[XPU]add sycl path for Mhc (#47245)
Signed-off-by: root <xiaolong.guo@intel.com> Co-authored-by: Kunshang Ji <kunshang.ji@intel.com>
Liangqiusong committed
Jul 20, 2026 at 23:32 UTC
b23bd73f540175f9e117eaee5029cd7d8df63964
1 file changed
+5
-20
vllm/model_executor/layers/mhc.py
+5
-20
@@ -175,7 +175,7 @@ class MHCPreOp(CustomOp):
175
norm_weight: torch.Tensor | None = None,
176
norm_eps: float = 0.0,
177
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
178
- return self.forward_native(
178
+ return torch.ops._xpu_C.mhc_pre(
179
residual,
180
fn,
181
hc_scale,
@@ -185,9 +185,6 @@ class MHCPreOp(CustomOp):
185
hc_sinkhorn_eps,
186
hc_post_mult_value,
187
sinkhorn_repeat,
188
- n_splits,
189
- norm_weight,
190
- norm_eps,
188
)
189
190
@@ -260,7 +257,7 @@ class MHCPostOp(CustomOp):
257
post_layer_mix: torch.Tensor,
258
comb_res_mix: torch.Tensor,
259
) -> torch.Tensor:
263
- return self.forward_native(
260
+ return torch.ops._xpu_C.mhc_post(
261
x,
262
residual,
263
post_layer_mix,
@@ -369,16 +366,8 @@ class HCHeadOp(CustomOp):
366
out = torch.empty(
367
num_tokens, hidden_size, dtype=torch.bfloat16, device=hidden_states.device
368
)
372
- torch.ops.vllm.hc_head_triton(
373
- hs_flat,
374
- hc_fn,
375
- hc_scale,
376
- hc_base,
377
- out,
378
- hidden_size,
379
- rms_norm_eps,
380
- hc_eps,
381
- hc_mult,
369
+ torch.ops._xpu_C.hc_head_fused(
370
+ hs_flat, hc_fn, hc_scale, hc_base, out, rms_norm_eps, hc_eps
371
)
372
return out.view(*outer_shape, hidden_size)
373
@@ -548,7 +537,7 @@ class MHCFusedPostPreOp(CustomOp):
537
norm_weight: torch.Tensor | None = None,
538
norm_eps: float = 0.0,
539
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
551
- return self.forward_native(
540
+ return torch.ops._xpu_C.mhc_fused_post_pre(
541
x,
542
residual,
543
post_layer_mix,
@@ -561,8 +550,4 @@ class MHCFusedPostPreOp(CustomOp):
550
hc_sinkhorn_eps,
551
hc_post_mult_value,
552
sinkhorn_repeat,
564
- n_splits,
565
- tile_n,
566
- norm_weight,
567
- norm_eps,
553
)