fix: auto-detect bf16 support

train.py

@@ -175,8 +175,11 @@ trainer = GRPOTrainer(
     model=model,
     processing_class=tokenizer,
     reward_funcs=[reward_fn],
+    import torch as _torch
     args=GRPOConfig(
         output_dir="/tmp/deceit-1.5b",
+        bf16=_torch.cuda.is_available() and _torch.cuda.is_bf16_supported(),
+        fp16=False,
         max_steps=150,
         per_device_train_batch_size=4,
         num_generations=4,