diff --git a/unsloth/models/falcon_h1.py b/unsloth/models/falcon_h1.py
index 70223a3c20..3010d37163 100644
--- a/unsloth/models/falcon_h1.py
+++ b/unsloth/models/falcon_h1.py
@@ -116,7 +116,7 @@ def FalconH1Attention_fast_forward(
     if past_key_value is not None:
         kv_seq_len += past_key_value[0].shape[-2]
 
-    if position_embeddings:
+    if position_embeddings and kv_seq_len <= position_embeddings[0].shape[0]:
         cos, sin = position_embeddings
     else:
         # Extend RoPE dynamically to fit in VRA
diff --git a/unsloth/models/llama.py b/unsloth/models/llama.py
index caf259c0bd..502cbec8bd 100644
--- a/unsloth/models/llama.py
+++ b/unsloth/models/llama.py
@@ -566,7 +566,7 @@ def LlamaAttention_fast_forward(
     if past_key_value is not None:
         kv_seq_len += past_key_value[0].shape[-2]
 
-    if position_embeddings:
+    if position_embeddings and kv_seq_len <= position_embeddings[0].shape[0]:
         cos, sin = position_embeddings
     else:
         # Extend RoPE dynamically to fit in VRA
diff --git a/unsloth/models/qwen3.py b/unsloth/models/qwen3.py
index 3d905fe06a..1a0d641552 100644
--- a/unsloth/models/qwen3.py
+++ b/unsloth/models/qwen3.py
@@ -110,7 +110,7 @@ def Qwen3Attention_fast_forward(
     if past_key_value is not None:
         kv_seq_len += past_key_value[0].shape[-2]
 
-    if position_embeddings:
+    if position_embeddings and kv_seq_len <= position_embeddings[0].shape[0]:
         cos, sin = position_embeddings
     else:
         # Extend RoPE dynamically to fit in VRA