fix adapter save path, smart LM fallback, compact training UI, remove Server Status

app.py

@@ -671,6 +671,9 @@ def gradio_main():
                     full_path, timeout=600,
                     cancel_check=lambda: _training_cancel.is_set(),
                 )
+                if not caption_data:
+                    use_understand = False
+                    _log(f"  {audio_fname}: GGUF LM too slow, switching to librosa for remaining files")
             if caption_data:
                 bpm_s = caption_data.get("bpm", "?")
                 key_s = caption_data.get("keyscale", caption_data.get("key", "?"))
@@ -685,7 +688,7 @@ def gradio_main():
                     fallback = {"caption": "", "bpm": bpm_val, "key": "", "signature": "", "lyrics": ""}
                     with open(sidecar_json, "w") as cj:
                         json.dump(fallback, cj)
-                    _log(f"  {audio_fname}: librosa fallback BPM={bpm_val}")
+                    _log(f"  {audio_fname}: librosa BPM={bpm_val}")
                 except Exception as cap_exc:
                     _log(f"  {audio_fname}: caption failed: {cap_exc}")
             yield _log_text(), gr.Button(visible=False), gr.Button(visible=True), gr.File()
@@ -921,9 +924,7 @@ def gradio_main():
                         value=DEFAULT_LM, scale=1,
                     )
 
-                with gr.Row(elem_classes="compact-row"):
-                    gen_btn = gr.Button("Generate Music", variant="primary", scale=2)
-                    status_btn = gr.Button("Server Status", scale=1)
+                gen_btn = gr.Button("Generate Music", variant="primary")
 
                 gen_btn.click(
                     fn=generate_music,
@@ -933,31 +934,23 @@ def gradio_main():
                     api_name="generate",
                 )
 
-                status_btn.click(
-                    fn=get_server_status,
-                    inputs=[],
-                    outputs=[status],
-                    api_name="server_status",
-                )
-
             # ============================================================
             # Tab 2: Train LoRA
             # ============================================================
             with gr.Tab("Train LoRA"):
-                gr.Markdown(
-                    "### LoRA Training\n"
-                    "Fine-tune ACE-Step on your audio. "
-                    "CPU training is slow -- ace-server stops during training."
-                )
-
                 with gr.Row(elem_classes="compact-row"):
-                    with gr.Column(scale=2):
-                        train_audio = gr.File(
-                            label="Training Audio + Caption Files",
-                            file_count="multiple",
-                            file_types=["audio", ".txt", ".json"],
+                    with gr.Column(scale=3):
+                        train_log = gr.Textbox(
+                            label="Training Log",
+                            interactive=False,
+                            lines=12,
+                            elem_classes="status-box",
                         )
-                    with gr.Column(scale=1):
+                        train_output_file = gr.File(label="Trained LoRA (download)", visible=False)
+                    with gr.Column(scale=2):
+                        with gr.Row(elem_classes="compact-row"):
+                            train_btn = gr.Button("Train", variant="primary", scale=2)
+                            cancel_btn = gr.Button("Cancel Training", variant="stop", visible=False, scale=1)
                         lora_name = gr.Textbox(label="LoRA Name", value="my-lora")
                         train_epochs = gr.Slider(
                             label="Epochs", minimum=1, maximum=1000,
@@ -968,18 +961,12 @@ def gradio_main():
                             label="Rank (r)", minimum=1, maximum=128,
                             value=32, step=1,
                         )
-
-                with gr.Row(elem_classes="compact-row"):
-                    train_btn = gr.Button("Train", variant="primary", scale=2)
-                    cancel_btn = gr.Button("Cancel Training", variant="stop", visible=False, scale=1)
-
-                train_output_file = gr.File(label="Trained LoRA (download)", visible=False)
-                train_log = gr.Textbox(
-                    label="Training Log",
-                    interactive=False,
-                    lines=10,
-                    elem_classes="status-box",
-                )
+                        train_audio = gr.File(
+                            label="Training Audio (optional caption .txt)",
+                            file_count="multiple",
+                            file_types=["audio", ".txt", ".json"],
+                            height=120,
+                        )
 
                 # Button swap on click (separate handler, like rvc-beatrice)
                 # This fires immediately so user sees Cancel even if training

train_engine.py

@@ -2539,10 +2539,9 @@ def train_lora_generator(
         if _training_cancel.is_set():
             _training_cancel.clear()
             if epoch > start_epoch:
-                early_path = str(out_path / "early_exit")
                 model.decoder.eval()
-                save_lora_adapter(model, early_path)
-                yield f"[OK] Cancelled at epoch {epoch + 1}, saved to {early_path}"
+                save_lora_adapter(model, str(out_path))
+                yield f"[OK] Cancelled at epoch {epoch + 1}, adapter saved"
             else:
                 yield f"[CANCELLED] Stopped before any epoch completed"
             yield "[DONE]"
@@ -2554,10 +2553,9 @@ def train_lora_generator(
         # Timeout check
         elapsed = time.time() - train_start
         if elapsed > MAX_TRAINING_TIME:
-            early_path = str(out_path / "timeout_exit")
             model.decoder.eval()
-            save_lora_adapter(model, early_path)
-            yield f"[WARN] Training timed out after {int(elapsed)}s, saved to {early_path}"
+            save_lora_adapter(model, str(out_path))
+            yield f"[WARN] Training timed out after {int(elapsed)}s, adapter saved"
             yield "[DONE]"
             _cuda_sync(device)
             unload_models(model)
@@ -2721,11 +2719,10 @@ def train_lora_generator(
             f"Loss: {avg_epoch_loss:.4f}{best_str}"
         )
 
-        # Save best
+        # Save best (directly to output dir so ace-server finds it)
         if is_best and epoch + 1 >= 10:
-            best_path = str(out_path / "best")
             model.decoder.eval()
-            save_lora_adapter(model, best_path)
+            save_lora_adapter(model, str(out_path))
             model.decoder.train()
             yield f"[OK] Best model saved (epoch {epoch + 1}, loss: {best_loss:.4f})"