disable flash_sdp on CPU, force attn_implementation=sdpa for training

app.py

@@ -335,6 +335,10 @@ time.sleep(2)
 gc.collect()
 
 try:
+    import torch
+    torch.backends.cuda.enable_flash_sdp(False)
+    os.environ["ATTN_BACKEND"] = "sdpa"
+
     import torchaudio
     _orig = torchaudio.load
     def _sf(p, *a, **kw):
@@ -371,10 +375,15 @@ try:
     from acestep.training_v2.trainer_fixed import FixedLoRATrainer
     from acestep.training_v2.configs import TrainingConfigV2, LoRAConfigV2
 
+    log("  Loading decoder (attn_implementation=sdpa)...")
     model = load_decoder_for_training(
         checkpoint_dir="{ACE_CHECKPOINT_DIR}", variant="turbo",
         device="cpu", precision="float32",
     ).float()
+    for m in model.modules():
+        if hasattr(m, 'config') and hasattr(m.config, '_attn_implementation'):
+            m.config._attn_implementation = "sdpa"
+    log("  Decoder loaded, applying LoRA...")
 
     trainer = FixedLoRATrainer(model,
         LoRAConfigV2(r={rank}, alpha={rank}, dropout=0.0),