Sync with upstream ollama/ollama and restore Tesla K80 (compute 3.7) support

This commit represents a complete rework after pulling the latest changes from official ollama/ollama repository and re-applying Tesla K80 compatibility patches. ## Key Changes ### CUDA Compute Capability 3.7 Support (Tesla K80) - Added sm_37 (compute 3.7) to CMAKE_CUDA_ARCHITECTURES in CMakeLists.txt - Updated CMakePresets.json to include compute 3.7 in "CUDA 11" preset - Using 37-virtual (PTX with JIT compilation) for maximum compatibility ### Legacy Toolchain Compatibility - **NVIDIA Driver**: 470.256.02 (last version supporting Kepler/K80) - **CUDA Version**: 11.4.4 (last CUDA 11.x supporting compute 3.7) - **GCC Version**: 10.5.0 (required by CUDA 11.4 host_config.h) ### CPU Architecture Trade-offs Due to GCC 10.5 limitation, sacrificed newer CPU optimizations: - Alderlake CPU variant enabled WITHOUT AVX_VNNI (requires GCC 11+) - Still supports: SSE4.2, AVX, F16C, AVX2, BMI2, FMA - Performance impact: ~3-7% on newer CPUs (acceptable for K80 compatibility) ### Build System Updates - Modified ml/backend/ggml/ggml/src/ggml-cuda/CMakeLists.txt for compute 3.7 - Added -Wno-deprecated-gpu-targets flag to suppress warnings - Updated ml/backend/ggml/ggml/src/CMakeLists.txt for Alderlake without AVX_VNNI ### Upstream Sync Merged latest llama.cpp changes including: - Enhanced KV cache management with ISWA and hybrid memory support - Improved multi-modal support (mtmd framework) - New model architectures (Gemma3, Llama4, Qwen3, etc.) - GPU backend improvements for CUDA, Metal, and ROCm - Updated quantization support and GGUF format handling ### Documentation - Updated CLAUDE.md with comprehensive build instructions - Documented toolchain constraints and CPU architecture trade-offs - Removed outdated CI/CD workflows (tesla-k80-*.yml) - Cleaned up temporary development artifacts ## Rationale This fork maintains Tesla K80 GPU support (compute 3.7) which was dropped in official Ollama due to legacy driver/CUDA requirements. The toolchain constraint creates a deadlock: - K80 → Driver 470 → CUDA 11.4 → GCC 10 → No AVX_VNNI We accept the loss of cutting-edge CPU optimizations to enable running modern LLMs on legacy but still capable Tesla K80 hardware (12GB VRAM per GPU). 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>
2025-12-18 03:37:09 +00:00 · 2025-11-05 14:03:05 +08:00
parent fabe2c5cb7
commit ef14fb5b26
817 changed files with 241634 additions and 70888 deletions
--- a/cmd/cmd.go
+++ b/cmd/cmd.go
@@ -47,6 +47,8 @@ import (
 	"github.com/ollama/ollama/version"
 )

+const ConnectInstructions = "To sign in, navigate to:\n    %s\n\n"
+
 // ensureThinkingSupport emits a warning if the model does not advertise thinking support
 func ensureThinkingSupport(ctx context.Context, client *api.Client, name string) {
 	if name == "" {
@@ -56,10 +58,8 @@ func ensureThinkingSupport(ctx context.Context, client *api.Client, name string)
 	if err != nil {
 		return
 	}
-	for _, cap := range resp.Capabilities {
-		if cap == model.CapabilityThinking {
-			return
-		}
+	if slices.Contains(resp.Capabilities, model.CapabilityThinking) {
+		return
 	}
 	fmt.Fprintf(os.Stderr, "warning: model %q does not support thinking output\n", name)
 }
@@ -280,6 +280,13 @@ func loadOrUnloadModel(cmd *cobra.Command, opts *runOptions) error {
 		return err
 	}

+	if info, err := client.Show(cmd.Context(), &api.ShowRequest{Model: opts.Model}); err != nil {
+		return err
+	} else if info.RemoteHost != "" {
+		// Cloud model, no need to load/unload
+		return nil
+	}
+
 	req := &api.GenerateRequest{
 		Model:     opts.Model,
 		KeepAlive: opts.KeepAlive,
@@ -288,7 +295,17 @@ func loadOrUnloadModel(cmd *cobra.Command, opts *runOptions) error {
 		Think: opts.Think,
 	}

-	return client.Generate(cmd.Context(), req, func(api.GenerateResponse) error { return nil })
+	return client.Generate(cmd.Context(), req, func(r api.GenerateResponse) error {
+		if r.RemoteModel != "" && opts.ShowConnect {
+			p.StopAndClear()
+			if strings.HasPrefix(r.RemoteHost, "https://ollama.com") {
+				fmt.Fprintf(os.Stderr, "Connecting to '%s' on 'ollama.com' ⚡\n", r.RemoteModel)
+			} else {
+				fmt.Fprintf(os.Stderr, "Connecting to '%s' on '%s'\n", r.RemoteModel, r.RemoteHost)
+			}
+		}
+		return nil
+	})
 }

 func StopHandler(cmd *cobra.Command, args []string) error {
@@ -309,9 +326,10 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 	interactive := true

 	opts := runOptions{
-		Model:    args[0],
-		WordWrap: os.Getenv("TERM") == "xterm-256color",
-		Options:  map[string]any{},
+		Model:       args[0],
+		WordWrap:    os.Getenv("TERM") == "xterm-256color",
+		Options:     map[string]any{},
+		ShowConnect: true,
 	}

 	format, err := cmd.Flags().GetString("format")
@@ -369,6 +387,7 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 		}

 		prompts = append([]string{string(in)}, prompts...)
+		opts.ShowConnect = false
 		opts.WordWrap = false
 		interactive = false
 	}
@@ -435,6 +454,15 @@ func RunHandler(cmd *cobra.Command, args []string) error {

 	if interactive {
 		if err := loadOrUnloadModel(cmd, &opts); err != nil {
+			var sErr api.AuthorizationError
+			if errors.As(err, &sErr) && sErr.StatusCode == http.StatusUnauthorized {
+				fmt.Printf("You need to be signed in to Ollama to run Cloud models.\n\n")
+
+				if sErr.SigninURL != "" {
+					fmt.Printf(ConnectInstructions, sErr.SigninURL)
+				}
+				return nil
+			}
 			return err
 		}

@@ -455,6 +483,59 @@ func RunHandler(cmd *cobra.Command, args []string) error {
 	return generate(cmd, opts)
 }

+func SigninHandler(cmd *cobra.Command, args []string) error {
+	client, err := api.ClientFromEnvironment()
+	if err != nil {
+		return err
+	}
+
+	user, err := client.Whoami(cmd.Context())
+	if err != nil {
+		var aErr api.AuthorizationError
+		if errors.As(err, &aErr) && aErr.StatusCode == http.StatusUnauthorized {
+			fmt.Println("You need to be signed in to Ollama to run Cloud models.")
+			fmt.Println()
+
+			if aErr.SigninURL != "" {
+				fmt.Printf(ConnectInstructions, aErr.SigninURL)
+			}
+			return nil
+		}
+		return err
+	}
+
+	if user != nil && user.Name != "" {
+		fmt.Printf("You are already signed in as user '%s'\n", user.Name)
+		fmt.Println()
+		return nil
+	}
+
+	return nil
+}
+
+func SignoutHandler(cmd *cobra.Command, args []string) error {
+	client, err := api.ClientFromEnvironment()
+	if err != nil {
+		return err
+	}
+
+	err = client.Signout(cmd.Context())
+	if err != nil {
+		var aErr api.AuthorizationError
+		if errors.As(err, &aErr) && aErr.StatusCode == http.StatusUnauthorized {
+			fmt.Println("You are not signed in to ollama.com")
+			fmt.Println()
+			return nil
+		} else {
+			return err
+		}
+	}
+
+	fmt.Println("You have signed out of ollama.com")
+	fmt.Println()
+	return nil
+}
+
 func PushHandler(cmd *cobra.Command, args []string) error {
 	client, err := api.ClientFromEnvironment()
 	if err != nil {
@@ -466,6 +547,25 @@ func PushHandler(cmd *cobra.Command, args []string) error {
 		return err
 	}

+	n := model.ParseName(args[0])
+	if strings.HasSuffix(n.Host, ".ollama.ai") || strings.HasSuffix(n.Host, ".ollama.com") {
+		_, err := client.Whoami(cmd.Context())
+		if err != nil {
+			var aErr api.AuthorizationError
+			if errors.As(err, &aErr) && aErr.StatusCode == http.StatusUnauthorized {
+				fmt.Println("You need to be signed in to push models to ollama.com.")
+				fmt.Println()
+
+				if aErr.SigninURL != "" {
+					fmt.Printf(ConnectInstructions, aErr.SigninURL)
+				}
+				return nil
+			}
+
+			return err
+		}
+	}
+
 	p := progress.NewProgress(os.Stderr)
 	defer p.Stop()

@@ -502,12 +602,12 @@ func PushHandler(cmd *cobra.Command, args []string) error {

 	request := api.PushRequest{Name: args[0], Insecure: insecure}

-	n := model.ParseName(args[0])
 	if err := client.Push(cmd.Context(), &request, fn); err != nil {
 		if spinner != nil {
 			spinner.Stop()
 		}
-		if strings.Contains(err.Error(), "access denied") {
+		errStr := strings.ToLower(err.Error())
+		if strings.Contains(errStr, "access denied") || strings.Contains(errStr, "unauthorized") {
 			return errors.New("you are not authorized to push to this namespace, create the model under a namespace you own")
 		}
 		return err
@@ -541,7 +641,14 @@ func ListHandler(cmd *cobra.Command, args []string) error {

 	for _, m := range models.Models {
 		if len(args) == 0 || strings.HasPrefix(strings.ToLower(m.Name), strings.ToLower(args[0])) {
-			data = append(data, []string{m.Name, m.Digest[:12], format.HumanBytes(m.Size), format.HumanTime(m.ModifiedAt, "Never")})
+			var size string
+			if m.RemoteModel != "" {
+				size = "-"
+			} else {
+				size = format.HumanBytes(m.Size)
+			}
+
+			data = append(data, []string{m.Name, m.Digest[:12], size, format.HumanTime(m.ModifiedAt, "Never")})
 		}
 	}

@@ -620,23 +727,21 @@ func DeleteHandler(cmd *cobra.Command, args []string) error {
 		return err
 	}

-	// Unload the model if it's running before deletion
-	opts := &runOptions{
-		Model:     args[0],
-		KeepAlive: &api.Duration{Duration: 0},
-	}
-	if err := loadOrUnloadModel(cmd, opts); err != nil {
-		if !strings.Contains(err.Error(), "not found") {
-			return fmt.Errorf("unable to stop existing running model \"%s\": %s", args[0], err)
+	for _, arg := range args {
+		// Unload the model if it's running before deletion
+		if err := loadOrUnloadModel(cmd, &runOptions{
+			Model:     args[0],
+			KeepAlive: &api.Duration{Duration: 0},
+		}); err != nil {
+			if !strings.Contains(strings.ToLower(err.Error()), "not found") {
+				fmt.Fprintf(os.Stderr, "Warning: unable to stop model '%s'\n", args[0])
+			}
 		}
-	}

-	for _, name := range args {
-		req := api.DeleteRequest{Name: name}
-		if err := client.Delete(cmd.Context(), &req); err != nil {
+		if err := client.Delete(cmd.Context(), &api.DeleteRequest{Name: arg}); err != nil {
 			return err
 		}
-		fmt.Printf("deleted '%s'\n", name)
+		fmt.Printf("deleted '%s'\n", arg)
 	}
 	return nil
 }
@@ -738,12 +843,36 @@ func showInfo(resp *api.ShowResponse, verbose bool, w io.Writer) error {
 	}

 	tableRender("Model", func() (rows [][]string) {
+		if resp.RemoteHost != "" {
+			rows = append(rows, []string{"", "Remote model", resp.RemoteModel})
+			rows = append(rows, []string{"", "Remote URL", resp.RemoteHost})
+		}
+
 		if resp.ModelInfo != nil {
 			arch := resp.ModelInfo["general.architecture"].(string)
 			rows = append(rows, []string{"", "architecture", arch})
-			rows = append(rows, []string{"", "parameters", format.HumanNumber(uint64(resp.ModelInfo["general.parameter_count"].(float64)))})
-			rows = append(rows, []string{"", "context length", strconv.FormatFloat(resp.ModelInfo[fmt.Sprintf("%s.context_length", arch)].(float64), 'f', -1, 64)})
-			rows = append(rows, []string{"", "embedding length", strconv.FormatFloat(resp.ModelInfo[fmt.Sprintf("%s.embedding_length", arch)].(float64), 'f', -1, 64)})
+
+			var paramStr string
+			if resp.Details.ParameterSize != "" {
+				paramStr = resp.Details.ParameterSize
+			} else if v, ok := resp.ModelInfo["general.parameter_count"]; ok {
+				if f, ok := v.(float64); ok {
+					paramStr = format.HumanNumber(uint64(f))
+				}
+			}
+			rows = append(rows, []string{"", "parameters", paramStr})
+
+			if v, ok := resp.ModelInfo[fmt.Sprintf("%s.context_length", arch)]; ok {
+				if f, ok := v.(float64); ok {
+					rows = append(rows, []string{"", "context length", strconv.FormatFloat(f, 'f', -1, 64)})
+				}
+			}
+
+			if v, ok := resp.ModelInfo[fmt.Sprintf("%s.embedding_length", arch)]; ok {
+				if f, ok := v.(float64); ok {
+					rows = append(rows, []string{"", "embedding length", strconv.FormatFloat(f, 'f', -1, 64)})
+				}
+			}
 		} else {
 			rows = append(rows, []string{"", "architecture", resp.Details.Family})
 			rows = append(rows, []string{"", "parameters", resp.Details.ParameterSize})
@@ -991,6 +1120,52 @@ type runOptions struct {
 	KeepAlive    *api.Duration
 	Think        *api.ThinkValue
 	HideThinking bool
+	ShowConnect  bool
+}
+
+func (r runOptions) Copy() runOptions {
+	var messages []api.Message
+	if r.Messages != nil {
+		messages = make([]api.Message, len(r.Messages))
+		copy(messages, r.Messages)
+	}
+
+	var images []api.ImageData
+	if r.Images != nil {
+		images = make([]api.ImageData, len(r.Images))
+		copy(images, r.Images)
+	}
+
+	var opts map[string]any
+	if r.Options != nil {
+		opts = make(map[string]any, len(r.Options))
+		for k, v := range r.Options {
+			opts[k] = v
+		}
+	}
+
+	var think *api.ThinkValue
+	if r.Think != nil {
+		cThink := *r.Think
+		think = &cThink
+	}
+
+	return runOptions{
+		Model:        r.Model,
+		ParentModel:  r.ParentModel,
+		Prompt:       r.Prompt,
+		Messages:     messages,
+		WordWrap:     r.WordWrap,
+		Format:       r.Format,
+		System:       r.System,
+		Images:       images,
+		Options:      opts,
+		MultiModal:   r.MultiModal,
+		KeepAlive:    r.KeepAlive,
+		Think:        think,
+		HideThinking: r.HideThinking,
+		ShowConnect:  r.ShowConnect,
+	}
 }

 type displayResponseState struct {
@@ -1546,6 +1721,22 @@ func NewCLI() *cobra.Command {

 	pushCmd.Flags().Bool("insecure", false, "Use an insecure registry")

+	signinCmd := &cobra.Command{
+		Use:     "signin",
+		Short:   "Sign in to ollama.com",
+		Args:    cobra.ExactArgs(0),
+		PreRunE: checkServerHeartbeat,
+		RunE:    SigninHandler,
+	}
+
+	signoutCmd := &cobra.Command{
+		Use:     "signout",
+		Short:   "Sign out from ollama.com",
+		Args:    cobra.ExactArgs(0),
+		PreRunE: checkServerHeartbeat,
+		RunE:    SignoutHandler,
+	}
+
 	listCmd := &cobra.Command{
 		Use:     "list",
 		Aliases: []string{"ls"},
@@ -1612,6 +1803,7 @@ func NewCLI() *cobra.Command {
 			appendEnvDocs(cmd, []envconfig.EnvVar{
 				envVars["OLLAMA_DEBUG"],
 				envVars["OLLAMA_HOST"],
+				envVars["OLLAMA_CONTEXT_LENGTH"],
 				envVars["OLLAMA_KEEP_ALIVE"],
 				envVars["OLLAMA_MAX_LOADED_MODELS"],
 				envVars["OLLAMA_MAX_QUEUE"],
@@ -1639,6 +1831,8 @@ func NewCLI() *cobra.Command {
 		stopCmd,
 		pullCmd,
 		pushCmd,
+		signinCmd,
+		signoutCmd,
 		listCmd,
 		psCmd,
 		copyCmd,