feat: write llama.ini for llama-server --models-preset llama.ini

Author: 0uep

.dockerignore

@@ -1,7 +1,9 @@
 *
 !conf
+!event
 !gie
 !go.mod
 !go.sum
 !goinfer.go
 !infer
+!proxy

.gitignore

@@ -8,13 +8,13 @@
 /.continue/
 /.old-code/
 /.roo/
+/*.ini*
+/*.log*
+/*.yml*
 /docsite/dist/
 /go.work*
 /goinfer
-/goinfer.ini*
 /infer/dist/
 /llama-server.exe
-/llama-swap.yml
 /template.jinja
 dist/
-llama.log

conf/conf.go

@@ -50,11 +50,6 @@ const (
 	// Hex code: C0ffee 15 C001, 50 C0ffee 15 900d. Bad C0de 15 Dead, 101. Cafe 91f7, 90 Cafe, 7e57 C0de.
 	debugAPIKey = "C0ffee15C00150C0ffee15900dBadC0de15Dead101Cafe91f790Cafe7e57C0de"
 	unsetAPIKey = "Please ⚠️ Set your private 64-hex-digit API key (32 bytes)"
-
-	// GoinferINI is the config filename.
-	GoinferINI = "goinfer.ini"
-	// LlamaSwapYML is the llama-swap config filename.
-	LlamaSwapYML = "llama-swap.yml"
 )
 
 // Do not use the bad ports: they are blocked by web browsers,

conf/conf_test.go

@@ -84,7 +84,7 @@ func TestWriteMainCfg(t *testing.T) {
 	t.Setenv("GI_MODELS_DIR", modelsDir)
 	t.Setenv("GI_LLAMA_EXE", llamaExe)
 
-	data, err := cfg.GenFileData(false, true)
+	data, err := cfg.GenGoinferINI(false, true)
 	if err != nil {
 		t.Fatalf("WriteMainCfg failed: %v", err)
 	}
@@ -107,7 +107,7 @@ func TestWriteSwapCfg(t *testing.T) {
 	}
 	cfg.ModelsDir = modelsDir
 
-	ymlData, err := cfg.GenSwapYAMLData(false, false)
+	ymlData, err := cfg.GenLlamaSwapYAML(false, false)
 	if err != nil {
 		t.Fatalf("WriteSwapCfg failed: %v ymlData=%s", err, string(ymlData))
 	}

conf/files.go

@@ -179,7 +179,7 @@ func nameWithDir(root, truncated, name string) string {
 }
 
 // extractFlags returns the truncated path and the llama-server flags from a file path.
-// It first checks for a companion ".args" file; if present, its contents are used as flags.
+// It first checks for a companion ".sh" file; if present, its contents are used as flags.
 // Otherwise, it parses flags encoded in the filename after an '&' delimiter.
 // Returns the truncated path (without extension) and a space-separated flag string.
 //
@@ -194,7 +194,7 @@ func extractFlags(path string) (truncated, flags_ string) {
 	}
 
 	// 1. Is there a file containing the command line arguments?
-	argsFn := filepath.Clean(truncated + ".args")
+	argsFn := filepath.Clean(truncated + ".sh")
 	args, err := os.ReadFile(argsFn)
 	if err == nil {
 		flags := oneLine(args)
@@ -226,7 +226,7 @@ func extractFlags(path string) (truncated, flags_ string) {
 	return truncated[:pos], strings.Join(flags, " ")
 }
 
-// oneLine converts the `.args` file into a single space-separated string,
+// oneLine converts the `.sh` file into a single space-separated string,
 // removing trailing backslashes, trimming whitespace, ignoring empty lines or comments.
 func oneLine(input []byte) string {
 	keep := make([]byte, 0, len(input))
@@ -238,13 +238,15 @@ func oneLine(input []byte) string {
 		}
 		// Remove leading/trailing whitespace
 		line = bytes.TrimSpace(line)
-		// Skip blank lines and comments
-		if len(line) == 0 || bytes.HasPrefix(line, []byte("#")) {
+		// Skip blank lines
+		if len(line) == 0 {
 			continue
 		}
 		// Convert the byte slice to a string before appending.
-		keep = append(keep, line...)
-		keep = append(keep, ' ')
+		if line[0] == '-' {
+			keep = append(keep, line...)
+			keep = append(keep, ' ')
+		}
 	}
 
 	return string(keep)

conf/files_test.go

@@ -85,9 +85,9 @@ func TestExtractFlags(t *testing.T) {
 	t.Parallel()
 	tmp := t.TempDir()
 
-	// .args file present
+	// .sh file present
 	modelPath := createGGUFFile(t, tmp, "model1.gguf", 2048)
-	argsPath := strings.TrimSuffix(modelPath, ".gguf") + ".args"
+	argsPath := strings.TrimSuffix(modelPath, ".gguf") + ".sh"
 	err := os.WriteFile(argsPath, []byte("-foo bar -baz qux"), 0o600)
 	if err != nil {
 		t.Fatalf("failed to write args file: %v", err)

conf/read.go conf/goinferini.goconf/read.go renamed to conf/goinferini.go

@@ -5,18 +5,22 @@
 package conf
 
 import (
+	"crypto/rand"
+	"encoding/hex"
 	"errors"
-	"io"
 	"log/slog"
 	"os"
+	"path/filepath"
 	"strings"
 	"syscall"
 
 	"github.com/LM4eu/goinfer/gie"
-	"github.com/LM4eu/goinfer/proxy/config"
 	"github.com/pelletier/go-toml/v2"
 )
 
+// GoinferINI is the config filename.
+const GoinferINI = "goinfer.ini"
+
 // ReadGoinferINI loads the configuration file, reads the env vars and verifies the settings.
 // Always return a valid configuration, because the receiver may want to write a valid config.
 func ReadGoinferINI(noAPIKey bool, extra, start string) (*Cfg, error) {
@@ -80,15 +84,39 @@ func ReadFileData(data []byte, noAPIKey bool, extra, start string) (*Cfg, error)
 	return cfg, err
 }
 
-// ReadSwapFromReader uses the LoadConfigFromReader() from llama-swap project.
-func (cfg *Cfg) ReadSwapFromReader(r io.Reader) error {
-	var err error
-	cfg.Swap, err = config.LoadConfigFromReader(r)
-	if err != nil {
-		slog.Error("Cannot load llama-swap config", "file", LlamaSwapYML, "error", err)
-		os.Exit(1)
+// WriteGoinferINI populates the configuration with defaults, applies environment variables,
+// writes the resulting configuration to the given file.
+func (cfg *Cfg) WriteGoinferINI(debug, noAPIKey bool) error {
+	data, err := cfg.GenGoinferINI(debug, noAPIKey)
+	er := writeWithHeader(GoinferINI, "# Configuration of https://github.com/LM4eu/goinfer\n\n", data)
+	if er != nil {
+		if err != nil {
+			return errors.Join(err, er)
+		}
+		return er
+	}
+	return err
+}
+
+// GenGoinferINI sets the API keys, reads the environment variables,
+// fix some settings and writes the result config to a buffer.
+func (cfg *Cfg) GenGoinferINI(debug, noAPIKey bool) ([]byte, error) {
+	cfg.setAPIKey(debug, noAPIKey)
+	cfg.applyEnvVars()
+	cfg.trimParamValues()
+	cfg.fixDefaultModel()
+
+	err := cfg.validate(noAPIKey)
+
+	data, er := toml.Marshal(&cfg)
+	if er != nil {
+		er = gie.Wrap(err, gie.ConfigErr, "failed to yaml.Marshal", "cfg", cfg)
+		if err != nil {
+			return data, errors.Join(err, er)
+		}
+		return data, er
 	}
-	return cfg.ValidateSwap()
+	return data, err
 }
 
 // load the configuration file (if filename not empty).
@@ -190,3 +218,55 @@ func (cfg *Cfg) trimParamValues() {
 	cfg.Llama.Common = strings.TrimSpace(cfg.Llama.Common)
 	cfg.Llama.Goinfer = strings.TrimSpace(cfg.Llama.Goinfer)
 }
+
+func writeWithHeader(path, header string, data []byte) error {
+	path = filepath.Clean(path)
+	file, err := os.Create(path)
+	if err != nil {
+		return gie.Wrap(err, gie.ConfigErr, "failed to create file="+path)
+	}
+
+	_, err = file.WriteString(header)
+	if err == nil {
+		_, err = file.Write(data)
+	}
+
+	er := file.Close()
+	if err != nil {
+		err = er
+	}
+	if err != nil {
+		return gie.Wrap(err, gie.ConfigErr, "failed to write file="+path)
+	}
+
+	return nil
+}
+
+func (cfg *Cfg) setAPIKey(debug, noAPIKey bool) {
+	switch {
+	case noAPIKey:
+		cfg.APIKey = unsetAPIKey
+		slog.Info("Flag -no-api-key => Do not generate API key")
+
+	case debug:
+		cfg.APIKey = debugAPIKey
+		slog.Warn("API key is DEBUG => security threat")
+
+	default:
+		cfg.APIKey = gen64HexDigits()
+		slog.Info("Generated random API key")
+	}
+}
+
+func gen64HexDigits() string {
+	buf := make([]byte, 32)
+	_, err := rand.Read(buf)
+	if err != nil {
+		slog.Warn("Failed to rand.Read", "error", err)
+		return ""
+	}
+
+	key := make([]byte, 64)
+	hex.Encode(key, buf)
+	return string(key)
+}

conf/llamacppini.go

@@ -0,0 +1,91 @@
+// Copyright 2025 The contributors of Goinfer.
+// This file is part of Goinfer, a LLM proxy under the MIT License.
+// SPDX-License-Identifier: MIT
+
+// Package conf reads/writes configuration
+package conf
+
+import (
+	"bytes"
+	"strings"
+)
+
+// LlamaINI is the llama.cpp config filename.
+const LlamaINI = "llama.ini"
+
+// WriteLlamaINI generates the llama.cpp configuration.
+func WriteLlamaINI(yml []byte) error {
+	header := `# DO NOT EDIT - This file is generated by Goinfer.
+#
+# llama.cpp configurations using Model Presets:
+#
+#     llama-server --models-preset ./llama.ini
+#
+# Each section in this file defines a new preset.
+# Keys within a section correspond to command-line arguments (without leading dashes).
+# For example, the argument --n-gpu-layer 123 is written as n-gpu-layer = 123.
+# Short argument forms (e.g., c, ngl) and environment variable names (e.g., LLAMA_ARG_N_GPU_LAYERS) are also supported as keys.
+#
+# Doc: https://github.com/ggml-org/llama.cpp/blob/master/tools/server/README.md#model-presets
+
+version = 1
+`
+	return writeWithHeader(LlamaINI, header, yml)
+}
+
+// GenSwapYAMLData generates the llama-swap configuration.
+func (cfg *Cfg) GenLlamaINI() []byte {
+	var out bytes.Buffer
+
+	// For each model, set two model settings:
+	// 1. for the OpenAI endpoints
+	// 2. for the /completion endpoint (prefix with A_ and hide the model)
+	for model, mi := range cfg.getInfo() {
+		out.Write(genModel(model, mi.Path, mi.Flags))
+		out.Write(genModel(model+":A", mi.Path, cfg.Llama.Goinfer+" "+mi.Flags))
+	}
+
+	return out.Bytes()
+}
+
+// Add the model settings within the llama-swap configuration.
+func genModel(name, path, flags string) []byte {
+	out := bytes.NewBufferString(`
+[` + name + `]
+model = ` + path)
+
+	boolean := false
+
+	for arg := range strings.FieldsSeq(flags) {
+		if arg[0] == '-' {
+			if boolean {
+				out.WriteString(" true")
+			}
+			boolean = true
+			out.WriteByte('\n')
+			if len(arg) == 2 {
+				out.WriteByte(arg[1])
+			} else {
+				i := 1
+				if arg[1] == '-' {
+					i = 2
+				}
+				out.WriteString(arg[i:])
+			}
+			out.WriteByte(' ')
+			out.WriteByte('=')
+			continue
+		}
+		boolean = false
+		out.WriteByte(' ')
+		out.WriteString(arg)
+	}
+
+	if boolean {
+		out.WriteString(" true\n")
+	} else {
+		out.WriteByte('\n')
+	}
+
+	return out.Bytes()
+}