[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:
nvjullin
2025-08-26 06:54:04 -07:00
committed by GitHub
co-authored by Michael Goin
parent b78bed1bc5
commit f66673a39d
9 changed files with 198 additions and 36 deletions
@@ -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,