diff --git a/comfy/text_encoders/llama.py b/comfy/text_encoders/llama.py
index 4c976058f..5e11956b5 100644
--- a/comfy/text_encoders/llama.py
+++ b/comfy/text_encoders/llama.py
@@ -128,11 +128,12 @@ def precompute_freqs_cis(head_dim, position_ids, theta, rope_dims=None, device=N
 
 
 def apply_rope(xq, xk, freqs_cis):
+    org_dtype = xq.dtype
     cos = freqs_cis[0]
     sin = freqs_cis[1]
     q_embed = (xq * cos) + (rotate_half(xq) * sin)
     k_embed = (xk * cos) + (rotate_half(xk) * sin)
-    return q_embed, k_embed
+    return q_embed.to(org_dtype), k_embed.to(org_dtype)
 
 
 class Attention(nn.Module):
diff --git a/comfy_extras/nodes_easycache.py b/comfy_extras/nodes_easycache.py
index 9d2988f5f..c170e9fd9 100644
--- a/comfy_extras/nodes_easycache.py
+++ b/comfy_extras/nodes_easycache.py
@@ -162,7 +162,12 @@ def easycache_sample_wrapper(executor, *args, **kwargs):
             logging.info(f"{easycache.name} [verbose] - output_change_rates {len(output_change_rates)}: {output_change_rates}")
             logging.info(f"{easycache.name} [verbose] - approx_output_change_rates {len(approx_output_change_rates)}: {approx_output_change_rates}")
         total_steps = len(args[3])-1
-        logging.info(f"{easycache.name} - skipped {easycache.total_steps_skipped}/{total_steps} steps ({total_steps/(total_steps-easycache.total_steps_skipped):.2f}x speedup).")
+        # catch division by zero for log statement; sucks to crash after all sampling is done
+        try:
+            speedup = total_steps/(total_steps-easycache.total_steps_skipped)
+        except ZeroDivisionError:
+            speedup = 1.0
+        logging.info(f"{easycache.name} - skipped {easycache.total_steps_skipped}/{total_steps} steps ({speedup:.2f}x speedup).")
         easycache.reset()
         guider.model_options = orig_model_options