fix(rollout): support non extra gpu placement when using rollout-external mode

[shinytang6]

Background

rollout-external mode is already supported in slime — it lets the training job talk to a pre-launched SGLang server instead of spawning one. While
smoke-testing this path end-to-end (machine A hosts SGLang, machine B runs train.py) I hit a few bugs and one piece of waste that this PR cleans up.

What this PR fixes:

1. External engine never registers with the router.
2. Sanity check rejects fields that must differ for an external server.
3. Redundant local GPU reservation for the rollout group.
   • For non-colocate runs, slime reserves actor_num_gpus + rollout_num_gpus GPU bundles in the placement group, with rollout starting at rollout_offset =actor_num_gpus.
   • But under --rollout-external the local rollout actor is a thin HTTP wrapper — it doesn't touch local GPU memory, so reserving rollout_num_gpus extra bundles on the training node is pure waste (and fails outright when the training node doesn't have that many spare GPUs). Reuse the actor PG and set rollout_offset = 0 in both create_placement_groups and _compute_rollout_offset.

How to reproduce

Two machines. Machine A hosts the pre-launched SGLang server; machine B runs the slime training job and points at it.

1. Machine A — launch the external SGLang server

python -m sglang.launch_server \
    --model-path /path/to/model \
    --host 0.0.0.0 \
    --port 10091

Verify reachable from B: curl http://<SGLANG_HOST>:10091/get_server_info.

2. Machine B — minimal GRPO smoke test driving the external server

The only --rollout-external-specific bit is the SGLANG_ARGS block:

SGLANG_ARGS=(
     --rollout-num-gpus-per-engine 1
     --sglang-mem-fraction-static 0.4
     --rollout-external
     --rollout-external-engine-addrs <SGLANG_HOST>:10091
  )

Full script:

  #!/bin/bash
  set -ex

  pkill -9 sglang 2>/dev/null || true
  ray stop --force 2>/dev/null || true
  pkill -9 ray python 2>/dev/null || true
  sleep 2

  export PYTHONBUFFERED=1
  SCRIPT_DIR="$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" &>/dev/null && pwd)"
  source "${SCRIPT_DIR}/models/qwen2.5-0.5B.sh"

  CKPT_ARGS=(
     --hf-checkpoint /path/to/Qwen2.5-0.5B-Instruct/
     --ref-load     /path/to/Qwen2.5-0.5B-Instruct_torch_dist/
     --save /tmp/slime_smoke_save/
     --save-interval 9999
  )

  ROLLOUT_ARGS=(
     --prompt-data /path/to/dapo-math-17k.jsonl
     --input-key prompt --label-key label
     --apply-chat-template --rollout-shuffle
     --rm-type deepscaler
     --num-rollout 3 --rollout-batch-size 2
     --n-samples-per-prompt 2 --num-steps-per-rollout 1
     --global-batch-size 4
     --rollout-max-response-len 256 --rollout-temperature 1
  )

  PERF_ARGS=(
     --tensor-model-parallel-size 1 --pipeline-model-parallel-size 1
     --context-parallel-size 1 --expert-model-parallel-size 1
     --expert-tensor-parallel-size 1
     --use-dynamic-batch-size --max-tokens-per-gpu 2048
  )

  GRPO_ARGS=(
     --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
  )

  OPTIMIZER_ARGS=(
     --optimizer adam --lr 1e-6 --lr-decay-style constant
     --weight-decay 0.1 --adam-beta1 0.9 --adam-beta2 0.98
  )

  SGLANG_ARGS=(
     --rollout-num-gpus-per-engine 1
     --sglang-mem-fraction-static 0.4
     --rollout-external
     --rollout-external-engine-addrs <SGLANG_HOST>:10091
  )

  MISC_ARGS=(
     --attention-dropout 0.0 --hidden-dropout 0.0
     --accumulate-allreduce-grads-in-fp32
     --attention-softmax-in-fp32
     --attention-backend flash
  )

  ray start --head --node-ip-address 127.0.0.1 --num-gpus 2 --disable-usage-stats

  ray job submit --address="http://127.0.0.1:8265" \
     --runtime-env-json='{
       "env_vars": {
          "PYTHONPATH": "/path/to/Megatron-LM/",
          "CUDA_DEVICE_MAX_CONNECTIONS": "1",
       }
     }' \
     -- python3 train.py \
     --actor-num-nodes 1 \
     --actor-num-gpus-per-node 2 \
     "${MODEL_ARGS[@]}" \
     "${CKPT_ARGS[@]}" "${ROLLOUT_ARGS[@]}" \
     "${OPTIMIZER_ARGS[@]}" "${GRPO_ARGS[@]}" \
     "${PERF_ARGS[@]}" "${SGLANG_ARGS[@]}" "${MISC_ARGS[@]}"

[shinytang6]

@zhuzilin could you please help take a look on this PR?