From 16a679c9112c4e02f75e77fca05a09ea7f859a7e Mon Sep 17 00:00:00 2001
From: Dipesh Babu <59379458+dipeshbabu@users.noreply.github.com>
Date: Sat, 28 Feb 2026 01:51:31 -0500
Subject: [PATCH] Disable growth and make the failure explicit

---
 nanochat/gpt.py | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/nanochat/gpt.py b/nanochat/gpt.py
index e7e1955..e7a6ce5 100644
--- a/nanochat/gpt.py
+++ b/nanochat/gpt.py
@@ -179,8 +179,7 @@ class GPT(nn.Module):
         # Precompute a reasonably large RoPE cache up front (cheap relative to model weights).
         # The cache may also grow lazily in forward() if generation exceeds this length.
         self.rotary_seq_len = config.sequence_len * 10
-        # Bound lazy growth to avoid unbounded memory usage during very long generation runs.
-        self.max_rotary_seq_len = max(self.rotary_seq_len, config.sequence_len * 64)
+        self.max_rotary_seq_len = self.rotary_seq_len
         
         head_dim = config.n_embd // config.n_head
         cos, sin = self._precompute_rotary_embeddings(self.rotary_seq_len, head_dim)