Update app.py

app.py

@@ -7,6 +7,18 @@
 import spaces
 
 import os
+
+os.putenv('PYTORCH_NVML_BASED_CUDA_CHECK','1')
+os.putenv('TORCH_LINALG_PREFER_CUSOLVER','1')
+alloc_conf_parts = [
+    'expandable_segments:True',
+    'pinned_use_background_threads:True'  # Specific to pinned memory.
+]
+os.environ['PYTORCH_CUDA_ALLOC_CONF'] = ','.join(alloc_conf_parts)
+os.environ["SAFETENSORS_FAST_GPU"] = "1"
+os.putenv('HF_HUB_ENABLE_HF_TRANSFER','1')
+
+
 import random
 import uuid
 import gradio as gr
@@ -28,25 +40,20 @@ from gradio import themes
 from image_gen_aux import UpscaleWithModel
 #from diffusers.models.attention_processor import AttnProcessor2_0
 
-torch.backends.cuda.matmul.allow_tf32 = False
+torch.backends.cuda.matmul.allow_tf32 = True
 torch.backends.cuda.matmul.allow_bf16_reduced_precision_reduction = False
 torch.backends.cuda.matmul.allow_fp16_reduced_precision_reduction = False
-torch.backends.cudnn.allow_tf32 = False
+torch.backends.cudnn.allow_tf32 = True
 torch.backends.cudnn.deterministic = False
-torch.backends.cudnn.benchmark = False
-#torch.backends.cuda.preferred_blas_library="cublas"
-# torch.backends.cuda.preferred_linalg_library="cusolver"
-
+torch.backends.cudnn.benchmark = True
+torch.backends.cuda.preferred_blas_library="cublas"
+torch.backends.cuda.preferred_linalg_library="cusolver"
 torch.set_float32_matmul_precision("highest")
-os.putenv("HF_HUB_ENABLE_HF_TRANSFER","1")
-
-os.environ["SAFETENSORS_FAST_GPU"] = "1"
 FTP_HOST = "1ink.us"
 FTP_USER = "ford442"
 FTP_PASS = os.getenv("FTP_PASS")
 FTP_DIR = "1ink.us/stable_diff/"  # Remote directory on FTP server
 
-    
 DESCRIPTIONXX = """
     ## ⚡⚡⚡⚡ REALVISXL V5.0 BF16 (Tester A) ⚡⚡⚡⚡
 """
@@ -116,7 +123,7 @@ def load_and_prepare_model():
     #vaeX = AutoencoderKL.from_pretrained('ford442/Juggernaut-XI-v11-fp32',subfolder='vae')   #  ,use_safetensors=True  FAILS
     #unetX = UNet2DConditionModel.from_pretrained('ford442/RealVisXL_V5.0_BF16',subfolder='unet').to(torch.bfloat16)   #  ,use_safetensors=True  FAILS
     #sched = EulerAncestralDiscreteScheduler.from_pretrained("SG161222/RealVisXL_V5.0", subfolder='scheduler',beta_schedule="scaled_linear", steps_offset=1,timestep_spacing="trailing"))
-    sched = EulerAncestralDiscreteScheduler.from_pretrained('ford442/RealVisXL_V5.0_BF16', subfolder='scheduler',beta_schedule="scaled_linear", beta_start=0.00085, beta_end=0.012, steps_offset=1,use_karras_sigmas=False)
+    sched = EulerAncestralDiscreteScheduler.from_pretrained('ford442/RealVisXL_V5.0_BF16', subfolder='scheduler') #,beta_schedule="scaled_linear", beta_start=0.00085, beta_end=0.012, steps_offset=1,use_karras_sigmas=False)
     #sched = EulerAncestralDiscreteScheduler.from_pretrained('ford442/RealVisXL_V5.0_BF16', subfolder='scheduler',beta_schedule="scaled_linear")
     #sched = DPMSolverSDEScheduler.from_pretrained('ford442/RealVisXL_V5.0_BF16', subfolder='scheduler')
     #pipeX = StableDiffusionXLPipeline.from_pretrained("SG161222/RealVisXL_V5.0").to(torch.bfloat16)
@@ -214,7 +221,8 @@ def load_and_prepare_model():
     #pipe.unet.set_default_attn_processor()
     
     #**** BETTER WAY ****#
-    pipe.to(device, torch.bfloat16)
+    #pipe.to(device, torch.bfloat16)
+    pipe.to(device)
     #**** BETTER WAY ****#
     
     #pipe.to(device)
@@ -326,7 +334,8 @@ def generate_30(
     #gc.collect()
     torch.set_float32_matmul_precision("medium")
     with torch.no_grad():
-        upscale = upscaler(rv_image, tiling=True, tile_width=256, tile_height=256)
+        upscalea = upscaler(rv_image, tiling=True, tile_width=256, tile_height=256)
+        upscale = upscaler(upscalea, tiling=True, tile_width=256, tile_height=256)
     downscale1 = upscale.resize((upscale.width // 4, upscale.height // 4), Image.LANCZOS)
     downscale_path = f"rv50_upscale_{timestamp}.png"
     downscale1.save(downscale_path,optimize=False,compress_level=0)
@@ -382,7 +391,8 @@ def generate_60(
     #gc.collect()
     torch.set_float32_matmul_precision("medium")
     with torch.no_grad():
-        upscale = upscaler(rv_image, tiling=True, tile_width=256, tile_height=256)
+        upscalea = upscaler(rv_image, tiling=True, tile_width=256, tile_height=256)
+        upscale = upscaler(upscalea, tiling=True, tile_width=256, tile_height=256)
     downscale1 = upscale.resize((upscale.width // 4, upscale.height // 4), Image.LANCZOS)
     downscale_path = f"rv50_upscale_{timestamp}.png"
     downscale1.save(downscale_path,optimize=False,compress_level=0)
@@ -438,7 +448,8 @@ def generate_90(
     #gc.collect()
     torch.set_float32_matmul_precision("medium")
     with torch.no_grad():
-        upscale = upscaler(rv_image, tiling=True, tile_width=256, tile_height=256)
+        upscalea = upscaler(rv_image, tiling=True, tile_width=256, tile_height=256)
+        upscale = upscaler(upscalea, tiling=True, tile_width=256, tile_height=256)
     downscale1 = upscale.resize((upscale.width // 4, upscale.height // 4), Image.LANCZOS)
     downscale_path = f"rv50_upscale_{timestamp}.png"
     downscale1.save(downscale_path,optimize=False,compress_level=0)