The large-scale redshift-space distortion (RSD) in galaxy clustering can probe $f\sigma_8$, a combination of the cosmic structure growth rate and the matter fluctuation amplitude, which can constrain dark energy models and test theories of gravity. While the RSD on small scales (e.g. a few to tens of $h^{-1}{\rm Mpc}$) can further tighten the $f\sigma_8$ constraints, galaxy assembly bias, if not correctly modelled, may introduce systematic uncertainties. Using a mock galaxy catalogue with built-in assembly bias, we perform a preliminary study on how assembly bias may affect the $f\sigma_8$ inference. We find good agreement on scales down to 8--9$h^{-1}{\rm Mpc}$ between a $f\sigma_8$ metric from the redshift-space two-point correlation function with the central-only mock catalogue and that with the shuffled catalogue free of assembly bias, implying that $f\sigma_8$ information can be extracted on such scales even with assembly bias. We then apply the halo occupation distribution (HOD) and three subhalo clustering and abundance matching (SCAM) models to model the redshift-space clustering with the mock. Only the SCAM model based on $V_{\rm peak}$ (used to create the mock) can reproduce the $f\sigma_8$ metric, and the other three could not. However, the $f\sigma_8$ metrics determined from central galaxies from all the models are able to match the expected one down to 8$h^{-1}{\rm Mpc}$. Our results suggest that halo models with no or incorrect assembly bias prescription could still be used to model the RSD down to scales of $\sim 8 h^{-1}{\rm Mpc}$ to tighten the $f\sigma_8$ constraint, with a sample of central galaxies or with a flexible satellite occupation prescription.