[Kernel] Added flashinfer fp8 per-tensor gemms (#22895)
Signed-off-by: Julien Lin <jullin@nvidia.com> Co-authored-by: Michael Goin <mgoin64@gmail.com>
This commit is contained in:
@@ -97,8 +97,8 @@ class PTPCFp8LinearMethod(Fp8LinearMethod):
|
||||
self.quant_config.is_checkpoint_fp8_serialized = False
|
||||
self.fp8_linear = Fp8LinearOp(
|
||||
act_quant_static=False,
|
||||
cutlass_fp8_supported=False,
|
||||
act_quant_group_shape=GroupShape.PER_TOKEN)
|
||||
act_quant_group_shape=GroupShape.PER_TOKEN,
|
||||
force_fp8_e4m3fnuz=True)
|
||||
|
||||
def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
|
||||
layer.weight = torch.nn.Parameter(layer.weight.data,
|
||||
|
||||
Reference in New Issue
Block a user