,
-			expectTemplate:  "user: [img-0]What's in this image?\n",
-			expectNumImages: 1,
-		},
-		{
-			name: "chat debug with tools",
-			request: api.ChatRequest{
-				Model: "test-model",
-				Messages: []api.Message{
-					{Role: "user", Content: "Get the weather"},
-				},
-				Tools: api.Tools{
-					{
-						Type: "function",
-						Function: api.ToolFunction{
-							Name:        "get_weather",
-							Description: "Get weather information",
-						},
-					},
-				},
-				DebugRenderOnly: true,
-			},
-			expectDebug:    true,
-			expectTemplate: "[{\"type\":\"function\",\"function\":{\"name\":\"get_weather\",\"description\":\"Get weather information\",\"parameters\":{\"type\":\"\",\"required\":null,\"properties\":null}}}]user: Get the weather\n",
-		},
-	}
-
-	for _, tt := range tests {
-		// Test both with and without streaming
-		streamValues := []bool{false, true}
-		for _, stream := range streamValues {
-			streamSuffix := ""
-			if stream {
-				streamSuffix = " (streaming)"
-			}
-			t.Run(tt.name+streamSuffix, func(t *testing.T) {
-				req := tt.request
-				req.Stream = &stream
-				w := createRequest(t, s.ChatHandler, req)
-
-				if tt.expectDebug {
-					if w.Code != http.StatusOK {
-						t.Errorf("expected status %d, got %d, body: %s", http.StatusOK, w.Code, w.Body.String())
-					}
-
-					var response api.DebugTemplateResponse
-					if err := json.Unmarshal(w.Body.Bytes(), &response); err != nil {
-						t.Fatalf("failed to unmarshal response: %v", err)
-					}
-
-					if response.Model != tt.request.Model {
-						t.Errorf("expected model %s, got %s", tt.request.Model, response.Model)
-					}
-
-					if tt.expectTemplate != "" && response.DebugInfo.RenderedTemplate != tt.expectTemplate {
-						t.Errorf("expected template %q, got %q", tt.expectTemplate, response.DebugInfo.RenderedTemplate)
-					}
-
-					if tt.expectNumImages > 0 && response.DebugInfo.ImageCount != tt.expectNumImages {
-						t.Errorf("expected image count %d, got %d", tt.expectNumImages, response.DebugInfo.ImageCount)
-					}
-				} else {
-					// When debug is disabled, it should attempt normal processing
-					if w.Code != http.StatusOK {
-						t.Errorf("expected status %d, got %d", http.StatusOK, w.Code)
-					}
-				}
-			})
-		}
-	}
-}

server/routes_generate_test.go

@@ -969,233 +969,3 @@ func TestGenerate(t *testing.T) {
 		}
 	})
 }
-
-func TestChatWithPromptEndingInThinkTag(t *testing.T) {
-	gin.SetMode(gin.TestMode)
-
-	// Helper to create a standard thinking test setup
-	setupThinkingTest := func(t *testing.T) (*mockRunner, *Server) {
-		mock := &mockRunner{
-			CompletionResponse: llm.CompletionResponse{
-				Done:               true,
-				DoneReason:         llm.DoneReasonStop,
-				PromptEvalCount:    1,
-				PromptEvalDuration: 1,
-				EvalCount:          1,
-				EvalDuration:       1,
-			},
-		}
-
-		s := &Server{
-			sched: &Scheduler{
-				pendingReqCh:  make(chan *LlmRequest, 1),
-				finishedReqCh: make(chan *LlmRequest, 1),
-				expiredCh:     make(chan *runnerRef, 1),
-				unloadedCh:    make(chan any, 1),
-				loaded:        make(map[string]*runnerRef),
-				newServerFn:   newMockServer(mock),
-				getGpuFn:      discover.GetGPUInfo,
-				getCpuFn:      discover.GetCPUInfo,
-				reschedDelay:  250 * time.Millisecond,
-				loadFn: func(req *LlmRequest, _ *ggml.GGML, _ discover.GpuInfoList, _ bool) bool {
-					time.Sleep(time.Millisecond)
-					req.successCh <- &runnerRef{llama: mock}
-					return false
-				},
-			},
-		}
-
-		go s.sched.Run(t.Context())
-
-		// Create a model with thinking support
-		_, digest := createBinFile(t, ggml.KV{
-			"general.architecture":          "llama",
-			"llama.block_count":             uint32(1),
-			"llama.context_length":          uint32(8192),
-			"llama.embedding_length":        uint32(4096),
-			"llama.attention.head_count":    uint32(32),
-			"llama.attention.head_count_kv": uint32(8),
-			"tokenizer.ggml.tokens":         []string{""},
-			"tokenizer.ggml.scores":         []float32{0},
-			"tokenizer.ggml.token_type":     []int32{0},
-		}, []*ggml.Tensor{
-			{Name: "token_embd.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
-			{Name: "blk.0.attn_norm.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
-			{Name: "blk.0.ffn_down.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
-			{Name: "blk.0.ffn_gate.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
-			{Name: "blk.0.ffn_up.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
-			{Name: "blk.0.ffn_norm.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
-			{Name: "blk.0.attn_k.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
-			{Name: "blk.0.attn_output.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
-			{Name: "blk.0.attn_q.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
-			{Name: "blk.0.attn_v.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
-			{Name: "output.weight", Shape: []uint64{1}, WriterTo: bytes.NewReader(make([]byte, 4))},
-		})
-
-		// Create model with thinking template that adds <think> at the end
-		w := createRequest(t, s.CreateHandler, api.CreateRequest{
-			Model: "test-thinking",
-			Files: map[string]string{"file.gguf": digest},
-			Template: `{{- range .Messages }}
-{{- if eq .Role "user" }}user: {{ .Content }}
-{{ else if eq .Role "assistant" }}assistant: {{ if .Thinking }}<think>{{ .Thinking }}</think>{{ end }}{{ .Content }}
-{{ end }}{{ end }}<think>`,
-			Stream: &stream,
-		})
-
-		if w.Code != http.StatusOK {
-			t.Fatalf("expected status 200, got %d", w.Code)
-		}
-
-		return mock, s
-	}
-
-	mock, s := setupThinkingTest(t)
-
-	// Helper to test chat responses
-	testChatRequest := func(t *testing.T, name string, userContent string, modelResponse string, expectedThinking string, expectedContent string, think bool) {
-		t.Run(name, func(t *testing.T) {
-			mock.CompletionResponse = llm.CompletionResponse{
-				Content:            modelResponse,
-				Done:               true,
-				DoneReason:         llm.DoneReasonStop,
-				PromptEvalCount:    1,
-				PromptEvalDuration: 1,
-				EvalCount:          1,
-				EvalDuration:       1,
-			}
-			mock.CompletionFn = nil
-
-			streamRequest := false
-			req := api.ChatRequest{
-				Model: "test-thinking",
-				Messages: []api.Message{
-					{Role: "user", Content: userContent},
-				},
-				Stream: &streamRequest,
-			}
-			if think {
-				req.Think = &api.ThinkValue{Value: think}
-			}
-
-			w := createRequest(t, s.ChatHandler, req)
-			if w.Code != http.StatusOK {
-				t.Fatalf("expected status 200, got %d", w.Code)
-			}
-
-			var resp api.ChatResponse
-			if err := json.NewDecoder(w.Body).Decode(&resp); err != nil {
-				t.Fatal(err)
-			}
-
-			if resp.Message.Thinking != expectedThinking {
-				t.Errorf("expected thinking %q, got %q", expectedThinking, resp.Message.Thinking)
-			}
-
-			if resp.Message.Content != expectedContent {
-				t.Errorf("expected content %q, got %q", expectedContent, resp.Message.Content)
-			}
-		})
-	}
-
-	// Test cases - Note: Template adds <think> at the end, and leading whitespace after <think> is eaten by the parser
-	testChatRequest(t, "basic thinking response",
-		"Help me solve this problem",
-		" Let me think about this step by step... </think> The answer is 42.",
-		"Let me think about this step by step... ",
-		"The answer is 42.",
-		true)
-
-	testChatRequest(t, "thinking with multiple sentences",
-		"Explain quantum computing",
-		" First, I need to understand the basics. Quantum bits can be in superposition. </think> Quantum computing uses quantum mechanics principles.",
-		"First, I need to understand the basics. Quantum bits can be in superposition. ",
-		"Quantum computing uses quantum mechanics principles.",
-		true)
-
-	testChatRequest(t, "no thinking content",
-		"What is 2+2?",
-		"</think> The answer is 4.",
-		"",
-		"The answer is 4.",
-		true)
-
-	testChatRequest(t, "thinking disabled but template still adds think tag",
-		"Simple question",
-		" My thoughts </think> The answer.",
-		"",
-		" My thoughts </think> The answer.",
-		false)
-
-	// Test streaming response with template-added <think>
-	t.Run("streaming with thinking", func(t *testing.T) {
-		var wg sync.WaitGroup
-		wg.Add(1)
-
-		mock.CompletionFn = func(ctx context.Context, r llm.CompletionRequest, fn func(r llm.CompletionResponse)) error {
-			defer wg.Done()
-
-			// Verify the prompt ends with <think> due to template
-			if !strings.HasSuffix(r.Prompt, "<think>") {
-				t.Errorf("expected prompt to end with <think>, got: %q", r.Prompt)
-			}
-
-			// Simulate streaming chunks
-			responses := []llm.CompletionResponse{
-				{Content: " I need to consider", Done: false, PromptEvalCount: 1, PromptEvalDuration: 1},
-				{Content: " multiple factors here...", Done: false, PromptEvalCount: 1, PromptEvalDuration: 1},
-				{Content: " </think> Based on my analysis,", Done: false, PromptEvalCount: 1, PromptEvalDuration: 1},
-				{Content: " the solution is straightforward.", Done: true, DoneReason: llm.DoneReasonStop, PromptEvalCount: 1, PromptEvalDuration: 1, EvalCount: 1, EvalDuration: 1},
-			}
-
-			for _, resp := range responses {
-				select {
-				case <-ctx.Done():
-					return ctx.Err()
-				default:
-					fn(resp)
-					time.Sleep(10 * time.Millisecond)
-				}
-			}
-			return nil
-		}
-
-		think := true
-		w := createRequest(t, s.ChatHandler, api.ChatRequest{
-			Model:    "test-thinking",
-			Messages: []api.Message{{Role: "user", Content: "Analyze this complex problem"}},
-			Think:    &api.ThinkValue{Value: think},
-			Stream:   &stream,
-		})
-
-		wg.Wait()
-
-		if w.Code != http.StatusOK {
-			t.Fatalf("expected status 200, got %d", w.Code)
-		}
-
-		// Parse streaming responses
-		decoder := json.NewDecoder(w.Body)
-		var allThinking, allContent strings.Builder
-
-		for {
-			var resp api.ChatResponse
-			if err := decoder.Decode(&resp); err == io.EOF {
-				break
-			} else if err != nil {
-				t.Fatal(err)
-			}
-			allThinking.WriteString(resp.Message.Thinking)
-			allContent.WriteString(resp.Message.Content)
-		}
-
-		// Note: Leading whitespace after <think> is eaten by the parser
-		if got := allThinking.String(); got != "I need to consider multiple factors here... " {
-			t.Errorf("expected thinking %q, got %q", "I need to consider multiple factors here... ", got)
-		}
-
-		if got := allContent.String(); got != "Based on my analysis, the solution is straightforward." {
-			t.Errorf("expected content %q, got %q", "Based on my analysis, the solution is straightforward.", got)
-		}
-	})
-}

thinking/parser.go

@@ -103,9 +103,7 @@ func eat(s *Parser) (string, string, bool) {
 			// note that we use the original content, not the trimmed one because we
 			// don't want to eat any whitespace in the real content if there were no
 			// thinking tags
-			untrimmed := s.acc.String()
-			s.acc.Reset()
-			return "", untrimmed, false
+			return "", s.acc.String(), false
 		}
 	case thinkingState_ThinkingStartedEatingWhitespace:
 		trimmed := strings.TrimLeftFunc(s.acc.String(), unicode.IsSpace)

thinking/parser_test.go

@@ -58,15 +58,6 @@ func TestThinkingStreaming(t *testing.T) {
 					wantContent:    "  abc",
 					wantStateAfter: thinkingState_ThinkingDone,
 				},
-				// regression test for a bug where we were transitioning directly to
-				// ThinkingDone without clearing the buffer. This would cuase the first
-				// step to be outputted twice
-				{
-					input:          "def",
-					wantThinking:   "",
-					wantContent:    "def",
-					wantStateAfter: thinkingState_ThinkingDone,
-				},
 			},
 		},
 		{

tools/tools.go

@@ -224,45 +224,22 @@ func findArguments(buffer []byte) (map[string]any, int) {
 		return nil, 0
 	}
 
-	start := -1
 	var braces int
-	var inString, escaped bool
-
-	for i := range buffer {
-		c := buffer[i]
-
-		if escaped {
-			escaped = false
-			continue
-		}
-
-		if c == '\\' {
-			escaped = true
-			continue
-		}
-
-		if c == '"' {
-			inString = !inString
-			continue
-		}
-
-		if inString {
-			continue
-		}
+	var start int = -1
 
+	for i, c := range buffer {
 		if c == '{' {
 			if braces == 0 {
 				start = i
 			}
 			braces++
-		} else if c == '}' {
+		} else if c == '}' && braces > 0 {
 			braces--
 			if braces == 0 && start != -1 {
 				object := buffer[start : i+1]
 
 				var data map[string]any
 				if err := json.Unmarshal(object, &data); err != nil {
-					// not a valid object, keep looking
 					start = -1
 					continue
 				}
@@ -305,10 +282,6 @@ func findArguments(buffer []byte) (map[string]any, int) {
 
 				return data, i
 			}
-
-			if braces < 0 {
-				braces = 0
-			}
 		}
 	}
 

tools/tools_test.go

@@ -1,7 +1,6 @@
 package tools
 
 import (
-	"strings"
 	"testing"
 	"text/template"
 
@@ -41,7 +40,13 @@ func TestParser(t *testing.T) {
 			Function: api.ToolFunction{
 				Name:        "get_temperature",
 				Description: "Retrieve the temperature for a given location",
-				Parameters: api.ToolFunctionParameters{
+				Parameters: struct {
+					Type       string                      `json:"type"`
+					Defs       any                         `json:"$defs,omitempty"`
+					Items      any                         `json:"items,omitempty"`
+					Required   []string                    `json:"required"`
+					Properties map[string]api.ToolProperty `json:"properties"`
+				}{
 					Type:     "object",
 					Required: []string{"city"},
 					Properties: map[string]api.ToolProperty{
@@ -63,7 +68,13 @@ func TestParser(t *testing.T) {
 			Function: api.ToolFunction{
 				Name:        "get_conditions",
 				Description: "Retrieve the current weather conditions for a given location",
-				Parameters: api.ToolFunctionParameters{
+				Parameters: struct {
+					Type       string                      `json:"type"`
+					Defs       any                         `json:"$defs,omitempty"`
+					Items      any                         `json:"items,omitempty"`
+					Required   []string                    `json:"required"`
+					Properties map[string]api.ToolProperty `json:"properties"`
+				}{
 					Type: "object",
 					Properties: map[string]api.ToolProperty{
 						"location": {
@@ -93,7 +104,13 @@ func TestParser(t *testing.T) {
 			Function: api.ToolFunction{
 				Name:        "get_address",
 				Description: "Get the address of a given location",
-				Parameters: api.ToolFunctionParameters{
+				Parameters: struct {
+					Type       string                      `json:"type"`
+					Defs       any                         `json:"$defs,omitempty"`
+					Items      any                         `json:"items,omitempty"`
+					Required   []string                    `json:"required"`
+					Properties map[string]api.ToolProperty `json:"properties"`
+				}{
 					Type: "object",
 					Properties: map[string]api.ToolProperty{
 						"location": {
@@ -109,7 +126,13 @@ func TestParser(t *testing.T) {
 			Function: api.ToolFunction{
 				Name:        "add",
 				Description: "Add two numbers",
-				Parameters: api.ToolFunctionParameters{
+				Parameters: struct {
+					Type       string                      `json:"type"`
+					Defs       any                         `json:"$defs,omitempty"`
+					Items      any                         `json:"items,omitempty"`
+					Required   []string                    `json:"required"`
+					Properties map[string]api.ToolProperty `json:"properties"`
+				}{
 					Type: "object",
 					Properties: map[string]api.ToolProperty{
 						"a": {
@@ -1117,163 +1140,11 @@ func TestFindArguments(t *testing.T) {
 		},
 		{
 			name:   "deepseek",
-			buffer: []byte(`"arguments": {"location": "Tokyo"}}</tool_call>`),
+			buffer: []byte(`", "arguments": {"location": "Tokyo"}}</tool_call>`),
 			want: map[string]any{
 				"location": "Tokyo",
 			},
 		},
-		{
-			name:   "string with braces",
-			buffer: []byte(`{"name": "process_code", "arguments": {"code": "if (x > 0) { return true; }"}}`),
-			want: map[string]any{
-				"code": "if (x > 0) { return true; }",
-			},
-		},
-		{
-			name:   "string with nested json",
-			buffer: []byte(`{"name": "send_data", "arguments": {"payload": "{\"nested\": {\"key\": \"value\"}}"}}`),
-			want: map[string]any{
-				"payload": `{"nested": {"key": "value"}}`,
-			},
-		},
-		{
-			name:   "string with escaped quotes and braces",
-			buffer: []byte(`{"name": "analyze", "arguments": {"text": "The JSON is: {\"key\": \"val{ue}\"}"}}`),
-			want: map[string]any{
-				"text": `The JSON is: {"key": "val{ue}"}`,
-			},
-		},
-		{
-			name:   "multiple objects with string containing braces",
-			buffer: []byte(`{"name": "test", "arguments": {"query": "find } in text"}} {"name": "other"}`),
-			want: map[string]any{
-				"query": "find } in text",
-			},
-		},
-		{
-			name:   "unmatched closing brace in string",
-			buffer: []byte(`{"name": "search", "arguments": {"pattern": "regex: }"}}`),
-			want: map[string]any{
-				"pattern": "regex: }",
-			},
-		},
-		{
-			name:   "complex nested with mixed braces",
-			buffer: []byte(`{"name": "analyze", "arguments": {"data": "{\"items\": [{\"value\": \"}\"}, {\"code\": \"if (x) { return y; }\"}]}"}}`),
-			want: map[string]any{
-				"data": `{"items": [{"value": "}"}, {"code": "if (x) { return y; }"}]}`,
-			},
-		},
-		{
-			name:   "string with newline and braces",
-			buffer: []byte(`{"name": "format", "arguments": {"template": "{\n  \"key\": \"value\"\n}"}}`),
-			want: map[string]any{
-				"template": "{\n  \"key\": \"value\"\n}",
-			},
-		},
-		{
-			name:   "string with unicode escape",
-			buffer: []byte(`{"name": "test", "arguments": {"text": "Unicode: \u007B and \u007D"}}`),
-			want: map[string]any{
-				"text": "Unicode: { and }",
-			},
-		},
-		{
-			name:   "array arguments",
-			buffer: []byte(`{"name": "batch", "arguments": ["item1", "item2", "{\"nested\": true}"]}`),
-			want:   nil, // This should return nil because arguments is not a map
-		},
-		{
-			name:   "escaped backslash before quote",
-			buffer: []byte(`{"name": "path", "arguments": {"dir": "C:\\Program Files\\{App}\\"}}`),
-			want: map[string]any{
-				"dir": `C:\Program Files\{App}\`,
-			},
-		},
-		{
-			name:   "single quotes not treated as string delimiters",
-			buffer: []byte(`{"name": "query", "arguments": {"sql": "SELECT * FROM users WHERE name = '{admin}'"}}`),
-			want: map[string]any{
-				"sql": "SELECT * FROM users WHERE name = '{admin}'",
-			},
-		},
-		{
-			name:   "incomplete json at buffer end",
-			buffer: []byte(`{"name": "test", "arguments": {"data": "some {"`),
-			want:   nil,
-		},
-		{
-			name:   "multiple escaped quotes",
-			buffer: []byte(`{"name": "echo", "arguments": {"msg": "He said \"Hello {World}\" loudly"}}`),
-			want: map[string]any{
-				"msg": `He said "Hello {World}" loudly`,
-			},
-		},
-		{
-			name:   "json with comments style string",
-			buffer: []byte(`{"name": "code", "arguments": {"snippet": "// This is a comment with { and }"}}`),
-			want: map[string]any{
-				"snippet": "// This is a comment with { and }",
-			},
-		},
-		{
-			name:   "consecutive escaped backslashes",
-			buffer: []byte(`{"name": "test", "arguments": {"path": "C:\\\\{folder}\\\\"}}`),
-			want: map[string]any{
-				"path": `C:\\{folder}\\`,
-			},
-		},
-		{
-			name:   "empty string with braces after",
-			buffer: []byte(`{"name": "test", "arguments": {"a": "", "b": "{value}"}}`),
-			want: map[string]any{
-				"a": "",
-				"b": "{value}",
-			},
-		},
-		{
-			name:   "unicode in key names",
-			buffer: []byte(`{"name": "test", "arguments": {"key{": "value", "key}": "value2"}}`),
-			want: map[string]any{
-				"key{": "value",
-				"key}": "value2",
-			},
-		},
-		{
-			name:   "very long string with braces",
-			buffer: []byte(`{"name": "test", "arguments": {"data": "` + strings.Repeat("a{b}c", 100) + `"}}`),
-			want: map[string]any{
-				"data": strings.Repeat("a{b}c", 100),
-			},
-		},
-		{
-			name:   "tab characters and braces",
-			buffer: []byte(`{"name": "test", "arguments": {"code": "\tif (true) {\n\t\treturn;\n\t}"}}`),
-			want: map[string]any{
-				"code": "\tif (true) {\n\t\treturn;\n\t}",
-			},
-		},
-		{
-			name:   "null byte in string",
-			buffer: []byte(`{"name": "test", "arguments": {"data": "before\u0000{after}"}}`),
-			want: map[string]any{
-				"data": "before\x00{after}",
-			},
-		},
-		{
-			name:   "escaped quote at end of string",
-			buffer: []byte(`{"name": "test", "arguments": {"data": "text with quote at end\\\""}}`),
-			want: map[string]any{
-				"data": `text with quote at end\"`,
-			},
-		},
-		{
-			name:   "mixed array and object in arguments",
-			buffer: []byte(`{"name": "test", "arguments": {"items": ["{", "}", {"key": "value"}]}}`),
-			want: map[string]any{
-				"items": []any{"{", "}", map[string]any{"key": "value"}},
-			},
-		},
 	}
 
 	for _, tt := range tests {