Interpretability twin: load the deployed model.bin into brainscope - #16
Open
moudrkat wants to merge 1 commit into
Open
Interpretability twin: load the deployed model.bin into brainscope#16moudrkat wants to merge 1 commit into
moudrkat wants to merge 1 commit into
Conversation
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Dequantizes the exact int4 artifact deploy.sh flashes (not the training checkpoint) into a transformers-loadable twin and serves it live in brainscope — logit lens, attention, per-layer activity. Verified against runtime/llm.h: max logit diff 8e-6 on a fixed prompt, the same tolerance verify.c holds the C port to, plus KV-cache parity (brainscope_adapter/verify_vs_c.py, with a small C driver reusing llm.h unmodified).
Everything is additive — one new directory, no existing file touched. brainscope itself is untouched too: the PLE architecture registers with transformers in-process.
Bonus: build_hf.py --zero-ple-table builds a 'flash unplugged' ablation — the 25M flash-resident parameters zeroed, the storyteller collapsing to a loop. A direct, visible demo of what the PLE table contributes; extract_direction.py adds a story-mood steering direction from contrast pairs.
Demo video/post coming — will link here.