### 0
mkdir -p /root/models /root/datasets

### 1
hf download zai-org/GLM-4.5
  --local-dir /root/models/GLM-4.5

### 2
hf download
  --repo-type dataset zhuzilin/dapo-math-17k
  --local-dir /root/datasets/dapo-math-17k

### 3
hf download
  --repo-type dataset zhuzilin/aime-2024
  --local-dir /root/datasets/aime-2024

### 4
hf download
  --repo-type dataset zhuzilin/aime-2025
  --local-dir /root/datasets/aime-2025

### 5
[multi_node num_nodes=1] PYTHONPATH=<REPO_ROOT>:/root/Megatron-LM:/frozen/pythonpath torchrun
  --nproc-per-node 4
  --master-addr {{master_addr}}
  --master-port 23456
  --nnodes={{nnodes}}
  --node-rank {{node_rank}} <REPO_ROOT>/tools/convert_hf_to_torch_dist.py
  --disable-bias-linear
  --qk-layernorm
  --group-query-attention
  --num-attention-heads 96
  --num-query-groups 8
  --kv-channels 128
  --num-layers 92
  --hidden-size 5120
  --ffn-hidden-size 12288
  --add-qkv-bias
  --normalization RMSNorm
  --position-embedding-type rope
  --rotary-percent 0.5
  --swiglu
  --untie-embeddings-and-output-weights
  --vocab-size 151552
  --rotary-base 1000000
  --moe-ffn-hidden-size 1536
  --moe-shared-expert-intermediate-size 1536
  --moe-router-pre-softmax
  --moe-router-score-function sigmoid
  --moe-router-enable-expert-bias
  --moe-router-bias-update-rate 0
  --moe-router-load-balancing-type seq_aux_loss
  --moe-token-dispatcher-type alltoall
  --moe-router-topk 8
  --moe-router-topk-scaling-factor 2.5
  --moe-layer-freq '[0]*3+[1]*89'
  --num-experts 160
  --moe-grouped-gemm
  --moe-router-dtype fp32
  --moe-permute-fusion
  --moe-aux-loss-coeff 0
  --hf-checkpoint /root/models/GLM-4.5
  --save /root/models/GLM-4.5_torch_dist 
