Views
No views yet
1CUDA_VISIBLE_DEVICES=5 python train_rotation.py \
2 --model_dir /models/Qwen_Qwen3-0.6B \
3 --quant_scheme mxfp4 \
4 --eval_batch_size 16 \
5 --num_samples 4000 \
6 --learning_rate 1.5 \
7 --rotation_algo_config_file ${ROTATION_TRAIN_CFG} \
8 --output_dir qwen3-0.6b-mxfp4-tuned-orthogonal \
9 --max_steps 5 \
10 --loss_type kl_top_1000 \
11 --model_attn_implementation sdpa \
12 --train_batch_size 8 \
13 --model_export hf_format \
14 --export_weight_format real_quantized