fix: proxy get_node_url raises TypeError when no node matches role

[Evolian-o]

Summary

NodeManager.get_node_url crashes with TypeError: cannot unpack non-iterable NoneType object when no node matches the requested model + role, returning HTTP 500 instead of an "unavailable model" response.

Root cause

get_matched_urls() (nested in get_node_url) returns None when there is no matching node:

all_matched_urls = urls_with_speeds + urls_without_speeds
if len(all_matched_urls) == 0:
    return None

But both the RANDOM and MIN_EXPECTED_LATENCY branches unpack the return value directly:

all_matched_urls, all_the_speeds = get_matched_urls()

so the None raises TypeError before the if len(all_matched_urls) == 0: return None guard is ever reached (it is dead code).

When this happens

In DistServe (PD-disaggregation) mode, a model can be registered on a node of one role while the engine of the requested role is unavailable (e.g. the Prefill engine is down or removed by the heartbeat check). check_request_model still passes because the model is in the global model list, but get_node_url(model, EngineRole.Prefill) finds no Prefill node and crashes instead of returning an "unavailable model" response.

Fix

Make get_matched_urls always return a (urls, speeds) pair (an empty pair instead of None), so the existing len(...) == 0 guard becomes reachable and get_node_url returns None as intended.

Tests

Added tests/test_lmdeploy/serve/test_proxy.py covering all three routing strategies for both "no node matches" (returns None) and "node matches" (returns the URL).