Motivation

If the first request to the PD engine carries disagg_prefill_dp_rank, the request fails with:

Prefill server with bootstrap_addr: {self.bootstrap_addr} is healthy before

However, if the first request does not contain disagg_prefill_dp_rank but a later request does, the subsequent requests work because the first request triggers the prefill info query and caches it.

Root cause

In DecodePreallocQueue._resolve_prefill_dp_rank, the req.disagg_prefill_dp_rank early-return is checked before self.kv_manager.prefill_info_table.get(_bootstrap_addr(req)). When the client explicitly supplies disagg_prefill_dp_rank, we short-circuit and never trigger the slow path that queries and caches the prefill info. The subsequent prefill-server health check then has no cached info to validate against, producing the "healthy before" error.

Modifications

Move prefill_info = self.kv_manager.prefill_info_table.get(_bootstrap_addr(req)) to the top of _resolve_prefill_dp_rank. If the lookup returns None, return None so the request falls through to the slow path (_ensure_prefill_info), which queries and caches the prefill info. Only after prefill info is available do we honor the client-provided req.disagg_prefill_dp_rank.

Reproduction

# Prefill
python3 -m sglang.launch_server --model-path Qwen/Qwen3-30B-A3B \
    --tp 4 --dp 4 --enable-dp-attention --disaggregation-mode prefill

# Decode
python3 -m sglang.launch_server --model-path Qwen/Qwen3-30B-A3B \
    --tp 4 --dp 4 --enable-dp-attention --disaggregation-mode decode \
    --base-gpu-id 4 --port 30010

# Load balancer
python -m sglang_router.launch_router --mini-lb --pd-disaggregation \
    --prefill http://127.0.0.1:30000 --decode http://127.0.0.1:30010 \
    --host 0.0.0.0 --port 8000

Send the very first request with disagg_prefill_dp_rank set — before the fix, this fails with the "healthy before" error; after the fix, it succeeds:

curl -X POST http://127.0.0.1:8000/generate \
    -H 'Content-Type: application/json' \
    -d '{
        "text": "Hello World How are you?",
        "sampling_params": {"max_new_tokens": 128, "temperature": 0.0},
        "stream": false,
        "routed_dp_rank": 0,
        "disagg_prefill_dp_rank": 0
    }'

Checklist

• Format your code according to the Code Formatting with Pre-Commit.
• Add unit tests as outlined in the Running Unit Tests.
• Update documentation / benchmarks / examples as needed.
• Ensure all CI checks pass before submission.
• For adding new models / features, please use slash commands.