tacit/nanochat
1
0
Fork 0
mirror of https://github.com/karpathy/nanochat.git synced 2026-05-19 14:17:36 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity

speedrun.sh: ratio=11.05 + --final-lr-frac=0.0 (CORE 0.2665 in 88 min)

Browse Source
This commit is contained in:
gio 2026-04-27 20:27:24 -05:00
parent cc2f2abdf0
commit 1e7810ddaa
1 changed files with 7 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

13
runs/speedrun.sh
View File

@ -69,12 +69,13 @@ python -m scripts.tok_eval
echo "Waiting for dataset download to complete..."
wait $DATASET_DOWNLOAD_PID
# d22 model trained for 6000 iterations at 1M tokens/iter = 6B tokens (~ratio=11 against d22's
# scaling params, mirror of Run 6's d24+ratio=8 strategy from the other side of compute-optimal —
# d22 is below GPT-2 capability so we overtrain). Combined with --warmdown-ratio=0.85 (longer
# low-LR tail) and --muon-qk-clip-tau=100 (Kimi K2 §A QK-Clip) the recipe crosses GPT-2 CORE
# in 88 min — ~10.8% less wall-clock than Run 6 — at CORE 0.2646, val_bpb 0.7241.
torchrun --standalone --nproc_per_node=8 -m scripts.base_train -- --depth=22 --num-iterations=6000 --total-batch-size=1048576 --device-batch-size=16 --warmdown-ratio=0.85 --muon-qk-clip-tau=100 --fp8 --run=$WANDB_RUN
# d22 model overtrained relative to compute-optimal 10.5 (mirror of Run 6's d24+ratio=8
# undertrained strategy from the other side of compute-optimal — d22 is below GPT-2
# capability so we overtrain at ratio=11.05). Combined with --warmdown-ratio=0.85 (longer
# low-LR tail), --final-lr-frac=0.0 (full LR decay floor; Hägele et al. arxiv 2405.18392),
# and --muon-qk-clip-tau=100 (Kimi K2 §A QK-Clip) the recipe crosses GPT-2 CORE in ~88 min
# — ~10.9% less wall-clock than Run 6 — at CORE 0.2665, val_bpb 0.7242.
torchrun --standalone --nproc_per_node=8 -m scripts.base_train -- --depth=22 --target-param-data-ratio=11.05 --total-batch-size=1048576 --device-batch-size=16 --warmdown-ratio=0.85 --final-lr-frac=0.0 --muon-qk-clip-tau=100 --fp8 --run=$WANDB_RUN
# evaluate the model: CORE metric, BPB on train/val, and draw samples
torchrun --standalone --nproc_per_node=8 -m scripts.base_eval -- --device-batch-size=16