serving: audit the 27B on exact tokens only until sparkinfer's /v1/score logprobs are fixed - #1760
Merged
Merged
Conversation
…ore logprobs are fixed
On sparkinfer 19ef39ec2, /v1/score with completion_token_ids returns the
right argmax at every position (argmax agreement 1.000 on every prompt
tried) but a wrong logprob VALUE at the first forced positions: ~20-30
nats off at position 1 on every call, position 0 on the first call after
a generation, and the occasional later position. Deterministic, on both
compute3 cards, 2026-09-08. Generation itself agrees bit-for-bit across
servers (0.0 drift), so the logprob check was failing every honest miner
on the first mainnet rounds ("reference disagreement: logprob drift").
The release's audit block keeps min_prefix_agreement 1.0 (the miner must
reproduce the reference's greedy tokens exactly) and lifts the two
logprob bands to 100 nats until upstream fixes the scoring path. The
validator already runs this from a loadout override.
Claude-Session: https://claude.ai/code/session_01VjhStJbNW6WzxucTcwuw4y
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
First mainnet rounds on the 27B failed every honest miner with
reference disagreement: logprob drift. Root cause, measured on compute3 (both cards, deterministic, repeatable): sparkinfer19ef39ec2's/v1/scorewithcompletion_token_idsreturns the correct argmax at every position (agreement 1.000 on every prompt tried) but a wrong logprob value at the first forced positions, ~20–30 nats off at position 1 on every call, position 0 on the first call after a generation, and occasionally a later position. Generation agrees bit-for-bit across the two servers (0.0 drift), so the model and pin are fine; the teacher-forced scoring path is not.This keeps
min_prefix_agreement1.0 (miners must reproduce the reference's greedy tokens exactly, which is the strong half of the check) and lifts the two logprob bands to 100 nats for this release only, via the release'sauditblock. The validator is already running this as a loadout override (SERVING_LOADOUT_PATH). Upstream issue to follow; revert the bands when the scoring path is fixed and re-qualified.https://claude.ai/code/session_01VjhStJbNW6WzxucTcwuw4y