[CPU][Perf]Added tanh AOR for faster gelu activations. (#44639)

Signed-off-by: Anna Mayne <anna.mayne@arm.com>
Signed-off-by: almayne <anna.mayne@arm.com>
Co-authored-by: Li, Jiang <jiang1.li@intel.com>
This commit is contained in:
almayne
2026-06-30 23:24:40 -07:00
committed by GitHub
co-authored by Li, Jiang <jiang1.li@intel.com>
parent b446792306
commit 89e99202f2
9 changed files with 328 additions and 3 deletions
+46 -3
View File
@@ -313,8 +313,10 @@ class GELU(CustomOp):
def __init__(self):
super().__init__()
if current_platform.get_cpu_architecture() == CpuArchEnum.ARM and hasattr(
torch.ops._C, "activation_lut_bf16"
if (
current_platform.is_cpu()
and current_platform.get_cpu_architecture() == CpuArchEnum.ARM
and hasattr(torch.ops._C, "activation_lut_bf16")
):
self.op = torch.ops._C.activation_lut_bf16
else:
@@ -334,6 +336,36 @@ class GELU(CustomOp):
return self.forward_native(x)
# --8<-- [start:gelu_tanh]
@CustomOp.register("gelu_tanh")
class GELUTanh(CustomOp):
# --8<-- [end:gelu_tanh]
def __init__(self):
super().__init__()
if (
current_platform.is_cpu()
and current_platform.get_cpu_architecture() == CpuArchEnum.ARM
and hasattr(torch.ops._C, "gelu_tanh")
):
self.op = torch.ops._C.gelu_tanh
else:
self.op = None
def forward_native(self, x: torch.Tensor) -> torch.Tensor:
return F.gelu(x, approximate="tanh")
def forward_cpu(self, x: torch.Tensor) -> torch.Tensor:
if self.op:
out = torch.empty_like(x)
self.op(out, x)
return out
return self.forward_native(x)
def forward_cuda(self, x: torch.Tensor) -> torch.Tensor:
return self.forward_native(x)
# --8<-- [start:gelu_and_mul]
@CustomOp.register("gelu_and_mul")
class GeluAndMul(CustomOp):
@@ -385,6 +417,11 @@ class GeluAndMul(CustomOp):
self.op(out, x)
return out
def forward_cpu(self, x: torch.Tensor) -> torch.Tensor:
if self.op:
return self.forward_cuda(x)
return self.native(x)
def forward_xpu(self, x: torch.Tensor) -> torch.Tensor:
return self.forward_cuda(x)
@@ -739,7 +776,8 @@ _ACTIVATION_REGISTRY = LazyDict(
def _get_gelu_pytorch_tanh() -> nn.Module:
"""Get PyTorch GELU with tanh approximation, with ROCm fallback."""
"""Get PyTorch GELU with tanh approximation, with ROCm fallback
and fast GELU for ARM."""
if current_platform.is_rocm():
# TODO:[ROCm] PyTorch native GELU with tanh is unstable with torch.compile
logger.warning_once(
@@ -747,6 +785,11 @@ def _get_gelu_pytorch_tanh() -> nn.Module:
"Falling back to GELU(approximate='none')."
)
return nn.GELU(approximate="none")
if (
current_platform.is_cpu()
and current_platform.get_cpu_architecture() == CpuArchEnum.ARM
):
return GELUTanh()
return nn.GELU(approximate="tanh")