### 0
pkill -9 sglang; sleep 3; ray stop
  --force; pkill -9 ray; pkill -9 miles; sleep 3; pkill -9 ray; pkill -9 miles; pkill -9 redis; true; 

### 1
export PYTHONUNBUFFERED=1 && ray start
  --head
  --node-ip-address 127.0.0.1
  --num-gpus 4
  --disable-usage-stats

### 2
[multi_node num_nodes=None] mkdir -p /root/local_data/GLM-4.5_torch_dist && rsync -a
  --info=progress2 /root/models/GLM-4.5_torch_dist/ /root/local_data/GLM-4.5_torch_dist && mkdir -p /root/local_data/GLM-4.5 && rsync -a
  --info=progress2 /root/models/GLM-4.5/ /root/local_data/GLM-4.5

### 3
nvidia-smi topo -m 2>/dev/null | grep -o 'NV[0-9][0-9]*' | wc -l

### 4
export no_proxy=127.0.0.1 && export PYTHONUNBUFFERED=1 && ray job submit
  --address="http://127.0.0.1:8265"
  --runtime-env-json='{"env_vars": {"PYTHONUNBUFFERED": "1", "CUDA_DEVICE_MAX_CONNECTIONS": "1", "NCCL_NVLS_ENABLE": "0", "no_proxy": "127.0.0.1,127.0.0.1", "MASTER_ADDR": "127.0.0.1", "PYTHONPATH": "<REPO_ROOT>:/root/Megatron-LM:/frozen/pythonpath"}}'
  -- python3 <REPO_ROOT>/train.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/local_data/GLM-4.5
  --ref-load /root/local_data/GLM-4.5_torch_dist
  --load /root/shared_data/260101-000000-000/checkpoints
  --save /root/shared_data/260101-000000-000/checkpoints
  --save-interval 10
  --save-retain-interval 10 
  --label-key label
  --apply-chat-template
  --rollout-shuffle
  --rm-type math
  --num-rollout 3000
  --rollout-batch-size 32
  --n-samples-per-prompt 8
  --rollout-temperature 1
  --num-steps-per-rollout 1
  --balance-data
  --rollout-stop-token-ids 151329 151336 151338
  --prompt-data /root/datasets/dapo-math-17k/dapo-math-17k.jsonl
  --input-key prompt
  --rollout-max-response-len 8192 
  --optimizer adam
  --lr 1e-6
  --lr-decay-style constant
  --weight-decay 0.1
  --adam-beta1 0.9
  --adam-beta2 0.98 
  --advantage-estimator grpo
  --kl-loss-coef 0.00
  --kl-loss-type low_var_kl
  --kl-coef 0.00
  --entropy-coef 0.00
  --eps-clip 1e-4
  --eps-clip-high 2e-4
  --use-tis 
  --use-wandb
  --wandb-project miles-run_glm45_355b_a32b
  --wandb-group 260101-000000-000
  --wandb-key 'frozen-wandb-api-key'
  --disable-wandb-random-suffix 
  --tensor-model-parallel-size 4
  --sequence-parallel
  --pipeline-model-parallel-size 1
  --context-parallel-size 1
  --expert-model-parallel-size 4
  --expert-tensor-parallel-size 1
  --recompute-granularity full
  --recompute-method uniform
  --recompute-num-layers 1
  --use-dynamic-batch-size
  --max-tokens-per-gpu 16384 
  --eval-interval 20
  --eval-top-p 1
  --eval-prompt-data aime /root/datasets/aime-2024/aime-2024.jsonl
  --n-samples-per-eval-prompt 8
  --eval-max-response-len 8192 
  --rollout-num-gpus-per-engine 4
  --sglang-mem-fraction-static 0.7
  --sglang-tp-size 4
  --sglang-chunked-prefill-size 8192
  --sglang-attention-backend trtllm_mha 
  --attention-dropout 0.0
  --hidden-dropout 0.0
  --accumulate-allreduce-grads-in-fp32
  --attention-softmax-in-fp32
  --attention-backend flash
  --update-weight-buffer-size 2147483648
  --actor-num-nodes 1
  --actor-num-gpus-per-node 4
  --num-gpus-per-node 4
  --colocate
  --use-fault-tolerance
  --dump-details /root/shared_data/260101-000000-000/dump_details
  --router-health-success-threshold 1
  --router-health-check-interval-secs 15   
