Commit Graph
17 Commits
Author SHA1 Message Date
Evandro Fonseca Junior 9b1e973e4e feat: detect template family from aliases 2026-03-26 18:28:37 -03:00
Evandro Fonseca Junior 9bf79ff68c chore: expand alias template benchmark coverage 2026-03-26 18:26:40 -03:00
Evandro Fonseca Junior 28d3fa7b51 feat: harden llama chat template rendering 2026-03-26 18:23:32 -03:00
Evandro Fonseca Junior f96b42c356 chore: expand template benchmark coverage 2026-03-26 18:20:35 -03:00
Evandro Fonseca Junior bad2bb33a5 feat: add family-based chat templates 2026-03-26 18:16:54 -03:00
Evandro Fonseca Junior b3001f6234 chore: initialize template autoresearch segment 2026-03-26 18:14:03 -03:00
Evandro Fonseca Junior 43e2fb01f2 perf: disable llama perf counters in hot path 2026-03-26 17:56:38 -03:00
Evandro Fonseca Junior 6eaf59dd74 Revert "perf: stop generation on default turn boundaries"
This reverts commit caf4654c0a2a856c8f71dd465d4a4b1d67cde5bc.
2026-03-26 17:54:48 -03:00
Evandro Fonseca Junior 653f5b5bc1 perf: stop generation on default turn boundaries 2026-03-26 17:52:59 -03:00
Evandro Fonseca Junior 1bd3365a31 Revert "perf: switch to qwen-style chat template"
This reverts commit db92212697efac74c50708139d7a5e5ab7948486.
2026-03-26 17:50:57 -03:00
Evandro Fonseca Junior 305c74e264 perf: switch to qwen-style chat template 2026-03-26 17:49:17 -03:00
Evandro Fonseca Junior f84570afe4 Revert "perf: slim turbo4b runtime profile"
This reverts commit 994e6071b61fecdbed34fde1ff871381eb48c840.
2026-03-26 17:41:08 -03:00
Evandro Fonseca Junior 0e6f5ca5a3 perf: slim turbo4b runtime profile 2026-03-26 17:39:37 -03:00
Evandro Fonseca Junior cac9227a6a Revert "perf: avoid heap allocs during token detokenization"
This reverts commit 5269f43e522ed03bfd888723c2ba492c51cf41b1.
2026-03-26 17:38:51 -03:00
Evandro Fonseca Junior 14ab598ae1 perf: avoid heap allocs during token detokenization 2026-03-26 17:37:01 -03:00
Evandro Fonseca Junior b38c77f875 chore: fix autoresearch benchmark model id 2026-03-26 17:31:35 -03:00
Evandro Fonseca Junior f83e8b83f6 chore: initialize autoresearch baseline 2026-03-26 17:28:25 -03:00