refactor(convert): route SafeTensors conversion through Python fallback

Use the official llama.cpp Python converter as the single conversion path so runtime behavior is consistent across architectures and users get clear dependency installation errors when Python env is missing.

Made-with: Cursor

Author: ganisback

internal/convert/convert.go

@@ -2,7 +2,6 @@ package convert
 
 import (
 	"encoding/binary"
-	"fmt"
 	"math"
 	"os"
 	"path/filepath"
@@ -13,87 +12,14 @@ import (
 type ProgressFunc func(step string, current, total int)
 
 // Convert converts SafeTensors model files in modelDir to a GGUF file.
-// Each architecture has a dedicated Go converter. For architectures without
-// a Go converter, it falls back to the official llama.cpp convert_hf_to_gguf.py.
+// Conversion is delegated to the official llama.cpp convert_hf_to_gguf.py
+// so users only need one consistent conversion path.
 func Convert(modelDir string, progress ProgressFunc) (string, error) {
 	if progress == nil {
 		progress = func(string, int, int) {}
 	}
 
-	progress("Reading model configuration", 0, 0)
-
-	cfg, err := loadModelConfig(modelDir)
-	if err != nil {
-		return "", fmt.Errorf("loading config: %w", err)
-	}
-
-	hfArch := cfg.Architectures[0]
-	ggufArch, ok := detectGGUFArch(hfArch)
-	if !ok {
-		return ConvertPython(modelDir, progress)
-	}
-
-	converter := getConverter(ggufArch, cfg)
-	if converter == nil {
-		return ConvertPython(modelDir, progress)
-	}
-
-	progress("Scanning SafeTensors files", 0, 0)
-
-	stFiles, err := scanSafeTensorsFiles(modelDir)
-	if err != nil {
-		return "", fmt.Errorf("scanning SafeTensors: %w", err)
-	}
-
-	sources := collectTensors(stFiles)
-
-	progress("Parsing tokenizer", 0, 0)
-
-	tok, err := parseTokenizer(modelDir, hfArch)
-	if err != nil {
-		return "", fmt.Errorf("parsing tokenizer: %w", err)
-	}
-
-	if cfg.VocabSize > len(tok.Tokens) {
-		for i := len(tok.Tokens); i < cfg.VocabSize; i++ {
-			tok.Tokens = append(tok.Tokens, fmt.Sprintf("[PAD%d]", i))
-			tok.Scores = append(tok.Scores, -1)
-			tok.Types = append(tok.Types, tokenTypeUserDefined)
-		}
-	}
-
-	progress("Building GGUF", 0, 0)
-
-	writer := newGGUFWriter()
-	converter.WriteKV(writer, cfg)
-	writeTokenizerKV(writer, tok, cfg)
-
-	if err := converter.ConvertTensors(writer, sources, cfg, progress); err != nil {
-		return "", fmt.Errorf("converting tensors: %w", err)
-	}
-
-	idx := findKVIndex(writer.kvs, "general.file_type")
-	if idx >= 0 {
-		writer.kvs[idx].value = uint32(1) // GGML_FTYPE_MOSTLY_F16
-	}
-
-	outputName := generateOutputName(modelDir, cfg)
-	outputPath := filepath.Join(modelDir, outputName)
-	tmpPath := outputPath + ".tmp"
-
-	progress("Writing GGUF file", 0, 0)
-
-	if err := writer.writeTo(tmpPath); err != nil {
-		os.Remove(tmpPath)
-		return "", fmt.Errorf("writing GGUF: %w", err)
-	}
-
-	if err := os.Rename(tmpPath, outputPath); err != nil {
-		os.Remove(tmpPath)
-		return "", fmt.Errorf("finalizing GGUF: %w", err)
-	}
-
-	return outputPath, nil
+	return ConvertPython(modelDir, progress)
 }
 
 // HasGGUF checks if a GGUF file already exists in the model directory.