fix: save_every_n_epochs=0, add demucs-infer to Dockerfile, debug adapter dir

Dockerfile

@@ -78,7 +78,7 @@ RUN pip3 install --no-cache-dir --extra-index-url https://download.pytorch.org/w
     "gradio[mcp]>=6.0.0,<7.0.0" requests torch safetensors \
     "transformers>=4.51.0,<4.58.0" peft>=0.18.0 \
     loguru "torchaudio==2.4.0" "diffusers==0.30.3" lightning numpy tensorboard soundfile \
-    einops vector_quantize_pytorch librosa mutagen
+    einops vector_quantize_pytorch librosa mutagen demucs-infer
 
 # Clone ACE-Step repo for training module
 RUN git clone --depth 1 https://github.com/ace-step/ACE-Step-1.5 /app/ace-step-source

app.py

@@ -730,7 +730,7 @@ def gradio_main():
                 warmup_steps=100,
                 weight_decay=0.01,
                 max_grad_norm=1.0,
-                save_every_n_epochs=max(1, epochs // 2),
+                save_every_n_epochs=0,
                 seed=42,
                 variant="turbo",
                 device="cpu",
@@ -769,6 +769,12 @@ def gradio_main():
                 _log("[OK] ace-server restarted successfully")
             else:
                 _log("[WARN] ace-server may not have restarted -- check logs")
+            # Debug: list what's in the adapter directory
+            if os.path.isdir(adapter_out):
+                contents = os.listdir(adapter_out)
+                _log(f"[DEBUG] Adapter dir {adapter_out}: {contents}")
+            else:
+                _log(f"[DEBUG] Adapter dir {adapter_out} does NOT exist!")
             adapter_safetensors = os.path.join(adapter_out, "adapter_model.safetensors")
             if os.path.isfile(adapter_safetensors):
                 # Copy to a temp file so Gradio doesn't try to validate /app paths