inf.py

from __future__ import annotations

import gc
import pathlib
import spaces

import gradio as gr
import PIL.Image
import torch
from diffusers import StableDiffusionXLPipeline
from huggingface_hub import ModelCard

from blora_utils import BLOCKS, filter_lora, scale_lora


class InferencePipeline:
    def __init__(self, hf_token: str | None = None):
        self.hf_token = hf_token
        self.base_model_id = "stabilityai/stable-diffusion-xl-base-1.0"
        self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
        if self.device.type == 'cpu':
            self.pipe = StableDiffusionXLPipeline.from_pretrained(
                self.base_model_id, use_auth_token=self.hf_token)
        else:
            self.pipe = StableDiffusionXLPipeline.from_pretrained(
                self.base_model_id,
                # torch_dtype=torch.float16,
                use_auth_token=self.hf_token)
            self.pipe = self.pipe.to(self.device)
        self.content_lora_model_id = None
        self.style_lora_model_id = None

    def clear(self) -> None:
        self.content_lora_model_id = None
        self.style_lora_model_id = None
        del self.pipe
        self.pipe = None
        torch.cuda.empty_cache()
        gc.collect()

    def load_b_lora_to_unet(self, content_lora_model_id: str, style_lora_model_id: str, content_alpha: float,
                            style_alpha: float) -> None:
        try:
            # Get Content B-LoRA SD
            if content_lora_model_id and content_lora_model_id != 'None':
                content_B_LoRA_sd, _ = self.pipe.lora_state_dict(content_lora_model_id, use_auth_token=self.hf_token)
                content_B_LoRA = filter_lora(content_B_LoRA_sd, BLOCKS['content'])
                content_B_LoRA = scale_lora(content_B_LoRA, content_alpha)
            else:
                content_B_LoRA = {}

            # Get Style B-LoRA SD
            if style_lora_model_id and style_lora_model_id != 'None':
                style_B_LoRA_sd, _ = self.pipe.lora_state_dict(style_lora_model_id, use_auth_token=self.hf_token)
                style_B_LoRA = filter_lora(style_B_LoRA_sd, BLOCKS['style'])
                style_B_LoRA = scale_lora(style_B_LoRA, style_alpha)
            else:
                style_B_LoRA = {}

            # Merge B-LoRAs SD
            res_lora = {**content_B_LoRA, **style_B_LoRA}

            # Load
            self.pipe.load_lora_into_unet(res_lora, None, self.pipe.unet)
        except Exception as e:
            raise type(e)(f'failed to load_b_lora_to_unet, due to: {e}')

    @staticmethod
    def check_if_model_is_local(lora_model_id: str) -> bool:
        return pathlib.Path(lora_model_id).exists()

    @staticmethod
    def get_model_card(model_id: str,
                       hf_token: str | None = None) -> ModelCard:
        if InferencePipeline.check_if_model_is_local(model_id):
            card_path = (pathlib.Path(model_id) / 'README.md').as_posix()
        else:
            card_path = model_id
        return ModelCard.load(card_path, token=hf_token)

    @staticmethod
    def get_base_model_info(lora_model_id: str,
                            hf_token: str | None = None) -> str:
        card = InferencePipeline.get_model_card(lora_model_id, hf_token)
        return card.data.base_model

    def load_pipe(self, content_lora_model_id: str, style_lora_model_id: str, content_alpha: float,
                  style_alpha: float) -> None:
        if content_lora_model_id == self.content_lora_model_id and style_lora_model_id == self.style_lora_model_id:
            return
        self.pipe.unload_lora_weights()

        self.load_b_lora_to_unet(content_lora_model_id, style_lora_model_id, content_alpha, style_alpha)

        self.content_lora_model_id = content_lora_model_id
        self.style_lora_model_id = style_lora_model_id
    
    @spaces.GPU
    def run(
            self,
            content_lora_model_id: str,
            style_lora_model_id: str,
            prompt: str,
            content_alpha: float,
            style_alpha: float,
            seed: int,
            n_steps: int,
            guidance_scale: float,
            num_images_per_prompt: int = 1
    ) -> PIL.Image.Image:
        if not torch.cuda.is_available():
            raise gr.Error('CUDA is not available.')

        print(f"running with {self.device}, {self.pipe.device}")

        self.load_pipe(content_lora_model_id, style_lora_model_id, content_alpha, style_alpha)

        generator = torch.Generator().manual_seed(seed)
        out = self.pipe(
            prompt,
            num_inference_steps=n_steps,
            guidance_scale=guidance_scale,
            generator=generator,
            num_images_per_prompt=num_images_per_prompt,
        )  # type: ignore
        return out.images