Fix: Dynamic RoPE Cache Expansion to Prevent Position-ID Out-of-Bounds in EAGLE + Long-Sequence Workloads (#10788)

python/sglang/srt/environ.py

@@ -222,6 +222,11 @@ class Envs:
     SGLANG_TRITON_PREFILL_TRUNCATION_ALIGN_SIZE = EnvInt(4096)
     SGLANG_TRITON_DECODE_SPLIT_TILE_SIZE = EnvInt(256)
 
+    # RoPE cache configuration
+    SGLANG_SPEC_EXPANSION_SAFETY_FACTOR = EnvInt(2)
+    SGLANG_ROPE_CACHE_SAFETY_MARGIN = EnvInt(256)
+    SGLANG_ROPE_CACHE_ALIGN = EnvInt(128)
+
     # Overlap Spec V2
     SGLANG_ENABLE_OVERLAP_PLAN_STREAM = EnvBool(False)
 

python/sglang/srt/layers/rotary_embedding.py

@@ -147,6 +147,36 @@ class RotaryEmbedding(CustomOp):
         cache = torch.cat((cos, sin), dim=-1)
         return cache
 
+    def _ensure_cos_sin_cache_length(self, needed_max_pos: int):
+        """Ensure cos_sin_cache length > needed_max_pos."""
+        cur_len = int(self.cos_sin_cache.shape[0])
+        if needed_max_pos < cur_len:
+            return
+
+        # Align to 128 to reduce realloc frequency
+        new_len = ((needed_max_pos + 128) // 128) * 128
+        device = self.cos_sin_cache.device
+        dtype = self.cos_sin_cache.dtype
+
+        # Compute inv_freq on same device
+        inv_freq = self._compute_inv_freq(self.base).to(device=device)
+
+        # Incremental computation for new positions only
+        start = cur_len
+        t_new = torch.arange(start, new_len, dtype=inv_freq.dtype, device=device)
+        if t_new.numel() == 0:
+            return
+
+        freqs_new = torch.einsum("i,j->ij", t_new, inv_freq)
+        cos_new = freqs_new.cos()
+        sin_new = freqs_new.sin()
+        new_rows = torch.cat((cos_new, sin_new), dim=-1).to(dtype=dtype)
+
+        # Update cache with new rows
+        self.cos_sin_cache = torch.cat((self.cos_sin_cache, new_rows), dim=0).to(
+            device=device, dtype=dtype
+        )
+
     def forward_native(
         self,
         positions: torch.Tensor,

python/sglang/srt/model_executor/model_runner.py

@@ -140,6 +140,7 @@ from sglang.srt.utils import (
     log_info_on_rank0,
     monkey_patch_p2p_access_check,
     monkey_patch_vllm_gguf_config,
+    reserve_rope_cache_for_long_sequences,
     set_cuda_arch,
     slow_rank_detector,
 )
@@ -898,6 +899,15 @@ class ModelRunner:
             f"mem usage={self.weight_load_mem_usage:.2f} GB."
         )
 
+        # Pre-expand RoPE cache before CUDA Graph capture
+        reserve_rope_cache_for_long_sequences(
+            self.model,
+            self.server_args,
+            self.model_config,
+            self.req_to_token_pool,
+            logger,
+        )
+
         if self.server_args.elastic_ep_backend == "mooncake":
             # Mooncake does not support `monitored_barrier`
             dist.barrier(group=get_tp_group().cpu_group)

python/sglang/srt/utils/common.py

@@ -3460,3 +3460,61 @@ def cached_triton_kernel(key_fn=None):
         return CachedKernel(fn, key_fn)
 
     return decorator
+
+
+def reserve_rope_cache_for_long_sequences(
+    model, server_args, model_config, req_to_token_pool=None, logger=None
+):
+    """Pre-expand RoPE cache for long sequences and speculative decoding."""
+    from sglang.srt.environ import envs
+
+    if logger is None:
+        import logging
+
+        logger = logging.getLogger(__name__)
+
+    SAFETY_FACTOR = envs.SGLANG_SPEC_EXPANSION_SAFETY_FACTOR.value
+    MARGIN = envs.SGLANG_ROPE_CACHE_SAFETY_MARGIN.value
+    ALIGN = envs.SGLANG_ROPE_CACHE_ALIGN.value
+
+    # 1) Estimate base context upper bound
+    base_ctx = (
+        getattr(server_args, "context_length", None)
+        or getattr(model_config, "context_len", None)
+        or getattr(model_config, "max_model_len", None)
+        or getattr(model_config.hf_text_config, "max_position_embeddings", None)
+        or 2048
+    )
+
+    # 2) Runtime input capacity (including extra_len from req_to_token_pool)
+    inferred_cap = getattr(req_to_token_pool, "max_context_len", None) or base_ctx
+
+    # 3) Speculative decoding expansion
+    steps = int(getattr(server_args, "speculative_num_steps", 0) or 0)
+    draft = int(getattr(server_args, "speculative_num_draft_tokens", 0) or 0)
+    reserve = inferred_cap + steps * draft * SAFETY_FACTOR + MARGIN
+
+    # 4) Align to reduce reallocation frequency
+    reserve = (reserve + ALIGN - 1) // ALIGN * ALIGN
+
+    logger.info(
+        f"RoPE cache reserve={reserve} (base={base_ctx}, cap={inferred_cap}, steps={steps}, draft={draft}, k={SAFETY_FACTOR}, margin={MARGIN})"
+    )
+
+    # Recursively expand all RoPE layers
+    def reserve_rope_cache_recursive(module):
+        for child in module.children():
+            if hasattr(child, "_ensure_cos_sin_cache_length") and hasattr(
+                child, "cos_sin_cache"
+            ):
+                old_len = child.cos_sin_cache.shape[0]
+                child._ensure_cos_sin_cache_length(reserve - 1)
+                new_len = child.cos_sin_cache.shape[0]
+                if new_len > old_len:
+                    logger.info(
+                        f"Expanded RoPE cache from {old_len} to {new_len} positions"
+                    )
+            else:
+                reserve_rope_cache_recursive(child)
+
+    reserve_rope_cache_recursive(model)