Update src/pipeline.py

src/pipeline.py

@@ -5,19 +5,20 @@ from pipelines.models import TextToImageRequest
 from torch import Generator
 #from time import perf_counter
 import os
-from diffusers import FluxPipeline, AutoencoderKL
+from diffusers import FluxPipeline, AutoencoderKL, AutoencoderTiny
 from diffusers.image_processor import VaeImageProcessor
 from transformers import T5EncoderModel, T5TokenizerFast, CLIPTokenizer, CLIPTextModel
 import diffusers
 #from optimum.quanto import freeze, qfloat8, quantize
 import gc
 from diffusers import FluxTransformer2DModel, DiffusionPipeline
-#from torchao.quantization import quantize_,int8_weight_only
+from torchao.quantization import quantize_,int8_weight_only
 
+os.environ["TOKENIZERS_PARALLELISM"] = "True"
 os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"
 Pipeline = None
 
-ckpt_id = "black-forest-labs/FLUX.1-schnell"
+ckpt_id = "blobers/tx"
 def empty_cache():
     gc.collect()
     torch.cuda.empty_cache()
@@ -27,9 +28,10 @@ def empty_cache():
 def load_pipeline() -> Pipeline:
     empty_cache()
     dtype, device = torch.bfloat16, "cuda"
-    text_encoder_2 = T5EncoderModel.from_pretrained("city96/t5-v1_1-xxl-encoder-bf16", torch_dtype=torch.bfloat16)
-    vae=AutoencoderKL.from_pretrained(ckpt_id, subfolder="vae", torch_dtype=dtype)
-    pipeline = DiffusionPipeline.from_pretrained(ckpt_id, vae=vae, text_encoder_2 = text_encoder_2, torch_dtype=dtype,)
+    vae = AutoencoderTiny.from_pretrained("aifeifei798/taef1", torch_dtype=dtype)
+    quantize_(vae, int8_weight_only())
+    model = FluxTransformer2DModel.from_pretrained("slobers/transgender", torch_dtype=dtype, use_safetensors=False)
+    pipeline = DiffusionPipeline.from_pretrained(ckpt_id, vae=vae, transformer = model, torch_dtype=dtype,)
     torch.backends.cudnn.benchmark = True
     torch.backends.cuda.matmul.allow_tf32 = True
     torch.cuda.set_per_process_memory_fraction(0.90)
@@ -38,7 +40,7 @@ def load_pipeline() -> Pipeline:
     pipeline.vae.to(memory_format=torch.channels_last)
     pipeline.vae.enable_tiling()
     pipeline._exclude_from_cpu_offload = ["vae"]
-    pipeline.enable_sequential_cpu_offload()
+    pipeline.to("cuda")
     for _ in range(2):
         pipeline(prompt="insensible, timbale, pothery, electrovital, actinogram, taxis, intracerebellar, centrodesmus", width=1024, height=1024, guidance_scale=0.0, num_inference_steps=4, max_sequence_length=256)
     empty_cache()