[Misc] Fix incorrect layer type annotation in Fp8LinearMethod (#46818)
Signed-off-by: shaojinjie.sjj <shaojinjiesjj@gmail.com> Co-authored-by: shaojinjie.sjj <shaojinjiesjj@gmail.com>
This commit is contained in:
co-authored by
shaojinjie.sjj
parent
56aa067bf0
commit
ea2ead1db3
@@ -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.
|
||||
|
||||
Reference in New Issue
Block a user