diff --git a/tests/unit/models/vllm/test_vllm_model.py b/tests/unit/models/vllm/test_vllm_model.py
index a2c68837c..e40572014 100644
--- a/tests/unit/models/vllm/test_vllm_model.py
+++ b/tests/unit/models/vllm/test_vllm_model.py
@@ -20,14 +20,19 @@
 # OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
 # SOFTWARE.
 
+import importlib
 import sys
+import types
 import unittest
 from types import ModuleType
-from unittest.mock import Mock, patch
+from unittest.mock import MagicMock, Mock, patch
 
+import pytest
 from transformers import AutoTokenizer
 
+from lighteval.models.model_input import GenerationParameters
 from lighteval.models.vllm.vllm_model import VLLMModel, VLLMModelConfig, build_vllm_token_prompts
+from lighteval.tasks.requests import Doc
 
 
 class TestVLLMPromptConstruction(unittest.TestCase):
@@ -84,3 +89,342 @@ def test_vllm_model_use_chat_template_with_different_model_names(self, mock_crea
 
                 self.assertEqual(model.use_chat_template, expected_result)
                 self.assertEqual(model.use_chat_template, model._tokenizer.chat_template is not None)
+
+
+# ---------------------------------------------------------------------------
+# Unit tests for the pure-Python logic of ``VLLMModel`` (issue #724).
+#
+# ``vllm`` is a heavy GPU dependency that is not available on CPU CI runners, and
+# ``VLLMModel`` is gated behind ``@requires("vllm")`` (it becomes a placeholder
+# that raises on instantiation when vllm is missing). To exercise the model's own
+# logic - the boundary with vllm's ``SamplingParams`` and the loglikelihood
+# math - without a GPU, the ``vllm_model_module`` fixture below injects a light
+# fake ``vllm`` package and reloads the module so the real class is available and
+# bound to a recording ``SamplingParams`` stand-in. Instances are built with
+# ``__new__`` so that ``__init__`` (which would spin up a real engine) is skipped
+# and only the attributes each method needs are set.
+# ---------------------------------------------------------------------------
+
+
+class FakeSamplingParams:
+    """Records the parameters ``VLLMModel`` passes to vllm's ``SamplingParams``.
+
+    Only the attributes touched by ``VLLMModel._generate`` are modelled. The
+    constructor stores everything it is given, and the model is free to set the
+    remaining fields afterwards, exactly as it does with the real class.
+    """
+
+    def __init__(self, **kwargs):
+        self.n = 1
+        self.temperature = 1.0
+        self.max_tokens = None
+        self.stop = None
+        self.logprobs = None
+        self.prompt_logprobs = None
+        self.detokenize = True
+        self.__dict__.update(kwargs)
+        self.init_kwargs = dict(kwargs)
+
+
+def _fake_vllm_modules():
+    """Build a minimal fake ``vllm`` (and ``ray``) package tree for imports."""
+    fake_vllm = types.ModuleType("vllm")
+    fake_vllm.LLM = MagicMock(name="LLM")
+    fake_vllm.RequestOutput = MagicMock(name="RequestOutput")
+    fake_vllm.SamplingParams = FakeSamplingParams
+
+    fake_inputs = types.ModuleType("vllm.inputs")
+    fake_inputs.TokensPrompt = lambda *, prompt_token_ids: {"prompt_token_ids": prompt_token_ids}
+    fake_vllm.inputs = fake_inputs
+
+    fake_parallel = types.ModuleType("vllm.distributed.parallel_state")
+    fake_parallel.destroy_distributed_environment = MagicMock()
+    fake_parallel.destroy_model_parallel = MagicMock()
+    fake_distributed = types.ModuleType("vllm.distributed")
+    fake_distributed.parallel_state = fake_parallel
+
+    fake_async = types.ModuleType("vllm.v1.engine.async_llm")
+    fake_async.AsyncLLM = MagicMock()
+    fake_async.AsyncEngineArgs = MagicMock()
+
+    fake_tokenizers = types.ModuleType("vllm.tokenizers")
+    fake_tokenizers.get_tokenizer = MagicMock()
+
+    fake_ray = types.ModuleType("ray")
+    fake_ray.remote = lambda *args, **kwargs: lambda fn: fn
+    fake_ray.get = MagicMock()
+    fake_ray.shutdown = MagicMock()
+
+    modules = {
+        "vllm": fake_vllm,
+        "vllm.inputs": fake_inputs,
+        "vllm.distributed": fake_distributed,
+        "vllm.distributed.parallel_state": fake_parallel,
+        "vllm.v1": types.ModuleType("vllm.v1"),
+        "vllm.v1.engine": types.ModuleType("vllm.v1.engine"),
+        "vllm.v1.engine.async_llm": fake_async,
+        "vllm.tokenizers": fake_tokenizers,
+        "ray": fake_ray,
+    }
+    try:
+        import more_itertools  # noqa: F401
+    except ImportError:
+        fake_more_itertools = types.ModuleType("more_itertools")
+        fake_more_itertools.distribute = lambda n, iterable: [list(iterable)]
+        modules["more_itertools"] = fake_more_itertools
+    return modules
+
+
+@pytest.fixture
+def vllm_model_module():
+    """Reload ``vllm_model`` with a fake ``vllm`` backend so the real class loads.
+
+    This keeps the tests runnable on CPU while still exercising ``VLLMModel``'s
+    own code. When vllm *is* installed (GPU CI) the fakes simply shadow it for the
+    duration of the test; the module is reloaded again on teardown to restore the
+    original import state for the rest of the suite.
+
+    Yields:
+        module: The reloaded ``vllm_model`` module exposing the real ``VLLMModel``.
+    """
+    import lighteval.models.vllm.vllm_model as vllm_model
+    import lighteval.utils.imports as imports_mod
+
+    real_is_available = imports_mod.is_package_available
+
+    def fake_is_available(package):
+        # ``package`` may be the raw string or a parsed ``Requirement`` object.
+        name = getattr(package, "name", package)
+        if name == "vllm":
+            return True
+        return real_is_available(package)
+
+    with patch.dict(sys.modules, _fake_vllm_modules()):
+        with patch.object(imports_mod, "is_package_available", fake_is_available):
+            module = importlib.reload(vllm_model)
+            try:
+                yield module
+            finally:
+                pass
+    # Restore the original (un-faked) module for any other tests importing it.
+    importlib.reload(vllm_model)
+
+
+def _new_model(module):
+    """Create a ``VLLMModel`` without running ``__init__``."""
+    return module.VLLMModel.__new__(module.VLLMModel)
+
+
+class FakeLogprob:
+    def __init__(self, logprob, rank):
+        self.logprob = logprob
+        self.rank = rank
+
+
+class FakeGenerateOutput:
+    def __init__(self, prompt_token_ids, prompt_logprobs):
+        self.prompt_token_ids = prompt_token_ids
+        self.prompt_logprobs = prompt_logprobs
+
+
+class TestVLLMProperties:
+    def test_max_length_returns_configured_value(self, vllm_model_module):
+        model = _new_model(vllm_model_module)
+        model._max_length = 2048
+        assert model.max_length == 2048
+
+    def test_add_special_tokens_returns_configured_value(self, vllm_model_module):
+        model = _new_model(vllm_model_module)
+        model._add_special_tokens = True
+        assert model.add_special_tokens is True
+
+    def test_tokenizer_property_returns_backing_tokenizer(self, vllm_model_module):
+        model = _new_model(vllm_model_module)
+        sentinel = object()
+        model._tokenizer = sentinel
+        assert model.tokenizer is sentinel
+
+
+class TestVLLMGenerateSamplingParams:
+    """The vllm boundary that PR #721 showed can break silently."""
+
+    def test_scoring_path_uses_greedy_sampling_params(self, vllm_model_module):
+        model = _new_model(vllm_model_module)
+        model.data_parallel_size = 1
+        model.model = MagicMock()
+        model.model.generate.return_value = ["sentinel"]
+
+        outputs = model._generate(inputs=[[1, 2, 3]], generate=False)
+
+        assert outputs == ["sentinel"]
+        call = model.model.generate.call_args
+        sampling_params = call.kwargs["sampling_params"]
+        assert sampling_params.temperature == 0.0
+        assert sampling_params.prompt_logprobs == 1
+        assert sampling_params.max_tokens == 1
+        assert sampling_params.detokenize is False
+        # Token ids are wrapped through build_vllm_token_prompts.
+        assert call.kwargs["prompts"] == [{"prompt_token_ids": [1, 2, 3]}]
+        assert call.kwargs["use_tqdm"] is True
+
+    def test_generation_path_maps_sampling_fields(self, vllm_model_module):
+        model = _new_model(vllm_model_module)
+        model.config = vllm_model_module.VLLMModelConfig(model_name="gpt2")
+        model.data_parallel_size = 1
+        model.model = MagicMock()
+        model.model.generate.return_value = []
+
+        model._generate(
+            inputs=[[1, 2]],
+            max_new_tokens=32,
+            stop_tokens=["</s>"],
+            returns_logits=True,
+            num_samples=1,
+            generate=True,
+        )
+
+        sampling_params = model.model.generate.call_args.kwargs["sampling_params"]
+        assert sampling_params.n == 1
+        assert sampling_params.max_tokens == 32
+        assert sampling_params.stop == ["</s>"]
+        assert sampling_params.logprobs == 1
+
+    def test_generation_path_disables_logprobs_when_not_requested(self, vllm_model_module):
+        model = _new_model(vllm_model_module)
+        model.config = vllm_model_module.VLLMModelConfig(model_name="gpt2")
+        model.data_parallel_size = 1
+        model.model = MagicMock()
+        model.model.generate.return_value = []
+
+        model._generate(inputs=[[1, 2]], returns_logits=False, num_samples=1, generate=True)
+
+        sampling_params = model.model.generate.call_args.kwargs["sampling_params"]
+        assert sampling_params.logprobs == 0
+
+    def test_multiple_samples_with_greedy_temperature_raises(self, vllm_model_module):
+        model = _new_model(vllm_model_module)
+        # Default generation parameters use temperature=0 (greedy).
+        model.config = vllm_model_module.VLLMModelConfig(model_name="gpt2")
+        model.data_parallel_size = 1
+        model.model = MagicMock()
+
+        with pytest.raises(ValueError):
+            model._generate(inputs=[[1, 2]], num_samples=2, generate=True)
+
+    def test_multiple_samples_allowed_with_positive_temperature(self, vllm_model_module):
+        model = _new_model(vllm_model_module)
+        model.config = vllm_model_module.VLLMModelConfig(
+            model_name="gpt2",
+            generation_parameters=GenerationParameters(temperature=0.7),
+        )
+        model.data_parallel_size = 1
+        model.model = MagicMock()
+        model.model.generate.return_value = []
+
+        model._generate(inputs=[[1, 2]], num_samples=3, generate=True)
+
+        sampling_params = model.model.generate.call_args.kwargs["sampling_params"]
+        assert sampling_params.n == 3
+        assert sampling_params.temperature == 0.7
+
+
+class TestVLLMCreateAutoModel:
+    def test_model_args_wiring_for_single_gpu(self, vllm_model_module):
+        model = _new_model(vllm_model_module)
+        model._max_length = 4096
+        config = vllm_model_module.VLLMModelConfig(model_name="gpt2", max_model_length=4096, seed=99)
+
+        returned = model._create_auto_model(config)
+
+        assert returned is not None  # LLM(**model_args) was constructed
+        args = model.model_args
+        assert args["model"] == "gpt2"
+        assert args["max_model_len"] == 4096
+        assert args["seed"] == 99
+        assert args["enforce_eager"] is True
+        assert args["tensor_parallel_size"] == 1
+        # Optional keys are only present when explicitly configured.
+        assert "quantization" not in args
+        assert "load_format" not in args
+
+    def test_optional_quantization_and_load_format_are_forwarded(self, vllm_model_module):
+        model = _new_model(vllm_model_module)
+        model._max_length = 4096
+        config = vllm_model_module.VLLMModelConfig(
+            model_name="gpt2",
+            max_model_length=4096,
+            quantization="fp8",
+            load_format="safetensors",
+        )
+
+        model._create_auto_model(config)
+
+        assert model.model_args["quantization"] == "fp8"
+        assert model.model_args["load_format"] == "safetensors"
+
+    def test_subfolder_is_appended_to_revision(self, vllm_model_module):
+        model = _new_model(vllm_model_module)
+        model._max_length = 4096
+        config = vllm_model_module.VLLMModelConfig(
+            model_name="gpt2", max_model_length=4096, revision="abc", subfolder="sub"
+        )
+
+        model._create_auto_model(config)
+
+        assert model.model_args["revision"] == "abc/sub"
+
+    def test_data_parallel_defers_model_creation(self, vllm_model_module):
+        model = _new_model(vllm_model_module)
+        model._max_length = 4096
+        config = vllm_model_module.VLLMModelConfig(model_name="gpt2", max_model_length=4096, data_parallel_size=2)
+
+        returned = model._create_auto_model(config)
+
+        # With data parallelism the model is built later by ray, so None is returned.
+        assert returned is None
+        assert model.model_args["distributed_executor_backend"] == "ray"
+        assert model._batch_size == "auto"
+
+
+class TestVLLMLoglikelihoodTokens:
+    def test_continuation_logprobs_and_argmax(self, vllm_model_module):
+        model = _new_model(vllm_model_module)
+        model.pairwise_tokenization = False
+        model.prompt_manager = MagicMock()
+        model.prompt_manager.prepare_prompt.return_value = "CTX"
+        # Context/continuation token ids for the two choices of the single doc.
+        model.tok_encode_pair = MagicMock(return_value=([[10, 11], [10, 11]], [[20], [21, 22]]))
+
+        # First choice: continuation [20] scored as top-1 (rank 1) -> logprob -0.5.
+        # Second choice: continuation [21, 22]; 22 is not top-1 -> argmax False, sum -3.0.
+        out0 = FakeGenerateOutput([10, 11, 20], [None, None, {20: FakeLogprob(-0.5, 1)}])
+        out1 = FakeGenerateOutput(
+            [10, 11, 21, 22],
+            [None, None, {21: FakeLogprob(-1.0, 1)}, {22: FakeLogprob(-2.0, 3)}],
+        )
+        model._generate = MagicMock(return_value=[out0, out1])
+
+        docs = [Doc(query="q", choices=["a", "bb"], gold_index=0)]
+        responses = model._loglikelihood_tokens(docs)
+
+        # Scoring must go through the greedy (generate=False) path.
+        assert model._generate.call_args.kwargs["generate"] is False
+        assert model._generate.call_args.args[0] == [[10, 11, 20], [10, 11, 21, 22]]
+
+        assert len(responses) == 1
+        response = responses[0]
+        assert response.logprobs == [-0.5, -3.0]
+        assert response.argmax_logits_eq_gold == [True, False]
+        assert response.output_tokens == [[20], [21, 22]]
+        assert response.input_tokens == [[10, 11], [10, 11]]
+        assert response.input == "CTX"
+
+
+class TestVLLMTokenPromptsBatching:
+    def test_build_vllm_token_prompts_preserves_order_and_batch(self, vllm_model_module):
+        prompts = vllm_model_module.build_vllm_token_prompts([[5, 6], [7], []])
+        assert prompts == [
+            {"prompt_token_ids": [5, 6]},
+            {"prompt_token_ids": [7]},
+            {"prompt_token_ids": []},
+        ]