Update src/pipeline.py

src/pipeline.py

@@ -131,8 +131,8 @@ def load_pipeline() -> Pipeline:
     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)
-    vae = torch.load('/root/.cache/huggingface/hub/compiled_vae.pth')
+    vae=AutoencoderKL.from_pretrained(ckpt_id, subfolder="vae", torch_dtype=dtype)
+    # vae = torch.load('/root/.cache/huggingface/hub/compiled_vae.pth')
     # transformer = FluxTransformer2DModel.from_pretrined("manbeast3b/transfomer-flux-schnell-int8") # torch_dtype=dtype
     pipeline = DiffusionPipeline.from_pretrained(
         ckpt_id,
@@ -144,7 +144,11 @@ def load_pipeline() -> Pipeline:
     # quantize_(pipeline.transformer, float8_dynamic_activation_float8_weight())
     
     torch.backends.cudnn.benchmark = True
+    torch.backends.cudnn.deterministic = False
+    torch.set_deterministic_debug_mode(0) 
     torch.backends.cuda.matmul.allow_tf32 = True
+    torch.backends.cuda.matmul.allow_fp16_reduced_precision_reduction = True
+    # torch.cuda.set_memory_growth(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)
@@ -153,12 +157,13 @@ def load_pipeline() -> Pipeline:
     # pipeline.transformer.save_pretrained("/root/.cache/huggingface/hub/transformer-flux")
     # exit()    
 
-    # pipeline.vae.to(memory_format=torch.channels_last)
-    # pipeline.vae = torch.compile(pipeline.vae)
+    pipeline.vae.to(memory_format=torch.channels_last)
+    pipeline.vae = torch.compile(pipeline.vae)
     # torch.save(pipeline.vae, '/root/.cache/huggingface/hub/compiled_vae.pth')
     # exit()
     
     
+    
     pipeline._exclude_from_cpu_offload = ["vae"]
     pipeline.enable_sequential_cpu_offload()
     for _ in range(2):