You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Your attached strata-iq3_xxs.log explains the stop: at every start the probe reports --ple-io direct measures slow here (237-656 rows/s, about 1-3 MB/s of random 4 KiB reads), then no progress for 60 s during a request (reading the prompt (batched): layer 1 ...) with 16 threads in disk wait (__iomap_dio_rw), and the watchdog restarts the engine. That also fits 0.1.43 working for you: until 0.1.43 the engine switched such a drive to mmap by itself, and since 0.1.43.1 it only warns and keeps direct, because on a SATA SSD mmap was 7x slower (#2040; docs/DETAILS.md, the --ple-io paragraph). On 0.1.43.1 or later, add --ple-io mmap to the config's args (or "STRATA_PLE_AUTO_MMAP": "1" in its env for the old automatic switch); --ple-io ram would need RAM for the whole table beside the experts. Moving the model to a faster internal drive also helps.
The "--ple-io mmap" argument was rerjected by server.py and "STRATA_PLE_AUTO_MMAP had no effect in the env. I've currently set "STRATA_WATCHDOG_S" = "180" and that allows it to complete.
Activity
lama7 commentedon Oct 13, 2026
I reverted to release 0.1.43 and am running without issue, fwiw.
qiuzan2001 commentedon Oct 15, 2026
Your attached
strata-iq3_xxs.logexplains the stop: at every start the probe reports--ple-io direct measures slow here (237-656 rows/s, about 1-3 MB/s of random 4 KiB reads), thenno progress for 60 s during a request (reading the prompt (batched): layer 1 ...)with 16 threads in disk wait (__iomap_dio_rw), and the watchdog restarts the engine. That also fits 0.1.43 working for you: until 0.1.43 the engine switched such a drive to mmap by itself, and since 0.1.43.1 it only warns and keeps direct, because on a SATA SSD mmap was 7x slower (#2040; docs/DETAILS.md, the--ple-ioparagraph). On 0.1.43.1 or later, add--ple-io mmapto the config'sargs(or"STRATA_PLE_AUTO_MMAP": "1"in itsenvfor the old automatic switch);--ple-io ramwould need RAM for the whole table beside the experts. Moving the model to a faster internal drive also helps.lama7 commentedon Oct 15, 2026
The "--ple-io mmap" argument was rerjected by server.py and "STRATA_PLE_AUTO_MMAP had no effect in the env. I've currently set "STRATA_WATCHDOG_S" = "180" and that allows it to complete.
serve: an engine option written as one entry of the config's args is …
Merge f-145-fix3: #2622 PLE gather, #2267 config args, #2644 split sh…
Merge upstream main (477ac13, 0.1.45): the server, setup, docs and up…
README: list v0.1.45's new startup parameters for a gfx906 run
serve: 設定の args の 1 項目に書いたエンジンのオプションを分け、JSON の誤りは場所を言う(Niko1221#2267、…
serve: an engine option written as one entry of the config's args is …
feat(server): model-card sampling defaults, several API keys, toleran…
serve: a switch reads the run config as the start does