diff --git a/nanochat/gpt.py b/nanochat/gpt.py
index 168ebb8..fb726be 100644
--- a/nanochat/gpt.py
+++ b/nanochat/gpt.py
@@ -184,7 +184,7 @@ class GPT(nn.Module):
         # To support meta device initialization, we init the rotary embeddings here, but it's just "fake" meta tensors only.
         # As for rotary_seq_len, these rotary embeddings are pretty small/cheap in memory,
         # so let's just over-compute them by 10X, but assert fail if we ever reach that amount.
-        # In the future we can dynamically grow the cache, for now it's fine.
+        # The cache grows dynamically in forward() with a 2x buffer if eval prompts exceed this.
         self.rotary_seq_len = config.sequence_len * 10 # 10X over-compute should be enough, TODO make nicer?
         head_dim = config.n_embd // config.n_head
         cos, sin = self._precompute_rotary_embeddings(self.rotary_seq_len, head_dim)
@@ -400,7 +400,7 @@ class GPT(nn.Module):
         T0 = 0 if kv_cache is None else kv_cache.get_pos()
         # Dynamically grow rotary cache if needed (e.g. few-shot eval prompts exceed training seq_len * 10)
         if T0 + T > self.cos.size(1):
-            self.rotary_seq_len = T0 + T
+            self.rotary_seq_len = max(T0 + T, self.cos.size(1) * 2)  # 2x buffer to avoid regrowing every call
             head_dim = self.config.n_embd // self.config.n_head
             self.cos, self.sin = self._precompute_rotary_embeddings(self.rotary_seq_len, head_dim)
         assert idx.device == self.cos.device, f"Rotary embeddings and idx are on different devices: {idx.device} != {self.cos.device}"