Allow top_k greater than vocab_size

[Hrishith Thadicherla (hthadicherla)]

Allow top_k greater than vocab_size

Generator::InitializeSamplingMethod rejected any top_k above the model's
vocab_size.
This caused failure in these two test cases whose parameters violate this condition by default.

SamplingTests.BatchedSamplingTopPNvTensorRtRtx
SamplingTests.RandomizedSamplingTopPNvTensorRtRtx

C++ exception with description "top_k (50) must be less than or equal to vocab_size (5)"   // BatchedSamplingTopP
C++ exception with description "top_k (50) must be less than or equal to vocab_size (21)"  // RandomizedSamplingTopP

That is not an error condition: a top_k larger than the vocabulary
simply means "consider all tokens", and it is what a common default (e.g. 50) does
on a small-vocab model, which failed generator creation outright.

Current PR removes the throw. All three sampling paths already clamp k to vocab_size
before use, so nothing can read out of bounds:

• CPU: ComputeSampledCategorical - src/sampling_distribution.h:245

const int m = std::min(top_k, vocab_size);

• CUDA: cuda::GetSample - src/cuda/cuda_sampling.cu:369

if (k <= 0 || k > vocab_size) { k = vocab_size; }

• MTP device top-k: MtpGenerator::TopKScoresRows - src/mtp_generator.cpp:419

topk_k_ = std::min(top_k_, vocab_size_);

The top_k < 0 check is unchanged.

[Copilot]

Pull request overview

This PR adjusts generator sampling initialization to allow top_k values greater than a model’s vocab_size, aligning generator creation behavior with existing sampling implementations that already clamp k to vocab_size at use sites.

Changes:

• Removed the top_k > vocab_size runtime error in Generator::InitializeSamplingMethod.
• Preserved existing validation for invalid top_k (negative values) and existing sampling-method selection logic.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.