| 32 | top_k: int = 1200 |
| 33 | |
| 34 | class TextToImageInference: |
| 35 | def __init__(self, config: T2IConfig): |
| 36 | self.config = config |
| 37 | self.device = torch.device(config.device) |
| 38 | self._load_models() |
| 39 | |
| 40 | def _load_models(self): |
| 41 | self.model = blip3oQwenForInferenceLM.from_pretrained(self.config.model_path, torch_dtype=self.config.dtype).to(self.device) |
| 42 | self.tokenizer = AutoTokenizer.from_pretrained(self.config.model_path) |
| 43 | |
| 44 | def generate_image(self, prompt: str,cfg_guidance,num_step) -> Image.Image: |
| 45 | |
| 46 | batch_messages = [] |
| 47 | |
| 48 | |
| 49 | messages = [ |
| 50 | {"role": "system", "content": "You are a helpful assistant."}, |
| 51 | {"role": "user", "content": f"Please generate image based on the following caption: {prompt}"} |
| 52 | ] |
| 53 | input_text = self.tokenizer.apply_chat_template( |
| 54 | messages, |
| 55 | tokenize=False, |
| 56 | add_generation_prompt=True) |
| 57 | input_text += f"<im_start><S{self.config.scale}>" |
| 58 | |
| 59 | batch_messages.append(input_text) |
| 60 | |
| 61 | # tokenize as a batch |
| 62 | inputs = self.tokenizer(batch_messages, return_tensors="pt", padding=True, truncation=True, padding_side="left") |
| 63 | _, output_image = self.model.generate_images( |
| 64 | inputs.input_ids.to(self.device), |
| 65 | inputs.attention_mask.to(self.device), |
| 66 | max_new_tokens=self.config.seq_len, |
| 67 | # image_sizes=512, |
| 68 | do_sample=True, |
| 69 | top_p=self.config.top_p, |
| 70 | top_k=self.config.top_k, |
| 71 | guidance_scale=cfg_guidance, |
| 72 | num_inference_steps=num_step,) |
| 73 | |
| 74 | print(output_image) |
| 75 | return output_image[0] |
| 76 | |
| 77 | |
| 78 | def main(): |