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

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)