Bf16 might be freezing your weight updates
2 comments
- Points8
- Comments2
2 comments
Story details from Hacker News ↗ · linked article on lacuna.tiptreesystems.com.
Bf16 might be freezing your weight updates was first seen on Hacker News · Front page on Nov 5, 2026. Since then it has appeared in 3 checks on 1 different day. Its best position was #29 of 30, reached on Nov 5, 01:07 UTC. It dropped off the list after Nov 5, 01:07 UTC.
Position on Hacker News · Front page at every check (lower is better).
Best rank per day (UTC).
Platform metric as reported by Hacker News.
A few top-level comments from Hacker News, shortened. Read the full thread on Hacker News.
There's a reason we keep around an fp32 copy of the weights? This issue was recognized a decade ago when we first started training with fp16.
If you fine-tune a model with weights stored in bf16 and let torch.optim.AdamW write to them directly, you get interesting behaviour. bf16 can only represent changes to weights if the change is at least ~1/512 of the weight itself, so at small learning rates only the smallest weights can update. As a result, you have…
Archived daily, weekly and monthly lists that include this entry.
Not at the moment. It was last on the list on Nov 5, 01:07 UTC, after 1 day of trending.
#29 on Hacker News · Front page, on Nov 5, 01:07 UTC. Trendoday checks this list every 15 minutes.
It was first seen on Nov 5, 2026 and has been on the list on 1 day (3 checks) since.
8 points and 2 comments; submitted by martincsweiss on Nov 2, 2026.