fix act fun for xpu (#23809)
This commit is contained in:
@@ -16,9 +16,11 @@ from sglang.multimodal_gen.runtime.platforms import current_platform
|
|||||||
_is_cuda = current_platform.is_cuda()
|
_is_cuda = current_platform.is_cuda()
|
||||||
_is_hip = current_platform.is_hip()
|
_is_hip = current_platform.is_hip()
|
||||||
_is_npu = current_platform.is_npu()
|
_is_npu = current_platform.is_npu()
|
||||||
|
_is_xpu = current_platform.is_xpu()
|
||||||
|
|
||||||
if _is_cuda:
|
if _is_cuda:
|
||||||
from sglang.jit_kernel.activation import silu_and_mul
|
from sglang.jit_kernel.activation import silu_and_mul
|
||||||
elif _is_hip:
|
elif _is_hip or _is_xpu:
|
||||||
from sgl_kernel import silu_and_mul
|
from sgl_kernel import silu_and_mul
|
||||||
|
|
||||||
|
|
||||||
@@ -61,6 +63,13 @@ class SiluAndMul(CustomOp):
|
|||||||
def forward_musa(self, x: torch.Tensor) -> torch.Tensor:
|
def forward_musa(self, x: torch.Tensor) -> torch.Tensor:
|
||||||
return nn.SwishGLU()(x)
|
return nn.SwishGLU()(x)
|
||||||
|
|
||||||
|
def forward_xpu(self, x: torch.Tensor) -> torch.Tensor:
|
||||||
|
d = x.shape[-1] // 2
|
||||||
|
output_shape = x.shape[:-1] + (d,)
|
||||||
|
out = torch.empty(output_shape, dtype=x.dtype, device=x.device)
|
||||||
|
silu_and_mul(x, out)
|
||||||
|
return out
|
||||||
|
|
||||||
|
|
||||||
@CustomOp.register("gelu_and_mul")
|
@CustomOp.register("gelu_and_mul")
|
||||||
class GeluAndMul(CustomOp):
|
class GeluAndMul(CustomOp):
|
||||||
|
|||||||
@@ -61,6 +61,9 @@ class RotaryEmbedding(CustomOp):
|
|||||||
def forward_cuda(self, *args, **kwargs):
|
def forward_cuda(self, *args, **kwargs):
|
||||||
return self.forward_native(*args, **kwargs)
|
return self.forward_native(*args, **kwargs)
|
||||||
|
|
||||||
|
def forward_xpu(self, *args, **kwargs):
|
||||||
|
return self.forward_native(*args, **kwargs)
|
||||||
|
|
||||||
def forward_native(
|
def forward_native(
|
||||||
self,
|
self,
|
||||||
positions: torch.Tensor,
|
positions: torch.Tensor,
|
||||||
|
|||||||
Reference in New Issue
Block a user