Maximal input size reduction between 1.5.2 and 1.5.5

Expected Behavior

The same input on the same hardware that fit into memory on an older version of ColabFold will still run on newer versions.

Current Behavior

On 1.5.2 a 4644 AA input ran fine but 1.5.5 runs out of memory, even after truncating to 4484 AA.

Steps to Reproduce

1. Install ColabFold 1.5.2 (3564546) via LocalColabFold.
2. Download this input and run as follows.

CUDA_VISIBLE_DEVICES=1 nohup colabfold_batch --templates --amber *.fasta . >&/dev/null &

3. Install ColabFold as above. This results in 1.5.5 9a5c316.
4. Run the same input as above.

I am unable to reproduce step 1.

I have also tried running 1.5.5 with a truncated input to no avail.

ColabFold Output

• log.txt from 3564546
• log.txt from 9a5c316
• log.txt from 9a5c316 with truncated input.

Context

I am running on GPU 1 because 0 is usually occupied. I used nvidia-smi to ensure no other processes were running on the GPU .

Environment

Server specifications

NVIDIA GeForce RTX 4090 video card running driver version 530.30.02 and CUDA 12.1.

Operating system and version

Ubuntu 22.04.3 LTS

Start by comparing the linked log.txt outputs for versions 1.5.2 and 1.5.5, then run the documented colabfold_batch command with the linked FASTA inputs on the stated RTX 4090 environment. Trace the change between commits 3564546 and 9a5c316 that affects memory use. Done means identifying the regression and confirming that the original input runs within the earlier memory limit.