feat(vamb/bin): add GPU support - #12876
Merged
Merged
Conversation
dialvarezs
force-pushed
the
vamb-gpu
branch
2 times, most recently
from
September 3, 2026 12:48
a354c96 to
4a9ee7e
Compare
Select the container and conda environment from `task.accelerator`, carry the `process_gpu` label, and pass `--cuda` when an accelerator is declared. The CUDA runtime version also goes to the versions topic. Vamb pins `torch ==2.6.0`, which conda-forge only builds against CUDA 12.6 on linux-64, so that is what the GPU environment gets. It carries no sm_120 kernels, but the driver JITs the `compute_90` PTX, so Blackwell cards still run. Two fixes the GPU path forced. The wave images are Amazon Linux based and ship neither `find` nor `gzip`, both of which the module used to rename and compress its bins. So `gzip` joins the GPU environment and the `find` call becomes a null glob, which tolerates zero bins the same way the old one did. That matters because `bins` is an optional output. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The existing checks were closures without `assert`, so `{ path(...).exists() }`
and the `Completed Vamb` line were evaluated and discarded. Adding `assert`
showed the log check never held: the lines are timestamped, so `contains` on
the bare string is false.
The tests now use the metagenome binning dataset, which ships per-sample
assemblies, BAMs and an abundance table, so BOWTIE2_BUILD, BOWTIE2_ALIGN,
COVERM_CONTIG and MMSEQS_CONTIG_TAXONOMY all go. The taxonomy test keeps one
GAWK, deriving the lineage from the contig labels the dataset documents. Four
tests that took minutes now take about fifteen seconds each.
Every file Vamb writes turns out to differ between machines, the `.npz`
archives included, so the snapshot keeps the versions and the bins are
checked as a partition instead: every contig Vamb clustered appears in
exactly one bin, read with the nft-fasta plugin. Note this means the
`abundance` snapshot the tests carried before only held because it was
recorded on a runner.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
dialvarezs
marked this pull request as ready for review
September 3, 2026 13:18
Joon-Klaps
approved these changes
Sep 4, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Same treatment SemiBin2 got in #12863, now for Vamb.
vamb/binpicks its container and conda environment fromtask.accelerator, carries theprocess_gpulabel, and passes--cudawhen an accelerator is declared. It also emits the CUDA runtime version to the versions topic.Vamb pins
torch ==2.6.0, which conda-forge only builds against CUDA 12.6, so that is what the GPU environment gets. That build has no sm_120 kernels, but the driver JITs thecompute_90PTX, so Blackwell cards still run.The wave images are Amazon Linux based and ship neither
findnorgzip, both of which the module used on its bins. Sogzipjoins the GPU environment and thefindcall becomes a null glob, which tolerates zero bins the same way.Tests
The existing checks were closures without
assert, so{ path(...).exists() }and theCompleted Vambline were evaluated and discarded. Addingassertshowed the log check never held, since the lines are timestamped.The tests now use the metagenome binning dataset, which ships per-sample assemblies, BAMs and an abundance table, so
BOWTIE2_BUILD,BOWTIE2_ALIGN,COVERM_CONTIGandMMSEQS_CONTIG_TAXONOMYall go. The taxonomy test keeps oneGAWK, deriving the lineage from the contig labels the dataset documents.What is asserted was measured over repeated runs. The bins, the cluster tables and the log move between runs, so the bins are checked as a partition instead: every contig Vamb clustered appears in exactly one bin.
abundance,compositionandlatent_encodingheld steady, so they are snapshotted alongside the versions.PR checklist
topic: versionslabelnf-core modules test vamb/bin --profile singularity🤖 Generated with Claude Code