Rna_fq2bam GPU-STAR: illegal memory access at ~1,000,000 collapsed junctions, independent of --max-collapsed-junctions

,

On junction-rich RNA-seq samples, pbrun rna_fq2bam aborts during GPU-STAR alignment with:

[src/stitchWindowAligns_gpu.cu:6488] cudaSafeCall() failed: an illegal memory access was encountered, exiting.

(process exits rc=255, no BAM produced).

We traced this to the number of collapsed splice junctions reaching ~1,000,000. Critically, raising --max-collapsed-junctions does NOT help — it crashes at the identical read offset whether the limit is the default 1,000,000 or 2,000,000. The flag only governs STAR’s per-chunk, host-side graceful check (ReadAlignChunk.cpp:570); there appears to be a separate, fixed-size GPU-side buffer in stitchWindowAligns_gpu.cu that overflows at ~1M collapsed junctions regardless of the flag.

Parabricks image nvcr.io/nvidia/clara/clara-parabricks:4.7.0-1
Bundled STAR 2.7.2a (genome index built with STAR 2.7.2a, --sjdbOverhang 250)
GPU NVIDIA H100 PCIe (80 GB)
NVIDIA driver 580.159.03
Host AMD EPYC 9124 (Genoa), 188 GB RAM
OS / kernel Ubuntu 24.04.4 LTS, kernel 6.8.0-124-generic
Container runtime Docker 29.6.0