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

### 1
hf download Qwen/Qwen3-4B
  --local-dir /root/models/Qwen3-4B

### 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
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; 

### 5
export PYTHONUNBUFFERED=1 && ray start
  --head
  --node-ip-address 127.0.0.1
  --num-gpus 8
  --disable-usage-stats

### 6
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
  --swiglu
  --num-layers 36
  --hidden-size 2560
  --ffn-hidden-size 9728
  --num-attention-heads 32
  --group-query-attention
  --num-query-groups 8
  --use-rotary-position-embeddings
  --disable-bias-linear
  --normalization RMSNorm
  --norm-epsilon 1e-6
  --rotary-base 1000000
  --vocab-size 151936
  --kv-channels 128
  --qk-layernorm
  --hf-checkpoint /root/models/Qwen3-4B
  --megatron-to-hf-mode bridge
  --save /root/shared_data/260101-000000-000/checkpoints
  --save-interval 50 
  --lora-rank 64
  --lora-alpha 32
  --lora-dropout 0.0
  --target-modules "all-linear"
  --lora-base-cpu-backup 
  --prompt-data /root/datasets/dapo-math-17k/dapo-math-17k.jsonl
  --input-key prompt
  --label-key label
  --apply-chat-template
  --rollout-shuffle
  --balance-data
  --rm-type deepscaler
  --num-rollout 100
  --rollout-batch-size 8
  --n-samples-per-prompt 8
  --rollout-max-response-len 8192
  --rollout-temperature 1
  --global-batch-size 64 
  --eval-interval 20
  --eval-prompt-data aime24 /root/datasets/aime-2024/aime-2024.jsonl
  --n-samples-per-eval-prompt 16
  --eval-max-response-len 16384
  --eval-top-p 1 
  --optimizer adam
  --lr 2e-05
  --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 0.2
  --eps-clip-high 0.28 
  --use-wandb
  --wandb-project miles-run_qwen3_4b_lora
  --wandb-group 260101-000000-000
  --wandb-key 'frozen-wandb-api-key'
  --disable-wandb-random-suffix 
  --rollout-num-gpus-per-engine 1
  --sglang-mem-fraction-static 0.4
  --sglang-decode-log-interval 1000
  --sglang-chunked-prefill-size 4096
  --sglang-lora-backend triton
  --sglang-attention-backend triton 
  --train-backend megatron
  --attention-dropout 0.0
  --hidden-dropout 0.0
  --accumulate-allreduce-grads-in-fp32
  --attention-softmax-in-fp32
  --calculate-per-token-loss
  --actor-num-nodes 1
  --actor-num-gpus-per-node 8
  --num-gpus-per-node 8
  --colocate   
