0.2.15
Patch Changes
-
ae7875aThanks @thrgreenwald! - The inference engine's Anthropic-compatible messages and token-counting endpoints now accept anyanthropic-versionheader value, or no header, instead of rejecting clients that do not send2023-06-01. -
ae7875aThanks @anerli! - Fix model loading on Metal exceeding its reserved weight-import memory when streaming weight tables (#231). Import batches now share the memory plan used to reserve their budget, accounting for mapped file spans and simultaneous staged uploads.