0.3.7 (2024-07-29)

Bug Fixes

• cpp: correct validateModelChatTemplate fn (f9ed84a)

Features

• example: add Chat Template in initial message (124cb6a)

0.3.6 (2024-07-28)

Features

• example: update json schema example (2f70192)
• support chat format (#72) (030ebaf)

0.3.5 (2024-07-27)

Bug Fixes

• android: revert Debug comment (7972f83)
• cpp: incomplete utf-8 character (3799cf8)
• not init sampling before get embedding (#69) (1088300)

Features

• sync llama.cpp (#71) (a579ce3)

0.3.3 (2024-07-22)

Bug Fixes

• add build for armv82 and v84 and sync llama.cpp (#67) (1942159)
• android: loadSession not taking paths with file:// (#66) (39d1e05)

0.3.2 (2024-05-29)

Features

• sync llama.cpp (#58) (dc6a47a)

0.3.0 (2024-04-26)

Features

• add model details in context (#52) (f2a4c53)
• android, example: disable repeat penalties by default (cf685ad)
• example: add missing commands help (4a53e68)
• migrate to new session API (725117c)

0.3.0-rc.18 (2024-04-25)

Features

• sync llama.cpp (#51) (0b8f9bb)
• ts: sync grammar converter (#48) (7834a73)

0.3.0-rc.9 (2023-12-20)

Bug Fixes

• android: use THREAD_POOL_EXECUTOR (c294171)

Features

• ios: use created dispatch_queue_t instead of main (b173d18)
• jest: add bench mock (9df745c)

0.3.0-rc.8 (2023-12-19)

Features

• add bench method (#35) (9d0c7df)
• sync llama.cpp (a57171e)

0.3.0-rc.0 (2023-10-11)

Bug Fixes

• android: default rope_freq_base / rope_freq_scale to 0 during recent breaking change (43e036e)
• cpp: incorrect num_tokens_predicted (6dd9865)

Features

• cpp: reuse llama_sample_token common util (#25) (95b4fb2)
• implement loadSession & saveSession methods (#23) (68fe09c)
• sync llama.cpp (81f9cb2)
• sync llama.cpp (#22) (8da7244)

0.2.0 (2023-09-18)

Bug Fixes

• example: incorrect proguard rule (ae7f262)

Features

• android: check if need to use the default ndkVersion (41b67d4)
• sync llama.cpp (#21) (50235c2)
• sync llama.cpp & remove ggml-metal patch (ad965e4)

0.2.0-rc.4 (2023-09-07)

Bug Fixes

• android: ignore android/build in npm package (178fab2)
• android: var name of context result (120e67e)
• ios: dispatch_release queue on refresh (614e909)
• ts: memory_f16 type (9af95bc)
• wait task finished on stop/release method (f61e18c)

Features

• add native source into git (6d01be7)
• android: avoid warning by use DeviceEventEmitter (b9487dd)

0.2.0-rc.3 (2023-09-04)

Bug Fixes

• android: arch check on init context (95519e4)
• example: typecheck (63a08cb)

Features

• android: build for x86_64 platform (#15) (d10bf0c)
• example: improve Bubble render (22a85bd)
• make partial completion callback optional (#17) (a43b5ba)
• sync llama.cpp (#16) (989858b)

0.2.0-rc.0 (2023-08-29)

Bug Fixes

• cpp: incorrect prob convert if the str is already a known token (d296d09)
• ios: do manual release due to disabled ARC in pod (6623e96)
• ios: incorrect sent_token_probs_index (ac7c554)
• send rest of content on stop (8a4b863)

Features

• add detokenize method (#12) (13b0a15)
• example: create dir for model cache (7cde772)
• initial android support (#10) (36f608e)

Breaking changes

• sync llama.cpp
  • A new file format has been introduced: GGUF
  • Tokenizer fixes and API change
  • GGUFv2 adds support for 64-bit sizes + backwards compatible
  • Added support for Falcon models

0.1.5 (2023-08-17)

Bug Fixes

• podspec: incorrect METAL_NDEBUG flag (e162586)

Features

• sync llama.cpp (1832694)
• sync llama.cpp changes (#7) (81257bf)

0.1.4 (2023-08-14)

Features

• add convertJsonSchemaToGrammar util (#6) (ce9201d)
• add tokenize & embedding function (fcf5a6d)

0.1.3 (2023-08-11)

Bug Fixes

• add prefix to avoid redef with other libraries using ggml (0b01bbb)
• prompt_n in timings (db9da00)
• use LM_GGML_USE_METAL to check ggml-metal not enabled (67f12ce)

Features

• add grammar support (#1) (f42b4fb)
• add metal patches (f736614)