Update src/pipeline.py

src/pipeline.py

@@ -25,17 +25,14 @@ def load_pipeline() -> Pipeline:
     dtype, device = torch.bfloat16, "cuda"
     
     vae = AutoencoderTiny.from_pretrained("manbeast3b/quantized1", torch_dtype=DTYPE)
-    ############ Text Encoder ############
-    text_encoder = CLIPTextModel.from_pretrained(
-        MODEL_ID, subfolder="text_encoder", torch_dtype=DTYPE
-    )
-    ############ Text Encoder 2 ############
     text_encoder_2 = T5EncoderModel.from_pretrained(
         "city96/t5-v1_1-xxl-encoder-bf16", torch_dtype=DTYPE
     )
-
-    pipeline = FluxPipeline.from_pretrained(MODEL_ID,vae=vae,text_encoder=text_encoder,text_encoder_2=text_encoder_2,
+    pipeline = FluxPipeline.from_pretrained(MODEL_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.text_encoder_2.to(memory_format=torch.channels_last)
     pipeline.transformer.to(memory_format=torch.channels_last)
@@ -43,14 +40,18 @@ def load_pipeline() -> Pipeline:
     pipeline.vae = torch.compile(pipeline.vae)
     pipeline._exclude_from_cpu_offload = ["vae"]
     pipeline.enable_sequential_cpu_offload()
+    clear()
     for _ in range(1):
         pipeline(prompt="unpervaded, unencumber, froggish, groundneedle, transnatural, fatherhood, outjump, cinerator", width=1024, height=1024, guidance_scale=0.1, num_inference_steps=4, max_sequence_length=256)
-    clear()
     return pipeline
 
+sample = True
 @torch.inference_mode()
 def infer(request: TextToImageRequest, pipeline: Pipeline) -> Image:
-    clear()
+    global sample
+    if sample:
+        clear()
+        sample = None
     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
+    return(image)