Merge branch 'ollama:main' into main

Dockerfile

@@ -104,8 +104,8 @@ COPY --from=cuda-12 dist/lib/ollama/cuda_v12 /lib/ollama/cuda_v12
 FROM --platform=linux/arm64 scratch AS arm64
 COPY --from=cuda-11 dist/lib/ollama/cuda_v11 /lib/ollama/cuda_v11
 COPY --from=cuda-12 dist/lib/ollama/cuda_v12 /lib/ollama/cuda_v12
-COPY --from=jetpack-5 dist/lib/ollama/cuda_v11 lib/ollama/cuda_jetpack5
+COPY --from=jetpack-5 dist/lib/ollama/cuda_v11 /lib/ollama/cuda_jetpack5
-COPY --from=jetpack-6 dist/lib/ollama/cuda_v12 lib/ollama/cuda_jetpack6
+COPY --from=jetpack-6 dist/lib/ollama/cuda_v12 /lib/ollama/cuda_jetpack6
 
 FROM scratch AS rocm
 COPY --from=rocm-6 dist/lib/ollama/rocm /lib/ollama/rocm

README.md

@@ -313,7 +313,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [Typescript UI](https://github.com/ollama-interface/Ollama-Gui?tab=readme-ov-file)
 - [Minimalistic React UI for Ollama Models](https://github.com/richawo/minimal-llm-ui)
 - [Ollamac](https://github.com/kevinhermawan/Ollamac)
-- [big-AGI](https://github.com/enricoros/big-AGI/blob/main/docs/config-local-ollama.md)
+- [big-AGI](https://github.com/enricoros/big-AGI) 
 - [Cheshire Cat assistant framework](https://github.com/cheshire-cat-ai/core)
 - [Amica](https://github.com/semperai/amica)
 - [chatd](https://github.com/BruceMacD/chatd)

api/types.go

@@ -163,6 +163,7 @@ func (t *ToolCallFunctionArguments) String() string {
 
 type Tool struct {
 	Type     string       `json:"type"`
+	Items    any          `json:"items,omitempty"`
 	Function ToolFunction `json:"function"`
 }
 
@@ -213,11 +214,14 @@ type ToolFunction struct {
 	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]struct {
 			Type        PropertyType `json:"type"`
+			Items       any          `json:"items,omitempty"`
 			Description string       `json:"description"`
-			Enum        []string     `json:"enum,omitempty"`
+			Enum        []any        `json:"enum,omitempty"`
 		} `json:"properties"`
 	} `json:"parameters"`
 }

api/types_test.go

@@ -232,6 +232,67 @@ func TestMessage_UnmarshalJSON(t *testing.T) {
 	}
 }
 
+func TestToolFunction_UnmarshalJSON(t *testing.T) {
+	tests := []struct {
+		name    string
+		input   string
+		wantErr string
+	}{
+		{
+			name: "valid enum with same types",
+			input: `{
+				"name": "test",
+				"description": "test function",
+				"parameters": {
+					"type": "object",
+					"required": ["test"],
+					"properties": {
+						"test": {
+							"type": "string",
+							"description": "test prop",
+							"enum": ["a", "b", "c"]
+						}
+					}
+				}
+			}`,
+			wantErr: "",
+		},
+		{
+			name: "empty enum array",
+			input: `{
+				"name": "test",
+				"description": "test function",
+				"parameters": {
+					"type": "object",
+					"required": ["test"],
+					"properties": {
+						"test": {
+							"type": "string",
+							"description": "test prop",
+							"enum": []
+						}
+					}
+				}
+			}`,
+			wantErr: "",
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			var tf ToolFunction
+			err := json.Unmarshal([]byte(tt.input), &tf)
+
+			if tt.wantErr != "" {
+				require.Error(t, err)
+				assert.Contains(t, err.Error(), tt.wantErr)
+			} else {
+				require.NoError(t, err)
+			}
+		})
+	}
+}
+
 func TestPropertyType_UnmarshalJSON(t *testing.T) {
 	tests := []struct {
 		name     string

cmd/cmd.go

@@ -1381,7 +1381,6 @@ func NewCLI() *cobra.Command {
 				envVars["OLLAMA_NOPRUNE"],
 				envVars["OLLAMA_ORIGINS"],
 				envVars["OLLAMA_SCHED_SPREAD"],
-				envVars["OLLAMA_TMPDIR"],
 				envVars["OLLAMA_FLASH_ATTENTION"],
 				envVars["OLLAMA_KV_CACHE_TYPE"],
 				envVars["OLLAMA_LLM_LIBRARY"],

docs/troubleshooting.md

@@ -26,7 +26,6 @@ When you run Ollama on **Windows**, there are a few different locations. You can
 - `explorer %LOCALAPPDATA%\Ollama` to view logs.  The most recent server logs will be in `server.log` and older logs will be in `server-#.log` 
 - `explorer %LOCALAPPDATA%\Programs\Ollama` to browse the binaries (The installer adds this to your user PATH)
 - `explorer %HOMEPATH%\.ollama` to browse where models and configuration is stored
-- `explorer %TEMP%` where temporary executable files are stored in one or more `ollama*` directories
 
 To enable additional debug logging to help troubleshoot problems, first **Quit the running app from the tray menu** then in a powershell terminal
 
@@ -69,10 +68,6 @@ If you run into problems on Linux and want to install an older version, or you'd
 curl -fsSL https://ollama.com/install.sh | OLLAMA_VERSION=0.5.7 sh
 ```
 
-## Linux tmp noexec 
-
-If your system is configured with the "noexec" flag where Ollama stores its temporary executable files, you can specify an alternate location by setting OLLAMA_TMPDIR to a location writable by the user ollama runs as. For example OLLAMA_TMPDIR=/usr/share/ollama/
-
 ## Linux docker
 
 If Ollama initially works on the GPU in a docker container, but then switches to running on CPU after some period of time with errors in the server log reporting GPU discovery failures, this can be resolved by disabling systemd cgroup management in Docker.  Edit `/etc/docker/daemon.json` on the host and add `"exec-opts": ["native.cgroupdriver=cgroupfs"]` to the docker configuration.

docs/windows.md

@@ -62,7 +62,6 @@ the explorer window by hitting `<Ctrl>+R` and type in:
     - *upgrade.log* contains log output for upgrades
 - `explorer %LOCALAPPDATA%\Programs\Ollama` contains the binaries (The installer adds this to your user PATH)
 - `explorer %HOMEPATH%\.ollama` contains models and configuration
-- `explorer %TEMP%` contains temporary executable files in one or more `ollama*` directories
 
 ## Uninstall
 

integration/max_queue_test.go

@@ -52,8 +52,8 @@ func TestMaxQueue(t *testing.T) {
 	embedCtx := ctx
 
 	var genwg sync.WaitGroup
+	genwg.Add(1)
 	go func() {
-		genwg.Add(1)
 		defer genwg.Done()
 		slog.Info("Starting generate request")
 		DoGenerate(ctx, t, client, req, resp, 45*time.Second, 5*time.Second)
@@ -71,8 +71,8 @@ func TestMaxQueue(t *testing.T) {
 	counterMu := sync.Mutex{}
 	var embedwg sync.WaitGroup
 	for i := 0; i < threadCount; i++ {
+		embedwg.Add(1)
 		go func(i int) {
-			embedwg.Add(1)
 			defer embedwg.Done()
 			slog.Info("embed started", "id", i)
 			embedReq := api.EmbeddingRequest{

kvcache/cache.go

@@ -56,8 +56,9 @@ type Cache interface {
 
 	// StartForward is called before the start of the model's forward pass.
 	// For each token in the coming batch, there must be a corresponding
-	// entry in positions and seqs.
+	// entry in positions and seqs. reserve is to preallocate memory
-	StartForward(ctx ml.Context, batch input.Batch) error
+	// without actually storing data in the cache.
+	StartForward(ctx ml.Context, batch input.Batch, reserve bool) error
 
 	// CopyPrefix copies tokens in the range [0, len) from srcSeq to dstSeq
 	CopyPrefix(srcSeq, dstSeq int, len int32)

kvcache/causal.go

@@ -146,51 +146,60 @@ func (c *Causal) Close() {
 	}
 }
 
-func (c *Causal) StartForward(ctx ml.Context, batch input.Batch) error {
+func (c *Causal) StartForward(ctx ml.Context, batch input.Batch, reserve bool) error {
 	c.curBatchSize = len(batch.Positions)
 	c.curSequences = batch.Sequences
 	c.curPositions = batch.Positions
 	c.opts.Except = nil
 
-	c.updateSlidingWindow()
+	if !reserve {
+		c.updateSlidingWindow()
+
+		var err error
+		c.curLoc, err = c.findStartLoc()
+		if errors.Is(err, ErrKvCacheFull) {
+			c.defrag()
+			c.curLoc, err = c.findStartLoc()
+		}
+		if err != nil {
+			return err
+		}
+
+		c.curCellRange = newRange()
+		for i, pos := range batch.Positions {
+			seq := batch.Sequences[i]
+
+			c.cells[c.curLoc+i] = cacheCell{pos: pos, sequences: []int{seq}}
+
+			seqRange, ok := c.cellRanges[seq]
+			if !ok {
+				seqRange = newRange()
+			}
+
+			if c.curLoc+i > seqRange.max {
+				seqRange.max = c.curLoc + i
+			}
+			if seqRange.max > c.curCellRange.max {
+				c.curCellRange.max = seqRange.max
+			}
+
+			if c.curLoc+i < seqRange.min {
+				seqRange.min = c.curLoc + i
+			}
+			if seqRange.min < c.curCellRange.min {
+				c.curCellRange.min = seqRange.min
+			}
+			c.cellRanges[seq] = seqRange
+		}
+	} else {
+		// If we are reserving memory, don't update any of the cache metadata but set the size
+		// to the worst case.
+		c.curLoc = 0
+		c.curCellRange.min = 0
+		c.curCellRange.max = len(c.cells) - 1
+	}
 
 	var err error
-	c.curLoc, err = c.findStartLoc()
-	if errors.Is(err, ErrKvCacheFull) {
-		c.defrag()
-		c.curLoc, err = c.findStartLoc()
-	}
-	if err != nil {
-		return err
-	}
-
-	c.curCellRange = newRange()
-	for i, pos := range batch.Positions {
-		seq := batch.Sequences[i]
-
-		c.cells[c.curLoc+i] = cacheCell{pos: pos, sequences: []int{seq}}
-
-		seqRange, ok := c.cellRanges[seq]
-		if !ok {
-			seqRange = newRange()
-		}
-
-		if c.curLoc+i > seqRange.max {
-			seqRange.max = c.curLoc + i
-		}
-		if seqRange.max > c.curCellRange.max {
-			c.curCellRange.max = seqRange.max
-		}
-
-		if c.curLoc+i < seqRange.min {
-			seqRange.min = c.curLoc + i
-		}
-		if seqRange.min < c.curCellRange.min {
-			c.curCellRange.min = seqRange.min
-		}
-		c.cellRanges[seq] = seqRange
-	}
-
 	c.curMask, err = c.buildMask(ctx)
 
 	return err

kvcache/causal_test.go

@@ -5,7 +5,6 @@ import (
 	"slices"
 	"testing"
 
-	"github.com/ollama/ollama/fs"
 	"github.com/ollama/ollama/ml"
 	"github.com/ollama/ollama/model/input"
 )
@@ -281,7 +280,7 @@ func testCache(t *testing.T, backend ml.Backend, cache Cache, tests []testCase)
 			context := backend.NewContext()
 			defer context.Close()
 
-			err := cache.StartForward(context, input.Batch{Positions: test.pos, Sequences: test.seqs})
+			err := cache.StartForward(context, input.Batch{Positions: test.pos, Sequences: test.seqs}, false)
 			if err != nil {
 				panic(err)
 			}
@@ -315,7 +314,7 @@ func TestCanResume(t *testing.T) {
 	err := cache.StartForward(context, input.Batch{
 		Positions: []int32{0, 1, 2, 3},
 		Sequences: []int{0, 0, 0, 0},
-	})
+	}, false)
 	if err != nil {
 		t.Fatalf("StartForward failed: %v", err)
 	}
@@ -342,7 +341,7 @@ func TestCanResume(t *testing.T) {
 	err = cache.StartForward(context, input.Batch{
 		Positions: []int32{4, 5},
 		Sequences: []int{0, 0},
-	})
+	}, false)
 	if err != nil {
 		t.Fatalf("StartForward failed: %v", err)
 	}
@@ -372,14 +371,8 @@ func TestCanResume(t *testing.T) {
 	}
 }
 
-type testBackend struct{}
+type testBackend struct {
-
+	ml.Backend
-func (b *testBackend) Config() fs.Config {
-	panic("not implemented")
-}
-
-func (b *testBackend) Get(name string) ml.Tensor {
-	panic("not implemented")
 }
 
 func (b *testBackend) NewContext() ml.Context {
@@ -390,12 +383,10 @@ func (b *testBackend) NewContextSize(int) ml.Context {
 	return &testContext{}
 }
 
-func (b *testBackend) SystemInfo() string {
+type testContext struct {
-	return "not implemented"
+	ml.Context
 }
 
-type testContext struct{}
-
 func (c *testContext) Empty(dtype ml.DType, shape ...int) ml.Tensor {
 	total := 0
 
@@ -440,6 +431,8 @@ func (c *testContext) Forward(...ml.Tensor) ml.Context { return c }
 
 func (c *testContext) Compute(...ml.Tensor) {}
 
+func (c *testContext) Reserve() error { return nil }
+
 func (c *testContext) MaxGraphNodes() int {
 	return 10
 }
@@ -447,6 +440,8 @@ func (c *testContext) MaxGraphNodes() int {
 func (c *testContext) Close() {}
 
 type testTensor struct {
+	ml.Tensor
+
 	dtype       ml.DType
 	elementSize int
 	data        []float32
@@ -474,10 +469,6 @@ func (t *testTensor) DType() ml.DType {
 	return t.dtype
 }
 
-func (t *testTensor) Bytes() []byte {
-	panic("not implemented")
-}
-
 func (t *testTensor) Floats() []float32 {
 	out := make([]float32, len(t.data))
 	copy(out, t.data)
@@ -502,64 +493,6 @@ func (t *testTensor) Add(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
 	return out
 }
 
-func (t *testTensor) Mul(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Mulmat(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) MulmatFullPrec(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Softmax(ctx ml.Context) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) LayerNorm(ctx ml.Context, weight, bias ml.Tensor, eps float32) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) RMSNorm(ctx ml.Context, weight ml.Tensor, eps float32) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Scale(ctx ml.Context, s float64) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) AvgPool1D(ctx ml.Context, k, s, p int) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) AvgPool2D(ctx ml.Context, k, s int, p float32) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Conv2D(ctx ml.Context, weight ml.Tensor, s0, s1, p0, p1, d0, d1 int) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) RoPE(ctx ml.Context, positionIDs, ropeFactors ml.Tensor, dim, ropeType uint32, base, scale float32) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) IM2Col(ctx ml.Context, weight ml.Tensor, s0, s1, p0, p1, d0, d1 int) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Cos(ctx ml.Context) ml.Tensor  { panic("not implemented") }
-func (t *testTensor) Sin(ctx ml.Context) ml.Tensor  { panic("not implemented") }
-func (t *testTensor) Tanh(ctx ml.Context) ml.Tensor { panic("not implemented") }
-func (t *testTensor) GELU(ctx ml.Context) ml.Tensor { panic("not implemented") }
-func (t *testTensor) SILU(ctx ml.Context) ml.Tensor { panic("not implemented") }
-
-func (t *testTensor) Reshape(ctx ml.Context, shape ...int) ml.Tensor {
-	panic("not implemented")
-}
-
 func (t *testTensor) View(ctx ml.Context, offset int, shape ...int) ml.Tensor {
 	offset /= t.elementSize
 
@@ -582,43 +515,7 @@ func (t *testTensor) View(ctx ml.Context, offset int, shape ...int) ml.Tensor {
 	return view
 }
 
-func (t *testTensor) Permute(ctx ml.Context, shape ...int) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Contiguous(ctx ml.Context) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Set(ctx ml.Context, t2 ml.Tensor, offset int, strides ...int) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Pad(ctx ml.Context, shape ...int) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Unpad(ctx ml.Context, shape ...int) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Stack(ctx ml.Context, dim int, s ...ml.Tensor) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Repeat(ctx ml.Context, dim, n int) ml.Tensor { panic("not implemented") }
-
-func (t *testTensor) Concat(ctx ml.Context, t2 ml.Tensor, dim int) ml.Tensor {
-	panic("not implemented")
-}
-
-func (t *testTensor) Rows(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
-	panic("not implemented")
-}
-
 func (t *testTensor) Copy(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
 	copy(t2.(*testTensor).data, t.data)
 	return nil
 }
-
-func (t *testTensor) Duplicate(ctx ml.Context) ml.Tensor { panic("not implemented") }

kvcache/encoder.go

@@ -27,6 +27,11 @@ type EncoderCache struct {
 	// anything will be stored)
 	curPos int32
 
+	// curReserve indicates that this forward pass is only for
+	// memory reservation and we should not update our metadata
+	// based on it.
+	curReserve bool
+
 	// ** cache metadata **
 
 	// was something stored in the cache?
@@ -83,12 +88,14 @@ func (c *EncoderCache) Close() {
 	}
 }
 
-func (c *EncoderCache) StartForward(ctx ml.Context, batch input.Batch) error {
+func (c *EncoderCache) StartForward(ctx ml.Context, batch input.Batch, reserve bool) error {
 	// We work with the most recent image
 	if len(batch.Multimodal) > 0 {
 		c.curPos = batch.Positions[batch.Multimodal[len(batch.Multimodal)-1].Index]
 	}
 
+	c.curReserve = reserve
+
 	return nil
 }
 
@@ -105,8 +112,10 @@ func (c *EncoderCache) Get(ctx ml.Context) (ml.Tensor, ml.Tensor, ml.Tensor) {
 }
 
 func (c *EncoderCache) Put(ctx ml.Context, key, value ml.Tensor) {
-	c.encoderPos = c.curPos
+	if !c.curReserve {
-	c.encoderCached = true
+		c.encoderPos = c.curPos
+		c.encoderCached = true
+	}
 
 	if c.config.PermutedV {
 		value = value.Permute(ctx, 1, 2, 0, 3)

kvcache/wrapper.go

@@ -41,9 +41,9 @@ func (c *WrapperCache) Close() {
 	}
 }
 
-func (c *WrapperCache) StartForward(ctx ml.Context, batch input.Batch) error {
+func (c *WrapperCache) StartForward(ctx ml.Context, batch input.Batch, reserve bool) error {
 	for i, cache := range c.caches {
-		err := cache.StartForward(ctx, batch)
+		err := cache.StartForward(ctx, batch, reserve)
 		if err != nil {
 			// unwind on error - Remove with endIndex set to math.MaxInt32 does not fail
 			for j := i - 1; j >= 0; j-- {

ml/backend.go

@@ -97,6 +97,13 @@ type Context interface {
 
 	Forward(...Tensor) Context
 	Compute(...Tensor)
+
+	// Reserve is analogous to Compute but rather than executing a
+	// graph, simply preallocates memory. Typically called with a
+	// worst case graph to ensure all resources are available for
+	// for future inference.
+	Reserve() error
+
 	MaxGraphNodes() int
 	Close()
 

ml/backend/ggml/ggml.go

@@ -10,6 +10,7 @@ import "C"
 
 import (
 	"context"
+	"errors"
 	"fmt"
 	"io"
 	"log/slog"
@@ -42,8 +43,12 @@ func devices() []*C.struct_ggml_backend_device {
 }
 
 type Backend struct {
-	meta    *fsggml.GGML
+	meta *fsggml.GGML
-	sched   *C.struct_ggml_backend_sched
+
+	sched         *C.struct_ggml_backend_sched
+	schedBackends []*C.struct_ggml_backend
+	schedBufts    []*C.struct_ggml_backend_buffer_type
+
 	tensors map[string]*C.struct_ggml_tensor
 
 	// input is the backend used for inputs
@@ -281,6 +286,10 @@ func New(ctx context.Context, r *os.File, params ml.BackendParams) (ml.Backend,
 		}
 
 		b := C.ggml_backend_alloc_ctx_tensors_from_buft(c, bt)
+		if b == nil {
+			return nil, fmt.Errorf("unable to allocate memory from device %v for model weights", C.GoString(C.ggml_backend_buft_name(bt)))
+		}
+
 		C.ggml_backend_buffer_set_usage(b, C.GGML_BACKEND_BUFFER_USAGE_WEIGHTS)
 		bbs[c] = b
 	}
@@ -385,8 +394,6 @@ func New(ctx context.Context, r *os.File, params ml.BackendParams) (ml.Backend,
 		schedBackends = append(schedBackends, b)
 		schedBufts = append(schedBufts, bt)
 
-		slog.Info("compute graph", "backend", C.GoString(C.ggml_backend_name(b)), "buffer_type", C.GoString(C.ggml_backend_buft_name(bt)))
-
 		if C.ggml_backend_is_cpu(b) {
 			// set number of threads for cpu backend
 			C.ggml_backend_cpu_set_n_threads(b, C.int(Threads(params.NumThreads)))
@@ -405,7 +412,9 @@ func New(ctx context.Context, r *os.File, params ml.BackendParams) (ml.Backend,
 			C.size_t(maxGraphNodes),
 			C._Bool(len(gpus) > 1 && slices.Contains(gpus, output.d)),
 		),
-		input: deviceBufferTypes[input.d],
+		schedBackends: schedBackends,
+		schedBufts:    schedBufts,
+		input:         deviceBufferTypes[input.d],
 		layers: func() map[int]*C.struct_ggml_backend_buffer_type {
 			m := make(map[int]*C.struct_ggml_backend_buffer_type)
 			for i, layer := range layers {
@@ -530,6 +539,24 @@ func (c Context) Compute(tensors ...ml.Tensor) {
 	}
 }
 
+func (c Context) Reserve() error {
+	if !C.ggml_backend_sched_reserve(c.b.sched, c.graph) {
+		C.ggml_backend_sched_reset(c.b.sched)
+		return errors.New("failed to reserve graph")
+	}
+
+	slog.Debug("compute graph", "nodes", C.ggml_graph_n_nodes(c.graph), "splits", C.ggml_backend_sched_get_n_splits(c.b.sched))
+	for i := range c.b.schedBackends {
+		size := C.ggml_backend_sched_get_buffer_size(c.b.sched, c.b.schedBackends[i])
+		slog.Info("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(size)))
+	}
+
+	C.ggml_backend_sched_reset(c.b.sched)
+
+	return nil
+}
+
 func (c Context) MaxGraphNodes() int {
 	return c.maxGraphNodes
 }
@@ -547,9 +574,9 @@ func pad(length, pad C.size_t) C.size_t {
 	return ((length + pad - 1) / pad) * pad
 }
 
-func (c Context) newTensor(dtype ml.DType, shape []int) ml.Tensor {
+func (c Context) newTensor(dtype ml.DType, shape []int) (ml.Tensor, error) {
 	if c.buft == nil {
-		panic("set Input, Output, or Layer before creating tensors")
+		panic("set Input or Layer before creating tensors")
 	}
 
 	var cdtype uint32
@@ -570,7 +597,7 @@ func (c Context) newTensor(dtype ml.DType, shape []int) ml.Tensor {
 
 	if len(shape) < 1 || shape[0] == 0 {
 		var shape C.int64_t = 0
-		return &Tensor{b: c.b, t: C.ggml_new_tensor(c.ctx, cdtype, 1, &shape)}
+		return &Tensor{b: c.b, t: C.ggml_new_tensor(c.ctx, cdtype, 1, &shape)}, nil
 	} else if len(shape) > 4 {
 		panic("unsupported number of dimensions")
 	}
@@ -584,16 +611,29 @@ func (c Context) newTensor(dtype ml.DType, shape []int) ml.Tensor {
 	t := C.ggml_new_tensor(c.ctx, cdtype, C.int(len(shape)), shapeToGGML(shape))
 	size := pad(C.ggml_backend_buft_get_alloc_size(c.buft, t), C.ggml_backend_buft_get_alignment(c.buft))
 	b := C.ggml_backend_buft_alloc_buffer(c.buft, size)
+	if b == nil {
+		return nil, fmt.Errorf("unable to allocate %v from device %v for new tensor", format.HumanBytes2(uint64(size)), C.GoString(C.ggml_backend_buft_name(c.buft)))
+	}
+
 	C.ggml_backend_tensor_alloc(b, t, C.ggml_backend_buffer_get_base(b))
-	return &Tensor{b: c.b, t: t}
+	return &Tensor{b: c.b, t: t}, nil
 }
 
 func (c Context) Empty(dtype ml.DType, shape ...int) ml.Tensor {
-	return c.newTensor(dtype, shape)
+	t, err := c.newTensor(dtype, shape)
+	if err != nil {
+		panic(err)
+	}
+
+	return t
 }
 
 func (c Context) Zeros(dtype ml.DType, shape ...int) ml.Tensor {
-	t := c.newTensor(dtype, shape)
+	t, err := c.newTensor(dtype, shape)
+	if err != nil {
+		panic(err)
+	}
+
 	C.ggml_set_zero(t.(*Tensor).t)
 	return t
 }
@@ -621,7 +661,11 @@ func (c Context) FromFloatSlice(s []float32, shape ...int) (ml.Tensor, error) {
 		return nil, err
 	}
 
-	t := c.newTensor(ml.DTypeF32, shape)
+	t, err := c.newTensor(ml.DTypeF32, shape)
+	if err != nil {
+		return nil, err
+	}
+
 	if len(s) > 0 {
 		C.ggml_backend_tensor_set(t.(*Tensor).t, unsafe.Pointer(&s[0]), 0, C.ggml_nbytes(t.(*Tensor).t))
 	}
@@ -634,7 +678,11 @@ func (c Context) FromIntSlice(s []int32, shape ...int) (ml.Tensor, error) {
 		return nil, err
 	}
 
-	t := c.newTensor(ml.DTypeI32, shape)
+	t, err := c.newTensor(ml.DTypeI32, shape)
+	if err != nil {
+		return nil, err
+	}
+
 	if len(s) > 0 {
 		C.ggml_backend_tensor_set(t.(*Tensor).t, unsafe.Pointer(&s[0]), 0, C.ggml_nbytes(t.(*Tensor).t))
 	}

model/model.go

@@ -299,7 +299,7 @@ func Forward(ctx ml.Context, m Model, inputs []int32, batch input.Batch) (ml.Ten
 
 	cache := m.Config().Cache
 	if cache != nil {
-		err := cache.StartForward(ctx, batch)
+		err := cache.StartForward(ctx, batch, false)
 		if err != nil {
 			return nil, err
 		}

openai/openai_test.go

@@ -281,19 +281,23 @@ func TestChatMiddleware(t *testing.T) {
 							Description: "Get the current weather",
 							Parameters: struct {
 								Type       string   `json:"type"`
+								Defs       any      `json:"$defs,omitempty"`
+								Items      any      `json:"items,omitempty"`
 								Required   []string `json:"required"`
 								Properties map[string]struct {
 									Type        api.PropertyType `json:"type"`
+									Items       any              `json:"items,omitempty"`
 									Description string           `json:"description"`
-									Enum        []string         `json:"enum,omitempty"`
+									Enum        []any            `json:"enum,omitempty"`
 								} `json:"properties"`
 							}{
 								Type:     "object",
 								Required: []string{"location"},
 								Properties: map[string]struct {
 									Type        api.PropertyType `json:"type"`
+									Items       any              `json:"items,omitempty"`
 									Description string           `json:"description"`
-									Enum        []string         `json:"enum,omitempty"`
+									Enum        []any            `json:"enum,omitempty"`
 								}{
 									"location": {
 										Type:        api.PropertyType{"string"},
@@ -301,7 +305,7 @@ func TestChatMiddleware(t *testing.T) {
 									},
 									"unit": {
 										Type: api.PropertyType{"string"},
-										Enum: []string{"celsius", "fahrenheit"},
+										Enum: []any{"celsius", "fahrenheit"},
 									},
 								},
 							},

runner/ollamarunner/cache_test.go

@@ -448,7 +448,7 @@ func (m *mockCache) Get(ctx ml.Context) (ml.Tensor, ml.Tensor, ml.Tensor)
 func (m *mockCache) Put(ctx ml.Context, key, value ml.Tensor)                                      {}
 func (m *mockCache) Init(backend ml.Backend, dtype ml.DType, maxSequences, capacity, maxBatch int) {}
 func (m *mockCache) Close()                                                                        {}
-func (m *mockCache) StartForward(ctx ml.Context, batch input.Batch) error                          { return nil }
+func (m *mockCache) StartForward(ctx ml.Context, batch input.Batch, reserve bool) error            { return nil }
 func (m *mockCache) CopyPrefix(srcSeq, dstSeq int, len int32)                                      {}
 func (m *mockCache) SetConfig(ml.CacheConfig)                                                      {}
 func (m *mockCache) CanResume(seq int, pos int32) bool                                             { return true }

runner/ollamarunner/runner.go

@@ -728,6 +728,51 @@ func (m *multiLPath) String() string {
 	return strings.Join(*m, ", ")
 }
 
+func (s *Server) reserveWorstCaseGraph() error {
+	ctx := s.model.Backend().NewContext()
+	defer ctx.Close()
+
+	var batch input.Batch
+
+	inputs := make([]int32, s.batchSize)
+	batch.Positions = make([]int32, len(inputs))
+	batch.Sequences = make([]int, len(inputs))
+	for i := range inputs {
+		batch.Positions[i] = int32(i)
+	}
+
+	batch.Outputs = make([]int32, s.parallel)
+	for i := range batch.Outputs {
+		batch.Outputs[i] = int32(i)
+	}
+
+	var err error
+	batch.Inputs, err = ctx.Input().FromIntSlice(inputs, len(inputs))
+	if err != nil {
+		return err
+	}
+
+	cache := s.model.Config().Cache
+	if cache != nil {
+		err := cache.StartForward(ctx, batch, true)
+		if err != nil {
+			return err
+		}
+	}
+
+	t, err := s.model.Forward(ctx, batch)
+	if err != nil {
+		return err
+	}
+
+	err = ctx.Forward(t).Reserve()
+	if err != nil {
+		return err
+	}
+
+	return nil
+}
+
 func (s *Server) loadModel(
 	ctx context.Context,
 	mpath string,
@@ -765,6 +810,11 @@ func (s *Server) loadModel(
 	s.seqs = make([]*Sequence, s.parallel)
 	s.seqsSem = semaphore.NewWeighted(int64(s.parallel))
 
+	err = s.reserveWorstCaseGraph()
+	if err != nil {
+		panic(err)
+	}
+
 	s.status = llm.ServerStatusReady
 	s.ready.Done()
 }

server/routes_generate_test.go

@@ -370,19 +370,23 @@ func TestGenerateChat(t *testing.T) {
 					Description: "Get the current weather",
 					Parameters: struct {
 						Type       string   `json:"type"`
+						Defs       any      `json:"$defs,omitempty"`
+						Items      any      `json:"items,omitempty"`
 						Required   []string `json:"required"`
 						Properties map[string]struct {
 							Type        api.PropertyType `json:"type"`
+							Items       any              `json:"items,omitempty"`
 							Description string           `json:"description"`
-							Enum        []string         `json:"enum,omitempty"`
+							Enum        []any            `json:"enum,omitempty"`
 						} `json:"properties"`
 					}{
 						Type:     "object",
 						Required: []string{"location"},
 						Properties: map[string]struct {
 							Type        api.PropertyType `json:"type"`
+							Items       any              `json:"items,omitempty"`
 							Description string           `json:"description"`
-							Enum        []string         `json:"enum,omitempty"`
+							Enum        []any            `json:"enum,omitempty"`
 						}{
 							"location": {
 								Type:        api.PropertyType{"string"},
@@ -390,7 +394,7 @@ func TestGenerateChat(t *testing.T) {
 							},
 							"unit": {
 								Type: api.PropertyType{"string"},
-								Enum: []string{"celsius", "fahrenheit"},
+								Enum: []any{"celsius", "fahrenheit"},
 							},
 						},
 					},
@@ -467,19 +471,23 @@ func TestGenerateChat(t *testing.T) {
 					Description: "Get the current weather",
 					Parameters: struct {
 						Type       string   `json:"type"`
+						Defs       any      `json:"$defs,omitempty"`
+						Items      any      `json:"items,omitempty"`
 						Required   []string `json:"required"`
 						Properties map[string]struct {
 							Type        api.PropertyType `json:"type"`
+							Items       any              `json:"items,omitempty"`
 							Description string           `json:"description"`
-							Enum        []string         `json:"enum,omitempty"`
+							Enum        []any            `json:"enum,omitempty"`
 						} `json:"properties"`
 					}{
 						Type:     "object",
 						Required: []string{"location"},
 						Properties: map[string]struct {
 							Type        api.PropertyType `json:"type"`
+							Items       any              `json:"items,omitempty"`
 							Description string           `json:"description"`
-							Enum        []string         `json:"enum,omitempty"`
+							Enum        []any            `json:"enum,omitempty"`
 						}{
 							"location": {
 								Type:        api.PropertyType{"string"},
@@ -487,7 +495,7 @@ func TestGenerateChat(t *testing.T) {
 							},
 							"unit": {
 								Type: api.PropertyType{"string"},
-								Enum: []string{"celsius", "fahrenheit"},
+								Enum: []any{"celsius", "fahrenheit"},
 							},
 						},
 					},

server/sched.go

@@ -667,13 +667,19 @@ func (runner *runnerRef) waitForVRAMRecovery() chan any {
 	return finished
 }
 
-type ByDuration []*runnerRef
+type ByDurationAndName []*runnerRef
 
-func (a ByDuration) Len() int      { return len(a) }
+func (a ByDurationAndName) Len() int      { return len(a) }
-func (a ByDuration) Swap(i, j int) { a[i], a[j] = a[j], a[i] }
+func (a ByDurationAndName) Swap(i, j int) { a[i], a[j] = a[j], a[i] }
-func (a ByDuration) Less(i, j int) bool {
+func (a ByDurationAndName) Less(i, j int) bool {
-	// uint64 to turn negative time (never unload) to largest
+	// Primary sort by session duration (uint64 to handle negatives)
-	return uint64(a[i].sessionDuration) < uint64(a[j].sessionDuration)
+	d1 := uint64(a[i].sessionDuration)
+	d2 := uint64(a[j].sessionDuration)
+	if d1 != d2 {
+		return d1 < d2
+	}
+	// Secondary sort by model path lex order
+	return a[i].modelPath < a[j].modelPath
 }
 
 // TODO - future consideration to pick runners based on size
@@ -775,7 +781,7 @@ func (s *Scheduler) findRunnerToUnload() *runnerRef {
 
 	// In the future we can enhance the algorithm to be smarter about picking the optimal runner to unload
 	// e.g., if we have multiple options, will one make room for the request?
-	sort.Sort(ByDuration(runnerList))
+	sort.Sort(ByDurationAndName(runnerList))
 
 	// First try to find a runner that's already idle
 	for _, runner := range runnerList {