perf: build KV slot mappings by page ranges

[T4t4KAU]

Summary

Build KV slot mappings one physical-page range at a time, and reuse the mapping helper for allocation and speculative append. Keep a dedicated single-token path. This removes repeated division/modulo for tokens within the same page.

Motivation

Long prefill and speculative intervals previously calculated each physical slot separately. The new implementation produces identical lists for partial pages and reordered physical blocks with fewer Python iterations.

Type of Change

• perf — performance improvement with unchanged behavior

Test Results of Involved Models on Supported Platforms

Local CPU validation covered boundary intervals, 2,000 seeded randomized intervals, cached-prefix allocation, speculative append/rollback, and allocation failure. These checks were performed during development; no unit-test or test-support files are included in the final PR diff. Repository formatting checks and git diff --check passed.

Full-model single-request, offline, sanity, and service tests were not run: this validation isolates Python cache metadata and has no local model weights. No accelerator throughput claim is made.

Benchmark / Performance Impact

Intel Core i5-14600KF, Python 3.12.13; Python integer metadata (dtype N/A). Baseline: 270feb3. Timed update_blocks_slot with reversed physical-page order and logical start offset 7; fixture creation excluded. Five warmup calls, median of seven rounds, max(10, 100000 // token_count) calls per round, perf_counter_ns; CPU not pinned. Outputs checked against baseline before timing. Temporary local harness, not committed.

Page size	Tokens	Before (µs/call)	After (µs/call)
16	128	3.719	1.703
16	4096	151.945	63.430
256	1	0.099	0.081
256	128	3.754	0.430
256	4096	153.170	26.265
256	32768	1258.954	190.769

These are CPU microbenchmarks, not end-to-end model results; gains depend on interval and page sizes.

Notes for Reviewers

Branch starts directly at 270feb3 on InfiniLM-v0.2.9c. Allocation, ownership, hash publication, and eviction behavior are unchanged.

CI / ChatOps

CI has not been triggered. Accelerator/model validation remains for maintainers with a configured environment.

Checklist

• Conventional Commits title and commit; matching perf/ branch; squashable commit history; no merge/fixup/WIP commits.
• Based directly on the requested InfiniLM-v0.2.9c release branch (instead of main).
• N/A — legacy issue-format exception.
• Minimal scope; no debug/dead code or unrelated formatting.
• Public signatures remain compatible.
• English comments/docstrings, trailing newlines, and whitespace checked.
• Python files checked with scripts/format.py; no changes to auto_config.py.
• N/A — C++ checks and native build: Python-only changes.
• Local correctness checks and before/after CPU benchmarks completed; omitted model/platform tests explained above.
• Additional platform validation and reviewer assignment.
• Manual CI run.
• N/A — user documentation and breaking-change notice: behavior and workflow unchanged.
• No secrets, new third-party code, or unsafe memory operations.