fix: train on XL turbo (matches XL GGUF for inference), add XL checkpoint download

Dockerfile

@@ -79,11 +79,17 @@ RUN pip3 install --no-cache-dir --extra-index-url https://download.pytorch.org/w
 # 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
 
-# Pre-download training checkpoints (avoids 10GB runtime download)
+# Pre-download training checkpoints (avoids runtime download)
+# Base repo has VAE + text encoder + standard turbo
 RUN python3 -c "from huggingface_hub import snapshot_download; \
     snapshot_download('ACE-Step/Ace-Step1.5', local_dir='/app/checkpoints', \
     ignore_patterns=['*.md', '*.txt', '.gitattributes'])"
 
+# XL turbo checkpoint for training (matches XL GGUF used for inference)
+RUN python3 -c "from huggingface_hub import snapshot_download; \
+    snapshot_download('ACE-Step/acestep-v15-xl-turbo', local_dir='/app/checkpoints/acestep-v15-xl-turbo', \
+    ignore_patterns=['*.md', '*.txt', '.gitattributes'])"
+
 # Copy application files
 COPY app.py /app/app.py
 COPY train_engine.py /app/train_engine.py

app.py

@@ -514,7 +514,7 @@ def gradio_main():
                 output_dir=preprocessed_dir,
                 checkpoint_dir=ACE_CHECKPOINT_DIR,
                 device="cpu",
-                variant="turbo",
+                variant="xl-turbo",
                 max_duration=float(MAX_AUDIO_DURATION),
                 progress_callback=preprocess_progress,
                 cancel_check=lambda: False,
@@ -554,7 +554,7 @@ def gradio_main():
                 max_grad_norm=1.0,
                 save_every_n_epochs=max(1, epochs // 2),
                 seed=42,
-                variant="turbo",
+                variant="xl-turbo",
                 device="cpu",
                 log_every=5,
             ):