Support intel xpu

pyproject.toml

@@ -36,6 +36,11 @@ mps = [
   "torch>=2.6.0",
   "torchvision",
 ]
+xpu = [
+  "torch>=2.6.0",
+  "torchvision",
+  "intel-extension-for-pytorch>=2.1.0",
+]
 detect = [
   # TODO: This feature is not enabled by default and the deps may require local compilation.
   # UV attempts to resolve all deps, including optionaly ones (https://github.com/astral-sh/uv/issues/6729).
@@ -50,6 +55,7 @@ conflicts = [
     { extra = "cu118" },
     { extra = "cu126" },
     { extra = "mps" },
+    { extra = "xpu" },
   ],
 ]
 
@@ -61,6 +67,7 @@ torch = [
   { index = "pytorch-cu124", extra = "cu126", marker = "platform_system == 'Windows'" },
   { index = "pytorch-cu126", extra = "cu126", marker = "platform_system != 'Windows'" },
   { index = "pytorch-cpu", extra = "mps" },
+  { index = "pytorch-cpu", extra = "xpu" },
 ]
 sam-2 = { git = "https://github.com/facebookresearch/sam2.git", rev = "c2ec8e14a185632b0a5d8b161928ceb50197eddc" }
 # For windows, prefer a prebuilt flash_attn wheel. https://huggingface.co/lldacing/flash-attention-windows-wheel/tree/main

src/enhance/lib/gpu.py

@@ -1,7 +1,13 @@
 import torch
 import pynvml
 import psutil
-from typing import List, Tuple
+from typing import Tuple
+
+try:
+    import intel_extension_for_pytorch as ipex
+    IPEX_AVAILABLE = True
+except ImportError:
+    IPEX_AVAILABLE = False
 
 class GPUInfo:
 
@@ -16,6 +22,12 @@ def __init__(self):
         except Exception as e:
             self.mpsAvailable = False
 
+        try:
+            # Check for Intel XPU support
+            self.xpuAvailable = IPEX_AVAILABLE and hasattr(torch, 'xpu') and torch.xpu.is_available()
+        except (AttributeError, Exception) as e:
+            self.xpuAvailable = False
+
         try:
             pynvml.nvmlInit()
             device_count = pynvml.nvmlDeviceGetCount()
@@ -33,6 +45,11 @@ def getGpuNames(self):
         try:
             if self.mpsAvailable:
                 return [["mps", "Apple Silicon GPU (MPS)"]]
+            if self.xpuAvailable:
+                return [
+                    [f"xpu:{i}", torch.xpu.get_device_name(i)]
+                    for i in range(torch.xpu.device_count())
+                ]
             if self.cudaAvailable:
                 return [
                     [f"cuda:{i}", torch.cuda.get_device_name(i)]
@@ -43,14 +60,17 @@ def getGpuNames(self):
         return []
 
     def getGpuPresent(self):
-        return self.cudaAvailable or self.mpsAvailable
+        return self.cudaAvailable or self.mpsAvailable or self.xpuAvailable
 
     def getGpuMemory(self) -> Tuple[float, float]:  # Return (total GB, available GB)
         try:
             if self.mpsAvailable:
                 total_bytes = psutil.virtual_memory().total
                 free_bytes = psutil.virtual_memory().available
                 return float(toGB(total_bytes)), float(toGB(free_bytes))
+            if self.xpuAvailable:
+                free_bytes, total_bytes = torch.xpu.mem_get_info()
+                return float(toGB(total_bytes)), float(toGB(free_bytes))
             if self.cudaAvailable:
                 free_bytes, total_bytes = torch.cuda.mem_get_info()
                 return float(toGB(total_bytes)), float(toGB(free_bytes))
@@ -63,6 +83,9 @@ def getGpuMemeoryTotal(self) -> float:
             if self.mpsAvailable:
                 total_bytes = psutil.virtual_memory().total
                 return float(toGB(total_bytes))
+            if self.xpuAvailable:
+                free_bytes, total_bytes = torch.xpu.mem_get_info()
+                return float(toGB(total_bytes))
             if self.cudaAvailable:
                 free_bytes, total_bytes = torch.cuda.mem_get_info()
                 return float(toGB(total_bytes))
@@ -75,6 +98,9 @@ def getGpuMemoryAvailable(self) -> float:
             if self.mpsAvailable:
                 free_bytes = psutil.virtual_memory().available
                 return float(toGB(free_bytes))
+            if self.xpuAvailable:
+                free_bytes, total_bytes = torch.xpu.mem_get_info()
+                return float(toGB(free_bytes))
             if self.cudaAvailable:
                 free_bytes, total_bytes = torch.cuda.mem_get_info()
                 return float(toGB(free_bytes))
@@ -86,6 +112,13 @@ def getGpuUtilization(self) -> float:
         try:
             if self.mpsAvailable:
                 return None
+            if self.xpuAvailable and IPEX_AVAILABLE:
+                try:
+                    xpu_utilization = ipex.xpu.get_device_utilization(0) / 100.0
+                    return xpu_utilization
+                except Exception as e:
+                    pass
+                return None
             if self.cudaAvailable and self.nvmlHandle:
                 gpu_utilization = pynvml.nvmlDeviceGetUtilizationRates(self.nvmlHandle).gpu / 100.0
                 return gpu_utilization

src/setup/probeGPU.py

@@ -41,11 +41,24 @@ def has_mps_support():
         pass
 
     return False
+
+def has_intel_gpu():
+    """Check if Intel GPU is available."""
+    try:
+        result = subprocess.run(['xpu-smi', 'discovery'], capture_output=True, text=True, timeout=5)
+        if result.returncode == 0 and 'Device' in result.stdout:
+            return True
+    except (subprocess.CalledProcessError, FileNotFoundError, subprocess.TimeoutExpired):
+        pass
+    return False
 
 if __name__ == "__main__":
     # First check for MPS on macOS
     if has_mps_support():
         print("mps", end="")
+    # Check for Intel GPU
+    elif has_intel_gpu():
+        print("xpu", end="")
     else:
         # Check for CUDA
         cuda_version = get_cuda_version()