### 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", "NCCL_TIMEOUT": "3600", "PYTHONPATH": "<REPO_ROOT>:/root/Megatron-LM:/root/sglang/python:/frozen/pythonpath", "SGLANG_JIT_ROUTE_RADIX": "1", "SGLANG_NUMA_BIND_V2": "0", "SGLANG_K3_AR_FUSION": "0"}}'
  -- python3 <REPO_ROOT>/train.py
  --spec miles_plugins.models.kimi_k3 get_kimi_k3_spec
  --disable-bias-linear
  --num-layers 4
  --hidden-size 7168
  --ffn-hidden-size 33792
  --num-attention-heads 96
  --num-query-groups 96
  --kv-channels 256
  --normalization RMSNorm
  --position-embedding-type none
  --rotary-base 10000
  --norm-epsilon 1e-5
  --hidden-dropout 0
  --attention-dropout 0
  --disable-bf16-reduced-precision-matmul
  --swiglu
  --untie-embeddings-and-output-weights
  --vocab-size 163840
  --make-vocab-size-divisible-by 128
  --multi-latent-attention
  --q-lora-rank 1536
  --kv-lora-rank 512
  --qk-head-dim 128
  --qk-pos-emb-head-dim 64
  --v-head-dim 128
  --qk-layernorm
  --attention-softmax-in-fp32
  --num-experts 64
  --moe-layer-freq '[0,1,1,1]'
  --moe-ffn-hidden-size 3072
  --moe-latent-size 3584
  --moe-shared-expert-intermediate-size 6144
  --moe-router-topk 16
  --moe-router-score-function sigmoid
  --moe-router-pre-softmax
  --moe-router-enable-expert-bias
  --moe-router-load-balancing-type none
  --moe-token-dispatcher-type alltoall
  --moe-router-bias-update-rate 0
  --moe-aux-loss-coeff 0
  --moe-router-topk-scaling-factor 1.0
  --moe-router-dtype fp32
  --moe-grouped-gemm
  --moe-permute-fusion
  --hf-checkpoint <SANDBOX>/models/Kimi-K3-4layer-64experts
  --ref-load <SANDBOX>/models/Kimi-K3-4layer-64experts-bf16_torch_dist
  --megatron-to-hf-mode raw
  --model-name kimi_k3 
  --lora-rank 16
  --lora-alpha 32
  --lora-dropout 0.0
  --target-modules "all-linear"
  --no-gradient-accumulation-fusion
  --lora-base-cpu-backup
  --experts-shared-outer-loras 
  --prompt-data /root/datasets/gsm8k/train.parquet
  --input-key messages
  --label-key label
  --apply-chat-template
  --rollout-shuffle
  --balance-data
  --rm-type math
  --num-rollout 2
  --rollout-batch-size 8
  --n-samples-per-prompt 2
  --rollout-max-response-len 256
  --rollout-temperature 1
  --global-batch-size 16
  --use-dynamic-global-batch-size 
  --optimizer sgd
  --sgd-momentum 0
  --lr 1e-6
  --lr-decay-style constant
  --weight-decay 0
  --use-distributed-optimizer 
  --advantage-estimator grpo
  --kl-loss-coef 0.0
  --kl-loss-type low_var_kl
  --entropy-coef 0.0
  --eps-clip 0.2
  --eps-clip-high 0.28 
  --use-wandb
  --wandb-project miles-run_kimi_k3
  --wandb-group 260101-000000-000
  --wandb-key 'frozen-wandb-api-key'
  --disable-wandb-random-suffix 
  --tensor-model-parallel-size 8
  --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 512
  --log-probs-chunk-size 512
  --distributed-timeout-minutes 10 
  --rollout-num-gpus-per-engine 8
  --sglang-tp-size 8
  --sglang-ep-size 1
  --sglang-server-concurrency 16
  --sglang-max-running-requests 16
  --sglang-max-mamba-cache-size 16
  --use-miles-router
  --sglang-lora-backend triton
  --sglang-lora-strict-loading
  --sglang-max-lora-rank 16
  --sglang-moe-runner-backend marlin
  --sglang-cuda-graph-bs-decode 1 2 4 8 16
  --sglang-mem-fraction-static 0.7
  --sglang-disable-shared-experts-fusion
  --sglang-context-length 8192 
  --attention-dropout 0.0
  --hidden-dropout 0.0
  --accumulate-allreduce-grads-in-fp32
  --colocate
  --offload-train
  --update-weight-buffer-size 2147483648
  --train-memory-margin-bytes 2147483648
  --actor-num-nodes 1
  --actor-num-gpus-per-node 8
  --num-gpus-per-node 8
  --no-check-for-nan-in-loss-and-grad
  --check-weight-update-skip-list vision_tower. mm_projector.
  --check-weight-update-allow-quant-error
  --save /personal/checkpoints/260101-000000-000
  --save-interval 50   
  --deploy-component all
