=== [Sat May 23 21:41:30 UTC 2026] adapter_v2 resume train (500 steps, seq=2048) === [transformers] `torch_dtype` is deprecated! Use `dtype` instead! The fast path is not available because on of `(selective_state_update, causal_conv1d_fn, causal_conv1d_update)` is None. Falling back to the naive implementation. To install follow https://github.com/state-spaces/mamba/#installation and https://github.com/Dao-AILab/causal-conv1d === TrainConfig === base_model: /home/lakesenberg/models/nemotron-3-nano-30b-a3b-bf16 data_path: /home/lakesenberg/nemotron_lora_work/cot.jsonl out_dir: /home/lakesenberg/nemotron_lora_work/adapter_v2 lora_rank: 32 lora_alpha: 32 lora_dropout: 0.0 target_modules: ('q_proj', 'k_proj', 'v_proj', 'o_proj', 'gate_proj', 'up_proj', 'down_proj', 'x_proj', 'in_proj', 'out_proj', 'dt_proj', 'experts.w1', 'experts.w2') lr: 5e-05 weight_decay: 0.0 betas: (0.9, 0.95) max_grad_norm: 1.0 warmup_steps: 50 total_steps: 500 grad_accum: 4 batch_size: 1 seq_len: 2048 save_every: 100 log_every: 10 seed: 0 bf16: True use_flash_attn: False gradient_checkpointing: True resume_adapter: /home/lakesenberg/nemotron_lora_work/adapter/final Loading /home/lakesenberg/models/nemotron-3-nano-30b-a3b-bf16 ... Loading weights: 0%| | 0/6243 [00:00 /home/lakesenberg/nemotron_lora_work/adapter_v2/step-00100 step 110/500 loss=0.0013 ema=0.0011 lr=4.78e-05 mem=69.1/86.8 GiB step 120/500 loss=0.0000 ema=0.0009 lr=4.71e-05 mem=69.1/86.8 GiB step 130/500 loss=0.0000 ema=0.0009 lr=4.62e-05 mem=68.8/86.8 GiB step 140/500 loss=0.0000 ema=0.0005 lr=4.52e-05 mem=69.0/86.8 GiB step 150/500 loss=0.0000 ema=0.0009 lr=4.42e-05 mem=68.9/86.8 GiB step 160/500 loss=0.0000 ema=0.0009 lr=4.30e-05 mem=68.9/86.8 GiB step 170/500 loss=0.0000 ema=0.0009 lr=4.17e-05 mem=68.8/86.8 GiB step 180/500 loss=0.0000 ema=0.0005 lr=4.04e-05 mem=68.8/86.8 GiB step 190/500 loss=0.0015 ema=0.0012 lr=3.90e-05 mem=69.1/86.8 GiB step 200/500 loss=0.0000 ema=0.0010 lr=3.75e-05 mem=68.8/86.8 GiB saved checkpoint -> /home/lakesenberg/nemotron_lora_work/adapter_v2/step-00200 step 210/500 loss=0.0000 ema=0.0010 lr=3.60e-05 mem=69.1/86.8 GiB step 220/500 loss=0.0000 ema=0.0005 lr=3.44e-05 mem=68.8/86.8 GiB step 230/500 loss=0.0000 ema=0.0005 lr=3.27e-05 mem=68.8/86.8 GiB step 240/500 loss=0.0000 ema=0.0002 lr=3.10e-05 mem=68.9/86.8 GiB step 250/500 loss=0.0000 ema=0.0004 lr=2.93e-05 mem=69.0/86.8 GiB step 260/500 loss=0.0000 ema=0.0004 lr=2.76e-05 mem=68.8/86.8 GiB step 270/500 loss=0.0000 ema=0.0009 lr=2.59e-05 mem=69.0/86.8 GiB step 280/500 loss=0.0000 ema=0.0008 lr=2.41e-05 mem=69.0/86.8 GiB step 290/500 loss=0.0007 ema=0.0009 lr=2.24e-05 mem=69.1/86.8 GiB step 300/500 loss=0.0000 ema=0.0004 lr=2.07e-05 mem=68.8/86.8 GiB saved checkpoint -> /home/lakesenberg/nemotron_lora_work/adapter_v2/step-00300 step 310/500 loss=0.0010 ema=0.0009 lr=1.90e-05 mem=69.5/86.8 GiB step 320/500 loss=0.0000 ema=0.0005 lr=1.73e-05 mem=68.8/86.8 GiB step 330/500 loss=0.0000 ema=0.0003 lr=1.56e-05 mem=69.0/86.8 GiB step 340/500 loss=0.0000 ema=0.0003 lr=1.40e-05 mem=68.8/86.8 GiB step 350/500 loss=0.0001 ema=0.0006 lr=1.25e-05 mem=69.0/86.8 GiB step 360/500 loss=0.0018 ema=0.0005 lr=1.10e-05 mem=68.8/86.8 GiB step 370/500 loss=0.0000 ema=0.0007 lr=9.61e-06 mem=68.9/86.8 GiB step 380/500 loss=0.0005 ema=0.0004 lr=8.27e-06 mem=68.8/86.8 GiB step 390/500 loss=0.0000 ema=0.0005 lr=7.02e-06 mem=69.1/86.8 GiB step 400/500 loss=0.0000 ema=0.0004 lr=5.85e-06 mem=69.0/86.8 GiB saved checkpoint -> /home/lakesenberg/nemotron_lora_work/adapter_v2/step-00400 step 410/500 loss=0.0045 ema=0.0008 lr=4.77e-06 mem=69.6/86.8 GiB step 420/500 loss=0.0014 ema=0.0012 lr=3.80e-06 mem=68.8/86.8 GiB step 430/500 loss=0.0000 ema=0.0005 lr=2.93e-06 mem=69.0/86.8 GiB step 440/500 loss=0.0000 ema=0.0004 lr=2.16e-06 mem=68.9/86.8 GiB step 450/500 loss=0.0000 ema=0.0003 lr=1.51e-06 mem=68.9/86.8 GiB step 460/500 loss=0.0008 ema=0.0004 lr=9.68e-07 mem=69.0/86.8 GiB step 470/500 loss=0.0000 ema=0.0006 lr=5.46e-07 mem=68.9/86.8 GiB step 480/500 loss=0.0000 ema=0.0005 lr=2.43e-07 mem=68.9/86.8 GiB step 490/500 loss=0.0000 ema=0.0004 lr=6.09e-08 mem=68.8/86.8 GiB step 500/500 loss=0.0020 ema=0.0008 lr=0.00e+00 mem=69.1/86.8 GiB saved checkpoint -> /home/lakesenberg/nemotron_lora_work/adapter_v2/step-00500 DONE. final adapter -> /home/lakesenberg/nemotron_lora_work/adapter_v2/final === [Sun May 24 01:02:23 UTC 2026] DONE adapter_v2 -> ~/nemotron_lora_work/adapter_v2/final ===