The variance-covariance matrix V is not positive definitive

I combined the binary files (.bed, .bim, .fam) of the two cohorts together with plink merge before using GCTA::GREML.

plink --bfile AA --bmerge EA --make-bed --out merged
plink --bfile EA --exclude merged.missnp --make-bed --out EA_tmp
plink --bfile AA --exclude merged.missnp --make-bed --out AA_tmp
plink --bfile EA_tmp --bmerge AA_tmp --make-bed --out merged

Next I continued with the commands from the documentation (including Step2 option#1): cnsgenomics.com/software/gcta/#GREMLinWGSorimputeddata until the "REML analysis with multiple GRMs" when I get "Error: the variance-covariance matrix V is not positive definite"

gcta64 --reml --mgrm multi_GRMs.txt --pheno phen.txt --out test
It's my understanding that if the variance-covariance V is negative definite this could be cause by a low sample number; however, this shouldn't be the case for positive definite. Furthermore, "for unrelated individuals and common SNPs, you will need at least 3160 unrelated samples to get a SE down to 0.1 (see Visscher et al. 2014 PLoS Genet). Once I combine Caucasian and Africans groups (2061 + 1131 = 3192) I should have sufficient power?

The advice here (gcta.freeforums.net/thread/191/gcta-reml-error) suggests that there may be something wrong with the GRMs and to check for any missing values in the GRM(s) with the following R function:

ReadGRMBin=function(prefix, AllN=F, size=4){
sum_i=function(i){
return((1+i)*i/2)
}

 BinFileName=paste(prefix,".grm.bin",sep="")
 NFileName=paste(prefix,".grm.N.bin",sep="")
 IDFileName=paste(prefix,".grm.id",sep="")
 id = read.table(IDFileName)
 n=dim(id)[1]
 BinFile=file(BinFileName, "rb");
 grm=readBin(BinFile, n=n*(n+1)/2, what=numeric(0), size=size)
 NFile=file(NFileName, "rb");
 if(AllN==T){
   N=readBin(NFile, n=n*(n+1)/2, what=numeric(0), size=size)
 }
 else N=readBin(NFile, n=1, what=numeric(0), size=size)
 i=sapply(1:n, sum_i)
 return(list(diag=grm[i], off=grm[-i], id=id, N=N))
}

So in GCTA::GREML Step #3 we created multiple GRMs (e.g.):

gcta64 --bfile merged --extract snp_group1.txt --make-grm --out merged_group1
So I check for missing values in all the GRMs like this (e.g. for GRM 1):

I assume I'm checking these GRMs correctly and since I get FALSE for each of the four GRMs these should be okay, right?

I tried two other suggestions as well. The --reml-no-constrain option (gcta.freeforums.net/thread/190/variance-covaraince-matrix-error), which gives the same error:

gcta64 --reml-no-constrain --mgrm multi_GRMs.txt --pheno merged.phen --out merged
and the --reml-bendV option (gcta.freeforums.net/thread/243/variance-covaraince-matrix-positive-definite) which seems to run (doesn't given any errors) but only produces the .log file and not the .hsq (results)

I'm currently trying to get around the problem using the Fisher scoring approach at this step:

gcta64 --bfile merged --extract snp_group1.txt --make-grm-alg 1 --out alg1_group1
Any feedback is greatly appreciated!