foldseek createdb --gpu 1 failed in slurm

[DUANZIHAOO]

Hi,
I found that the foldseek function (createdb) is not working properly with --gpu 1 on the slurm.
My foldseek is using the foldseek linux-gpu.tar.gz version. Prostt5-model is prostt5-f16.gguf.
NVIDIA-SMI 535.154.05 Driver Version: 535.154.05 CUDA Version: 12.2 NVIDIA A100-SXM4-40GB
So I run the code (foldseek createdb test.faa testDB --prostt5-model /project/duanzihao/spacedust/weights --threads 1 --gpu 1), however, the result file testDB_ss is not the normal 3Di seq but garbled.

My code here. I can't even be sure if my GPU is functioning properly.
`module load nvidia/cuda/12.2
foldseek createdb test.faa testDB --prostt5-model /project/duanzihao/spacedust/weights --threads 1 --gpu 1
createdb test.faa testDB --prostt5-model /project/duanzihao/spacedust/weights --threads 1 --gpu 1

MMseqs Version: d620467
Use GPU 1
Path to ProstT5 /project/duanzihao/spacedust/weights
Chain name mode 0
Model name mode 0
Createdb extraction mode 0
Interface distance threshold 10
Write mapping file 0
Write Foldcomp 0
Mask b-factor threshold 0
Coord store mode 2
Write lookup file 1
Input format 0
File Inclusion Regex .*
File Exclusion Regex ^$
Threads 1
Verbosity 3

Converting sequences
[
Sort single files in 0h 0m 0s 479ms
Merge all files 0h 0m 0s 107ms
Database type: Aminoacid
CUDA0
CPU
[=================================================================] 2 1s 131ms
Time for merging to testDB_ss: 0h 0m 0s 4ms
Time for merging to testDB_ss_tmp: 0h 0m 0s 3ms
Time for processing: 0h 0m 2s 270ms`

Thanks!

[milot-mirdita]

Could you upload the output database? The output looks normal.

[DUANZIHAOO]

Sure. testDB_ss file is not allowed to upload in github, so I rename testDB_ss to testDB_ss.txt. Use "head -n 10 testDB", the file is displayed as garbled.

spacedust createsetdb testDB testSetDB tmpFolder
testSetDB is also garbled.

spacedust clustersearch testSetDB targetSetDB result.tsv tmpFolder
The k-mers, all AAAAAA, seems unnormal. Because my test.faa is normal.

I guess my step foldseek createdb maybe something wrong.

testDB_ss.txt

[DUANZIHAOO]

head -n 10 testDB_ss, 3Di seq (foldseek createdb cpu-version)
head -n 10 testDB_ss, garbled (foldseek createdb gpu-version)

[DUANZIHAOO]

foldseek createdb test.faa testDB --prostt5-model weights --threads 1
head -n 5 testDB_ss
DDDDFDDDDDDQKDKKDVVDDDKDWDDDPPDDDDD......

foldseek createdb test.faa testDB --prostt5-model weights --threads 1 --gpu 1
head -n 5 testDB_ss
ʉʉʉʉʉʉʉʉʉʉʉʉʉʉʉʉʉʉʉʉʉəʉʉʉ
׮..ʉ.??.¹.¹.¹.ښ......

[milot-mirdita]

Thanks a lot, that was very helpful.

That was a recent MMseqs2 change breaking foldseek's createdb. Should be fixed now in the latest git commit.

[DUANZIHAOO]

Thanks a lot, that was very helpful.

That was a recent MMseqs2 change breaking foldseek's createdb. Should be fixed now in the latest git commit.

In fect, I downloaded the foldseek (foldseek-linux-gpu.tar.gz, 2025.1.19) today, and it worked without garbled errors mentioned above. Surprisingly.