[ROCm][Bugfix] Fix init-time bias dtype cast when gate.out_dtype is None (#41405)

Signed-off-by: Rita Brugarolas Brufau <rita.brugarolasbrufau@amd.com>
This commit is contained in:
Rita Brugarolas
2026-05-02 00:13:15 -04:00
committed by GitHub
parent d58c42e19c
commit c293ccc58e
+2 -1
View File
@@ -351,8 +351,9 @@ class DeepseekV2MoE(nn.Module):
self.is_rocm_aiter_moe_enabled
and self.gate.e_score_correction_bias is not None
):
gate_out_dtype = self.gate.out_dtype or self.gate.weight.dtype
self.gate.e_score_correction_bias.data = (
self.gate.e_score_correction_bias.data.to(self.gate.out_dtype)
self.gate.e_score_correction_bias.data.to(gate_out_dtype)
)
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: