sync from github

src/backend/tasks/arena_hard/task.py

@@ -72,7 +72,7 @@ class ArenaHard(ConfigurableTask):
         super().__init__(config={"metadata": {"version": self.VERSION}})
         # these end tokens are hard coded because of the current limitaion of the llm-eval.
         # self.generation_kwargs = {"until": ["\n\n", "<unk>", "<|im_end|>", "</s>", "<|endoftext|>"], "max_length": 512}
-        self.generation_kwargs = {"until": ["</s>", "<|im_end|>"], "max_length": 4096}
+        self.generation_kwargs = {"until": ["</s>", "<|im_end|>"], "max_gen_toks": 4096}
         # self.generation_kwargs_sampling_number = 5  # the number of sampling for self-consistence
         # self.generation_kwargs_sampling = {
         #     "temperature": 0.99,

src/display/utils.py

@@ -188,6 +188,7 @@ class InferenceFramework(Enum):
     HF_Chat = ModelDetails("hf-chat")
     VLLM = ModelDetails("vllm_moe")
     TRTLLM = ModelDetails("tensorrt_llm")
+    VLLM_FIX = ModelDetails("vllm_moe_fixbs")
     Unknown = ModelDetails("?")
 
     def to_str(self):
@@ -203,6 +204,8 @@ class InferenceFramework(Enum):
             return InferenceFramework.HF_Chat
         if inference_framework in ["vllm_moe"]:
             return InferenceFramework.VLLM
+        if inference_framework in ["vllm_moe_fixbs"]:
+            return InferenceFramework.VLLM_FIX
         return InferenceFramework.Unknown
 
 class GPUType(Enum):