Pull requests / #813
#813 rtx6000fix
closed · @Fringe210 · 0 comments · View on GitHub
NVIDIA / CUDADocumentationWindows
Description
`--ple-io ram` refused to start on Windows (`no mlock; use --ple-io mmap`). This wires the existing `strata::platform::lock_resident` helper (working-set minimum + `VirtualLock`, chunked per GiB) into `PleTable::open`, so the whole n-gram table is locked in RAM at startup on Windows too. A lock failure still falls back to touching the pages (plain mmap behaviour) with a warning, same as the POSIX path. Changes - `src/kernels/ngram.cpp`: Windows branch of the Mmap+lock path calls `lock_resident`; sets `locked()` on success. - `src/program/generate.cpp`: drop the Windows `--ple-io ram` refusal; usage text now says VirtualLock on Windows, mlock elsewhere. - `include/strata/kernels/ngram.hpp`: doc comments updated (28.8 GB IQ4_NL / 51.2 GB FP8, Windows semantics of `locked()`). Measured on Windows, RTX PRO 6000 Blackwell 96 GB, 128 GB RAM, MSVC 19.44 - `platform_memory_test`: `locked 256 MiB via working-set minimum + VirtualLock`, OK. - Full server start, 320,001,536-row IQ4_NL table (28.8 GB): `PLE table locked in RAM (--ple-io ram) in 95.9 s`, no fallback warning.
Related on strata.com
Editorial links to help you install, pick models, or read release notes — not part of the upstream thread.