Upload src/pipeline.py with huggingface_hub

src/pipeline.py

@@ -10,15 +10,18 @@ from pipelines.models import TextToImageRequest
 from torch import Generator
 from torchao.quantization import quantize_, int8_weight_only
 from transformers import T5EncoderModel, CLIPTextModel, logging 
-
+import torch._dynamo
+torch._dynamo.config.suppress_errors = True
 
 Pipeline: TypeAlias = FluxPipeline
-torch.backends.cudnn.benchmark = True
+
 torch.backends.cudnn.benchmark = True
 torch._inductor.config.conv_1x1_as_mm = True
 torch._inductor.config.coordinate_descent_tuning = True
 torch._inductor.config.epilogue_fusion = False
 torch._inductor.config.coordinate_descent_check_all_directions = True
+
+
 os.environ['PYTORCH_CUDA_ALLOC_CONF']="expandable_segments:True"
 os.environ["TOKENIZERS_PARALLELISM"] = "True"
 CHECKPOINT = "jokerbit/flux.1-schnell-Robert-int8wo"
@@ -29,29 +32,38 @@ TinyVAE_REV = "2d552378e58c9c94201075708d7de4e1163b2689"
 
 
 def load_pipeline() -> Pipeline:
-    path = os.path.join(HF_HUB_CACHE, "models--jokerbit--flux.1-schnell-Robert-int8wo/snapshots/5ef0012f11a863e5111ec56540302a023bc8587b/transformer")
+    path = os.path.join(HF_HUB_CACHE, "models--jokerbit--flux.1-schnell-Robert-int8wo/snapshots/5ef0012f11a863e5111ec56540302a023bc8587b/transformer") 
     transformer = FluxTransformer2DModel.from_pretrained(
             path, 
             use_safetensors=False,
             local_files_only=True,
             torch_dtype=torch.bfloat16)
+    vae = AutoencoderTiny.from_pretrained(
+            TinyVAE,
+            TinyVAE_REV,
+            local_files_only=True,
+            torch_dtype=torch.bfloat16)
 
     pipeline = FluxPipeline.from_pretrained(
         CHECKPOINT,
         revision=REVISION,
         transformer=transformer,
+        vae=vae,
         local_files_only=True,
         torch_dtype=torch.bfloat16,
     ).to("cuda")
 
     pipeline.transformer.to(memory_format=torch.channels_last)
-    # pipeline.vae.to(memory_format=torch.channels_last)
-    # quantize_(pipeline.vae, int8_weight_only())
+    pipeline.transformer = torch.compile(pipeline.transformer, mode="max-autotune", fullgraph=False)
+    pipeline.vae.to(memory_format=torch.channels_last)
+    quantize_(pipeline.vae, int8_weight_only())
     pipeline.vae = torch.compile(pipeline.vae, fullgraph=True, mode="max-autotune")
-      
-    for _ in range(2):
-        pipeline("cat", num_inference_steps=4)
-
+    
+    PROMPT = 'semiconformity, peregrination, quip, twineless, emotionless, tawa, depickle'
+    with torch.inference_mode():  
+        for _ in range(4):
+            pipeline(prompt="onomancy, aftergo, spirantic, Platyhelmia, modificator, drupaceous, jobbernowl, hereness", width=1024, height=1024, guidance_scale=0.0, num_inference_steps=4, max_sequence_length=256)
+    torch.cuda.empty_cache()
     return pipeline
 
 @torch.inference_mode()
@@ -75,13 +87,14 @@ if __name__ == "__main__":
             height=None,
             width=None,
             seed=666)
+    generator = torch.Generator(device="cuda")
     start_time = perf_counter()
     pipe_ = load_pipeline()
     stop_time = perf_counter()
     print(f"Pipeline is loaded in {stop_time - start_time}s")
     for _ in range(4):
         start_time = perf_counter()
-        infer(request, pipe_)
+        infer(request, pipe_, generator=generator.manual_seed(request.seed))
         stop_time = perf_counter()
         print(f"Request in {stop_time - start_time}s")