diff --git a/model_training_nnn/rnn_trainer.py b/model_training_nnn/rnn_trainer.py
index feea704..07cc2a2 100644
--- a/model_training_nnn/rnn_trainer.py
+++ b/model_training_nnn/rnn_trainer.py
@@ -59,8 +59,13 @@ class BrainToTextDecoder_Trainer:
         )
 
         # Initialize Accelerator for TPU/multi-device support
+        use_xla = bool(xm.get_xla_supported_devices())
+        mixed_precision_mode = 'no'
+        if not use_xla and args.get('use_amp', True):
+            mixed_precision_mode = 'bf16'
+
         self.accelerator = Accelerator(
-            mixed_precision='bf16' if args.get('use_amp', True) else 'no',
+            mixed_precision=mixed_precision_mode,
             gradient_accumulation_steps=args.get('gradient_accumulation_steps', 1),
             log_with=None,  # We'll use our own logging
             project_dir=args.get('output_dir', './output'),
@@ -126,6 +131,8 @@ class BrainToTextDecoder_Trainer:
         self.logger.info(f'Accelerator state: {self.accelerator.state}')
         if self.accelerator.num_processes > 1:
             self.logger.info(f'Distributed training on {self.accelerator.num_processes} processes')
+        if mixed_precision_mode == 'no' and use_xla and args.get('use_amp', True):
+            self.logger.info('AMP requested but disabled on XLA to avoid dtype mismatches; running in float32 on TPU.')
 
         # Set seed if provided (using Accelerator's set_seed for proper distributed seeding)
         if self.args['seed'] != -1: