add dp 2

qwen3_6_scripts/serving_chat.py

@@ -243,6 +243,7 @@ class OpenAIServingChat(OpenAIServing):
 
         # set up reasoning parser
         self.reasoning_parser_cls = None
+        self._template_patched = False  # [BI100] one-shot chat template patch
         if reasoning_parser:
             try:
                 from vllm.reasoning import ReasoningParserManager
@@ -309,6 +310,19 @@ class OpenAIServingChat(OpenAIServing):
             model_config = self.model_config
             tokenizer = await self.engine_client.get_tokenizer(lora_request)
 
+            # [BI100] Runtime equivalent of patch_chat_template.py.
+            # Qwen3.5/3.6-MoE templates inject '<think>\n\n</think>\n\n'
+            # when enable_thinking=false, which causes the model to
+            # degenerate (outputs '!!!!!'). Patch it out once.
+            if not self._template_patched and hasattr(tokenizer, 'chat_template'):
+                _tgt = r"{{- '<think>\n\n</think>\n\n' }}"
+                if isinstance(tokenizer.chat_template, str) and _tgt in tokenizer.chat_template:
+                    tokenizer.chat_template = tokenizer.chat_template.replace(
+                        _tgt, "{{- '' }}", 1)
+                    logger.info("[BI100] Patched chat_template: removed empty "
+                                "<think></think> block for non-thinking mode")
+                self._template_patched = True
+
             conversation, mm_data_future = parse_chat_messages_futures(
                 request.messages, model_config, tokenizer)
 

qwen3_6_scripts/vendor_overrides/vllm/engine/arg_utils.py

@@ -1,6 +1,7 @@
 import argparse
 import dataclasses
 import json
+import os
 from dataclasses import dataclass
 from typing import (TYPE_CHECKING, Any, Dict, List, Literal, Mapping, Optional,
                     Tuple, Type, Union)
@@ -105,7 +106,7 @@ class EngineArgs:
                                                  Type[ExecutorBase]]] = None
     pipeline_parallel_size: int = 1
     tensor_parallel_size: int = 1
-    data_parallel_size: int = 1
+    data_parallel_size: int = int(os.environ.get("VLLM_DATA_PARALLEL_SIZE", "1"))
     max_parallel_loading_workers: Optional[int] = None
     block_size: int = 16
     enable_prefix_caching: bool = False
@@ -349,12 +350,6 @@ class EngineArgs:
                             type=int,
                             default=EngineArgs.tensor_parallel_size,
                             help='Number of tensor parallel replicas.')
-        parser.add_argument('--data-parallel-size',
-                            '-dp',
-                            type=int,
-                            default=EngineArgs.data_parallel_size,
-                            help='Number of data parallel replicas. '
-                            'Total GPUs = tp * pp * dp.')
         parser.add_argument(
             '--max-parallel-loading-workers',
             type=int,