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

### 1
hf download zai-org/GLM-4.7-Flash
  --local-dir /root/models/GLM-4.7-Flash

### 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
PYTHONPATH=<REPO_ROOT>:/root/Megatron-LM:/frozen/pythonpath torchrun
  --nproc-per-node 8 <REPO_ROOT>/tools/convert_hf_to_torch_dist.py
  --moe-layer-freq '[0]*1+[1]*46'
  --num-experts 64
  --moe-shared-expert-intermediate-size 1536
  --moe-router-topk 4
  --moe-grouped-gemm
  --moe-permute-fusion
  --moe-ffn-hidden-size 1536
  --moe-router-score-function sigmoid
  --moe-router-pre-softmax
  --moe-router-enable-expert-bias
  --moe-router-bias-update-rate 0
  --moe-router-load-balancing-type seq_aux_loss
  --moe-router-topk-scaling-factor 1.8
  --moe-aux-loss-coeff 0
  --moe-router-dtype fp32
  --make-vocab-size-divisible-by 64
  --num-layers 47
  --hidden-size 2048
  --ffn-hidden-size 10240
  --num-attention-heads 20
  --disable-bias-linear
  --add-qkv-bias
  --swiglu
  --untie-embeddings-and-output-weights
  --position-embedding-type rope
  --rope-type rope
  --no-position-embedding
  --normalization RMSNorm
  --norm-epsilon 1e-5
  --qk-layernorm
  --multi-latent-attention
  --q-lora-rank 768
  --kv-lora-rank 512
  --qk-head-dim 192
  --v-head-dim 256
  --kv-channels 192
  --qk-pos-emb-head-dim 64
  --vocab-size 154880
  --rotary-base 1000000
  --no-rope-fusion
  --mtp-num-layers 1
  --hf-checkpoint /root/models/GLM-4.7-Flash
  --save /root/models/GLM-4.7-Flash_torch_dist 
