Merge pull request #14171 from Nuullll/ipex

Initial IPEX support for Intel Arc GPU

5 files changed, 88 insertions, 2 deletions

diff --git a/modules/cmd_args.py b/modules/cmd_args.py
index a9fb9bfa..da93eb26 100644
--- a/modules/cmd_args.py
+++ b/modules/cmd_args.py
@@ -70,6 +70,7 @@ parser.add_argument("--opt-sdp-no-mem-attention", action='store_true', help="pre
 parser.add_argument("--disable-opt-split-attention", action='store_true', help="prefer no cross-attention layer optimization for automatic choice of optimization")

 parser.add_argument("--disable-nan-check", action='store_true', help="do not check if produced images/latent spaces have nans; useful for running without a checkpoint in CI")

 parser.add_argument("--use-cpu", nargs='+', help="use CPU as torch device for specified modules", default=[], type=str.lower)

+parser.add_argument("--use-ipex", action="store_true", help="use Intel XPU as torch device")

 parser.add_argument("--disable-model-loading-ram-optimization", action='store_true', help="disable an optimization that reduces RAM use when loading a model")

 parser.add_argument("--listen", action='store_true', help="launch gradio with 0.0.0.0 as server name, allowing to respond to network requests")

 parser.add_argument("--port", type=int, help="launch gradio with given server port, you need root/admin rights for ports < 1024, defaults to 7860 if available", default=None)

diff --git a/modules/devices.py b/modules/devices.py
index f97f9cfb..ea1f712f 100644
--- a/modules/devices.py
+++ b/modules/devices.py
@@ -8,6 +8,13 @@ from modules import errors, shared
 if sys.platform == "darwin":
     from modules import mac_specific
 
+if shared.cmd_opts.use_ipex:
+    from modules import xpu_specific
+
+
+def has_xpu() -> bool:
+    return shared.cmd_opts.use_ipex and xpu_specific.has_xpu
+
 
 def has_mps() -> bool:
     if sys.platform != "darwin":
@@ -30,6 +37,9 @@ def get_optimal_device_name():
     if has_mps():
         return "mps"
 
+    if has_xpu():
+        return xpu_specific.get_xpu_device_string()
+
     return "cpu"
 
 
@@ -54,6 +64,9 @@ def torch_gc():
     if has_mps():
         mac_specific.torch_mps_gc()
 
+    if has_xpu():
+        xpu_specific.torch_xpu_gc()
+
 
 def enable_tf32():
     if torch.cuda.is_available():
diff --git a/modules/launch_utils.py b/modules/launch_utils.py
index 264ec9ca..586cdc7e 100644
--- a/modules/launch_utils.py
+++ b/modules/launch_utils.py
@@ -310,6 +310,26 @@ def requirements_met(requirements_file):
 def prepare_environment():

     torch_index_url = os.environ.get('TORCH_INDEX_URL', "https://download.pytorch.org/whl/cu118")

     torch_command = os.environ.get('TORCH_COMMAND', f"pip install torch==2.0.1 torchvision==0.15.2 --extra-index-url {torch_index_url}")

+    if args.use_ipex:

+        if platform.system() == "Windows":

+            # The "Nuullll/intel-extension-for-pytorch" wheels were built from IPEX source for Intel Arc GPU: https://github.com/intel/intel-extension-for-pytorch/tree/xpu-main

+            # This is NOT an Intel official release so please use it at your own risk!!

+            # See https://github.com/Nuullll/intel-extension-for-pytorch/releases/tag/v2.0.110%2Bxpu-master%2Bdll-bundle for details.

+            #

+            # Strengths (over official IPEX 2.0.110 windows release):

+            #   - AOT build (for Arc GPU only) to eliminate JIT compilation overhead: https://github.com/intel/intel-extension-for-pytorch/issues/399

+            #   - Bundles minimal oneAPI 2023.2 dependencies into the python wheels, so users don't need to install oneAPI for the whole system.

+            #   - Provides a compatible torchvision wheel: https://github.com/intel/intel-extension-for-pytorch/issues/465

+            # Limitation:

+            #   - Only works for python 3.10

+            url_prefix = "https://github.com/Nuullll/intel-extension-for-pytorch/releases/download/v2.0.110%2Bxpu-master%2Bdll-bundle"

+            torch_command = os.environ.get('TORCH_COMMAND', f"pip install {url_prefix}/torch-2.0.0a0+gite9ebda2-cp310-cp310-win_amd64.whl {url_prefix}/torchvision-0.15.2a0+fa99a53-cp310-cp310-win_amd64.whl {url_prefix}/intel_extension_for_pytorch-2.0.110+gitc6ea20b-cp310-cp310-win_amd64.whl")

+        else:

+            # Using official IPEX release for linux since it's already an AOT build.

+            # However, users still have to install oneAPI toolkit and activate oneAPI environment manually.

+            # See https://intel.github.io/intel-extension-for-pytorch/index.html#installation for details.

+            torch_index_url = os.environ.get('TORCH_INDEX_URL', "https://pytorch-extension.intel.com/release-whl/stable/xpu/us/")

+            torch_command = os.environ.get('TORCH_COMMAND', f"pip install torch==2.0.0a0 intel-extension-for-pytorch==2.0.110+gitba7f6c1 --extra-index-url {torch_index_url}")

     requirements_file = os.environ.get('REQS_FILE', "requirements_versions.txt")

 

     xformers_package = os.environ.get('XFORMERS_PACKAGE', 'xformers==0.0.20')

@@ -352,6 +372,8 @@ def prepare_environment():
         run(f'"{python}" -m {torch_command}', "Installing torch and torchvision", "Couldn't install torch", live=True)

         startup_timer.record("install torch")

 

+    if args.use_ipex:

+        args.skip_torch_cuda_test = True

     if not args.skip_torch_cuda_test and not check_run_python("import torch; assert torch.cuda.is_available()"):

         raise RuntimeError(

             'Torch is not able to use GPU; '

diff --git a/modules/sd_samplers_timesteps_impl.py b/modules/sd_samplers_timesteps_impl.py
index a72daafd..930a64af 100644
--- a/modules/sd_samplers_timesteps_impl.py
+++ b/modules/sd_samplers_timesteps_impl.py
@@ -11,7 +11,7 @@ from modules.models.diffusion.uni_pc import uni_pc
 def ddim(model, x, timesteps, extra_args=None, callback=None, disable=None, eta=0.0):

     alphas_cumprod = model.inner_model.inner_model.alphas_cumprod

     alphas = alphas_cumprod[timesteps]

-    alphas_prev = alphas_cumprod[torch.nn.functional.pad(timesteps[:-1], pad=(1, 0))].to(torch.float64 if x.device.type != 'mps' else torch.float32)

+    alphas_prev = alphas_cumprod[torch.nn.functional.pad(timesteps[:-1], pad=(1, 0))].to(torch.float64 if x.device.type != 'mps' and x.device.type != 'xpu' else torch.float32)

     sqrt_one_minus_alphas = torch.sqrt(1 - alphas)

     sigmas = eta * np.sqrt((1 - alphas_prev.cpu().numpy()) / (1 - alphas.cpu()) * (1 - alphas.cpu() / alphas_prev.cpu().numpy()))

 

@@ -43,7 +43,7 @@ def ddim(model, x, timesteps, extra_args=None, callback=None, disable=None, eta=
 def plms(model, x, timesteps, extra_args=None, callback=None, disable=None):

     alphas_cumprod = model.inner_model.inner_model.alphas_cumprod

     alphas = alphas_cumprod[timesteps]

-    alphas_prev = alphas_cumprod[torch.nn.functional.pad(timesteps[:-1], pad=(1, 0))].to(torch.float64 if x.device.type != 'mps' else torch.float32)

+    alphas_prev = alphas_cumprod[torch.nn.functional.pad(timesteps[:-1], pad=(1, 0))].to(torch.float64 if x.device.type != 'mps' and x.device.type != 'xpu' else torch.float32)

     sqrt_one_minus_alphas = torch.sqrt(1 - alphas)

 

     extra_args = {} if extra_args is None else extra_args

diff --git a/modules/xpu_specific.py b/modules/xpu_specific.py
new file mode 100644
index 00000000..d933c790
--- /dev/null
+++ b/modules/xpu_specific.py
@@ -0,0 +1,50 @@
+from modules import shared
+from modules.sd_hijack_utils import CondFunc
+
+has_ipex = False
+try:
+    import torch
+    import intel_extension_for_pytorch as ipex # noqa: F401
+    has_ipex = True
+except Exception:
+    pass
+
+
+def check_for_xpu():
+    return has_ipex and hasattr(torch, 'xpu') and torch.xpu.is_available()
+
+
+def get_xpu_device_string():
+    if shared.cmd_opts.device_id is not None:
+        return f"xpu:{shared.cmd_opts.device_id}"
+    return "xpu"
+
+
+def torch_xpu_gc():
+    with torch.xpu.device(get_xpu_device_string()):
+        torch.xpu.empty_cache()
+
+
+has_xpu = check_for_xpu()
+
+if has_xpu:
+    # W/A for https://github.com/intel/intel-extension-for-pytorch/issues/452: torch.Generator API doesn't support XPU device
+    CondFunc('torch.Generator',
+        lambda orig_func, device=None: torch.xpu.Generator(device),
+        lambda orig_func, device=None: device is not None and device.type == "xpu")
+
+    # W/A for some OPs that could not handle different input dtypes
+    CondFunc('torch.nn.functional.layer_norm',
+        lambda orig_func, input, normalized_shape=None, weight=None, *args, **kwargs:
+        orig_func(input.to(weight.data.dtype), normalized_shape, weight, *args, **kwargs),
+        lambda orig_func, input, normalized_shape=None, weight=None, *args, **kwargs:
+        weight is not None and input.dtype != weight.data.dtype)
+    CondFunc('torch.nn.modules.GroupNorm.forward',
+        lambda orig_func, self, input: orig_func(self, input.to(self.weight.data.dtype)),
+        lambda orig_func, self, input: input.dtype != self.weight.data.dtype)
+    CondFunc('torch.nn.modules.linear.Linear.forward',
+        lambda orig_func, self, input: orig_func(self, input.to(self.weight.data.dtype)),
+        lambda orig_func, self, input: input.dtype != self.weight.data.dtype)
+    CondFunc('torch.nn.modules.conv.Conv2d.forward',
+        lambda orig_func, self, input: orig_func(self, input.to(self.weight.data.dtype)),
+        lambda orig_func, self, input: input.dtype != self.weight.data.dtype)