A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 15, 18:30 UTC
Fix non-contiguous input passed to Marlin kernel (#15319)
Qubitium-ModelCloud committed
Mar 24, 2025 at 11:09 UTC
d20e261199d3e4bffe9ee0e495ba7f7454229f11
1 file changed
+4
vllm/model_executor/layers/quantization/kernels/mixed_precision/marlin.py
+4
@@ -115,6 +115,10 @@ class MarlinLinearKernel(MPLinearKernel):
115
layer: torch.nn.Module,
116
x: torch.Tensor,
117
bias: Optional[torch.Tensor] = None) -> torch.Tensor:
118
+ # marlin requires contiguous memory layout
119
+ # prefix caching may cause x to be non-contiguous
120
+ x = x.contiguous() # no-op if already contiguous
121
+
122
c = self.config
123
w_q, w_s, w_zp, w_gidx = self._get_weight_params(layer)
124