fix(tests): use bare --router-request-timeout-secs in glm4.7 mooncake test

[aoshen02]

What

One-line fix: tests/test_glm4.7_30B_A3B_pd_mooncake.py passed the pre-hybrid flag --vllm-router-request-timeout-secs 1200. Renamed to bare --router-request-timeout-secs 1200.

Why

The hybrid router-arg convention settled in #18 / #40:

• Endpoint-locating flags keep the vllm_ prefix: --vllm-router-ip, --vllm-router-port (vime owns where the router lives; RouterArgs excludes host/port).
• Every other vllm-router knob is bare --router-* (dest router_*), e.g. --router-request-timeout-secs, --router-policy.

This GPU-only integration test still used the old --vllm-router-request-timeout-secs, which is no longer a registered flag. The 1200s timeout therefore never reached router_args.request_timeout_secs (slime/ray/rollout.py:978 reads bare args.router_request_timeout_secs). Because the test is outside the unit suite, #40's rename sweep missed it.

Scope

• Single-line change in one GPU integration test.
• Verified no other stale --vllm-router-* non-ip/port flags remain in tests/.
• Unit suite (tests/unit/backends/vllm_utils/) stays green: 66 passed.

🤖 Generated with Claude Code

[gemini-code-assist]

Code Review

This pull request updates the test script tests/test_glm4.7_30B_A3B_pd_mooncake.py by renaming the command-line argument --vllm-router-request-timeout-secs to --router-request-timeout-secs. There are no review comments, and I have no feedback to provide.

[gcanlin]

I remember that we plan to just replace sglang by vllm. And slime is using --sglang-router-request-timeout-secs. See https://github.com/THUDM/slime/blob/7a7aba4eea36513e3c076594d3aa9aa6263ce558/slime/backends/sglang_utils/arguments.py#L25

[aoshen02]

I remember that we plan to just replace sglang by vllm. And slime is using --sglang-router-request-timeout-secs. See https://github.com/THUDM/slime/blob/7a7aba4eea36513e3c076594d3aa9aa6263ce558/slime/backends/sglang_utils/arguments.py#L25

So --router-request-timeout-secs is a genuine vllm-router knob. We register the router package's entire CLI surface (~30 args — policy, cache_threshold, retries, health-check, timeout, …) via RouterArgs.add_cli_args(parser, use_router_prefix=True), which namespaces them all under --router-* and reads them back from args.router_. So this flag's name and dest come straight from vllm-router. Renaming it --vllm-router- would desync it from the other ~30 router knobs and from the dest that RouterArgs.from_cli_args reads.

--vllm-router-ip / --vllm-router-port are different: they aren't vllm-router CLI args at all (we pass exclude_host_port=True) — they're vime's own "where do I reach the router" flags, so they live in vime's --vllm-* namespace.

So the rule is: --router-* = pass-through to the router package, --vllm-* = vime-owned. What do you think, I'm also consider renaming vllm-router-port and ip to router-port and ip as well.

[gcanlin]

I agree with remove some redundant vllm- prefix.