Public/automatic
1
0
Fork 0
mirror of https://github.com/vladmandic/automatic synced 2026-09-20 01:31:13 +02:00
Code Issues Packages Projects Releases Wiki Activity

improve zluda installation

Browse Source
This commit is contained in:
Seunghoon Lee
2024-02-23 15:14:13 +09:00
committed by Vladimir Mandic
parent d6a086c454
commit a02052cd6a
4 changed files with 64 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files
installer.py
+42 -4
View File
@@ -405,10 +405,14 @@ def check_torch():
log.debug(f'Torch allowed: cuda={allow_cuda} rocm={allow_rocm} ipex={allow_ipex} diml={allow_directml} openvino={allow_openvino}')
torch_command = os.environ.get('TORCH_COMMAND', '')
xformers_package = os.environ.get('XFORMERS_PACKAGE', 'none')
zluda_need_dll_patch = False
def is_rocm_available():
if not allow_rocm:
return False
if installed('torch-directml'):
log.debug('DirectML installation is detected. Skipping HIP SDK check.')
return False
if platform.system() == 'Windows':
hip_path = os.environ.get('HIP_PATH', None)
return hip_path is not None and os.path.exists(os.path.join(hip_path, 'bin'))
@@ -474,7 +478,15 @@ def check_torch():
except Exception as e:
log.debug(f'ROCm hipconfig failed: {e}')
rocm_ver = None
if not is_windows: # remove after PyTorch built with ROCm for Windows is launched
if args.use_zluda: # ZLUDA is available on both Linux and Windows
torch_command = os.environ.get('TORCH_COMMAND', 'torch==2.2.0 torchvision --index-url https://download.pytorch.org/whl/cu118')
log.warning("Currently, ZLUDA support is experimental and unstable.")
zluda_need_dll_patch = is_windows and not installed('torch')
elif is_windows: # remove this check after PyTorch built with ROCm for Windows is released
log.warning("HIP SDK is detected, but there's no PyTorch release for Windows at this moment. If you are trying ZLUDA, please add '--use-zluda'.")
log.info('Using CPU-only torch')
torch_command = os.environ.get('TORCH_COMMAND', 'torch torchvision')
else:
if rocm_ver in {"5.7"}:
torch_command = os.environ.get('TORCH_COMMAND', f'torch torchvision --pre --index-url https://download.pytorch.org/whl/nightly/rocm{rocm_ver}')
elif rocm_ver in {"5.5", "5.6"}:
@@ -484,9 +496,6 @@ def check_torch():
torch_command = os.environ.get('TORCH_COMMAND', 'torch torchvision --index-url https://download.pytorch.org/whl/rocm5.5')
if rocm_ver is not None:
install(os.environ.get('ONNXRUNTIME_PACKAGE', get_onnxruntime_source_for_rocm(arr)), "onnxruntime-training built with ROCm", ignore=True)
else:
log.info('Using CPU-only Torch')
torch_command = os.environ.get('TORCH_COMMAND', 'torch torchvision')
xformers_package = os.environ.get('XFORMERS_PACKAGE', 'none')
elif allow_ipex and (args.use_ipex or shutil.which('sycl-ls') is not None or shutil.which('sycl-ls.exe') is not None or os.environ.get('ONEAPI_ROOT') is not None or os.path.exists('/opt/intel/oneapi') or os.path.exists("C:/Program Files (x86)/Intel/oneAPI") or os.path.exists("C:/oneAPI")):
args.use_ipex = True # pylint: disable=attribute-defined-outside-init
@@ -539,11 +548,15 @@ def check_torch():
install(torch_command, 'torch torchvision')
install('onnxruntime-directml', 'onnxruntime-directml', ignore=True)
else:
if args.use_zluda:
log.warning("Failed to initialize ZLUDA. There's no HIP SDK found in PATH.")
log.info('Using CPU-only Torch')
torch_command = os.environ.get('TORCH_COMMAND', 'torch torchvision')
if 'torch' in torch_command and not args.version:
log.debug(f'Installing torch: {torch_command}')
install(torch_command, 'torch torchvision')
if zluda_need_dll_patch:
patch_dlls_for_zluda()
else:
try:
import torch
@@ -896,6 +909,31 @@ def get_onnxruntime_source_for_rocm(rocm_ver):
return 'onnxruntime-gpu'
def patch_dlls_for_zluda():
zluda_path = os.environ.get('ZLUDA', None)
if zluda_path is None:
paths = os.environ.get('PATH', '').split(';')
for path in paths:
if os.path.exists(os.path.join(path, 'zluda_redirect.dll')):
zluda_path = path
break
if zluda_path is None:
log.warning('Failed to automatically patch torch with ZLUDA. Could not find ZLUDA from PATH.')
return
venv_path = os.path.dirname(shutil.which('python'))
dlls_to_patch = {
'cublas.dll': 'cublas64_11.dll',
'cudnn.dll': 'cudnn64_8.dll',
'cusparse.dll': 'cusparse64_11.dll',
'nvrtc.dll': 'nvrtc64_112_0.dll',
}
try:
for k, v in dlls_to_patch.items():
shutil.copyfile(os.path.join(zluda_path, k), os.path.join(venv_path, 'Lib', 'site-packages', 'torch', 'lib', v))
except Exception as e:
log.warning(f'Failed to automatically patch torch with ZLUDA: {e}')
# check version of the main repo and optionally upgrade it
def check_version(offline=False, reset=True): # pylint: disable=unused-argument
if args.skip_all:
modules/cmd_args.py
+1
View File
@@ -41,6 +41,7 @@ group.add_argument("--profile", default=os.environ.get("SD_PROFILE", False), act
group.add_argument("--disable-queue", default=os.environ.get("SD_DISABLEQUEUE", False), action='store_true', help="Disable queues, default: %(default)s")
group.add_argument('--debug', default=os.environ.get("SD_DEBUG", False), action='store_true', help = "Run installer with debug logging, default: %(default)s")
group.add_argument('--use-directml', default=os.environ.get("SD_USEDIRECTML", False), action='store_true', help = "Use DirectML if no compatible GPU is detected, default: %(default)s")
group.add_argument('--use-zluda', default=os.environ.get("SD_USEZLUDA", False), action='store_true', help = "Force use ZLUDA, AMD GPUs only, default: %(default)s")
group.add_argument("--use-openvino", default=os.environ.get("SD_USEOPENVINO", False), action='store_true', help="Use Intel OpenVINO backend, default: %(default)s")
group.add_argument("--use-ipex", default=os.environ.get("SD_USEIPX", False), action='store_true', help="Force use Intel OneAPI XPU backend, default: %(default)s")
group.add_argument("--use-cuda", default=os.environ.get("SD_USECUDA", False), action='store_true', help="Force use nVidia CUDA backend, default: %(default)s")
modules/shared.py
-3
View File
@@ -407,9 +407,6 @@ options_templates.update(options_section(('cuda', "Compute Settings"), {
"olive_vae_encoder_float32": OptionInfo(False, 'Olive force FP32 for VAE Encoder'),
"olive_static_dims": OptionInfo(True, 'Olive use static dimensions'),
"olive_cache_optimized": OptionInfo(True, 'Olive cache optimized models'),
"zluda_sep": OptionInfo("<h2>ZLUDA</h2>(experimental)", "", gr.HTML, {"visible": devices.backend == "cuda"}),
"zluda_enable_cudnn": OptionInfo(False, 'ZLUDA enable cuDNN (restart required)', gr.Checkbox, {"visible": devices.backend == "cuda"}),
}))
options_templates.update(options_section(('advanced', "Inference Settings"), {
modules/zluda.py
+21 -5
View File
@@ -3,13 +3,29 @@ import torch
from modules import shared, devices
def test(device: torch.device):
try:
ten1 = torch.randn((2, 4,), device=device)
ten2 = torch.randn((4, 8,), device=device)
out = torch.mm(ten1, ten2)
return out.sum().is_nonzero()
except Exception:
return False
def initialize_zluda():
if platform.system() == "Windows" and devices.cuda_ok and torch.cuda.get_device_name(devices.get_optimal_device()).endswith("[ZLUDA]"):
shared.log.warning("Detected ZLUDA device. Currently, ZLUDA support is experimental and unstable.")
torch.backends.cudnn.enabled = shared.opts.zluda_enable_cudnn
if torch.backends.cudnn.enabled:
shared.log.warning("cuDNN with ZLUDA won't work at this moment. Please wait for future update.")
device = devices.get_optimal_device()
if platform.system() == "Windows" and devices.cuda_ok and torch.cuda.get_device_name(device).endswith("[ZLUDA]"):
torch.backends.cudnn.enabled = False
torch.backends.cuda.enable_flash_sdp(False)
torch.backends.cuda.enable_math_sdp(True)
torch.backends.cuda.enable_mem_efficient_sdp(False)
shared.opts.sdp_options = ['Math attention']
devices.device_codeformer = devices.cpu
if not test(device):
shared.log.error(f'ZLUDA device failed to pass basic operation test: index={device.index}, device_name={torch.cuda.get_device_name(device)}')
torch.cuda.is_available = lambda: False
devices.cuda_ok = False
devices.backend = 'cpu'
devices.device = devices.device_esrgan = devices.device_gfpgan = devices.device_interrogate = devices.cpu