test stable fast

app_onediff.py

@@ -12,7 +12,9 @@ from diffusers import (
     AutoencoderKL,
 )
 
-from DeepCache import DeepCacheSDHelper
+import xformers
+import triton
+from sfast.compilers.diffusion_pipeline_compiler import (compile, CompilationConfig)
 
 def runcmd(cmd, verbose = False, *args, **kwargs):
 
@@ -41,9 +43,9 @@ def runcmd(cmd, verbose = False, *args, **kwargs):
 
 # from onediffx import compile_pipe, save_pipe
 
-def nexfort_compile(torch_module: torch.nn.Module):
-    options = json.loads('{"mode": "max-optimize:max-autotune:low-precision", "memory_format": "channels_last", "dynamic": true}')
-    return compile_pipe(torch_module, backend="nexfort", options=options, fuse_qkv_projections=True)
+# def nexfort_compile(torch_module: torch.nn.Module):
+#     options = json.loads('{"mode": "max-optimize:max-autotune:low-precision", "memory_format": "channels_last", "dynamic": true}')
+#     return compile_pipe(torch_module, backend="nexfort", options=options, fuse_qkv_projections=True)
 
 BASE_MODEL = "stabilityai/sdxl-turbo"
 device = "cuda"
@@ -67,6 +69,14 @@ base_pipe.to(device)
 
 # base_pipe = nexfort_compile(base_pipe)
 
+ccnf = CompilationConfig.Default()
+
+ccnf.enable_xformers = True
+ccnf.enable_triton = True
+ccnf.enable_cuda_graph = True
+
+base_pipe = compile(base_pipe, ccnf)
+
 from gfpgan.utils import GFPGANer
 from basicsr.archs.srvgg_arch import SRVGGNetCompact
 from realesrgan.utils import RealESRGANer

requirements.txt

@@ -9,4 +9,6 @@ git+https://github.com/XPixelGroup/BasicSR@master
 gfpgan
 facexlib
 realesrgan
-deepcache
+stable-fast
+triton
+xformers