Views
No views yet
agillm35.py. The helper modules are folded into that one file so the runtime can be cloned, inspected, and launched without restoring the whole AGILLM4 source tree.deepseek-ai/DeepSeek-V3.2large (d=1024, layers=24, heads=16, rank=128)--agillm3_compat--dblock1python agillm35.py --help
2python agillm35.py status --ckpt /path/to/pretrain_step00051081.pt
3python agillm35.py infer --ckpt /path/to/pretrain_step00051081.pt --prompt "Hello"1python agillm35.py train \
2 --agillm3_compat \
3 --preset large \
4 --resume /path/to/pretrain_step00051081.pt \
5 --block 512 \
6 --batch_size 1 \
7 --source HuggingFaceFW/fineweb-edu \
8 --save_dir ckpts \
9 --dblock \
10 --dblock_blocks 8 \
11 --nat_every 0 \
12 --dblock_nat_weight 0loss plus the actual EDM-weighted training objective as weighted. The weighted value is the optimization target; the raw value is the sanity-check number to compare with ordinary AR/SAT loss.