Fix PD bootstrap failure handling

[yhyang201]

Summary

• Set self.bootstrap_infos = None on bootstrap info fetch failure so downstream code hits the None-check instead of AttributeError
• Skip update_status(WaitingForInput) when _setup_bootstrap_infos already marked the request as Failed

Ported from 48135b2

Test plan

• PD disaggregation with bootstrap failure (e.g. prefill server down)

[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!

[yhyang201]

/tag-and-rerun-ci

[ShangmingCai]

Not sure about this. If routed_dp_rank is not None, we should respect it?

If this would be assigned when the strategy is follow_bootstrap_room, then maybe we should add this in the condition as well, something like:

Suggested change

	# In PD-prefill mode the cross-engine contract is `bootstrap_room`:
	# the decode-side KV receiver locates the prefill DP rank via
	# `bootstrap_room % prefill_dp_size`. Honoring an externally-set
	# `routed_dp_rank` here breaks that contract whenever the two
	# diverge (e.g., dynamo's KV router picks a rank for load-balance
	# reasons that has no relation to `bootstrap_room`). Fall through
	# to follow_bootstrap_room dispatch to keep prefill ↔ decode aligned.
	if (
	self.server_args.disaggregation_mode == "prefill"
	and req.bootstrap_room is not None
	):
	return False
	# In PD-prefill mode the cross-engine contract is `bootstrap_room`:
	# the decode-side KV receiver locates the prefill DP rank via
	# `bootstrap_room % prefill_dp_size`. Honoring an externally-set
	# `routed_dp_rank` here breaks that contract whenever the two
	# diverge (e.g., dynamo's KV router picks a rank for load-balance
	# reasons that has no relation to `bootstrap_room`). Fall through
	# to follow_bootstrap_room dispatch to keep prefill ↔ decode aligned.
	if (
	self.server_args.disaggregation_mode == "prefill"
	and req.bootstrap_room is not None
	and self.load_balance_method = "follow_bootstrap_room"
	):
	return False

I am not sure about this, should ping liangsheng for this.

[yhyang201]

Thanks for the review! You're right — this is already handled better on main via #23882, so I've reverted it.

[ShangmingCai]

The common backend modification looks good.

[ShangmingCai]

LGTM

[ShangmingCai]

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

[github-actions]

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

[ShangmingCai]

/rerun-test test/registered/disaggregation/test_disaggregation_basic.py

[github-actions]

✅ 2-gpu-h100 (1 test): View workflow run

cd test/ && python3 registered/disaggregation/test_disaggregation_basic.py

[ShangmingCai]

No need to run full CI, we only need these two, which should be enough.