Update src/pipeline.py

src/pipeline.py

@@ -1,53 +1,45 @@
+from torch import Generator
 import torch
 from PIL.Image import Image
-from diffusers import FluxPipeline
 from pipelines.models import TextToImageRequest
 from torch import Generator
-#from time import perf_counter
+from diffusers import FluxTransformer2DModel, DiffusionPipeline
+import gc
 import os
-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
+import torch._dynamo
+from torchao.quantization import quantize_, int8_weight_only, fpx_weight_only
+os.environ['PYTORCH_CUDA_ALLOC_CONF']="expandable_segments:True"
 HOME = os.environ["HOME"]
-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 = "slobers/tx"
 def empty_cache():
     gc.collect()
     torch.cuda.empty_cache()
     torch.cuda.reset_max_memory_allocated()
     torch.cuda.reset_peak_memory_stats()
-
+    
 def load_pipeline() -> Pipeline:
     empty_cache()
     dtype, device = torch.bfloat16, "cuda"
-    vae = AutoencoderTiny.from_pretrained("aifeifei798/taef1", torch_dtype=dtype)
+    text_encoder = CLIPTextModel.from_pretrained(ckpt_id, subfolder="text_encoder", torch_dtype=torch.bfloat16)
+    quantize_(text_encoder, int8_weight_only())
+    vae = AutoencoderTiny.from_pretrained("RobertML/FLUX.1-schnell-vae_e3m2", torch_dtype=torch.bfloat16)
     quantize_(vae, int8_weight_only())
-    model = FluxTransformer2DModel.from_pretrained(f"{HOME}/.cache/huggingface/hub/models--slobers--transgender/snapshots/cb99836efa0ed55856970269c42fafdaa0e44c5d", 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)
-    pipeline.text_encoder.to(memory_format=torch.channels_last)
-    pipeline.transformer.to(memory_format=torch.channels_last)
-    pipeline.vae.to(memory_format=torch.channels_last)
-    pipeline.vae.enable_tiling()
+    text_encoder_2 = T5EncoderModel.from_pretrained("city96/t5-v1_1-xxl-encoder-bf16", torch_dtype=torch.bfloat16)
+    quantize_(text_encoder_2, int8_weight_only())
+    model = FluxTransformer2DModel.from_pretrained(f"{HOME}/.cache/huggingface/hub/models--slobers--transgender/snapshots/cb99836efa0ed55856970269c42fafdaa0e44c5d", torch_dtype=torch.bfloat16, use_safetensors=False)
+    pipeline = DiffusionPipeline.from_pretrained(ckpt_id, text_encoder=text_encoder, transformer=model, text_encoder_2=text_encoder_2, torch_dtype=torch.bfloat16)
     pipeline.to("cuda")
+    
     for _ in range(2):
+        empty_cache()
         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()
     return pipeline
 
-
-@torch.inference_mode()
 def infer(request: TextToImageRequest, pipeline: Pipeline) -> Image:
+    empty_cache()
     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]
+    image=pipeline(request.prompt,generator=generator, guidance_scale=0.0, num_inference_steps=3, max_sequence_length=256, height=request.height, width=request.width, output_type="pil").images[0]
     return(image)