fix(serve): restore n-gram loop-guard DEFAULT OFF (master regressed to 8 → lobotomizes reasoning)#523
Merged
Conversation
…ster regressed to 8 Master regressed HIPFIRE_NGRAM_LOOP_THRESHOLD back to default 8 — the content-blind 4-gram×8 counter false-positives on legitimate enumeration-heavy reasoning at temp>0 (a repeating markdown bullet scaffold `\n - **` trips it INSIDE <think>, force-EOSes mid-think, and EMPTIES the answer channel — the qwen3.6 temp=1.0 registry default hits this). Restores the validated 7eac3c8 behavior: ngram_loop_threshold default 8→0 (OFF); runaway thinking is handled by the think-budget force-close instead. Opt back in with HIPFIRE_NGRAM_LOOP_THRESHOLD=N. Exactly the n-gram hunk (config.rs default+comment, loop_guard.rs doc) 3-way-merged onto current master, nothing else. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Coherence regression restore. Master's
HIPFIRE_NGRAM_LOOP_THRESHOLDis back at default 8 — the content-blind 4-gram×8 loop guard false-positives on legitimate enumeration-heavy reasoning at temp>0: a repeating\n - **markdown-bullet scaffold trips it inside<think>, force-EOSes mid-think, and empties the answer channel (the qwen3.6 temp=1.0 registry default hits this). Restores the validated 7eac3c8 behavior: default 8 → 0 (OFF); runaway thinking is handled by the think-budget force-close instead. Opt back in withHIPFIRE_NGRAM_LOOP_THRESHOLD=N. Exactly the n-gram hunk (config.rs default+comment + loop_guard.rs doc), 3-way-merged onto current master — no other 7eac3c8 changes. Compiles (cargo check -p hipfire-runtime).