test: port test_value_temperature from slime #1928 (mirror-slime gap)

[aoshen02]

Summary

slime #1928 landed both the critic-temperature fix (loss.py) and tests/test_value_temperature.py. vime #95 ported only the fix — the test was missed. This restores it (mirror-slime).

• tests/test_value_temperature.py: slime verbatim, only slime.→vime. (import + sys.modules paths). NUM_GPUS=0, 0 sglang refs. Passes locally (1 passed).
• Wired into the num_gpus:0 unit matrix in pr-test.yml + pr-test.yml.j2. vime CI runs explicit per-file matrix entries (not bare pytest), so an unlisted test would never run.

AI-assisted. Signoff item 3.

[gemini-code-assist]

Code Review

This pull request adds a new test file tests/test_value_temperature.py to verify that get_values does not apply rollout temperature. The review feedback suggests a more idiomatic and less error-prone approach to mock and restore sys.modules using pytest's monkeypatch.delitem instead of manual try...finally blocks.

[gemini-code-assist]

Instead of manually popping modules from sys.modules and restoring them in a try...finally block, you can use pytest's monkeypatch.delitem with raising=False. This is more idiomatic, less error-prone, and automatically handles the cleanup/restoration of sys.modules after the test completes.

    monkeypatch.delitem(sys.modules, "vime.backends.megatron_utils.loss", raising=False)
    monkeypatch.delitem(sys.modules, "vime.backends.megatron_utils.cp_utils", raising=False)

    mpu_stub = types.SimpleNamespace(
        get_context_parallel_world_size=lambda: 1,
        get_context_parallel_rank=lambda: 0,
    )
    megatron_mod = types.ModuleType("megatron")
    core_mod = types.ModuleType("megatron.core")
    core_mod.mpu = mpu_stub
    monkeypatch.setitem(sys.modules, "megatron", megatron_mod)
    monkeypatch.setitem(sys.modules, "megatron.core", core_mod)

    from vime.backends.megatron_utils.loss import get_values

    args = Namespace(qkv_format="thd", rollout_temperature=0.5, allgather_cp=False)
    logits = torch.tensor([[[1.0], [2.0], [3.0], [4.0]]], dtype=torch.float32)
    tokens = [torch.tensor([10, 11, 12, 13], dtype=torch.long)]

    _, result = get_values(
        logits,
        args=args,
        unconcat_tokens=tokens,
        total_lengths=[4],
        response_lengths=[2],
    )

    torch.testing.assert_close(result["values"][0], torch.tensor([2.0, 3.0]))

[CalvinXKY]

LGTM