From 17a81b9c86cfedd70812a0e83f33798b64c1678e Mon Sep 17 00:00:00 2001 From: yurenh Date: Mon, 31 Aug 2026 18:16:31 -0500 Subject: data prep (FineWeb-Edu->GPT2 BPE), rho diagnostics, tests, measured-cost notes Co-Authored-By: Claude Fable 5 Claude-Session: https://claude.ai/code/session_01GkgLsACEF6CCP7EUfA5fZe --- README.md | 6 ++++++ 1 file changed, 6 insertions(+) (limited to 'README.md') diff --git a/README.md b/README.md index c9f5821..9c038d5 100644 --- a/README.md +++ b/README.md @@ -21,3 +21,9 @@ python scripts/prepare_data.py --dataset fineweb-edu --tokens 3e9 --out data/fin torchrun --nproc_per_node=8 scripts/train.py --model configs/model/m124.yaml --train configs/train/zbp_n16.yaml ``` Global batch is fixed in the train config; per-rank micro-batch and accumulation adapt to world size. + +## Measured cost (48 GB Ampere-class, d=512 L=8 seq 1024, shared GPU) +BP 312 ms/step; ZBP n=16 **8.7x**, n=64 **29x** (FLOPs-bound: probe batching is already saturated at +probe_chunk 8; the score core uses its own chunk <= 4 since its memory goes as chunk*B*H*T^2). +Ladder arms: **bp / zbp_n16 / zbp_n64** per size (n=4 optional). Headroom if needed: torch.compile on the +query path; forward differences (n+1 instead of 2n queries) as a cheaper biased arm. -- cgit v1.2.3