Spaces:
Running
Running
| # !usr/bin/env python | |
| # -*- coding:utf-8 -*- | |
| ''' | |
| Description : | |
| Version : 1.0 | |
| Author : MrYXJ | |
| Mail : [email protected] | |
| Github : https://github.com/MrYxJ | |
| Date : 2023-09-05 23:28:32 | |
| LastEditTime : 2023-09-09 19:14:20 | |
| Copyright (C) 2023 mryxj. All rights reserved. | |
| ''' | |
| import gradio as gr | |
| import torch | |
| from accelerate.commands.estimate import check_has_model | |
| from urllib.parse import urlparse | |
| from huggingface_hub.utils import GatedRepoError | |
| from huggingface_hub.utils import RepositoryNotFoundError | |
| from calflops import create_empty_model | |
| from calflops import calculate_flops_hf | |
| from calflops import flops_to_string | |
| from calflops import macs_to_string | |
| from calflops import params_to_string | |
| def calculate_flops_in_hugging_space(model_name: str, | |
| empty_model: torch.nn.modules, | |
| access_token: str, | |
| input_shape: tuple, | |
| bp_factor: float, | |
| output_unit: str): | |
| "Calculates the FLOPs and Params usage for a model init on `meta` device" | |
| try: | |
| # print("model_name:", model_name) | |
| # print("empty_model:", empty_model) | |
| # print("access_token:", access_token) | |
| # print("input_shape:", input_shape) | |
| flops, macs, params, return_print = calculate_flops_hf(model_name=model_name, | |
| empty_model=empty_model, | |
| input_shape=input_shape, | |
| access_token=access_token, | |
| output_as_string=False, | |
| return_results=True) | |
| except Exception as e: | |
| print("Error info:", e) | |
| raise gr.Error( | |
| f"Model `{model_name}` does not support inference on the meta device, You can download the complete model parameters to your local and using the python package calflops to calculate FLOPs and Params of model `{model_name}`." | |
| ) | |
| fw_bp_flops = flops * (1.0 + bp_factor) | |
| fw_bp_macs = macs * (1.0 + bp_factor) | |
| if output_unit == "": | |
| pass | |
| elif output_unit == "auto": | |
| params = params_to_string(params, units=None, precision=3) | |
| flops = flops_to_string(flops, units=None, precision=3) | |
| macs = macs_to_string(macs, units=None, precision=3) | |
| fw_bp_flops = flops_to_string(fw_bp_flops, units=None, precision=3) | |
| fw_bp_macs = macs_to_string(fw_bp_macs, units=None, precision=3) | |
| elif output_unit == "T" or output_unit == "G" or output_unit == "M" or output_unit == "K" or output_unit == "m" or output_unit == "u": | |
| params = params_to_string(params, units=output_unit, precision=3) | |
| flops = flops_to_string(flops, units=output_unit, precision=3) | |
| macs = macs_to_string(macs, units=output_unit, precision=3) | |
| fw_bp_flops = flops_to_string(fw_bp_flops, units=output_unit, precision=3) | |
| fw_bp_macs = macs_to_string(fw_bp_macs, units=output_unit, precision=3) | |
| return_lines = return_print.split("\n") | |
| return_start = False | |
| return_print = "" | |
| for line in return_lines[:-2]: | |
| if return_start: | |
| return_print += line + "\n" | |
| if "Detailed" in line: | |
| return_start = True | |
| data = [] | |
| data.append( | |
| { "Total Training Params": params, | |
| "Forward FLOPs": flops, | |
| "Forward MACs": macs, | |
| "Forward+Backward FLOPs": fw_bp_flops, | |
| "Forward+Backward MACs": fw_bp_macs | |
| } | |
| ) | |
| return data, return_print | |
| def extract_from_url(name: str): | |
| "Checks if `name` is a URL, and if so converts it to a model name" | |
| is_url = False | |
| try: | |
| result = urlparse(name) | |
| is_url = all([result.scheme, result.netloc]) | |
| except Exception: | |
| is_url = False | |
| # Pass through if not a URL | |
| if not is_url: | |
| return name | |
| else: | |
| path = result.path | |
| return path[1:] | |
| def translate_llama2(text): | |
| "Translates llama-2 to its hf counterpart" | |
| if not text.endswith("-hf"): | |
| return text + "-hf" | |
| return text | |
| def get_mode_from_hf(model_name: str, library: str, access_token: str): | |
| "Finds and grabs model from the Hub, and initializes on `meta`" | |
| if "meta-llama" in model_name: | |
| model_name = translate_llama2(model_name) | |
| if library == "auto": | |
| library = None | |
| model_name = extract_from_url(model_name) | |
| try: | |
| model = create_empty_model(model_name, library_name=library, trust_remote_code=True, access_token=access_token) | |
| except GatedRepoError: | |
| raise gr.Error( | |
| f"Model `{model_name}` is a gated model, please ensure to pass in your access token and try again if you have access. You can find your access token here : https://huggingface.co/settings/tokens. " | |
| ) | |
| except RepositoryNotFoundError: | |
| raise gr.Error(f"Model `{model_name}` was not found on the Hub, please try another model name.") | |
| except ValueError: | |
| raise gr.Error( | |
| f"Model `{model_name}` does not have any library metadata on the Hub, please manually select a library_name to use (such as `transformers`)" | |
| ) | |
| except (RuntimeError, OSError) as e: | |
| library = check_has_model(e) | |
| if library != "unknown": | |
| raise gr.Error( | |
| f"Tried to load `{model_name}` with `{library}` but a possible model to load was not found inside the repo." | |
| ) | |
| raise gr.Error( | |
| f"Model `{model_name}` had an error, please open a discussion on the model's page with the error message and name: `{e}`" | |
| ) | |
| except ImportError: | |
| # hacky way to check if it works with `trust_remote_code=False` | |
| model = create_empty_model( | |
| model_name, library_name=library, trust_remote_code=False, access_token=access_token | |
| ) | |
| except Exception as e: | |
| raise gr.Error( | |
| f"Model `{model_name}` had an error, please open a discussion on the model's page with the error message and name: `{e}`" | |
| ) | |
| return model | |