inferencer.py

#!/usr/bin/env python3
# -*- coding: utf-8 -*-

import torch
import math
from PIL import Image
from typing import List, Optional
from transformers import Qwen2_5_VLForConditionalGeneration, Qwen2_5_VLProcessor
from diffusers import FlowMatchEulerDiscreteScheduler, AutoencoderKL, BitsAndBytesConfig
from unipicv2.pipeline_stable_diffusion_3_kontext import StableDiffusion3KontextPipeline
from unipicv2.transformer_sd3_kontext import SD3Transformer2DKontextModel
from unipicv2.stable_diffusion_3_conditioner import StableDiffusion3Conditioner
import spaces

class UniPicV2Inferencer:
    def __init__(
        self,
        model_path: str,
        qwen_vl_path: str,
        quant: str = "int4",  # {"int4", "fp16"}
        image_size: int = 512,
        default_negative_prompt: str = "blurry, low quality, low resolution, distorted, deformed, broken content, missing parts, damaged details, artifacts, glitch, noise, pixelated, grainy, compression artifacts, bad composition, wrong proportion, incomplete editing, unfinished, unedited areas."
    ):
        self.model_path = model_path
        self.qwen_vl_path = qwen_vl_path
        self.quant = quant
        self.image_size = image_size
        self.default_negative_prompt = default_negative_prompt
        self.device = torch.device("cuda")
        self.pipeline = None#self._init_pipeline()

    def _init_pipeline(self) -> StableDiffusion3KontextPipeline:
        print("Initializing UniPicV2 pipeline...")
        
        # ===== 1. Initialize BNB Config =====
        bnb4 = BitsAndBytesConfig(
            load_in_4bit=True,
            bnb_4bit_use_double_quant=True,
            bnb_4bit_quant_type="nf4",
            bnb_4bit_compute_dtype=torch.float16,
        )

        # ===== 2. Load SD3 Transformer =====
        if self.quant == "int4":
            transformer = SD3Transformer2DKontextModel.from_pretrained(
                self.model_path, subfolder="transformer",
                quantization_config=bnb4, device_map="auto", low_cpu_mem_usage=True
            )
        else:
            transformer = SD3Transformer2DKontextModel.from_pretrained(
                self.model_path, subfolder="transformer",
                torch_dtype=torch.float16, device_map="auto", low_cpu_mem_usage=True
            )

        # ===== 3. Load VAE =====
        vae = AutoencoderKL.from_pretrained(
            self.model_path, subfolder="vae",
            torch_dtype=torch.float16, device_map="auto", low_cpu_mem_usage=True
        ).to(self.device)

        # ===== 4. Load Qwen2.5-VL (LMM) =====
        try:
            self.lmm = Qwen2_5_VLForConditionalGeneration.from_pretrained(
                self.qwen_vl_path,
                torch_dtype=torch.float16,
                attn_implementation="flash_attention_2",
                device_map="auto",
            ).to(self.device)
            print("**"*20)
        except Exception:
            self.lmm = Qwen2_5_VLForConditionalGeneration.from_pretrained(
                self.qwen_vl_path,
                torch_dtype=torch.float16,
                attn_implementation="sdpa",
                device_map="auto",
            ).to(self.device)

        # ===== 5. Load Processor =====
        self.processor = Qwen2_5_VLProcessor.from_pretrained(self.qwen_vl_path, use_fast=False)
        
        if hasattr(self.processor, "chat_template") and self.processor.chat_template:
            self.processor.chat_template = self.processor.chat_template.replace(
                "{% if loop.first and message['role'] != 'system' %}<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n{% endif %}",
                ""
            )

        # ===== 6. Load Conditioner =====
        self.conditioner = StableDiffusion3Conditioner.from_pretrained(
            self.model_path, subfolder="conditioner",
            torch_dtype=torch.float16, low_cpu_mem_usage=True
        ).to(self.device)

        # ===== 7. Load Scheduler =====
        scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained(
            self.model_path, subfolder="scheduler"
        )

        # ===== 8. Create Pipeline =====
        pipeline = StableDiffusion3KontextPipeline(
            transformer=transformer,
            vae=vae,
            text_encoder=None,
            tokenizer=None,
            text_encoder_2=None,
            tokenizer_2=None,
            text_encoder_3=None,
            tokenizer_3=None,
            scheduler=scheduler
        )

        try:
            pipeline.enable_vae_slicing()
            pipeline.enable_vae_tiling()
            pipeline.enable_model_cpu_offload()
        except Exception:
            print("Note: Could not enable all memory-saving features")

        print("Pipeline initialization complete!")
        return pipeline

    def _prepare_text_inputs(self, prompt: str, negative_prompt: str = None):
        negative_prompt = negative_prompt or self.default_negative_prompt
        
        messages = [
            [{"role": "user", "content": [{"type": "text", "text": prompt}]}],
            [{"role": "user", "content": [{"type": "text", "text": negative_prompt}]}]
        ]
        
        texts = [
            self.processor.apply_chat_template(m, tokenize=False, add_generation_prompt=True) 
            for m in messages
        ]
        
        inputs = self.processor(
            text=texts,
            images=None,
            padding=True,
            return_tensors="pt"
        )
        
        return inputs

    def _prepare_image_inputs(self, image: Image.Image, prompt: str, negative_prompt: str = None):
        negative_prompt = negative_prompt or self.default_negative_prompt
        
        messages = [
            [{"role": "user", "content": [{"type": "image", "image": image}, {"type": "text", "text": prompt}]}],
            [{"role": "user", "content": [{"type": "image", "image": image}, {"type": "text", "text": negative_prompt}]}]
        ]
        
        texts = [
            self.processor.apply_chat_template(m, tokenize=False, add_generation_prompt=True) 
            for m in messages
        ]
        
        min_pixels = max_pixels = int(image.height * 28 / 32 * image.width * 28 / 32)
        
        inputs = self.processor(
            text=texts,
            images=[image] * 2,
            min_pixels=min_pixels,
            max_pixels=max_pixels,
            padding=True,
            return_tensors="pt"
        )
        
        return inputs

    def _process_inputs(self, inputs: dict, num_queries: int):
        # Ensure all tensors are on the correct device
        inputs = {
            k: v.to(self.device) if isinstance(v, torch.Tensor) else v
            for k, v in inputs.items()
        }
        
        input_ids = inputs["input_ids"]
        attention_mask = inputs["attention_mask"]
        
        # Pad with meta queries
        pad_ids = torch.zeros((input_ids.size(0), num_queries), 
                            dtype=input_ids.dtype, device=self.device)
        pad_mask = torch.ones((attention_mask.size(0), num_queries), 
                            dtype=attention_mask.dtype, device=self.device)
        
        input_ids = torch.cat([input_ids, pad_ids], dim=1)
        attention_mask = torch.cat([attention_mask, pad_mask], dim=1)
        
        # Get input embeddings
        
        # 获取 embedding 权重所在设备
        embed_device = self.lmm.get_input_embeddings().weight.device
        
        # 确保 input_ids 在同一设备
        input_ids = input_ids.to(embed_device)
        
        inputs_embeds = self.lmm.get_input_embeddings()(input_ids)
        
        # Ensure meta queries are on correct device
        self.conditioner.meta_queries.data = self.conditioner.meta_queries.data.to(self.device)
        inputs_embeds[:, -num_queries:] = self.conditioner.meta_queries[None].expand(2, -1, -1)
        
        # Handle image embeddings if present
        if "pixel_values" in inputs:
            image_embeds = self.lmm.visual(
                inputs["pixel_values"].to(self.device), 
                grid_thw=inputs["image_grid_thw"].to(self.device)
            )
            image_token_id = self.processor.tokenizer.convert_tokens_to_ids('<|image_pad|>')
            mask_img = (input_ids == image_token_id)
            inputs_embeds[mask_img] = image_embeds
        
        # Forward through LMM
        if hasattr(self.lmm.model, "rope_deltas"):
            self.lmm.model.rope_deltas = None

        #model_device = self.lmm.model.embed_tokens.weight.device
        # 强制将所有 tensor 输入搬到这个设备
        for k, v in inputs.items():
            if isinstance(v, torch.Tensor):
                inputs[k] = v.to(self.device)
                
        outputs = self.lmm.model(
            inputs_embeds=inputs_embeds.to(self.device),
            attention_mask=attention_mask.to(self.device),
            image_grid_thw=inputs.get("image_grid_thw", None),
            use_cache=False
        )
        
        hidden_states = outputs.last_hidden_state[:, -num_queries:]
        hidden_states = hidden_states.to(self.device)
        
        # Get prompt embeds
        prompt_embeds, pooled_prompt_embeds = self.conditioner(hidden_states)
        
        return {
            "prompt_embeds": prompt_embeds[:1],
            "pooled_prompt_embeds": pooled_prompt_embeds[:1],
            "negative_prompt_embeds": prompt_embeds[1:],
            "negative_pooled_prompt_embeds": pooled_prompt_embeds[1:]
        }

    def _resize_image(self, image: Image.Image, size: int) -> Image.Image:
        w, h = image.size
        if w >= h:
            new_w = size
            new_h = int(h * (new_w / w))
            new_h = (new_h // 32) * 32
        else:
            new_h = size
            new_w = int(w * (new_h / h))
            new_w = (new_w // 32) * 32
            
        return image.resize((new_w, new_h))

    @spaces.GPU(duration=120)
    def generate_image(
        self,
        prompt: str,
        negative_prompt: Optional[str] = None,
        height: Optional[int] = None,
        width: Optional[int] = None,
        num_inference_steps: int = 36,
        guidance_scale: float = 3.0,
        seed: int = 42
    ) -> Image.Image:
        if not self.pipeline:
            self.pipeline = self._init_pipeline()
        height = height or self.image_size
        width = width or self.image_size
        
        inputs = self._prepare_text_inputs(prompt, negative_prompt)
        num_queries = self.conditioner.config.num_queries
        embeds = self._process_inputs(inputs, num_queries)
        
        generator = torch.Generator(device=self.device).manual_seed(seed)
        
        image = self.pipeline(
            prompt_embeds=embeds["prompt_embeds"].to(self.device),
            pooled_prompt_embeds=embeds["pooled_prompt_embeds"].to(self.device),
            negative_prompt_embeds=embeds["negative_prompt_embeds"].to(self.device),
            negative_pooled_prompt_embeds=embeds["negative_pooled_prompt_embeds"].to(self.device),
            height=height,
            width=width,
            num_inference_steps=num_inference_steps,
            guidance_scale=guidance_scale,
            generator=generator
        ).images
        
        return image

    @spaces.GPU(duration=120)
    def edit_image(
        self,
        image: Image.Image,
        prompt: str,
        negative_prompt: Optional[str] = None,
        height: Optional[int] = None,
        width: Optional[int] = None,
        num_inference_steps: int = 36,
        guidance_scale: float = 3.0,
        seed: int = 42
    ) -> Image.Image:
        if not self.pipeline:
            self.pipeline = self._init_pipeline()
        original_size = image.size
        image = self._resize_image(image, self.image_size)
        if height is None or width is None:
            height, width = image.height, image.width
        
        inputs = self._prepare_image_inputs(image, prompt, negative_prompt)
        num_queries = self.conditioner.config.num_queries
        embeds = self._process_inputs(inputs, num_queries)
        
        generator = torch.Generator(device=self.device).manual_seed(seed)
        
        edited_image = self.pipeline(
            image=image,
            prompt_embeds=embeds["prompt_embeds"].to(self.device),
            pooled_prompt_embeds=embeds["pooled_prompt_embeds"].to(self.device),
            negative_prompt_embeds=embeds["negative_prompt_embeds"].to(self.device),
            negative_pooled_prompt_embeds=embeds["negative_pooled_prompt_embeds"].to(self.device),
            height=height,
            width=width,
            num_inference_steps=num_inference_steps,
            guidance_scale=guidance_scale,
            generator=generator
        ).images
        
        return edited_image

    @spaces.GPU(duration=120)
    def understand_image(
        self,
        image: Image.Image,
        prompt: str,
        max_new_tokens: int = 512
    ) -> str:
        """
        Understand the content of an image and answer questions about it.
        
        Args:
            image: Input image to understand
            prompt: Question or instruction about the image
            max_new_tokens: Maximum number of tokens to generate
            
        Returns:
            str: The model's response to the prompt
        """
        # Prepare messages in Qwen-VL format
        if not self.pipeline:
            self.pipeline = self._init_pipeline()
        messages = [
            {
                "role": "user",
                "content": [
                    {"type": "image", "image": image},
                    {"type": "text", "text": prompt},
                ],
            },
        ]
        
        # Apply chat template
        text = self.processor.apply_chat_template(
            messages, 
            tokenize=False, 
            add_generation_prompt=True
        )
        
        # Calculate appropriate image size for processing
        min_pixels = max_pixels = int(image.height * 28 / 32 * image.width * 28 / 32)
        
        # Process inputs
        inputs = self.processor(
            text=[text],
            images=[image],
            min_pixels=min_pixels,
            max_pixels=max_pixels,
            padding=True,
            return_tensors="pt"
        ).to(self.device)
        
        # Generate response
        generated_ids = self.lmm.generate(
            **inputs,
            max_new_tokens=max_new_tokens
        )
        
        # Trim input tokens from output
        generated_ids_trimmed = [
            out_ids[len(in_ids):] 
            for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
        ]
        
        # Decode the response
        output_text = self.processor.batch_decode(
            generated_ids_trimmed,
            skip_special_tokens=True,
            clean_up_tokenization_spaces=False
        )[0]
        
        return output_text