From 8b40f475a31109cc6ecbdc0d14a0cee9e0303291 Mon Sep 17 00:00:00 2001 From: Nuullll Date: Fri, 10 Nov 2023 11:06:26 +0800 Subject: Initial IPEX support --- modules/xpu_specific.py | 42 ++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 42 insertions(+) create mode 100644 modules/xpu_specific.py (limited to 'modules/xpu_specific.py') diff --git a/modules/xpu_specific.py b/modules/xpu_specific.py new file mode 100644 index 00000000..6417dd2d --- /dev/null +++ b/modules/xpu_specific.py @@ -0,0 +1,42 @@ +import contextlib +from modules import shared +from modules.sd_hijack_utils import CondFunc + +has_ipex = False +try: + import torch + import intel_extension_for_pytorch as ipex + has_ipex = True +except Exception: + pass + +def check_for_xpu(): + if not has_ipex: + return False + + return hasattr(torch, 'xpu') and torch.xpu.is_available() + +has_xpu = check_for_xpu() + +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 return_null_context(*args, **kwargs): # pylint: disable=unused-argument + return contextlib.nullcontext() + +if has_xpu: + CondFunc('torch.Generator', + lambda orig_func, device=None: torch.xpu.Generator(device), + lambda orig_func, device=None: device is not None and device != torch.device("cpu") and device != "cpu") + + 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) -- cgit v1.2.3 From 7499148ad4dbd3444215c843d02453f68c459707 Mon Sep 17 00:00:00 2001 From: Nuullll Date: Sat, 2 Dec 2023 14:00:46 +0800 Subject: Disable ipex autocast due to its bad perf --- modules/xpu_specific.py | 28 ++++++++++++++++++---------- 1 file changed, 18 insertions(+), 10 deletions(-) (limited to 'modules/xpu_specific.py') diff --git a/modules/xpu_specific.py b/modules/xpu_specific.py index 6417dd2d..2df68665 100644 --- a/modules/xpu_specific.py +++ b/modules/xpu_specific.py @@ -1,4 +1,3 @@ -import contextlib from modules import shared from modules.sd_hijack_utils import CondFunc @@ -10,33 +9,42 @@ try: except Exception: pass -def check_for_xpu(): - if not has_ipex: - return False - return hasattr(torch, 'xpu') and torch.xpu.is_available() +def check_for_xpu(): + return has_ipex and hasattr(torch, 'xpu') and torch.xpu.is_available() -has_xpu = check_for_xpu() 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 return_null_context(*args, **kwargs): # pylint: disable=unused-argument - return contextlib.nullcontext() + +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 != torch.device("cpu") and device != "cpu") + 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) -- cgit v1.2.3 From 87cd07b3af74c447b02570bf3963ba83ade2e203 Mon Sep 17 00:00:00 2001 From: Nuullll Date: Sat, 2 Dec 2023 15:54:25 +0800 Subject: Fix fp64 --- modules/xpu_specific.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'modules/xpu_specific.py') diff --git a/modules/xpu_specific.py b/modules/xpu_specific.py index 2df68665..d933c790 100644 --- a/modules/xpu_specific.py +++ b/modules/xpu_specific.py @@ -4,7 +4,7 @@ from modules.sd_hijack_utils import CondFunc has_ipex = False try: import torch - import intel_extension_for_pytorch as ipex + import intel_extension_for_pytorch as ipex # noqa: F401 has_ipex = True except Exception: pass -- cgit v1.2.3 From b7e0d4a7e171ee1cef73684b8423fe4a20ca7e34 Mon Sep 17 00:00:00 2001 From: AUTOMATIC1111 <16777216c@gmail.com> Date: Thu, 14 Dec 2023 09:52:23 +0300 Subject: Merge pull request #14229 from Nuullll/ipex-embedding [IPEX] Fix embedding and ControlNet --- modules/xpu_specific.py | 9 +++++++++ 1 file changed, 9 insertions(+) (limited to 'modules/xpu_specific.py') diff --git a/modules/xpu_specific.py b/modules/xpu_specific.py index d933c790..d8da94a0 100644 --- a/modules/xpu_specific.py +++ b/modules/xpu_specific.py @@ -48,3 +48,12 @@ if has_xpu: 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) + CondFunc('torch.bmm', + lambda orig_func, input, mat2, out=None: orig_func(input.to(mat2.dtype), mat2, out=out), + lambda orig_func, input, mat2, out=None: input.dtype != mat2.dtype) + CondFunc('torch.cat', + lambda orig_func, tensors, dim=0, out=None: orig_func([t.to(tensors[0].dtype) for t in tensors], dim=dim, out=out), + lambda orig_func, tensors, dim=0, out=None: not all(t.dtype == tensors[0].dtype for t in tensors)) + CondFunc('torch.nn.functional.scaled_dot_product_attention', + lambda orig_func, query, key, value, attn_mask=None, dropout_p=0.0, is_causal=False: orig_func(query, key.to(query.dtype), value.to(query.dtype), attn_mask, dropout_p, is_causal), + lambda orig_func, query, key, value, attn_mask=None, dropout_p=0.0, is_causal=False: query.dtype != key.dtype or query.dtype != value.dtype) -- cgit v1.2.3