Update src/pipeline.py

src/pipeline.py

@@ -19,11 +19,10 @@ os.environ['PYTORCH_CUDA_ALLOC_CONF']="expandable_segments:True"
 torch.backends.cuda.matmul.allow_tf32 = True
 torch.backends.cudnn.enabled = True
 torch.backends.cudnn.benchmark = True
-
+torch.backends.cudnn.deterministic = False
 ckpt_id = "black-forest-labs/FLUX.1-schnell"
 ckpt_revision = "741f7c3ce8b383c54771c7003378a50191e9efe9" 
-
-
+sample = 0
 Pipeline = None
 def empty_cache():
     gc.collect()
@@ -52,15 +51,18 @@ def load_pipeline() -> Pipeline:
         ).to(device)
     quantize_(pipeline.vae, int8_weight_only())
 
-    pipeline(prompt="imprisonable, forechamber, demagogic, monotropic, blandiloquious, blechnoid", width=1024, height=1024, guidance_scale=0.0, num_inference_steps=4, max_sequence_length=256)
-    
+    pipeline(prompt="imprisonable, forechamber, demagogic, monotropic, blandiloquious, blechnoid, blechnoid, blechnoid, blechnoid", width=1024, height=1024, guidance_scale=0.0, num_inference_steps=4, max_sequence_length=256)
     empty_cache()
+    pipeline(prompt="", 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, generator: Generator) -> Image:
-    
+    global sample
+    if not sample:
+        sample = 1
+        empty_cache()
+        
     image=pipeline(request.prompt,
                 generator=generator, 
                 guidance_scale=0.0,