[Perf] [XPU] diffusion: Use the AOT SYCL fused residual-norm-scale-shift kernel (#41401)

Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
This commit is contained in:
Chetan Kumar Verma
2026-10-23 14:32:25 +08:00
committed by GitHub
co-authored by Claude Opus 5.5
parent 904691e0b5
commit 76c382e80f
3 changed files with 114 additions and 0 deletions
@@ -636,8 +636,10 @@ _EXPORTS: dict[str, str] = {
"can_use_flux2_gated_resnorm": "norm.flux2_gated_resnorm_jit",
"flux2_gated_resnorm_raw": "norm.flux2_gated_resnorm_jit",
"FLYDSL_NORM_MIN_ALIGNED_DIM": "norm.fused_residual_norm_flydsl",
"can_use_fused_scale_residual_norm_scale_shift_sycl": "norm.scale_residual_norm_scale_shift_sycl",
"can_use_fused_scale_residual_norm_scale_shift_triton": "norm.scale_residual_norm_scale_shift_triton",
"flydsl_fused_residual_norm_scale_shift": "norm.fused_residual_norm_flydsl",
"fused_scale_residual_norm_scale_shift_sycl": "norm.scale_residual_norm_scale_shift_sycl",
"fused_scale_residual_norm_scale_shift_triton": "norm.scale_residual_norm_scale_shift_triton",
"flydsl_norm_scale_shift": "norm.fused_residual_norm_flydsl",
"apply_group_norm_silu": "norm.group_norm_silu",
@@ -0,0 +1,88 @@
import sgl_kernel # noqa: F401
import torch
from sglang.kernels.ops.diffusion.norm.scale_residual_norm_scale_shift_triton import (
can_use_fused_scale_residual_norm_scale_shift_triton,
)
from sglang.srt.utils.patch_torch import register_fake_if_exists
_fused_scale_residual_norm_scale_shift = (
torch.ops.sgl_kernel.fused_scale_residual_norm_scale_shift
)
def _param_dtype(
*, gate: torch.Tensor | int, weight: torch.Tensor | None, shift: torch.Tensor
) -> torch.dtype:
if weight is not None:
return weight.dtype
if isinstance(gate, torch.Tensor):
return gate.dtype
return shift.dtype
def can_use_fused_scale_residual_norm_scale_shift_sycl(
*,
residual: torch.Tensor,
x: torch.Tensor,
gate: torch.Tensor | int,
shift: torch.Tensor,
scale: torch.Tensor,
weight: torch.Tensor | None,
bias: torch.Tensor | None,
) -> bool:
if not can_use_fused_scale_residual_norm_scale_shift_triton(
residual=residual,
x=x,
gate=gate,
shift=shift,
scale=scale,
weight=weight,
bias=bias,
):
return False
dtypes = {p.dtype for p in (gate, weight, bias) if isinstance(p, torch.Tensor)}
if len(dtypes) > 1:
return False
param_dtype = _param_dtype(gate=gate, weight=weight, shift=shift)
return all(
m.dtype == param_dtype or param_dtype == torch.float32 for m in (shift, scale)
)
def fused_scale_residual_norm_scale_shift_sycl(
*,
residual: torch.Tensor,
x: torch.Tensor,
gate: torch.Tensor | int,
shift: torch.Tensor,
scale: torch.Tensor,
weight: torch.Tensor | None,
bias: torch.Tensor | None,
eps: float,
) -> tuple[torch.Tensor, torch.Tensor]:
param_dtype = _param_dtype(gate=gate, weight=weight, shift=shift)
return _fused_scale_residual_norm_scale_shift(
residual,
x,
gate if isinstance(gate, torch.Tensor) else None,
weight,
bias,
scale.to(param_dtype),
shift.to(param_dtype),
eps,
)
@register_fake_if_exists("sgl_kernel::fused_scale_residual_norm_scale_shift")
def _(
residual: torch.Tensor,
x: torch.Tensor,
gate: torch.Tensor | None,
weight: torch.Tensor | None,
bias: torch.Tensor | None,
scale: torch.Tensor,
shift: torch.Tensor,
eps: float,
) -> tuple[torch.Tensor, torch.Tensor]:
return torch.empty_like(x), torch.empty_like(x)
@@ -90,6 +90,11 @@ if USE_AITER:
if _is_xpu:
from sgl_kernel import fused_inplace_qknorm_rope
from sglang.kernels.ops.diffusion import (
can_use_fused_scale_residual_norm_scale_shift_sycl,
fused_scale_residual_norm_scale_shift_sycl,
)
if not _is_cpu:
from sglang.kernels.ops.diffusion import norm_infer, rms_norm_fn
@@ -727,6 +732,25 @@ class _ScaleResidualNormScaleShift(CustomOp):
if self.norm_type == "layer":
weight = self.norm.weight
bias = self.norm.bias
if can_use_fused_scale_residual_norm_scale_shift_sycl(
residual=residual,
x=x,
gate=gate,
shift=shift,
scale=scale,
weight=weight,
bias=bias,
):
return fused_scale_residual_norm_scale_shift_sycl(
residual=residual,
x=x,
gate=gate,
shift=shift,
scale=scale,
weight=weight,
bias=bias,
eps=self.eps,
)
if can_use_fused_scale_residual_norm_scale_shift_triton(
residual=residual,
x=x,