Replace control-token KV hiding with token-exchange (#8)#34

Open

AlonMalach wants to merge 16 commits into

mainfrom

feature/reduce-control-token-kv-overhead

Commits on May 15, 2026

Replace control-token KV hiding with token-exchange by default (#8)
AlonMalach
committed
Add token-exchange parity eval harness (#8)
AlonMalach
committed
Use <|start_of_role|> instead of BOS for LoRA/builtin substitute (#8)
AlonMalach
committed
Template: drop the first role marker after each adapter control token (#8)
AlonMalach
committed
Template: drop first char of in-message ALoRA invocation (#8)
AlonMalach
committed
Derive LoRA substitute from the tokenizer's chat template (#8)
AlonMalach
committed
Move token-exchange rewrite into the switch (#8)
AlonMalach
committed
Remove the legacy KV-hiding code path (#8)
AlonMalach
committed
Fix base-weight validator rejecting control_to_substitute_lut buffer (#8)
AlonMalach
committed
Remove dead hiding-constant report from compose output (#8)
AlonMalach
committed
Update tests for removed legacy hiding fields (#8) — partial
AlonMalach
committed
Strip remaining legacy hiding-field references from tests/docs (#8)
AlonMalach
committed
Drop tensor.any() gate from switch token-exchange rewrite (#8)
AlonMalach
committed
Fix vLLM test runners after switch tuple-return + dead-class purge (#8)
AlonMalach
committed

Commits on May 16, 2026

Document the LoRA substitute probe's data-independence assumption (#8)
AlonMalach
committed

Commits on May 18, 2026

Remove residual position-correction references (#8)
AlonMalach
committed