diff --git a/results/retrieval_VLACombined_256.csv b/results/retrieval_VLACombined_256.csv
new file mode 100644
index 0000000..2f06fcc
--- /dev/null
+++ b/results/retrieval_VLACombined_256.csv
@@ -0,0 +1 @@
+epoch,train_loss,train_acc,test_acc,epoch_time_sec
diff --git a/results/retrieval_VLAKVExploding_256.csv b/results/retrieval_VLAKVExploding_256.csv
new file mode 100644
index 0000000..e69de29
diff --git a/src/models/attention/fast_vla.py b/src/models/attention/fast_vla.py
index 70b3b5c..54ad7ee 100644
--- a/src/models/attention/fast_vla.py
+++ b/src/models/attention/fast_vla.py
@@ -6,15 +6,15 @@
 
 # Try importing Triton
 try:
-    import triton
-    import triton.language as tl
+    import triton  # type: ignore
+    import triton.language as tl  # type: ignore
     HAS_TRITON = True
 except ImportError:
     HAS_TRITON = False
 
 def _combine(F_l, G_l, F_r, G_r):
     """Associative operator for the (F, G) recurrence."""
-    return torch.matmul(F_r, F_l), torch.matmul(F_r, G_l) + G_r
+    return torch.matmul(F_r, F_l).to(F_l.dtype), (torch.matmul(F_r, G_l) + G_r).to(G_l.dtype)
 
 def parallel_scan(Fs, Gs):
     """