Merge pull request #3933 from dhiltgen/ci_fixes

Move cuda/rocm dependency gathering into generate script

.github/workflows/release.yaml

@@ -311,29 +311,18 @@ jobs:
       - uses: actions/download-artifact@v4
         with:
           name: generate-windows-cpu
-          path: |
-            llm/build
-            dist/windows-amd64
       - uses: actions/download-artifact@v4
         with:
           name: generate-windows-cuda
-          path: |
-            llm/build
-            dist/windows-amd64
       - uses: actions/download-artifact@v4
         with:
           name: windows-cuda-deps
-          path: dist/deps
       - uses: actions/download-artifact@v4
         with:
           name: windows-rocm-deps
-          path: dist/deps
       - uses: actions/download-artifact@v4
         with:
           name: generate-windows-rocm
-          path: |
-            llm/build
-            dist/windows-amd64
       - run: dir llm/build
       - run: |
           $gopath=(get-command go).source | split-path -parent
@@ -342,8 +331,6 @@ jobs:
           $env:CMAKE_SYSTEM_VERSION="10.0.22621.0"
           $env:PATH="$gopath;$env:PATH"
           $env:OLLAMA_SKIP_GENERATE="1"
-          $env:NVIDIA_DIR=$(resolve-path ".\dist\deps")
-          $env:HIP_PATH=$(resolve-path ".\dist\deps")
           & .\scripts\build_windows.ps1
       - uses: actions/upload-artifact@v4
         with:

api/types.go

@@ -396,8 +396,10 @@ func (opts *Options) FromMap(m map[string]interface{}) error {
 func DefaultOptions() Options {
 	return Options{
 		// options set on request to runner
-		NumPredict:       -1,
-		NumKeep:          0,
+		NumPredict: -1,
+
+		// set a minimal num_keep to avoid issues on context shifts
+		NumKeep:          4,
 		Temperature:      0.8,
 		TopK:             40,
 		TopP:             0.9,

app/ollama.iss

@@ -92,12 +92,8 @@ Source: "..\dist\windows-amd64\*.dll"; DestDir: "{app}"; Flags: ignoreversion 64
 Source: "..\dist\windows-amd64\ollama_runners\*"; DestDir: "{app}\ollama_runners"; Flags: ignoreversion 64bit recursesubdirs
 Source: "..\dist\ollama_welcome.ps1"; DestDir: "{app}"; Flags: ignoreversion
 Source: ".\assets\app.ico"; DestDir: "{app}"; Flags: ignoreversion
-; Assumes v5.7, may need adjustments for v6
-#if GetEnv("HIP_PATH") != ""
-  Source: "{#GetEnv('HIP_PATH')}\bin\hipblas.dll"; DestDir: "{app}\rocm\"; Flags: ignoreversion
-  Source: "{#GetEnv('HIP_PATH')}\bin\rocblas.dll"; DestDir: "{app}\rocm\"; Flags: ignoreversion
-  ; amdhip64.dll dependency comes from the driver and must be installed already
-  Source: "{#GetEnv('HIP_PATH')}\bin\rocblas\library\*"; DestDir: "{app}\rocm\rocblas\library\"; Flags: ignoreversion
+#if DirExists("..\dist\windows-amd64\rocm")
+  Source: "..\dist\windows-amd64\rocm\*"; DestDir: "{app}\rocm\"; Flags: ignoreversion recursesubdirs
 #endif
 
 

cmd/cmd.go

@@ -17,6 +17,7 @@ import (
 	"os"
 	"os/signal"
 	"path/filepath"
+	"regexp"
 	"runtime"
 	"strings"
 	"syscall"
@@ -53,8 +54,6 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 	p := progress.NewProgress(os.Stderr)
 	defer p.Stop()
 
-	bars := make(map[string]*progress.Bar)
-
 	modelfile, err := os.ReadFile(filename)
 	if err != nil {
 		return err
@@ -95,95 +94,16 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 				return err
 			}
 
-			// TODO make this work w/ adapters
 			if fi.IsDir() {
-				tf, err := os.CreateTemp("", "ollama-tf")
+				// this is likely a safetensors or pytorch directory
+				// TODO make this work w/ adapters
+				tempfile, err := tempZipFiles(path)
 				if err != nil {
 					return err
 				}
-				defer os.RemoveAll(tf.Name())
+				defer os.RemoveAll(tempfile)
 
-				zf := zip.NewWriter(tf)
-
-				files := []string{}
-
-				tfiles, err := filepath.Glob(filepath.Join(path, "pytorch_model-*.bin"))
-				if err != nil {
-					return err
-				} else if len(tfiles) == 0 {
-					tfiles, err = filepath.Glob(filepath.Join(path, "model-*.safetensors"))
-					if err != nil {
-						return err
-					}
-				}
-
-				files = append(files, tfiles...)
-
-				if len(files) == 0 {
-					return fmt.Errorf("no models were found in '%s'", path)
-				}
-
-				// add the safetensor/torch config file + tokenizer
-				files = append(files, filepath.Join(path, "config.json"))
-				files = append(files, filepath.Join(path, "params.json"))
-				files = append(files, filepath.Join(path, "added_tokens.json"))
-				files = append(files, filepath.Join(path, "tokenizer.model"))
-
-				for _, fn := range files {
-					f, err := os.Open(fn)
-
-					// just skip whatever files aren't there
-					if os.IsNotExist(err) {
-						if strings.HasSuffix(fn, "tokenizer.model") {
-							// try the parent dir before giving up
-							parentDir := filepath.Dir(path)
-							newFn := filepath.Join(parentDir, "tokenizer.model")
-							f, err = os.Open(newFn)
-							if os.IsNotExist(err) {
-								continue
-							} else if err != nil {
-								return err
-							}
-						} else {
-							continue
-						}
-					} else if err != nil {
-						return err
-					}
-
-					fi, err := f.Stat()
-					if err != nil {
-						return err
-					}
-
-					h, err := zip.FileInfoHeader(fi)
-					if err != nil {
-						return err
-					}
-
-					h.Name = filepath.Base(fn)
-					h.Method = zip.Store
-
-					w, err := zf.CreateHeader(h)
-					if err != nil {
-						return err
-					}
-
-					_, err = io.Copy(w, f)
-					if err != nil {
-						return err
-					}
-
-				}
-
-				if err := zf.Close(); err != nil {
-					return err
-				}
-
-				if err := tf.Close(); err != nil {
-					return err
-				}
-				path = tf.Name()
+				path = tempfile
 			}
 
 			digest, err := createBlob(cmd, client, path)
@@ -191,10 +111,17 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 				return err
 			}
 
-			modelfile = bytes.ReplaceAll(modelfile, []byte(c.Args), []byte("@"+digest))
+			name := c.Name
+			if c.Name == "model" {
+				name = "from"
+			}
+
+			re := regexp.MustCompile(fmt.Sprintf(`(?im)^(%s)\s+%s\s*$`, name, c.Args))
+			modelfile = re.ReplaceAll(modelfile, []byte("$1 @"+digest))
 		}
 	}
 
+	bars := make(map[string]*progress.Bar)
 	fn := func(resp api.ProgressResponse) error {
 		if resp.Digest != "" {
 			spinner.Stop()
@@ -228,6 +155,88 @@ func CreateHandler(cmd *cobra.Command, args []string) error {
 	return nil
 }
 
+func tempZipFiles(path string) (string, error) {
+	tempfile, err := os.CreateTemp("", "ollama-tf")
+	if err != nil {
+		return "", err
+	}
+	defer tempfile.Close()
+
+	zipfile := zip.NewWriter(tempfile)
+	defer zipfile.Close()
+
+	tfiles, err := filepath.Glob(filepath.Join(path, "pytorch_model-*.bin"))
+	if err != nil {
+		return "", err
+	} else if len(tfiles) == 0 {
+		tfiles, err = filepath.Glob(filepath.Join(path, "model-*.safetensors"))
+		if err != nil {
+			return "", err
+		}
+	}
+
+	files := []string{}
+	files = append(files, tfiles...)
+
+	if len(files) == 0 {
+		return "", fmt.Errorf("no models were found in '%s'", path)
+	}
+
+	// add the safetensor/torch config file + tokenizer
+	files = append(files, filepath.Join(path, "config.json"))
+	files = append(files, filepath.Join(path, "params.json"))
+	files = append(files, filepath.Join(path, "added_tokens.json"))
+	files = append(files, filepath.Join(path, "tokenizer.model"))
+
+	for _, fn := range files {
+		f, err := os.Open(fn)
+
+		// just skip whatever files aren't there
+		if os.IsNotExist(err) {
+			if strings.HasSuffix(fn, "tokenizer.model") {
+				// try the parent dir before giving up
+				parentDir := filepath.Dir(path)
+				newFn := filepath.Join(parentDir, "tokenizer.model")
+				f, err = os.Open(newFn)
+				if os.IsNotExist(err) {
+					continue
+				} else if err != nil {
+					return "", err
+				}
+			} else {
+				continue
+			}
+		} else if err != nil {
+			return "", err
+		}
+
+		fi, err := f.Stat()
+		if err != nil {
+			return "", err
+		}
+
+		h, err := zip.FileInfoHeader(fi)
+		if err != nil {
+			return "", err
+		}
+
+		h.Name = filepath.Base(fn)
+		h.Method = zip.Store
+
+		w, err := zipfile.CreateHeader(h)
+		if err != nil {
+			return "", err
+		}
+
+		_, err = io.Copy(w, f)
+		if err != nil {
+			return "", err
+		}
+	}
+
+	return tempfile.Name(), nil
+}
+
 func createBlob(cmd *cobra.Command, client *api.Client, path string) (string, error) {
 	bin, err := os.Open(path)
 	if err != nil {

llm/generate/gen_common.sh

@@ -21,7 +21,7 @@ init_vars() {
         # TODO - add additional optimization flags...
         CMAKE_DEFS="-DCMAKE_BUILD_TYPE=Release -DLLAMA_SERVER_VERBOSE=off ${CMAKE_DEFS}"
     fi
-    case $(uname -s) in 
+    case $(uname -s) in
     "Darwin")
         LIB_EXT="dylib"
         WHOLE_ARCHIVE="-Wl,-force_load"

llm/generate/gen_linux.sh

@@ -165,11 +165,11 @@ if [ -d "${CUDA_LIB_DIR}" ]; then
     fi
     if [ "${ARCH}" == "arm64" ]; then
         echo "ARM CPU detected - disabling unsupported AVX instructions"
-        
+
         # ARM-based CPUs such as M1 and Tegra do not support AVX extensions.
         #
-        # CUDA compute < 6.0 lacks proper FP16 support on ARM. 
-        # Disabling has minimal performance effect while maintaining compatibility. 
+        # 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"
     fi
     # Users building from source can tune the exact flags we pass to cmake for configuring llama.cpp

llm/generate/gen_windows.ps1

@@ -252,6 +252,11 @@ if ($null -ne $script:CUDA_LIB_DIR) {
     build
     sign
     install
+
+    write-host "copying CUDA dependencies to ${script:SRC_DIR}\dist\windows-${script:ARCH}\"
+    cp "${script:CUDA_LIB_DIR}\cudart64_*.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\"
+    cp "${script:CUDA_LIB_DIR}\cublas64_*.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\"
+    cp "${script:CUDA_LIB_DIR}\cublasLt64_*.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\"
 }
 
 if ($null -ne $env:HIP_PATH) {
@@ -295,6 +300,14 @@ if ($null -ne $env:HIP_PATH) {
     }
     sign
     install
+
+    # Assumes v5.7, may need adjustments for v6
+    rm -ea 0 -recurse -force -path "${script:SRC_DIR}\dist\windows-${script:ARCH}\rocm\"
+    md "${script:SRC_DIR}\dist\windows-${script:ARCH}\rocm\rocblas\library\" -ea 0 > $null
+    cp "${env:HIP_PATH}\bin\hipblas.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\rocm\"
+    cp "${env:HIP_PATH}\bin\rocblas.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\rocm\"
+    # amdhip64.dll dependency comes from the driver and must be installed on the host to use AMD GPUs
+    cp "${env:HIP_PATH}\bin\rocblas\library\*" "${script:SRC_DIR}\dist\windows-${script:ARCH}\rocm\rocblas\library\"
 }
 
 

llm/memory.go

@@ -5,7 +5,6 @@ import (
 	"log/slog"
 	"os"
 	"strconv"
-	"strings"
 
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/format"
@@ -100,8 +99,22 @@ func EstimateGPULayers(gpus []gpu.GpuInfo, ggml *GGML, projectors []string, opts
 		return 0, 0
 	}
 
-	var layerCount int
 	layers := ggml.Tensors().Layers()
+
+	var memoryLayerOutput uint64
+	for k, v := range layers {
+		if k == "output" || k == "output_norm" {
+			memoryLayerOutput += v.size()
+		}
+	}
+
+	if gpus[0].Library == "metal" && opts.UseMMap {
+		// memory is preallocated for output tensors
+		memoryRequiredTotal += memoryLayerOutput
+		memoryRequiredPartial += memoryLayerOutput
+	}
+
+	var layerCount int
 	for i := 0; i < int(ggml.KV().BlockCount()); i++ {
 		memoryLayer := layers[fmt.Sprintf("blk.%d", i)].size()
 
@@ -115,15 +128,11 @@ func EstimateGPULayers(gpus []gpu.GpuInfo, ggml *GGML, projectors []string, opts
 		}
 	}
 
-	var memoryLayerOutput uint64
-	for k, v := range layers {
-		if !strings.HasPrefix(k, "blk.") {
-			memoryLayerOutput += v.size()
-		}
+	if gpus[0].Library != "metal" || !opts.UseMMap {
+		// memory was not preallocated for output tensors
+		memoryRequiredTotal += memoryLayerOutput
 	}
 
-	memoryRequiredTotal += memoryLayerOutput
-
 	if memoryAvailable > memoryRequiredTotal {
 		layerCount = int(ggml.KV().BlockCount()) + 1
 		memoryRequiredPartial = memoryRequiredTotal

llm/patches/04-metal.diff

@@ -0,0 +1,45 @@
+diff --git a/ggml-metal.m b/ggml-metal.m
+index 0207b787..b5e9884b 100644
+--- a/ggml-metal.m
++++ b/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;
+ 
+-#if 0
+                         // the numbers below are measured on M2 Ultra for 7B and 13B models
+                         // these numbers do not translate to other devices or model sizes
+                         // TODO: need to find a better approach
+-                        if ([ctx->device.name isEqualToString:@"Apple M2 Ultra"]) {
+-                            switch (src0t) {
+-                                case GGML_TYPE_F16:  ne11_mm_min = 2;  break;
+-                                case GGML_TYPE_Q8_0: ne11_mm_min = 7;  break;
+-                                case GGML_TYPE_Q2_K: ne11_mm_min = 15; break;
+-                                case GGML_TYPE_Q3_K: ne11_mm_min = 7;  break;
+-                                case GGML_TYPE_Q4_0:
+-                                case GGML_TYPE_Q4_1: ne11_mm_min = 15; break;
+-                                case GGML_TYPE_Q4_K: ne11_mm_min = 11; break;
+-                                case GGML_TYPE_Q5_0:                          // not tested yet
+-                                case GGML_TYPE_Q5_1: ne11_mm_min = 13; break; // not tested yet
+-                                case GGML_TYPE_Q5_K: ne11_mm_min = 7;  break;
+-                                case GGML_TYPE_Q6_K: ne11_mm_min = 7;  break;
+-                                default:             ne11_mm_min = 1;  break;
+-                            }
++                        switch (src0t) {
++                            case GGML_TYPE_F16:  ne11_mm_min = 2;  break;
++                            case GGML_TYPE_Q8_0: ne11_mm_min = 7;  break;
++                            case GGML_TYPE_Q2_K: ne11_mm_min = 15; break;
++                            case GGML_TYPE_Q3_K: ne11_mm_min = 7;  break;
++                            case GGML_TYPE_Q4_0:
++                            case GGML_TYPE_Q4_1: ne11_mm_min = 15; break;
++                            case GGML_TYPE_Q4_K: ne11_mm_min = 11; break;
++                            case GGML_TYPE_Q5_0:                          // not tested yet
++                            case GGML_TYPE_Q5_1: ne11_mm_min = 13; break; // not tested yet
++                            case GGML_TYPE_Q5_K: ne11_mm_min = 7;  break;
++                            case GGML_TYPE_Q6_K: ne11_mm_min = 7;  break;
++                            default:             ne11_mm_min = 1;  break;
+                         }
+-#endif
+ 
+                         // for now the matrix-matrix multiplication kernel only works on A14+/M1+ SoCs
+                         // AMD GPU and older A-chips will reuse matrix-vector multiplication kernel

llm/server.go

@@ -560,6 +560,13 @@ func (s *llmServer) Completion(ctx context.Context, req CompletionRequest, fn fu
 		return err
 	}
 	defer s.sem.Release(1)
+
+	// only allow maximum 10 "context shifts" to avoid infinite generation
+	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{
 		"prompt":            req.Prompt,
 		"stream":            true,

scripts/build_windows.ps1

@@ -109,9 +109,6 @@ function gatherDependencies() {
     cp "${env:VCToolsRedistDir}\x64\Microsoft.VC*.CRT\vcruntime140.dll" "${script:DEPS_DIR}\"
     cp "${env:VCToolsRedistDir}\x64\Microsoft.VC*.CRT\vcruntime140_1.dll" "${script:DEPS_DIR}\"
 
-    cp "${script:NVIDIA_DIR}\cudart64_*.dll" "${script:DEPS_DIR}\"
-    cp "${script:NVIDIA_DIR}\cublas64_*.dll" "${script:DEPS_DIR}\"
-    cp "${script:NVIDIA_DIR}\cublasLt64_*.dll" "${script:DEPS_DIR}\"
 
     cp "${script:SRC_DIR}\app\ollama_welcome.ps1" "${script:SRC_DIR}\dist\"
     if ("${env:KEY_CONTAINER}") {
@@ -123,15 +120,6 @@ function gatherDependencies() {
             if ($LASTEXITCODE -ne 0) { exit($LASTEXITCODE)}
         }
     }
-    if ($null -ne $env:HIP_PATH) {
-        # Assumes v5.7, may need adjustments for v6
-        rm -ea 0 -recurse -force -path "${script:DEPS_DIR}\rocm\"
-        md "${script:DEPS_DIR}\rocm\rocblas\library\" -ea 0 > $null
-        cp "${env:HIP_PATH}\bin\hipblas.dll" "${script:DEPS_DIR}\rocm\"
-        cp "${env:HIP_PATH}\bin\rocblas.dll" "${script:DEPS_DIR}\rocm\"
-        # amdhip64.dll dependency comes from the driver and must be installed on the host to use AMD GPUs
-        cp "${env:HIP_PATH}\bin\rocblas\library\*" "${script:DEPS_DIR}\rocm\rocblas\library\"
-    }
 }
 
 function buildInstaller() {

server/images.go

@@ -29,6 +29,7 @@ import (
 	"github.com/ollama/ollama/format"
 	"github.com/ollama/ollama/llm"
 	"github.com/ollama/ollama/parser"
+	"github.com/ollama/ollama/types/model"
 	"github.com/ollama/ollama/version"
 )
 
@@ -701,36 +702,32 @@ func convertModel(name, path string, fn func(resp api.ProgressResponse)) (string
 	return path, nil
 }
 
-func CopyModel(src, dest string) error {
-	srcModelPath := ParseModelPath(src)
-	srcPath, err := srcModelPath.GetManifestPath()
+func CopyModel(src, dst model.Name) error {
+	manifests, err := GetManifestPath()
 	if err != nil {
 		return err
 	}
 
-	destModelPath := ParseModelPath(dest)
-	destPath, err := destModelPath.GetManifestPath()
-	if err != nil {
-		return err
-	}
-	if err := os.MkdirAll(filepath.Dir(destPath), 0o755); err != nil {
+	dstpath := filepath.Join(manifests, dst.FilepathNoBuild())
+	if err := os.MkdirAll(filepath.Dir(dstpath), 0o755); err != nil {
 		return err
 	}
 
-	// copy the file
-	input, err := os.ReadFile(srcPath)
+	srcpath := filepath.Join(manifests, src.FilepathNoBuild())
+	srcfile, err := os.Open(srcpath)
 	if err != nil {
-		fmt.Println("Error reading file:", err)
 		return err
 	}
+	defer srcfile.Close()
 
-	err = os.WriteFile(destPath, input, 0o644)
+	dstfile, err := os.Create(dstpath)
 	if err != nil {
-		fmt.Println("Error reading file:", err)
 		return err
 	}
+	defer dstfile.Close()
 
-	return nil
+	_, err = io.Copy(dstfile, srcfile)
+	return err
 }
 
 func deleteUnusedLayers(skipModelPath *ModelPath, deleteMap map[string]struct{}, dryRun bool) error {

server/routes.go

@@ -29,6 +29,7 @@ import (
 	"github.com/ollama/ollama/llm"
 	"github.com/ollama/ollama/openai"
 	"github.com/ollama/ollama/parser"
+	"github.com/ollama/ollama/types/model"
 	"github.com/ollama/ollama/version"
 )
 
@@ -788,34 +789,34 @@ func (s *Server) ListModelsHandler(c *gin.Context) {
 }
 
 func (s *Server) CopyModelHandler(c *gin.Context) {
-	var req api.CopyRequest
-	err := c.ShouldBindJSON(&req)
-	switch {
-	case errors.Is(err, io.EOF):
+	var r api.CopyRequest
+	if err := c.ShouldBindJSON(&r); errors.Is(err, io.EOF) {
 		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "missing request body"})
 		return
-	case err != nil:
+	} else if err != nil {
 		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": err.Error()})
 		return
 	}
 
-	if req.Source == "" || req.Destination == "" {
-		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "source add destination are required"})
+	src := model.ParseName(r.Source)
+	if !src.IsValid() {
+		_ = c.Error(fmt.Errorf("source %q is invalid", r.Source))
+	}
+
+	dst := model.ParseName(r.Destination)
+	if !dst.IsValid() {
+		_ = c.Error(fmt.Errorf("destination %q is invalid", r.Destination))
+	}
+
+	if len(c.Errors) > 0 {
+		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": c.Errors.Errors()})
 		return
 	}
 
-	if err := ParseModelPath(req.Destination).Validate(); err != nil {
-		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": err.Error()})
-		return
-	}
-
-	if err := CopyModel(req.Source, req.Destination); err != nil {
-		if os.IsNotExist(err) {
-			c.JSON(http.StatusNotFound, gin.H{"error": fmt.Sprintf("model '%s' not found", req.Source)})
-		} else {
-			c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
-		}
-		return
+	if err := CopyModel(src, dst); errors.Is(err, os.ErrNotExist) {
+		c.JSON(http.StatusNotFound, gin.H{"error": fmt.Sprintf("model %q not found", r.Source)})
+	} else if err != nil {
+		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
 	}
 }
 

server/sched.go

@@ -23,7 +23,6 @@ import (
 type LlmRequest struct {
 	ctx             context.Context //nolint:containedctx
 	model           *Model
-	ggml            *llm.GGML // TODO - how large is this, and do we need to free it after we've finished loading?
 	opts            api.Options
 	sessionDuration time.Duration
 	successCh       chan *runnerRef
@@ -39,7 +38,7 @@ type Scheduler struct {
 	loaded   map[string]*runnerRef
 	loadedMu sync.Mutex
 
-	loadFn      func(req *LlmRequest, gpus gpu.GpuInfoList)
+	loadFn      func(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList)
 	newServerFn func(gpus gpu.GpuInfoList, model string, ggml *llm.GGML, adapters []string, projectors []string, opts api.Options) (llm.LlamaServer, error)
 	getGpuFn    func() gpu.GpuInfoList
 }
@@ -47,6 +46,7 @@ type Scheduler struct {
 // TODO set this to zero after a release or two, to enable multiple models by default
 var loadedMax = 1          // Maximum runners; < 1 maps to as many as will fit in VRAM (unlimited for CPU runners)
 var maxQueuedRequests = 10 // TODO configurable
+var numParallel = 1
 
 func InitScheduler(ctx context.Context) *Scheduler {
 	maxRunners := os.Getenv("OLLAMA_MAX_LOADED_MODELS")
@@ -58,6 +58,14 @@ func InitScheduler(ctx context.Context) *Scheduler {
 			loadedMax = m
 		}
 	}
+	if onp := os.Getenv("OLLAMA_NUM_PARALLEL"); onp != "" {
+		p, err := strconv.Atoi(onp)
+		if err != nil || p <= 0 {
+			slog.Error("invalid parallel setting, must be greater than zero", "OLLAMA_NUM_PARALLEL", onp, "error", err)
+		} else {
+			numParallel = p
+		}
+	}
 
 	sched := &Scheduler{
 		pendingReqCh:  make(chan *LlmRequest, maxQueuedRequests),
@@ -74,20 +82,16 @@ 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) {
-	ggml, err := llm.LoadModel(model.ModelPath)
 	req := &LlmRequest{
 		ctx:             c,
 		model:           model,
-		ggml:            ggml,
 		opts:            opts,
 		sessionDuration: sessionDuration,
 		successCh:       make(chan *runnerRef),
 		errCh:           make(chan error, 1),
 	}
-	if err != nil {
-		req.errCh <- err
-		return req.successCh, req.errCh
-	}
+	// context split across parallel threads
+	opts.NumCtx = opts.NumCtx * numParallel
 	select {
 	case s.pendingReqCh <- req:
 	default:
@@ -130,28 +134,39 @@ func (s *Scheduler) processPending(ctx context.Context) {
 						pending.useLoadedRunner(runner, s.finishedReqCh)
 						break
 					}
-				} else if loadedCount == 0 {
-					slog.Debug("loading first model", "model", pending.model.ModelPath)
-					gpus := s.getGpuFn()
-					g := pickBestFitGPUs(pending, gpus)
-					if g != nil {
-						gpus = g
-					}
-					s.loadFn(pending, gpus)
-					break
 				} else if loadedMax > 0 && loadedCount >= loadedMax {
 					slog.Debug("max runners achieved, unloading one to make room", "runner_count", loadedCount)
 					runnerToExpire = s.findRunnerToUnload(pending)
 				} else {
-					// More than one loaded model, so we have to see if the new one fits
+					// Either no models are loaded or below loadedMax
 					// Get a refreshed GPU list
 					gpus := s.getGpuFn()
+
+					// Load model for fitting
+					ggml, err := llm.LoadModel(pending.model.ModelPath)
+					if err != nil {
+						pending.errCh <- err
+						break
+					}
+
+					// No models loaded. Load the model but prefer the best fit.
+					if loadedCount == 0 {
+						slog.Debug("loading first model", "model", pending.model.ModelPath)
+						g := pickBestFitGPUs(pending, ggml, gpus)
+						if g != nil {
+							gpus = g
+						}
+						s.loadFn(pending, ggml, gpus)
+						break
+					}
+
+					// More than one loaded model, so we have to see if the new one fits
 					// Update free memory from currently loaded models
 					s.updateFreeSpace(gpus)
-					gpus = pickBestFitGPUs(pending, gpus)
+					gpus = pickBestFitGPUs(pending, ggml, gpus)
 					if gpus != nil {
 						slog.Debug("new model fits with existing models, loading")
-						s.loadFn(pending, gpus)
+						s.loadFn(pending, ggml, gpus)
 						break
 					}
 					runnerToExpire = s.findRunnerToUnload(pending)
@@ -282,8 +297,8 @@ func (pending *LlmRequest) useLoadedRunner(runner *runnerRef, finished chan *Llm
 	}()
 }
 
-func (s *Scheduler) load(req *LlmRequest, gpus gpu.GpuInfoList) {
-	llama, err := s.newServerFn(gpus, req.model.ModelPath, req.ggml, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts)
+func (s *Scheduler) load(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList) {
+	llama, err := s.newServerFn(gpus, req.model.ModelPath, ggml, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts)
 	if err != nil {
 		// some older models are not compatible with newer versions of llama.cpp
 		// show a generalized compatibility error until there is a better way to
@@ -417,16 +432,21 @@ func (runner *runnerRef) needsReload(ctx context.Context, req *LlmRequest) bool
 	slog.Debug("evaluating already loaded", "model", req.model.ModelPath)
 	runner.refMu.Lock()
 	defer runner.refMu.Unlock()
-	// Ignore the NumGPU settings for comparison
-	optsExisting := runner.Options.Runner
-	optsExisting.NumGPU = -1
-	optsNew := req.opts.Runner
-	optsNew.NumGPU = -1
+
 	timeout := 10 * time.Second
 	if runner.loading {
 		timeout = 2 * time.Minute // Initial load can take a long time for big models on slow systems...
 	}
-	ctx, cancel := context.WithTimeout(ctx, timeout) // BUG -
+
+	// Don't reload runner if num_gpu=-1 was provided
+	optsExisting := runner.Options.Runner
+	optsNew := req.opts.Runner
+	if optsNew.NumGPU < 0 {
+		optsExisting.NumGPU = -1
+		optsNew.NumGPU = -1
+	}
+
+	ctx, cancel := context.WithTimeout(ctx, timeout)
 	defer cancel()
 	if !reflect.DeepEqual(runner.adapters, req.model.AdapterPaths) || // have the adapters changed?
 		!reflect.DeepEqual(runner.projectors, req.model.ProjectorPaths) || // have the projectors changed?
@@ -434,6 +454,7 @@ func (runner *runnerRef) needsReload(ctx context.Context, req *LlmRequest) bool
 		runner.llama.Ping(ctx) != nil {
 		return true
 	}
+
 	return false
 }
 
@@ -454,7 +475,7 @@ func (a ByDuration) Less(i, j int) bool {
 
 // pickBestFitGPUs will try to find the optimal placement of the model in the available GPUs where the model fully fits
 // If the model can not be fit fully within the available GPU(s) nil is returned
-func pickBestFitGPUs(req *LlmRequest, gpus gpu.GpuInfoList) gpu.GpuInfoList {
+func pickBestFitGPUs(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList) gpu.GpuInfoList {
 	var estimatedVRAM uint64
 	for _, gl := range gpus.ByLibrary() {
 		var ok bool
@@ -466,7 +487,7 @@ func pickBestFitGPUs(req *LlmRequest, gpus gpu.GpuInfoList) gpu.GpuInfoList {
 
 		// First attempt to fit the model into a single GPU
 		for _, g := range sgl {
-			if ok, estimatedVRAM = llm.PredictServerFit([]gpu.GpuInfo{g}, req.ggml, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts); ok {
+			if ok, estimatedVRAM = llm.PredictServerFit([]gpu.GpuInfo{g}, ggml, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts); ok {
 				slog.Debug("new model will fit in available VRAM in single GPU, loading", "model", req.model.ModelPath, "gpu", g.ID, "available", g.FreeMemory, "required", format.HumanBytes2(estimatedVRAM))
 				return []gpu.GpuInfo{g}
 			}
@@ -477,7 +498,7 @@ func pickBestFitGPUs(req *LlmRequest, gpus gpu.GpuInfoList) gpu.GpuInfoList {
 		// - try subsets of GPUs instead of just falling back to 1 or all in a family
 
 		// Now try all the GPUs
-		if ok, estimatedVRAM = llm.PredictServerFit(gl, req.ggml, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts); ok {
+		if ok, estimatedVRAM = llm.PredictServerFit(gl, ggml, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts); ok {
 			slog.Debug("new model will fit in available VRAM, loading", "model", req.model.ModelPath, "library", gl[0].Library, "required", format.HumanBytes2(estimatedVRAM))
 			return gl
 		}

server/sched_test.go

@@ -47,6 +47,7 @@ func TestLoad(t *testing.T) {
 	ctx, done := context.WithTimeout(context.Background(), 5*time.Millisecond)
 	defer done()
 	s := InitScheduler(ctx)
+	var ggml *llm.GGML // value not used in tests
 	req := &LlmRequest{
 		ctx:             ctx,
 		model:           &Model{ModelPath: "foo"},
@@ -59,7 +60,7 @@ func TestLoad(t *testing.T) {
 		return nil, fmt.Errorf("something failed to load model blah")
 	}
 	gpus := gpu.GpuInfoList{}
-	s.load(req, gpus)
+	s.load(req, ggml, gpus)
 	require.Len(t, req.successCh, 0)
 	require.Len(t, req.errCh, 1)
 	require.Len(t, s.loaded, 0)
@@ -70,7 +71,7 @@ func TestLoad(t *testing.T) {
 	s.newServerFn = func(gpus gpu.GpuInfoList, model string, ggml *llm.GGML, adapters []string, projectors []string, opts api.Options) (llm.LlamaServer, error) {
 		return server, nil
 	}
-	s.load(req, gpus)
+	s.load(req, ggml, gpus)
 	select {
 	case err := <-req.errCh:
 		require.NoError(t, err)
@@ -82,7 +83,7 @@ func TestLoad(t *testing.T) {
 
 	req.model.ModelPath = "dummy_model_path"
 	server.waitResp = fmt.Errorf("wait failure")
-	s.load(req, gpus)
+	s.load(req, ggml, gpus)
 	select {
 	case err := <-req.errCh:
 		require.Contains(t, err.Error(), "wait failure")
@@ -101,6 +102,7 @@ type bundle struct {
 	ctxDone func()
 	srv     *mockLlm
 	req     *LlmRequest
+	ggml    *llm.GGML
 }
 
 func (scenario *bundle) newServer(gpus gpu.GpuInfoList, model string, ggml *llm.GGML, adapters []string, projectors []string, opts api.Options) (llm.LlamaServer, error) {
@@ -132,14 +134,15 @@ func newScenario(t *testing.T, ctx context.Context, modelName string, estimatedV
 		{Name: "blk.0.attn.weight", Kind: uint32(0), Offset: uint64(0), Shape: []uint64{1, 1, 1, 1}, WriterTo: &bytes.Reader{}},
 	})
 	assert.Nil(t, err)
+
 	fname := f.Name()
 	model := &Model{Name: modelName, ModelPath: fname}
-	ggml, err := llm.LoadModel(model.ModelPath)
+	scenario.ggml, err = llm.LoadModel(model.ModelPath)
 	require.NoError(t, err)
+
 	scenario.req = &LlmRequest{
 		ctx:             scenario.ctx,
 		model:           model,
-		ggml:            ggml,
 		sessionDuration: 5 * time.Millisecond,
 		successCh:       make(chan *runnerRef, 1),
 		errCh:           make(chan error, 1),
@@ -157,13 +160,13 @@ func TestRequests(t *testing.T) {
 	scenario1a.req.sessionDuration = 0
 	scenario1b := newScenario(t, ctx, "ollama-model-1", 11)
 	scenario1b.req.model = scenario1a.req.model
-	scenario1b.req.ggml = scenario1a.req.ggml
+	scenario1b.ggml = scenario1a.ggml
 	scenario1b.req.sessionDuration = 0
 
 	// simple reload of same model
 	scenario2a := newScenario(t, ctx, "ollama-model-1", 20)
 	scenario2a.req.model = scenario1a.req.model
-	scenario2a.req.ggml = scenario1a.req.ggml
+	scenario2a.ggml = scenario1a.ggml
 
 	// Multiple loaded models
 	scenario3a := newScenario(t, ctx, "ollama-model-3a", 1*format.GigaByte)
@@ -322,13 +325,14 @@ func TestGetRunner(t *testing.T) {
 	successCh1c, errCh1c := s.GetRunner(scenario1c.ctx, scenario1c.req.model, scenario1c.req.opts, scenario1c.req.sessionDuration)
 	require.Len(t, s.pendingReqCh, 0)
 	require.Len(t, successCh1c, 0)
+	require.Len(t, errCh1c, 0)
+
+	time.Sleep(5 * time.Millisecond)
+	require.Len(t, s.loaded, 0)
 	require.Len(t, errCh1c, 1)
 	err = <-errCh1c
 	require.Contains(t, err.Error(), "bad path")
 	scenario1b.ctxDone()
-
-	time.Sleep(5 * time.Millisecond)
-	require.Len(t, s.loaded, 0)
 }
 
 // TODO - add one scenario that triggers the bogus finished event with positive ref count
@@ -366,7 +370,9 @@ func TestPrematureExpired(t *testing.T) {
 	require.LessOrEqual(t, len(s.finishedReqCh), 1)
 	time.Sleep(10 * time.Millisecond)
 	require.Len(t, s.finishedReqCh, 0)
+	s.loadedMu.Lock()
 	require.Len(t, s.loaded, 0)
+	s.loadedMu.Unlock()
 
 	// also shouldn't happen in real life
 	s.finishedReqCh <- scenario1a.req
@@ -426,7 +432,6 @@ func TestUpdateFreeSpace(t *testing.T) {
 	s.updateFreeSpace(gpus)
 	require.Equal(t, uint64(850), gpus[0].FreeMemory)
 	require.Equal(t, uint64(1850), gpus[1].FreeMemory)
-
 }
 
 func TestFindRunnerToUnload(t *testing.T) {
@@ -485,6 +490,9 @@ func TestNeedsReload(t *testing.T) {
 	require.False(t, resp)
 	req.opts.NumGPU = 99
 	resp = runner.needsReload(ctx, req)
+	require.True(t, resp)
+	req.opts.NumGPU = -1
+	resp = runner.needsReload(ctx, req)
 	require.False(t, resp)
 }
 

types/model/name.go

@@ -686,10 +686,7 @@ func IsValidNamePart(kind PartKind, s string) bool {
 		return false
 	}
 	var consecutiveDots int
-	for i, c := range []byte(s) {
-		if i == 0 && !isAlphaNumeric(c) {
-			return false
-		}
+	for _, c := range []byte(s) {
 		if c == '.' {
 			if consecutiveDots++; consecutiveDots >= 2 {
 				return false
@@ -704,10 +701,6 @@ func IsValidNamePart(kind PartKind, s string) bool {
 	return true
 }
 
-func isAlphaNumeric(c byte) bool {
-	return c >= 'a' && c <= 'z' || c >= 'A' && c <= 'Z' || c >= '0' && c <= '9'
-}
-
 func isValidByteFor(kind PartKind, c byte) bool {
 	if kind == PartNamespace && c == '.' {
 		return false

types/model/name_test.go

@@ -101,8 +101,6 @@ var testNames = map[string]fields{
 	"./../passwd":    {},
 	"./0+..":         {},
 
-	"-h": {},
-
 	strings.Repeat("a", MaxNamePartLen):   {model: strings.Repeat("a", MaxNamePartLen)},
 	strings.Repeat("a", MaxNamePartLen+1): {},
 }
@@ -119,7 +117,7 @@ func TestIsValidNameLen(t *testing.T) {
 // preventing path traversal.
 func TestNameConsecutiveDots(t *testing.T) {
 	for i := 1; i < 10; i++ {
-		s := "a" + strings.Repeat(".", i)
+		s := strings.Repeat(".", i)
 		if i > 1 {
 			if g := ParseNameFill(s, FillNothing).DisplayLong(); g != "" {
 				t.Errorf("ParseName(%q) = %q; want empty string", s, g)
@@ -341,17 +339,17 @@ func TestDisplayShortest(t *testing.T) {
 		want      string
 		wantPanic bool
 	}{
-		{"example.com/library/mistral:latest+Q4_0", "example.com/library/?:latest", "mistral", false},
-		{"example.com/library/mistral:latest+Q4_0", "example.com/?/?:latest", "library/mistral", false},
+		{"example.com/library/mistral:latest+Q4_0", "example.com/library/_:latest", "mistral", false},
+		{"example.com/library/mistral:latest+Q4_0", "example.com/_/_:latest", "library/mistral", false},
 		{"example.com/library/mistral:latest+Q4_0", "", "example.com/library/mistral", false},
 		{"example.com/library/mistral:latest+Q4_0", "", "example.com/library/mistral", false},
 
 		// case-insensitive
-		{"Example.com/library/mistral:latest+Q4_0", "example.com/library/?:latest", "mistral", false},
-		{"example.com/Library/mistral:latest+Q4_0", "example.com/library/?:latest", "mistral", false},
-		{"example.com/library/Mistral:latest+Q4_0", "example.com/library/?:latest", "Mistral", false},
-		{"example.com/library/mistral:Latest+Q4_0", "example.com/library/?:latest", "mistral", false},
-		{"example.com/library/mistral:Latest+q4_0", "example.com/library/?:latest", "mistral", false},
+		{"Example.com/library/mistral:latest+Q4_0", "example.com/library/_:latest", "mistral", false},
+		{"example.com/Library/mistral:latest+Q4_0", "example.com/library/_:latest", "mistral", false},
+		{"example.com/library/Mistral:latest+Q4_0", "example.com/library/_:latest", "Mistral", false},
+		{"example.com/library/mistral:Latest+Q4_0", "example.com/library/_:latest", "mistral", false},
+		{"example.com/library/mistral:Latest+q4_0", "example.com/library/_:latest", "mistral", false},
 
 		// zero value
 		{"", MaskDefault, "", true},
@@ -363,10 +361,10 @@ func TestDisplayShortest(t *testing.T) {
 		{"registry.ollama.ai/library/mistral:latest+Q4_0", MaskDefault, "mistral", false},
 
 		// Auto-Fill
-		{"x", "example.com/library/?:latest", "x", false},
-		{"x", "example.com/library/?:latest+Q4_0", "x", false},
-		{"x/y:z", "a.com/library/?:latest+Q4_0", "x/y:z", false},
-		{"x/y:z", "a.com/library/?:latest+Q4_0", "x/y:z", false},
+		{"x", "example.com/library/_:latest", "x", false},
+		{"x", "example.com/library/_:latest+Q4_0", "x", false},
+		{"x/y:z", "a.com/library/_:latest+Q4_0", "x/y:z", false},
+		{"x/y:z", "a.com/library/_:latest+Q4_0", "x/y:z", false},
 	}
 
 	for _, tt := range cases {
@@ -697,10 +695,10 @@ func ExampleName_completeAndResolved() {
 func ExampleName_DisplayShortest() {
 	name := ParseNameFill("example.com/jmorganca/mistral:latest+Q4_0", FillNothing)
 
-	fmt.Println(name.DisplayShortest("example.com/jmorganca/?:latest"))
-	fmt.Println(name.DisplayShortest("example.com/?/?:latest"))
-	fmt.Println(name.DisplayShortest("example.com/?/?:?"))
-	fmt.Println(name.DisplayShortest("?/?/?:?"))
+	fmt.Println(name.DisplayShortest("example.com/jmorganca/_:latest"))
+	fmt.Println(name.DisplayShortest("example.com/_/_:latest"))
+	fmt.Println(name.DisplayShortest("example.com/_/_:_"))
+	fmt.Println(name.DisplayShortest("_/_/_:_"))
 
 	// Default
 	name = ParseNameFill("registry.ollama.ai/library/mistral:latest+Q4_0", FillNothing)