Method forward

models/quantization.py:278–282

(ctx, input, qweight, scales, qzeros, g_idx, bits, maxq)

    @staticmethod
    @custom_fwd(cast_inputs=torch.float16)
    def forward(ctx, input, qweight, scales, qzeros, g_idx, bits, maxq):
        output = matmul248(input, qweight, scales, qzeros, g_idx, bits, maxq)
        ctx.save_for_backward(qweight, scales, qzeros, g_idx)
        ctx.bits, ctx.maxq = bits, maxq
        return output

    @staticmethod
    @custom_bwd