Add unit tests for VLLMModel core methods (#724)

[sneha4175]

What does this PR do?

Adds unit tests for the VLLMModel class methods, as requested by @lewtun in #724 (unit tests to "catch subtle bugs like [#721] and also protect us a bit from silent, yet breaking changes on the vllm side").

The existing tests/unit/models/vllm/test_vllm_model.py only exercised tokenizer/chat-template wiring. This PR adds coverage for the pure-Python logic of the model, focusing on the boundary with vllm's SamplingParams (exactly the area #721 broke) and the loglikelihood math.

Methods now covered

• _generate — the vllm boundary:
  • greedy scoring path (generate=False): temperature=0, prompt_logprobs=1, max_tokens=1, detokenize=False, and prompts wrapped via build_vllm_token_prompts
  • generation path: n, max_tokens, stop, and logprobs mapping (incl. returns_logits on/off)
  • the num_samples > 1 + temperature == 0 guard raising ValueError, and that sampling is allowed with a positive temperature
• _create_auto_model — model-args wiring (max_model_len, seed casting, enforce_eager, tensor-parallel size), optional quantization/load_format forwarding, subfolder→revision handling, and data-parallel deferral (returns None, sets the ray executor backend)
• _loglikelihood_tokens — continuation logprob slicing, summation, and rank-1 argmax bookkeeping, plus the resulting ModelResponse fields
• max_length, add_special_tokens, tokenizer properties
• build_vllm_token_prompts batching/empty-input behaviour

How vllm is mocked

vllm is a heavy GPU dependency and VLLMModel is gated behind @requires("vllm") (it degrades to a placeholder when vllm is absent). To keep these tests CPU-runnable while still exercising the model's own code, a vllm_model_module fixture injects a lightweight fake vllm (and ray) package into sys.modules — including a recording SamplingParams stand-in — patches is_package_available and reloads the module so the real class loads, then reloads again on teardown to restore the original import state. Instances are built with __new__ to skip __init__ (no real engine), setting only the attributes each method needs. When vllm is installed (the dev-gpu CI env), the fakes simply shadow it for the duration of each test.

Test / lint results

• pytest tests/unit/models/vllm/test_vllm_model.py -k "Properties or GenerateSamplingParams or CreateAutoModel or LoglikelihoodTokens or TokenPromptsBatching" → 14 passed
• ruff check and ruff format --check → clean

Closes #724 (or fixes part of it — happy to adjust scope).

Before submitting

• Did you read the contributor guideline?
• Was this discussed/approved via a GitHub issue first? [FT] Add tests for VLLMModel base methods #724
• Did you make sure to update the documentation with your changes?
• Did you write any new necessary tests?