[XPU]add sycl path for Mhc (#47245)

Signed-off-by: root <xiaolong.guo@intel.com>
Co-authored-by: Kunshang Ji <kunshang.ji@intel.com>
This commit is contained in:
Liangqiusong
2026-07-20 15:32:54 +00:00
committed by GitHub
co-authored by Kunshang Ji
parent e2d7adeb64
commit b23bd73f54
+5 -20
View File
@@ -175,7 +175,7 @@ class MHCPreOp(CustomOp):
norm_weight: torch.Tensor | None = None,
norm_eps: float = 0.0,
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
return self.forward_native(
return torch.ops._xpu_C.mhc_pre(
residual,
fn,
hc_scale,
@@ -185,9 +185,6 @@ class MHCPreOp(CustomOp):
hc_sinkhorn_eps,
hc_post_mult_value,
sinkhorn_repeat,
n_splits,
norm_weight,
norm_eps,
)
@@ -260,7 +257,7 @@ class MHCPostOp(CustomOp):
post_layer_mix: torch.Tensor,
comb_res_mix: torch.Tensor,
) -> torch.Tensor:
return self.forward_native(
return torch.ops._xpu_C.mhc_post(
x,
residual,
post_layer_mix,
@@ -369,16 +366,8 @@ class HCHeadOp(CustomOp):
out = torch.empty(
num_tokens, hidden_size, dtype=torch.bfloat16, device=hidden_states.device
)
torch.ops.vllm.hc_head_triton(
hs_flat,
hc_fn,
hc_scale,
hc_base,
out,
hidden_size,
rms_norm_eps,
hc_eps,
hc_mult,
torch.ops._xpu_C.hc_head_fused(
hs_flat, hc_fn, hc_scale, hc_base, out, rms_norm_eps, hc_eps
)
return out.view(*outer_shape, hidden_size)
@@ -548,7 +537,7 @@ class MHCFusedPostPreOp(CustomOp):
norm_weight: torch.Tensor | None = None,
norm_eps: float = 0.0,
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
return self.forward_native(
return torch.ops._xpu_C.mhc_fused_post_pre(
x,
residual,
post_layer_mix,
@@ -561,8 +550,4 @@ class MHCFusedPostPreOp(CustomOp):
hc_sinkhorn_eps,
hc_post_mult_value,
sinkhorn_repeat,
n_splits,
tile_n,
norm_weight,
norm_eps,
)