[PD-Disagg] Fix bootstrap server race condition when prefill workers not yet registered

[hnyls2002]

Summary

Bootstrap server

• Add _is_ready() check: track registered worker count, return 503 until all dp_size * tp_size * pp_size workers have registered
• Guard both metadata query (-1,-1,-1) and specific worker query with _is_ready()
• Catch KeyError on prefill_port_table lookup instead of crashing the server
• Remove unreachable None check after dict lookup (dead code)

Decode client

• Add retry logic in ensure_parallel_info (20 retries, 1s interval) so decode waits for prefill registration instead of failing immediately

Error Log

From CI failure:

File ".../sglang/srt/disaggregation/common/conn.py", line 675, in _handle_route_get
    info = PrefillServerInfo(
File ".../sglang/srt/disaggregation/common/conn.py", line 56, in __post_init__
    self.attn_tp_size = int(self.attn_tp_size)
TypeError: int() argument must be a string, a bytes-like object or a real number, not 'NoneType'

Root Cause

CommonKVBootstrapServer starts its HTTP server before any prefill worker has registered. Fields like attn_tp_size are None until a PUT arrives. A decode GET during this window crashes with TypeError in PrefillServerInfo.__post_init__.

This race always existed but was silently masked: previously _handle_route_get returned a plain dict (serializing None as JSON null), and the decode-side int(None) was swallowed by except Exception. After #19195 introduced PrefillServerInfo with __post_init__, the crash moved server-side and became visible.

Repro Patch

Delay register_to_bootstrap() to widen the race window:

def register_to_bootstrap(self):
    timer = threading.Timer(10, self._do_register_to_bootstrap)
    timer.daemon = True
    timer.start()
    return

def _do_register_to_bootstrap(self):
    # ... original body ...

Test plan

• test_disaggregation_basic.py passes
• test_disaggregation_dp_attention.py passes

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[ShangmingCai]

LGTM

[hnyls2002]

/rerun-stage stage-c-test-8-gpu-h20

[hnyls2002]

/rerun-stage stage-b-test-large-2-gpu

[github-actions]

✅ Triggered stage-c-test-8-gpu-h20 to run independently (skipping dependencies).

[github-actions]

✅ Triggered stage-b-test-large-2-gpu to run independently (skipping dependencies).

[github-actions]

🔗 View workflow run

[github-actions]

🔗 View workflow run

[hnyls2002]

🔗 View workflow run

PD-Disaggregation tests passed.