Merge pull request #270 from unslothai/fix/gguf-export-relocation

Fix GGUF exports saving to wrong directory and missing from chat model selector

studio/backend/core/export/export.py

@@ -2,9 +2,11 @@
 """
 Export backend - handles model exporting in various formats
 """
+import glob
 import json
 import logging
 import os
+import shutil
 from pathlib import Path
 from typing import Optional, Tuple, List
 from peft import PeftModel, PeftModelForCausalLM
@@ -409,9 +411,15 @@ class ExportBackend:
                 # On WSL, patch out sudo check before llama.cpp build
                 _apply_wsl_sudo_patch()
 
+                # Snapshot existing .gguf files in cwd before conversion.
+                # unsloth's convert_to_gguf writes output files relative to
+                # cwd (repo root), so we diff afterwards and relocate them.
+                cwd = os.getcwd()
+                pre_existing_ggufs = set(glob.glob(os.path.join(cwd, "*.gguf")))
+
                 # Pass absolute path — no os.chdir needed.
-                # unsloth saves model files into this directory, while
-                # check_llama_cpp("llama.cpp") resolves against cwd (repo root)
+                # unsloth saves intermediate HF model files into model_save_path,
+                # while check_llama_cpp("llama.cpp") resolves against cwd (repo root)
                 # where setup.sh already built llama.cpp with quantizer.
                 model_save_path = os.path.join(abs_save_dir, "model")
                 self.current_model.save_pretrained_gguf(
@@ -420,6 +428,32 @@ class ExportBackend:
                     quantization_method=quant_method
                 )
 
+                # Relocate GGUF artifacts into the export directory.
+                # convert_to_gguf writes .gguf files to cwd (repo root)
+                # because --outfile is a relative path like "model.Q4_K_M.gguf".
+                new_ggufs = set(glob.glob(os.path.join(cwd, "*.gguf"))) - pre_existing_ggufs
+                for src in sorted(new_ggufs):
+                    dest = os.path.join(abs_save_dir, os.path.basename(src))
+                    shutil.move(src, dest)
+                    logger.info(f"Relocated GGUF: {os.path.basename(src)} → {abs_save_dir}/")
+
+                # Flatten any .gguf files from subdirectories into abs_save_dir.
+                # save_pretrained_gguf may create subdirs (e.g. model_gguf/)
+                # with a name different from model_save_path.
+                for sub in list(Path(abs_save_dir).iterdir()):
+                    if not sub.is_dir():
+                        continue
+                    for src in sub.glob("*.gguf"):
+                        dest = os.path.join(abs_save_dir, src.name)
+                        shutil.move(str(src), dest)
+                        logger.info(f"Relocated GGUF: {src.name} → {abs_save_dir}/")
+                    # Clean up the subdirectory (intermediate HF files, etc.)
+                    shutil.rmtree(str(sub), ignore_errors=True)
+                    logger.info(f"Cleaned up subdirectory: {sub.name}")
+
+                # Write export metadata so the Chat page can identify the base model
+                self._write_export_metadata(abs_save_dir)
+
                 logger.info(f"GGUF model saved successfully in {abs_save_dir}")
 
             # Push to hub if requested

studio/backend/models/models.py

@@ -63,7 +63,7 @@ class LoRAInfo(BaseModel):
     adapter_path: str = Field(..., description="Path to the LoRA adapter or exported model")
     base_model: Optional[str] = Field(None, description="Base model identifier")
     source: Optional[str] = Field(None, description="'training' or 'exported'")
-    export_type: Optional[str] = Field(None, description="'lora' or 'merged' (for exports)")
+    export_type: Optional[str] = Field(None, description="'lora', 'merged', or 'gguf' (for exports)")
 
 
 class LoRAScanResponse(BaseModel):

studio/backend/utils/models/model_config.py

@@ -640,14 +640,15 @@ def scan_trained_loras(outputs_dir: str = "./outputs") -> List[Tuple[str, str]]:
 
 def scan_exported_models(exports_dir: str = "./exports") -> List[Tuple[str, str, str, Optional[str]]]:
     """
-    Scan exports folder for exported models (merged, LoRA, base).
-    Skips GGUF-only exports (not loadable by Unsloth inference backend).
+    Scan exports folder for exported models (merged, LoRA, GGUF).
 
-    The exports directory is two levels deep: {run}/{checkpoint}/
+    Supports two directory layouts:
+      - Two-level: {run}/{checkpoint}/  (merged & LoRA exports)
+      - Flat:      {name}-finetune-gguf/  (GGUF exports)
 
     Returns:
         List of tuples: [(display_name, model_path, export_type, base_model), ...]
-        export_type: "lora" | "merged"
+        export_type: "lora" | "merged" | "gguf"
     """
     results = []
     exports_path = Path(exports_dir)
@@ -659,6 +660,26 @@ def scan_exported_models(exports_dir: str = "./exports") -> List[Tuple[str, str,
         for run_dir in exports_path.iterdir():
             if not run_dir.is_dir():
                 continue
+
+            # Check for flat GGUF export (e.g. exports/gemma-3-4b-it-finetune-gguf/)
+            gguf_files = list(run_dir.glob("*.gguf"))
+            if gguf_files:
+                base_model = None
+                export_meta = run_dir / "export_metadata.json"
+                try:
+                    if export_meta.exists():
+                        meta = json.loads(export_meta.read_text())
+                        base_model = meta.get("base_model")
+                except Exception:
+                    pass
+
+                display_name = run_dir.name
+                model_path = str(gguf_files[0])  # path to the .gguf file
+                results.append((display_name, model_path, "gguf", base_model))
+                logger.debug(f"Found GGUF export: {display_name}")
+                continue
+
+            # Two-level: {run}/{checkpoint}/
             for checkpoint_dir in run_dir.iterdir():
                 if not checkpoint_dir.is_dir():
                     continue
@@ -683,7 +704,6 @@ def scan_exported_models(exports_dir: str = "./exports") -> List[Tuple[str, str,
                         pass
                 elif config_file.exists() and has_weights:
                     export_type = "merged"
-                    # Read base model from export_metadata.json (written at export time)
                     export_meta = checkpoint_dir / "export_metadata.json"
                     try:
                         if export_meta.exists():
@@ -692,7 +712,25 @@ def scan_exported_models(exports_dir: str = "./exports") -> List[Tuple[str, str,
                     except Exception:
                         pass
                 elif has_gguf:
-                    # GGUF-only — not loadable by current inference backend
+                    export_type = "gguf"
+                    gguf_list = list(checkpoint_dir.glob("*.gguf"))
+                    # Check checkpoint_dir first, then fall back to parent run_dir
+                    # (export.py writes metadata to the top-level export directory)
+                    for meta_dir in (checkpoint_dir, run_dir):
+                        export_meta = meta_dir / "export_metadata.json"
+                        try:
+                            if export_meta.exists():
+                                meta = json.loads(export_meta.read_text())
+                                base_model = meta.get("base_model")
+                                if base_model:
+                                    break
+                        except Exception:
+                            pass
+
+                    display_name = f"{run_dir.name} / {checkpoint_dir.name}"
+                    model_path = str(gguf_list[0]) if gguf_list else str(checkpoint_dir)
+                    results.append((display_name, model_path, export_type, base_model))
+                    logger.debug(f"Found GGUF export: {display_name}")
                     continue
                 else:
                     continue

studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx

@@ -525,9 +525,12 @@ export function LoraModelPicker({
                 {adapters.map((adapter) => {
                   const isExported = adapter.source === "exported";
                   const isMerged = adapter.exportType === "merged";
-                  const tag = isExported
-                    ? isMerged ? "Merged" : "LoRA"
-                    : "LoRA";
+                  const isGguf = adapter.exportType === "gguf";
+                  const tag = isGguf
+                    ? "GGUF"
+                    : isExported
+                      ? isMerged ? "Merged" : "LoRA"
+                      : "LoRA";
                   const meta = isExported ? `${tag} · Exported` : tag;
                   return (
                     <ModelRow
@@ -537,7 +540,7 @@ export function LoraModelPicker({
                       selected={value === adapter.id}
                       onClick={() => onSelect(adapter.id, {
                         source: isExported ? "exported" : "lora",
-                        isLora: !isMerged,
+                        isLora: !isMerged && !isGguf,
                       })}
                     />
                   );

studio/frontend/src/components/assistant-ui/model-selector/types.ts

@@ -11,7 +11,7 @@ export interface LoraModelOption extends ModelOption {
   baseModel?: string;
   updatedAt?: number;
   source?: "training" | "exported";
-  exportType?: "lora" | "merged";
+  exportType?: "lora" | "merged" | "gguf";
 }
 
 export interface ModelSelectorChangeMeta {

studio/frontend/src/features/chat/hooks/use-chat-model-runtime.ts

@@ -75,7 +75,7 @@ function toLoraSummary(lora: {
   adapter_path: string;
   base_model?: string | null;
   source?: "training" | "exported" | null;
-  export_type?: "lora" | "merged" | null;
+  export_type?: "lora" | "merged" | "gguf" | null;
 }): ChatLoraSummary {
   const idTail = lora.adapter_path.split("/").filter(Boolean).at(-1) ?? "";
   const updatedAt =

studio/frontend/src/features/chat/types/api.ts

@@ -16,7 +16,7 @@ export interface BackendLoraInfo {
   adapter_path: string;
   base_model?: string | null;
   source?: "training" | "exported" | null;
-  export_type?: "lora" | "merged" | null;
+  export_type?: "lora" | "merged" | "gguf" | null;
 }
 
 export interface ListLorasResponse {

studio/frontend/src/features/chat/types/runtime.ts

@@ -35,5 +35,5 @@ export interface ChatLoraSummary {
   baseModel: string;
   updatedAt?: number;
   source?: "training" | "exported";
-  exportType?: "lora" | "merged";
+  exportType?: "lora" | "merged" | "gguf";
 }

studio/frontend/src/features/export/export-page.tsx

@@ -155,7 +155,12 @@ export function ExportPage() {
     setExportError(null);
     setExportSuccess(false);
 
-    const saveDir = `./exports/${selectedModelIdx ?? "model"}/${checkpoint}`;
+    // For GGUF, use a flat folder like "exports/gemma-3-4b-it-finetune-gguf"
+    // For other formats, nest under training-run/checkpoint
+    const saveDir =
+      exportMethod === "gguf"
+        ? `./exports/${(baseModelName.split("/").pop() ?? selectedModelIdx ?? "model")}-finetune-gguf`
+        : `./exports/${selectedModelIdx ?? "model"}/${checkpoint}`;
     const pushToHub = destination === "hub";
     const repoId = pushToHub && hfUsername && modelName
       ? `${hfUsername}/${modelName}`