grobidOrg / grobid / 32764834259

43%

master: 44%

push

github

perf(docker): tune torch threading and bump delft to a7dc7df

Default the torch intra-op pool to 4 threads. Leaving OMP_NUM_THREADS unset
handed sizing to delft's configure_threads(), which widens the pool to the
whole cpuset -- right for a single document, wrong under concurrent load,
where a 20-core container gives every one of 10 in-flight documents a
20-thread pool and they spend the width fighting each other.

The 2026-08 grid5000 sweeps (cpus x concurrency x threads, 10/20-core
containers, 1000- and 10000-PDF corpora) put the plateau at 4 threads on
every width: ~35% faster than 1 thread at the throughput optimum
(concurrency = cores/2), and the same wall time as 8 threads for ~23%
less CPU. ENV so it binds at "import torch" (when OpenMP reads it) and
stays overridable with -e at "docker run".

Also bump DELFT_REF to a7dc7df, "chore: retrain with embeddings" (model
configs re-enable glove-840B word embeddings), which sits on top of the
"Improve pytorch performances" merge (#230) and the padded-token masking
fixes (#231-#234). configure_threads() is unchanged at this revision, so
the OMP_NUM_THREADS=4 default applies exactly as measured.

10128 of 25690 branches covered (39.42%)

Branch coverage included in aggregate %.

21217 of 46365 relevant lines covered (45.76%)

1.86 hits per line