[Bugfix][Core] Fix request-bound KV cache sizing

[lesj0610]

Purpose

Fix KV cache sizing for hybrid models that combine token-proportional attention KV with request-bound KV groups such as Mamba. The startup sizing path now accounts for the fixed request-bound blocks consumed from the shared BlockPool, and the auto-fit path reserves the BlockPool null block so derived capacity matches runtime admission.

The implementation keeps request-bound groups in the shared pool. That lets inactive request slots remain available to token KV instead of reserving a dedicated request-constant pool that this fix does not need. The PR also removes the inactive dedicated-pool fallback so the changed code path is the one that is actually exercised.

AI assistance: Codex was used during implementation and PR preparation; the submitter reviewed the changes.

Test Plan

• Run syntax, lint, and format checks for the modified KV cache utility.
• Run the KV cache utility and single-type KV cache manager tests that cover shared-pool request-bound sizing, CPU-offload variants, and null-block capacity accounting.
• Check the patch for whitespace errors.

Test Result

Syntax, lint, format, and whitespace checks:

.venv/bin/python -m py_compile vllm/v1/core/kv_cache_utils.py
.venv/bin/python -m ruff check vllm/v1/core/kv_cache_utils.py
.venv/bin/python -m ruff format --check vllm/v1/core/kv_cache_utils.py
git diff --check

Result: passed.

Targeted tests:

.venv/bin/python -m pytest \
  tests/v1/core/test_kv_cache_utils.py \
  tests/v1/core/test_single_type_kv_cache_manager.py \
  -q

Result: 74 passed.

Commit hooks also passed for the latest cleanup commit, including ruff, format, mypy, SPDX, config validation, and related repository checks.

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results.
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model. Not needed; this is an internal KV cache sizing fix.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.