Update src/pipeline.py

src/pipeline.py

@@ -1,44 +1,39 @@
+#7
+from huggingface_hub.constants import HF_HUB_CACHE
+from transformers import T5EncoderModel, T5TokenizerFast, CLIPTokenizer, CLIPTextModel
+import torch
+import torch._dynamo
 import gc
 import os
-from typing import TypeAlias
-
-import torch
+from diffusers import FluxPipeline, AutoencoderKL, AutoencoderTiny
 from PIL.Image import Image
-from diffusers import FluxPipeline, FluxTransformer2DModel, AutoencoderTiny, DiffusionPipeline
-from huggingface_hub.constants import HF_HUB_CACHE
 from pipelines.models import TextToImageRequest
 from torch import Generator
-from transformers import T5EncoderModel, CLIPTextModel
+from diffusers import FluxTransformer2DModel, DiffusionPipeline
+from torchao.quantization import quantize_, int8_weight_only, fpx_weight_only
 
+os.environ['PYTORCH_CUDA_ALLOC_CONF']="expandable_segments:True"
 os.environ["TOKENIZERS_PARALLELISM"] = "True"
 torch._dynamo.config.suppress_errors = True
-os.environ['PYTORCH_CUDA_ALLOC_CONF']="expandable_segments:True"
 
 Pipeline = None
-
-ids = "slobers/Flux.1.Schnella"
-Revision = "e34d670e44cecbbc90e4962e7aada2ac5ce8b55b"
-
+ids = "black-forest-labs/FLUX.1-schnell"
+Revision = "741f7c3ce8b383c54771c7003378a50191e9efe9"
 
 def load_pipeline() -> Pipeline:
-    text_encoder = CLIPTextModel.from_pretrained(ids, revision=Revision, subfolder="text_encoder", local_files_only=True, torch_dtype=torch.bfloat16)
-
-    text_encoder_2 = T5EncoderModel.from_pretrained(ids, revision=Revision, subfolder="text_encoder_2", torch_dtype=torch.bfloat16)
-
-    vae = AutoencoderTiny.from_pretrained(ids, revision=Revision, subfolder="vae", torch_dtype=torch.bfloat16)
-
-    path = os.path.join(HF_HUB_CACHE, "models--slobers--Flux.1.Schnella/snapshots/e34d670e44cecbbc90e4962e7aada2ac5ce8b55b/transformer")
-
-    transformer = FluxTransformer2DModel.from_pretrained(path, torch_dtype=torch.bfloat16, use_safetensors=False)
-
-    pipeline = DiffusionPipeline.from_pretrained(ids, vae=vae, revision=Revision, transformer=transformer, text_encoder_2=text_encoder_2, torch_dtype=torch.bfloat16)
+    vae = AutoencoderKL.from_pretrained(ids,revision=Revision, subfolder="vae", local_files_only=True, torch_dtype=torch.bfloat16,)
+    quantize_(vae, int8_weight_only())
+    text_encoder_2 = T5EncoderModel.from_pretrained("city96/t5-v1_1-xxl-encoder-bf16", revision = "1b9c856aadb864af93c1dcdc226c2774fa67bc86", torch_dtype=torch.bfloat16).to(memory_format=torch.channels_last)
+    path = os.path.join(HF_HUB_CACHE, "models--RobertML--FLUX.1-schnell-int8wo/snapshots/307e0777d92df966a3c0f99f31a6ee8957a9857a")
+    transformer = FluxTransformer2DModel.from_pretrained(path, torch_dtype=torch.bfloat16, use_safetensors=False).to(memory_format=torch.channels_last)
+    pipeline = DiffusionPipeline.from_pretrained(ids, revision=Revision, transformer=transformer, text_encoder_2=text_encoder_2, torch_dtype=torch.bfloat16,)
     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=5, max_sequence_length=256)
-
+    for _ in range(3):
+        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.no_grad()
 def infer(request: TextToImageRequest, pipeline: Pipeline) -> Image:
     generator = Generator(pipeline.device).manual_seed(request.seed)
 
@@ -50,4 +45,4 @@ def infer(request: TextToImageRequest, pipeline: Pipeline) -> Image:
         max_sequence_length=256,
         height=request.height,
         width=request.width,
-    ).images[0]
+    ).images[0]