File size: 2,961 Bytes
de54bfa 6bc8605 59e7d86 6bc8605 de54bfa 6bc8605 de54bfa 6bc8605 de54bfa 6bc8605 de54bfa 6bc8605 de54bfa 6bc8605 de54bfa 6bc8605 de54bfa 6bc8605 de54bfa |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 |
from diffusers import AutoencoderKL
from diffusers.image_processor import VaeImageProcessor
import torch
import torch._dynamo
import gc
from PIL import Image
from pipelines.models import TextToImageRequest
from torch import Generator
from diffusers import DiffusionPipeline
from torchao.quantization import quantize_, int8_weight_only
Pipeline = None
MODEL_ID = "black-forest-labs/FLUX.1-schnell"
def clear():
gc.collect()
torch.cuda.empty_cache()
torch.cuda.reset_max_memory_allocated()
torch.cuda.reset_peak_memory_stats()
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) #q
qtensor = torch.clamp(qtensor, self.qmin, self.qmax)
tensor_q = (qtensor - zero_point) * scale #d
return tensor_q, scale, zero_point
class SDXLQuantization:
def __init__(self, model, bit_number=16):
self.model = model
self.quant = BasicQuantization(bit_number)
def quantize_model(self, save_name=None):
quantized_layers_state = {}
for name, module in self.model.named_modules():
if isinstance(module, (torch.nn.Linear)): # nn.Conv2d
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 load_pipeline() -> Pipeline:
clear()
dtype, device = torch.bfloat16, "cuda"
vae = AutoencoderKL.from_pretrained(
MODEL_ID, subfolder="vae", torch_dtype=torch.bfloat16
)
instance = SDXLQuantization(vae, 9)
instance.quantize_model()
pipeline = DiffusionPipeline.from_pretrained(
MODEL_ID,
vae=vae,
torch_dtype=dtype,
)
pipeline.enable_sequential_cpu_offload()
for _ in range(2):
pipeline(prompt="unpervaded, unencumber, froggish, groundneedle, transnatural, fatherhood, outjump, cinerator", width=1024, height=1024, guidance_scale=0.1, num_inference_steps=4, max_sequence_length=256)
clear()
return pipeline
@torch.inference_mode()
def infer(request: TextToImageRequest, pipeline: Pipeline) -> Image:
clear()
generator = Generator("cuda").manual_seed(request.seed)
image=pipeline(request.prompt,generator=generator, guidance_scale=0.0, num_inference_steps=4, max_sequence_length=256, height=request.height, width=request.width, output_type="pil").images[0]
return image |