Fix RoPE cache overflow for long prompts with KV cache

nanochat/gpt.py

@@ -12,7 +12,6 @@
 - Flash Attention 3 integration
 """
 
-from functools import partial
 from dataclasses import dataclass
 
 import torch
@@ -22,6 +21,7 @@
 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
 
@@ -176,15 +176,60 @@ def __init__(self, config, pad_vocab_size_to=64):
         kv_dim = config.n_kv_head * head_dim
         self.value_embeds = nn.ModuleDict({str(i): nn.Embedding(padded_vocab_size, kv_dim) for i in range(config.n_layer) if has_ve(i, config.n_layer)})
         # 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.
-        self.rotary_seq_len = config.sequence_len * 10 # 10X over-compute should be enough, TODO make nicer?
+        # 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
+        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)
         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):
+        """
+        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.
+
+        Growth is bounded by self.max_rotary_seq_len to avoid unbounded memory usage.
+        """
+        cur_len = self.cos.size(1)
+        if needed_seq_len <= cur_len:
+            return
+
+        if needed_seq_len > self.max_rotary_seq_len:
+            raise RuntimeError(
+                f"RoPE cache request exceeds max_rotary_seq_len: need {needed_seq_len}, "
+                f"have {cur_len}, cap {self.max_rotary_seq_len}. "
+                "Increase max_rotary_seq_len for longer-context generation."
+            )
+
+        # Safety: mutating buffers during torch.compile tracing is unsafe.
+        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}). "
+                "Increase initial rotary_seq_len/max_rotary_seq_len or avoid compiled generation."
+            )
+
+        # Next power-of-two >= needed_seq_len (amortized growth), bounded by cap
+        new_len = min(self.max_rotary_seq_len, 1 << (needed_seq_len - 1).bit_length())
+
+        head_dim = self.config.n_embd // self.config.n_head
+        device = self.cos.device
+        cos, sin = self._precompute_rotary_embeddings(seq_len=new_len, head_dim=head_dim, device=device)
+
+        # Preserve dtype/device invariants (precompute already returns bf16, but keep explicit)
+        cos = cos.to(dtype=self.cos.dtype, device=device)
+        sin = sin.to(dtype=self.sin.dtype, device=device)
+
+        # Overwrite existing registered buffers (persistent=False remains from initial registration)
+        self.cos = cos
+        self.sin = sin
+        self.rotary_seq_len = new_len
+
     @torch.no_grad()
     def init_weights(self):
         """
@@ -387,14 +432,16 @@ def setup_optimizer(self, unembedding_lr=0.004, embedding_lr=0.2, matrix_lr=0.02
 
     def forward(self, idx, targets=None, kv_cache=None, loss_reduction='mean'):
         B, T = idx.size()
-
-        # 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)}"
+        T0 = 0 if kv_cache is None else kv_cache.get_pos()
+
+        # Ensure RoPE buffers cover absolute positions [T0, T0+T)
+        self._ensure_rope_cache(T0 + T)
+
         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
+
+        # If kv cache exists, offset RoPE by current absolute position
+        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