spec-b300 / checkpoints /train_command.txt
khazic's picture
Archive three-epoch run: logs and provenance part 1
a181ec9 verified
Raw History Blame Contribute Delete
1.9 kB
# Timestamp: 2026-09-11T20:26:30.779193+00:00
# Git SHA: cd5091edb48f3de74eee315e508161f494ef3cb1
# World size: 3
# speculators: 0.8.0.dev0
# vllm: 0.27.1
# transformers: 5.15.0
# torch: 2.13.0+cu130
# compressed-tensors: 0.17.0
/lustre/fs1/portfolios/coreai/projects/coreai_dlalgo_llm/users/huiyingl/sd/speculators/logs/ngram-candidates-3epoch-b300-20260910/source/scripts/train.py --verifier-name-or-path /lustre/fs1/portfolios/coreai/projects/coreai_dlalgo_llm/users/huiyingl/sd/model_weights/Qwen--Qwen3.6-35B-A3B --data-path /lustre/fs1/portfolios/coreai/projects/coreai_dlalgo_llm/users/huiyingl/sd/datasets/qwen3.6-35b-a3b/qwen3.6-35b-a3b_train_spec_len3072_fullvocab --save-path /lustre/fs1/portfolios/coreai/projects/coreai_dlalgo_llm/users/huiyingl/sd/model_weights/ngram_candidates_qwen3_6-35b-a3b-perfectblend/ngram-candidates-3epoch-b300-20260910/checkpoints --epochs 3 --train-data-ratio 0.98 --optimizer muon --muon-lr 2e-4 --lr 1e-4 --weight-decay 0.01 --noise-std 0 --scheduler-type cosine --scheduler-warmup-ratio 0.01 --total-seq-len 16384 --speculator-type ngram_candidates --block-size 7 --sample-from-anchor --max-anchors 4096 --num-layers 5 --dflash-decay-gamma 7 --target-layer-ids 1 10 19 28 37 --candidate-top-k 256 --proposal-loss-weight 1.0 --ngram-rank 256 --ngram-num-buckets 65536 --per-position-loss-weight dpace --loss-fn ce --vllm-endpoint http://127.0.0.1:8000/v1 --request-timeout 180 --on-missing generate --on-generate delete --logger wandb --log-dir /lustre/fs1/portfolios/coreai/projects/coreai_dlalgo_llm/users/huiyingl/sd/speculators/logs/ngram-candidates-3epoch-b300-20260910/segments/7086788 --checkpoint-freq 0.1 --run-name ngram-candidates-3epoch-b300-20260910 --hidden-states-backend file --hidden-states-path /tmp/root/speculators-hidden-states/ngram_candidates_7086788.wi11j8 --full-attention-indices 0 1 2 3 4 --checkpoint-freq 0.01 --log-freq 10