performance-tensor-parallel-alig-runtimeerror-size-k-must-divisible-by-block-size-k-579b764c

THEN Align the K dimension of activation and weight tensors to the kernel's BLOCK_SIZE_K (typically 64) before calling the MoE WNA16 GEMM. This can be done by padding the activation tensor's K dimension in Python using torch.nn.functional.pad, and by padding the weight tensors (B, B_scale, B_zp) during model loading or offline transformation to avoid runtime overhead.

if size_k % BLOCK_SIZE_K != 0:
    pad_amount = BLOCK_SIZE_K - (size_k % BLOCK_SIZE_K)
    A = torch.nn.functional.pad(A, (0, pad_amount), 'constant', 0)  # Pad activation
    B = torch.nn.functional.pad(B, (0, pad_amount), 'constant', 0)  # Pad weight (preferably once, offline)