From 26c754d8475e753928d57b71c2398515eabe0a0d Mon Sep 17 00:00:00 2001 From: Ma Jian Date: Mon, 6 Jul 2026 17:15:26 +0800 Subject: [PATCH] [XPU][Bugfix] Do not transpose weight_scale_inv at load time (#47116) Signed-off-by: Ma Jian Co-authored-by: Kunshang Ji --- .../kernels/linear/scaled_mm/xpu.py | 26 ++----------------- 1 file changed, 2 insertions(+), 24 deletions(-) diff --git a/vllm/model_executor/kernels/linear/scaled_mm/xpu.py b/vllm/model_executor/kernels/linear/scaled_mm/xpu.py index 3d23b02bea..f30d6ced1d 100644 --- a/vllm/model_executor/kernels/linear/scaled_mm/xpu.py +++ b/vllm/model_executor/kernels/linear/scaled_mm/xpu.py @@ -197,29 +197,6 @@ class XPUFp8BlockScaledMMKernel(Fp8BlockScaledMMLinearKernel): return False, "XPUFp8BlockScaledMM only support on XPU" return True, None - def process_weights_after_loading(self, layer: torch.nn.Module): - super().process_weights_after_loading(layer) - scale_attr = ( - "weight_scale_inv" if hasattr(layer, "weight_scale_inv") else "weight_scale" - ) - scale = getattr(layer, scale_attr) - # Models with scale_fmt=ue8m0 (e.g. DeepSeek-V4) store weight scales - # as float8_e8m0fnu. The oneDNN fp8_gemm kernel dispatches to its - # "block quant" path only when NEITHER scale is e8m0: - # - # is_block_quant = (m1_sc != e8m0) && (m2_sc != e8m0) && ... - # - # Since activation scales are always float32 (use_ue8m0=False on XPU, - # DeepGEMM requires Hopper/Blackwell), an e8m0 weight scale causes - # is_block_quant=false and falls into the wrong per-channel path, - # producing NaN. Converting e8m0→float32 here at load time (one-time, - # negligible overhead for small scale tensors) ensures the kernel sees - # matching dtypes and correctly enters the block-quant path with the - # actual group_size derived from scale tensor shapes. - if scale.dtype == torch.float8_e8m0fnu: - scale = scale.to(torch.float32) - replace_parameter(layer, scale_attr, scale.data.t().contiguous()) - def apply_block_scaled_mm( self, A: torch.Tensor, @@ -228,11 +205,12 @@ class XPUFp8BlockScaledMMKernel(Fp8BlockScaledMMLinearKernel): Bs: torch.Tensor, ) -> torch.Tensor: # Weight is [N, K]. Use .t() to create a [K, N] view without copying. + # Bs is [N/128, K/128] — transpose to [K/128, N/128] for oneDNN. return torch.ops._xpu_C.fp8_gemm( A, B.t(), self.config.out_dtype, As, - Bs, + Bs.t().contiguous(), torch.Tensor(), )