[XPU] Register fake meta kernel for fp4_gemm (#50946)
Signed-off-by: Chaojun Zhang <chaojun.zhang@intel.com> Co-authored-by: Kunshang Ji <kunshang.ji@intel.com>
Chaojun Zhang committed
Aug 6, 2026 at 17:18 UTC
276f0bb5c1afef985f3f8729fc7794db6cc177ed
1 file changed
+18
vllm/_xpu_ops.py
+18
@@ -56,6 +56,24 @@ if hasattr(torch.ops._xpu_C, "fp8_gemm_w8a16"):
56
return torch.empty((M, N), dtype=input.dtype, device=input.device)
57
58
59
+if hasattr(torch.ops._xpu_C, "fp4_gemm"):
60
+
61
+ @register_fake("_xpu_C::fp4_gemm")
62
+ def _fp4_gemm_fake(
63
+ input: torch.Tensor,
64
+ weight: torch.Tensor,
65
+ scale_act: torch.Tensor,
66
+ scale_wei: torch.Tensor,
67
+ out_dtype: torch.dtype | None = None,
68
+ bias: torch.Tensor | None = None,
69
+ ) -> torch.Tensor:
70
+ input_2d = input.view(-1, input.shape[-1])
71
+ M = input_2d.size(0)
72
+ N = weight.size(-1)
73
+ result_dtype = out_dtype if out_dtype is not None else torch.get_default_dtype()
74
+ return torch.empty((M, N), dtype=result_dtype, device=input.device)
75
+
76
+
77
if hasattr(torch.ops._xpu_C, "int4_gemm_w4a8"):
78
79
@register_fake("_xpu_C::int4_gemm_w4a8")