How to configure for run-to-run replicability?

[psadil]

The documentation for the --skull-strip-fixed-seed flag indicates that setting the flag and specifying --omp-nthreads 1` ensures run-to-run replicability. But I'm not able to get that to work (using v0.14.2) with the following call

singularity run -e \
-B $BIND_DIR:$BIND_DIR \
 docker://pennbbl/qsiprep:0.14.2 \
 $BIDS_DIR ${OUT_DIR}-1 \
 participant \
 --output-resolution 1.7 \
 --denoise-method patch2self \
 --unringing-method mrdegibbs \
 --hmc_model eddy --eddy_config eddy_params.json \
 --fs-license-file license.txt -w ${WORK_DIR}-1 \
 --participant_label $SUBID -v \
 --skip_bids_validation --omp-nthreads 1 --skull-strip-fixed-seed &

singularity run -e \
-B $BIND_DIR:$BIND_DIR \
 docker://pennbbl/qsiprep:0.14.2 \
 $BIDS_DIR ${OUT_DIR}-2 \
 participant \
 --output-resolution 1.7 \
 --denoise-method patch2self \
 --unringing-method mrdegibbs \
 --hmc_model eddy --eddy_config eddy_params.json \
 --fs-license-file license.txt -w ${WORK_DIR}-2 \
 --participant_label $SUBID -v \
 --skip_bids_validation --omp-nthreads 1 --skull-strip-fixed-seed &

wait

The contents of eddy_params.json are

$ cat eddy_params.json 
{
  "flm": "linear",
  "slm": "linear",
  "fep": false,
  "interp": "spline",
  "nvoxhp": 1000,
  "fudge_factor": 10,
  "dont_sep_offs_move": false,
  "dont_peas": false,
  "niter": 5,
  "method": "jac",
  "repol": true,
  "num_threads": 1,
  "is_shelled": true,
  "use_cuda": false,
  "cnr_maps": true,
  "residuals": false,
  "output_type": "NIFTI_GZ",
  "args": ""
}

(tested with both "use_cuda": true and "use_cuda": false)

I haven't noticed substantial differences in the output of two runs, but there are differences. For example, a diff on the dwiqc.jsons gives

44c44
<       "raw_neighbor_corr": 0.703825,
---
>       "raw_neighbor_corr": 0.703779,
47,48c47,48
<       "raw_coherence_index": 896.185,
<       "raw_incoherence_index": 8035.92,
---
>       "raw_coherence_index": 895.598,
>       "raw_incoherence_index": 8027.47,
56c56
<       "t1_neighbor_corr": 0.904242,
---
>       "t1_neighbor_corr": 0.915362,
59,67c59,67
<       "t1_coherence_index": 1580.51,
<       "t1_incoherence_index": 6836.46,
<       "mean_fd": 0.4759555044,
<       "max_fd": 1.2247645825,
<       "max_rotation": 0.0027262706,
<       "max_translation": 1.027283256,
<       "max_rel_rotation": 0.0020498587,
<       "max_rel_translation": 0.8929094262,
<       "t1_dice_distance": 0.5678131472,
---
>       "t1_coherence_index": 1500.87,
>       "t1_incoherence_index": 6798.68,
>       "mean_fd": 0.4717088062,
>       "max_fd": 1.2484832519,
>       "max_rotation": 0.0028351869,
>       "max_translation": 1.032040753,
>       "max_rel_rotation": 0.0020856578,
>       "max_rel_translation": 0.8992222859,
>       "t1_dice_distance": 0.5686687337,

I see that 0.16.0RC3 includes a patch for ensuring omp-threads is respected in all pipelines, so perhaps the lack of replicability is caused by multithreading? Or, are there other parameters that I should be setting?

Thanks,