In recent years, there has been a substantial effort directed toward applying large-eddy simulation techniques to scramjet combustion processes. The choice of an appropriate turbulent combustion model to use is complicated by the nature of the problem, which can involve compressibility effects, shock-flame interactions, high turbulence levels, and short residence times. At high Mach numbers, finite-rate chemistry effects are predominant, and most successful combustion models include multi-step reaction chemistry as part of the formulation. As non-premixed combustion is necessarily preceded by macro-and micromixing, it is generally accepted that some type of subgrid-scale model for micro-mixing and the effects of unresolved fluctuations on the filtered production rates is necessary. The exact form of this model and the range of influence that it must have as a function of the chosen mesh resolution is a matter of debate, and as yet, no clear consensus has emerged. In the present work, the influences of several finite rate chemistry type subgrid turbulencechemistry interaction models ('laminar chemistry', three Partially Strirred Reactor (PaSR) type models, and one Scale-Similarity (SS) model) on the predictive capability of NCSU's hybrid large-eddy / Reynolds-averaged Navier-Stokes (LES/RANS) solver (REACTMB) are studied, quantified, and presented. As test cases, reactive flows through two different scramjet configurations are simulated: the model scramjet investigated at the Institute for Chemical Propulsion of the German Aerospace Center (DLR) and the Configuration A tested at the University of Virginia's (UVa) Scramjet Combustion Facility. The overall influence of the subgrid closures for filtered species production rates on the reactive flow in the DLR combustor is not substantial. All subgrid closures predict results which closely resemble those predicted by the 'laminar chemistry' (where the turbulence-chemistry interactions are ignored) assumption. For the UVa combustor, the subgrid closures have a more pronounced effect on the predictions, most evident in temperature distributions. Error analysis and performance quantification of various models reveal that, relative to a baseline model, 'laminar chemistry' when used with the Jachimowski reaction mechanism and scale-similarity model when used with the Burke et al. reaction mechanism provide the most improvement, but the degree of improvement is quite modest.