EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

adapt to dsv32 on dcu

Browse Source
This commit is contained in:
maxiao
2025-09-30 18:37:31 +08:00
parent 8f7453e3af
commit 852a49c5cc
159 changed files with 7211 additions and 7782 deletions
Download Patch File Download Diff File Expand all files Collapse all files

55
python/sglang/srt/layers/layernorm.py
View File

@@ -127,45 +127,21 @@ class RMSNorm(CustomOp):
return output, residual_out
return rms_norm(x, self.weight.data, self.variance_epsilon)
# def forward_hip(
# self,
# x: torch.Tensor,
# residual: Optional[torch.Tensor] = None,
# ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
# if not x.is_contiguous():
# # NOTE: Remove this if aiter kernel supports discontinuous input
# x = x.contiguous()
# if residual is not None:
# if _vllm_version < Version("0.9"):
# fused_add_rms_norm(x, residual, self.weight.data, self.variance_epsilon)
# return x, residual
# else:
# residual_out = torch.empty_like(x)
# output = torch.empty_like(x)
# fused_add_rms_norm(
# output,
# x,
# residual_out,
# residual,
# self.weight.data,
# self.variance_epsilon,
# )
# return output, residual_out
# out = torch.empty_like(x)
# rms_norm(out, x, self.weight.data, self.variance_epsilon)
# return out
def forward_hip(
self,
x: torch.Tensor,
residual: Optional[torch.Tensor] = None,
):
self,
x: torch.Tensor,
residual: Optional[torch.Tensor] = None,
) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
if not x.is_contiguous():
# NOTE: Remove this if aiter kernel supports discontinuous input
x = x.contiguous()
if residual is not None:
try:
output = torch.empty_like(x)
if _vllm_version < Version("0.9"):
fused_add_rms_norm(x, residual, self.weight.data, self.variance_epsilon)
return x, residual
else:
residual_out = torch.empty_like(x)
output = torch.empty_like(x)
fused_add_rms_norm(
output,
x,
@@ -175,21 +151,10 @@ class RMSNorm(CustomOp):
self.variance_epsilon,
)
return output, residual_out
except TypeError:
fused_add_rms_norm(
x,
residual,
self.weight.data,
self.variance_epsilon,
)
return x, residual
out = torch.empty_like(x)
rms_norm(out, x, self.weight.data, self.variance_epsilon)
return out
def forward_native(
self,
x: torch.Tensor,