SN
Sam Nguyen
5 hours ago
lm studio remapped qwen2.5-14b to an old gguf and i demoted the wrong one
opened LM Studio this morning and the chat still said qwen2.5-14b-instruct. Activity Monitor said 9.1GB resident, which is fine for that tag on the Mac Studio.
then the answers started sounding like the 7B I deleted in July. checked the Files tab — same alias, different file path, old Q4_K_M from a leftover Downloads folder. i had "demoted" the good weights yesterday thinking i was clearing the stale ones.
anyone else treat the alias as truth, or do you open the absolute path every time before a client demo?
5 comments
Join the discussion
Log in to comment.
sorry, english not perfect — this happens a lot with aliases.
i always record the sha256 of the GGUF next to the client config. if Activity Monitor RAM looks "right" but answers feel thin, the weights moved under you.
also check
num_ctx. old 7B file with context still set for 14B will thrash and look slow, not just dumb. i got burned once in a live demo.sha256 next to the config is the move. i got burned demoting "stale" weights too — deleted the good Q4 and kept the Downloads leftover.
wifi was fine. coffee was cold. model was just the wrong file wearing the right name.
Alias-as-truth is how eval baselines silently rot. We pin absolute paths in a small manifest and refuse to start the FastAPI shim if the file hash drifts.
Curious — do you keep a one-line golden prompt you run before demos, or only notice when the client call goes weird?
one-line golden prompt, every time. mine is just "count the r's in strawberry and print the loaded model id".
if the id doesn't match the tag, i abort before the client demo. hash drift checks in the FastAPI shim are nicer but the sticky note prompt still catches half my mistakes.
yeah lm studio aliases lie. i keep a sticky note of the absolute path + the working num_ctx for each tag.
last week qwen2.5-14b was pointing at a leftover Q3 from april and Activity Monitor still said ~9GB because Metal was caching the old map. reboot fixed the RAM number. the answers stayed dumb until i re-linked the file.
do you pin hashes in a sidecar json or just open Files every time?