switch training to standard turbo (11s/epoch), auto-select standard GGUF for LoRA inference

Dockerfile

@@ -53,10 +53,14 @@ RUN mkdir -p /app/models /app/adapters /app/outputs
 # Download GGUF models at build time (NOT via Git LFS)
 # --fail ensures curl returns non-zero on HTTP errors (catches 404)
 
-# XL DiT turbo Q4_K_M (~2.8GB) - best quality/size for CPU
+# XL DiT turbo Q4_K_M (~2.8GB) - best quality for no-LoRA inference
 RUN curl -fL --retry 3 --retry-delay 5 -o /app/models/acestep-v15-xl-turbo-Q4_K_M.gguf \
     "https://huggingface.co/Serveurperso/ACE-Step-1.5-GGUF/resolve/main/acestep-v15-xl-turbo-Q4_K_M.gguf"
 
+# Standard DiT turbo Q4_K_M (~1.1GB) - used when LoRA adapter is selected
+RUN curl -fL --retry 3 --retry-delay 5 -o /app/models/acestep-v15-turbo-Q4_K_M.gguf \
+    "https://huggingface.co/Serveurperso/ACE-Step-1.5-GGUF/resolve/main/acestep-v15-turbo-Q4_K_M.gguf"
+
 # LM 1.7B Q8_0 (~1.7GB) - best speed/quality on CPU
 RUN curl -fL --retry 3 --retry-delay 5 -o /app/models/acestep-5Hz-lm-1.7B-Q8_0.gguf \
     "https://huggingface.co/Serveurperso/ACE-Step-1.5-GGUF/resolve/main/acestep-5Hz-lm-1.7B-Q8_0.gguf"

app.py

@@ -148,6 +148,7 @@ def _run_pipeline(caption, lyrics, bpm, duration, seed, steps, output_format,
     synth_request["output_format"] = synth_fmt
     if adapter:
         synth_request["adapter"] = adapter
+        synth_request["synth_model"] = "acestep-v15-turbo-Q4_K_M.gguf"
     if progress_cb:
         progress_cb("synth_submit", None)
     r = requests.post(f"{ACE_SERVER}/synth", json=synth_request, timeout=30)
@@ -514,7 +515,7 @@ def gradio_main():
                 output_dir=preprocessed_dir,
                 checkpoint_dir=ACE_CHECKPOINT_DIR,
                 device="cpu",
-                variant="xl-turbo",
+                variant="turbo",
                 max_duration=float(MAX_AUDIO_DURATION),
                 progress_callback=preprocess_progress,
                 cancel_check=lambda: False,
@@ -554,7 +555,7 @@ def gradio_main():
                 max_grad_norm=1.0,
                 save_every_n_epochs=max(1, epochs // 2),
                 seed=42,
-                variant="xl-turbo",
+                variant="turbo",
                 device="cpu",
                 log_every=5,
             ):