Update app.py
Browse files
app.py
CHANGED
|
@@ -16,12 +16,12 @@ import numpy as np
|
|
| 16 |
import random
|
| 17 |
|
| 18 |
import torch
|
| 19 |
-
torch.backends.cuda.matmul.allow_tf32 =
|
| 20 |
torch.backends.cuda.matmul.allow_bf16_reduced_precision_reduction = False
|
| 21 |
torch.backends.cuda.matmul.allow_fp16_reduced_precision_reduction = False
|
| 22 |
-
torch.backends.cudnn.allow_tf32 =
|
| 23 |
torch.backends.cudnn.deterministic = False
|
| 24 |
-
torch.backends.cudnn.benchmark =
|
| 25 |
torch.backends.cuda.preferred_blas_library="cublas"
|
| 26 |
torch.backends.cuda.preferred_linalg_library="cusolver"
|
| 27 |
torch.set_float32_matmul_precision("highest")
|
|
@@ -150,7 +150,7 @@ def infer_60(
|
|
| 150 |
sd35_path = f"sd35ll_{timestamp}.png"
|
| 151 |
sd_image.save(sd35_path,optimize=False,compress_level=0)
|
| 152 |
pyx.upload_to_ftp(sd35_path)
|
| 153 |
-
upscaler_2.to(torch.device('cuda'))
|
| 154 |
with torch.no_grad():
|
| 155 |
upscale = upscaler_2(sd_image, tiling=True, tile_width=256, tile_height=256)
|
| 156 |
upscale2 = upscaler_2(upscale, tiling=True, tile_width=256, tile_height=256)
|
|
@@ -200,7 +200,7 @@ def infer_90(
|
|
| 200 |
sd35_path = f"sd35ll_{timestamp}.png"
|
| 201 |
sd_image.save(sd35_path,optimize=False,compress_level=0)
|
| 202 |
pyx.upload_to_ftp(sd35_path)
|
| 203 |
-
upscaler_2.to(torch.device('cuda'))
|
| 204 |
with torch.no_grad():
|
| 205 |
upscale = upscaler_2(sd_image, tiling=True, tile_width=256, tile_height=256)
|
| 206 |
upscale2 = upscaler_2(upscale, tiling=True, tile_width=256, tile_height=256)
|
|
@@ -250,7 +250,7 @@ def infer_110(
|
|
| 250 |
sd35_path = f"sd35ll_{timestamp}.png"
|
| 251 |
sd_image.save(sd35_path,optimize=False,compress_level=0)
|
| 252 |
pyx.upload_to_ftp(sd35_path)
|
| 253 |
-
upscaler_2.to(torch.device('cuda'))
|
| 254 |
with torch.no_grad():
|
| 255 |
upscale = upscaler_2(sd_image, tiling=True, tile_width=256, tile_height=256)
|
| 256 |
upscale2 = upscaler_2(upscale, tiling=True, tile_width=256, tile_height=256)
|
|
|
|
| 16 |
import random
|
| 17 |
|
| 18 |
import torch
|
| 19 |
+
torch.backends.cuda.matmul.allow_tf32 = True
|
| 20 |
torch.backends.cuda.matmul.allow_bf16_reduced_precision_reduction = False
|
| 21 |
torch.backends.cuda.matmul.allow_fp16_reduced_precision_reduction = False
|
| 22 |
+
torch.backends.cudnn.allow_tf32 = True
|
| 23 |
torch.backends.cudnn.deterministic = False
|
| 24 |
+
torch.backends.cudnn.benchmark = True
|
| 25 |
torch.backends.cuda.preferred_blas_library="cublas"
|
| 26 |
torch.backends.cuda.preferred_linalg_library="cusolver"
|
| 27 |
torch.set_float32_matmul_precision("highest")
|
|
|
|
| 150 |
sd35_path = f"sd35ll_{timestamp}.png"
|
| 151 |
sd_image.save(sd35_path,optimize=False,compress_level=0)
|
| 152 |
pyx.upload_to_ftp(sd35_path)
|
| 153 |
+
upscaler_2.to(torch.device('cuda',non_blocking=True))
|
| 154 |
with torch.no_grad():
|
| 155 |
upscale = upscaler_2(sd_image, tiling=True, tile_width=256, tile_height=256)
|
| 156 |
upscale2 = upscaler_2(upscale, tiling=True, tile_width=256, tile_height=256)
|
|
|
|
| 200 |
sd35_path = f"sd35ll_{timestamp}.png"
|
| 201 |
sd_image.save(sd35_path,optimize=False,compress_level=0)
|
| 202 |
pyx.upload_to_ftp(sd35_path)
|
| 203 |
+
upscaler_2.to(torch.device('cuda',non_blocking=True))
|
| 204 |
with torch.no_grad():
|
| 205 |
upscale = upscaler_2(sd_image, tiling=True, tile_width=256, tile_height=256)
|
| 206 |
upscale2 = upscaler_2(upscale, tiling=True, tile_width=256, tile_height=256)
|
|
|
|
| 250 |
sd35_path = f"sd35ll_{timestamp}.png"
|
| 251 |
sd_image.save(sd35_path,optimize=False,compress_level=0)
|
| 252 |
pyx.upload_to_ftp(sd35_path)
|
| 253 |
+
upscaler_2.to(torch.device('cuda',non_blocking=True))
|
| 254 |
with torch.no_grad():
|
| 255 |
upscale = upscaler_2(sd_image, tiling=True, tile_width=256, tile_height=256)
|
| 256 |
upscale2 = upscaler_2(upscale, tiling=True, tile_width=256, tile_height=256)
|