Signed-off-by: Andreas Karatzas <akaratza@amd.com>
This commit is contained in:
@@ -158,6 +158,12 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
|
||||
intermediate_size_per_partition_after_pad
|
||||
)
|
||||
|
||||
# CK (gfx950) padding info for rocm_aiter_ops.fused_moe()
|
||||
self.hidden_pad = extra_weight_attrs.get("hidden_pad", 0)
|
||||
self.intermediate_pad = (
|
||||
intermediate_size_per_partition_after_pad - intermediate_size_per_partition
|
||||
)
|
||||
|
||||
# Fused gate_up_proj (column parallel)
|
||||
w13_weight = torch.nn.Parameter(
|
||||
torch.zeros(
|
||||
@@ -362,6 +368,8 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
|
||||
w2_scale=w2_scale,
|
||||
w1_bias=w1_bias,
|
||||
w2_bias=w2_bias,
|
||||
hidden_pad=self.hidden_pad,
|
||||
intermediate_pad=self.intermediate_pad,
|
||||
)
|
||||
|
||||
def select_gemm_impl(
|
||||
|
||||
Reference in New Issue
Block a user