adapt long sequence input when rerope is disabled

Author: xinSky00

ucm/integration/vllm/patch/0.9.2/vllm-rerope-adapt.patch

@@ -1430,7 +1430,7 @@ index 5a26e88db..df86287c8 100644
          max_num_scheduled_tokens = max(tokens)
 
 +        # Get use_rerope
-+        if len(scheduler_output.scheduled_new_reqs) > 0:
++        if envs.VLLM_USE_REROPE and len(scheduler_output.scheduled_new_reqs) > 0:
 +            self.use_rerope = max([len(req.prompt_token_ids) 
 +                                   for req in scheduler_output.scheduled_new_reqs]) > envs.REROPE_WINDOW
 +

ucm/integration/vllm/patch/patch_funcs/v092/vllm_rerope_patch.py

@@ -164,6 +164,7 @@ def add_init(
 
         from typing import TYPE_CHECKING, Any, Optional
 
+        import os
         import numpy as np
         from vllm.v1.attention.backends.utils import CommonAttentionMetadata
         from vllm.v1.spec_decode.metadata import SpecDecodeMetadata
@@ -202,7 +203,13 @@ def _prepare_inputs_modify(
 
             ###################### rerope patch ###############
             # Get use_rerope
-            if len(scheduler_output.scheduled_new_reqs) > 0:
+            vllm_use_rerope = os.getenv("VLLM_USE_REROPE", "0").lower() in (
+                                        "1",
+                                        "true",
+                                        "yes",
+                                        "on",
+                                    )
+            if vllm_use_rerope and len(scheduler_output.scheduled_new_reqs) > 0:
                 self.use_rerope = (
                     max(
                         [