model/model_manager.py

import concurrent.futures 
import random
import gradio as gr
import requests, os
import io, base64, json
import spaces
import torch
from PIL import Image
from openai import OpenAI
from .models import IMAGE_GENERATION_MODELS, VIDEO_GENERATION_MODELS, B2I_MODELS, load_pipeline
from serve.upload import get_random_mscoco_prompt, get_random_video_prompt, get_ssh_random_video_prompt, get_ssh_random_image_prompt
from serve.constants import SSH_CACHE_OPENSOURCE, SSH_CACHE_ADVANCE, SSH_CACHE_PIKA, SSH_CACHE_SORA, SSH_CACHE_IMAGE


class ModelManager:
    def __init__(self):
        self.model_ig_list = IMAGE_GENERATION_MODELS
        self.model_ie_list = [] #IMAGE_EDITION_MODELS
        self.model_vg_list = VIDEO_GENERATION_MODELS
        self.model_b2i_list = B2I_MODELS
        self.loaded_models = {}

    def load_model_pipe(self, model_name):
        if not model_name in self.loaded_models:
            pipe = load_pipeline(model_name)
            self.loaded_models[model_name] = pipe
        else:
            pipe = self.loaded_models[model_name]
        return pipe
    
    @spaces.GPU(duration=120)
    def generate_image_ig(self, prompt, model_name):
        pipe = self.load_model_pipe(model_name)
        if 'Stable-cascade' not in model_name:
            result = pipe(prompt=prompt).images[0]
        else:
            prior, decoder = pipe
            prior.enable_model_cpu_offload()
            prior_output = prior(
                prompt=prompt,
                height=512,
                width=512,
                negative_prompt='',
                guidance_scale=4.0,
                num_images_per_prompt=1,
                num_inference_steps=20
            )
            decoder.enable_model_cpu_offload()
            result = decoder(
                image_embeddings=prior_output.image_embeddings.to(torch.float16),
                prompt=prompt,
                negative_prompt='',
                guidance_scale=0.0,
                output_type="pil",
                num_inference_steps=10
            ).images[0]
        return result

    def generate_image_ig_api(self, prompt, model_name):
        pipe = self.load_model_pipe(model_name)
        result = pipe(prompt=prompt)
        return result

    def generate_image_ig_parallel_anony(self, prompt, model_A, model_B, model_C, model_D):
        if model_A == "" and model_B == "" and model_C == "" and model_D == "":
            from .matchmaker import matchmaker
            not_run = [20,21,22, 25,26, 30] #12,13,14,15,16,17,18,19,20,21,22, #23,24,
            model_ids = matchmaker(num_players=len(self.model_ig_list), not_run=not_run)
            print(model_ids)
            model_names = [self.model_ig_list[i] for i in model_ids]
            print(model_names)
        else:
            model_names = [model_A, model_B, model_C, model_D]

        with concurrent.futures.ThreadPoolExecutor() as executor:
            futures = [executor.submit(self.generate_image_ig, prompt, model) if model.startswith("huggingface")
                    else executor.submit(self.generate_image_ig_api, prompt, model) for model in model_names]
            results = [future.result() for future in futures]

        return results[0], results[1], results[2], results[3], \
            model_names[0], model_names[1], model_names[2], model_names[3]
    
    def generate_image_b2i(self, prompt, grounding_instruction, bbox, model_name):
        pipe = self.load_model_pipe(model_name)
        if model_name == "local_MIGC_b2i":
            from model_bbox.MIGC.inference_single_image import inference_image
            result = inference_image(pipe, prompt, grounding_instruction, bbox)
        elif model_name == "huggingface_ReCo_b2i":
            from model_bbox.ReCo.inference import inference_image
            result = inference_image(pipe, prompt, grounding_instruction, bbox)
        elif model_name == "huggingface_Gligen_b2i":
            from model_bbox.Gligen.inference import inference_image
            result = inference_image(pipe, prompt, grounding_instruction, bbox)
        return result
    

    def generate_image_b2i_parallel_anony(self, prompt, grounding_instruction, bbox, model_A, model_B, model_C, model_D):
            if model_A == "" and model_B == "" and model_C == "" and model_D == "":
                from .matchmaker import matchmaker
                not_run = [] #12,13,14,15,16,17,18,19,20,21,22, #23,24,
                # model_ids = matchmaker(num_players=len(self.model_ig_list), not_run=not_run)
                model_ids = [0, 1, 2]
                print(model_ids)
                model_names = [self.model_b2i_list[i] for i in model_ids]
                print(model_names)
            else:
                model_names = [model_A, model_B, model_C, model_D]

            from concurrent.futures import ProcessPoolExecutor
            with ProcessPoolExecutor() as executor:
                futures = [executor.submit(self.generate_image_b2i, prompt, grounding_instruction, bbox, model) 
                        for model in model_names]
                results = [future.result() for future in futures]

            # with concurrent.futures.ThreadPoolExecutor() as executor:
            #     futures = [executor.submit(self.generate_image_b2i, prompt, grounding_instruction, bbox, model) for model in model_names]
            #     results = [future.result() for future in futures]

            blank_image = None
            final_results = []
            for i in range(4):
                if i < len(model_ids):
                    # 如果是有效模型，返回相应的生成结果
                    final_results.append(results[i])
                else:
                    # 如果没有生成结果，则返回空白图像
                    final_results.append(blank_image)
            final_model_names = []
            for i in range(4):
                if i < len(model_ids):
                    final_model_names.append(model_names[i])
                else:
                    final_model_names.append("")

            return final_results[0], final_results[1], final_results[2], final_results[3], \
                final_model_names[0], final_model_names[1], final_model_names[2], final_model_names[3]

    def generate_image_ig_cache_anony(self, model_A, model_B, model_C, model_D):
        if model_A == "" and model_B == "" and model_C == "" and model_D == "":
            from .matchmaker import matchmaker
            not_run = [20,21,22]
            model_ids = matchmaker(num_players=len(self.model_ig_list), not_run=not_run)
            print(model_ids)
            model_names = [self.model_ig_list[i] for i in model_ids]
            print(model_names)
        else:
            model_names = [model_A, model_B, model_C, model_D]

        root_dir = SSH_CACHE_IMAGE
        local_dir = "./cache_image"
        if not os.path.exists(local_dir):
            os.makedirs(local_dir)
        prompt, results = get_ssh_random_image_prompt(root_dir, local_dir, model_names)

        return results[0], results[1], results[2], results[3], \
            model_names[0], model_names[1], model_names[2], model_names[3], prompt

    def generate_video_vg_parallel_anony(self, model_A, model_B, model_C, model_D):
        if model_A == "" and model_B == "" and model_C == "" and model_D == "":
            # model_names = random.sample([model for model in self.model_vg_list], 4)

            from .matchmaker_video import matchmaker_video
            model_ids = matchmaker_video(num_players=len(self.model_vg_list))
            print(model_ids)
            model_names = [self.model_vg_list[i] for i in model_ids]
            print(model_names)
        else:
            model_names = [model_A, model_B, model_C, model_D]
        
        root_dir = SSH_CACHE_OPENSOURCE
        for name in model_names:
            if "Runway-Gen3" in name or "Runway-Gen2" in name or "Pika-v1.0" in name:
                root_dir = SSH_CACHE_ADVANCE
            elif "Pika-beta" in name:
                root_dir = SSH_CACHE_PIKA      
            elif "Sora" in name and "OpenSora" not in name: 
                root_dir = SSH_CACHE_SORA  
        
        local_dir = "./cache_video"
        if not os.path.exists(local_dir):
            os.makedirs(local_dir)
        prompt, results = get_ssh_random_video_prompt(root_dir, local_dir, model_names)
        cache_dir = local_dir

        return results[0], results[1], results[2], results[3], \
            model_names[0], model_names[1], model_names[2], model_names[3], prompt, cache_dir

    def generate_image_ig_museum_parallel_anony(self, model_A, model_B, model_C, model_D):
        if model_A == "" and model_B == "" and model_C == "" and model_D == "":
            # model_names = random.sample([model for model in self.model_ig_list], 4)

            from .matchmaker import matchmaker
            model_ids = matchmaker(num_players=len(self.model_ig_list))
            print(model_ids)
            model_names = [self.model_ig_list[i] for i in model_ids]
            print(model_names)
        else:
            model_names = [model_A, model_B, model_C, model_D]

        prompt = get_random_mscoco_prompt()
        print(prompt)

        with concurrent.futures.ThreadPoolExecutor() as executor:
            futures = [executor.submit(self.generate_image_ig, prompt, model) if model.startswith("huggingface")
                       else executor.submit(self.generate_image_ig_api, prompt, model) for model in model_names]
            results = [future.result() for future in futures]

        return results[0], results[1], results[2], results[3], \
            model_names[0], model_names[1], model_names[2], model_names[3], prompt

    def generate_image_ig_parallel(self, prompt, model_A, model_B):
        model_names = [model_A, model_B]
        with concurrent.futures.ThreadPoolExecutor() as executor:
            futures = [executor.submit(self.generate_image_ig, prompt, model) if model.startswith("imagenhub")
                       else executor.submit(self.generate_image_ig_api, prompt, model) for model in model_names]
            results = [future.result() for future in futures]
        return results[0], results[1]

    @spaces.GPU(duration=200)
    def generate_image_ie(self, textbox_source, textbox_target, textbox_instruct, source_image, model_name):
        pipe = self.load_model_pipe(model_name)
        result = pipe(src_image = source_image, src_prompt = textbox_source, target_prompt = textbox_target, instruct_prompt = textbox_instruct)
        return result

    def generate_image_ie_parallel(self, textbox_source, textbox_target, textbox_instruct, source_image, model_A, model_B):
        model_names = [model_A, model_B]
        with concurrent.futures.ThreadPoolExecutor() as executor:
            futures = [
                executor.submit(self.generate_image_ie, textbox_source, textbox_target, textbox_instruct, source_image,
                                model) for model in model_names]
            results = [future.result() for future in futures]
        return results[0], results[1]

    def generate_image_ie_parallel_anony(self, textbox_source, textbox_target, textbox_instruct, source_image, model_A, model_B):
        if model_A == "" and model_B == "":
            model_names = random.sample([model for model in self.model_ie_list], 2)
        else:
            model_names = [model_A, model_B]
        with concurrent.futures.ThreadPoolExecutor() as executor:
            futures = [executor.submit(self.generate_image_ie, textbox_source, textbox_target, textbox_instruct, source_image, model) for model in model_names]
            results = [future.result() for future in futures]
        return results[0], results[1], model_names[0], model_names[1]