Fix: Safe RoPE Cache Expansion to Prevent Position-ID Out-of-Bounds in EAGLE + Long-Sequence Workloads (#11871)

python/sglang/srt/model_executor/model_runner.py

@@ -155,6 +155,7 @@ from sglang.srt.utils import (
     is_npu,
     log_info_on_rank0,
     monkey_patch_p2p_access_check,
+    reserve_rope_cache_for_long_sequences,
     set_cuda_arch,
     slow_rank_detector,
     xpu_has_xmx_support,
@@ -866,6 +867,14 @@ class ModelRunner:
                 self.pp_rank,
             )
 
+        # Pre-expand RoPE cache before CUDA Graph capture
+        reserve_rope_cache_for_long_sequences(
+            self.model,
+            self.server_args,
+            self.model_config,
+            logger,
+        )
+
         if self.server_args.elastic_ep_backend == "mooncake":
             # Mooncake does not support `monitored_barrier`
             dist.barrier(group=get_tp_group().cpu_group)