Skip to content

Non-record: Entropy-regularized QAT (WIP, pending compute) #930

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
lamb356 wants to merge 3 commits into
base: main
Choose a base branch
from
Open

Non-record: Entropy-regularized QAT (WIP, pending compute) #930

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
58fce58
Non-record: Entropy-regularized QAT (WIP, pending compute)
Mar 27, 2026
8dd3709
Add Legal TTT + entropy penalty to submission script
Mar 27, 2026
ecd1e23
Update README with TTT support
Mar 27, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
20 changes: 20 additions & 0 deletions records/track_10min_16mb/2026-03-27_EntropyRegularized/README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,20 @@
# Entropy-Regularized QAT + Legal TTT (WIP - pending compute)

Built on the merged 1.1233 record (PR #374) with techniques from the 1.1194 record (PR #549).

## Changes from baseline

**Training:**
- LeakyReLU(0.5) activation (proven -0.003 BPB from PR #493)
- Differentiable entropy penalty on int6 quantized MLP weight symbols during late QAT window
- Entropy penalty gated by ENTROPY_LAMBDA env var (default 0, no behavior change)

**Eval:**
- Legal score-first TTT from PR #461/#549 (gated by TTT_ENABLED, default off)
- Sliding window eval stride=64

## Run command
ENTROPY_LAMBDA=0.003 TTT_ENABLED=1 SEED=1337 torchrun --standalone --nproc_per_node=8 train_gpt.py

## Status
Code complete, tested locally. Awaiting 8H100 runs for results.
1 change: 1 addition & 0 deletions records/track_10min_16mb/2026-03-27_EntropyRegularized/submission.json
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1 @@
{"author":"Carson","github_id":"lamb356","name":"Non-record: Entropy-Regularized QAT (WIP)","blurb":"Compressibility-aware training via entropy penalty on quantized weight symbols","date":"2026-03-27T00:00:00Z","val_loss":0,"val_bpb":0,"bytes_total":0}
Loading