fix: adapter saved to clean dir, LM dropdown no 'Default', on-demand download

app.py

@@ -179,14 +179,27 @@ def _run_pipeline(caption, lyrics, bpm, duration, seed, steps, output_format,
 # LM model scanning & on-demand download
 # ---------------------------------------------------------------------------
 
+AVAILABLE_LM_MODELS = [
+    "acestep-5Hz-lm-0.6B-Q8_0.gguf",
+    "acestep-5Hz-lm-1.7B-Q8_0.gguf",
+    "acestep-5Hz-lm-4B-Q5_K_M.gguf",
+]
+
 def _scan_lm_models():
-    """Scan /app/models for *-lm-*.gguf files, return list of filenames."""
-    models = []
+    """Return all available LM models (installed get a checkmark)."""
+    installed = set()
     if os.path.isdir(MODELS_DIR):
-        for f in sorted(os.listdir(MODELS_DIR)):
+        for f in os.listdir(MODELS_DIR):
             if "-lm-" in f and f.endswith(".gguf"):
-                models.append(f)
-    return models
+                installed.add(f)
+    choices = []
+    for m in AVAILABLE_LM_MODELS:
+        label = m.replace(".gguf", "")
+        if m in installed:
+            choices.append(m)
+        else:
+            choices.append(f"{m} (download)")
+    return choices
 
 
 def _download_lm_model(filename):
@@ -372,7 +385,10 @@ def gradio_main():
 
         actual_seed = None if seed is None or int(seed) < 0 else int(seed)
         adapter = None if lora_select == "None (no LoRA)" else lora_select
-        lm_model = None if not lm_model_select or lm_model_select == "Default" else lm_model_select
+        lm_model_file = lm_model_select.replace(" (download)", "") if lm_model_select else None
+        if lm_model_file and "(download)" in (lm_model_select or ""):
+            _download_lm_model(lm_model_file)
+        lm_model = lm_model_file
 
         progress_map = {
             "lm_submit": (0.05, "Submitting LM job..."),
@@ -450,10 +466,12 @@ def gradio_main():
         lr = float(lr)
         rank = max(1, min(int(rank), 64))
 
-        output_dir = os.path.join(ADAPTER_DIR, lora_name)
-        os.makedirs(output_dir, exist_ok=True)
-        audio_dir = os.path.join(output_dir, "audio_input")
+        work_dir = os.path.join(OUTPUT_DIR, "train_workspace", lora_name)
+        os.makedirs(work_dir, exist_ok=True)
+        audio_dir = os.path.join(work_dir, "audio_input")
         os.makedirs(audio_dir, exist_ok=True)
+        adapter_out = os.path.join(ADAPTER_DIR, lora_name)
+        os.makedirs(adapter_out, exist_ok=True)
 
         # Copy uploaded audio files
         _log(f"[INFO] Preparing {len(audio_files)} audio files...")
@@ -478,7 +496,7 @@ def gradio_main():
             _log("[Step 1/2] Preprocessing audio...")
             yield _log_text(), gr.update(visible=False), gr.update(visible=True)
 
-            preprocessed_dir = os.path.join(output_dir, "preprocessed_tensors")
+            preprocessed_dir = os.path.join(work_dir, "preprocessed_tensors")
 
             def preprocess_progress(current, total, desc):
                 _log(f"  {desc} ({current}/{total})")
@@ -514,7 +532,7 @@ def gradio_main():
 
             for msg in train_lora_generator(
                 dataset_dir=preprocessed_dir,
-                output_dir=output_dir,
+                output_dir=adapter_out,
                 checkpoint_dir=ACE_CHECKPOINT_DIR,
                 epochs=epochs,
                 lr=lr,
@@ -580,10 +598,7 @@ def gradio_main():
 
     # -- Build LM model choices --
     def _lm_model_choices():
-        models = _scan_lm_models()
-        choices = ["Default"]
-        choices.extend(models)
-        return choices
+        return _scan_lm_models()
 
     # -- Build UI --
     CSS = """
@@ -646,7 +661,7 @@ def gradio_main():
                     )
                     lm_model_select = gr.Dropdown(
                         label="LM Model", choices=_lm_model_choices(),
-                        value="Default", scale=1,
+                        value=_lm_model_choices()[0] if _lm_model_choices() else None, scale=1,
                     )
 
                 with gr.Row(elem_classes="compact-row"):

train_engine.py

@@ -1316,18 +1316,17 @@ def train_lora_generator(
         unload_models(model)
         return
 
-    # Final save
-    final_path = str(out_path / "final")
+    # Final save (directly to output_dir, not a subdirectory)
     model.decoder.eval()
-    save_lora_adapter(model, final_path)
+    save_lora_adapter(model, str(out_path))
 
     final_loss = avg_epoch_loss if num_updates > 0 else 0.0
     best_note = ""
     if best_epoch > 0 and Path(out_path / "best").exists():
         best_note = f"\n  Best: {out_path / 'best'} (epoch {best_epoch}, loss: {best_loss:.4f})"
     yield (
-        f"[OK] Training complete! LoRA saved to {final_path}{best_note}\n"
-        f"  For inference, set your LoRA path to: {final_path}"
+        f"[OK] Training complete! LoRA saved to {out_path}{best_note}\n"
+        f"  Adapter ready for inference."
     )
     yield "[DONE]"
     unload_models(model)