Update src/pipeline.py

src/pipeline.py

@@ -1,26 +1,62 @@
-from transformers import T5EncoderModel, T5TokenizerFast, CLIPTokenizer, CLIPTextModel
-import gc
+import torch
 from PIL.Image import Image
+from diffusers import FluxPipeline
 from pipelines.models import TextToImageRequest
 from torch import Generator
-import torch
-from diffusers import FluxPipeline, AutoencoderTiny
+#from time import perf_counter
+import os
+from diffusers import FluxPipeline, AutoencoderKL
+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
 
+os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"
 Pipeline = None
 
+ckpt_id = "black-forest-labs/FLUX.1-schnell"
+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:
-    pipeline = FluxPipeline.from_pretrained(
-    "black-forest-labs/FLUX.1-schnell", torch_dtype=torch.bfloat16)
-    pipeline.vae = AutoencoderTiny.from_pretrained("madebyollin/taef1", torch_dtype=torch.bfloat16)
-    pipeline.enable_sequential_cpu_offload()
+    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,
+        )
+#    torch.backends.cudnn.benchmark = True
+#    torch.backends.cuda.matmul.allow_tf32 = True
+#    torch.cuda.set_per_process_memory_fraction(0.99)
+#    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()
+#    pipeline.vae = torch.compile(pipeline.vae)
+#    pipeline._exclude_from_cpu_offload = ["vae"]
+    pipeline.enable_sequential_cpu_offload()
     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)
-
     return pipeline
 
+
+@torch.inference_mode()
 def infer(request: TextToImageRequest, pipeline: Pipeline) -> Image:
+#    torch.cuda.reset_peak_memory_stats()
+#    torch.backends.cuda.matmul.allow_tf32 = True
     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)