It is not clear that the other respondents at the Pi forum understand what you are asking about. Using awb_gains will allow every photo to exaggerate the values in the red (or blue) channel. I assume that the camera continues to use automatic white balance (AWB) and then also modify the values based on the awb_gains parameters. This is potentially a very good feature for using the Pi NoIR with a filter to capture photos that can each be converted to a facsimile of NDVI. However, if the AWB algorithm continues to adjust for every photo, the ratio of blue to red could vary among photos in unpredictable ways. For example, AWB will modify the ratio of blue to red if the background color (behind the plants of interest) differs between two photos. AWB will modify the ratio of blue to red if the sun goes behind a cloud. AWB could modify the ratio of blue to red if the camera angle changes relative to the sun or light source. So two photos of the same plants could have different red and blue values and therefore different NDVI values. This is not ideal, but awb_gains will probably allow better results than AWB alone or one of the white balance presets (sunny, tungsten, etc.). It would be very interesting to see a comparison of results.
The problem of variable lighting color affecting the balance of red to blue also applies when a custom white balance is done with a Canon PowerShot. In that case, the camera always applies the same adjustment to each photo, but if the external lighting color changes, the ratio of red to blue in the scene and also in the image of plants will change. (If only the background color changes, the red:blue ratio in plants will not change.) So even a custom white balance does not offer sufficient control to get consistent NDVI results unless the photo scene is carefully controlled.
This is why a calibration procedure is required to get reliable results. This requires that targets of known reflectivity in the visible and NIR be placed in the photos and then that a procedure be done to adjust the red and blue values in all pixels based on the red and blue values in the pixels of the targets. The calibration procedure does not solve the other problem with single camera NDVI systems that at least one of the channels will capture a mix of visible and NIR light. The proportion of NIR light contaminating the visible channel (or vice versa) can be estimated and accounted for (as Ned Horning's calibration procedure allows), but that assumes that the proportion is the same in every pixel. I assume it will vary a lot from pixel to pixel, so it is not clear how effective the VIS/NIR-mixture fudge factor will be.
So the only way to solve both problems is to use two cameras, one for visible light and one for NIR, and put calibration targets in photos from both cameras. The potential value of doing this with Pi cameras is that the Pi could trigger both cameras, align the two photos, and produce an NDVI image on the fly. I don't know whether the Pi is powerful enough to do this, and the calibration part might require some human intervention.