diff --git a/vllm/model_executor/layers/quantization/fp8.py b/vllm/model_executor/layers/quantization/fp8.py index 869fbf7523..d4ec1c093a 100644 --- a/vllm/model_executor/layers/quantization/fp8.py +++ b/vllm/model_executor/layers/quantization/fp8.py @@ -320,7 +320,7 @@ class Fp8LinearMethod(LinearMethodBase): def create_weights( self, - layer: RoutedExperts, + layer: torch.nn.Module, input_size_per_partition: int, output_partition_sizes: list[int], input_size: int, @@ -394,7 +394,7 @@ class Fp8LinearMethod(LinearMethodBase): self.use_marlin = isinstance(self.fp8_linear, MarlinFP8ScaledMMLinearKernel) - def process_weights_after_loading(self, layer: RoutedExperts) -> None: + def process_weights_after_loading(self, layer: torch.nn.Module) -> None: if self.use_marlin: if not self.block_quant: # Canonicalize to (K, N) for the kernel.