[bingdongni]

What changed

Keep a singleton chunk without scoring it. ChunkPress always retains at least one token per chunk, but a final one-token chunk currently reaches scorers such as TOVAPress, which fail when computing attention scores for an empty history. For example, a 9-token prefill with a chunk length of 4 fails on its last chunk.

The new tests cover a single-token input, one-token tails and a chunk length of 1 with TOVA, SnapKV and Knorm, checking both cache length and the retained key/value. They use a small randomly initialized Llama model.

Validation

• GitHub Actions run / redacted job log: 24 passed (12 CPU, 12 CUDA) on a self-hosted RTX 4060 Laptop GPU, PyTorch 2.6.0+cu124 and transformers 4.57.6.
• black --check, isort --check-only and flake8 on the two changed files.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.