fix(test): update vllm_generate_endpoint to post-#68 launch_server_process API

[aoshen02]

Problem

tests/test_vllm_generate_endpoint.py raised at runtime:

TypeError: launch_server_process() got an unexpected keyword argument 'bind_host'

The test still called the pre-#68 multi-kwarg form
launch_server_process(bind_host=, server_port=, args=, rank=, visible_devices=, model_path=).
PR #68 (multi-node rollout topology) refactored launch_server_process to take a
single server_args dict produced by _compute_server_args(...) — see
VLLMEngine itself at vime/backends/vllm_utils/vllm_engine.py (_compute_server_args(...) → launch_server_process(self._server_args)).

So this test has been failing since #68 and went unnoticed: the "Unit & utils"
job is gated behind pre-commit, which currently fails on every PR, so the job
is skipped. (Surfaced while running the full CI matrix for #121 on gb200; the
failure is not caused by #121 — the bind_host line predates it, from #16.)

Fix

Test-only change — mirror how VLLMEngine launches the server:

• build server_args via _compute_server_args(args, rank=0, dist_init_addr=None, host="127.0.0.1", port=server_port, base_gpu_id=0), then launch_server_process(server_args).
• base_gpu_id=0 pins this single-engine rank-0 server to the first visible GPU, sidestepping the rollout-placement math in get_base_gpu_id (which needs a full actor/critic placement Namespace).
• expand the args Namespace with the attrs _compute_server_args / build_vllm_cmd_and_env read (num_gpus_per_node, hf_checkpoint, vllm_enable_sleep_mode, vllm_dp_size).
• drop the now-unused _visible_devices helper.

No production code change (deliberately not re-adding bind_host to launch_server_process, which would regress the #68 dict API).

Verification

test_qwen3_0_6b_vllm_inference_generate_endpoint now passes on a single GPU:

1 passed, 2 warnings in 89.49s

(Qwen3-0.6B, real vllm serve subprocess + /inference/v1/generate rollout path, arm64 sync1916 image.)

🤖 Generated with Claude Code

[gemini-code-assist]

Code Review

This pull request refactors tests/test_vllm_generate_endpoint.py to use _compute_server_args for constructing server arguments instead of the removed _visible_devices helper, aligning with a multi-node topology refactor. Feedback points out that this change can cause CUDA initialization failures in multi-GPU environments when CUDA_VISIBLE_DEVICES contains non-consecutive or non-zero indices, and suggests explicitly overriding server_args["visible_devices"] to restore the correct physical GPU mapping.

[gemini-code-assist]

By removing the _visible_devices helper and relying solely on _compute_server_args, the test inherits a limitation in _compute_server_args (specifically in _to_local_gpu_id). When CUDA_VISIBLE_DEVICES is set to non-zero/non-consecutive indices (e.g., "2,3" in a multi-GPU or containerized environment), _to_local_gpu_id maps the GPU ID to a local index (e.g., 0), resulting in server_args["visible_devices"] being set to "0". When the vllm serve subprocess is launched, its CUDA_VISIBLE_DEVICES is set to "0", causing it to attempt to run on physical GPU 0 instead of the allocated physical GPU 2. This can lead to CUDA initialization failures or resource conflicts in multi-GPU environments. To prevent this, we should explicitly override server_args["visible_devices"] with the correct physical GPU IDs from CUDA_VISIBLE_DEVICES.

    server_args = _compute_server_args(
        args,
        rank=0,
        dist_init_addr=None,
        host="127.0.0.1",
        port=server_port,
        base_gpu_id=0,
    )
    cvd = os.environ.get("CUDA_VISIBLE_DEVICES")
    if cvd:
        server_args["visible_devices"] = ",".join(cvd.split(",")[:case.num_gpus])
    else:
        server_args["visible_devices"] = ",".join(str(i) for i in range(case.num_gpus))
    process = launch_server_process(server_args)

[CalvinXKY]

LGTM