Add dtype=torch.float32 for gate bias#963

dmtri35 wants to merge 2 commits intodeepseek-ai:maindeepseek-ai/DeepSeek-V3:mainfrom

dmtri35:trid/f32-gate-biasdmtri35/DeepSeek-V3:trid/f32-gate-biasCopy head branch name to clipboard