Skip to content

Non-record: 4090 single-GPU ablations on ValCalib GPTQ + XSA stack (partial logs)#1226

Open
Wolfie8935 wants to merge 6 commits intoopenai:mainfrom
Wolfie8935:chore/runpod-winning-harness
Open

Non-record: 4090 single-GPU ablations on ValCalib GPTQ + XSA stack (partial logs)#1226
Wolfie8935 wants to merge 6 commits intoopenai:mainfrom
Wolfie8935:chore/runpod-winning-harness

Conversation

@Wolfie8935
Copy link
Copy Markdown

This PR adds a non-record submission under records/track_non_record_16mb/2026-04-01_Wolfie8935_4090_ValCalib_ablations/.

Contents

README.md — context, hardware limits (1×RTX 4090, TRAIN_BATCH_TOKENS=196608), which ablations are included, and known log caveats.
submission.json — metadata; no claim of a new 8×H100 / 10-minute record.
train_gpt.py + requirements.txt — same lineage as the reference ValCalib record (2026-03-25_ValCalib_GPTQ_XSA_BigramHash3072).
train_*.log — Runpod logs for ctrl, a1, a2, b1 (seed 314). Ablations c1 / c2 / d1 / d2 are not included in this PR; ctrl may lack final_int6_sliding_window_exact if the captured run was interrupted (documented in README).
Intent
Document budget-GPU exploration and methodology on the public stack; scores are not comparable to 8×H100 SOTA due to throughput and batch settings.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant