| import torch |
| from PIL.Image import Image |
| from diffusers import StableDiffusionXLPipeline |
| import torch.nn as nn |
| from pipelines.models import TextToImageRequest |
| from diffusers import DDIMScheduler |
| from torch import Generator |
| from loss import SchedulerWrapper |
| from onediffx import compile_pipe, save_pipe, load_pipe |
|
|
|
|
|
|
| class BasicQuantization: |
| def __init__(self, bits=1): |
| self.bits = bits |
| self.qmin = -(2**(bits-1)) |
| self.qmax = 2**(bits-1) - 1 |
| |
| def quantize_tensor(self, tensor): |
| scale = (tensor.max() - tensor.min()) / (self.qmax - self.qmin) |
| zero_point = self.qmin - torch.round(tensor.min() / scale) |
| |
| |
| qtensor = torch.round(tensor / scale + zero_point) |
| qtensor = torch.clamp(qtensor, self.qmin, self.qmax) |
| |
| |
| tensor_q = (qtensor - zero_point) * scale |
| return tensor_q, scale, zero_point |
|
|
|
|
| class SDXLQuantization: |
| def __init__(self, model, bit_number=6): |
| self.model = model |
| self.quant = BasicQuantization(bit_number) |
| |
| def quantize_model(self, save_name='quantized_layers.pth'): |
| quantized_layers_state = {} |
| for name, module in self.model.named_modules(): |
| used = False |
| if isinstance(module, (nn.Linear)): |
| |
| if hasattr(module, 'weight'): |
| quantized_weight, _, _ = self.quant.quantize_tensor(module.weight) |
| module.weight = torch.nn.Parameter(quantized_weight) |
| |
| |
| if hasattr(module, 'bias') and module.bias is not None: |
| quantized_bias, _, _ = self.quant.quantize_tensor(module.bias) |
| module.bias = torch.nn.Parameter(quantized_bias) |
| |
|
|
|
|
| def callback_dynamic_cfg(pipe, step_index, timestep, callback_kwargs): |
| if step_index == int(pipe.num_timesteps * 0.78): |
| callback_kwargs['prompt_embeds'] = callback_kwargs['prompt_embeds'].chunk(2)[-1] |
| callback_kwargs['add_text_embeds'] = callback_kwargs['add_text_embeds'].chunk(2)[-1] |
| callback_kwargs['add_time_ids'] = callback_kwargs['add_time_ids'].chunk(2)[-1] |
| pipe._guidance_scale = 0.1 |
|
|
| return callback_kwargs |
|
|
| def load_pipeline(pipeline=None) -> StableDiffusionXLPipeline: |
| if not pipeline: |
| pipeline = StableDiffusionXLPipeline.from_pretrained( |
| "stablediffusionapi/newdream-sdxl-20", |
| torch_dtype=torch.float16, |
| ).to("cuda") |
|
|
| quantizer = SDXLQuantization(pipeline.unet, 14) |
| quantizer.quantize_model() |
| pipeline.scheduler = SchedulerWrapper(DDIMScheduler.from_config(pipeline.scheduler.config)) |
|
|
| pipeline = compile_pipe(pipeline) |
| load_pipe(pipeline, dir="/home/sandbox/.cache/huggingface/hub/models--RobertML--cached-pipe-02/snapshots/58d70deae87034cce351b780b48841f9746d4ad7") |
|
|
| for _ in range(1): |
| deepcache_output = pipeline(prompt="telestereography, unstrengthen, preadministrator, copatroness, hyperpersonal, paramountness, paranoid, guaniferous", output_type="pil", num_inference_steps=20) |
| pipeline.scheduler.prepare_loss() |
| for _ in range(2): |
| pipeline(prompt="telestereography, unstrengthen, preadministrator, copatroness, hyperpersonal, paramountness, paranoid, guaniferous", output_type="pil", num_inference_steps=20) |
| return pipeline |
|
|
| def infer(request: TextToImageRequest, pipeline: StableDiffusionXLPipeline) -> Image: |
| if request.seed is None: |
| generator = None |
| else: |
| generator = Generator(pipeline.device).manual_seed(request.seed) |
|
|
| return pipeline( |
| prompt=request.prompt, |
| negative_prompt=request.negative_prompt, |
| width=request.width, |
| height=request.height, |
| generator=generator, |
| num_inference_steps=17, |
| cache_interval=1, |
| cache_layer_id=1, |
| cache_block_id=0, |
| eta=1.0, |
| guidance_scale = 5.0, |
| guidance_rescale = 0.0, |
| callback_on_step_end=callback_dynamic_cfg, |
| callback_on_step_end_tensor_inputs=['prompt_embeds', 'add_text_embeds', 'add_time_ids'], |
| ).images[0] |
|
|
|
|
