The per-token Python loops that built the sparse {token_id: weight} dict and
the colbert vector lists held the GIL between forward passes, starving the GPU
(steady ~10-36% util with brief bursts). Replace them with GPU tensor ops:

- sparse: relu(sparse_linear·h) with invalid positions (specials/padding)
  zeroed, then max-pooled per token id via a single scatter_reduce(amax) into a
  (B, vocab) matrix; per-row nonzero extraction is O(nnz), not O(L).
- colbert: one C-level tolist per row over its valid tokens (vs per-token).
- valid mask computed on-device via torch.isin.

Output byte-matches the previous per-token logic (dense/sparse/colbert within
1e-4). Benefit grows with batch size, so larger client batches translate to GPU
work instead of Python-loop time.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Mw2KQiswmD69T45fTfjKwW