Merge branch 'main' into royh-batchembed

.github/workflows/release.yaml

@@ -304,6 +304,11 @@ jobs:
           write-host "Installing plugin"
           & "${env:RUNNER_TEMP}\plugin\*\kmscng.msi" /quiet
           write-host "plugin installed"
+      - name: remove unwanted mingw dll.a files
+        run: |
+          Get-ChildItem -Path "C:\mingw64" -Recurse -Filter "libpthread.dll.a" -File | Remove-Item -Force
+          Get-ChildItem -Path "C:\mingw64" -Recurse -Filter "libwinpthread.dll.a" -File | Remove-Item -Force
+          Get-ChildItem -Path "C:\mingw64" -Recurse -Filter "libstdc++.dll.a" -File | Remove-Item -Force
       - uses: actions/setup-go@v5
         with:
           go-version-file: go.mod

.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

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"`
+	UseNUMA   bool  `json:"numa,omitempty"`
-	NumCtx    int      `json:"num_ctx,omitempty"`
+	NumCtx    int   `json:"num_ctx,omitempty"`
-	NumBatch  int      `json:"num_batch,omitempty"`
+	NumBatch  int   `json:"num_batch,omitempty"`
-	NumGPU    int      `json:"num_gpu,omitempty"`
+	NumGPU    int   `json:"num_gpu,omitempty"`
-	MainGPU   int      `json:"main_gpu,omitempty"`
+	MainGPU   int   `json:"main_gpu,omitempty"`
-	LowVRAM   bool     `json:"low_vram,omitempty"`
+	LowVRAM   bool  `json:"low_vram,omitempty"`
-	F16KV     bool     `json:"f16_kv,omitempty"`
+	F16KV     bool  `json:"f16_kv,omitempty"`
-	LogitsAll bool     `json:"logits_all,omitempty"`
+	LogitsAll bool  `json:"logits_all,omitempty"`
-	VocabOnly bool     `json:"vocab_only,omitempty"`
+	VocabOnly bool  `json:"vocab_only,omitempty"`
-	UseMMap   TriState `json:"use_mmap,omitempty"`
+	UseMMap   *bool `json:"use_mmap,omitempty"`
-	UseMLock  bool     `json:"use_mlock,omitempty"`
+	UseMLock  bool  `json:"use_mlock,omitempty"`
-	NumThread int      `json:"num_thread,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)
 }
 
 // EmbedRequest is the request passed to [Client.Embed].
@@ -468,19 +437,6 @@ 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) {
@@ -527,6 +483,17 @@ 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())
 			}
@@ -569,7 +536,7 @@ func DefaultOptions() Options {
 			LowVRAM:   false,
 			F16KV:     true,
 			UseMLock:  false,
-			UseMMap:   TriStateUndefined,
+			UseMMap:   nil,
 			UseNUMA:   false,
 		},
 	}
@@ -639,19 +606,6 @@ 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)
@@ -679,6 +633,17 @@ 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)
 				}

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))
 			}
 		})
 	}

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 .
 ```
 

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.
+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)
-- Is the uvm driver not loaded? `sudo nvidia-modprobe -u`
+
+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.

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
 

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
@@ -132,6 +134,7 @@ func init() {
 	NumParallel = 0 // Autoselect
 	MaxRunners = 0  // Autoselect
 	MaxQueuedRequests = 512
+	KeepAlive = 5 * time.Minute
 
 	LoadConfig()
 }
@@ -266,7 +269,10 @@ func LoadConfig() {
 		}
 	}
 
-	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
+		}
+	}
+}

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) {

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

gpu/gpu_darwin.go

@@ -56,7 +56,7 @@ func GetCPUInfo() GpuInfoList {
 func GetCPUMem() (memInfo, error) {
 	return memInfo{
 		TotalMemory: uint64(C.getPhysicalMemory()),
-		FreeMemory:  0,
+		FreeMemory:  uint64(C.getFreeMemory()),
 	}, nil
 }
 

gpu/gpu_info_darwin.h

@@ -2,3 +2,4 @@
 #include <stdint.h>
 uint64_t getRecommendedMaxVRAM();
 uint64_t getPhysicalMemory();
+uint64_t getFreeMemory();

gpu/gpu_info_darwin.m

@@ -1,4 +1,5 @@
-// go:build darwin
+#import <Foundation/Foundation.h>
+#import <mach/mach.h>
 #include "gpu_info_darwin.h"
 
 uint64_t getRecommendedMaxVRAM() {
@@ -8,6 +9,27 @@ uint64_t getRecommendedMaxVRAM() {
   return result;
 }
 
+// getPhysicalMemory returns the total physical memory in bytes
 uint64_t getPhysicalMemory() {
-  return [[NSProcessInfo processInfo] physicalMemory];
+  return [NSProcessInfo processInfo].physicalMemory;
+}
+
+// getFreeMemory returns the total free memory in bytes, including inactive
+// memory that can be reclaimed by the system.
+uint64_t getFreeMemory() {
+  mach_port_t host_port = mach_host_self();
+  mach_msg_type_number_t host_size = sizeof(vm_statistics64_data_t) / sizeof(integer_t);
+  vm_size_t pagesize;
+  vm_statistics64_data_t vm_stat;
+
+  host_page_size(host_port, &pagesize);
+  if (host_statistics64(host_port, HOST_VM_INFO64, (host_info64_t)&vm_stat, &host_size) != KERN_SUCCESS) {
+    return 0;
+  }
+
+  uint64_t free_memory = (uint64_t)vm_stat.free_count * pagesize;
+  free_memory += (uint64_t)vm_stat.speculative_count * pagesize;
+  free_memory += (uint64_t)vm_stat.inactive_count * pagesize;
+
+  return free_memory;
 }

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) {
+    snprintf(buf, buflen, "cuda driver library init failure: %d", ret);
-      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);
     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;

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);

llm/ext_server/CMakeLists.txt

@@ -1,4 +1,3 @@
-
 set(TARGET ollama_llama_server)
 option(LLAMA_SERVER_VERBOSE "Build verbose logging option for Server" ON)
 include_directories(${CMAKE_CURRENT_SOURCE_DIR})
@@ -7,7 +6,7 @@ install(TARGETS ${TARGET} RUNTIME)
 target_compile_definitions(${TARGET} PRIVATE
     SERVER_VERBOSE=$<BOOL:${LLAMA_SERVER_VERBOSE}>
 )
-target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE ggml llama common llava ${CMAKE_THREAD_LIBS_INIT})
 if (WIN32)
     TARGET_LINK_LIBRARIES(${TARGET} PRIVATE ws2_32)
 endif()

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_DEBUG("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
@@ -1650,22 +1688,8 @@ struct llama_server_context
                     }
                     slot.params.n_keep = std::min(slot.n_ctx - 4, slot.params.n_keep);
 
-                    char buf[256];
-                    llama_model_meta_val_str(model, "general.architecture", buf, 256);
-                    bool gemma2 = strcmp(buf, "gemma2") == 0;
-
-                    int32_t truncate_at = slot.n_ctx;
-
-                    // truncate at 2/3 of the context length for gemma2 models
-                    // as they do not support context shifts (from the sliding window implementation).
-                    // this way, prompts that almost fit the context length can still generate a full
-                    // response without a sudden stop from hitting the context limit
-                    if (gemma2) {
-                        truncate_at = 2 * slot.n_ctx / 3;
-                    }
-
                     // if input prompt is too big, truncate it, if group attention self-extend is disabled
-                    if (slot.ga_n == 1 && slot.n_prompt_tokens >= truncate_at)
+                    if (slot.ga_n == 1 && slot.n_prompt_tokens >= slot.n_ctx)
                     {
                         const int n_left = slot.n_ctx - slot.params.n_keep;
                         const int n_shift = n_left / 2;
@@ -1693,19 +1717,6 @@ struct llama_server_context
                         GGML_ASSERT(slot.n_prompt_tokens < slot.n_ctx);
                     }
 
-                    // Models with sliding window attention do not work with context shifts, so
-                    // limit their prediction to the context length
-                    if (gemma2) {
-                        int32_t limit = slot.n_ctx - slot.n_prompt_tokens;
-                        slot.n_predict = limit;
-                        slot.params.n_predict = limit;
-                        LOG_INFO("model does not support sliding window, limiting generation", {
-                            {"n_ctx", slot.n_ctx},
-                            {"n_prompt_tokens", slot.n_prompt_tokens},
-                            {"n_predict", slot.n_predict}
-                        });
-                    }
-
                     if (!slot.params.cache_prompt)
                     {
                         llama_sampling_reset(slot.ctx_sampling);
@@ -1732,7 +1743,7 @@ 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)
                         {

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

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
@@ -77,29 +77,29 @@ if [ -z "${OLLAMA_SKIP_CPU_GENERATE}" ]; then
     if [ -n "${OLLAMA_CUSTOM_CPU_DEFS}" ]; then
         init_vars
         echo "OLLAMA_CUSTOM_CPU_DEFS=\"${OLLAMA_CUSTOM_CPU_DEFS}\""
-        CMAKE_DEFS="${OLLAMA_CUSTOM_CPU_DEFS} -DCMAKE_POSITION_INDEPENDENT_CODE=on ${CMAKE_DEFS}"
+        CMAKE_DEFS="${OLLAMA_CUSTOM_CPU_DEFS} -DBUILD_SHARED_LIBS=off -DCMAKE_POSITION_INDEPENDENT_CODE=on ${CMAKE_DEFS}"
         BUILD_DIR="../build/linux/${ARCH}/cpu"
         echo "Building custom CPU"
         build
         compress
     else
         # Darwin Rosetta x86 emulation does NOT support AVX, AVX2, AVX512
-        # -DLLAMA_AVX -- 2011 Intel Sandy Bridge & AMD Bulldozer
+        # -DGGML_AVX -- 2011 Intel Sandy Bridge & AMD Bulldozer
-        # -DLLAMA_F16C -- 2012 Intel Ivy Bridge & AMD 2011 Bulldozer (No significant improvement over just AVX)
+        # -DGGML_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
+        # -DGGML_AVX2 -- 2013 Intel Haswell & 2015 AMD Excavator / 2017 AMD Zen
-        # -DLLAMA_FMA (FMA3) -- 2013 Intel Haswell & 2012 AMD Piledriver
+        # -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)
+        # -DGGML_AVX512 -- 2017 Intel Skylake and High End DeskTop (HEDT)
-        # -DLLAMA_AVX512_VBMI -- 2018 Intel Cannon Lake
+        # -DGGML_AVX512_VBMI -- 2018 Intel Cannon Lake
-        # -DLLAMA_AVX512_VNNI -- 2021 Intel Alder 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="-DBUILD_SHARED_LIBS=off -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}\""

llm/generate/gen_windows.ps1

@@ -39,8 +39,8 @@ function init_vars {
     }
     $script:cmakeDefs = @(
         "-DBUILD_SHARED_LIBS=on",
-        "-DLLAMA_NATIVE=off",
+        "-DGGML_NATIVE=off",
-        "-DLLAMA_OPENMP=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
+# -DGGML_AVX -- 2011 Intel Sandy Bridge & AMD Bulldozer
-# -DLLAMA_AVX2 -- 2013 Intel Haswell & 2015 AMD Excavator / 2017 AMD Zen
+# -DGGML_AVX2 -- 2013 Intel Haswell & 2015 AMD Excavator / 2017 AMD Zen
-# -DLLAMA_FMA (FMA3) -- 2013 Intel Haswell & 2012 AMD Piledriver
+# -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",
+            "-DGGML_NATIVE=off",
-            "-DLLAMA_AVX=off",
+            "-DGGML_AVX=off",
-            "-DLLAMA_AVX2=off",
+            "-DGGML_AVX2=off",
-            "-DLLAMA_AVX512=off",
+            "-DGGML_AVX512=off",
-            "-DLLAMA_F16C=off",
+            "-DGGML_F16C=off",
-            "-DLLAMA_FMA=off",
+            "-DGGML_FMA=off",
-            "-DLLAMA_OPENMP=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"
         $script:cmakeDefs += @(
             "-A", "x64",
-            "-DLLAMA_CUDA=ON",
+            "-DGGML_CUDA=ON",
-            "-DLLAMA_AVX=on",
+            "-DGGML_AVX=on",
-            "-DLLAMA_AVX2=off",
+            "-DGGML_AVX2=off",
             "-DCUDAToolkit_INCLUDE_DIR=$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"
     $script:cmakeDefs += @(
       "-G", "MinGW Makefiles",
-      "-DLLAMA_SYCL=ON",
+      "-DGGML_SYCL=ON",
       "-DCMAKE_C_COMPILER=icx",
       "-DCMAKE_CXX_COMPILER=icx",
       "-DCMAKE_BUILD_TYPE=Release"
@@ -365,10 +365,10 @@ function build_rocm() {
             "-G", "Ninja", 
             "-DCMAKE_C_COMPILER=clang.exe",
             "-DCMAKE_CXX_COMPILER=clang++.exe",
-            "-DLLAMA_HIPBLAS=on",
+            "-DGGML_HIPBLAS=on",
             "-DHIP_PLATFORM=amd",
-            "-DLLAMA_AVX=on",
+            "-DGGML_AVX=on",
-            "-DLLAMA_AVX2=off",
+            "-DGGML_AVX2=off",
             "-DCMAKE_POSITION_INDEPENDENT_CODE=on",
             "-DAMDGPU_TARGETS=$(amdGPUs)",
             "-DGPU_TARGETS=$(amdGPUs)"

llm/llama.cpp

@@ -1 +1 @@
-Subproject commit 7c26775adb579e92b59c82e8084c07a1d0f75e9c
+Subproject commit a8db2a9ce64cd4417f6a312ab61858f17f0f8584

llm/llm.go

@@ -1,12 +1,13 @@
 package llm
 
-// #cgo CFLAGS: -Illama.cpp
+// #cgo CFLAGS: -Illama.cpp -Illama.cpp/include -Illama.cpp/ggml/include
-// #cgo darwin,arm64 LDFLAGS: ${SRCDIR}/build/darwin/arm64_static/libllama.a -lstdc++
+// #cgo LDFLAGS: -lllama -lggml -lstdc++ -lpthread
-// #cgo darwin,amd64 LDFLAGS: ${SRCDIR}/build/darwin/x86_64_static/libllama.a -lstdc++
+// #cgo darwin,arm64 LDFLAGS: -L${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 windows,amd64 LDFLAGS: ${SRCDIR}/build/windows/amd64_static/libllama.a -static -lstdc++
+// #cgo darwin,amd64 LDFLAGS: -L${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,arm64 LDFLAGS: ${SRCDIR}/build/windows/arm64_static/libllama.a -static -lstdc++
+// #cgo windows,amd64 LDFLAGS: -L${SRCDIR}/build/windows/amd64_static -L${SRCDIR}/build/windows/amd64_static/src -L${SRCDIR}/build/windows/amd64_static/ggml/src
-// #cgo linux,amd64 LDFLAGS: ${SRCDIR}/build/linux/x86_64_static/libllama.a -lstdc++
+// #cgo windows,arm64 LDFLAGS: -L${SRCDIR}/build/windows/arm64_static -L${SRCDIR}/build/windows/arm64_static/src -L${SRCDIR}/build/windows/arm64_static/ggml/src
-// #cgo linux,arm64 LDFLAGS: ${SRCDIR}/build/linux/arm64_static/libllama.a -lstdc++
+// #cgo linux,amd64 LDFLAGS: -L${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"

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
+     // embedding
-     int32_t port           = 8080;         // server listens on this network port
+     bool embedding         = false; // get only sentence embedding
-     int32_t timeout_read   = 600;          // http read timeout in seconds
+     int32_t embd_normalize = 2;     // normalisation for embendings (-1=none, 0=max absolute int16, 1=taxicab, 2=euclidean, >2=p-norm)

llm/patches/03-load_exception.diff

@@ -1,17 +1,8 @@
-From 544a2d2e646d39e878d87dfbb3398a356bc560ab Mon Sep 17 00:00:00 2001
+diff --git a/src/llama.cpp b/src/llama.cpp
-From: Michael Yang <mxyng@pm.me>
+index 73f52435..58a00fb1 100644
-Date: Thu, 23 May 2024 11:18:45 -0700
+--- a/src/llama.cpp
-Subject: [PATCH] throw exception on load errors
++++ b/src/llama.cpp
-
+@@ -7241,7 +7241,7 @@ static int llama_model_load(const std::string & fname, llama_model & model, llam
----
- 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
          }
      } 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
-

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
+--- a/ggml/src/ggml-metal.m
-+++ b/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;

llm/patches/05-default-pretokenizer.diff

@@ -1,11 +1,11 @@
-diff --git a/llama.cpp b/llama.cpp
+diff --git a/src/llama.cpp b/src/llama.cpp
-index 61948751..4b72a293 100644
+index 2b9ace28..172640e2 100644
---- a/llama.cpp
+--- a/src/llama.cpp
-+++ b/llama.cpp
++++ b/src/llama.cpp
-@@ -4824,16 +4824,7 @@ static void llm_load_vocab(
+@@ -5357,16 +5357,7 @@ static void llm_load_vocab(
- 
-         // for now, only BPE models have pre-tokenizers
          if (vocab.type == LLAMA_VOCAB_TYPE_BPE) {
+             vocab.tokenizer_add_space_prefix = false;
+             vocab.tokenizer_clean_spaces = true;
 -            if (tokenizer_pre.empty()) {
 -                LLAMA_LOG_WARN("%s: missing pre-tokenizer type, using: 'default'\n", __func__);
 -                LLAMA_LOG_WARN("%s:                                             \n", __func__);
@@ -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(
+@@ -5439,7 +5430,8 @@ static void llm_load_vocab(
-                 tokenizer_pre == "poro-chat") {
+                 tokenizer_pre == "jais") {
-                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_PORO;
+                 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;

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
+--- a/src/llama.cpp
-+++ b/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);

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);

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
-

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);

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;
+             }

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 !(runtime.GOOS == "darwin" && runtime.GOARCH == "arm64") {
-	if info.Library != "cpu" {
+		// Load up the best CPU variant if not primary requested
-		variant := gpu.GetCPUCapability()
+		if info.Library != "cpu" {
-		// If no variant, then we fall back to default
+			variant := gpu.GetCPUCapability()
-		// If we have a variant, try that if we find an exact match
+			// If no variant, then we fall back to default
-		// Attempting to run the wrong CPU instructions will panic the
+			// If we have a variant, try that if we find an exact match
-		// process
+			// Attempting to run the wrong CPU instructions will panic the
-		if variant != gpu.CPUCapabilityNone {
+			// process
-			for cmp := range availableServers {
+			if variant != gpu.CPUCapabilityNone {
-				if cmp == "cpu_"+variant.String() {
+				for cmp := range availableServers {
-					servers = append(servers, cmp)
+					if cmp == "cpu_"+variant.String() {
-					break
+						servers = append(servers, cmp)
+						break
+					}
 				}
+			} else {
+				servers = append(servers, "cpu")
 			}
-		} else {
-			servers = append(servers, "cpu")
 		}
-	}
 
-	if len(servers) == 0 {
+		if len(servers) == 0 {
-		servers = []string{"cpu"}
+			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() {

llm/server.go

@@ -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
 		}
 	}
 
@@ -219,10 +233,10 @@ 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
 	// For CPU loads we want the memory to be allocated, not FS cache
-	if (runtime.GOOS == "windows" && gpus[0].Library == "cuda" && opts.UseMMap == api.TriStateUndefined) ||
+	if (runtime.GOOS == "windows" && gpus[0].Library == "cuda" && opts.UseMMap == nil) ||
-		(runtime.GOOS == "linux" && systemFreeMemory < estimate.TotalSize && opts.UseMMap == api.TriStateUndefined) ||
+		(runtime.GOOS == "linux" && systemFreeMemory < estimate.TotalSize && opts.UseMMap == nil) ||
-		(gpus[0].Library == "cpu" && opts.UseMMap == api.TriStateUndefined) ||
+		(gpus[0].Library == "cpu" && opts.UseMMap == nil) ||
-		opts.UseMMap == api.TriStateFalse {
+		(opts.UseMMap != nil && !*opts.UseMMap) {
 		params = append(params, "--no-mmap")
 	}
 
@@ -685,10 +699,9 @@ func (s *llmServer) Completion(ctx context.Context, req CompletionRequest, fn fu
 	}
 	defer s.sem.Release(1)
 
-	// only allow maximum 10 "context shifts" to avoid infinite generation
+	// put an upper limit on num_predict to avoid the model running on forever
 	if req.Options.NumPredict < 0 || req.Options.NumPredict > 10*s.options.NumCtx {
 		req.Options.NumPredict = 10 * s.options.NumCtx
-		slog.Debug("setting token limit to 10x num_ctx", "num_ctx", s.options.NumCtx, "num_predict", req.Options.NumPredict)
 	}
 
 	request := map[string]any{

llm/status.go

@@ -25,7 +25,7 @@ var errorPrefixes = []string{
 	"CUDA error",
 	"cudaMalloc failed",
 	"\"ERR\"",
-	"architecture",
+	"error loading model",
 }
 
 func (w *StatusWriter) Write(b []byte) (int, error) {

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()
+	dir := envconfig.ModelsDir
-	if err != nil {
-		return "", err
-	}
 
 	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()
+	dir := envconfig.ModelsDir
-	if err != nil {
-		return "", err
-	}
 
 	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()
+	dir := envconfig.ModelsDir
-	if err != nil {
-		return "", err
-	}
 
 	// only accept actual sha256 digests
 	pattern := "^sha256[:-][0-9a-fA-F]{64}$"

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"
@@ -56,8 +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 {
@@ -133,14 +129,7 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 		return
 	}
 
-	var sessionDuration time.Duration
+	rCh, eCh := s.sched.GetRunner(c.Request.Context(), model, opts, req.KeepAlive)
-	if req.KeepAlive == nil {
-		sessionDuration = getDefaultSessionDuration()
-	} else {
-		sessionDuration = req.KeepAlive.Duration
-	}
-
-	rCh, eCh := s.sched.GetRunner(c.Request.Context(), model, opts, sessionDuration)
 	var runner *runnerRef
 	select {
 	case runner = <-rCh:
@@ -536,14 +525,7 @@ func (s *Server) EmbeddingsHandler(c *gin.Context) {
 		return
 	}
 
-	var sessionDuration time.Duration
+	rCh, eCh := s.sched.GetRunner(c.Request.Context(), model, opts, req.KeepAlive)
-	if req.KeepAlive == nil {
-		sessionDuration = getDefaultSessionDuration()
-	} else {
-		sessionDuration = req.KeepAlive.Duration
-	}
-
-	rCh, eCh := s.sched.GetRunner(c.Request.Context(), model, opts, sessionDuration)
 	var runner *runnerRef
 	select {
 	case runner = <-rCh:
@@ -1475,14 +1457,7 @@ func (s *Server) ChatHandler(c *gin.Context) {
 		return
 	}
 
-	var sessionDuration time.Duration
+	rCh, eCh := s.sched.GetRunner(c.Request.Context(), model, opts, req.KeepAlive)
-	if req.KeepAlive == nil {
-		sessionDuration = getDefaultSessionDuration()
-	} else {
-		sessionDuration = req.KeepAlive.Duration
-	}
-
-	rCh, eCh := s.sched.GetRunner(c.Request.Context(), model, opts, sessionDuration)
 	var runner *runnerRef
 	select {
 	case runner = <-rCh:

server/sched.go

@@ -24,7 +24,7 @@ type LlmRequest struct {
 	model           *Model
 	opts            api.Options
 	origNumCtx      int // Track the initial ctx request
-	sessionDuration time.Duration
+	sessionDuration *api.Duration
 	successCh       chan *runnerRef
 	errCh           chan error
 	schedAttempts   uint
@@ -75,7 +75,7 @@ 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) {
+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
 	}
@@ -139,6 +139,11 @@ func (s *Scheduler) processPending(ctx context.Context) {
 			}
 
 			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]
@@ -192,6 +197,38 @@ func (s *Scheduler) processPending(ctx context.Context) {
 						break
 					}
 
+					estimate := llm.EstimateGPULayers(gpus, ggml, pending.model.ProjectorPaths, pending.opts)
+					maxSize := systemMem.FreeMemory
+
+					// Add available GPU memory to the total pool
+					// macOS hardware has unified memory so don't double count
+					if runtime.GOOS != "darwin" {
+						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
+							}
+						}
+					}
+
+					// Block attempting to load a model larger than system memory + GPU memory
+					if estimate.TotalSize > maxSize {
+						slog.Warn("model request too large for system", "requested", format.HumanBytes2(estimate.TotalSize), "system", format.HumanBytes2(maxSize))
+
+						// Linux will crash if over-allocating memory - return an error to the user.
+						// TODO (jmorganca): add reasonable upper limits for darwin and windows as well
+						if runtime.GOOS == "linux" {
+							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
@@ -389,7 +426,9 @@ 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()
@@ -402,6 +441,10 @@ func (s *Scheduler) load(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList,
 	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
@@ -419,7 +462,7 @@ 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(),

server/sched_test.go

@@ -44,7 +44,7 @@ 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, numParallel int) (llm.LlamaServer, error) {
@@ -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)
@@ -199,6 +199,8 @@ 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.Fatal("timeout")
 	}
@@ -212,6 +214,8 @@ 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.Fatal("timeout")
 	}
@@ -230,6 +234,8 @@ 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.Fatal("timeout")
 	}
@@ -246,6 +252,8 @@ 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.Fatal("timeout")
 	}
@@ -262,6 +270,8 @@ 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.Fatal("timeout")
 	}
@@ -278,6 +288,8 @@ 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.Fatal("timeout")
 	}
@@ -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 {
@@ -402,7 +414,7 @@ func TestPrematureExpired(t *testing.T) {
 	case <-ctx.Done():
 		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)
@@ -423,7 +435,7 @@ 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{}}
@@ -614,7 +626,7 @@ func TestAlreadyCanceled(t *testing.T) {
 	dctx, done2 := context.WithCancel(ctx)
 	done2()
 	scenario1a := newScenario(t, dctx, "ollama-model-1", 10)
-	scenario1a.req.sessionDuration = 0
+	scenario1a.req.sessionDuration = &api.Duration{Duration: 0}
 	s := InitScheduler(ctx)
 	slog.Info("scenario1a")
 	s.pendingReqCh <- scenario1a.req

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

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": true, // bare minimum part sizes
-	"h/nn/mm:t@sha256:1000000000000000000000000000000000000000000000000000000000000000": 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": false,
-	"hh/-nn/mm:tt@dd": false,
+	"hh/-nn/mm:tt": false,
-	"hh/nn/-mm:tt@dd": false,
+	"hh/nn/-mm:tt": false,
-	"hh/nn/mm:-tt@dd": false,
+	"hh/nn/mm:-tt": false,
-	"hh/nn/mm:tt@-dd": 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")