MultiScaleDeformableAttention的OP已经编译成功 Gradio 界面启动时怎么默认启用MultiScaleDeformableAttention

自己在实际的使用过程 中 经常的进行动物的面部表情检测的时候总是会遇到 启动参数: python app.py --server_name 0.0.0.0 --flag-use-half-precision --share --server_port 7860 [13:31:56] Load appearance_feature_extractor from live_portrait_wrapper.py:46 /workspace/liveportrait_env/LivePortrait/pretrained_weights/liveportrait/base_models/appearance_feature_extractor.pth done. [13:31:57] Load motion_extractor from live_portrait_wrapper.py:49 /workspace/liveportrait_env/LivePortrait/pretrained_weights/liveportrait/base_models/motion_extractor.pth done. [13:31:58] Load warping_module from live_portrait_wrapper.py:52 /workspace/liveportrait_env/LivePortrait/pretrained_weights/liveportrait/base_models/warping_module.pth done. [13:32:00] Load spade_generator from live_portrait_wrapper.py:55 /workspace/liveportrait_env/LivePortrait/pretrained_weights/liveportrait/base_models/spade_generator.pth done. Load stitching_retargeting_module from live_portrait_wrapper.py:59 /workspace/liveportrait_env/LivePortrait/pretrained_weights/liveportrait/retargeting_models/stitching_retargeting_module.pth done. [13:32:06] FaceAnalysisDIY warmup time: 4.090s face_analysis_diy.py:79 [13:32:08] LandmarkRunner warmup time: 1.392s human_landmark_runner.py:95 Running on local URL: http://0.0.0.0:7860 Running on public URL: https://1a1129f399bd5dc3b4.gradio.live This share link expires in 72 hours. For free permanent hosting and GPU upgrades, run `gradio deploy` from Terminal to deploy to Spaces (https://huggingface.co/spaces) [13:39:25] Load source image from /tmp/gradio/1ca02b4ec2bfff7aad2bf53f004a631faa911f931b36257dd72fbb8d723238d0/head.png live_portrait_pipeline.py:90 Load driving video from: /tmp/gradio/63f6f9962e1fdf6e6722172e7a18155204858d5d5ce3b1e0646c150360c33bed/d0.mp4, FPS is 25 live_portrait_pipeline.py:133 [13:39:26] Start making driving motion template... live_portrait_pipeline.py:144 Making motion templates... ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 100% 0:00:04 [13:39:50] Dump motion template to /tmp/gradio/63f6f9962e1fdf6e6722172e7a18155204858d5d5ce3b1e0646c150360c33bed/d0.pkl live_portrait_pipeline.py:172 Prepared pasteback mask done. live_portrait_pipeline.py:183 No face detected in the source image. cropper.py:107 [13:41:15] Load source image from /tmp/gradio/1ca02b4ec2bfff7aad2bf53f004a631faa911f931b36257dd72fbb8d723238d0/head.png live_portrait_pipeline.py:90 Load driving video from: /tmp/gradio/63f6f9962e1fdf6e6722172e7a18155204858d5d5ce3b1e0646c150360c33bed/d0.mp4, FPS is 25 live_portrait_pipeline.py:133 [13:41:16] Start making driving motion template... live_portrait_pipeline.py:144 Making motion templates... ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 100% 0:00:01 [13:41:19] Dump motion template to /tmp/gradio/63f6f9962e1fdf6e6722172e7a18155204858d5d5ce3b1e0646c150360c33bed/d0.pkl live_portrait_pipeline.py:172 Prepared pasteback mask done. live_portrait_pipeline.py:183 No face detected in the source image. 自己想在使用动物进行表情检测的时候，是不是启用MultiScaleDeformableAttention的OP此库就可以增加检测的成功率，现在很难检测动物的表情 怎么使用MultiScaleDeformableAttention的OP这个库，希望给指导下