Could not predict many-chain complexes

Expected Behavior

Use ColabFold to predict two-protein complexes with trimmed sequences (intrachain gaps).

Current Behavior

I am running my fasta file using colabfold_batch, which to my knowledge does not have a distinction between intrachain breaks ('/' in Google Colab) and interchain breaks (':'). If there are too many intrachain breaks (the exact number of chains depends on sequence length), then the ColabFold prediction never completes.

Steps to Reproduce (for bugs)

Cannot share exact sequences, but this error occurs for example with a heterodimer of length 1012 and 6 total chains (one of the proteins trimmed into 5 chains) in the format X:X:X:X:X:Y.

ColabFold Output (for bugs)

2022-07-14 11:50:03,960 Sleeping for 7s. Reason: RUNNING
2022-07-14 11:50:11,458 Sleeping for 6s. Reason: RUNNING
2022-07-14 11:50:17,932 Sleeping for 7s. Reason: RUNNING
2022-07-14 11:50:25,403 Sleeping for 7s. Reason: RUNNING
2022-07-14 11:50:32,888 Sleeping for 9s. Reason: RUNNING
2022-07-14 11:50:44,767 Sleeping for 5s. Reason: PENDING
2022-07-14 11:50:50,260 Sleeping for 7s. Reason: RUNNING
2022-07-14 11:50:57,762 Sleeping for 6s. Reason: RUNNING
2022-07-14 11:51:04,242 Sleeping for 9s. Reason: RUNNING
2022-07-14 11:51:13,722 Sleeping for 5s. Reason: RUNNING
2022-07-14 11:51:19,208 Sleeping for 8s. Reason: RUNNING
2022-07-14 11:51:31,228 Running model_3
-- Stops running here until Timeout (no errors)

Your Environment

Linux CUDA9.2.88 V100

Start at the colabfold_batch entry point and reproduce the reported case using a 1012-residue heterodimer formatted as X:X:X:X:X:Y, with Linux CUDA9.2.88 on a V100. Compare behavior for interchain and intrachain breaks, focusing on the run that stops at model_3. Done means trimmed sequences with many intrachain breaks complete or fail with a clear error instead of timing out.