diff --git a/plugins/huggingface/modelgauge/suts/huggingface_chat_completion.py b/plugins/huggingface/modelgauge/suts/huggingface_chat_completion.py index 378be005..12e1add4 100644 --- a/plugins/huggingface/modelgauge/suts/huggingface_chat_completion.py +++ b/plugins/huggingface/modelgauge/suts/huggingface_chat_completion.py @@ -1,5 +1,6 @@ from dataclasses import asdict from typing import Dict, List, Optional + from huggingface_hub import get_inference_endpoint, InferenceClient, InferenceEndpointStatus # type: ignore from huggingface_hub.utils import HfHubHTTPError # type: ignore from pydantic import BaseModel @@ -128,7 +129,7 @@ def translate_response( SUTS.register( HuggingFaceChatCompletionSUT, "gemma-2-9b-it-hf", - "gemma-2-9b-it-qfa", + "gemma-2-9b-it-plf", HF_SECRET, )