From 55842a8d6961204f5adbcb5ca07da8cf79d85c33 Mon Sep 17 00:00:00 2001 From: Xinyu Chen Date: Sat, 18 Apr 2026 08:53:56 +0800 Subject: [PATCH] [XPU]fake impl for xpu fp8_gemm (#39984) Signed-off-by: Xinyu Chen Co-authored-by: Kunshang Ji --- vllm/_xpu_ops.py | 17 +++++++++++++++++ 1 file changed, 17 insertions(+) diff --git a/vllm/_xpu_ops.py b/vllm/_xpu_ops.py index a0a321173d..7db074bf92 100644 --- a/vllm/_xpu_ops.py +++ b/vllm/_xpu_ops.py @@ -22,6 +22,23 @@ else: except ImportError: from torch.library import impl_abstract as register_fake +if hasattr(torch.ops._xpu_C, "fp8_gemm"): + + @register_fake("_xpu_C::fp8_gemm") + def _fp8_gemm_fake( + q_input: torch.Tensor, + q_weight: torch.Tensor, + out_dtype: torch.dtype, + input_scales: torch.Tensor, + weight_scale: torch.Tensor, + bias: torch.Tensor | None = None, + ) -> torch.Tensor: + input_2d = q_input.view(-1, q_input.shape[-1]) + M = input_2d.size(0) + N = q_weight.size(1) + return torch.empty((M, N), dtype=out_dtype, device=q_input.device) + + if hasattr(torch.ops._xpu_C, "fp8_gemm_w8a16"): @register_fake("_xpu_C::fp8_gemm_w8a16")