ComfyUI/comfy/customzluda/zluda.py

700 lines
28 KiB
Python

# ------------------- Hide ROCm/HIP -------------------
import sys
import os
os.environ.pop("ROCM_HOME", None)
os.environ.pop("HIP_HOME", None)
os.environ.pop("ROCM_VERSION", None)
#triton fix?
os.environ["FLASH_ATTENTION_TRITON_AMD_ENABLE"] = "TRUE"
os.environ["FLASH_ATTENTION_TRITON_AMD_AUTOTUNE"] = "TRUE"
os.environ["TRITON_DEBUG"] = "1" # Verbose logging
paths = os.environ["PATH"].split(";")
paths_no_rocm = [p for p in paths if "rocm" not in p.lower()]
os.environ["PATH"] = ";".join(paths_no_rocm)
# ------------------- End ROCm/HIP Hiding -------------
# Fix for cublasLt errors on newer ZLUDA (if no hipblaslt)
os.environ['DISABLE_ADDMM_CUDA_LT'] = '1'
# ------------------- main imports -------------------
# main imports
import torch
torch._dynamo.config.suppress_errors = True # Skip compilation errors
torch._dynamo.config.optimize_ddp = False # Disable distributed optimizations
import ctypes
import shutil
import subprocess
import importlib.metadata
from functools import wraps
from typing import Union, List
from enum import Enum
# ------------------- main imports -------------------
# ------------------- gfx detection -------------------
import os
import re
def detect_amd_gpu_architecture():
"""
Detect AMD GPU architecture on Windows and return the appropriate gfx code for TRITON_OVERRIDE_ARCH
"""
try:
# Method 1: Try Windows registry
try:
import winreg
key_path = r"SYSTEM\CurrentControlSet\Control\Class\{4d36e968-e325-11ce-bfc1-08002be10318}"
with winreg.OpenKey(winreg.HKEY_LOCAL_MACHINE, key_path) as key:
i = 0
while True:
try:
subkey_name = winreg.EnumKey(key, i)
with winreg.OpenKey(key, subkey_name) as subkey:
try:
desc = winreg.QueryValueEx(subkey, "DriverDesc")[0]
if "AMD" in desc or "Radeon" in desc:
print(f" :: Detected GPU via Windows registry: {desc}")
return gpu_name_to_gfx(desc)
except FileNotFoundError:
pass
i += 1
except OSError:
break
except ImportError:
pass
# Method 2: Try WMIC command
try:
import subprocess
result = subprocess.run(['wmic', 'path', 'win32_VideoController', 'get', 'name'],
capture_output=True, text=True, timeout=10)
if result.returncode == 0:
for line in result.stdout.split('\n'):
line = line.strip()
if line and "AMD" in line or "Radeon" in line:
print(f" :: Detected GPU via WMIC: {line}")
return gpu_name_to_gfx(line)
except (FileNotFoundError, subprocess.TimeoutExpired):
pass
print(" :: Could not detect AMD GPU architecture automatically")
return None
except Exception as e:
print(f" :: GPU detection failed: {str(e)}")
return None
def gpu_name_to_gfx(gpu_name):
"""
Map GPU names to their corresponding gfx architecture codes
"""
gpu_name_lower = gpu_name.lower()
# RDNA3 (gfx11xx)
if any(x in gpu_name_lower for x in ['rx 7900', 'rx 7800', 'rx 7700', 'rx 7600', 'rx 7500']):
if 'rx 7900' in gpu_name_lower:
return 'gfx1100' # Navi 31
elif 'rx 7800' in gpu_name_lower or 'rx 7700' in gpu_name_lower:
return 'gfx1101' # Navi 32
elif 'rx 7600' in gpu_name_lower or 'rx 7500' in gpu_name_lower:
return 'gfx1102' # Navi 33
# RDNA2 (gfx10xx)
elif any(x in gpu_name_lower for x in ['rx 6950', 'rx 6900', 'rx 6800', 'rx 6750', 'rx 6700']):
return 'gfx1030' # Navi 21/22
elif any(x in gpu_name_lower for x in ['rx 6650', 'rx 6600', 'rx 6500', 'rx 6400']):
return 'gfx1032' # Navi 23/24
# RDNA1 (gfx10xx)
elif any(x in gpu_name_lower for x in ['rx 5700', 'rx 5600', 'rx 5500']):
return 'gfx1010' # Navi 10
# Vega (gfx9xx)
elif any(x in gpu_name_lower for x in ['vega 64', 'vega 56', 'vega 20', 'radeon vii']):
return 'gfx900' # Vega 10/20
elif 'vega 11' in gpu_name_lower or 'vega 8' in gpu_name_lower:
return 'gfx902' # Raven Ridge APU
# Polaris (gfx8xx)
elif any(x in gpu_name_lower for x in ['rx 580', 'rx 570', 'rx 480', 'rx 470']):
return 'gfx803' # Polaris 10/20
elif any(x in gpu_name_lower for x in ['rx 560', 'rx 550', 'rx 460']):
return 'gfx803' # Polaris 11/12
# Default fallback - try to extract numbers and make educated guess
if 'rx 9' in gpu_name_lower: # Future RDNA4?
return 'gfx1200' # Anticipated next gen
elif 'rx 8' in gpu_name_lower: # Future RDNA4?
return 'gfx1150' # Anticipated next gen
elif 'rx 7' in gpu_name_lower:
return 'gfx1100' # Default RDNA3
elif 'rx 6' in gpu_name_lower:
return 'gfx1030' # Default RDNA2
elif 'rx 5' in gpu_name_lower:
return 'gfx1010' # Default RDNA1
print(f" :: Unknown GPU model: {gpu_name}, using default gfx1030")
return 'gfx1030' # Safe default for most modern AMD GPUs
def set_triton_arch_override():
"""
Automatically detect and set TRITON_OVERRIDE_ARCH environment variable
"""
# Check if already set by user
if 'TRITON_OVERRIDE_ARCH' in os.environ:
print(f" :: TRITON_OVERRIDE_ARCH already set to: {os.environ['TRITON_OVERRIDE_ARCH']}")
return
print(" :: Auto-detecting AMD GPU architecture for Triton...")
gfx_arch = detect_amd_gpu_architecture()
if gfx_arch:
os.environ['TRITON_OVERRIDE_ARCH'] = gfx_arch
print(f" :: Set TRITON_OVERRIDE_ARCH={gfx_arch}")
else:
# Fallback to a common architecture
fallback_arch = 'gfx1030'
os.environ['TRITON_OVERRIDE_ARCH'] = fallback_arch
print(f" :: Using fallback TRITON_OVERRIDE_ARCH={fallback_arch}")
print(" :: If Triton fails, you may need to manually set TRITON_OVERRIDE_ARCH in your environment")
# ------------------- gfx detection -------------------
# ------------------- ComfyUI Package Version Check -------------------
def get_package_version(package_name):
try:
from importlib.metadata import version
return version(package_name)
except ImportError:
from importlib_metadata import version
return version(package_name)
def parse_requirements_file(requirements_path):
"""Parse requirements.txt file and extract package versions."""
requirements = {}
try:
with open(requirements_path, 'r') as f:
for line in f:
line = line.strip()
if line and not line.startswith('#'):
# Handle different version specifiers
if '==' in line:
pkg, version = line.split('==', 1)
requirements[pkg] = version.strip()
elif '>=' in line:
pkg, version = line.split('>=', 1)
requirements[pkg] = version.strip()
elif '~=' in line: # Compatible release operator
pkg, version = line.split('~=', 1)
requirements[pkg] = version.strip()
# You can add more operators if needed (>, <, <=, !=)
except FileNotFoundError:
print(f" :: Warning: requirements.txt not found at {requirements_path}")
return requirements
def is_compatible_version(installed_version, required_version, operator='>='):
"""Check if installed version meets requirement based on operator."""
try:
from packaging import version
installed_v = version.parse(installed_version)
required_v = version.parse(required_version)
if operator == '>=':
return installed_v >= required_v
elif operator == '==':
return installed_v == required_v
elif operator == '~=':
# Compatible release: ~=2.0 means >=2.0, <3.0
# ~=2.1 means >=2.1, <2.2
required_parts = required_v.release
if len(required_parts) == 1:
# ~=2 means >=2.0, <3.0
return (installed_v >= required_v and
installed_v.release[0] == required_parts[0])
else:
# ~=2.1 means >=2.1, <2.2
return (installed_v >= required_v and
installed_v.release[:len(required_parts)-1] == required_parts[:-1] and
installed_v.release[len(required_parts)-1] >= required_parts[-1])
else:
# Default to >= for unknown operators
return installed_v >= required_v
except Exception as e:
print(f" :: Version comparison error for {installed_version} vs {required_version}: {e}")
return False
def uninstall_package(package_name):
"""Uninstall a package quietly"""
import subprocess
import sys
try:
args = [sys.executable, '-m', 'pip', 'uninstall', package_name, '-y', '--quiet']
subprocess.check_call(args)
return True
except subprocess.CalledProcessError:
return False
def check_pydantic_compatibility():
"""Check if current pydantic packages are compatible, return True if they need reinstalling"""
try:
# Try to import the problematic class that causes the error
from pydantic_settings import TomlConfigSettingsSource
# If we get here, the packages are compatible
return False
except ImportError:
# Import failed, packages are incompatible
return True
except Exception:
# Any other error, assume incompatible
return True
def handle_pydantic_packages(required_packages):
"""Special handling for pydantic packages to ensure compatibility"""
import subprocess
import sys
pydantic_packages = ['pydantic', 'pydantic-settings']
packages_in_requirements = [pkg for pkg in pydantic_packages if pkg in required_packages]
if not packages_in_requirements:
return # No pydantic packages to handle
# Check if both packages are available and what versions
pydantic_installed = None
pydantic_settings_installed = None
try:
pydantic_installed = get_package_version('pydantic')
except:
pass
try:
pydantic_settings_installed = get_package_version('pydantic-settings')
except:
pass
# If both are installed, check compatibility
if pydantic_installed and pydantic_settings_installed:
print(f"Found pydantic: {pydantic_installed}, pydantic-settings: {pydantic_settings_installed}")
# Check if they're compatible by testing the import
if not check_pydantic_compatibility():
print(" :: Pydantic packages are compatible, skipping reinstall")
return
else:
print(" :: Pydantic packages are incompatible, need to reinstall")
# If we get here, we need to install/reinstall pydantic packages
print(" :: Setting up pydantic packages for compatibility...")
# Uninstall existing versions to avoid conflicts
if pydantic_installed:
print(f" :: Uninstalling existing pydantic {pydantic_installed}")
uninstall_package('pydantic')
if pydantic_settings_installed:
print(f" :: Uninstalling existing pydantic-settings {pydantic_settings_installed}")
uninstall_package('pydantic-settings')
# Install both packages together
try:
print(" :: Installing compatible pydantic packages...")
combined_args = [sys.executable, '-m', 'pip', 'install',
'pydantic~=2.0',
'pydantic-settings~=2.0',
'--quiet',
'--disable-pip-version-check']
subprocess.check_call(combined_args)
# Verify installation
new_pydantic = get_package_version('pydantic')
new_pydantic_settings = get_package_version('pydantic-settings')
print(f" :: Successfully installed pydantic: {new_pydantic}, pydantic-settings: {new_pydantic_settings}")
except subprocess.CalledProcessError as e:
print(f" :: Failed to install pydantic packages: {e}")
def install_package(package_name, version_spec, upgrade=False):
import subprocess
import sys
# For ~= operator, install with the compatible release syntax
if '~=' in version_spec:
package_spec = f'{package_name}~={version_spec}'
else:
package_spec = f'{package_name}=={version_spec}'
args = [sys.executable, '-m', 'pip', 'install',
package_spec,
'--quiet',
'--disable-pip-version-check']
if upgrade:
args.append('--upgrade')
try:
subprocess.check_call(args)
except subprocess.CalledProcessError as e:
print(f" :: Failed to install {package_name}: {e}")
# Try installing without version constraint as fallback
if upgrade and '~=' in package_spec:
try:
print(f" :: Retrying {package_name} installation without version constraint...")
fallback_args = [sys.executable, '-m', 'pip', 'install',
package_name,
'--upgrade',
'--quiet',
'--disable-pip-version-check']
subprocess.check_call(fallback_args)
print(f" :: {package_name} installed successfully without version constraint")
except subprocess.CalledProcessError as e2:
print(f" :: Fallback installation also failed: {e2}")
def ensure_package(package_name, required_version, operator='>='):
# Skip individual pydantic package handling - they're handled together
if package_name in ['pydantic', 'pydantic-settings']:
return
try:
installed_version = get_package_version(package_name)
print(f"Installed version of {package_name}: {installed_version}")
if not is_compatible_version(installed_version, required_version, operator):
install_package(package_name, required_version, upgrade=True)
print(f"\n{package_name} outdated. Upgraded to {required_version}.")
except Exception as e:
print(f" :: {package_name} not found or error checking version: {e}")
install_package(package_name, required_version)
print(f"\n{package_name} was missing. Installed it.")
# Determine operator from requirements.txt
def get_version_operator(requirements_path, package_name):
"""Extract the version operator used for a package in requirements.txt"""
try:
with open(requirements_path, 'r') as f:
for line in f:
line = line.strip()
if line and not line.startswith('#') and line.startswith(package_name):
if '~=' in line:
return '~='
elif '==' in line:
return '=='
elif '>=' in line:
return '>='
except FileNotFoundError:
pass
return '>=' # Default
import os
requirements_path = os.path.join(os.path.dirname(os.path.dirname(__file__)), 'requirements.txt')
required_packages = parse_requirements_file(requirements_path)
packages_to_monitor = [
"comfyui-frontend-package",
"comfyui-workflow-templates",
"av",
"comfyui-embedded-docs",
"pydantic",
"pydantic-settings",
]
print("\n :: Checking package versions...")
# Handle pydantic packages first with special logic
handle_pydantic_packages(required_packages)
# Handle other packages
for package_name in packages_to_monitor:
if package_name in required_packages and package_name not in ['pydantic', 'pydantic-settings']:
operator = get_version_operator(requirements_path, package_name)
ensure_package(package_name, required_packages[package_name], operator)
elif package_name not in ['pydantic', 'pydantic-settings']:
print(f" :: Warning: {package_name} not found in requirements.txt")
print(" :: Package version check complete.")
# ------------------- End Version Check -------------------
# ------------------- Triton Setup -------------------
print("\n :: ------------------------ ZLUDA ----------------------- :: ")
# identify device and set triton arch override
zluda_device_name = torch.cuda.get_device_name() if torch.cuda.is_available() else ""
is_zluda = zluda_device_name.endswith("[ZLUDA]")
if is_zluda:
set_triton_arch_override()
try:
import triton
import triton.language as tl
print(" :: Triton core imported successfully")
# This needs to be up here, so it can disable cudnn before anything can even think about using it
torch.backends.cudnn.enabled = os.environ.get("TORCH_BACKENDS_CUDNN_ENABLED", "1").strip().lower() not in {"0", "off", "false", "disable", "disabled", "no"}
if torch.backends.cudnn.enabled:
print(" :: Enabled cuDNN")
else:
print(" :: Disabled cuDNN")
@triton.jit
def _zluda_kernel_test(x_ptr, y_ptr, n_elements, BLOCK_SIZE: tl.constexpr):
pid = tl.program_id(axis=0)
offsets = pid * BLOCK_SIZE + tl.arange(0, BLOCK_SIZE)
mask = offsets < n_elements
x = tl.load(x_ptr + offsets, mask=mask)
tl.store(y_ptr + offsets, x + 1, mask=mask)
def _verify_triton() -> bool:
try:
print(" :: Running Triton kernel test...")
x = torch.ones(64, device='cuda')
y = torch.empty_like(x)
_zluda_kernel_test[(1,)](x, y, x.numel(), BLOCK_SIZE=64)
if torch.allclose(y, x + 1):
print(" :: Triton kernel test passed successfully")
return True
print(" :: Triton kernel test failed (incorrect output)")
return False
except Exception as e:
print(f" :: Triton test failed: {str(e)}")
return False
triton_available = _verify_triton()
if triton_available:
print(" :: Triton initialized successfully")
os.environ['FLASH_ATTENTION_TRITON_AMD_AUTOTUNE'] = 'TRUE'
else:
print(" :: Triton available but failed verification")
except ImportError:
print(" :: Triton not installed")
triton_available = False
except Exception as e:
print(f" :: Triton initialization failed: {str(e)}")
triton_available = False
# ------------------- End Triton Verification -------------------
# # ------------------- ZLUDA Core Implementation -------------------
MEM_BUS_WIDTH = {
"AMD Radeon RX 9070 XT": 256,
"AMD Radeon RX 9070": 256,
"AMD Radeon RX 9070 GRE": 192,
"AMD Radeon RX 9060 XT": 128,
"AMD Radeon RX 9060": 128,
"AMD Radeon RX 7900 XTX": 384,
"AMD Radeon RX 7900 XT": 320,
"AMD Radeon RX 7900 GRE": 256,
"AMD Radeon RX 7800 XT": 256,
"AMD Radeon RX 7700 XT": 192,
"AMD Radeon RX 7700": 192,
"AMD Radeon RX 7650 GRE": 128,
"AMD Radeon RX 7600 XT": 128,
"AMD Radeon RX 7600": 128,
"AMD Radeon RX 7400": 128,
"AMD Radeon RX 6950 XT": 256,
"AMD Radeon RX 6900 XT": 256,
"AMD Radeon RX 6800 XT": 256,
"AMD Radeon RX 6800": 256,
"AMD Radeon RX 6750 XT": 192,
"AMD Radeon RX 6750 GRE 12GB": 192,
"AMD Radeon RX 6750 GRE 10GB": 160,
"AMD Radeon RX 6700 XT": 192,
"AMD Radeon RX 6700": 160,
"AMD Radeon RX 6650 XT": 128,
"AMD Radeon RX 6600 XT": 128,
"AMD Radeon RX 6600": 128,
"AMD Radeon RX 6500 XT": 64,
"AMD Radeon RX 6400": 64,
"AMD Radeon RX 5700 XT": 256,
"AMD Radeon RX 5700": 256,
"AMD Radeon RX 5600 XT": 192,
"AMD Radeon RX 5500 XT": 128,
"AMD Radeon RX 5500": 128,
"AMD Radeon RX 5300": 96,
# AMD Radeon Pro R9000/W7000/W6000/W5000 series, Apple exclusive WX series not listed
"AMD Radeon AI PRO R9700": 256,
"AMD Radeon PRO W7900": 384,
"AMD Radeon PRO W7800 48GB": 384,
"AMD Radeon PRO W7800": 256,
"AMD Radeon PRO W7700": 256,
"AMD Radeon PRO W7600": 128,
"AMD Radeon PRO W7500": 128,
"AMD Radeon PRO W7400": 128,
"AMD Radeon PRO W6800": 256,
"AMD Radeon PRO W6600": 128,
"AMD Radeon PRO W6400": 64,
"AMD Radeon PRO W5700": 256,
"AMD Radeon PRO W5500": 128,
}
# ------------------- Device Properties Implementation -------------------
class DeviceProperties:
PROPERTIES_OVERRIDE = {"regs_per_multiprocessor": 65535, "gcnArchName": "UNKNOWN ARCHITECTURE"}
internal: torch._C._CudaDeviceProperties
def __init__(self, props: torch._C._CudaDeviceProperties):
self.internal = props
def __getattr__(self, name):
if name in DeviceProperties.PROPERTIES_OVERRIDE:
return DeviceProperties.PROPERTIES_OVERRIDE[name]
return getattr(self.internal, name)
# # ------------------- Audio Ops Patch -------------------
# if is_zluda:
# _torch_stft = torch.stft
# _torch_istft = torch.istft
# def z_stft(input: torch.Tensor, window: torch.Tensor, *args, **kwargs):
# return _torch_stft(input=input.cpu(), window=window.cpu(), *args, **kwargs).to(input.device)
# def z_istft(input: torch.Tensor, window: torch.Tensor, *args, **kwargs):
# return _torch_istft(input=input.cpu(), window=window.cpu(), *args, **kwargs).to(input.device)
# def z_jit(f, *_, **__):
# f.graph = torch._C.Graph()
# return f
# torch._dynamo.config.suppress_errors = True
# torch.stft = z_stft
# torch.istft = z_istft
# torch.jit.script = z_jit
# # ------------------- End Audio Patch -------------------
# ------------------- Top-K Fallback Patch -------------------
if is_zluda:
_topk = torch.topk
def safe_topk(input: torch.Tensor, *args, **kwargs):
device = input.device
values, indices = _topk(input.cpu(), *args, **kwargs)
return torch.return_types.topk((values.to(device), indices.to(device),))
torch.topk = safe_topk
# ------------------- End Top-K Patch -------------------
# ------------------- ONNX Runtime Patch -------------------
try:
import onnxruntime as ort
if is_zluda:
print(" :: Patching ONNX Runtime for ZLUDA — disabling CUDA EP.")
# Store original get_available_providers
original_get_available_providers = ort.get_available_providers
def filtered_providers():
return [ep for ep in original_get_available_providers() if ep != "CUDAExecutionProvider"]
# Patch ONLY the _pybind_state version (used during session creation)
ort.capi._pybind_state.get_available_providers = filtered_providers
# Wrap InferenceSession to force CPU provider when CUDA is explicitly requested
OriginalSession = ort.InferenceSession
class SafeInferenceSession(OriginalSession):
def __init__(self, *args, providers=None, **kwargs):
if providers and "CUDAExecutionProvider" in providers:
print(" :: Forcing ONNX to use CPUExecutionProvider instead of CUDA.")
providers = ["CPUExecutionProvider"]
super().__init__(*args, providers=providers, **kwargs)
ort.InferenceSession = SafeInferenceSession
except ImportError:
print(" :: ONNX Runtime not installed — skipping patch.")
except Exception as e:
print(" :: Failed to patch ONNX Runtime:", e)
# ------------------- End ONNX Patch -------------------
# ------------------- ZLUDA hijack ---------------------
do_nothing = lambda _: None
def do_hijack():
if not is_zluda:
return
print(f" :: Using ZLUDA with device: {zluda_device_name}")
print(" :: Applying core ZLUDA patches...")
# 2. Triton optimizations
if triton_available:
print(" :: Initializing Triton optimizations")
try:
# General Triton config
print(" :: Configuring Triton device properties...")
_get_props = triton.runtime.driver.active.utils.get_device_properties
def patched_props(device):
props = _get_props(device)
name = torch.cuda.get_device_name()[:-8] # Remove [ZLUDA]
props["mem_bus_width"] = MEM_BUS_WIDTH.get(name, 128)
if name not in MEM_BUS_WIDTH:
print(f' :: Using default mem_bus_width=128 for {name}')
return props
triton.runtime.driver.active.utils.get_device_properties = patched_props
print(" :: Triton device properties configured")
# Flash Attention
flash_enabled = False
try:
from comfy.flash_attn_triton_amd import interface_fa
print(" :: Flash attention components found")
original_sdpa = torch.nn.functional.scaled_dot_product_attention
def amd_flash_wrapper(query, key, value, attn_mask=None, dropout_p=0.0, is_causal=False, scale=None):
try:
if (query.shape[-1] <= 128 and
attn_mask is None and # fix flash-attention error : "Flash attention error: Boolean value of Tensor with more than one value is ambiguous"
query.dtype != torch.float32):
if scale is None:
scale = query.shape[-1] ** -0.5
return interface_fa.fwd(
query.transpose(1, 2),
key.transpose(1, 2),
value.transpose(1, 2),
None, None, dropout_p, scale,
is_causal, -1, -1, 0.0, False, None
)[0].transpose(1, 2)
except Exception as e:
print(f' :: Flash attention error: {str(e)}')
return original_sdpa(query=query, key=key, value=value, attn_mask=attn_mask, dropout_p=dropout_p, is_causal=is_causal, scale=scale)
torch.nn.functional.scaled_dot_product_attention = amd_flash_wrapper
flash_enabled = True
print(" :: AMD flash attention enabled successfully")
except ImportError:
print(" :: Flash attention components not installed")
except Exception as e:
print(f" :: Flash attention setup failed: {str(e)}")
# Other Triton optimizations
if not flash_enabled:
print(" :: Applying basic Triton optimizations")
# Add other Triton optimizations here
# ...
except Exception as e:
print(f" :: Triton optimization failed: {str(e)}")
else:
print(" :: Triton optimizations skipped (not available)")
# 3. Common configurations
print(" :: Configuring PyTorch backends...")
torch.backends.cuda.enable_mem_efficient_sdp(False)
torch.backends.cuda.enable_mem_efficient_sdp = do_nothing
if hasattr(torch.backends.cuda, "enable_flash_sdp"):
torch.backends.cuda.enable_flash_sdp(True)
print(" :: Disabled CUDA flash attention")
if hasattr(torch.backends.cuda, "enable_math_sdp"):
torch.backends.cuda.enable_math_sdp(True)
print(" :: Enabled math attention fallback")
print(" :: ZLUDA initialization complete")
print(" :: ------------------------ ZLUDA ----------------------- :: \n")
if is_zluda:
do_hijack()
else:
print(f" :: CUDA device detected: {zluda_device_name or 'None'}")