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, 17:15 UTC
Revert "Fix non-contiguous input passed to Marlin kernel (#15319)" (#15398)
Tyler Michael Smith committed
Mar 24, 2025 at 23:43 UTC
b5269db959e8f33a89b6d4670f688bf8033aa52b
1 file changed
-4
vllm/model_executor/layers/quantization/kernels/mixed_precision/marlin.py
-4
@@ -115,10 +115,6 @@ 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
-
118
c = self.config
119
w_q, w_s, w_zp, w_gidx = self._get_weight_params(layer)
120