Environment

variable	what it does
MHCMATCH_PMHC_DIR	a local mirror of the isalgo/pmhc_data dataset root, used before any download. Stage it once with mhcmatch bootstrap --reference; essential on compute nodes with no outbound network
MHCMATCH_CALIBRATION_CACHE	two caches, one variable. The shared per-allele %rank calibration — measured 15× on a 25-allele sweep (13.3 s → 0.9 s) — and, in a mimicry_refs/ subdirectory, the mimicry reference indexes, which are masked-face projections and so cannot be answered by a text index. Both are derived data keyed by their inputs — the calibration cache also on the scoring source itself, from 1.20.0, so a change to the scoring code invalidates it — and both are safe to delete; reusing the variable means a cluster that already points it at shared storage gets the second for free. Safe to share under concurrency without a lock — entries are written to a tempfile and moved with os.replace. Set it to 0 / off / none to disable both. A proteome_index/ subdirectory used to be the third thing here. Peptide origin search needs no cache now: one text index over the proteome builds in 0.7 s and serves every length. A *published* index never lands here — see the next two rows
HF_HOME	where huggingface_hub puts what it downloads. MHCMATCH_PMHC_DIR is a read override consulted first, not a download destination — when a file is missing the fetch goes through hf_hub_download and ignores it — so on a cluster this is the variable that keeps ~250 MB out of a home quota
MHCMATCH_PMHC	the directory holding pmhc_<tier>.tsv.gz, rather than the dataset root. Overrides MHCMATCH_PMHC_DIR for the panel specifically
MHCMATCH_EXPRESSION / MHCMATCH_STRUCTURES	local overrides for the expression tables and structure templates

Fanning out over more than one machine, point MHCMATCH_PMHC_DIR, MHCMATCH_CALIBRATION_CACHE and HF_HOME at one shared directory every worker can see, and set them in your own executor config — the workflow modules ship local-only and carry no scheduler profile. Sharing the calibration cache under concurrency is safe by construction: an entry is written to a tempfile in the same directory and moved into place with os.replace, which is atomic on POSIX, so there is no lock and nothing to leak when a task is killed. Running a cohort is the whole cohort story — the two arms, the samplesheet contract, and what a fan-out gets wrong first.