Spaces:
Running
Running
| import gradio as gr | |
| def get_base_backend_config(backend_name="pytorch"): | |
| return [ | |
| # seed | |
| gr.Textbox( | |
| value=42, | |
| label=f"{backend_name}.seed", | |
| info="Sets seed for reproducibility", | |
| ), | |
| # inter_op_num_threads | |
| gr.Textbox( | |
| value="null", | |
| label=f"{backend_name}.inter_op_num_threads", | |
| info="Use null for default and -1 for cpu_count()", | |
| ), | |
| # intra_op_num_threads | |
| gr.Textbox( | |
| value="null", | |
| label=f"{backend_name}.intra_op_num_threads", | |
| info="Use null for default and -1 for cpu_count()", | |
| ), | |
| # initial_isolation_check | |
| gr.Checkbox( | |
| value=True, | |
| label=f"{backend_name}.initial_isolation_check", | |
| info="Makes sure that initially, no other process is running on the target device", | |
| ), | |
| # continous_isolation_check | |
| gr.Checkbox( | |
| value=True, | |
| label=f"{backend_name}.continous_isolation_check", | |
| info="Makes sure that throughout the benchmark, no other process is running on the target device", | |
| ), | |
| # delete_cache | |
| gr.Checkbox( | |
| value=False, | |
| label=f"{backend_name}.delete_cache", | |
| info="Deletes model cache (weights & configs) after benchmark is done", | |
| ), | |
| ] | |
| def get_pytorch_config(): | |
| return get_base_backend_config(backend_name="pytorch") + [ | |
| # no_weights | |
| gr.Checkbox( | |
| value=False, | |
| label="pytorch.no_weights", | |
| info="Generates random weights instead of downloading pretrained ones", | |
| ), | |
| # # device_map | |
| # gr.Dropdown( | |
| # value="null", | |
| # | |
| # label="pytorch.device_map", | |
| # choices=["null", "auto", "sequential"], | |
| # info="Use null for default and `auto` or `sequential` the same way as in `from_pretrained`", | |
| # ), | |
| # torch_dtype | |
| gr.Dropdown( | |
| value="null", | |
| label="pytorch.torch_dtype", | |
| choices=["null", "bfloat16", "float16", "float32", "auto"], | |
| info="Use null for default and `auto` for automatic dtype selection", | |
| ), | |
| # amp_autocast | |
| gr.Checkbox( | |
| value=False, | |
| label="pytorch.amp_autocast", | |
| info="Enables Pytorch's native Automatic Mixed Precision", | |
| ), | |
| # amp_dtype | |
| gr.Dropdown( | |
| value="null", | |
| label="pytorch.amp_dtype", | |
| info="Use null for default", | |
| choices=["null", "bfloat16", "float16"], | |
| ), | |
| # torch_compile | |
| gr.Checkbox( | |
| value=False, | |
| label="pytorch.torch_compile", | |
| info="Compiles the model with torch.compile", | |
| ), | |
| # bettertransformer | |
| gr.Checkbox( | |
| value=False, | |
| label="pytorch.bettertransformer", | |
| info="Applies optimum.BetterTransformer for fastpath anf optimized attention", | |
| ), | |
| # quantization_scheme | |
| gr.Dropdown( | |
| value="null", | |
| choices=["null", "gptq", "bnb"], | |
| label="pytorch.quantization_scheme", | |
| info="Use null for no quantization", | |
| ), | |
| # # use_ddp | |
| # gr.Checkbox( | |
| # value=False, | |
| # | |
| # label="pytorch.use_ddp", | |
| # info="Uses DistributedDataParallel for multi-gpu training", | |
| # ), | |
| # peft_strategy | |
| gr.Textbox( | |
| value="null", | |
| label="pytorch.peft_strategy", | |
| ), | |
| ] | |
| def get_onnxruntime_config(): | |
| return get_base_backend_config(backend_name="onnxruntime") | |
| # no_weights | |
| # no_weights: bool = False | |
| # # export options | |
| # export: bool = True | |
| # use_cache: bool = True | |
| # use_merged: bool = False | |
| # torch_dtype: Optional[str] = None | |
| # # provider options | |
| # provider: str = "${infer_provider:${device}}" | |
| # device_id: Optional[int] = "${oc.deprecated:backend.provider_options.device_id}" | |
| # provider_options: Dict[str, Any] = field(default_factory=lambda: {"device_id": "${infer_device_id:${device}}"}) | |
| # # inference options | |
| # use_io_binding: bool = "${is_gpu:${device}}" | |
| # enable_profiling: bool = "${oc.deprecated:backend.session_options.enable_profiling}" | |
| # session_options: Dict[str, Any] = field( | |
| # default_factory=lambda: {"enable_profiling": "${is_profiling:${benchmark.name}}"} | |
| # ) | |
| # # optimization options | |
| # optimization: bool = False | |
| # optimization_config: Dict[str, Any] = field(default_factory=dict) | |
| # # quantization options | |
| # quantization: bool = False | |
| # quantization_config: Dict[str, Any] = field(default_factory=dict) | |
| # # calibration options | |
| # calibration: bool = False | |
| # calibration_config: Dict[str, Any] = field(default_factory=dict) | |
| # # null, O1, O2, O3, O4 | |
| # auto_optimization: Optional[str] = None | |
| # auto_optimization_config: Dict[str, Any] = field(default_factory=dict) | |
| # # null, arm64, avx2, avx512, avx512_vnni, tensorrt | |
| # auto_quantization: Optional[str] = None | |
| # auto_quantization_config: Dict[str, Any] = field(default_factory=dict) | |
| # # ort-training is basically a different package so we might need to seperate these two backends in the future | |
| # use_inference_session: bool = "${is_inference:${benchmark.name}}" | |
| # # training options | |
| # use_ddp: bool = False | |
| # ddp_config: Dict[str, Any] = field(default_factory=dict) | |
| # # peft options | |
| # peft_strategy: Optional[str] = None | |
| # peft_config: Dict[str, Any] = field(default_factory=dict) | |
| def get_openvino_config(): | |
| return get_base_backend_config(backend_name="openvino") | |
| def get_neural_compressor_config(): | |
| return get_base_backend_config(backend_name="neural-compressor") | |
| def get_text_generation_inference_config(): | |
| return get_base_backend_config(backend_name="text-generation-inference") | |
| def get_inference_config(): | |
| return [ | |
| # duration | |
| gr.Textbox( | |
| value=10, | |
| label="inference.duration", | |
| info="Minimum duration of benchmark in seconds", | |
| ), | |
| # warmup runs | |
| gr.Textbox( | |
| value=10, | |
| label="inference.warmup_runs", | |
| info="Number of warmup runs before measurements", | |
| ), | |
| # memory | |
| gr.Checkbox( | |
| value=False, | |
| label="inference.memory", | |
| info="Measures the peak memory footprint", | |
| ), | |
| # energy | |
| gr.Checkbox( | |
| value=False, | |
| label="inference.energy", | |
| info="Measures energy consumption and carbon emissions", | |
| ), | |
| # input_shapes | |
| gr.Dataframe( | |
| type="array", | |
| value=[[2, 16]], | |
| row_count=(1, "static"), | |
| col_count=(2, "dynamic"), | |
| label="inference.input_shapes", | |
| headers=["batch_size", "sequence_length"], | |
| info="Controllable input shapes, add more columns for more inputs", | |
| ), | |
| # forward kwargs | |
| gr.Dataframe( | |
| type="array", | |
| value=[[False]], | |
| headers=["return_dict"], | |
| row_count=(1, "static"), | |
| col_count=(1, "dynamic"), | |
| label="inference.forward_kwargs", | |
| info="Keyword arguments for the forward pass, add more columns for more arguments", | |
| ), | |
| ] | |
| def get_training_config(): | |
| return [ | |
| # warmup steps | |
| gr.Textbox( | |
| value=40, | |
| label="training.warmup_steps", | |
| ), | |
| # dataset_shapes | |
| gr.Dataframe( | |
| type="array", | |
| value=[[500, 16]], | |
| headers=["dataset_size", "sequence_length"], | |
| row_count=(1, "static"), | |
| col_count=(2, "dynamic"), | |
| label="training.dataset_shapes", | |
| ), | |
| # training_arguments | |
| gr.Dataframe( | |
| value=[[2]], | |
| type="array", | |
| row_count=(1, "static"), | |
| col_count=(1, "dynamic"), | |
| label="training.training_arguments", | |
| headers=["per_device_train_batch_size"], | |
| ), | |
| ] | |