Merge https://github.com/ollama/ollama

Signed-off-by: baalajimaestro <me@baalajimaestro.me>
2024-07-06 23:37:14 +05:30 · 2024-07-06 23:37:14 +05:30 · 415d9f0f15
commit 415d9f0f15
parent 110deb68cf 5796bfc401
79 changed files with 1964 additions and 980 deletions
--- a/.github/workflows/release.yaml
+++ b/.github/workflows/release.yaml
@ -304,6 +304,10 @@ jobs:
          write-host "Installing plugin"
          & "${env:RUNNER_TEMP}\plugin\*\kmscng.msi" /quiet
          write-host "plugin installed"
+      - name: remove unwanted mingw dll.a files
+        run: |
+          Remove-Item "C:\mingw64\x86_64-w64-mingw32\lib\libpthread.dll.a"
+          Remove-Item "C:\mingw64\x86_64-w64-mingw32\lib\libwinpthread.dll.a"
      - uses: actions/setup-go@v5
        with:
          go-version-file: go.mod
--- a/.github/workflows/test.yaml
+++ b/.github/workflows/test.yaml
@ -58,6 +58,7 @@ jobs:
    runs-on: ${{ matrix.os }}
    env:
      GOARCH: ${{ matrix.arch }}
+      CGO_ENABLED: '1'
    steps:
      - uses: actions/checkout@v4
      - uses: actions/setup-go@v5
@ -79,6 +80,7 @@ jobs:
      - run: go generate -x ./...
        if: ${{ ! startsWith(matrix.os, 'windows-') }}
        name: 'Unix Go Generate'
+      - run: go build .
      - uses: actions/upload-artifact@v4
        with:
          name: ${{ matrix.os }}-${{ matrix.arch }}-libraries
--- a/api/types.go
+++ b/api/types.go
@ -159,49 +159,18 @@ type Options struct {

 // Runner options which must be set when the model is loaded into memory
 type Runner struct {
-	UseNUMA   bool     `json:"numa,omitempty"`
-	NumCtx    int      `json:"num_ctx,omitempty"`
-	NumBatch  int      `json:"num_batch,omitempty"`
-	NumGPU    int      `json:"num_gpu,omitempty"`
-	MainGPU   int      `json:"main_gpu,omitempty"`
-	LowVRAM   bool     `json:"low_vram,omitempty"`
-	F16KV     bool     `json:"f16_kv,omitempty"`
-	LogitsAll bool     `json:"logits_all,omitempty"`
-	VocabOnly bool     `json:"vocab_only,omitempty"`
-	UseMMap   TriState `json:"use_mmap,omitempty"`
-	UseMLock  bool     `json:"use_mlock,omitempty"`
-	NumThread int      `json:"num_thread,omitempty"`
-}
-
-type TriState int
-
-const (
-	TriStateUndefined TriState = -1
-	TriStateFalse     TriState = 0
-	TriStateTrue      TriState = 1
-)
-
-func (b *TriState) UnmarshalJSON(data []byte) error {
-	var v bool
-	if err := json.Unmarshal(data, &v); err != nil {
-		return err
-	}
-	if v {
-		*b = TriStateTrue
-	}
-	*b = TriStateFalse
-	return nil
-}
-
-func (b *TriState) MarshalJSON() ([]byte, error) {
-	if *b == TriStateUndefined {
-		return nil, nil
-	}
-	var v bool
-	if *b == TriStateTrue {
-		v = true
-	}
-	return json.Marshal(v)
+	UseNUMA   bool  `json:"numa,omitempty"`
+	NumCtx    int   `json:"num_ctx,omitempty"`
+	NumBatch  int   `json:"num_batch,omitempty"`
+	NumGPU    int   `json:"num_gpu,omitempty"`
+	MainGPU   int   `json:"main_gpu,omitempty"`
+	LowVRAM   bool  `json:"low_vram,omitempty"`
+	F16KV     bool  `json:"f16_kv,omitempty"`
+	LogitsAll bool  `json:"logits_all,omitempty"`
+	VocabOnly bool  `json:"vocab_only,omitempty"`
+	UseMMap   *bool `json:"use_mmap,omitempty"`
+	UseMLock  bool  `json:"use_mlock,omitempty"`
+	NumThread int   `json:"num_thread,omitempty"`
 }

 // EmbeddingRequest is the request passed to [Client.Embeddings].
@ -345,6 +314,13 @@ type ProcessModelResponse struct {
 	SizeVRAM  int64        `json:"size_vram"`
 }

+type RetrieveModelResponse struct {
+	Id      string `json:"id"`
+	Object  string `json:"object"`
+	Created int64  `json:"created"`
+	OwnedBy string `json:"owned_by"`
+}
+
 type TokenResponse struct {
 	Token string `json:"token"`
 }
@ -437,19 +413,6 @@ func (opts *Options) FromMap(m map[string]interface{}) error {
 				continue
 			}

-			if reflect.PointerTo(field.Type()) == reflect.TypeOf((*TriState)(nil)) {
-				val, ok := val.(bool)
-				if !ok {
-					return fmt.Errorf("option %q must be of type boolean", key)
-				}
-				if val {
-					field.SetInt(int64(TriStateTrue))
-				} else {
-					field.SetInt(int64(TriStateFalse))
-				}
-				continue
-			}
-
 			switch field.Kind() {
 			case reflect.Int:
 				switch t := val.(type) {
@ -496,6 +459,17 @@ func (opts *Options) FromMap(m map[string]interface{}) error {
 					slice[i] = str
 				}
 				field.Set(reflect.ValueOf(slice))
+			case reflect.Pointer:
+				var b bool
+				if field.Type() == reflect.TypeOf(&b) {
+					val, ok := val.(bool)
+					if !ok {
+						return fmt.Errorf("option %q must be of type boolean", key)
+					}
+					field.Set(reflect.ValueOf(&val))
+				} else {
+					return fmt.Errorf("unknown type loading config params: %v %v", field.Kind(), field.Type())
+				}
 			default:
 				return fmt.Errorf("unknown type loading config params: %v", field.Kind())
 			}
@ -538,7 +512,7 @@ func DefaultOptions() Options {
 			LowVRAM:   false,
 			F16KV:     true,
 			UseMLock:  false,
-			UseMMap:   TriStateUndefined,
+			UseMMap:   nil,
 			UseNUMA:   false,
 		},
 	}
@ -608,19 +582,6 @@ func FormatParams(params map[string][]string) (map[string]interface{}, error) {
 		} else {
 			field := valueOpts.FieldByName(opt.Name)
 			if field.IsValid() && field.CanSet() {
-				if reflect.PointerTo(field.Type()) == reflect.TypeOf((*TriState)(nil)) {
-					boolVal, err := strconv.ParseBool(vals[0])
-					if err != nil {
-						return nil, fmt.Errorf("invalid bool value %s", vals)
-					}
-					if boolVal {
-						out[key] = TriStateTrue
-					} else {
-						out[key] = TriStateFalse
-					}
-					continue
-				}
-
 				switch field.Kind() {
 				case reflect.Float32:
 					floatVal, err := strconv.ParseFloat(vals[0], 32)
@ -648,6 +609,17 @@ func FormatParams(params map[string][]string) (map[string]interface{}, error) {
 				case reflect.Slice:
 					// TODO: only string slices are supported right now
 					out[key] = vals
+				case reflect.Pointer:
+					var b bool
+					if field.Type() == reflect.TypeOf(&b) {
+						boolVal, err := strconv.ParseBool(vals[0])
+						if err != nil {
+							return nil, fmt.Errorf("invalid bool value %s", vals)
+						}
+						out[key] = &boolVal
+					} else {
+						return nil, fmt.Errorf("unknown type %s for %s", field.Kind(), key)
+					}
 				default:
 					return nil, fmt.Errorf("unknown type %s for %s", field.Kind(), key)
 				}
--- a/api/types_test.go
+++ b/api/types_test.go
@ -108,25 +108,27 @@ func TestDurationMarshalUnmarshal(t *testing.T) {
 }

 func TestUseMmapParsingFromJSON(t *testing.T) {
+	tr := true
+	fa := false
 	tests := []struct {
 		name string
 		req  string
-		exp  TriState
+		exp  *bool
 	}{
 		{
 			name: "Undefined",
 			req:  `{ }`,
-			exp:  TriStateUndefined,
+			exp:  nil,
 		},
 		{
 			name: "True",
 			req:  `{ "use_mmap": true }`,
-			exp:  TriStateTrue,
+			exp:  &tr,
 		},
 		{
 			name: "False",
 			req:  `{ "use_mmap": false }`,
-			exp:  TriStateFalse,
+			exp:  &fa,
 		},
 	}

@ -144,50 +146,52 @@ func TestUseMmapParsingFromJSON(t *testing.T) {
 }

 func TestUseMmapFormatParams(t *testing.T) {
+	tr := true
+	fa := false
 	tests := []struct {
 		name string
 		req  map[string][]string
-		exp  TriState
+		exp  *bool
 		err  error
 	}{
 		{
 			name: "True",
 			req: map[string][]string{
-				"use_mmap": []string{"true"},
+				"use_mmap": {"true"},
 			},
-			exp: TriStateTrue,
+			exp: &tr,
 			err: nil,
 		},
 		{
 			name: "False",
 			req: map[string][]string{
-				"use_mmap": []string{"false"},
+				"use_mmap": {"false"},
 			},
-			exp: TriStateFalse,
+			exp: &fa,
 			err: nil,
 		},
 		{
 			name: "Numeric True",
 			req: map[string][]string{
-				"use_mmap": []string{"1"},
+				"use_mmap": {"1"},
 			},
-			exp: TriStateTrue,
+			exp: &tr,
 			err: nil,
 		},
 		{
 			name: "Numeric False",
 			req: map[string][]string{
-				"use_mmap": []string{"0"},
+				"use_mmap": {"0"},
 			},
-			exp: TriStateFalse,
+			exp: &fa,
 			err: nil,
 		},
 		{
 			name: "invalid string",
 			req: map[string][]string{
-				"use_mmap": []string{"foo"},
+				"use_mmap": {"foo"},
 			},
-			exp: TriStateUndefined,
+			exp: nil,
 			err: fmt.Errorf("invalid bool value [foo]"),
 		},
 	}
@ -195,11 +199,11 @@ func TestUseMmapFormatParams(t *testing.T) {
 	for _, test := range tests {
 		t.Run(test.name, func(t *testing.T) {
 			resp, err := FormatParams(test.req)
-			require.Equal(t, err, test.err)
+			require.Equal(t, test.err, err)
 			respVal, ok := resp["use_mmap"]
-			if test.exp != TriStateUndefined {
+			if test.exp != nil {
 				assert.True(t, ok, "resp: %v", resp)
-				assert.Equal(t, test.exp, respVal)
+				assert.Equal(t, *test.exp, *respVal.(*bool))
 			}
 		})
 	}
--- a/docs/development.md
+++ b/docs/development.md
@ -104,7 +104,7 @@ like to use. For example, to compile an optimized binary for an Intel i9-9880H,
 you might use:

 ```
-OLLAMA_CUSTOM_CPU_DEFS="-DLLAMA_AVX=on -DLLAMA_AVX2=on -DLLAMA_F16C=on -DLLAMA_FMA=on" go generate ./...
+OLLAMA_CUSTOM_CPU_DEFS="-DGGML_AVX=on -DGGML_AVX2=on -DGGML_F16C=on -DGGML_FMA=on" go generate ./...
 go build .
 ```

--- a/docs/faq.md
+++ b/docs/faq.md
@ -257,3 +257,19 @@ If you wish to override the `OLLAMA_KEEP_ALIVE` setting, use the `keep_alive` AP
 ## How do I manage the maximum number of requests the Ollama server can queue?

 If too many requests are sent to the server, it will respond with a 503 error indicating the server is overloaded.  You can adjust how many requests may be queue by setting `OLLAMA_MAX_QUEUE`.
+
+## How does Ollama handle concurrent requests?
+
+Ollama supports two levels of concurrent processing.  If your system has sufficient available memory (system memory when using CPU inference, or VRAM for GPU inference) then multiple models can be loaded at the same time.  For a given model, if there is sufficient available memory when the model is loaded, it is configured to allow parallel request processing.
+
+If there is insufficient available memory to load a new model request while one or more models are already loaded, all new requests will be queued until the new model can be loaded.  As prior models become idle, one or more will be unloaded to make room for the new model.  Queued requests will be processed in order.  When using GPU inference new models must be able to completely fit in VRAM to allow concurrent model loads.
+
+Parallel request processing for a given model results in increasing the context size by the number of parallel requests.  For example, a 2K context with 4 parallel requests will result in an 8K context and additional memory allocation.
+
+The following server settings may be used to adjust how Ollama handles concurrent requests on most platforms:
+
+- `OLLAMA_MAX_LOADED_MODELS` - The maximum number of models that can be loaded concurrently provided they fit in available memory.  The default is 3 * the number of GPUs or 3 for CPU inference.
+- `OLLAMA_NUM_PARALLEL` - The maximum number of parallel requests each model will process at the same time.  The default will auto-select either 4 or 1 based on available memory.
+- `OLLAMA_MAX_QUEUE` - The maximum number of requests Ollama will queue when busy before rejecting additional requests. The default is 512
+
+Note: Windows with Radeon GPUs currently default to 1 model maximum due to limitations in ROCm v5.7 for available VRAM reporting.  Once ROCm v6 is available, Windows Radeon will follow the defaults above.  You may enable concurrent model loads on Radeon on Windows, but ensure you don't load more models than will fit into your GPUs VRAM.
--- a/docs/openai.md
+++ b/docs/openai.md
@ -65,6 +65,7 @@ curl http://localhost:11434/v1/chat/completions \
            }
        ]
    }'
+
 ```

 ## Endpoints
--- a/docs/troubleshooting.md
+++ b/docs/troubleshooting.md
@ -70,14 +70,18 @@ curl -fsSL https://ollama.com/install.sh | OLLAMA_VERSION="0.1.29" sh

 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/

-## Container fails to run on NVIDIA GPU
+## NVIDIA GPU Discovery

-Make sure you've set up the container runtime first as described in [docker.md](./docker.md)
+When Ollama starts up, it takes inventory of the GPUs present in the system to determine compatibility and how much VRAM is available.  Sometimes this discovery can fail to find your GPUs.  In general, running the latest driver will yield the best results.

-Sometimes the container runtime can have difficulties initializing the GPU. When you check the server logs, this can show up as various error codes, such as "3" (not initialized), "46" (device unavailable), "100" (no device), "999" (unknown), or others. The following troubleshooting techniques may help resolve the problem
+### Linux NVIDIA Troubleshooting

- Is the container runtime working?  Try `docker run --gpus all ubuntu nvidia-smi` - if this doesn't work, Ollama wont be able to see your NVIDIA GPU.
- Is the uvm driver not loaded? `sudo nvidia-modprobe -u`
+If you are using a container to run Ollama, make sure you've set up the container runtime first as described in [docker.md](./docker.md)
+
+Sometimes the Ollama can have difficulties initializing the GPU. When you check the server logs, this can show up as various error codes, such as "3" (not initialized), "46" (device unavailable), "100" (no device), "999" (unknown), or others. The following troubleshooting techniques may help resolve the problem
+
+- If you are using a container, is the container runtime working?  Try `docker run --gpus all ubuntu nvidia-smi` - if this doesn't work, Ollama wont be able to see your NVIDIA GPU.
+- Is the uvm driver loaded? `sudo nvidia-modprobe -u`
 - Try reloading the nvidia_uvm driver - `sudo rmmod nvidia_uvm` then `sudo modprobe nvidia_uvm`
 - Try rebooting
 - Make sure you're running the latest nvidia drivers
@ -85,3 +89,8 @@ Sometimes the container runtime can have difficulties initializing the GPU. When
 If none of those resolve the problem, gather additional information and file an issue:
 - Set `CUDA_ERROR_LEVEL=50` and try again to get more diagnostic logs
 - Check dmesg for any errors `sudo dmesg | grep -i nvrm` and `sudo dmesg | grep -i nvidia`
+
+
+## Windows Terminal Errors
+
+Older versions of Windows 10 (e.g., 21H1) are known to have a bug where the standard terminal program does not display control characters correctly.  This can result in a long string of strings like `←[?25h←[?25l` being displayed, sometimes erroring with `The parameter is incorrect`  To resolve this problem, please update to Win 10 22H1 or newer.
--- a/docs/windows.md
+++ b/docs/windows.md
@ -19,7 +19,7 @@ Logs will often be helpful in diagnosing the problem (see

 ## System Requirements

-* Windows 10 or newer, Home or Pro
+* Windows 10 22H2 or newer, Home or Pro
 * NVIDIA 452.39 or newer Drivers if you have an NVIDIA card
 * AMD Radeon Driver https://www.amd.com/en/support if you have a Radeon card

--- a/envconfig/config.go
+++ b/envconfig/config.go
@ -4,12 +4,14 @@ import (
 	"errors"
 	"fmt"
 	"log/slog"
+	"math"
 	"net"
 	"os"
 	"path/filepath"
 	"runtime"
 	"strconv"
 	"strings"
+	"time"
 )

 type OllamaHost struct {
@ -34,17 +36,17 @@ var (
 	// Set via OLLAMA_HOST in the environment
 	Host *OllamaHost
 	// Set via OLLAMA_KEEP_ALIVE in the environment
-	KeepAlive string
+	KeepAlive time.Duration
 	// Set via OLLAMA_LLM_LIBRARY in the environment
 	LLMLibrary string
 	// Set via OLLAMA_MAX_LOADED_MODELS in the environment
 	MaxRunners int
 	// Set via OLLAMA_MAX_QUEUE in the environment
 	MaxQueuedRequests int
-	// Set via OLLAMA_MODELS in the environment
-	ModelsDir string
 	// Set via OLLAMA_MAX_VRAM in the environment
 	MaxVRAM uint64
+	// Set via OLLAMA_MODELS in the environment
+	ModelsDir string
 	// Set via OLLAMA_NOHISTORY in the environment
 	NoHistory bool
 	// Set via OLLAMA_NOPRUNE in the environment
@ -85,13 +87,13 @@ func AsMap() map[string]EnvVar {
 		"OLLAMA_HOST":              {"OLLAMA_HOST", Host, "IP Address for the ollama server (default 127.0.0.1:11434)"},
 		"OLLAMA_KEEP_ALIVE":        {"OLLAMA_KEEP_ALIVE", KeepAlive, "The duration that models stay loaded in memory (default \"5m\")"},
 		"OLLAMA_LLM_LIBRARY":       {"OLLAMA_LLM_LIBRARY", LLMLibrary, "Set LLM library to bypass autodetection"},
-		"OLLAMA_MAX_LOADED_MODELS": {"OLLAMA_MAX_LOADED_MODELS", MaxRunners, "Maximum number of loaded models (default 1)"},
+		"OLLAMA_MAX_LOADED_MODELS": {"OLLAMA_MAX_LOADED_MODELS", MaxRunners, "Maximum number of loaded models per GPU"},
 		"OLLAMA_MAX_QUEUE":         {"OLLAMA_MAX_QUEUE", MaxQueuedRequests, "Maximum number of queued requests"},
 		"OLLAMA_MAX_VRAM":          {"OLLAMA_MAX_VRAM", MaxVRAM, "Maximum VRAM"},
 		"OLLAMA_MODELS":            {"OLLAMA_MODELS", ModelsDir, "The path to the models directory"},
 		"OLLAMA_NOHISTORY":         {"OLLAMA_NOHISTORY", NoHistory, "Do not preserve readline history"},
 		"OLLAMA_NOPRUNE":           {"OLLAMA_NOPRUNE", NoPrune, "Do not prune model blobs on startup"},
-		"OLLAMA_NUM_PARALLEL":      {"OLLAMA_NUM_PARALLEL", NumParallel, "Maximum number of parallel requests (default 1)"},
+		"OLLAMA_NUM_PARALLEL":      {"OLLAMA_NUM_PARALLEL", NumParallel, "Maximum number of parallel requests"},
 		"OLLAMA_ORIGINS":           {"OLLAMA_ORIGINS", AllowOrigins, "A comma separated list of allowed origins"},
 		"OLLAMA_RUNNERS_DIR":       {"OLLAMA_RUNNERS_DIR", RunnersDir, "Location for runners"},
 		"OLLAMA_SCHED_SPREAD":      {"OLLAMA_SCHED_SPREAD", SchedSpread, "Always schedule model across all GPUs"},
@ -129,9 +131,10 @@ func clean(key string) string {

 func init() {
 	// default values
-	NumParallel = 1
-	MaxRunners = 1
+	NumParallel = 0 // Autoselect
+	MaxRunners = 0  // Autoselect
 	MaxQueuedRequests = 512
+	KeepAlive = 5 * time.Minute

 	LoadConfig()
 }
@ -205,8 +208,8 @@ func LoadConfig() {

 	if onp := clean("OLLAMA_NUM_PARALLEL"); onp != "" {
 		val, err := strconv.Atoi(onp)
-		if err != nil || val <= 0 {
-			slog.Error("invalid setting must be greater than zero", "OLLAMA_NUM_PARALLEL", onp, "error", err)
+		if err != nil {
+			slog.Error("invalid setting, ignoring", "OLLAMA_NUM_PARALLEL", onp, "error", err)
 		} else {
 			NumParallel = val
 		}
@ -251,7 +254,7 @@ func LoadConfig() {
 	if maxRunners != "" {
 		m, err := strconv.Atoi(maxRunners)
 		if err != nil {
-			slog.Error("invalid setting", "OLLAMA_MAX_LOADED_MODELS", maxRunners, "error", err)
+			slog.Error("invalid setting, ignoring", "OLLAMA_MAX_LOADED_MODELS", maxRunners, "error", err)
 		} else {
 			MaxRunners = m
 		}
@ -260,13 +263,16 @@ func LoadConfig() {
 	if onp := os.Getenv("OLLAMA_MAX_QUEUE"); onp != "" {
 		p, err := strconv.Atoi(onp)
 		if err != nil || p <= 0 {
-			slog.Error("invalid setting", "OLLAMA_MAX_QUEUE", onp, "error", err)
+			slog.Error("invalid setting, ignoring", "OLLAMA_MAX_QUEUE", onp, "error", err)
 		} else {
 			MaxQueuedRequests = p
 		}
 	}

-	KeepAlive = clean("OLLAMA_KEEP_ALIVE")
+	ka := clean("OLLAMA_KEEP_ALIVE")
+	if ka != "" {
+		loadKeepAlive(ka)
+	}

 	var err error
 	ModelsDir, err = getModelsDir()
@ -344,3 +350,24 @@ func getOllamaHost() (*OllamaHost, error) {
 		Port:   port,
 	}, nil
 }
+
+func loadKeepAlive(ka string) {
+	v, err := strconv.Atoi(ka)
+	if err != nil {
+		d, err := time.ParseDuration(ka)
+		if err == nil {
+			if d < 0 {
+				KeepAlive = time.Duration(math.MaxInt64)
+			} else {
+				KeepAlive = d
+			}
+		}
+	} else {
+		d := time.Duration(v) * time.Second
+		if d < 0 {
+			KeepAlive = time.Duration(math.MaxInt64)
+		} else {
+			KeepAlive = d
+		}
+	}
+}
--- a/envconfig/config_test.go
+++ b/envconfig/config_test.go
@ -2,8 +2,10 @@ package envconfig

 import (
 	"fmt"
+	"math"
 	"net"
 	"testing"
+	"time"

 	"github.com/stretchr/testify/assert"
 	"github.com/stretchr/testify/require"
@ -23,6 +25,21 @@ func TestConfig(t *testing.T) {
 	t.Setenv("OLLAMA_FLASH_ATTENTION", "1")
 	LoadConfig()
 	require.True(t, FlashAttention)
+	t.Setenv("OLLAMA_KEEP_ALIVE", "")
+	LoadConfig()
+	require.Equal(t, 5*time.Minute, KeepAlive)
+	t.Setenv("OLLAMA_KEEP_ALIVE", "3")
+	LoadConfig()
+	require.Equal(t, 3*time.Second, KeepAlive)
+	t.Setenv("OLLAMA_KEEP_ALIVE", "1h")
+	LoadConfig()
+	require.Equal(t, 1*time.Hour, KeepAlive)
+	t.Setenv("OLLAMA_KEEP_ALIVE", "-1s")
+	LoadConfig()
+	require.Equal(t, time.Duration(math.MaxInt64), KeepAlive)
+	t.Setenv("OLLAMA_KEEP_ALIVE", "-1")
+	LoadConfig()
+	require.Equal(t, time.Duration(math.MaxInt64), KeepAlive)
 }

 func TestClientFromEnvironment(t *testing.T) {
--- a/gpu/amd_windows.go
+++ b/gpu/amd_windows.go
@ -115,8 +115,6 @@ func AMDGetGPUInfo() []RocmGPUInfo {
 			continue
 		}

-		// TODO revisit this once ROCm v6 is available on windows.
-		// v5.7 only reports VRAM used by this process, so it's completely wrong and unusable
 		slog.Debug("amdgpu memory", "gpu", i, "total", format.HumanBytes2(totalMemory))
 		slog.Debug("amdgpu memory", "gpu", i, "available", format.HumanBytes2(freeMemory))
 		gpuInfo := RocmGPUInfo{
@ -126,6 +124,9 @@ func AMDGetGPUInfo() []RocmGPUInfo {
 					TotalMemory: totalMemory,
 					FreeMemory:  freeMemory,
 				},
+				// Free memory reporting on Windows is not reliable until we bump to ROCm v6.2
+				UnreliableFreeMemory: true,
+
 				ID:             strconv.Itoa(i), // TODO this is probably wrong if we specify visible devices
 				DependencyPath: libDir,
 				MinimumMemory:  rocmMinimumMemory,
--- a/gpu/gpu.go
+++ b/gpu/gpu.go
@ -202,7 +202,7 @@ func GetGPUInfo() GpuInfoList {
 	}()

 	if !bootstrapped {
-		slog.Debug("Detecting GPUs")
+		slog.Info("looking for compatible GPUs")
 		needRefresh = false
 		cpuCapability = GetCPUCapability()
 		var memInfo C.mem_info_t
@ -320,6 +320,9 @@ func GetGPUInfo() GpuInfoList {

 		rocmGPUs = AMDGetGPUInfo()
 		bootstrapped = true
+		if len(cudaGPUs) == 0 && len(rocmGPUs) == 0 && len(oneapiGPUs) == 0 {
+			slog.Info("no compatible GPUs were discovered")
+		}
 	}

 	// For detected GPUs, load library if not loaded
@ -514,7 +517,23 @@ func LoadNVCUDAMgmt(nvcudaLibPaths []string) (int, *C.nvcuda_handle_t, string) {
 		defer C.free(unsafe.Pointer(lib))
 		C.nvcuda_init(lib, &resp)
 		if resp.err != nil {
-			slog.Debug("Unable to load nvcuda", "library", libPath, "error", C.GoString(resp.err))
+			// Decide what log level based on the type of error message to help users understand why
+			msg := C.GoString(resp.err)
+			switch resp.cudaErr {
+			case C.CUDA_ERROR_INSUFFICIENT_DRIVER, C.CUDA_ERROR_SYSTEM_DRIVER_MISMATCH:
+				slog.Warn("version mismatch between driver and cuda driver library - reboot or upgrade may be required", "library", libPath, "error", msg)
+			case C.CUDA_ERROR_NO_DEVICE:
+				slog.Info("no nvidia devices detected", "library", libPath)
+			case C.CUDA_ERROR_UNKNOWN:
+				slog.Warn("unknown error initializing cuda driver library", "library", libPath, "error", msg)
+				slog.Warn("see https://github.com/ollama/ollama/blob/main/docs/troubleshooting.md for more information")
+			default:
+				if strings.Contains(msg, "wrong ELF class") {
+					slog.Debug("skipping 32bit library", "library", libPath)
+				} else {
+					slog.Info("unable to load cuda driver library", "library", libPath, "error", msg)
+				}
+			}
 			C.free(unsafe.Pointer(resp.err))
 		} else {
 			return int(resp.num_devices), &resp.ch, libPath
--- a/gpu/gpu_info_nvcuda.c
+++ b/gpu/gpu_info_nvcuda.c
@ -7,6 +7,7 @@ void nvcuda_init(char *nvcuda_lib_path, nvcuda_init_resp_t *resp) {
  CUresult ret;
  resp->err = NULL;
  resp->num_devices = 0;
+  resp->cudaErr = CUDA_SUCCESS;
  const int buflen = 256;
  char buf[buflen + 1];
  int i;
@ -38,6 +39,7 @@ void nvcuda_init(char *nvcuda_lib_path, nvcuda_init_resp_t *resp) {
            nvcuda_lib_path, msg);
    free(msg);
    resp->err = strdup(buf);
+    resp->cudaErr = -1;
    return;
  }

@ -52,6 +54,7 @@ void nvcuda_init(char *nvcuda_lib_path, nvcuda_init_resp_t *resp) {
              msg);
      free(msg);
      resp->err = strdup(buf);
+      resp->cudaErr = -1;
      return;
    }
  }
@ -61,12 +64,9 @@ void nvcuda_init(char *nvcuda_lib_path, nvcuda_init_resp_t *resp) {
    LOG(resp->ch.verbose, "cuInit err: %d\n", ret);
    UNLOAD_LIBRARY(resp->ch.handle);
    resp->ch.handle = NULL;
-    if (ret == CUDA_ERROR_INSUFFICIENT_DRIVER) {
-      resp->err = strdup("your nvidia driver is too old or missing.  If you have a CUDA GPU please upgrade to run ollama");
-      return;
-    }
-    snprintf(buf, buflen, "nvcuda init failure: %d", ret);
+    snprintf(buf, buflen, "cuda driver library init failure: %d", ret);
    resp->err = strdup(buf);
+    resp->cudaErr = ret;
    return;
  }

@ -91,6 +91,7 @@ void nvcuda_init(char *nvcuda_lib_path, nvcuda_init_resp_t *resp) {
    resp->ch.handle = NULL;
    snprintf(buf, buflen, "unable to get device count: %d", ret);
    resp->err = strdup(buf);
+    resp->cudaErr = ret;
    return;
  }
 }
@ -106,13 +107,13 @@ void nvcuda_bootstrap(nvcuda_handle_t h, int i, mem_info_t *resp) {
  CUuuid uuid = {0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0};

  if (h.handle == NULL) {
-    resp->err = strdup("nvcuda handle isn't initialized");
+    resp->err = strdup("cuda driver library handle isn't initialized");
    return;
  }

  ret = (*h.cuDeviceGet)(&device, i);
  if (ret != CUDA_SUCCESS) {
-    snprintf(buf, buflen, "nvcuda device failed to initialize");
+    snprintf(buf, buflen, "cuda driver library device failed to initialize");
    resp->err = strdup(buf);
    return;
  }
@ -168,14 +169,14 @@ void nvcuda_bootstrap(nvcuda_handle_t h, int i, mem_info_t *resp) {
  // To get memory we have to set (and release) a context
  ret = (*h.cuCtxCreate_v3)(&ctx, NULL, 0, 0, device);
  if (ret != CUDA_SUCCESS) {
-    snprintf(buf, buflen, "nvcuda failed to get device context %d", ret);
+    snprintf(buf, buflen, "cuda driver library failed to get device context %d", ret);
    resp->err = strdup(buf);
    return;
  }

  ret = (*h.cuMemGetInfo_v2)(&memInfo.free, &memInfo.total);
  if (ret != CUDA_SUCCESS) {
-    snprintf(buf, buflen, "nvcuda device memory info lookup failure %d", ret);
+    snprintf(buf, buflen, "cuda driver library device memory info lookup failure %d", ret);
    resp->err = strdup(buf);
    // Best effort on failure...
    (*h.cuCtxDestroy)(ctx);
@ -193,7 +194,7 @@ void nvcuda_bootstrap(nvcuda_handle_t h, int i, mem_info_t *resp) {

  ret = (*h.cuCtxDestroy)(ctx);
  if (ret != CUDA_SUCCESS) {
-    LOG(1, "nvcuda failed to release device context %d", ret);
+    LOG(1, "cuda driver library failed to release device context %d", ret);
  }
 }

@ -206,7 +207,7 @@ void nvcuda_get_free(nvcuda_handle_t h, int i, uint64_t *free, uint64_t *total)

  ret = (*h.cuDeviceGet)(&device, i);
  if (ret != CUDA_SUCCESS) {
-    LOG(1, "nvcuda device failed to initialize");
+    LOG(1, "cuda driver library device failed to initialize");
    return;
  }

@ -214,13 +215,13 @@ void nvcuda_get_free(nvcuda_handle_t h, int i, uint64_t *free, uint64_t *total)
  // To get memory we have to set (and release) a context
  ret = (*h.cuCtxCreate_v3)(&ctx, NULL, 0, 0, device);
  if (ret != CUDA_SUCCESS) {
-    LOG(1, "nvcuda failed to get device context %d", ret);
+    LOG(1, "cuda driver library failed to get device context %d", ret);
    return;
  }

  ret = (*h.cuMemGetInfo_v2)(free, total);
  if (ret != CUDA_SUCCESS) {
-    LOG(1, "nvcuda device memory info lookup failure %d", ret);
+    LOG(1, "cuda driver library device memory info lookup failure %d", ret);
    // Best effort on failure...
    (*h.cuCtxDestroy)(ctx);
    return;
@ -228,12 +229,12 @@ void nvcuda_get_free(nvcuda_handle_t h, int i, uint64_t *free, uint64_t *total)

  ret = (*h.cuCtxDestroy)(ctx);
  if (ret != CUDA_SUCCESS) {
-    LOG(1, "nvcuda failed to release device context %d", ret);
+    LOG(1, "cuda driver library failed to release device context %d", ret);
  }
 }

 void nvcuda_release(nvcuda_handle_t h) {
-  LOG(h.verbose, "releasing nvcuda library\n");
+  LOG(h.verbose, "releasing cuda driver library\n");
  UNLOAD_LIBRARY(h.handle);
  // TODO and other context release logic?
  h.handle = NULL;
--- a/gpu/gpu_info_nvcuda.h
+++ b/gpu/gpu_info_nvcuda.h
@ -7,9 +7,12 @@
 typedef enum cudaError_enum {
  CUDA_SUCCESS = 0,
  CUDA_ERROR_INVALID_VALUE = 1,
-  CUDA_ERROR_MEMORY_ALLOCATION = 2,
+  CUDA_ERROR_OUT_OF_MEMORY = 2,
  CUDA_ERROR_NOT_INITIALIZED = 3,
  CUDA_ERROR_INSUFFICIENT_DRIVER = 35,
+  CUDA_ERROR_NO_DEVICE = 100,
+  CUDA_ERROR_SYSTEM_DRIVER_MISMATCH = 803,
+  CUDA_ERROR_UNKNOWN = 999,
  // Other values omitted for now...
 } CUresult;

@ -64,6 +67,7 @@ typedef struct nvcuda_init_resp {
  char *err;  // If err is non-null handle is invalid
  nvcuda_handle_t ch;
  int num_devices;
+  CUresult cudaErr;
 } nvcuda_init_resp_t;

 void nvcuda_init(char *nvcuda_lib_path, nvcuda_init_resp_t *resp);
--- a/gpu/types.go
+++ b/gpu/types.go
@ -29,6 +29,11 @@ type GpuInfo struct {
 	// 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
+	// False indicates FreeMemory can generally be trusted on this GPU
+	UnreliableFreeMemory bool
+
 	// GPU information
 	ID      string `json:"gpu_id"`  // string to use for selection of this specific GPU
 	Name    string `json:"name"`    // user friendly name if available
--- a/llm/ext_server/CMakeLists.txt
+++ b/llm/ext_server/CMakeLists.txt
@ -1,14 +1,13 @@
-
-set(TARGET ollama_llama_server)
-option(LLAMA_SERVER_VERBOSE "Build verbose logging option for Server" ON)
-include_directories(${CMAKE_CURRENT_SOURCE_DIR})
-add_executable(${TARGET} server.cpp utils.hpp json.hpp httplib.h)
-install(TARGETS ${TARGET} RUNTIME)
-target_compile_definitions(${TARGET} PRIVATE
-    SERVER_VERBOSE=$<BOOL:${LLAMA_SERVER_VERBOSE}>
-)
-target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT})
-if (WIN32)
-    TARGET_LINK_LIBRARIES(${TARGET} PRIVATE ws2_32)
-endif()
+set(TARGET ollama_llama_server)
+option(LLAMA_SERVER_VERBOSE "Build verbose logging option for Server" ON)
+include_directories(${CMAKE_CURRENT_SOURCE_DIR})
+add_executable(${TARGET} server.cpp utils.hpp json.hpp httplib.h)
+install(TARGETS ${TARGET} RUNTIME)
+target_compile_definitions(${TARGET} PRIVATE
+    SERVER_VERBOSE=$<BOOL:${LLAMA_SERVER_VERBOSE}>
+)
+target_link_libraries(${TARGET} PRIVATE ggml llama common llava ${CMAKE_THREAD_LIBS_INIT})
+if (WIN32)
+    TARGET_LINK_LIBRARIES(${TARGET} PRIVATE ws2_32)
+endif()
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
--- a/llm/ext_server/server.cpp
+++ b/llm/ext_server/server.cpp
@ -1382,12 +1382,50 @@ struct llama_server_context
        }
    }

+    std::string common_prefix(const std::string& str1, const std::string& str2) {
+        auto mismatch_pair = std::mismatch(str1.begin(), str1.end(), str2.begin());
+        return std::string(str1.begin(), mismatch_pair.first);
+    }
+
+    // Find the slot that has the greatest common prefix
+    server_slot *prefix_slot(const json &prompt) {
+        if (!prompt.is_string()) {
+            return nullptr;
+        }
+
+        std::string prompt_str = prompt.get<std::string>();
+        server_slot *slot = nullptr;
+        size_t longest = 0;
+
+        for (server_slot &s : slots) {
+            if (s.available() && s.prompt.is_string()) {
+                std::string s_prompt = s.prompt.get<std::string>();
+                std::string prefix = common_prefix(s_prompt, prompt_str);
+
+                if (prefix.size() > longest) {
+                    slot = &s;
+                    longest = prefix.size();
+                }
+            }
+        }
+
+        if (!slot) {
+            return get_slot(-1);
+        }
+
+        LOG_INFO("slot with common prefix found", {{
+            "slot_id", slot->id,
+            "characters", longest
+        }});
+        return slot;
+    }
+
    void process_single_task(task_server& task)
    {
        switch (task.type)
        {
            case TASK_TYPE_COMPLETION: {
-                server_slot *slot = get_slot(json_value(task.data, "slot_id", -1));
+                server_slot *slot = prefix_slot(task.data["prompt"]);
                if (slot == nullptr)
                {
                    // if no slot is available, we defer this task for processing later
@ -1732,7 +1770,7 @@ struct llama_server_context
                            slot.n_past -= 1;
                        }

-                        slot.n_prompt_tokens_processed = slot.n_prompt_tokens - slot.n_past;
+                        slot.n_prompt_tokens_processed = slot.n_prompt_tokens;

                        if (slot.ga_n != 1)
                        {
--- a/llm/generate/gen_darwin.sh
+++ b/llm/generate/gen_darwin.sh
@ -18,16 +18,16 @@ sign() {
    fi
 }

-COMMON_DARWIN_DEFS="-DCMAKE_OSX_DEPLOYMENT_TARGET=11.3 -DLLAMA_METAL_MACOSX_VERSION_MIN=11.3 -DCMAKE_SYSTEM_NAME=Darwin -DLLAMA_METAL_EMBED_LIBRARY=on -DLLAMA_OPENMP=off"
+COMMON_DARWIN_DEFS="-DBUILD_SHARED_LIBS=off -DCMAKE_OSX_DEPLOYMENT_TARGET=11.3 -DLLAMA_METAL_MACOSX_VERSION_MIN=11.3 -DCMAKE_SYSTEM_NAME=Darwin -DGGML_METAL_EMBED_LIBRARY=on -DGGML_OPENMP=off"

 case "${GOARCH}" in
 "amd64")
-    COMMON_CPU_DEFS="${COMMON_DARWIN_DEFS} -DCMAKE_SYSTEM_PROCESSOR=${ARCH} -DCMAKE_OSX_ARCHITECTURES=${ARCH} -DLLAMA_METAL=off -DLLAMA_NATIVE=off"
+    COMMON_CPU_DEFS="${COMMON_DARWIN_DEFS} -DCMAKE_SYSTEM_PROCESSOR=${ARCH} -DCMAKE_OSX_ARCHITECTURES=${ARCH} -DGGML_METAL=off -DGGML_NATIVE=off"

    # Static build for linking into the Go binary
    init_vars
    CMAKE_TARGETS="--target llama --target ggml"
-    CMAKE_DEFS="${COMMON_CPU_DEFS} -DBUILD_SHARED_LIBS=off -DLLAMA_BLAS=off -DLLAMA_ACCELERATE=off -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
+    CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_BLAS=off -DGGML_ACCELERATE=off -DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off ${CMAKE_DEFS}"
    BUILD_DIR="../build/darwin/${ARCH}_static"
    echo "Building static library"
    build
@ -37,7 +37,7 @@ case "${GOARCH}" in
        # CPU first for the default library, set up as lowest common denominator for maximum compatibility (including Rosetta)
        #
        init_vars
-        CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_ACCELERATE=off -DLLAMA_BLAS=off -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
+        CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_ACCELERATE=off -DGGML_BLAS=off -DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off ${CMAKE_DEFS}"
        BUILD_DIR="../build/darwin/${ARCH}/cpu"
        echo "Building LCD CPU"
        build
@ -49,7 +49,7 @@ case "${GOARCH}" in
        # Approximately 400% faster than LCD on same CPU
        #
        init_vars
-        CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_ACCELERATE=off -DLLAMA_BLAS=off -DLLAMA_AVX=on -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
+        CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_ACCELERATE=off -DGGML_BLAS=off -DGGML_AVX=on -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off ${CMAKE_DEFS}"
        BUILD_DIR="../build/darwin/${ARCH}/cpu_avx"
        echo "Building AVX CPU"
        build
@ -61,7 +61,7 @@ case "${GOARCH}" in
        # Approximately 10% faster than AVX on same CPU
        #
        init_vars
-        CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_ACCELERATE=on -DLLAMA_BLAS=off -DLLAMA_AVX=on -DLLAMA_AVX2=on -DLLAMA_AVX512=off -DLLAMA_FMA=on -DLLAMA_F16C=on ${CMAKE_DEFS}"
+        CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_ACCELERATE=on -DGGML_BLAS=off -DGGML_AVX=on -DGGML_AVX2=on -DGGML_AVX512=off -DGGML_FMA=on -DGGML_F16C=on ${CMAKE_DEFS}"
        BUILD_DIR="../build/darwin/${ARCH}/cpu_avx2"
        echo "Building AVX2 CPU"
        EXTRA_LIBS="${EXTRA_LIBS} -framework Accelerate -framework Foundation"
@ -75,14 +75,14 @@ case "${GOARCH}" in
    # Static build for linking into the Go binary
    init_vars
    CMAKE_TARGETS="--target llama --target ggml"
-    CMAKE_DEFS="-DCMAKE_OSX_DEPLOYMENT_TARGET=11.3 -DLLAMA_BLAS=off -DCMAKE_SYSTEM_NAME=Darwin -DBUILD_SHARED_LIBS=off -DCMAKE_SYSTEM_PROCESSOR=${ARCH} -DCMAKE_OSX_ARCHITECTURES=${ARCH} -DLLAMA_METAL=off -DLLAMA_ACCELERATE=off -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
+    CMAKE_DEFS="${COMMON_DARWIN_DEFS} -DCMAKE_OSX_DEPLOYMENT_TARGET=11.3 -DCMAKE_SYSTEM_NAME=Darwin -DCMAKE_SYSTEM_PROCESSOR=${ARCH} -DCMAKE_OSX_ARCHITECTURES=${ARCH} ${CMAKE_DEFS}"
    BUILD_DIR="../build/darwin/${ARCH}_static"
    echo "Building static library"
    build

    if [ -z "$OLLAMA_SKIP_METAL_GENERATE" ]; then
        init_vars
-        CMAKE_DEFS="${COMMON_DARWIN_DEFS} -DLLAMA_ACCELERATE=on -DCMAKE_SYSTEM_PROCESSOR=${ARCH} -DCMAKE_OSX_ARCHITECTURES=${ARCH} -DLLAMA_METAL=on ${CMAKE_DEFS}"
+        CMAKE_DEFS="${COMMON_DARWIN_DEFS} -DCMAKE_SYSTEM_PROCESSOR=${ARCH} -DCMAKE_OSX_ARCHITECTURES=${ARCH} ${CMAKE_DEFS}"
        BUILD_DIR="../build/darwin/${ARCH}/metal"
        EXTRA_LIBS="${EXTRA_LIBS} -framework Accelerate -framework Foundation -framework Metal -framework MetalKit -framework MetalPerformanceShaders"
        build
--- a/llm/generate/gen_linux.sh
+++ b/llm/generate/gen_linux.sh
@ -51,7 +51,7 @@ if [ -z "${CUDACXX}" ]; then
        export CUDACXX=$(command -v nvcc)
    fi
 fi
-COMMON_CMAKE_DEFS="-DCMAKE_POSITION_INDEPENDENT_CODE=on -DLLAMA_NATIVE=off -DLLAMA_AVX=on -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off -DLLAMA_OPENMP=off"
+COMMON_CMAKE_DEFS="-DBUILD_SHARED_LIBS=off -DCMAKE_POSITION_INDEPENDENT_CODE=on -DGGML_NATIVE=off -DGGML_AVX=on -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off -DGGML_OPENMP=off"
 source $(dirname $0)/gen_common.sh
 init_vars
 git_module_setup
@ -64,7 +64,7 @@ if [ -z "${OLLAMA_SKIP_STATIC_GENERATE}" -o "${OLLAMA_CPU_TARGET}" = "static" ];
    # Static build for linking into the Go binary
    init_vars
    CMAKE_TARGETS="--target llama --target ggml"
-    CMAKE_DEFS="-DBUILD_SHARED_LIBS=off -DLLAMA_NATIVE=off -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off -DLLAMA_OPENMP=off ${CMAKE_DEFS}"
+    CMAKE_DEFS="-DBUILD_SHARED_LIBS=off -DGGML_NATIVE=off -DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off -DGGML_OPENMP=off ${CMAKE_DEFS}"
    BUILD_DIR="../build/linux/${ARCH}_static"
    echo "Building static library"
    build
@ -84,22 +84,22 @@ if [ -z "${OLLAMA_SKIP_CPU_GENERATE}" ]; then
        compress
    else
        # Darwin Rosetta x86 emulation does NOT support AVX, AVX2, AVX512
-        # -DLLAMA_AVX -- 2011 Intel Sandy Bridge & AMD Bulldozer
-        # -DLLAMA_F16C -- 2012 Intel Ivy Bridge & AMD 2011 Bulldozer (No significant improvement over just AVX)
-        # -DLLAMA_AVX2 -- 2013 Intel Haswell & 2015 AMD Excavator / 2017 AMD Zen
-        # -DLLAMA_FMA (FMA3) -- 2013 Intel Haswell & 2012 AMD Piledriver
+        # -DGGML_AVX -- 2011 Intel Sandy Bridge & AMD Bulldozer
+        # -DGGML_F16C -- 2012 Intel Ivy Bridge & AMD 2011 Bulldozer (No significant improvement over just AVX)
+        # -DGGML_AVX2 -- 2013 Intel Haswell & 2015 AMD Excavator / 2017 AMD Zen
+        # -DGGML_FMA (FMA3) -- 2013 Intel Haswell & 2012 AMD Piledriver
        # Note: the following seem to yield slower results than AVX2 - ymmv
-        # -DLLAMA_AVX512 -- 2017 Intel Skylake and High End DeskTop (HEDT)
-        # -DLLAMA_AVX512_VBMI -- 2018 Intel Cannon Lake
-        # -DLLAMA_AVX512_VNNI -- 2021 Intel Alder Lake
+        # -DGGML_AVX512 -- 2017 Intel Skylake and High End DeskTop (HEDT)
+        # -DGGML_AVX512_VBMI -- 2018 Intel Cannon Lake
+        # -DGGML_AVX512_VNNI -- 2021 Intel Alder Lake

-        COMMON_CPU_DEFS="-DCMAKE_POSITION_INDEPENDENT_CODE=on -DLLAMA_NATIVE=off -DLLAMA_OPENMP=off"
+        COMMON_CPU_DEFS="-DCMAKE_POSITION_INDEPENDENT_CODE=on -DGGML_NATIVE=off -DGGML_OPENMP=off"
        if [ -z "${OLLAMA_CPU_TARGET}" -o "${OLLAMA_CPU_TARGET}" = "cpu" ]; then
            #
            # CPU first for the default library, set up as lowest common denominator for maximum compatibility (including Rosetta)
            #
            init_vars
-            CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
+            CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off ${CMAKE_DEFS}"
            BUILD_DIR="../build/linux/${ARCH}/cpu"
            echo "Building LCD CPU"
            build
@ -116,7 +116,7 @@ if [ -z "${OLLAMA_SKIP_CPU_GENERATE}" ]; then
                # Approximately 400% faster than LCD on same CPU
                #
                init_vars
-                CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_AVX=on -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
+                CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_AVX=on -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off ${CMAKE_DEFS}"
                BUILD_DIR="../build/linux/${ARCH}/cpu_avx"
                echo "Building AVX CPU"
                build
@ -129,7 +129,7 @@ if [ -z "${OLLAMA_SKIP_CPU_GENERATE}" ]; then
                # Approximately 10% faster than AVX on same CPU
                #
                init_vars
-                CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_AVX=on -DLLAMA_AVX2=on -DLLAMA_AVX512=off -DLLAMA_FMA=on -DLLAMA_F16C=on ${CMAKE_DEFS}"
+                CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_AVX=on -DGGML_AVX2=on -DGGML_AVX512=off -DGGML_FMA=on -DGGML_F16C=on ${CMAKE_DEFS}"
                BUILD_DIR="../build/linux/${ARCH}/cpu_avx2"
                echo "Building AVX2 CPU"
                build
@ -170,15 +170,15 @@ if [ -z "${OLLAMA_SKIP_CUDA_GENERATE}" -a -d "${CUDA_LIB_DIR}" ]; then
        #
        # CUDA compute < 6.0 lacks proper FP16 support on ARM.
        # Disabling has minimal performance effect while maintaining compatibility.
-        ARM64_DEFS="-DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_CUDA_F16=off"
+        ARM64_DEFS="-DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_CUDA_F16=off"
    fi
    # Users building from source can tune the exact flags we pass to cmake for configuring llama.cpp
    if [ -n "${OLLAMA_CUSTOM_CUDA_DEFS}" ]; then
        echo "OLLAMA_CUSTOM_CUDA_DEFS=\"${OLLAMA_CUSTOM_CUDA_DEFS}\""
-        CMAKE_CUDA_DEFS="-DLLAMA_CUDA=on -DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES} ${OLLAMA_CUSTOM_CUDA_DEFS}"
+        CMAKE_CUDA_DEFS="-DGGML_CUDA=on -DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES} ${OLLAMA_CUSTOM_CUDA_DEFS}"
        echo "Building custom CUDA GPU"
    else
-        CMAKE_CUDA_DEFS="-DLLAMA_CUDA=on -DCMAKE_CUDA_FLAGS=-t8 -DLLAMA_CUDA_FORCE_MMQ=on -DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES}"
+        CMAKE_CUDA_DEFS="-DGGML_CUDA=on -DCMAKE_CUDA_FLAGS=-t8 -DGGML_CUDA_FORCE_MMQ=on -DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES} -DCMAKE_LIBRARY_PATH=/usr/local/cuda/compat"
    fi
    CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} ${ARM64_DEFS} ${CMAKE_CUDA_DEFS}"
    BUILD_DIR="../build/linux/${ARCH}/cuda${CUDA_VARIANT}"
@ -216,7 +216,7 @@ if [ -z "${OLLAMA_SKIP_ONEAPI_GENERATE}" -a -d "${ONEAPI_ROOT}" ]; then
    init_vars
    source ${ONEAPI_ROOT}/setvars.sh --force # set up environment variables for oneAPI
    CC=icx
-    CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_SYCL=ON -DLLAMA_SYCL_F16=OFF"
+    CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL=ON -DGGML_SYCL_F16=OFF"
    BUILD_DIR="../build/linux/${ARCH}/oneapi"
    EXTRA_LIBS="-fsycl -Wl,-rpath,${ONEAPI_ROOT}/compiler/latest/lib,-rpath,${ONEAPI_ROOT}/mkl/latest/lib,-rpath,${ONEAPI_ROOT}/tbb/latest/lib,-rpath,${ONEAPI_ROOT}/compiler/latest/opt/oclfpga/linux64/lib -lOpenCL -lmkl_core -lmkl_sycl_blas -lmkl_intel_ilp64 -lmkl_tbb_thread -ltbb"
    DEBUG_FLAGS="" # icx compiles with -O0 if we pass -g, so we must remove it
@ -254,7 +254,7 @@ if [ -z "${OLLAMA_SKIP_ROCM_GENERATE}" -a -d "${ROCM_PATH}" ]; then
        ROCM_VARIANT=_v$(ls ${ROCM_PATH}/lib/librocblas.so.*.*.????? | cut -f5 -d. || true)
    fi
    init_vars
-    CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} -DLLAMA_HIPBLAS=on -DCMAKE_C_COMPILER=$ROCM_PATH/llvm/bin/clang -DCMAKE_CXX_COMPILER=$ROCM_PATH/llvm/bin/clang++ -DAMDGPU_TARGETS=$(amdGPUs) -DGPU_TARGETS=$(amdGPUs)"
+    CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} -DGGML_HIPBLAS=on -DCMAKE_C_COMPILER=$ROCM_PATH/llvm/bin/clang -DCMAKE_CXX_COMPILER=$ROCM_PATH/llvm/bin/clang++ -DAMDGPU_TARGETS=$(amdGPUs) -DGPU_TARGETS=$(amdGPUs)"
    # Users building from source can tune the exact flags we pass to cmake for configuring llama.cpp
    if [ -n "${OLLAMA_CUSTOM_ROCM_DEFS}" ]; then
        echo "OLLAMA_CUSTOM_ROCM_DEFS=\"${OLLAMA_CUSTOM_ROCM_DEFS}\""
--- a/llm/generate/gen_windows.ps1
+++ b/llm/generate/gen_windows.ps1
@ -39,8 +39,8 @@ function init_vars {
    }
    $script:cmakeDefs = @(
        "-DBUILD_SHARED_LIBS=on",
-        "-DLLAMA_NATIVE=off",
-        "-DLLAMA_OPENMP=off"
+        "-DGGML_NATIVE=off",
+        "-DGGML_OPENMP=off"
        )
    $script:commonCpuDefs = @("-DCMAKE_POSITION_INDEPENDENT_CODE=on")
    $script:ARCH = $Env:PROCESSOR_ARCHITECTURE.ToLower()
@ -182,9 +182,9 @@ function cleanup {
 }


-# -DLLAMA_AVX -- 2011 Intel Sandy Bridge & AMD Bulldozer
-# -DLLAMA_AVX2 -- 2013 Intel Haswell & 2015 AMD Excavator / 2017 AMD Zen
-# -DLLAMA_FMA (FMA3) -- 2013 Intel Haswell & 2012 AMD Piledriver
+# -DGGML_AVX -- 2011 Intel Sandy Bridge & AMD Bulldozer
+# -DGGML_AVX2 -- 2013 Intel Haswell & 2015 AMD Excavator / 2017 AMD Zen
+# -DGGML_FMA (FMA3) -- 2013 Intel Haswell & 2012 AMD Piledriver


 function build_static() {
@ -204,13 +204,13 @@ function build_static() {
            "-DCMAKE_C_COMPILER=gcc.exe",
            "-DCMAKE_CXX_COMPILER=g++.exe",
            "-DBUILD_SHARED_LIBS=off",
-            "-DLLAMA_NATIVE=off",
-            "-DLLAMA_AVX=off",
-            "-DLLAMA_AVX2=off",
-            "-DLLAMA_AVX512=off",
-            "-DLLAMA_F16C=off",
-            "-DLLAMA_FMA=off",
-            "-DLLAMA_OPENMP=off")
+            "-DGGML_NATIVE=off",
+            "-DGGML_AVX=off",
+            "-DGGML_AVX2=off",
+            "-DGGML_AVX512=off",
+            "-DGGML_F16C=off",
+            "-DGGML_FMA=off",
+            "-DGGML_OPENMP=off")
        $script:buildDir="../build/windows/${script:ARCH}_static"
        write-host "Building static library"
        build
@ -224,7 +224,7 @@ function build_cpu($gen_arch) {
    if ((-not "${env:OLLAMA_SKIP_CPU_GENERATE}" ) -and ((-not "${env:OLLAMA_CPU_TARGET}") -or ("${env:OLLAMA_CPU_TARGET}" -eq "cpu"))) {
        # remaining llama.cpp builds use MSVC 
        init_vars
-        $script:cmakeDefs = $script:commonCpuDefs + @("-A", $gen_arch, "-DLLAMA_AVX=off", "-DLLAMA_AVX2=off", "-DLLAMA_AVX512=off", "-DLLAMA_FMA=off", "-DLLAMA_F16C=off") + $script:cmakeDefs
+        $script:cmakeDefs = $script:commonCpuDefs + @("-A", $gen_arch, "-DGGML_AVX=off", "-DGGML_AVX2=off", "-DGGML_AVX512=off", "-DGGML_FMA=off", "-DGGML_F16C=off") + $script:cmakeDefs
        $script:buildDir="../build/windows/${script:ARCH}/cpu"
        $script:distDir="$script:DIST_BASE\cpu"
        write-host "Building LCD CPU"
@ -239,7 +239,7 @@ function build_cpu($gen_arch) {
 function build_cpu_avx() {
    if ((-not "${env:OLLAMA_SKIP_CPU_GENERATE}" ) -and ((-not "${env:OLLAMA_CPU_TARGET}") -or ("${env:OLLAMA_CPU_TARGET}" -eq "cpu_avx"))) {
        init_vars
-        $script:cmakeDefs = $script:commonCpuDefs + @("-A", "x64", "-DLLAMA_AVX=on", "-DLLAMA_AVX2=off", "-DLLAMA_AVX512=off", "-DLLAMA_FMA=off", "-DLLAMA_F16C=off") + $script:cmakeDefs
+        $script:cmakeDefs = $script:commonCpuDefs + @("-A", "x64", "-DGGML_AVX=on", "-DGGML_AVX2=off", "-DGGML_AVX512=off", "-DGGML_FMA=off", "-DGGML_F16C=off") + $script:cmakeDefs
        $script:buildDir="../build/windows/${script:ARCH}/cpu_avx"
        $script:distDir="$script:DIST_BASE\cpu_avx"
        write-host "Building AVX CPU"
@ -254,7 +254,7 @@ function build_cpu_avx() {
 function build_cpu_avx2() {
    if ((-not "${env:OLLAMA_SKIP_CPU_GENERATE}" ) -and ((-not "${env:OLLAMA_CPU_TARGET}") -or ("${env:OLLAMA_CPU_TARGET}" -eq "cpu_avx2"))) {
        init_vars
-        $script:cmakeDefs = $script:commonCpuDefs + @("-A", "x64", "-DLLAMA_AVX=on", "-DLLAMA_AVX2=on", "-DLLAMA_AVX512=off", "-DLLAMA_FMA=on", "-DLLAMA_F16C=on") + $script:cmakeDefs
+        $script:cmakeDefs = $script:commonCpuDefs + @("-A", "x64", "-DGGML_AVX=on", "-DGGML_AVX2=on", "-DGGML_AVX512=off", "-DGGML_FMA=on", "-DGGML_F16C=on") + $script:cmakeDefs
        $script:buildDir="../build/windows/${script:ARCH}/cpu_avx2"
        $script:distDir="$script:DIST_BASE\cpu_avx2"
        write-host "Building AVX2 CPU"
@ -279,9 +279,9 @@ function build_cuda() {
        $script:distDir="$script:DIST_BASE\cuda$script:CUDA_VARIANT"
        $script:cmakeDefs += @(
            "-A", "x64",
-            "-DLLAMA_CUDA=ON",
-            "-DLLAMA_AVX=on",
-            "-DLLAMA_AVX2=off",
+            "-DGGML_CUDA=ON",
+            "-DGGML_AVX=on",
+            "-DGGML_AVX2=off",
            "-DCUDAToolkit_INCLUDE_DIR=$script:CUDA_INCLUDE_DIR",
            "-DCMAKE_CUDA_FLAGS=-t8",
            "-DCMAKE_CUDA_ARCHITECTURES=${script:CMAKE_CUDA_ARCHITECTURES}"
@ -319,7 +319,7 @@ function build_oneapi() {
    $script:distDir ="$script:DIST_BASE\oneapi$script:ONEAPI_VARIANT"
    $script:cmakeDefs += @(
      "-G", "MinGW Makefiles",
-      "-DLLAMA_SYCL=ON",
+      "-DGGML_SYCL=ON",
      "-DCMAKE_C_COMPILER=icx",
      "-DCMAKE_CXX_COMPILER=icx",
      "-DCMAKE_BUILD_TYPE=Release"
@ -365,10 +365,10 @@ function build_rocm() {
            "-G", "Ninja", 
            "-DCMAKE_C_COMPILER=clang.exe",
            "-DCMAKE_CXX_COMPILER=clang++.exe",
-            "-DLLAMA_HIPBLAS=on",
+            "-DGGML_HIPBLAS=on",
            "-DHIP_PLATFORM=amd",
-            "-DLLAMA_AVX=on",
-            "-DLLAMA_AVX2=off",
+            "-DGGML_AVX=on",
+            "-DGGML_AVX2=off",
            "-DCMAKE_POSITION_INDEPENDENT_CODE=on",
            "-DAMDGPU_TARGETS=$(amdGPUs)",
            "-DGPU_TARGETS=$(amdGPUs)"
--- a/llm/llama.cpp
+++ b/llm/llama.cpp
@ -1 +1 @@
-Subproject commit 7c26775adb579e92b59c82e8084c07a1d0f75e9c
+Subproject commit d7fd29fff16456ce9c3a23fd2d09a66256b05aff
--- a/llm/llm.go
+++ b/llm/llm.go
@ -1,12 +1,14 @@
 package llm

-// #cgo CFLAGS: -Illama.cpp
-// #cgo darwin,arm64 LDFLAGS: ${SRCDIR}/build/darwin/arm64_static/libllama.a -lstdc++
-// #cgo darwin,amd64 LDFLAGS: ${SRCDIR}/build/darwin/x86_64_static/libllama.a -lstdc++
-// #cgo windows,amd64 LDFLAGS: ${SRCDIR}/build/windows/amd64_static/libllama.a -static -lstdc++
-// #cgo windows,arm64 LDFLAGS: ${SRCDIR}/build/windows/arm64_static/libllama.a -static -lstdc++
-// #cgo linux,amd64 LDFLAGS: ${SRCDIR}/build/linux/x86_64_static/libllama.a -lstdc++
-// #cgo linux,arm64 LDFLAGS: ${SRCDIR}/build/linux/arm64_static/libllama.a -lstdc++
+// #cgo CFLAGS: -Illama.cpp -Illama.cpp/include -Illama.cpp/ggml/include
+// #cgo windows LDFLAGS: -static-libstdc++
+// #cgo LDFLAGS: -lllama -lggml -lstdc++ -lpthread
+// #cgo darwin,arm64 LDFLAGS: -L${SRCDIR}/build/darwin/arm64_static -L${SRCDIR}/build/darwin/arm64_static/src -L${SRCDIR}/build/darwin/arm64_static/ggml/src -framework Accelerate -framework Metal
+// #cgo darwin,amd64 LDFLAGS: -L${SRCDIR}/build/darwin/x86_64_static -L${SRCDIR}/build/darwin/x86_64_static/src -L${SRCDIR}/build/darwin/x86_64_static/ggml/src
+// #cgo windows,amd64 LDFLAGS: -L${SRCDIR}/build/windows/amd64_static -L${SRCDIR}/build/windows/amd64_static/src -L${SRCDIR}/build/windows/amd64_static/ggml/src
+// #cgo windows,arm64 LDFLAGS: -L${SRCDIR}/build/windows/arm64_static -L${SRCDIR}/build/windows/arm64_static/src -L${SRCDIR}/build/windows/arm64_static/ggml/src
+// #cgo linux,amd64 LDFLAGS: -L${SRCDIR}/build/linux/x86_64_static -L${SRCDIR}/build/linux/x86_64_static/src -L${SRCDIR}/build/linux/x86_64_static/ggml/src
+// #cgo linux,arm64 LDFLAGS: -L${SRCDIR}/build/linux/arm64_static -L${SRCDIR}/build/linux/arm64_static/src -L${SRCDIR}/build/linux/arm64_static/ggml/src
 // #include <stdlib.h>
 // #include "llama.h"
 import "C"
--- a/llm/patches/01-load-progress.diff
+++ b/llm/patches/01-load-progress.diff
@ -1,8 +1,8 @@
 diff --git a/common/common.cpp b/common/common.cpp
-index 73ff0e85..6adb1a92 100644
+index 2c05a4d4..927f0e3d 100644
 --- a/common/common.cpp
 +++ b/common/common.cpp
-@@ -2447,6 +2447,8 @@ struct llama_model_params llama_model_params_from_gpt_params(const gpt_params &
+@@ -2093,6 +2093,8 @@ struct llama_model_params llama_model_params_from_gpt_params(const gpt_params &
     mparams.use_mmap        = params.use_mmap;
     mparams.use_mlock       = params.use_mlock;
     mparams.check_tensors   = params.check_tensors;
@ -12,10 +12,10 @@ index 73ff0e85..6adb1a92 100644
         mparams.kv_overrides = NULL;
     } else {
 diff --git a/common/common.h b/common/common.h
-index 58ed72f4..0bb2605e 100644
+index 65c0ef81..ebca2c77 100644
 --- a/common/common.h
 +++ b/common/common.h
-@@ -180,6 +180,13 @@ struct gpt_params {
+@@ -184,6 +184,13 @@ struct gpt_params {
     std::string mmproj = "";        // path to multimodal projector
     std::vector<std::string> image; // path to image file(s)
 
@ -26,6 +26,6 @@ index 58ed72f4..0bb2605e 100644
 +    // context pointer passed to the progress callback
 +    void * progress_callback_user_data;
 +
-     // server params
-     int32_t port           = 8080;         // server listens on this network port
-     int32_t timeout_read   = 600;          // http read timeout in seconds
+     // embedding
+     bool embedding         = false; // get only sentence embedding
+     int32_t embd_normalize = 2;     // normalisation for embendings (-1=none, 0=max absolute int16, 1=taxicab, 2=euclidean, >2=p-norm)
--- a/llm/patches/03-load_exception.diff
+++ b/llm/patches/03-load_exception.diff
@ -1,17 +1,8 @@
-From 544a2d2e646d39e878d87dfbb3398a356bc560ab Mon Sep 17 00:00:00 2001
-From: Michael Yang <mxyng@pm.me>
-Date: Thu, 23 May 2024 11:18:45 -0700
-Subject: [PATCH] throw exception on load errors
-
---
- llama.cpp | 25 ++++++++++++++++---------
- 1 file changed, 16 insertions(+), 9 deletions(-)
-
-diff --git a/llama.cpp b/llama.cpp
-index 15c66077..8ba90b6a 100644
--- a/llama.cpp
-+++ b/llama.cpp
-@@ -6346,7 +6346,7 @@ static int llama_model_load(const std::string & fname, llama_model & model, llam
+diff --git a/src/llama.cpp b/src/llama.cpp
+index 73f52435..58a00fb1 100644
+--- a/src/llama.cpp
+++ b/src/llama.cpp
+@@ -7241,7 +7241,7 @@ static int llama_model_load(const std::string & fname, llama_model & model, llam
         }
     } catch (const std::exception & err) {
         LLAMA_LOG_ERROR("%s: error loading model: %s\n", __func__, err.what());
@ -20,7 +11,7 @@ index 15c66077..8ba90b6a 100644
     }
 
     return 0;
-@@ -15600,16 +15600,23 @@ struct llama_model * llama_load_model_from_file(
+@@ -17564,16 +17564,23 @@ struct llama_model * llama_load_model_from_file(
         }
         model->rpc_servers.push_back(servers);
     }
@ -52,6 +43,3 @@ index 15c66077..8ba90b6a 100644
     }
 
     return model;
-- 
-2.45.1
-
--- a/llm/patches/04-metal.diff
+++ b/llm/patches/04-metal.diff
@ -1,7 +1,7 @@
-diff --git a/ggml-metal.m b/ggml-metal.m
+diff --git a/ggml/src/ggml-metal.m b/ggml/src/ggml-metal.m
 index 0207b787..b5e9884b 100644
--- a/ggml-metal.m
-+++ b/ggml-metal.m
+--- a/ggml/src/ggml-metal.m
+++ b/ggml/src/ggml-metal.m
@@ -1396,27 +1396,23 @@ static enum ggml_status ggml_metal_graph_compute(
                         // to the matrix-vector kernel
                         int ne11_mm_min = 1;
--- a/llm/patches/05-default-pretokenizer.diff
+++ b/llm/patches/05-default-pretokenizer.diff
@ -1,8 +1,8 @@
-diff --git a/llama.cpp b/llama.cpp
-index 61948751..4b72a293 100644
--- a/llama.cpp
-+++ b/llama.cpp
-@@ -4824,16 +4824,7 @@ static void llm_load_vocab(
+diff --git a/src/llama.cpp b/src/llama.cpp
+index 73f52435..2b81b4bd 100644
+--- a/src/llama.cpp
+++ b/src/llama.cpp
+@@ -5092,16 +5092,7 @@ static void llm_load_vocab(
 
         // for now, only BPE models have pre-tokenizers
         if (vocab.type == LLAMA_VOCAB_TYPE_BPE) {
@ -20,13 +20,13 @@ index 61948751..4b72a293 100644
                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
             } else if (
                     tokenizer_pre == "llama3"   ||
-@@ -4888,7 +4879,8 @@ static void llm_load_vocab(
-                 tokenizer_pre == "poro-chat") {
-                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_PORO;
+@@ -5164,7 +5155,8 @@ static void llm_load_vocab(
+                 tokenizer_pre == "jais") {
+                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_JAIS;
             } else {
 -                throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str()));
 +                LLAMA_LOG_WARN("%s: missing or unrecognized pre-tokenizer type, using: 'default'\n", __func__);
 +                vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
             }
-         } else {
+         } else if (vocab.type == LLAMA_VOCAB_TYPE_SPM) {
             vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
--- a/llm/patches/06-qwen2.diff
+++ b/llm/patches/06-qwen2.diff
@ -1,7 +1,7 @@
-diff --git a/llama.cpp b/llama.cpp
+diff --git a/src/llama.cpp b/src/llama.cpp
 index 40d2ec2c..f34eb79a 100644
--- a/llama.cpp
-+++ b/llama.cpp
+--- a/src/llama.cpp
+++ b/src/llama.cpp
@@ -6943,7 +6943,7 @@ static struct ggml_tensor * llm_build_kqv(
         struct ggml_tensor * kq = ggml_mul_mat(ctx, k, q);
         cb(kq, "kq", il);
--- a/llm/patches/07-embeddings.diff
+++ b/llm/patches/07-embeddings.diff
@ -0,0 +1,45 @@
+diff --git a/src/llama.cpp b/src/llama.cpp
+index 1fe2b9f7..a43312a7 100644
+--- a/src/llama.cpp
+++ b/src/llama.cpp
+@@ -13689,7 +13689,7 @@ static size_t llama_output_reserve(llama_context & lctx, size_t n_outputs) {
+     const auto n_embd  = hparams.n_embd;
+ 
+     // TODO: use a per-batch flag for logits presence instead
+-    const bool has_logits = !cparams.embeddings;
+    const bool has_logits =  cparams.causal_attn;
+     const bool has_embd   =  lctx.is_encoding || (cparams.embeddings && (cparams.pooling_type == LLAMA_POOLING_TYPE_NONE));
+ 
+     const size_t logits_size = has_logits ? n_vocab*n_outputs_max : 0;
+@@ -13959,17 +13959,25 @@ static int llama_decode_internal(
+             // no output
+             res  = nullptr;
+             embd = nullptr;
+-        } else if (cparams.embeddings) {
+-            res = nullptr; // do not extract logits for embedding case
+-            embd = gf->nodes[gf->n_nodes - 1];
+-            if (strcmp(embd->name, "result_embd_pooled") != 0) {
+-                embd = gf->nodes[gf->n_nodes - 2];
+        }
+
+        if (cparams.embeddings) {
+            for (int i = gf->n_nodes - 1; i >= 0; --i) {
+                embd = gf->nodes[i];
+                if (strcmp(embd->name, "result_embd_pooled") == 0) {
+                    break;
+                }
+             }
+             GGML_ASSERT(strcmp(embd->name, "result_embd_pooled") == 0 && "missing embeddings tensor");
+-        } else {
+         } else {
+             embd = nullptr; // do not extract embeddings when not needed
+             GGML_ASSERT(strcmp(res->name, "result_output") == 0 && "missing result_output tensor");
+         }
+
+        if (!cparams.causal_attn) {
+            res = nullptr; // do not extract logits when not needed
+        }
+
+         // LLAMA_LOG_INFO("graph build time: %.3f ms (%d nodes, %d leafs)\n", (ggml_time_us() - t_start_us)/1000.0, gf->n_nodes, gf->n_leafs);
+ 
+         ggml_backend_sched_alloc_graph(lctx.sched, gf);
--- a/llm/patches/07-gemma.diff
+++ b/llm/patches/07-gemma.diff
@ -1,305 +0,0 @@
-From 5cadb45f39d001ffbad95b690d6cf0abcb4a6d96 Mon Sep 17 00:00:00 2001
-From: Ollama maintainers <hello@ollama.com>
-Date: Wed, 26 Jun 2024 16:18:09 -0700
-Subject: [PATCH] Architecture support
-
---
- llama.cpp | 194 +++++++++++++++++++++++++++++++++++++++++++++++++++++-
- 1 file changed, 193 insertions(+), 1 deletion(-)
-
-diff --git a/llama.cpp b/llama.cpp
-index 61948751..3b4196f5 100644
--- a/llama.cpp
-+++ b/llama.cpp
-@@ -217,6 +217,7 @@ enum llm_arch {
-     LLM_ARCH_INTERNLM2,
-     LLM_ARCH_MINICPM,
-     LLM_ARCH_GEMMA,
-+    LLM_ARCH_GEMMA2,
-     LLM_ARCH_STARCODER2,
-     LLM_ARCH_MAMBA,
-     LLM_ARCH_XVERSE,
-@@ -255,6 +256,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
-     { LLM_ARCH_INTERNLM2,       "internlm2"    },
-     { LLM_ARCH_MINICPM,         "minicpm"      },
-     { LLM_ARCH_GEMMA,           "gemma"        },
-+    { LLM_ARCH_GEMMA2,          "gemma2"       },
-     { LLM_ARCH_STARCODER2,      "starcoder2"   },
-     { LLM_ARCH_MAMBA,           "mamba"        },
-     { LLM_ARCH_XVERSE,          "xverse"       },
-@@ -464,10 +466,12 @@ enum llm_tensor {
-     LLM_TENSOR_ATTN_NORM,
-     LLM_TENSOR_ATTN_NORM_2,
-     LLM_TENSOR_ATTN_OUT_NORM,
-+    LLM_TENSOR_ATTN_POST_NORM,
-     LLM_TENSOR_ATTN_ROT_EMBD,
-     LLM_TENSOR_FFN_GATE_INP,
-     LLM_TENSOR_FFN_GATE_INP_SHEXP,
-     LLM_TENSOR_FFN_NORM,
-+    LLM_TENSOR_FFN_POST_NORM,
-     LLM_TENSOR_FFN_GATE,
-     LLM_TENSOR_FFN_DOWN,
-     LLM_TENSOR_FFN_UP,
-@@ -960,6 +964,24 @@ static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NA
-             { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
-         },
-     },
-+    {
-+        LLM_ARCH_GEMMA2,
-+        {
-+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
-+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
-+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
-+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
-+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
-+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
-+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
-+            { LLM_TENSOR_ATTN_POST_NORM,  "blk.%d.post_attention_norm" },
-+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
-+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
-+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
-+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
-+            { LLM_TENSOR_FFN_POST_NORM,   "blk.%d.post_ffw_norm" },
-+        },
-+    },
-     {
-         LLM_ARCH_STARCODER2,
-         {
-@@ -1941,6 +1963,8 @@ enum e_model {
-     MODEL_8x22B,
-     MODEL_16x12B,
-     MODEL_10B_128x3_66B,
-+    MODEL_9B,
-+    MODEL_27B,
- };
- 
- static const size_t kiB = 1024;
-@@ -2114,6 +2138,7 @@ struct llama_layer {
-     struct ggml_tensor * attn_out_norm_b;
-     struct ggml_tensor * attn_q_a_norm;
-     struct ggml_tensor * attn_kv_a_norm;
-+    struct ggml_tensor * attn_post_norm;
- 
-     // attention
-     struct ggml_tensor * wq;
-@@ -2136,6 +2161,7 @@ struct llama_layer {
-     // normalization
-     struct ggml_tensor * ffn_norm;
-     struct ggml_tensor * ffn_norm_b;
-+    struct ggml_tensor * ffn_post_norm;
-     struct ggml_tensor * layer_out_norm;
-     struct ggml_tensor * layer_out_norm_b;
-     struct ggml_tensor * ffn_norm_exps;
-@@ -4529,6 +4555,16 @@ static void llm_load_hparams(
-                 }
-             } break;
-         case LLM_ARCH_GEMMA:
-+            {
-+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
-+
-+                switch (hparams.n_layer) {
-+                    case 18: model.type = e_model::MODEL_9B; break;
-+                    case 28: model.type = e_model::MODEL_27B; break;
-+                    default: model.type = e_model::MODEL_UNKNOWN;
-+               }
-+            } break;
-+        case LLM_ARCH_GEMMA2:
-             {
-                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
- 
-@@ -6305,6 +6341,40 @@ static bool llm_load_tensors(
-                         layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd});
-                     }
-                 } break;
-+            case LLM_ARCH_GEMMA2:
-+                {
-+                    model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
-+
-+                    // output
-+                    model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
-+                    model.output      = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD,  "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED); // same as tok_embd, duplicated to allow offloading
-+
-+                    const int64_t n_ff          = hparams.n_ff;
-+                    const int64_t n_embd_head_k = hparams.n_embd_head_k;
-+                    const int64_t n_embd_k_gqa  = hparams.n_embd_k_gqa();
-+                    const int64_t n_embd_v_gqa  = hparams.n_embd_v_gqa();
-+
-+                    for (uint32_t i = 0; i < n_layer; ++i) {
-+                        ggml_context * ctx_layer = ctx_for_layer(i);
-+                        ggml_context * ctx_split = ctx_for_layer_split(i);
-+
-+                        auto & layer = model.layers[i];
-+
-+                        layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
-+
-+                        layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd_head_k * hparams.n_head});
-+                        layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_k_gqa});
-+                        layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_v_gqa});
-+                        layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * hparams.n_head, n_embd});
-+                        layer.attn_post_norm = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd});
-+
-+                        layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
-+                        layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff});
-+                        layer.ffn_up   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff});
-+                        layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd});
-+                        layer.ffn_post_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd});
-+                    }
-+                } break;
-             case LLM_ARCH_STARCODER2:
-                 {
-                     model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
-@@ -10614,6 +10684,123 @@ struct llm_build_context {
-         return gf;
-     }
- 
-+    struct ggml_cgraph * build_gemma2() {
-+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
-+
-+        const int64_t n_embd_head_k = hparams.n_embd_head_k;
-+
-+        struct ggml_tensor * cur;
-+        struct ggml_tensor * inpL;
-+
-+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
-+
-+        inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
-+        cb(inpL, "inp_scaled", -1);
-+
-+        // inp_pos - contains the positions
-+        struct ggml_tensor * inp_pos = build_inp_pos();
-+
-+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
-+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
-+
-+        for (int il = 0; il < n_layer; ++il) {
-+            // norm
-+            cur = llm_build_norm(ctx0, inpL, hparams,
-+                    model.layers[il].attn_norm, NULL,
-+                    LLM_NORM_RMS, cb, il);
-+            cb(cur, "attn_norm", il);
-+
-+            // self-attention
-+            {
-+                // compute Q and K and RoPE them
-+                struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
-+                cb(Qcur, "Qcur", il);
-+
-+                struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur);
-+                cb(Kcur, "Kcur", il);
-+
-+                struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur);
-+                cb(Vcur, "Vcur", il);
-+
-+                Qcur = ggml_rope_ext(
-+                        ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head_k, n_head,    n_tokens), inp_pos, nullptr,
-+                        n_embd_head_k, rope_type, n_ctx_orig, freq_base, freq_scale,
-+                        ext_factor, attn_factor, beta_fast, beta_slow);
-+                cb(Qcur, "Qcur", il);
-+
-+                Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head_k)));
-+                cb(Qcur, "Qcur_scaled", il);
-+
-+                Kcur = ggml_rope_ext(
-+                        ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head_k, n_head_kv, n_tokens), inp_pos, nullptr,
-+                        n_embd_head_k, rope_type, n_ctx_orig, freq_base, freq_scale,
-+                        ext_factor, attn_factor, beta_fast, beta_slow);
-+                cb(Kcur, "Kcur", il);
-+
-+                cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
-+                        model.layers[il].wo, NULL,
-+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f, cb, il);
-+            }
-+
-+            if (il == n_layer - 1) {
-+                // skip computing output for unused tokens
-+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-+                cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
-+                inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
-+            }
-+
-+            cur = llm_build_norm(ctx0, cur, hparams,
-+                    model.layers[il].attn_post_norm, NULL,
-+                    LLM_NORM_RMS, cb, il);
-+            cb(cur, "attn_post_norm", il);
-+
-+            struct ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL);
-+            cb(sa_out, "sa_out", il);
-+
-+            cur = llm_build_norm(ctx0, sa_out, hparams,
-+                    model.layers[il].ffn_norm, NULL,
-+                    LLM_NORM_RMS, cb, il);
-+            cb(cur, "ffn_norm", il);
-+
-+            // feed-forward network
-+            {
-+                cur = llm_build_ffn(ctx0, cur,
-+                        model.layers[il].ffn_up, NULL,
-+                        model.layers[il].ffn_gate, NULL,
-+                        model.layers[il].ffn_down, NULL,
-+                        NULL,
-+                        LLM_FFN_GELU, LLM_FFN_PAR, cb, il);
-+                cb(cur, "ffn_out", il);
-+            }
-+
-+            cur = llm_build_norm(ctx0, cur, hparams,
-+                model.layers[il].ffn_post_norm, NULL,
-+                LLM_NORM_RMS, cb, -1);
-+            cb(cur, "ffn_post_norm", -1);
-+
-+            cur = ggml_add(ctx0, cur, sa_out);
-+            cb(cur, "l_out", il);
-+
-+            // input for next layer
-+            inpL = cur;
-+        }
-+
-+        cur = inpL;
-+
-+        cur = llm_build_norm(ctx0, cur, hparams,
-+                model.output_norm, NULL,
-+                LLM_NORM_RMS, cb, -1);
-+        cb(cur, "result_norm", -1);
-+
-+        // lm_head
-+        cur = ggml_mul_mat(ctx0, model.output, cur);
-+        cb(cur, "result_output", -1);
-+
-+        ggml_build_forward_expand(gf, cur);
-+
-+        return gf;
-+    }
-+
-     struct ggml_cgraph * build_starcoder2() {
-         struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
- 
-@@ -11847,6 +12034,10 @@ static struct ggml_cgraph * llama_build_graph(
-             {
-                 result = llm.build_gemma();
-             } break;
-+        case LLM_ARCH_GEMMA2:
-+            {
-+                result = llm.build_gemma2();
-+            } break;
-         case LLM_ARCH_STARCODER2:
-             {
-                 result = llm.build_starcoder2();
-@@ -16671,6 +16862,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
-         case LLM_ARCH_PHI2:
-         case LLM_ARCH_PHI3:
-         case LLM_ARCH_GEMMA:
-+        case LLM_ARCH_GEMMA2:
-         case LLM_ARCH_STARCODER2:
-         case LLM_ARCH_GPTNEOX:
-             return LLAMA_ROPE_TYPE_NEOX;
-@@ -18551,7 +18743,7 @@ static int32_t llama_chat_apply_template_internal(
-         if (add_ass) {
-             ss << "<s>assistant\n";
-         }
-    } else if (tmpl == "gemma" || tmpl.find("<start_of_turn>") != std::string::npos) {
-+    } else if (tmpl == "gemma" || tmpl == "gemma2" || tmpl.find("<start_of_turn>") != std::string::npos) {
-         // google/gemma-7b-it
-         std::string system_prompt = "";
-         for (auto message : chat) {
-- 
-2.45.2
-
--- a/llm/patches/08-clip-unicode.diff
+++ b/llm/patches/08-clip-unicode.diff
@ -0,0 +1,42 @@
+diff --git a/examples/llava/clip.cpp b/examples/llava/clip.cpp
+index 95fbe3d0..5a02a6ec 100644
+--- a/examples/llava/clip.cpp
+++ b/examples/llava/clip.cpp
+@@ -32,6 +33,14 @@
+ #include <cinttypes>
+ #include <limits>
+ 
+#if defined(_WIN32)
+#define WIN32_LEAN_AND_MEAN
+#ifndef NOMINMAX
+    #define NOMINMAX
+#endif
+#include <windows.h>
+#endif
+
+ //#define CLIP_DEBUG_FUNCTIONS
+ 
+ // RGB uint8 image
+@@ -1055,7 +1064,22 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
+             return nullptr;
+         }
+ 
+#ifdef _WIN32
+        int wlen = MultiByteToWideChar(CP_UTF8, 0, fname, -1, NULL, 0);
+        if (!wlen) {
+            return NULL;
+        }
+        wchar_t * wbuf = (wchar_t *) malloc(wlen * sizeof(wchar_t));
+        wlen = MultiByteToWideChar(CP_UTF8, 0, fname, -1, wbuf, wlen);
+        if (!wlen) {
+            free(wbuf);
+            return NULL;
+        }
+        auto fin = std::ifstream(wbuf, std::ios::binary);
+        free(wbuf);
+#else
+         auto fin = std::ifstream(fname, std::ios::binary);
+#endif
+         if (!fin) {
+             LOG_TEE("cannot open model file for loading tensors\n");
+             clip_free(new_clip);
--- a/llm/patches/09-pooling.diff
+++ b/llm/patches/09-pooling.diff
@ -0,0 +1,60 @@
+diff --git a/src/llama.cpp b/src/llama.cpp
+index 721b8f4e..cfe7ac40 100644
+--- a/src/llama.cpp
+++ b/src/llama.cpp
+@@ -8420,14 +8420,14 @@ struct llm_build_context {
+     }
+ 
+     struct ggml_tensor * build_inp_mean() {
+-        lctx.inp_mean = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_tokens, n_tokens);
+        lctx.inp_mean = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_tokens, cparams.n_seq_max);
+         cb(lctx.inp_mean, "inp_mean", -1);
+         ggml_set_input(lctx.inp_mean);
+         return lctx.inp_mean;
+     }
+ 
+     struct ggml_tensor * build_inp_cls() {
+-        lctx.inp_cls = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);
+        lctx.inp_cls = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, cparams.n_seq_max);
+         cb(lctx.inp_cls, "inp_cls", -1);
+         ggml_set_input(lctx.inp_cls);
+         return lctx.inp_cls;
+@@ -13847,19 +13847,16 @@ static void llama_set_inputs(llama_context & lctx, const llama_batch & batch) {
+         GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_mean->buffer));
+ 
+         float * data = (float *) lctx.inp_mean->data;
+-        memset(lctx.inp_mean->data, 0, n_tokens * n_tokens * ggml_element_size(lctx.inp_mean));
+        memset(lctx.inp_mean->data, 0, n_tokens * cparams.n_seq_max * ggml_element_size(lctx.inp_mean));
+ 
+         std::vector<uint64_t> sum(n_tokens, 0);
+         for (int i = 0; i < n_tokens; ++i) {
+             const llama_seq_id seq_id = batch.seq_id[i][0];
+-
+-            GGML_ASSERT(seq_id < n_tokens && "seq_id cannot be larger than n_tokens with pooling_type == MEAN");
+-
+             sum[seq_id] += 1;
+         }
+ 
+-        std::vector<float> div(n_tokens, 0.0f);
+-        for (int i = 0; i < n_tokens; ++i) {
+        std::vector<float> div(cparams.n_seq_max, 0.0f);
+        for (uint32_t i = 0; i < cparams.n_seq_max; ++i) {
+             const uint64_t s = sum[i];
+             if (s > 0) {
+                 div[i] = 1.0f/float(s);
+@@ -13879,14 +13876,11 @@ static void llama_set_inputs(llama_context & lctx, const llama_batch & batch) {
+         GGML_ASSERT(ggml_backend_buffer_is_host(lctx.inp_cls->buffer));
+ 
+         uint32_t * data = (uint32_t *) lctx.inp_cls->data;
+-        memset(lctx.inp_cls->data, 0, n_tokens * ggml_element_size(lctx.inp_cls));
+        memset(lctx.inp_cls->data, 0, cparams.n_seq_max * ggml_element_size(lctx.inp_cls));
+ 
+         for (int i = 0; i < n_tokens; ++i) {
+             const llama_seq_id seq_id = batch.seq_id[i][0];
+             const llama_pos    pos    = batch.pos[i];
+-
+-            GGML_ASSERT(seq_id < n_tokens && "seq_id cannot be larger than n_tokens with pooling_type == CLS");
+-
+             if (pos == 0) {
+                 data[seq_id] = i;
+             }
--- a/llm/payload.go
+++ b/llm/payload.go
@ -38,7 +38,7 @@ func Init() error {
 	}

 	var variants []string
-	for v := range availableServers() {
+	for v := range getAvailableServers() {
 		variants = append(variants, v)
 	}
 	slog.Info(fmt.Sprintf("Dynamic LLM libraries %v", variants))
@ -50,7 +50,7 @@ func Init() error {
 // binary names may contain an optional variant separated by '_'
 // For example, "ollama_rocm_v6" and "ollama_rocm_v5" or "ollama_cpu" and "ollama_cpu_avx2"
 // Any library without a variant is the lowest common denominator
-func availableServers() map[string]string {
+func getAvailableServers() map[string]string {
 	payloadsDir, err := gpu.PayloadsDir()
 	if err != nil {
 		slog.Error("payload lookup error", "error", err)
@ -80,7 +80,7 @@ func availableServers() map[string]string {
 // TODO - switch to metadata based mapping
 func serversForGpu(info gpu.GpuInfo) []string {
 	// glob workDir for files that start with ollama_
-	availableServers := availableServers()
+	availableServers := getAvailableServers()
 	requested := info.Library
 	if info.Variant != gpu.CPUCapabilityNone {
 		requested += "_" + info.Variant.String()
@ -115,27 +115,29 @@ func serversForGpu(info gpu.GpuInfo) []string {
 		servers = append(servers, alt...)
 	}

-	// Load up the best CPU variant if not primary requested
-	if info.Library != "cpu" {
-		variant := gpu.GetCPUCapability()
-		// If no variant, then we fall back to default
-		// If we have a variant, try that if we find an exact match
-		// Attempting to run the wrong CPU instructions will panic the
-		// process
-		if variant != gpu.CPUCapabilityNone {
-			for cmp := range availableServers {
-				if cmp == "cpu_"+variant.String() {
-					servers = append(servers, cmp)
-					break
+	if !(runtime.GOOS == "darwin" && runtime.GOARCH == "arm64") {
+		// Load up the best CPU variant if not primary requested
+		if info.Library != "cpu" {
+			variant := gpu.GetCPUCapability()
+			// If no variant, then we fall back to default
+			// If we have a variant, try that if we find an exact match
+			// Attempting to run the wrong CPU instructions will panic the
+			// process
+			if variant != gpu.CPUCapabilityNone {
+				for cmp := range availableServers {
+					if cmp == "cpu_"+variant.String() {
+						servers = append(servers, cmp)
+						break
+					}
 				}
+			} else {
+				servers = append(servers, "cpu")
 			}
-		} else {
-			servers = append(servers, "cpu")
 		}
-	}

-	if len(servers) == 0 {
-		servers = []string{"cpu"}
+		if len(servers) == 0 {
+			servers = []string{"cpu"}
+		}
 	}

 	return servers
@ -147,7 +149,7 @@ func serverForCpu() string {
 		return "metal"
 	}
 	variant := gpu.GetCPUCapability()
-	availableServers := availableServers()
+	availableServers := getAvailableServers()
 	if variant != gpu.CPUCapabilityNone {
 		for cmp := range availableServers {
 			if cmp == "cpu_"+variant.String() {
--- a/llm/server.go
+++ b/llm/server.go
@ -82,7 +82,7 @@ func LoadModel(model string, maxArraySize int) (*GGML, error) {

 // NewLlamaServer will run a server for the given GPUs
 // The gpu list must be a single family.
-func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, projectors []string, opts api.Options) (LlamaServer, error) {
+func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, projectors []string, opts api.Options, numParallel int) (LlamaServer, error) {
 	var err error
 	var cpuRunner string
 	var estimate MemoryEstimate
@ -131,7 +131,20 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 		return nil, errors.New("ollama supports only one lora adapter, but multiple were provided")
 	}

-	availableServers := availableServers()
+	availableServers := getAvailableServers()
+	if len(availableServers) == 0 {
+		if runtime.GOOS != "windows" {
+			slog.Warn("llama server binary disappeared, reinitializing payloads")
+			err = Init()
+			if err != nil {
+				slog.Warn("failed to reinitialize payloads", "error", err)
+				return nil, err
+			}
+			availableServers = getAvailableServers()
+		} else {
+			return nil, finalErr
+		}
+	}
 	var servers []string
 	if cpuRunner != "" {
 		servers = []string{cpuRunner}
@ -208,7 +221,8 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 		if g.Library == "metal" &&
 			uint64(opts.NumGPU) > 0 &&
 			uint64(opts.NumGPU) < ggml.KV().BlockCount()+1 {
-			opts.UseMMap = api.TriStateFalse
+			opts.UseMMap = new(bool)
+			*opts.UseMMap = false
 		}
 	}

@ -218,9 +232,11 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr

 	// Windows CUDA should not use mmap for best performance
 	// Linux  with a model larger than free space, mmap leads to thrashing
-	if (runtime.GOOS == "windows" && gpus[0].Library == "cuda" && opts.UseMMap == api.TriStateUndefined) ||
-		(runtime.GOOS == "linux" && systemFreeMemory < estimate.TotalSize && opts.UseMMap == api.TriStateUndefined) ||
-		opts.UseMMap == api.TriStateFalse {
+	// For CPU loads we want the memory to be allocated, not FS cache
+	if (runtime.GOOS == "windows" && gpus[0].Library == "cuda" && opts.UseMMap == nil) ||
+		(runtime.GOOS == "linux" && systemFreeMemory < estimate.TotalSize && opts.UseMMap == nil) ||
+		(gpus[0].Library == "cpu" && opts.UseMMap == nil) ||
+		(opts.UseMMap != nil && !*opts.UseMMap) {
 		params = append(params, "--no-mmap")
 	}

@ -232,15 +248,6 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 		params = append(params, "--numa")
 	}

-	numParallel := envconfig.NumParallel
-
-	// TODO (jmorganca): multimodal models don't support parallel yet
-	// see https://github.com/ollama/ollama/issues/4165
-	if len(projectors) > 0 {
-		numParallel = 1
-		slog.Warn("multimodal models don't support parallel requests yet")
-	}
-
 	params = append(params, "--parallel", fmt.Sprintf("%d", numParallel))

 	if estimate.TensorSplit != "" {
@ -567,6 +574,9 @@ func (s *llmServer) WaitUntilRunning(ctx context.Context) error {
 			if s.status != nil && s.status.LastErrMsg != "" {
 				msg = s.status.LastErrMsg
 			}
+			if strings.Contains(msg, "unknown model") {
+				return fmt.Errorf("this model is not supported by your version of Ollama. You may need to upgrade")
+			}
 			return fmt.Errorf("llama runner process has terminated: %v %s", err, msg)
 		default:
 		}
--- a/llm/status.go
+++ b/llm/status.go
@ -25,6 +25,7 @@ var errorPrefixes = []string{
 	"CUDA error",
 	"cudaMalloc failed",
 	"\"ERR\"",
+	"error loading model",
 }

 func (w *StatusWriter) Write(b []byte) (int, error) {
--- a/openai/openai.go
+++ b/openai/openai.go
@ -12,6 +12,7 @@ import (

 	"github.com/gin-gonic/gin"
 	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/types/model"
 )

 type Error struct {
@ -42,6 +43,12 @@ type ChunkChoice struct {
 	FinishReason *string `json:"finish_reason"`
 }

+type CompleteChunkChoice struct {
+	Text         string  `json:"text"`
+	Index        int     `json:"index"`
+	FinishReason *string `json:"finish_reason"`
+}
+
 type Usage struct {
 	PromptTokens     int `json:"prompt_tokens"`
 	CompletionTokens int `json:"completion_tokens"`
@ -85,6 +92,51 @@ type ChatCompletionChunk struct {
 	Choices           []ChunkChoice `json:"choices"`
 }

+// TODO (https://github.com/ollama/ollama/issues/5259): support []string, []int and [][]int
+type CompletionRequest struct {
+	Model            string   `json:"model"`
+	Prompt           string   `json:"prompt"`
+	FrequencyPenalty float32  `json:"frequency_penalty"`
+	MaxTokens        *int     `json:"max_tokens"`
+	PresencePenalty  float32  `json:"presence_penalty"`
+	Seed             *int     `json:"seed"`
+	Stop             any      `json:"stop"`
+	Stream           bool     `json:"stream"`
+	Temperature      *float32 `json:"temperature"`
+	TopP             float32  `json:"top_p"`
+}
+
+type Completion struct {
+	Id                string                `json:"id"`
+	Object            string                `json:"object"`
+	Created           int64                 `json:"created"`
+	Model             string                `json:"model"`
+	SystemFingerprint string                `json:"system_fingerprint"`
+	Choices           []CompleteChunkChoice `json:"choices"`
+	Usage             Usage                 `json:"usage,omitempty"`
+}
+
+type CompletionChunk struct {
+	Id                string                `json:"id"`
+	Object            string                `json:"object"`
+	Created           int64                 `json:"created"`
+	Choices           []CompleteChunkChoice `json:"choices"`
+	Model             string                `json:"model"`
+	SystemFingerprint string                `json:"system_fingerprint"`
+}
+
+type Model struct {
+	Id      string `json:"id"`
+	Object  string `json:"object"`
+	Created int64  `json:"created"`
+	OwnedBy string `json:"owned_by"`
+}
+
+type ListCompletion struct {
+	Object string  `json:"object"`
+	Data   []Model `json:"data"`
+}
+
 func NewError(code int, message string) ErrorResponse {
 	var etype string
 	switch code {
@ -145,7 +197,79 @@ func toChunk(id string, r api.ChatResponse) ChatCompletionChunk {
 	}
 }

-func fromRequest(r ChatCompletionRequest) api.ChatRequest {
+func toCompletion(id string, r api.GenerateResponse) Completion {
+	return Completion{
+		Id:                id,
+		Object:            "text_completion",
+		Created:           r.CreatedAt.Unix(),
+		Model:             r.Model,
+		SystemFingerprint: "fp_ollama",
+		Choices: []CompleteChunkChoice{{
+			Text:  r.Response,
+			Index: 0,
+			FinishReason: func(reason string) *string {
+				if len(reason) > 0 {
+					return &reason
+				}
+				return nil
+			}(r.DoneReason),
+		}},
+		Usage: Usage{
+			// TODO: ollama returns 0 for prompt eval if the prompt was cached, but openai returns the actual count
+			PromptTokens:     r.PromptEvalCount,
+			CompletionTokens: r.EvalCount,
+			TotalTokens:      r.PromptEvalCount + r.EvalCount,
+		},
+	}
+}
+
+func toCompleteChunk(id string, r api.GenerateResponse) CompletionChunk {
+	return CompletionChunk{
+		Id:                id,
+		Object:            "text_completion",
+		Created:           time.Now().Unix(),
+		Model:             r.Model,
+		SystemFingerprint: "fp_ollama",
+		Choices: []CompleteChunkChoice{{
+			Text:  r.Response,
+			Index: 0,
+			FinishReason: func(reason string) *string {
+				if len(reason) > 0 {
+					return &reason
+				}
+				return nil
+			}(r.DoneReason),
+		}},
+	}
+}
+
+func toListCompletion(r api.ListResponse) ListCompletion {
+	var data []Model
+	for _, m := range r.Models {
+		data = append(data, Model{
+			Id:      m.Name,
+			Object:  "model",
+			Created: m.ModifiedAt.Unix(),
+			OwnedBy: model.ParseName(m.Name).Namespace,
+		})
+	}
+
+	return ListCompletion{
+		Object: "list",
+		Data:   data,
+	}
+}
+
+func toModel(r api.ShowResponse, m string) Model {
+	return Model{
+		Id:      m,
+		Object:  "model",
+		Created: r.ModifiedAt.Unix(),
+		OwnedBy: model.ParseName(m).Namespace,
+	}
+}
+
+func fromChatRequest(r ChatCompletionRequest) api.ChatRequest {
 	var messages []api.Message
 	for _, msg := range r.Messages {
 		messages = append(messages, api.Message{Role: msg.Role, Content: msg.Content})
@ -156,7 +280,7 @@ func fromRequest(r ChatCompletionRequest) api.ChatRequest {
 	switch stop := r.Stop.(type) {
 	case string:
 		options["stop"] = []string{stop}
-	case []interface{}:
+	case []any:
 		var stops []string
 		for _, s := range stop {
 			if str, ok := s.(string); ok {
@ -208,13 +332,78 @@ func fromRequest(r ChatCompletionRequest) api.ChatRequest {
 	}
 }

-type writer struct {
-	stream bool
-	id     string
+func fromCompleteRequest(r CompletionRequest) (api.GenerateRequest, error) {
+	options := make(map[string]any)
+
+	switch stop := r.Stop.(type) {
+	case string:
+		options["stop"] = []string{stop}
+	case []string:
+		options["stop"] = stop
+	default:
+		if r.Stop != nil {
+			return api.GenerateRequest{}, fmt.Errorf("invalid type for 'stop' field: %T", r.Stop)
+		}
+	}
+
+	if r.MaxTokens != nil {
+		options["num_predict"] = *r.MaxTokens
+	}
+
+	if r.Temperature != nil {
+		options["temperature"] = *r.Temperature * 2.0
+	} else {
+		options["temperature"] = 1.0
+	}
+
+	if r.Seed != nil {
+		options["seed"] = *r.Seed
+	}
+
+	options["frequency_penalty"] = r.FrequencyPenalty * 2.0
+
+	options["presence_penalty"] = r.PresencePenalty * 2.0
+
+	if r.TopP != 0.0 {
+		options["top_p"] = r.TopP
+	} else {
+		options["top_p"] = 1.0
+	}
+
+	return api.GenerateRequest{
+		Model:   r.Model,
+		Prompt:  r.Prompt,
+		Options: options,
+		Stream:  &r.Stream,
+	}, nil
+}
+
+type BaseWriter struct {
 	gin.ResponseWriter
 }

-func (w *writer) writeError(code int, data []byte) (int, error) {
+type ChatWriter struct {
+	stream bool
+	id     string
+	BaseWriter
+}
+
+type CompleteWriter struct {
+	stream bool
+	id     string
+	BaseWriter
+}
+
+type ListWriter struct {
+	BaseWriter
+}
+
+type RetrieveWriter struct {
+	BaseWriter
+	model string
+}
+
+func (w *BaseWriter) writeError(code int, data []byte) (int, error) {
 	var serr api.StatusError
 	err := json.Unmarshal(data, &serr)
 	if err != nil {
@ -230,7 +419,7 @@ func (w *writer) writeError(code int, data []byte) (int, error) {
 	return len(data), nil
 }

-func (w *writer) writeResponse(data []byte) (int, error) {
+func (w *ChatWriter) writeResponse(data []byte) (int, error) {
 	var chatResponse api.ChatResponse
 	err := json.Unmarshal(data, &chatResponse)
 	if err != nil {
@ -270,7 +459,7 @@ func (w *writer) writeResponse(data []byte) (int, error) {
 	return len(data), nil
 }

-func (w *writer) Write(data []byte) (int, error) {
+func (w *ChatWriter) Write(data []byte) (int, error) {
 	code := w.ResponseWriter.Status()
 	if code != http.StatusOK {
 		return w.writeError(code, data)
@ -279,7 +468,176 @@ func (w *writer) Write(data []byte) (int, error) {
 	return w.writeResponse(data)
 }

-func Middleware() gin.HandlerFunc {
+func (w *CompleteWriter) writeResponse(data []byte) (int, error) {
+	var generateResponse api.GenerateResponse
+	err := json.Unmarshal(data, &generateResponse)
+	if err != nil {
+		return 0, err
+	}
+
+	// completion chunk
+	if w.stream {
+		d, err := json.Marshal(toCompleteChunk(w.id, generateResponse))
+		if err != nil {
+			return 0, err
+		}
+
+		w.ResponseWriter.Header().Set("Content-Type", "text/event-stream")
+		_, err = w.ResponseWriter.Write([]byte(fmt.Sprintf("data: %s\n\n", d)))
+		if err != nil {
+			return 0, err
+		}
+
+		if generateResponse.Done {
+			_, err = w.ResponseWriter.Write([]byte("data: [DONE]\n\n"))
+			if err != nil {
+				return 0, err
+			}
+		}
+
+		return len(data), nil
+	}
+
+	// completion
+	w.ResponseWriter.Header().Set("Content-Type", "application/json")
+	err = json.NewEncoder(w.ResponseWriter).Encode(toCompletion(w.id, generateResponse))
+	if err != nil {
+		return 0, err
+	}
+
+	return len(data), nil
+}
+
+func (w *CompleteWriter) Write(data []byte) (int, error) {
+	code := w.ResponseWriter.Status()
+	if code != http.StatusOK {
+		return w.writeError(code, data)
+	}
+
+	return w.writeResponse(data)
+}
+
+func (w *ListWriter) writeResponse(data []byte) (int, error) {
+	var listResponse api.ListResponse
+	err := json.Unmarshal(data, &listResponse)
+	if err != nil {
+		return 0, err
+	}
+
+	w.ResponseWriter.Header().Set("Content-Type", "application/json")
+	err = json.NewEncoder(w.ResponseWriter).Encode(toListCompletion(listResponse))
+	if err != nil {
+		return 0, err
+	}
+
+	return len(data), nil
+}
+
+func (w *ListWriter) Write(data []byte) (int, error) {
+	code := w.ResponseWriter.Status()
+	if code != http.StatusOK {
+		return w.writeError(code, data)
+	}
+
+	return w.writeResponse(data)
+}
+
+func (w *RetrieveWriter) writeResponse(data []byte) (int, error) {
+	var showResponse api.ShowResponse
+	err := json.Unmarshal(data, &showResponse)
+	if err != nil {
+		return 0, err
+	}
+
+	// retrieve completion
+	w.ResponseWriter.Header().Set("Content-Type", "application/json")
+	err = json.NewEncoder(w.ResponseWriter).Encode(toModel(showResponse, w.model))
+	if err != nil {
+		return 0, err
+	}
+
+	return len(data), nil
+}
+
+func (w *RetrieveWriter) Write(data []byte) (int, error) {
+	code := w.ResponseWriter.Status()
+	if code != http.StatusOK {
+		return w.writeError(code, data)
+	}
+
+	return w.writeResponse(data)
+}
+
+func ListMiddleware() gin.HandlerFunc {
+	return func(c *gin.Context) {
+		w := &ListWriter{
+			BaseWriter: BaseWriter{ResponseWriter: c.Writer},
+		}
+
+		c.Writer = w
+
+		c.Next()
+	}
+}
+
+func RetrieveMiddleware() gin.HandlerFunc {
+	return func(c *gin.Context) {
+		var b bytes.Buffer
+		if err := json.NewEncoder(&b).Encode(api.ShowRequest{Name: c.Param("model")}); err != nil {
+			c.AbortWithStatusJSON(http.StatusInternalServerError, NewError(http.StatusInternalServerError, err.Error()))
+			return
+		}
+
+		c.Request.Body = io.NopCloser(&b)
+
+		// response writer
+		w := &RetrieveWriter{
+			BaseWriter: BaseWriter{ResponseWriter: c.Writer},
+			model:      c.Param("model"),
+		}
+
+		c.Writer = w
+
+		c.Next()
+	}
+}
+
+func CompletionsMiddleware() gin.HandlerFunc {
+	return func(c *gin.Context) {
+		var req CompletionRequest
+		err := c.ShouldBindJSON(&req)
+		if err != nil {
+			c.AbortWithStatusJSON(http.StatusBadRequest, NewError(http.StatusBadRequest, err.Error()))
+			return
+		}
+
+		var b bytes.Buffer
+		genReq, err := fromCompleteRequest(req)
+		if err != nil {
+			c.AbortWithStatusJSON(http.StatusBadRequest, NewError(http.StatusBadRequest, err.Error()))
+			return
+		}
+
+		if err := json.NewEncoder(&b).Encode(genReq); err != nil {
+			c.AbortWithStatusJSON(http.StatusInternalServerError, NewError(http.StatusInternalServerError, err.Error()))
+			return
+		}
+
+		c.Request.Body = io.NopCloser(&b)
+
+		w := &CompleteWriter{
+			BaseWriter: BaseWriter{ResponseWriter: c.Writer},
+			stream:     req.Stream,
+			id:         fmt.Sprintf("cmpl-%d", rand.Intn(999)),
+		}
+
+		c.Writer = w
+
+		c.Next()
+	}
+}
+
+func ChatMiddleware() gin.HandlerFunc {
 	return func(c *gin.Context) {
 		var req ChatCompletionRequest
 		err := c.ShouldBindJSON(&req)
@ -294,17 +652,17 @@ func Middleware() gin.HandlerFunc {
 		}

 		var b bytes.Buffer
-		if err := json.NewEncoder(&b).Encode(fromRequest(req)); err != nil {
+		if err := json.NewEncoder(&b).Encode(fromChatRequest(req)); err != nil {
 			c.AbortWithStatusJSON(http.StatusInternalServerError, NewError(http.StatusInternalServerError, err.Error()))
 			return
 		}

 		c.Request.Body = io.NopCloser(&b)

-		w := &writer{
-			ResponseWriter: c.Writer,
-			stream:         req.Stream,
-			id:             fmt.Sprintf("chatcmpl-%d", rand.Intn(999)),
+		w := &ChatWriter{
+			BaseWriter: BaseWriter{ResponseWriter: c.Writer},
+			stream:     req.Stream,
+			id:         fmt.Sprintf("chatcmpl-%d", rand.Intn(999)),
 		}

 		c.Writer = w
--- a/openai/openai_test.go
+++ b/openai/openai_test.go
@ -0,0 +1,298 @@
+package openai
+
+import (
+	"bytes"
+	"encoding/json"
+	"fmt"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/gin-gonic/gin"
+	"github.com/ollama/ollama/api"
+	"github.com/stretchr/testify/assert"
+)
+
+func TestMiddleware(t *testing.T) {
+	type testCase struct {
+		Name     string
+		Method   string
+		Path     string
+		TestPath string
+		Handler  func() gin.HandlerFunc
+		Endpoint func(c *gin.Context)
+		Setup    func(t *testing.T, req *http.Request)
+		Expected func(t *testing.T, resp *httptest.ResponseRecorder)
+	}
+
+	testCases := []testCase{
+		{
+			Name:     "chat handler",
+			Method:   http.MethodPost,
+			Path:     "/api/chat",
+			TestPath: "/api/chat",
+			Handler:  ChatMiddleware,
+			Endpoint: func(c *gin.Context) {
+				var chatReq api.ChatRequest
+				if err := c.ShouldBindJSON(&chatReq); err != nil {
+					c.JSON(http.StatusBadRequest, gin.H{"error": "invalid request"})
+					return
+				}
+
+				userMessage := chatReq.Messages[0].Content
+				var assistantMessage string
+
+				switch userMessage {
+				case "Hello":
+					assistantMessage = "Hello!"
+				default:
+					assistantMessage = "I'm not sure how to respond to that."
+				}
+
+				c.JSON(http.StatusOK, api.ChatResponse{
+					Message: api.Message{
+						Role:    "assistant",
+						Content: assistantMessage,
+					},
+				})
+			},
+			Setup: func(t *testing.T, req *http.Request) {
+				body := ChatCompletionRequest{
+					Model:    "test-model",
+					Messages: []Message{{Role: "user", Content: "Hello"}},
+				}
+
+				bodyBytes, _ := json.Marshal(body)
+
+				req.Body = io.NopCloser(bytes.NewReader(bodyBytes))
+				req.Header.Set("Content-Type", "application/json")
+			},
+			Expected: func(t *testing.T, resp *httptest.ResponseRecorder) {
+				assert.Equal(t, http.StatusOK, resp.Code)
+
+				var chatResp ChatCompletion
+				if err := json.NewDecoder(resp.Body).Decode(&chatResp); err != nil {
+					t.Fatal(err)
+				}
+
+				if chatResp.Object != "chat.completion" {
+					t.Fatalf("expected chat.completion, got %s", chatResp.Object)
+				}
+
+				if chatResp.Choices[0].Message.Content != "Hello!" {
+					t.Fatalf("expected Hello!, got %s", chatResp.Choices[0].Message.Content)
+				}
+			},
+		},
+		{
+			Name:     "completions handler",
+			Method:   http.MethodPost,
+			Path:     "/api/generate",
+			TestPath: "/api/generate",
+			Handler:  CompletionsMiddleware,
+			Endpoint: func(c *gin.Context) {
+				c.JSON(http.StatusOK, api.GenerateResponse{
+					Response: "Hello!",
+				})
+			},
+			Setup: func(t *testing.T, req *http.Request) {
+				body := CompletionRequest{
+					Model:  "test-model",
+					Prompt: "Hello",
+				}
+
+				bodyBytes, _ := json.Marshal(body)
+
+				req.Body = io.NopCloser(bytes.NewReader(bodyBytes))
+				req.Header.Set("Content-Type", "application/json")
+			},
+			Expected: func(t *testing.T, resp *httptest.ResponseRecorder) {
+				assert.Equal(t, http.StatusOK, resp.Code)
+				var completionResp Completion
+				if err := json.NewDecoder(resp.Body).Decode(&completionResp); err != nil {
+					t.Fatal(err)
+				}
+
+				if completionResp.Object != "text_completion" {
+					t.Fatalf("expected text_completion, got %s", completionResp.Object)
+				}
+
+				if completionResp.Choices[0].Text != "Hello!" {
+					t.Fatalf("expected Hello!, got %s", completionResp.Choices[0].Text)
+				}
+			},
+		},
+		{
+			Name:     "completions handler with params",
+			Method:   http.MethodPost,
+			Path:     "/api/generate",
+			TestPath: "/api/generate",
+			Handler:  CompletionsMiddleware,
+			Endpoint: func(c *gin.Context) {
+				var generateReq api.GenerateRequest
+				if err := c.ShouldBindJSON(&generateReq); err != nil {
+					c.JSON(http.StatusBadRequest, gin.H{"error": "invalid request"})
+					return
+				}
+
+				temperature := generateReq.Options["temperature"].(float64)
+				var assistantMessage string
+
+				switch temperature {
+				case 1.6:
+					assistantMessage = "Received temperature of 1.6"
+				default:
+					assistantMessage = fmt.Sprintf("Received temperature of %f", temperature)
+				}
+
+				c.JSON(http.StatusOK, api.GenerateResponse{
+					Response: assistantMessage,
+				})
+			},
+			Setup: func(t *testing.T, req *http.Request) {
+				temp := float32(0.8)
+				body := CompletionRequest{
+					Model:       "test-model",
+					Prompt:      "Hello",
+					Temperature: &temp,
+				}
+
+				bodyBytes, _ := json.Marshal(body)
+
+				req.Body = io.NopCloser(bytes.NewReader(bodyBytes))
+				req.Header.Set("Content-Type", "application/json")
+			},
+			Expected: func(t *testing.T, resp *httptest.ResponseRecorder) {
+				assert.Equal(t, http.StatusOK, resp.Code)
+				var completionResp Completion
+				if err := json.NewDecoder(resp.Body).Decode(&completionResp); err != nil {
+					t.Fatal(err)
+				}
+
+				if completionResp.Object != "text_completion" {
+					t.Fatalf("expected text_completion, got %s", completionResp.Object)
+				}
+
+				if completionResp.Choices[0].Text != "Received temperature of 1.6" {
+					t.Fatalf("expected Received temperature of 1.6, got %s", completionResp.Choices[0].Text)
+				}
+			},
+		},
+		{
+			Name:     "completions handler with error",
+			Method:   http.MethodPost,
+			Path:     "/api/generate",
+			TestPath: "/api/generate",
+			Handler:  CompletionsMiddleware,
+			Endpoint: func(c *gin.Context) {
+				c.JSON(http.StatusBadRequest, gin.H{"error": "invalid request"})
+			},
+			Setup: func(t *testing.T, req *http.Request) {
+				body := CompletionRequest{
+					Model:  "test-model",
+					Prompt: "Hello",
+				}
+
+				bodyBytes, _ := json.Marshal(body)
+
+				req.Body = io.NopCloser(bytes.NewReader(bodyBytes))
+				req.Header.Set("Content-Type", "application/json")
+			},
+			Expected: func(t *testing.T, resp *httptest.ResponseRecorder) {
+				if resp.Code != http.StatusBadRequest {
+					t.Fatalf("expected 400, got %d", resp.Code)
+				}
+
+				if !strings.Contains(resp.Body.String(), `"invalid request"`) {
+					t.Fatalf("error was not forwarded")
+				}
+			},
+		},
+		{
+			Name:     "list handler",
+			Method:   http.MethodGet,
+			Path:     "/api/tags",
+			TestPath: "/api/tags",
+			Handler:  ListMiddleware,
+			Endpoint: func(c *gin.Context) {
+				c.JSON(http.StatusOK, api.ListResponse{
+					Models: []api.ListModelResponse{
+						{
+							Name: "Test Model",
+						},
+					},
+				})
+			},
+			Expected: func(t *testing.T, resp *httptest.ResponseRecorder) {
+				assert.Equal(t, http.StatusOK, resp.Code)
+
+				var listResp ListCompletion
+				if err := json.NewDecoder(resp.Body).Decode(&listResp); err != nil {
+					t.Fatal(err)
+				}
+
+				if listResp.Object != "list" {
+					t.Fatalf("expected list, got %s", listResp.Object)
+				}
+
+				if len(listResp.Data) != 1 {
+					t.Fatalf("expected 1, got %d", len(listResp.Data))
+				}
+
+				if listResp.Data[0].Id != "Test Model" {
+					t.Fatalf("expected Test Model, got %s", listResp.Data[0].Id)
+				}
+			},
+		},
+		{
+			Name:     "retrieve model",
+			Method:   http.MethodGet,
+			Path:     "/api/show/:model",
+			TestPath: "/api/show/test-model",
+			Handler:  RetrieveMiddleware,
+			Endpoint: func(c *gin.Context) {
+				c.JSON(http.StatusOK, api.ShowResponse{
+					ModifiedAt: time.Date(2024, 6, 17, 13, 45, 0, 0, time.UTC),
+				})
+			},
+			Expected: func(t *testing.T, resp *httptest.ResponseRecorder) {
+				var retrieveResp Model
+				if err := json.NewDecoder(resp.Body).Decode(&retrieveResp); err != nil {
+					t.Fatal(err)
+				}
+
+				if retrieveResp.Object != "model" {
+					t.Fatalf("Expected object to be model, got %s", retrieveResp.Object)
+				}
+
+				if retrieveResp.Id != "test-model" {
+					t.Fatalf("Expected id to be test-model, got %s", retrieveResp.Id)
+				}
+			},
+		},
+	}
+
+	gin.SetMode(gin.TestMode)
+	router := gin.New()
+
+	for _, tc := range testCases {
+		t.Run(tc.Name, func(t *testing.T) {
+			router = gin.New()
+			router.Use(tc.Handler())
+			router.Handle(tc.Method, tc.Path, tc.Endpoint)
+			req, _ := http.NewRequest(tc.Method, tc.TestPath, nil)
+
+			if tc.Setup != nil {
+				tc.Setup(t, req)
+			}
+
+			resp := httptest.NewRecorder()
+			router.ServeHTTP(resp, req)
+
+			tc.Expected(t, resp)
+		})
+	}
+}
--- a/parser/parser.go
+++ b/parser/parser.go
@ -124,7 +124,7 @@ func ParseFile(r io.Reader) (*File, error) {
 			case stateComment, stateNil:
 				// pass
 			case stateValue:
-				s, ok := unquote(b.String())
+				s, ok := unquote(strings.TrimSpace(b.String()))
 				if !ok || isSpace(r) {
 					if _, err := b.WriteRune(r); err != nil {
 						return nil, err
@ -158,7 +158,7 @@ func ParseFile(r io.Reader) (*File, error) {
 	case stateComment, stateNil:
 		// pass; nothing to flush
 	case stateValue:
-		s, ok := unquote(b.String())
+		s, ok := unquote(strings.TrimSpace(b.String()))
 		if !ok {
 			return nil, io.ErrUnexpectedEOF
 		}
--- a/parser/parser_test.go
+++ b/parser/parser_test.go
@ -22,7 +22,13 @@ ADAPTER adapter1
 LICENSE MIT
 PARAMETER param1 value1
 PARAMETER param2 value2
-TEMPLATE template1
+TEMPLATE """{{ if .System }}<|start_header_id|>system<|end_header_id|>
+
+{{ .System }}<|eot_id|>{{ end }}{{ if .Prompt }}<|start_header_id|>user<|end_header_id|>
+
+{{ .Prompt }}<|eot_id|>{{ end }}<|start_header_id|>assistant<|end_header_id|>
+
+{{ .Response }}<|eot_id|>"""    
 `

 	reader := strings.NewReader(input)
@ -36,7 +42,40 @@ TEMPLATE template1
 		{Name: "license", Args: "MIT"},
 		{Name: "param1", Args: "value1"},
 		{Name: "param2", Args: "value2"},
-		{Name: "template", Args: "template1"},
+		{Name: "template", Args: "{{ if .System }}<|start_header_id|>system<|end_header_id|>\n\n{{ .System }}<|eot_id|>{{ end }}{{ if .Prompt }}<|start_header_id|>user<|end_header_id|>\n\n{{ .Prompt }}<|eot_id|>{{ end }}<|start_header_id|>assistant<|end_header_id|>\n\n{{ .Response }}<|eot_id|>"},
+	}
+
+	assert.Equal(t, expectedCommands, modelfile.Commands)
+}
+
+func TestParseFileTrimSpace(t *testing.T) {
+	input := `
+FROM "     model 1"
+ADAPTER      adapter3
+LICENSE "MIT       "
+PARAMETER param1        value1
+PARAMETER param2    value2
+TEMPLATE """   {{ if .System }}<|start_header_id|>system<|end_header_id|>
+
+{{ .System }}<|eot_id|>{{ end }}{{ if .Prompt }}<|start_header_id|>user<|end_header_id|>
+
+{{ .Prompt }}<|eot_id|>{{ end }}<|start_header_id|>assistant<|end_header_id|>
+
+{{ .Response }}<|eot_id|>   """    
+`
+
+	reader := strings.NewReader(input)
+
+	modelfile, err := ParseFile(reader)
+	require.NoError(t, err)
+
+	expectedCommands := []Command{
+		{Name: "model", Args: "     model 1"},
+		{Name: "adapter", Args: "adapter3"},
+		{Name: "license", Args: "MIT       "},
+		{Name: "param1", Args: "value1"},
+		{Name: "param2", Args: "value2"},
+		{Name: "template", Args: "   {{ if .System }}<|start_header_id|>system<|end_header_id|>\n\n{{ .System }}<|eot_id|>{{ end }}{{ if .Prompt }}<|start_header_id|>user<|end_header_id|>\n\n{{ .Prompt }}<|eot_id|>{{ end }}<|start_header_id|>assistant<|end_header_id|>\n\n{{ .Response }}<|eot_id|>   "},
 	}

 	assert.Equal(t, expectedCommands, modelfile.Commands)
@ -48,6 +87,26 @@ func TestParseFileFrom(t *testing.T) {
 		expected []Command
 		err      error
 	}{
+		{
+			"FROM \"FOO  BAR  \"",
+			[]Command{{Name: "model", Args: "FOO  BAR  "}},
+			nil,
+		},
+		{
+			"FROM \"FOO BAR\"\nPARAMETER param1 value1",
+			[]Command{{Name: "model", Args: "FOO BAR"}, {Name: "param1", Args: "value1"}},
+			nil,
+		},
+		{
+			"FROM     FOOO BAR    ",
+			[]Command{{Name: "model", Args: "FOOO BAR"}},
+			nil,
+		},
+		{
+			"FROM /what/is/the path ",
+			[]Command{{Name: "model", Args: "/what/is/the path"}},
+			nil,
+		},
 		{
 			"FROM foo",
 			[]Command{{Name: "model", Args: "foo"}},
@ -86,6 +145,11 @@ func TestParseFileFrom(t *testing.T) {
 			[]Command{{Name: "param1", Args: "value1"}, {Name: "model", Args: "foo"}},
 			nil,
 		},
+		{
+			"PARAMETER what the \nFROM lemons make lemonade ",
+			[]Command{{Name: "what", Args: "the"}, {Name: "model", Args: "lemons make lemonade"}},
+			nil,
+		},
 	}

 	for _, c := range cases {
@ -399,7 +463,7 @@ func TestParseFileParameters(t *testing.T) {
 		"mirostat_eta 1.0":             {"mirostat_eta", "1.0"},
 		"penalize_newline true":        {"penalize_newline", "true"},
 		"stop ### User:":               {"stop", "### User:"},
-		"stop ### User: ":              {"stop", "### User: "},
+		"stop ### User: ":              {"stop", "### User:"},
 		"stop \"### User:\"":           {"stop", "### User:"},
 		"stop \"### User: \"":          {"stop", "### User: "},
 		"stop \"\"\"### User:\"\"\"":   {"stop", "### User:"},
--- a/run_model.sh
+++ b/run_model.sh
@ -9,7 +9,8 @@ if [ -n "$INIT_MODELS" ]; then
        echo "FROM /models/$MODEL_NAME" > /tmp/Modelfile
        echo "PARAMETER temperature 1" >> /tmp/Modelfile
        echo "PARAMETER num_ctx 4096" >> /tmp/Modelfile
-        echo 'PARAMETER stop ["<|im_start|>","<|im_end|>"]' >> /tmp/Modelfile
+        echo 'PARAMETER stop "<|im_start|>"' >> /tmp/Modelfile
+        echo 'PARAMETER stop "<|im_end|>"' >> /tmp/Modelfile
 		echo 'TEMPLATE """{{ if .System }}<|im_start|>system' >> /tmp/Modelfile
 		echo "{{ .System }}<|im_end|>" >> /tmp/Modelfile
 		echo "{{ end }}{{ if .Prompt }}<|im_start|>user" >> /tmp/Modelfile
--- a/scripts/rh_linux_deps.sh
+++ b/scripts/rh_linux_deps.sh
@ -6,10 +6,21 @@ set -ex
 MACHINE=$(uname -m)

 if grep -i "centos" /etc/system-release >/dev/null; then
+    # As of 7/1/2024 mirrorlist.centos.org has been taken offline, so adjust accordingly
+    sed -i s/mirror.centos.org/vault.centos.org/g /etc/yum.repos.d/*.repo
+    sed -i s/^#.*baseurl=http/baseurl=http/g /etc/yum.repos.d/*.repo
+    sed -i s/^mirrorlist=http/#mirrorlist=http/g /etc/yum.repos.d/*.repo
+
    # Centos 7 derivatives have too old of a git version to run our generate script
    # uninstall and ignore failures
    yum remove -y git
    yum -y install epel-release centos-release-scl
+
+    # The release packages reinstate the mirrors, undo that again
+    sed -i s/mirror.centos.org/vault.centos.org/g /etc/yum.repos.d/*.repo
+    sed -i s/^#.*baseurl=http/baseurl=http/g /etc/yum.repos.d/*.repo
+    sed -i s/^mirrorlist=http/#mirrorlist=http/g /etc/yum.repos.d/*.repo
+
    yum -y install dnf
    if [ "${MACHINE}" = "x86_64" ]; then
        yum -y install https://repo.ius.io/ius-release-el7.rpm
--- a/server/images.go
+++ b/server/images.go
@ -28,11 +28,16 @@ import (
 	"github.com/ollama/ollama/format"
 	"github.com/ollama/ollama/llm"
 	"github.com/ollama/ollama/parser"
+	"github.com/ollama/ollama/template"
 	"github.com/ollama/ollama/types/errtypes"
 	"github.com/ollama/ollama/types/model"
 	"github.com/ollama/ollama/version"
 )

+type Capability string
+
+const CapabilityCompletion = Capability("completion")
+
 type registryOptions struct {
 	Insecure bool
 	Username string
@ -48,16 +53,43 @@ type Model struct {
 	ParentModel    string
 	AdapterPaths   []string
 	ProjectorPaths []string
-	Template       string
 	System         string
 	License        []string
 	Digest         string
 	Options        map[string]interface{}
 	Messages       []Message
+
+	Template *template.Template
 }

-func (m *Model) IsEmbedding() bool {
-	return slices.Contains(m.Config.ModelFamilies, "bert") || slices.Contains(m.Config.ModelFamilies, "nomic-bert")
+func (m *Model) Has(caps ...Capability) bool {
+	for _, cap := range caps {
+		switch cap {
+		case CapabilityCompletion:
+			f, err := os.Open(m.ModelPath)
+			if err != nil {
+				slog.Error("couldn't open model file", "error", err)
+				continue
+			}
+			defer f.Close()
+
+			// TODO(mxyng): decode the GGML into model to avoid doing this multiple times
+			ggml, _, err := llm.DecodeGGML(f, 0)
+			if err != nil {
+				slog.Error("couldn't decode ggml", "error", err)
+				continue
+			}
+
+			if _, ok := ggml.KV()[fmt.Sprintf("%s.pooling_type", ggml.KV().Architecture())]; ok {
+				return false
+			}
+		default:
+			slog.Error("unknown capability", "capability", cap)
+			return false
+		}
+	}
+
+	return true
 }

 func (m *Model) String() string {
@ -82,10 +114,10 @@ func (m *Model) String() string {
 		})
 	}

-	if m.Template != "" {
+	if m.Template != nil {
 		modelfile.Commands = append(modelfile.Commands, parser.Command{
 			Name: "template",
-			Args: m.Template,
+			Args: m.Template.String(),
 		})
 	}

@ -135,13 +167,6 @@ type Message struct {
 	Content string `json:"content"`
 }

-type ManifestV2 struct {
-	SchemaVersion int      `json:"schemaVersion"`
-	MediaType     string   `json:"mediaType"`
-	Config        *Layer   `json:"config"`
-	Layers        []*Layer `json:"layers"`
-}
-
 type ConfigV2 struct {
 	ModelFormat   string   `json:"model_format"`
 	ModelFamily   string   `json:"model_family"`
@ -160,7 +185,7 @@ type RootFS struct {
 	DiffIDs []string `json:"diff_ids"`
 }

-func GetManifest(mp ModelPath) (*ManifestV2, string, error) {
+func GetManifest(mp ModelPath) (*Manifest, string, error) {
 	fp, err := mp.GetManifestPath()
 	if err != nil {
 		return nil, "", err
@ -170,7 +195,7 @@ func GetManifest(mp ModelPath) (*ManifestV2, string, error) {
 		return nil, "", err
 	}

-	var manifest *ManifestV2
+	var manifest *Manifest

 	bts, err := os.ReadFile(fp)
 	if err != nil {
@ -198,8 +223,7 @@ func GetModel(name string) (*Model, error) {
 		Name:      mp.GetFullTagname(),
 		ShortName: mp.GetShortTagname(),
 		Digest:    digest,
-		Template:  "{{ .Prompt }}",
-		License:   []string{},
+		Template:  template.DefaultTemplate,
 	}

 	filename, err := GetBlobsPath(manifest.Config.Digest)
@ -235,13 +259,17 @@ func GetModel(name string) (*Model, error) {
 			model.AdapterPaths = append(model.AdapterPaths, filename)
 		case "application/vnd.ollama.image.projector":
 			model.ProjectorPaths = append(model.ProjectorPaths, filename)
-		case "application/vnd.ollama.image.template":
+		case "application/vnd.ollama.image.prompt",
+			"application/vnd.ollama.image.template":
 			bts, err := os.ReadFile(filename)
 			if err != nil {
 				return nil, err
 			}

-			model.Template = string(bts)
+			model.Template, err = template.Parse(string(bts))
+			if err != nil {
+				return nil, err
+			}
 		case "application/vnd.ollama.image.system":
 			bts, err := os.ReadFile(filename)
 			if err != nil {
@ -249,13 +277,6 @@ func GetModel(name string) (*Model, error) {
 			}

 			model.System = string(bts)
-		case "application/vnd.ollama.image.prompt":
-			bts, err := os.ReadFile(filename)
-			if err != nil {
-				return nil, err
-			}
-
-			model.Template = string(bts)
 		case "application/vnd.ollama.image.params":
 			params, err := os.Open(filename)
 			if err != nil {
@ -822,7 +843,7 @@ func PushModel(ctx context.Context, name string, regOpts *registryOptions, fn fu
 func PullModel(ctx context.Context, name string, regOpts *registryOptions, fn func(api.ProgressResponse)) error {
 	mp := ParseModelPath(name)

-	var manifest *ManifestV2
+	var manifest *Manifest
 	var err error
 	var noprune string

@ -929,7 +950,7 @@ func PullModel(ctx context.Context, name string, regOpts *registryOptions, fn fu
 	return nil
 }

-func pullModelManifest(ctx context.Context, mp ModelPath, regOpts *registryOptions) (*ManifestV2, error) {
+func pullModelManifest(ctx context.Context, mp ModelPath, regOpts *registryOptions) (*Manifest, error) {
 	requestURL := mp.BaseURL().JoinPath("v2", mp.GetNamespaceRepository(), "manifests", mp.Tag)

 	headers := make(http.Header)
@ -940,7 +961,7 @@ func pullModelManifest(ctx context.Context, mp ModelPath, regOpts *registryOptio
 	}
 	defer resp.Body.Close()

-	var m *ManifestV2
+	var m *Manifest
 	if err := json.NewDecoder(resp.Body).Decode(&m); err != nil {
 		return nil, err
 	}
--- a/server/manifest.go
+++ b/server/manifest.go
@ -14,7 +14,10 @@ import (
 )

 type Manifest struct {
-	ManifestV2
+	SchemaVersion int      `json:"schemaVersion"`
+	MediaType     string   `json:"mediaType"`
+	Config        *Layer   `json:"config"`
+	Layers        []*Layer `json:"layers"`

 	filepath string
 	fi       os.FileInfo
@ -66,7 +69,7 @@ func ParseNamedManifest(n model.Name) (*Manifest, error) {

 	p := filepath.Join(manifests, n.Filepath())

-	var m ManifestV2
+	var m Manifest
 	f, err := os.Open(p)
 	if err != nil {
 		return nil, err
@ -83,12 +86,11 @@ func ParseNamedManifest(n model.Name) (*Manifest, error) {
 		return nil, err
 	}

-	return &Manifest{
-		ManifestV2: m,
-		filepath:   p,
-		fi:         fi,
-		digest:     fmt.Sprintf("%x", sha256sum.Sum(nil)),
-	}, nil
+	m.filepath = p
+	m.fi = fi
+	m.digest = fmt.Sprintf("%x", sha256sum.Sum(nil))
+
+	return &m, nil
 }

 func WriteManifest(name model.Name, config *Layer, layers []*Layer) error {
@ -108,7 +110,7 @@ func WriteManifest(name model.Name, config *Layer, layers []*Layer) error {
 	}
 	defer f.Close()

-	m := ManifestV2{
+	m := Manifest{
 		SchemaVersion: 2,
 		MediaType:     "application/vnd.docker.distribution.manifest.v2+json",
 		Config:        config,
--- a/server/manifest_test.go
+++ b/server/manifest_test.go
@ -25,7 +25,7 @@ func createManifest(t *testing.T, path, name string) {
 	}
 	defer f.Close()

-	if err := json.NewEncoder(f).Encode(ManifestV2{}); err != nil {
+	if err := json.NewEncoder(f).Encode(Manifest{}); err != nil {
 		t.Fatal(err)
 	}
 }
--- a/server/model.go
+++ b/server/model.go
@ -11,12 +11,11 @@ import (
 	"net/http"
 	"os"
 	"path/filepath"
-	"strings"

 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/convert"
 	"github.com/ollama/ollama/llm"
-	"github.com/ollama/ollama/templates"
+	"github.com/ollama/ollama/template"
 	"github.com/ollama/ollama/types/model"
 )

@ -91,12 +90,11 @@ func extractFromZipFile(p string, file *os.File, fn func(api.ProgressResponse))

 	fn(api.ProgressResponse{Status: "unpacking model metadata"})
 	for _, f := range r.File {
-		n := filepath.Join(p, f.Name)
-		if !strings.HasPrefix(n, p) {
-			slog.Warn("skipped extracting file outside of context", "name", f.Name)
-			continue
+		if !filepath.IsLocal(f.Name) {
+			return fmt.Errorf("%w: %s", zip.ErrInsecurePath, f.Name)
 		}

+		n := filepath.Join(p, f.Name)
 		if err := os.MkdirAll(filepath.Dir(n), 0o750); err != nil {
 			return err
 		}
@ -258,7 +256,7 @@ func parseFromFile(ctx context.Context, file *os.File, digest string, fn func(ap
 func detectChatTemplate(layers []*layerGGML) ([]*layerGGML, error) {
 	for _, layer := range layers {
 		if s := layer.GGML.KV().ChatTemplate(); s != "" {
-			if t, err := templates.NamedTemplate(s); err != nil {
+			if t, err := template.Named(s); err != nil {
 				slog.Debug("template detection", "error", err)
 			} else {
 				tmpl, err := NewLayer(t.Reader(), "application/vnd.ollama.image.template")
--- a/server/model_test.go
+++ b/server/model_test.go
@ -3,10 +3,12 @@ package server
 import (
 	"archive/zip"
 	"bytes"
+	"errors"
 	"io"
 	"os"
 	"path/filepath"
 	"slices"
+	"strings"
 	"testing"

 	"github.com/ollama/ollama/api"
@ -39,13 +41,31 @@ func TestExtractFromZipFile(t *testing.T) {
 	cases := []struct {
 		name   string
 		expect []string
+		err    error
 	}{
 		{
 			name:   "good",
 			expect: []string{"good"},
 		},
 		{
-			name: filepath.Join("..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "bad"),
+			name:   strings.Join([]string{"path", "..", "to", "good"}, string(os.PathSeparator)),
+			expect: []string{filepath.Join("to", "good")},
+		},
+		{
+			name:   strings.Join([]string{"path", "..", "to", "..", "good"}, string(os.PathSeparator)),
+			expect: []string{"good"},
+		},
+		{
+			name:   strings.Join([]string{"path", "to", "..", "..", "good"}, string(os.PathSeparator)),
+			expect: []string{"good"},
+		},
+		{
+			name: strings.Join([]string{"..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "bad"}, string(os.PathSeparator)),
+			err:  zip.ErrInsecurePath,
+		},
+		{
+			name: strings.Join([]string{"path", "..", "..", "to", "bad"}, string(os.PathSeparator)),
+			err:  zip.ErrInsecurePath,
 		},
 	}

@ -55,7 +75,7 @@ func TestExtractFromZipFile(t *testing.T) {
 			defer f.Close()

 			tempDir := t.TempDir()
-			if err := extractFromZipFile(tempDir, f, func(api.ProgressResponse) {}); err != nil {
+			if err := extractFromZipFile(tempDir, f, func(api.ProgressResponse) {}); !errors.Is(err, tt.err) {
 				t.Fatal(err)
 			}

--- a/server/modelpath.go
+++ b/server/modelpath.go
@ -103,18 +103,9 @@ func (mp ModelPath) GetShortTagname() string {
 	return fmt.Sprintf("%s/%s/%s:%s", mp.Registry, mp.Namespace, mp.Repository, mp.Tag)
 }

-// modelsDir returns the value of the OLLAMA_MODELS environment variable or the user's home directory if OLLAMA_MODELS is not set.
-// The models directory is where Ollama stores its model files and manifests.
-func modelsDir() (string, error) {
-	return envconfig.ModelsDir, nil
-}
-
 // GetManifestPath returns the path to the manifest file for the given model path, it is up to the caller to create the directory if it does not exist.
 func (mp ModelPath) GetManifestPath() (string, error) {
-	dir, err := modelsDir()
-	if err != nil {
-		return "", err
-	}
+	dir := envconfig.ModelsDir

 	return filepath.Join(dir, "manifests", mp.Registry, mp.Namespace, mp.Repository, mp.Tag), nil
 }
@ -127,10 +118,7 @@ func (mp ModelPath) BaseURL() *url.URL {
 }

 func GetManifestPath() (string, error) {
-	dir, err := modelsDir()
-	if err != nil {
-		return "", err
-	}
+	dir := envconfig.ModelsDir

 	path := filepath.Join(dir, "manifests")
 	if err := os.MkdirAll(path, 0o755); err != nil {
@ -141,10 +129,7 @@ func GetManifestPath() (string, error) {
 }

 func GetBlobsPath(digest string) (string, error) {
-	dir, err := modelsDir()
-	if err != nil {
-		return "", err
-	}
+	dir := envconfig.ModelsDir

 	// only accept actual sha256 digests
 	pattern := "^sha256[:-][0-9a-fA-F]{64}$"
--- a/server/prompt.go
+++ b/server/prompt.go
@ -4,10 +4,11 @@ import (
 	"fmt"
 	"log/slog"
 	"strings"
-	"text/template"
+
 	"text/template/parse"

 	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/template"
 )

 // isResponseNode checks if the node contains .Response
@ -53,13 +54,8 @@ func formatTemplateForResponse(tmpl *template.Template, generate bool) {

 // Prompt renders a prompt from a template. If generate is set to true,
 // the response and parts of the template following it are not rendered
-func Prompt(tmpl, system, prompt, response string, generate bool) (string, error) {
-	parsed, err := template.New("").Option("missingkey=zero").Parse(tmpl)
-	if err != nil {
-		return "", err
-	}
-
-	formatTemplateForResponse(parsed, generate)
+func Prompt(tmpl *template.Template, system, prompt, response string, generate bool) (string, error) {
+	formatTemplateForResponse(tmpl, generate)

 	vars := map[string]any{
 		"System":   system,
@ -68,14 +64,14 @@ func Prompt(tmpl, system, prompt, response string, generate bool) (string, error
 	}

 	var sb strings.Builder
-	if err := parsed.Execute(&sb, vars); err != nil {
+	if err := tmpl.Execute(&sb, vars); err != nil {
 		return "", err
 	}

 	return sb.String(), nil
 }

-func countTokens(tmpl string, system string, prompt string, response string, encode func(string) ([]int, error)) (int, error) {
+func countTokens(tmpl *template.Template, system string, prompt string, response string, encode func(string) ([]int, error)) (int, error) {
 	rendered, err := Prompt(tmpl, system, prompt, response, false)
 	if err != nil {
 		return 0, err
@ -91,7 +87,7 @@ func countTokens(tmpl string, system string, prompt string, response string, enc
 }

 // ChatPrompt builds up a prompt from a series of messages, truncating based on context window size
-func ChatPrompt(tmpl string, messages []api.Message, window int, encode func(string) ([]int, error)) (string, error) {
+func ChatPrompt(tmpl *template.Template, messages []api.Message, window int, encode func(string) ([]int, error)) (string, error) {
 	type prompt struct {
 		System   string
 		Prompt   string
--- a/server/prompt_test.go
+++ b/server/prompt_test.go
@ -5,6 +5,7 @@ import (
 	"testing"

 	"github.com/ollama/ollama/api"
+	"github.com/ollama/ollama/template"
 )

 func TestPrompt(t *testing.T) {
@ -61,7 +62,12 @@ func TestPrompt(t *testing.T) {

 	for _, tc := range tests {
 		t.Run(tc.name, func(t *testing.T) {
-			got, err := Prompt(tc.template, tc.system, tc.prompt, tc.response, tc.generate)
+			tmpl, err := template.Parse(tc.template)
+			if err != nil {
+				t.Fatal(err)
+			}
+
+			got, err := Prompt(tmpl, tc.system, tc.prompt, tc.response, tc.generate)
 			if err != nil {
 				t.Errorf("error = %v", err)
 			}
@ -192,7 +198,12 @@ func TestChatPrompt(t *testing.T) {

 	for _, tc := range tests {
 		t.Run(tc.name, func(t *testing.T) {
-			got, err := ChatPrompt(tc.template, tc.messages, tc.window, encode)
+			tmpl, err := template.Parse(tc.template)
+			if err != nil {
+				t.Fatal(err)
+			}
+
+			got, err := ChatPrompt(tmpl, tc.messages, tc.window, encode)
 			if err != nil {
 				t.Errorf("error = %v", err)
 			}
--- a/server/routes.go
+++ b/server/routes.go
@ -9,7 +9,6 @@ import (
 	"io"
 	"io/fs"
 	"log/slog"
-	"math"
 	"net"
 	"net/http"
 	"net/netip"
@ -17,7 +16,6 @@ import (
 	"os/signal"
 	"path/filepath"
 	"slices"
-	"strconv"
 	"strings"
 	"syscall"
 	"time"
@ -31,6 +29,7 @@ import (
 	"github.com/ollama/ollama/llm"
 	"github.com/ollama/ollama/openai"
 	"github.com/ollama/ollama/parser"
+	"github.com/ollama/ollama/template"
 	"github.com/ollama/ollama/types/errtypes"
 	"github.com/ollama/ollama/types/model"
 	"github.com/ollama/ollama/version"
@ -55,8 +54,6 @@ func init() {
 	gin.SetMode(mode)
 }

-var defaultSessionDuration = 5 * time.Minute
-
 func modelOptions(model *Model, requestOpts map[string]interface{}) (api.Options, error) {
 	opts := api.DefaultOptions()
 	if err := opts.FromMap(model.Options); err != nil {
@ -121,8 +118,8 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 		return
 	}

-	if model.IsEmbedding() {
-		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "embedding models do not support generate"})
+	if !model.Has(CapabilityCompletion) {
+		c.JSON(http.StatusBadRequest, gin.H{"error": fmt.Sprintf("%s does not support generate", req.Model)})
 		return
 	}

@ -132,14 +129,7 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 		return
 	}

-	var sessionDuration time.Duration
-	if req.KeepAlive == nil {
-		sessionDuration = getDefaultSessionDuration()
-	} else {
-		sessionDuration = req.KeepAlive.Duration
-	}
-
-	rCh, eCh := s.sched.GetRunner(c.Request.Context(), model, opts, sessionDuration)
+	rCh, eCh := s.sched.GetRunner(c.Request.Context(), model, opts, req.KeepAlive)
 	var runner *runnerRef
 	select {
 	case runner = <-rCh:
@ -161,6 +151,12 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 		return
 	}

+	tmpl, err := template.Parse(req.Template)
+	if err != nil {
+		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+		return
+	}
+
 	checkpointLoaded := time.Now()

 	var prompt string
@ -169,7 +165,7 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 		prompt = req.Prompt
 	case req.Prompt != "":
 		if req.Template == "" {
-			req.Template = model.Template
+			tmpl = model.Template
 		}

 		if req.System == "" {
@ -187,7 +183,7 @@ func (s *Server) GenerateHandler(c *gin.Context) {

 		sb.WriteString(req.Prompt)

-		p, err := Prompt(req.Template, req.System, sb.String(), "", true)
+		p, err := Prompt(tmpl, req.System, sb.String(), "", true)
 		if err != nil {
 			c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
 			return
@ -242,7 +238,7 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 				resp.LoadDuration = checkpointLoaded.Sub(checkpointStart)

 				if !req.Raw {
-					p, err := Prompt(req.Template, req.System, req.Prompt, generated.String(), false)
+					p, err := Prompt(tmpl, req.System, req.Prompt, generated.String(), false)
 					if err != nil {
 						c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
 						return
@ -313,32 +309,6 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 	streamResponse(c, ch)
 }

-func getDefaultSessionDuration() time.Duration {
-	if envconfig.KeepAlive != "" {
-		v, err := strconv.Atoi(envconfig.KeepAlive)
-		if err != nil {
-			d, err := time.ParseDuration(envconfig.KeepAlive)
-			if err != nil {
-				return defaultSessionDuration
-			}
-
-			if d < 0 {
-				return time.Duration(math.MaxInt64)
-			}
-
-			return d
-		}
-
-		d := time.Duration(v) * time.Second
-		if d < 0 {
-			return time.Duration(math.MaxInt64)
-		}
-		return d
-	}
-
-	return defaultSessionDuration
-}
-
 func (s *Server) EmbeddingsHandler(c *gin.Context) {
 	var req api.EmbeddingRequest
 	err := c.ShouldBindJSON(&req)
@ -373,14 +343,7 @@ func (s *Server) EmbeddingsHandler(c *gin.Context) {
 		return
 	}

-	var sessionDuration time.Duration
-	if req.KeepAlive == nil {
-		sessionDuration = getDefaultSessionDuration()
-	} else {
-		sessionDuration = req.KeepAlive.Duration
-	}
-
-	rCh, eCh := s.sched.GetRunner(c.Request.Context(), model, opts, sessionDuration)
+	rCh, eCh := s.sched.GetRunner(c.Request.Context(), model, opts, req.KeepAlive)
 	var runner *runnerRef
 	select {
 	case runner = <-rCh:
@ -680,7 +643,10 @@ func GetModelInfo(req api.ShowRequest) (*api.ShowResponse, error) {
 	}

 	if req.Template != "" {
-		m.Template = req.Template
+		m.Template, err = template.Parse(req.Template)
+		if err != nil {
+			return nil, err
+		}
 	}

 	msgs := make([]api.Message, 0)
@ -701,7 +667,7 @@ func GetModelInfo(req api.ShowRequest) (*api.ShowResponse, error) {
 	resp := &api.ShowResponse{
 		License:    strings.Join(m.License, "\n"),
 		System:     m.System,
-		Template:   m.Template,
+		Template:   m.Template.String(),
 		Details:    modelDetails,
 		Messages:   msgs,
 		ModifiedAt: manifest.fi.ModTime(),
@ -1039,7 +1005,10 @@ func (s *Server) GenerateRoutes() http.Handler {
 	r.GET("/api/ps", s.ProcessHandler)

 	// Compatibility endpoints
-	r.POST("/v1/chat/completions", openai.Middleware(), s.ChatHandler)
+	r.POST("/v1/chat/completions", openai.ChatMiddleware(), s.ChatHandler)
+	r.POST("/v1/completions", openai.CompletionsMiddleware(), s.GenerateHandler)
+	r.GET("/v1/models", openai.ListMiddleware(), s.ListModelsHandler)
+	r.GET("/v1/models/:model", openai.RetrieveMiddleware(), s.ShowModelHandler)

 	for _, method := range []string{http.MethodGet, http.MethodHead} {
 		r.Handle(method, "/", func(c *gin.Context) {
@ -1237,11 +1206,16 @@ func (s *Server) ProcessHandler(c *gin.Context) {
 		models = append(models, mr)
 	}

+	slices.SortStableFunc(models, func(i, j api.ProcessModelResponse) int {
+		// longest duration remaining listed first
+		return cmp.Compare(j.ExpiresAt.Unix(), i.ExpiresAt.Unix())
+	})
+
 	c.JSON(http.StatusOK, api.ProcessResponse{Models: models})
 }

 // ChatPrompt builds up a prompt from a series of messages for the currently `loaded` model
-func chatPrompt(ctx context.Context, runner *runnerRef, template string, messages []api.Message, numCtx int) (string, error) {
+func chatPrompt(ctx context.Context, runner *runnerRef, template *template.Template, messages []api.Message, numCtx int) (string, error) {
 	encode := func(s string) ([]int, error) {
 		return runner.llama.Tokenize(ctx, s)
 	}
@ -1289,8 +1263,8 @@ func (s *Server) ChatHandler(c *gin.Context) {
 		return
 	}

-	if model.IsEmbedding() {
-		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "embedding models do not support chat"})
+	if !model.Has(CapabilityCompletion) {
+		c.JSON(http.StatusBadRequest, gin.H{"error": fmt.Sprintf("%s does not support chat", req.Model)})
 		return
 	}

@ -1300,14 +1274,7 @@ func (s *Server) ChatHandler(c *gin.Context) {
 		return
 	}

-	var sessionDuration time.Duration
-	if req.KeepAlive == nil {
-		sessionDuration = getDefaultSessionDuration()
-	} else {
-		sessionDuration = req.KeepAlive.Duration
-	}
-
-	rCh, eCh := s.sched.GetRunner(c.Request.Context(), model, opts, sessionDuration)
+	rCh, eCh := s.sched.GetRunner(c.Request.Context(), model, opts, req.KeepAlive)
 	var runner *runnerRef
 	select {
 	case runner = <-rCh:
--- a/server/routes_test.go
+++ b/server/routes_test.go
@ -20,6 +20,7 @@ import (
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/llm"
+	"github.com/ollama/ollama/openai"
 	"github.com/ollama/ollama/parser"
 	"github.com/ollama/ollama/types/model"
 	"github.com/ollama/ollama/version"
@ -105,6 +106,24 @@ func Test_Routes(t *testing.T) {
 				assert.Empty(t, len(modelList.Models))
 			},
 		},
+		{
+			Name:   "openai empty list",
+			Method: http.MethodGet,
+			Path:   "/v1/models",
+			Expected: func(t *testing.T, resp *http.Response) {
+				contentType := resp.Header.Get("Content-Type")
+				assert.Equal(t, "application/json", contentType)
+				body, err := io.ReadAll(resp.Body)
+				require.NoError(t, err)
+
+				var modelList openai.ListCompletion
+				err = json.Unmarshal(body, &modelList)
+				require.NoError(t, err)
+
+				assert.Equal(t, "list", modelList.Object)
+				assert.Empty(t, modelList.Data)
+			},
+		},
 		{
 			Name:   "Tags Handler (yes tags)",
 			Method: http.MethodGet,
@ -128,6 +147,25 @@ func Test_Routes(t *testing.T) {
 				assert.Equal(t, "test-model:latest", modelList.Models[0].Name)
 			},
 		},
+		{
+			Name:   "openai list models with tags",
+			Method: http.MethodGet,
+			Path:   "/v1/models",
+			Expected: func(t *testing.T, resp *http.Response) {
+				contentType := resp.Header.Get("Content-Type")
+				assert.Equal(t, "application/json", contentType)
+				body, err := io.ReadAll(resp.Body)
+				require.NoError(t, err)
+
+				var modelList openai.ListCompletion
+				err = json.Unmarshal(body, &modelList)
+				require.NoError(t, err)
+
+				assert.Len(t, modelList.Data, 1)
+				assert.Equal(t, "test-model:latest", modelList.Data[0].Id)
+				assert.Equal(t, "library", modelList.Data[0].OwnedBy)
+			},
+		},
 		{
 			Name:   "Create Model Handler",
 			Method: http.MethodPost,
@ -216,6 +254,24 @@ func Test_Routes(t *testing.T) {
 				assert.InDelta(t, 0, showResp.ModelInfo["general.parameter_count"], 1e-9, "Parameter count should be 0")
 			},
 		},
+		{
+			Name:   "openai retrieve model handler",
+			Method: http.MethodGet,
+			Path:   "/v1/models/show-model",
+			Expected: func(t *testing.T, resp *http.Response) {
+				contentType := resp.Header.Get("Content-Type")
+				assert.Equal(t, "application/json", contentType)
+				body, err := io.ReadAll(resp.Body)
+				require.NoError(t, err)
+
+				var retrieveResp api.RetrieveModelResponse
+				err = json.Unmarshal(body, &retrieveResp)
+				require.NoError(t, err)
+
+				assert.Equal(t, "show-model", retrieveResp.Id)
+				assert.Equal(t, "library", retrieveResp.OwnedBy)
+			},
+		},
 	}

 	t.Setenv("OLLAMA_MODELS", t.TempDir())
--- a/server/sched.go
+++ b/server/sched.go
@ -23,7 +23,8 @@ type LlmRequest struct {
 	ctx             context.Context //nolint:containedctx
 	model           *Model
 	opts            api.Options
-	sessionDuration time.Duration
+	origNumCtx      int // Track the initial ctx request
+	sessionDuration *api.Duration
 	successCh       chan *runnerRef
 	errCh           chan error
 	schedAttempts   uint
@ -38,13 +39,23 @@ type Scheduler struct {
 	loaded   map[string]*runnerRef
 	loadedMu sync.Mutex

-	loadFn       func(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList)
-	newServerFn  func(gpus gpu.GpuInfoList, model string, ggml *llm.GGML, adapters []string, projectors []string, opts api.Options) (llm.LlamaServer, error)
+	loadFn       func(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList, numParallel int)
+	newServerFn  func(gpus gpu.GpuInfoList, model string, ggml *llm.GGML, adapters []string, projectors []string, opts api.Options, numParallel int) (llm.LlamaServer, error)
 	getGpuFn     func() gpu.GpuInfoList
 	getCpuFn     func() gpu.GpuInfoList
 	reschedDelay time.Duration
 }

+// Default automatic value for number of models we allow per GPU
+// Model will still need to fit in VRAM, but loading many small models
+// on a large GPU can cause stalling
+var defaultModelsPerGPU = 3
+
+// Default automatic value for parallel setting
+// Model will still need to fit in VRAM.  If this setting wont fit
+// we'll back off down to 1 to try to get it to fit
+var defaultParallel = 4
+
 var ErrMaxQueue = fmt.Errorf("server busy, please try again.  maximum pending requests exceeded")

 func InitScheduler(ctx context.Context) *Scheduler {
@ -64,14 +75,11 @@ func InitScheduler(ctx context.Context) *Scheduler {
 }

 // context must be canceled to decrement ref count and release the runner
-func (s *Scheduler) GetRunner(c context.Context, model *Model, opts api.Options, sessionDuration time.Duration) (chan *runnerRef, chan error) {
-	// allocate a large enough kv cache for all parallel requests
+func (s *Scheduler) GetRunner(c context.Context, model *Model, opts api.Options, sessionDuration *api.Duration) (chan *runnerRef, chan error) {
 	if opts.NumCtx < 4 {
 		opts.NumCtx = 4
 	}

-	opts.NumCtx *= envconfig.NumParallel
-
 	req := &LlmRequest{
 		ctx:             c,
 		model:           model,
@ -110,13 +118,32 @@ func (s *Scheduler) processPending(ctx context.Context) {
 		case pending := <-s.pendingReqCh:
 			// Block other requests until we get this pending request running
 			pending.schedAttempts++
+			if pending.origNumCtx == 0 {
+				pending.origNumCtx = pending.opts.NumCtx
+			}

 			if pending.ctx.Err() != nil {
 				slog.Debug("pending request cancelled or timed out, skipping scheduling")
 				continue
 			}
+			numParallel := envconfig.NumParallel
+			// TODO (jmorganca): multimodal models don't support parallel yet
+			// see https://github.com/ollama/ollama/issues/4165
+			if len(pending.model.ProjectorPaths) > 0 && numParallel != 1 {
+				numParallel = 1
+				slog.Warn("multimodal models don't support parallel requests yet")
+			}
+			// Keep NumCtx and numParallel in sync
+			if numParallel > 1 {
+				pending.opts.NumCtx = pending.origNumCtx * numParallel
+			}

 			for {
+				cpus := s.getCpuFn()
+				var systemMem gpu.GpuInfo
+				if len(cpus) > 0 {
+					systemMem = cpus[0]
+				}
 				var runnerToExpire *runnerRef
 				s.loadedMu.Lock()
 				runner := s.loaded[pending.model.ModelPath]
@ -143,6 +170,26 @@ func (s *Scheduler) processPending(ctx context.Context) {
 						gpus = s.getGpuFn()
 					}

+					if envconfig.MaxRunners <= 0 {
+						// No user specified MaxRunners, so figure out what automatic setting to use
+						// If all GPUs have reliable free memory reporting, defaultModelsPerGPU * the number of GPUs
+						// if any GPU has unreliable free memory reporting, 1x the number of GPUs
+						allReliable := true
+						for _, gpu := range gpus {
+							if gpu.UnreliableFreeMemory {
+								allReliable = false
+								break
+							}
+						}
+						if allReliable {
+							envconfig.MaxRunners = defaultModelsPerGPU * len(gpus)
+							slog.Debug("updating default concurrency", "OLLAMA_MAX_LOADED_MODELS", envconfig.MaxRunners, "gpu_count", len(gpus))
+						} else {
+							slog.Info("one or more GPUs detected that are unable to accurately report free memory - disabling default concurrency")
+							envconfig.MaxRunners = len(gpus)
+						}
+					}
+
 					// Load model for fitting
 					ggml, err := llm.LoadModel(pending.model.ModelPath, 0)
 					if err != nil {
@ -150,28 +197,55 @@ func (s *Scheduler) processPending(ctx context.Context) {
 						break
 					}

+					// Block attempting to load a model larger than system memory + GPU memory
+					estimate := llm.EstimateGPULayers(gpus, ggml, pending.model.ProjectorPaths, pending.opts)
+					maxSize := systemMem.FreeMemory
+					for _, gpu := range gpus {
+						if gpu.Library == "cpu" {
+							continue
+						}
+						if loadedCount == 0 {
+							// If no other models are loaded, set the limit based on what's available
+							maxSize += gpu.FreeMemory
+						} else {
+							// Other models could be unloaded, favor total memory for limit
+							maxSize += gpu.TotalMemory
+						}
+					}
+					if estimate.TotalSize > maxSize {
+						slog.Warn("model request too large for system", "requested", format.HumanBytes2(estimate.TotalSize), "system", format.HumanBytes2(maxSize))
+						pending.errCh <- fmt.Errorf("requested model (%s) is too large for this system (%s)", format.HumanBytes2(estimate.TotalSize), format.HumanBytes2(maxSize))
+						break
+					}
+
 					// Evaluate if the model will fit in the available system memory, or if we should unload a model first
 					if len(gpus) == 1 && gpus[0].Library == "cpu" {
+						// simplifying assumption of defaultParallel when in CPU mode
+						if numParallel <= 0 {
+							numParallel = defaultParallel
+							pending.opts.NumCtx = pending.origNumCtx * numParallel
+						}
+
 						if loadedCount == 0 {
 							slog.Debug("cpu mode with first model, loading")
-							s.loadFn(pending, ggml, gpus)
+							s.loadFn(pending, ggml, gpus, numParallel)
 							break
 						}
 						runnerToExpire = s.maybeFindCPURunnerToUnload(pending, ggml, gpus)
 						if runnerToExpire == nil {
 							slog.Debug("cpu mode with available system memory or first model, loading")
-							s.loadFn(pending, ggml, gpus)
+							s.loadFn(pending, ggml, gpus, numParallel)
 							break
 						}
 						// else we need to expire a runner
 					} else if loadedCount == 0 {
 						// No models loaded. Load the model but prefer the best fit.
 						slog.Debug("loading first model", "model", pending.model.ModelPath)
-						g := pickBestFitGPUs(pending, ggml, gpus)
+						g := pickBestFitGPUs(pending, ggml, gpus, &numParallel)
 						if g != nil {
 							gpus = g
 						}
-						s.loadFn(pending, ggml, gpus)
+						s.loadFn(pending, ggml, gpus, numParallel)
 						break
 					}

@ -186,10 +260,10 @@ func (s *Scheduler) processPending(ctx context.Context) {

 						// Update free memory from currently loaded models
 						s.updateFreeSpace(availGpus)
-						fitGpus := pickBestFitGPUs(pending, ggml, availGpus)
+						fitGpus := pickBestFitGPUs(pending, ggml, availGpus, &numParallel)
 						if fitGpus != nil {
 							slog.Debug("new model fits with existing models, loading")
-							s.loadFn(pending, ggml, fitGpus)
+							s.loadFn(pending, ggml, fitGpus, numParallel)
 							break
 						}

@ -341,7 +415,9 @@ func (pending *LlmRequest) useLoadedRunner(runner *runnerRef, finished chan *Llm
 		runner.expireTimer.Stop()
 		runner.expireTimer = nil
 	}
-	runner.sessionDuration = pending.sessionDuration
+	if pending.sessionDuration != nil {
+		runner.sessionDuration = pending.sessionDuration.Duration
+	}
 	pending.successCh <- runner
 	go func() {
 		<-pending.ctx.Done()
@ -350,8 +426,15 @@ func (pending *LlmRequest) useLoadedRunner(runner *runnerRef, finished chan *Llm
 	}()
 }

-func (s *Scheduler) load(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList) {
-	llama, err := s.newServerFn(gpus, req.model.ModelPath, ggml, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts)
+func (s *Scheduler) load(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList, numParallel int) {
+	if numParallel < 1 {
+		numParallel = 1
+	}
+	sessionDuration := envconfig.KeepAlive
+	if req.sessionDuration != nil {
+		sessionDuration = req.sessionDuration.Duration
+	}
+	llama, err := s.newServerFn(gpus, req.model.ModelPath, ggml, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts, numParallel)
 	if err != nil {
 		// some older models are not compatible with newer versions of llama.cpp
 		// show a generalized compatibility error until there is a better way to
@ -368,13 +451,14 @@ func (s *Scheduler) load(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList)
 		modelPath:       req.model.ModelPath,
 		llama:           llama,
 		Options:         &req.opts,
-		sessionDuration: req.sessionDuration,
+		sessionDuration: sessionDuration,
 		gpus:            gpus,
 		estimatedVRAM:   llama.EstimatedVRAM(),
 		estimatedTotal:  llama.EstimatedTotal(),
 		loading:         true,
 		refCount:        1,
 	}
+	runner.numParallel = numParallel
 	runner.refMu.Lock()

 	s.loadedMu.Lock()
@ -483,8 +567,9 @@ type runnerRef struct {
 	expireTimer     *time.Timer
 	expiresAt       time.Time

-	model     *Model
-	modelPath string
+	model       *Model
+	modelPath   string
+	numParallel int
 	*api.Options
 }

@ -525,6 +610,9 @@ func (runner *runnerRef) needsReload(ctx context.Context, req *LlmRequest) bool
 		optsNew.NumGPU = -1
 	}

+	// Normalize the NumCtx for parallelism
+	optsExisting.NumCtx = optsExisting.NumCtx / runner.numParallel
+
 	ctx, cancel := context.WithTimeout(ctx, timeout)
 	defer cancel()
 	if !reflect.DeepEqual(runner.model.AdapterPaths, req.model.AdapterPaths) || // have the adapters changed?
@ -611,22 +699,38 @@ func (a ByDuration) Less(i, j int) bool {

 // pickBestFitGPUs will try to find the optimal placement of the model in the available GPUs where the model fully fits
 // If the model can not be fit fully within the available GPU(s) nil is returned
-func pickBestFitGPUs(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList) gpu.GpuInfoList {
+// If numParallel is <= 0, this will attempt try to optimize parallism based on available VRAM, and adjust
+// opts.NumCtx accordingly
+func pickBestFitGPUs(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList, numParallel *int) gpu.GpuInfoList {
 	var estimatedVRAM uint64
+
+	var numParallelToTry []int
+	if *numParallel <= 0 {
+		// If no specific parallel setting was provided, try larger then smaller, always end with 1
+		numParallelToTry = append(numParallelToTry, defaultParallel, 1)
+	} else {
+		numParallelToTry = []int{*numParallel}
+	}
+
 	for _, gl := range gpus.ByLibrary() {
 		var ok bool
 		sgl := append(make(gpu.GpuInfoList, 0, len(gl)), gl...)

 		// TODO - potentially sort by performance capability, existing models loaded, etc.
+		// TODO - Eliminate any GPUs that already have envconfig.MaxRunners loaded on them
 		// Note: at present, this will favor more VRAM over faster GPU speed in mixed setups
 		sort.Sort(sort.Reverse(gpu.ByFreeMemory(sgl)))

 		// First attempt to fit the model into a single GPU
-		if !envconfig.SchedSpread {
-			for _, g := range sgl {
-				if ok, estimatedVRAM = llm.PredictServerFit([]gpu.GpuInfo{g}, ggml, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts); ok {
-					slog.Debug("new model will fit in available VRAM in single GPU, loading", "model", req.model.ModelPath, "gpu", g.ID, "available", g.FreeMemory, "required", format.HumanBytes2(estimatedVRAM))
-					return []gpu.GpuInfo{g}
+		for _, p := range numParallelToTry {
+			req.opts.NumCtx = req.origNumCtx * p
+			if !envconfig.SchedSpread {
+				for _, g := range sgl {
+					if ok, estimatedVRAM = llm.PredictServerFit([]gpu.GpuInfo{g}, ggml, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts); ok {
+						slog.Info("new model will fit in available VRAM in single GPU, loading", "model", req.model.ModelPath, "gpu", g.ID, "parallel", p, "available", g.FreeMemory, "required", format.HumanBytes2(estimatedVRAM))
+						*numParallel = p
+						return []gpu.GpuInfo{g}
+					}
 				}
 			}
 		}
@ -636,9 +740,13 @@ func pickBestFitGPUs(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList) gpu.
 		// - try subsets of GPUs instead of just falling back to 1 or all in a family

 		// Now try all the GPUs
-		if ok, estimatedVRAM = llm.PredictServerFit(sgl, ggml, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts); ok {
-			slog.Debug("new model will fit in available VRAM, loading", "model", req.model.ModelPath, "library", sgl[0].Library, "required", format.HumanBytes2(estimatedVRAM))
-			return sgl
+		for _, p := range numParallelToTry {
+			req.opts.NumCtx = req.origNumCtx * p
+			if ok, estimatedVRAM = llm.PredictServerFit(sgl, ggml, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts); ok {
+				slog.Info("new model will fit in available VRAM, loading", "model", req.model.ModelPath, "library", sgl[0].Library, "parallel", p, "required", format.HumanBytes2(estimatedVRAM))
+				*numParallel = p
+				return sgl
+			}
 		}
 	}
 	return nil
--- a/server/sched_test.go
+++ b/server/sched_test.go
@ -44,14 +44,14 @@ func TestLoad(t *testing.T) {
 		opts:            api.DefaultOptions(),
 		successCh:       make(chan *runnerRef, 1),
 		errCh:           make(chan error, 1),
-		sessionDuration: 2,
+		sessionDuration: &api.Duration{Duration: 2 * time.Second},
 	}
 	// Fail to load model first
-	s.newServerFn = func(gpus gpu.GpuInfoList, model string, ggml *llm.GGML, adapters []string, projectors []string, opts api.Options) (llm.LlamaServer, error) {
+	s.newServerFn = func(gpus gpu.GpuInfoList, model string, ggml *llm.GGML, adapters []string, projectors []string, opts api.Options, numParallel int) (llm.LlamaServer, error) {
 		return nil, fmt.Errorf("something failed to load model blah")
 	}
 	gpus := gpu.GpuInfoList{}
-	s.load(req, ggml, gpus)
+	s.load(req, ggml, gpus, 0)
 	require.Empty(t, req.successCh)
 	require.Len(t, req.errCh, 1)
 	s.loadedMu.Lock()
@ -61,10 +61,10 @@ func TestLoad(t *testing.T) {
 	require.Contains(t, err.Error(), "this model may be incompatible")

 	server := &mockLlm{estimatedVRAM: 10, estimatedVRAMByGPU: map[string]uint64{}}
-	s.newServerFn = func(gpus gpu.GpuInfoList, model string, ggml *llm.GGML, adapters []string, projectors []string, opts api.Options) (llm.LlamaServer, error) {
+	s.newServerFn = func(gpus gpu.GpuInfoList, model string, ggml *llm.GGML, adapters []string, projectors []string, opts api.Options, numParallel int) (llm.LlamaServer, error) {
 		return server, nil
 	}
-	s.load(req, ggml, gpus)
+	s.load(req, ggml, gpus, 0)
 	select {
 	case err := <-req.errCh:
 		require.NoError(t, err)
@ -78,12 +78,12 @@ func TestLoad(t *testing.T) {

 	req.model.ModelPath = "dummy_model_path"
 	server.waitResp = fmt.Errorf("wait failure")
-	s.load(req, ggml, gpus)
+	s.load(req, ggml, gpus, 0)
 	select {
 	case err := <-req.errCh:
 		require.Contains(t, err.Error(), "wait failure")
 	case resp := <-req.successCh:
-		t.Errorf("unexpected success %v", resp)
+		t.Fatalf("unexpected success %v", resp)
 	}
 	s.loadedMu.Lock()
 	runner := s.loaded["dummy_model_path"]
@ -102,7 +102,7 @@ type bundle struct {
 	ggml    *llm.GGML
 }

-func (scenario *bundle) newServer(gpus gpu.GpuInfoList, model string, ggml *llm.GGML, adapters []string, projectors []string, opts api.Options) (llm.LlamaServer, error) {
+func (scenario *bundle) newServer(gpus gpu.GpuInfoList, model string, ggml *llm.GGML, adapters []string, projectors []string, opts api.Options, numParallel int) (llm.LlamaServer, error) {
 	return scenario.srv, nil
 }

@ -142,7 +142,7 @@ func newScenario(t *testing.T, ctx context.Context, modelName string, estimatedV
 		ctx:             scenario.ctx,
 		model:           model,
 		opts:            api.DefaultOptions(),
-		sessionDuration: 5 * time.Millisecond,
+		sessionDuration: &api.Duration{Duration: 5 * time.Millisecond},
 		successCh:       make(chan *runnerRef, 1),
 		errCh:           make(chan error, 1),
 	}
@ -156,18 +156,18 @@ func TestRequests(t *testing.T) {

 	// Same model, same request
 	scenario1a := newScenario(t, ctx, "ollama-model-1", 10)
-	scenario1a.req.sessionDuration = 5 * time.Millisecond
+	scenario1a.req.sessionDuration = &api.Duration{Duration: 5 * time.Millisecond}
 	scenario1b := newScenario(t, ctx, "ollama-model-1", 11)
 	scenario1b.req.model = scenario1a.req.model
 	scenario1b.ggml = scenario1a.ggml
-	scenario1b.req.sessionDuration = 0
+	scenario1b.req.sessionDuration = &api.Duration{Duration: 0}

 	// simple reload of same model
 	scenario2a := newScenario(t, ctx, "ollama-model-1", 20)
 	tmpModel := *scenario1a.req.model
 	scenario2a.req.model = &tmpModel
 	scenario2a.ggml = scenario1a.ggml
-	scenario2a.req.sessionDuration = 5 * time.Millisecond
+	scenario2a.req.sessionDuration = &api.Duration{Duration: 5 * time.Millisecond}

 	// Multiple loaded models
 	scenario3a := newScenario(t, ctx, "ollama-model-3a", 1*format.GigaByte)
@ -199,8 +199,10 @@ func TestRequests(t *testing.T) {
 		require.Equal(t, resp.llama, scenario1a.srv)
 		require.Empty(t, s.pendingReqCh)
 		require.Empty(t, scenario1a.req.errCh)
+	case err := <-scenario1a.req.errCh:
+		t.Fatal(err.Error())
 	case <-ctx.Done():
-		t.Errorf("timeout")
+		t.Fatal("timeout")
 	}

 	// Same runner as first request due to not needing a reload
@ -212,8 +214,10 @@ func TestRequests(t *testing.T) {
 		require.Equal(t, resp.llama, scenario1a.srv)
 		require.Empty(t, s.pendingReqCh)
 		require.Empty(t, scenario1b.req.errCh)
+	case err := <-scenario1b.req.errCh:
+		t.Fatal(err.Error())
 	case <-ctx.Done():
-		t.Errorf("timeout")
+		t.Fatal("timeout")
 	}

 	// Trigger a reload
@ -230,8 +234,10 @@ func TestRequests(t *testing.T) {
 		require.Equal(t, resp.llama, scenario2a.srv)
 		require.Empty(t, s.pendingReqCh)
 		require.Empty(t, scenario2a.req.errCh)
+	case err := <-scenario2a.req.errCh:
+		t.Fatal(err.Error())
 	case <-ctx.Done():
-		t.Errorf("timeout")
+		t.Fatal("timeout")
 	}

 	envconfig.MaxRunners = 1
@ -246,8 +252,10 @@ func TestRequests(t *testing.T) {
 		require.Equal(t, resp.llama, scenario3a.srv)
 		require.Empty(t, s.pendingReqCh)
 		require.Empty(t, scenario3a.req.errCh)
+	case err := <-scenario3a.req.errCh:
+		t.Fatal(err.Error())
 	case <-ctx.Done():
-		t.Errorf("timeout")
+		t.Fatal("timeout")
 	}
 	s.loadedMu.Lock()
 	require.Len(t, s.loaded, 1)
@ -262,8 +270,10 @@ func TestRequests(t *testing.T) {
 		require.Equal(t, resp.llama, scenario3b.srv)
 		require.Empty(t, s.pendingReqCh)
 		require.Empty(t, scenario3b.req.errCh)
+	case err := <-scenario3b.req.errCh:
+		t.Fatal(err.Error())
 	case <-ctx.Done():
-		t.Errorf("timeout")
+		t.Fatal("timeout")
 	}
 	s.loadedMu.Lock()
 	require.Len(t, s.loaded, 2)
@ -278,8 +288,10 @@ func TestRequests(t *testing.T) {
 		require.Equal(t, resp.llama, scenario3c.srv)
 		require.Empty(t, s.pendingReqCh)
 		require.Empty(t, scenario3c.req.errCh)
+	case err := <-scenario3c.req.errCh:
+		t.Fatal(err.Error())
 	case <-ctx.Done():
-		t.Errorf("timeout")
+		t.Fatal("timeout")
 	}
 	s.loadedMu.Lock()
 	require.Len(t, s.loaded, 3)
@ -306,7 +318,7 @@ func TestRequests(t *testing.T) {
 		require.Empty(t, s.pendingReqCh)
 		require.Empty(t, scenario3d.req.errCh)
 	case <-ctx.Done():
-		t.Errorf("timeout")
+		t.Fatal("timeout")
 	}
 	s.loadedMu.Lock()
 	require.Len(t, s.loaded, 2)
@ -318,11 +330,11 @@ func TestGetRunner(t *testing.T) {
 	defer done()

 	scenario1a := newScenario(t, ctx, "ollama-model-1a", 10)
-	scenario1a.req.sessionDuration = 0
+	scenario1a.req.sessionDuration = &api.Duration{Duration: 0}
 	scenario1b := newScenario(t, ctx, "ollama-model-1b", 10)
-	scenario1b.req.sessionDuration = 0
+	scenario1b.req.sessionDuration = &api.Duration{Duration: 0}
 	scenario1c := newScenario(t, ctx, "ollama-model-1c", 10)
-	scenario1c.req.sessionDuration = 0
+	scenario1c.req.sessionDuration = &api.Duration{Duration: 0}
 	envconfig.MaxQueuedRequests = 1
 	s := InitScheduler(ctx)
 	s.getGpuFn = func() gpu.GpuInfoList {
@ -349,7 +361,7 @@ func TestGetRunner(t *testing.T) {
 		require.Empty(t, s.pendingReqCh)
 		require.Empty(t, errCh1a)
 	case <-ctx.Done():
-		t.Errorf("timeout")
+		t.Fatal("timeout")
 	}
 	scenario1a.ctxDone()
 	s.loadedMu.Lock()
@ -400,9 +412,9 @@ func TestPrematureExpired(t *testing.T) {
 		slog.Info("sending premature expired event now")
 		s.expiredCh <- resp // Shouldn't happen in real life, but make sure its safe
 	case <-ctx.Done():
-		t.Errorf("timeout")
+		t.Fatal("timeout")
 	}
-	time.Sleep(scenario1a.req.sessionDuration)
+	time.Sleep(scenario1a.req.sessionDuration.Duration)
 	scenario1a.ctxDone()
 	time.Sleep(20 * time.Millisecond)
 	require.LessOrEqual(t, len(s.finishedReqCh), 1)
@ -423,11 +435,11 @@ func TestUseLoadedRunner(t *testing.T) {
 		ctx:             ctx,
 		opts:            api.DefaultOptions(),
 		successCh:       make(chan *runnerRef, 1),
-		sessionDuration: 2,
+		sessionDuration: &api.Duration{Duration: 2},
 	}
 	finished := make(chan *LlmRequest)
 	llm1 := &mockLlm{estimatedVRAMByGPU: map[string]uint64{}}
-	r1 := &runnerRef{llama: llm1, sessionDuration: 1}
+	r1 := &runnerRef{llama: llm1, sessionDuration: 1, numParallel: 1}
 	req.useLoadedRunner(r1, finished)
 	require.Equal(t, uint(1), r1.refCount)
 	require.Equal(t, time.Duration(2), r1.sessionDuration)
@ -435,7 +447,7 @@ func TestUseLoadedRunner(t *testing.T) {
 	case success := <-req.successCh:
 		require.Equal(t, r1, success)
 	case <-ctx.Done():
-		t.Errorf("timeout")
+		t.Fatal("timeout")
 	}
 	done()
 	fin := <-finished
@ -461,8 +473,8 @@ func TestUpdateFreeSpace(t *testing.T) {
 	gpus[1].FreeMemory = 1900
 	llm1 := &mockLlm{estimatedVRAMByGPU: map[string]uint64{"1": 50, "2": 50}}
 	llm2 := &mockLlm{estimatedVRAMByGPU: map[string]uint64{"1": 125, "2": 75}}
-	r1 := &runnerRef{llama: llm1, gpus: gpus}
-	r2 := &runnerRef{llama: llm2, gpus: gpus}
+	r1 := &runnerRef{llama: llm1, gpus: gpus, numParallel: 1}
+	r2 := &runnerRef{llama: llm2, gpus: gpus, numParallel: 1}

 	s := InitScheduler(ctx)
 	s.loadedMu.Lock()
@ -513,8 +525,8 @@ func TestFindRunnerToUnload(t *testing.T) {
 	ctx, done := context.WithTimeout(context.Background(), 100*time.Millisecond)
 	defer done()

-	r1 := &runnerRef{refCount: 1, sessionDuration: 1}
-	r2 := &runnerRef{sessionDuration: 2}
+	r1 := &runnerRef{refCount: 1, sessionDuration: 1, numParallel: 1}
+	r2 := &runnerRef{sessionDuration: 2, numParallel: 1}

 	s := InitScheduler(ctx)
 	s.loadedMu.Lock()
@ -536,9 +548,13 @@ func TestNeedsReload(t *testing.T) {
 	llm := &mockLlm{estimatedVRAMByGPU: map[string]uint64{}}
 	do := api.DefaultOptions()
 	runner := &runnerRef{
-		model:   &Model{AdapterPaths: []string{"adapter1"}, ProjectorPaths: []string{"projector1"}},
-		Options: &do,
-		llama:   llm,
+		model: &Model{
+			AdapterPaths:   []string{"adapter1"},
+			ProjectorPaths: []string{"projector1"},
+		},
+		Options:     &do,
+		llama:       llm,
+		numParallel: 1,
 	}
 	req := &LlmRequest{
 		model: &Model{
@ -581,8 +597,8 @@ func TestUnloadAllRunners(t *testing.T) {
 	s := InitScheduler(ctx)
 	s.unloadAllRunners()

-	r1 := &runnerRef{llama: llm1}
-	r2 := &runnerRef{llama: llm2}
+	r1 := &runnerRef{llama: llm1, numParallel: 1}
+	r2 := &runnerRef{llama: llm2, numParallel: 1}

 	s.loadedMu.Lock()
 	s.loaded["a"] = r1
@ -596,14 +612,32 @@ func TestUnloadAllRunners(t *testing.T) {

 func TestUnload(t *testing.T) {
 	llm1 := &mockLlm{estimatedVRAMByGPU: map[string]uint64{}}
-	r1 := &runnerRef{llama: llm1}
-	r2 := &runnerRef{model: &Model{AdapterPaths: []string{"A"}}}
+	r1 := &runnerRef{llama: llm1, numParallel: 1}
+	r2 := &runnerRef{model: &Model{AdapterPaths: []string{"A"}}, numParallel: 1}
 	r1.unload()
 	require.True(t, llm1.closeCalled)
 	r2.unload()
 	require.Nil(t, r2.model)
 }

+func TestAlreadyCanceled(t *testing.T) {
+	ctx, done := context.WithTimeout(context.Background(), 500*time.Millisecond)
+	defer done()
+	dctx, done2 := context.WithCancel(ctx)
+	done2()
+	scenario1a := newScenario(t, dctx, "ollama-model-1", 10)
+	scenario1a.req.sessionDuration = &api.Duration{Duration: 0}
+	s := InitScheduler(ctx)
+	slog.Info("scenario1a")
+	s.pendingReqCh <- scenario1a.req
+	require.Len(t, s.pendingReqCh, 1)
+	s.Run(ctx)
+	time.Sleep(5 * time.Millisecond)
+	require.Empty(t, s.pendingReqCh)
+	require.Empty(t, scenario1a.req.errCh)
+	require.Empty(t, scenario1a.req.successCh)
+}
+
 type mockLlm struct {
 	pingResp           error
 	waitResp           error
--- a/templates/alfred.gotmpl
+++ b/templates/alfred.gotmpl
--- a/templates/alpaca.gotmpl
+++ b/templates/alpaca.gotmpl
--- a/templates/chatml.gotmpl
+++ b/templates/chatml.gotmpl
--- a/templates/chatqa.gotmpl
+++ b/templates/chatqa.gotmpl
--- a/templates/codellama-70b-instruct.gotmpl
+++ b/templates/codellama-70b-instruct.gotmpl
--- a/templates/falcon-instruct.gotmpl
+++ b/templates/falcon-instruct.gotmpl
--- a/templates/gemma-instruct.gotmpl
+++ b/templates/gemma-instruct.gotmpl
--- a/templates/granite-instruct.gotmpl
+++ b/templates/granite-instruct.gotmpl
--- a/templates/index.json
+++ b/templates/index.json
--- a/templates/llama2-chat.gotmpl
+++ b/templates/llama2-chat.gotmpl
--- a/templates/llama3-instruct.gotmpl
+++ b/templates/llama3-instruct.gotmpl
--- a/templates/magicoder.gotmpl
+++ b/templates/magicoder.gotmpl
--- a/templates/mistral-instruct.gotmpl
+++ b/templates/mistral-instruct.gotmpl
--- a/templates/openchat.gotmpl
+++ b/templates/openchat.gotmpl
--- a/templates/phi-3.gotmpl
+++ b/templates/phi-3.gotmpl
--- a/templates/solar-instruct.gotmpl
+++ b/templates/solar-instruct.gotmpl
--- a/templates/starcoder2-instruct.gotmpl
+++ b/templates/starcoder2-instruct.gotmpl
--- a/template/template.go
+++ b/template/template.go
@ -0,0 +1,158 @@
+package template
+
+import (
+	"bytes"
+	"embed"
+	"encoding/json"
+	"errors"
+	"io"
+	"math"
+	"slices"
+	"strings"
+	"sync"
+	"text/template"
+	"text/template/parse"
+
+	"github.com/agnivade/levenshtein"
+	"golang.org/x/exp/maps"
+)
+
+//go:embed index.json
+var indexBytes []byte
+
+//go:embed *.gotmpl
+var templatesFS embed.FS
+
+var templatesOnce = sync.OnceValues(func() ([]*named, error) {
+	var templates []*named
+	if err := json.Unmarshal(indexBytes, &templates); err != nil {
+		return nil, err
+	}
+
+	for _, t := range templates {
+		bts, err := templatesFS.ReadFile(t.Name + ".gotmpl")
+		if err != nil {
+			return nil, err
+		}
+
+		// normalize line endings
+		t.Bytes = bytes.ReplaceAll(bts, []byte("\r\n"), []byte("\n"))
+	}
+
+	return templates, nil
+})
+
+type named struct {
+	Name     string `json:"name"`
+	Template string `json:"template"`
+	Bytes    []byte
+}
+
+func (t named) Reader() io.Reader {
+	return bytes.NewReader(t.Bytes)
+}
+
+func Named(s string) (*named, error) {
+	templates, err := templatesOnce()
+	if err != nil {
+		return nil, err
+	}
+
+	var template *named
+	score := math.MaxInt
+	for _, t := range templates {
+		if s := levenshtein.ComputeDistance(s, t.Template); s < score {
+			score = s
+			template = t
+		}
+	}
+
+	if score < 100 {
+		return template, nil
+	}
+
+	return nil, errors.New("no matching template found")
+}
+
+type Template struct {
+	*template.Template
+	raw string
+}
+
+func (t *Template) String() string {
+	return t.raw
+}
+
+var DefaultTemplate, _ = Parse("{{ .Prompt }}")
+
+func Parse(s string) (*Template, error) {
+	t, err := template.New("").Option("missingkey=zero").Parse(s)
+	if err != nil {
+		return nil, err
+	}
+
+	return &Template{Template: t, raw: s}, nil
+}
+
+func (t *Template) Vars() []string {
+	var vars []string
+	for _, n := range t.Tree.Root.Nodes {
+		vars = append(vars, parseNode(n)...)
+	}
+
+	set := make(map[string]struct{})
+	for _, n := range vars {
+		set[strings.ToLower(n)] = struct{}{}
+	}
+
+	vars = maps.Keys(set)
+	slices.Sort(vars)
+	return vars
+}
+
+func parseNode(n parse.Node) []string {
+	switch n := n.(type) {
+	case *parse.ActionNode:
+		return parseNode(n.Pipe)
+	case *parse.IfNode:
+		names := parseNode(n.Pipe)
+		names = append(names, parseNode(n.List)...)
+		if n.ElseList != nil {
+			names = append(names, parseNode(n.ElseList)...)
+		}
+		return names
+	case *parse.RangeNode:
+		names := parseNode(n.Pipe)
+		names = append(names, parseNode(n.List)...)
+		if n.ElseList != nil {
+			names = append(names, parseNode(n.ElseList)...)
+		}
+		return names
+	case *parse.WithNode:
+		names := parseNode(n.Pipe)
+		names = append(names, parseNode(n.List)...)
+		if n.ElseList != nil {
+			names = append(names, parseNode(n.ElseList)...)
+		}
+		return names
+	case *parse.PipeNode:
+		var names []string
+		for _, c := range n.Cmds {
+			for _, a := range c.Args {
+				names = append(names, parseNode(a)...)
+			}
+		}
+		return names
+	case *parse.ListNode:
+		var names []string
+		for _, n := range n.Nodes {
+			names = append(names, parseNode(n)...)
+		}
+
+		return names
+	case *parse.FieldNode:
+		return n.Ident
+	}
+
+	return nil
+}
--- a/template/template_test.go
+++ b/template/template_test.go
@ -0,0 +1,89 @@
+package template
+
+import (
+	"bufio"
+	"bytes"
+	"encoding/json"
+	"io"
+	"os"
+	"path/filepath"
+	"slices"
+	"testing"
+	"text/template"
+
+	"github.com/ollama/ollama/llm"
+)
+
+func TestNamed(t *testing.T) {
+	f, err := os.Open(filepath.Join("testdata", "templates.jsonl"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer f.Close()
+
+	scanner := bufio.NewScanner(f)
+	for scanner.Scan() {
+		var ss map[string]string
+		if err := json.Unmarshal(scanner.Bytes(), &ss); err != nil {
+			t.Fatal(err)
+		}
+
+		for k, v := range ss {
+			t.Run(k, func(t *testing.T) {
+				kv := llm.KV{"tokenizer.chat_template": v}
+				s := kv.ChatTemplate()
+				r, err := Named(s)
+				if err != nil {
+					t.Fatal(err)
+				}
+
+				if r.Name != k {
+					t.Errorf("expected %q, got %q", k, r.Name)
+				}
+
+				var b bytes.Buffer
+				if _, err := io.Copy(&b, r.Reader()); err != nil {
+					t.Fatal(err)
+				}
+
+				tmpl, err := template.New(s).Parse(b.String())
+				if err != nil {
+					t.Fatal(err)
+				}
+
+				if tmpl.Tree.Root.String() == "" {
+					t.Errorf("empty %s template", k)
+				}
+			})
+		}
+	}
+}
+
+func TestParse(t *testing.T) {
+	cases := []struct {
+		template string
+		vars     []string
+	}{
+		{"{{ .Prompt }}", []string{"prompt"}},
+		{"{{ .System }} {{ .Prompt }}", []string{"prompt", "system"}},
+		{"{{ .System }} {{ .Prompt }} {{ .Response }}", []string{"prompt", "response", "system"}},
+		{"{{ with .Tools }}{{ . }}{{ end }} {{ .System }} {{ .Prompt }}", []string{"prompt", "system", "tools"}},
+		{"{{ range .Messages }}{{ .Role }} {{ .Content }}{{ end }}", []string{"content", "messages", "role"}},
+		{"{{ range .Messages }}{{ if eq .Role \"system\" }}SYSTEM: {{ .Content }}{{ else if eq .Role \"user\" }}USER: {{ .Content }}{{ else if eq .Role \"assistant\" }}ASSISTANT: {{ .Content }}{{ end }}{{ end }}", []string{"content", "messages", "role"}},
+		{"{{ .Prompt }} {{ .Suffix }}", []string{"prompt", "suffix"}},
+	}
+
+	for _, tt := range cases {
+		t.Run("", func(t *testing.T) {
+			tmpl, err := Parse(tt.template)
+			if err != nil {
+				t.Fatal(err)
+			}
+
+			vars := tmpl.Vars()
+			if !slices.Equal(tt.vars, vars) {
+				t.Errorf("expected %v, got %v", tt.vars, vars)
+			}
+		})
+	}
+}
--- a/templates/testdata/templates.jsonl
+++ b/templates/testdata/templates.jsonl
--- a/templates/vicuna.gotmpl
+++ b/templates/vicuna.gotmpl
--- a/templates/zephyr.gotmpl
+++ b/templates/zephyr.gotmpl
--- a/templates/template.go
+++ b/templates/template.go
@ -1,70 +0,0 @@
-package templates
-
-import (
-	"bytes"
-	"embed"
-	"encoding/json"
-	"errors"
-	"io"
-	"math"
-	"sync"
-
-	"github.com/agnivade/levenshtein"
-)
-
-//go:embed index.json
-var indexBytes []byte
-
-//go:embed *.gotmpl
-var templatesFS embed.FS
-
-var templatesOnce = sync.OnceValues(func() ([]*Template, error) {
-	var templates []*Template
-	if err := json.Unmarshal(indexBytes, &templates); err != nil {
-		return nil, err
-	}
-
-	for _, t := range templates {
-		bts, err := templatesFS.ReadFile(t.Name + ".gotmpl")
-		if err != nil {
-			return nil, err
-		}
-
-		// normalize line endings
-		t.Bytes = bytes.ReplaceAll(bts, []byte("\r\n"), []byte("\n"))
-	}
-
-	return templates, nil
-})
-
-type Template struct {
-	Name     string `json:"name"`
-	Template string `json:"template"`
-	Bytes []byte
-}
-
-func (t Template) Reader() io.Reader {
-	return bytes.NewReader(t.Bytes)
-}
-
-func NamedTemplate(s string) (*Template, error) {
-	templates, err := templatesOnce()
-	if err != nil {
-		return nil, err
-	}
-
-	var template *Template
-	score := math.MaxInt
-	for _, t := range templates {
-		if s := levenshtein.ComputeDistance(s, t.Template); s < score {
-			score = s
-			template = t
-		}
-	}
-
-	if score < 100 {
-		return template, nil
-	}
-
-	return nil, errors.New("no matching template found")
-}
--- a/templates/template_test.go
+++ b/templates/template_test.go
@ -1,59 +0,0 @@
-package templates
-
-import (
-	"bufio"
-	"bytes"
-	"encoding/json"
-	"io"
-	"os"
-	"path/filepath"
-	"testing"
-	"text/template"
-
-	"github.com/ollama/ollama/llm"
-)
-
-func TestKVChatTemplate(t *testing.T) {
-	f, err := os.Open(filepath.Join("testdata", "templates.jsonl"))
-	if err != nil {
-		t.Fatal(err)
-	}
-	defer f.Close()
-
-	scanner := bufio.NewScanner(f)
-	for scanner.Scan() {
-		var ss map[string]string
-		if err := json.Unmarshal(scanner.Bytes(), &ss); err != nil {
-			t.Fatal(err)
-		}
-
-		for k, v := range ss {
-			t.Run(k, func(t *testing.T) {
-				kv := llm.KV{"tokenizer.chat_template": v}
-				s := kv.ChatTemplate()
-				r, err := NamedTemplate(s)
-				if err != nil {
-					t.Fatal(err)
-				}
-
-				if r.Name != k {
-					t.Errorf("expected %q, got %q", k, r.Name)
-				}
-
-				var b bytes.Buffer
-				if _, err := io.Copy(&b, r.Reader()); err != nil {
-					t.Fatal(err)
-				}
-
-				tmpl, err := template.New(s).Parse(b.String())
-				if err != nil {
-					t.Fatal(err)
-				}
-
-				if tmpl.Tree.Root.String() == "" {
-					t.Errorf("empty %s template", k)
-				}
-			})
-		}
-	}
-}
--- a/types/model/name.go
+++ b/types/model/name.go
@ -91,7 +91,6 @@ type Name struct {
 	Namespace string
 	Model     string
 	Tag       string
-	RawDigest string
 }

 // ParseName parses and assembles a Name from a name string. The
@ -143,11 +142,6 @@ func ParseNameBare(s string) Name {
 	var n Name
 	var promised bool

-	s, n.RawDigest, promised = cutLast(s, "@")
-	if promised && n.RawDigest == "" {
-		n.RawDigest = MissingPart
-	}
-
 	// "/" is an illegal tag character, so we can use it to split the host
 	if strings.LastIndex(s, ":") > strings.LastIndex(s, "/") {
 		s, n.Tag, _ = cutPromised(s, ":")
@ -222,10 +216,6 @@ func (n Name) String() string {
 		b.WriteByte(':')
 		b.WriteString(n.Tag)
 	}
-	if n.RawDigest != "" {
-		b.WriteByte('@')
-		b.WriteString(n.RawDigest)
-	}
 	return b.String()
 }

@ -250,16 +240,18 @@ func (n Name) DisplayShortest() string {
 	return sb.String()
 }

-func IsValidNamespace(namespace string) bool {
-	return isValidPart(kindNamespace, namespace)
+// IsValidNamespace reports whether the provided string is a valid
+// namespace.
+func IsValidNamespace(s string) bool {
+	return isValidPart(kindNamespace, s)
 }

 // IsValid reports whether all parts of the name are present and valid. The
 // digest is a special case, and is checked for validity only if present.
+//
+// Note: The digest check has been removed as is planned to be added back in
+// at a later time.
 func (n Name) IsValid() bool {
-	if n.RawDigest != "" && !isValidPart(kindDigest, n.RawDigest) {
-		return false
-	}
 	return n.IsFullyQualified()
 }

--- a/types/model/name_test.go
+++ b/types/model/name_test.go
@ -122,21 +122,6 @@ func TestParseNameParts(t *testing.T) {
 			},
 			wantFilepath: filepath.Join(part350, part80, part80, part80),
 		},
-		{
-			in: "@digest",
-			want: Name{
-				RawDigest: "digest",
-			},
-			wantValidDigest: false,
-		},
-		{
-			in: "model@sha256:123",
-			want: Name{
-				Model:     "model",
-				RawDigest: "sha256:123",
-			},
-			wantValidDigest: true,
-		},
 	}

 	for _, tt := range cases {
@ -160,22 +145,18 @@ var testCases = map[string]bool{ // name -> valid
 	"_why/_the/_lucky:_stiff": true,

 	// minimal
-	"h/n/m:t@d": true,
+	"h/n/m:t": true,

 	"host/namespace/model:tag": true,
 	"host/namespace/model":     false,
 	"namespace/model":          false,
 	"model":                    false,
-	"@sha256-1000000000000000000000000000000000000000000000000000000000000000":      false,
-	"model@sha256-1000000000000000000000000000000000000000000000000000000000000000": false,
-	"model@sha256:1000000000000000000000000000000000000000000000000000000000000000": false,

 	// long (but valid)
 	part80 + "/" + part80 + "/" + part80 + ":" + part80:  true,
 	part350 + "/" + part80 + "/" + part80 + ":" + part80: true,

-	"h/nn/mm:t@sha256-1000000000000000000000000000000000000000000000000000000000000000": true, // bare minimum part sizes
-	"h/nn/mm:t@sha256:1000000000000000000000000000000000000000000000000000000000000000": true, // bare minimum part sizes
+	"h/nn/mm:t": true, // bare minimum part sizes

 	// unqualified
 	"m":     false,
@ -196,11 +177,10 @@ var testCases = map[string]bool{ // name -> valid
 	"@":      false,

 	// not starting with alphanum
-	"-hh/nn/mm:tt@dd": false,
-	"hh/-nn/mm:tt@dd": false,
-	"hh/nn/-mm:tt@dd": false,
-	"hh/nn/mm:-tt@dd": false,
-	"hh/nn/mm:tt@-dd": false,
+	"-hh/nn/mm:tt": false,
+	"hh/-nn/mm:tt": false,
+	"hh/nn/-mm:tt": false,
+	"hh/nn/mm:-tt": false,

 	// hosts
 	"host:https/namespace/model:tag": true,
@ -334,7 +314,7 @@ func FuzzName(f *testing.F) {
 	f.Fuzz(func(t *testing.T, s string) {
 		n := ParseNameBare(s)
 		if n.IsValid() {
-			parts := [...]string{n.Host, n.Namespace, n.Model, n.Tag, n.RawDigest}
+			parts := [...]string{n.Host, n.Namespace, n.Model, n.Tag}
 			for _, part := range parts {
 				if part == ".." {
 					t.Errorf("unexpected .. as valid part")