From d20e261199d3e4bffe9ee0e495ba7f7454229f11 Mon Sep 17 00:00:00 2001 From: Qubitium-ModelCloud Date: Mon, 24 Mar 2025 11:09:44 +0800 Subject: [PATCH] Fix non-contiguous input passed to Marlin kernel (#15319) --- .../layers/quantization/kernels/mixed_precision/marlin.py | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/vllm/model_executor/layers/quantization/kernels/mixed_precision/marlin.py b/vllm/model_executor/layers/quantization/kernels/mixed_precision/marlin.py index e21801cf6a..b030e1484a 100644 --- a/vllm/model_executor/layers/quantization/kernels/mixed_precision/marlin.py +++ b/vllm/model_executor/layers/quantization/kernels/mixed_precision/marlin.py @@ -115,6 +115,10 @@ class MarlinLinearKernel(MPLinearKernel): layer: torch.nn.Module, x: torch.Tensor, bias: Optional[torch.Tensor] = None) -> torch.Tensor: + # marlin requires contiguous memory layout + # prefix caching may cause x to be non-contiguous + x = x.contiguous() # no-op if already contiguous + c = self.config w_q, w_s, w_zp, w_gidx = self._get_weight_params(layer)