0
0
Fork 0
mirror of https://github.com/matrix-construct/construct synced 2024-07-01 16:28:19 +02:00
Commit graph

41 commits

Author SHA1 Message Date
Jason Volk 4a8302038a ircd::gpt::vocab: Add tokenization and detokenization count() convenience. 2022-09-21 16:42:07 -07:00
Jason Volk 1f947f982c ircd::gpt: Cache pipe::code instance in shared_ptr for task reuse. 2022-07-31 02:46:05 -07:00
Jason Volk 33094ece07 ircd::gpt: Fix doc comment on count vs. tokens. [ci skip] 2022-07-30 19:25:32 -07:00
Jason Volk a44e180ff7 ircd::gpt: Fix stale comment. [ci skip]. 2022-07-22 19:11:51 -07:00
Jason Volk 6d2da3b4f1 ircd::gpt::task: Refactor generator interface to member functions. 2022-07-01 20:17:56 -07:00
Jason Volk 56d944f33e ircd::gpt::vocab: Add space-prefix convenience argument. 2022-07-01 17:36:45 -07:00
Jason Volk d68f261036 ircd::gpt::pipe: Simplify init/fini w/ run::level::QUIT handler. 2022-06-30 16:04:31 -07:00
Jason Volk 78848925ee ircd::gpt: Various refactoring. 2022-06-19 20:14:22 -07:00
Jason Volk b7b1328352 ircd::gpt::pipe: Reuse logsm buffer for logexp intermediate values. 2022-06-17 21:11:53 -07:00
Jason Volk 26ce330d6d ircd::gpt: Add layer count to model section of opts. 2022-06-17 21:11:53 -07:00
Jason Volk d844ec895c ircd::gpt: Simplify aperature unions. 2022-06-17 21:11:53 -07:00
Jason Volk a7560b5cfc ircd::gpt::pipe::code: Preliminary reorg for legacy NIR build cache. 2022-06-17 21:11:53 -07:00
Jason Volk aea6c79fc2 ircd::gpt: Add top N and target label result register control block. 2021-10-06 13:13:47 -07:00
Jason Volk 8bd78af128 ircd::gpt: Additional task header/interface simplification. 2021-10-06 13:13:47 -07:00
Jason Volk b72a4590b1 ircd::gpt: Task header reorganization/resimplification. 2021-10-06 13:13:47 -07:00
Jason Volk c1f3e580c3 ircd::gpt: Add top_p lmhead selector, quantized for now. 2021-10-06 13:13:47 -07:00
Jason Volk 8a3eeb46f9 ircd::gpt::pipe: Optimize pipeline to cache attention state for generations. 2021-10-06 13:13:47 -07:00
Jason Volk 79b6d4b9da ircd::gpt::pipe: Unify descriptor allocations. 2021-10-06 13:13:47 -07:00
Jason Volk ce9abfb321 ircd::gpt::model: Optimize left-attention mask. 2021-10-06 13:13:47 -07:00
Jason Volk 20162fd7d5 ircd::gpt: Splits and renames; various reorg. 2021-09-15 01:44:36 -07:00
Jason Volk 6e7e56f081 ircd::gpt::pipe: Add conf item for default source paths and opts for now. 2021-09-14 23:39:55 -07:00
Jason Volk 1eec9bfd64 ircd::gpt: Remove unnecessary packed attribute. 2021-08-01 03:34:27 -07:00
Jason Volk c3cb249f78 ircd::gpt: Split debug related; improve flush options; minor cleanup. 2021-05-14 05:57:47 -07:00
Jason Volk 37b1d47c8d ircd::gpt: Reorganize task options and control blocks. 2021-05-11 20:56:09 -07:00
Jason Volk f61239a52c ircd::gpt::pipe: Various statistical instrumentation. 2021-04-22 12:27:57 -07:00
Jason Volk 2a3c54afa2 ircd::gpt::pipe: Add model invalidation for device resynchronization. 2021-04-22 12:27:57 -07:00
Jason Volk bd460a80ce ircd::gpt::pipe::model: Additional sub-buffering for lmhead. 2021-04-22 12:27:57 -07:00
Jason Volk d5eb1e3a87 ircd::gpt: Add backpropagation pipe. 2021-04-22 12:27:57 -07:00
Jason Volk 14a1561cad ircd::gpt: Add adaptive moment state to pipe model; move task prelim defs. 2021-04-22 12:27:57 -07:00
Jason Volk eba5d7f586 ircd::gpt: Improve build related; debug/assertions. 2021-04-22 12:27:57 -07:00
Jason Volk 9c062d9c3f ircd::gpt::pipe: Correctness; compute loss, statistics; pipeline optimize. 2021-04-22 12:27:57 -07:00
Jason Volk 1870c364f4 ircd::gpt: Reorganize interface. 2021-04-22 12:27:57 -07:00
Jason Volk 29e74ec9e1 ircd::gpt: Transformer Pipe. 2021-04-22 12:27:56 -07:00
Jason Volk 29fb7910b7 ircd::gpt: Model structural tweaks; task structure; various. 2021-04-22 12:27:56 -07:00
Jason Volk 4da7d2ae43 ircd::gpt: Add task struct; mmap cached model directly; improve init. 2021-03-16 12:58:22 -07:00
Jason Volk 734948863f ircd::gpt::vocab: Add token debug string tool. 2021-03-09 04:50:19 -08:00
Jason Volk 33a1ffd4bf ircd::gpt: Add basic interface; add options, context. 2021-03-09 04:50:19 -08:00
Jason Volk 4458235dfa ircd::gpt: More Matrix Than Matrix. 2021-03-09 04:50:19 -08:00
Jason Volk 53c4260a21 ircd::gpt: Add Basic Latin (lower) and C0 replacement LUT; various. 2021-03-09 04:50:19 -08:00
Jason Volk 29b99dcf4d ircd::gpt: Split vocab related into separate unit. 2021-03-02 11:13:59 -08:00
Jason Volk 57f9d3fdfb ircd::gpt: Add vocabulary tokenization; byte-pair merge encoding for natural language. 2021-02-26 14:52:33 -08:00