diff --git a/nanochat/gpt.py b/nanochat/gpt.py
index 208acd1..3e12424 100644
--- a/nanochat/gpt.py
+++ b/nanochat/gpt.py
@@ -12,7 +12,6 @@ Notable features:
 - Flash Attention 3 integration
 """
 
-from functools import partial
 from dataclasses import dataclass
 
 import torch
@@ -22,6 +21,7 @@ import torch.nn.functional as F
 from nanochat.common import get_dist_info, print0
 from nanochat.optim import MuonAdamW, DistMuonAdamW
 
+
 # Our custom Flash Attention module that automatically uses FA3 on Hopper+ and SDPA fallback elsewhere
 from nanochat.flash_attention import flash_attn
 
@@ -185,6 +185,47 @@ class GPT(nn.Module):
         self.register_buffer("cos", cos, persistent=False) # persistent=False means it's not saved to the checkpoint
         self.register_buffer("sin", sin, persistent=False)
 
+    def _ensure_rope_cache(self, needed_seq_len: int, device: torch.device):
+        """
+        Ensure rotary embedding cache (cos/sin) is long enough for absolute positions [0, needed_seq_len).
+
+        We grow lazily to avoid crashes for long prompts / long KV-cache generation.
+        Growth is amortized by rounding up to the next power of two.
+
+        NOTE: We avoid register_buffer() here; we simply overwrite the existing buffers.
+        """
+        cur_len = self.cos.size(1)
+        if needed_seq_len <= cur_len:
+            return
+
+        # Safety: mutating buffers during torch.compile tracing is unsafe.
+        try:
+            import torch._dynamo
+            if torch._dynamo.is_compiling():
+                raise RuntimeError(
+                    f"RoPE cache too small during torch.compile (need {needed_seq_len}, have {cur_len}). "
+                    f"Increase initial rotary_seq_len or disable compile for generation."
+                )
+        except Exception:
+            # torch._dynamo may not exist in older torch; ignore.
+            pass
+
+        # Next power-of-two >= needed_seq_len
+        new_len = 1 << (needed_seq_len - 1).bit_length()
+
+        head_dim = self.config.n_embd // self.config.n_head
+        cos, sin = self._precompute_rotary_embeddings(
+            seq_len=new_len, head_dim=head_dim, device=device)
+
+        # Preserve dtype/device invariants (precompute returns bf16 already)
+        cos = cos.to(dtype=self.cos.dtype, device=device)
+        sin = sin.to(dtype=self.sin.dtype, device=device)
+
+        # Overwrite existing registered buffers (no re-register)
+        self.cos = cos
+        self.sin = sin
+        self.rotary_seq_len = new_len  # keep metadata consistent
+
     @torch.no_grad()
     def init_weights(self):
         """
@@ -387,14 +428,16 @@ class GPT(nn.Module):
 
     def forward(self, idx, targets=None, kv_cache=None, loss_reduction='mean'):
         B, T = idx.size()
+        T0 = 0 if kv_cache is None else kv_cache.get_pos()
 
-        # Grab the rotary embeddings for the current sequence length (they are of shape (1, seq_len, 1, head_dim/2))
-        assert T <= self.cos.size(1), f"Sequence length grew beyond the rotary embeddings cache: {T} > {self.cos.size(1)}"
+        # Ensure RoPE buffers cover absolute positions [T0, T0+T)
+        self._ensure_rope_cache(T0 + T, device=idx.device)
+
+        # Now it's safe to slice
         assert idx.device == self.cos.device, f"Rotary embeddings and idx are on different devices: {idx.device} != {self.cos.device}"
         assert self.cos.dtype == torch.bfloat16, "Rotary embeddings must be in bfloat16"
-        # if kv cache exists, we need to offset the rotary embeddings to the current position in the cache
-        T0 = 0 if kv_cache is None else kv_cache.get_pos()
-        cos_sin = self.cos[:, T0:T0+T], self.sin[:, T0:T0+T] # truncate cache to current sequence length
+
+        cos_sin = self.cos[:, T0:T0+T], self.sin[:, T0:T0+T]
 
         # Forward the trunk of the Transformer
         x = self.transformer.wte(idx) # embed current token