Skip to content- Sources: primary, discussion
- Summary: PR 59958, merged 2026-10-30, adds an engram-config disk_offload option for Qwen4Exp so the PLE n-gram table stays in the safetensors files. On a GB10, Qwen3.8-Flash-Next NVFP4 loaded 27.0 GiB less weight memory with GSM8K unchanged, while a cold 8K prefill from NFS was 2 to 4.5x slower than from local NVMe. A reviewer notes the saving is reclaimable page cache rather than freed memory.
- Why it matters: Unified-memory machines can serve PLE models that did not fit before, provided the checkpoint sits on local NVMe.