[DSV4][XPU] Add MHC fused_post_pre support (#44144)

Signed-off-by: Ma Jian <jian1.ma@intel.com>
This commit is contained in:
Ma Jian
2026-06-09 17:23:17 +08:00
committed by GitHub
parent 2385e140d6
commit 70db1488c5
2 changed files with 112 additions and 17 deletions
+72 -3
View File
@@ -466,7 +466,76 @@ class MHCFusedPostPreOp(CustomOp):
norm_eps,
)
def forward_native(self, *args, **kwargs):
raise NotImplementedError(
"Native implementation of mhc_fused_post_pre is not available"
def forward_native(
self,
x: torch.Tensor,
residual: torch.Tensor,
post_layer_mix: torch.Tensor,
comb_res_mix: torch.Tensor,
fn: torch.Tensor,
hc_scale: torch.Tensor,
hc_base: torch.Tensor,
rms_eps: float,
hc_pre_eps: float,
hc_sinkhorn_eps: float,
hc_post_mult_value: float,
sinkhorn_repeat: int,
n_splits: int = 1,
tile_n: int = 1,
norm_weight: torch.Tensor | None = None,
norm_eps: float = 0.0,
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
# Decompose into post + pre (no fused kernel available).
residual_cur = mhc_kernels.mhc_post_torch(
x, residual, post_layer_mix, comb_res_mix
)
post_mix_cur, comb_mix_cur, layer_input_cur = mhc_kernels.mhc_pre_torch(
residual_cur,
fn,
hc_scale,
hc_base,
rms_eps,
hc_pre_eps,
hc_sinkhorn_eps,
hc_post_mult_value,
sinkhorn_repeat,
)
return residual_cur, post_mix_cur, comb_mix_cur, layer_input_cur
def forward_xpu(
self,
x: torch.Tensor,
residual: torch.Tensor,
post_layer_mix: torch.Tensor,
comb_res_mix: torch.Tensor,
fn: torch.Tensor,
hc_scale: torch.Tensor,
hc_base: torch.Tensor,
rms_eps: float,
hc_pre_eps: float,
hc_sinkhorn_eps: float,
hc_post_mult_value: float,
sinkhorn_repeat: int,
n_splits: int = 1,
tile_n: int = 1,
norm_weight: torch.Tensor | None = None,
norm_eps: float = 0.0,
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
return self.forward_native(
x,
residual,
post_layer_mix,
comb_res_mix,
fn,
hc_scale,
hc_base,
rms_eps,
hc_pre_eps,
hc_sinkhorn_eps,
hc_post_mult_value,
sinkhorn_repeat,
n_splits,
tile_n,
norm_weight,
norm_eps,
)