### 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 8
  --disable-usage-stats

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

### 3
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
  --group-query-attention
  --num-attention-heads 16
  --num-query-groups 8
  --kv-channels 256
  --num-layers 30
  --hidden-size 2816
  --ffn-hidden-size 2112
  --normalization RMSNorm
  --norm-epsilon 1e-06
  --position-embedding-type rope
  --rotary-base 1000000
  --vocab-size 262144
  --make-vocab-size-divisible-by 128
  --max-position-embeddings 262144
  --num-experts 128
  --moe-router-topk 8
  --moe-ffn-hidden-size 704
  --moe-router-score-function softmax
  --moe-grouped-gemm
  --moe-router-dtype fp32
  --moe-router-num-groups 1
  --moe-router-group-topk 1
  --moe-router-load-balancing-type seq_aux_loss
  --moe-router-bias-update-rate 0
  --moe-aux-loss-coeff 0
  --hf-checkpoint /root/models/gemma-4-26B-A4B-it
  --ref-load /root/models/gemma-4-26B-A4B-it
  --megatron-to-hf-mode bridge
  --load /root/shared_data/260101-000000-000/checkpoints
  --save /root/shared_data/260101-000000-000/checkpoints
  --save-interval 2 
  --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 gemma_math
  --num-rollout 3000
  --rollout-batch-size 32
  --n-samples-per-prompt 8
  --rollout-max-response-len 256
  --rollout-temperature 1
  --global-batch-size 256 
  --optimizer adam
  --lr 1e-6
  --lr-decay-style constant
  --weight-decay 0.1
  --adam-beta1 0.9
  --adam-beta2 0.98 
  --advantage-estimator grpo
  --use-kl-loss
  --kl-loss-coef 0.00
  --kl-loss-type low_var_kl
  --entropy-coef 0.00
  --eps-clip 0.2
  --eps-clip-high 0.28 
  --use-wandb
  --wandb-project miles-run_gemma_4_26b_a4b
  --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 8
  --expert-tensor-parallel-size 1
  --recompute-granularity full
  --recompute-method uniform
  --recompute-num-layers 1
  --use-dynamic-batch-size
  --max-tokens-per-gpu 1024  
  --rollout-num-gpus-per-engine 4
  --sglang-mem-fraction-static 0.55
  --sglang-attention-backend triton
  --sglang-moe-runner-backend triton
  --sglang-disable-custom-all-reduce
  --sglang-disable-cuda-graph
  --sglang-disable-overlap-schedule
  --sglang-disable-radix-cache
  --no-offload-train
  --no-offload-rollout
  --use-rollout-routing-replay 
  --attention-dropout 0.0
  --hidden-dropout 0.0
  --accumulate-allreduce-grads-in-fp32
  --no-gradient-accumulation-fusion
  --no-check-for-nan-in-loss-and-grad
  --attention-softmax-in-fp32
  --attention-backend unfused
  --qkv-format bshd
  --colocate
  --actor-num-nodes 1
  --actor-num-gpus-per-node 8
  --num-gpus-per-node 8   
