From b23bd73f540175f9e117eaee5029cd7d8df63964 Mon Sep 17 00:00:00 2001 From: Liangqiusong Date: Mon, 20 Jul 2026 23:32:54 +0800 Subject: [PATCH] [XPU]add sycl path for Mhc (#47245) Signed-off-by: root Co-authored-by: Kunshang Ji --- vllm/model_executor/layers/mhc.py | 25 +++++-------------------- 1 file changed, 5 insertions(+), 20 deletions(-) diff --git a/vllm/model_executor/layers/mhc.py b/vllm/model_executor/layers/mhc.py index 23733f7693..da6bbf2e00 100644 --- a/vllm/model_executor/layers/mhc.py +++ b/vllm/model_executor/layers/mhc.py @@ -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, )