fix(mocker): use logical KV tokens for decode timing [DYN-3851]

[PeaBrane]

Summary

• Align Mocker decode timing with the scheduled logical KV-token semantics used by profiler-derived performance data.
• Stop using distinct physically resident blocks as a decode-compute proxy, which undercounts heterogeneous batches when prefix caching deduplicates shared blocks.
• Keep the full logical sum uncapped for interpolation while saturating only the legacy polynomial model's utilization ratio at 1.0.

This does not change KV allocation, prefix-cache ownership, routing, or direct AIC decode prediction.

Details

• vLLM ordinary and speculative decode now pass the summed context lengths of the requests scheduled for decode.
• SGLang uses the same uncapped logical-KV meaning.
• Removed reservation-size accessors that were only needed by the physical-block timing proxy.

Where should the reviewer start?

Start with lib/mocker/src/scheduler/vllm/core.rs for the semantic change and lib/mocker/src/common/perf_model.rs for the timing-model boundary.

Related Issues

• Confirmed — no related issue

Validation

• cargo test -p dynamo-mocker test_fpm_prefill_and_decode_mixed_batch -- --nocapture (vLLM and SGLang cases passed locally and on a clean Linux checkout)

• cargo test -p dynamo-mocker --features kvbm-offload speculative_decode_reservation_waits_without_preempting -- --nocapture

• cargo fmt --all -- --check, commit hooks, and git diff --check

• ComputeLab SC-01: single H100, Qwen3-1.7B, vLLM TP1, 38,863 runtime GPU KV blocks, four prefix groups at 68.68-74.85% measured prefix-cache token hit rate, three trials each at concurrency 8 and 128, zero request errors or output-length mismatches.

• At concurrency 128, candidate median ITL was 21.69 ms versus live 22.12 ms; baseline was 8.98 ms. Candidate output-throughput error was -19.6% versus baseline +40.8%.

• Paired trace at scheduler batch 106/context 4,103 showed candidate active_kv_tokens=434,972 (106.01 logical contexts) versus baseline 143,968 (35.09 contexts), confirming the corrected timing input tracks the scheduled batch under prefix dedup.

• Lower-level attribution on the exact live stack (H100 80GB, pinned vLLM 0.23.1 development image, Qwen3-1.7B BF16 TP1, FlashAttention 3) reproduced the scheduler shape at batch 106 with 434,976 logical KV tokens. At fixed batch and logical total, the actual ragged request vector was within 0.5% of the homogeneous attention control.

• The old batch-35 proxy underpriced direct paged FlashAttention by 58.2% and steady-state engine decode by 48.6% versus the actual batch-106 workload. This directly validates removing the dedup-aware scheduler cap.

• Physical prefix-block aliasing reduced attention time by 35.7% versus an unaliased replay. That is a separate attention-model fidelity opportunity; it does not make deduplicated physical blocks a valid proxy for the scheduled decode batch.

The live image reports vLLM 0.23.1rc1.dev1502; the available AIC interpolation data is for vLLM 0.24.0. Candidate mean and upper-tail simulated ITL remained conservative, so the campaign validates the accounting semantics and center-of-distribution parity rather than exact end-to-end calibration across those versions.

---

[devin-ai-integration]

Devin Review found 1 potential issue.

[coderabbitai]

Walkthrough

Changes

Decode modeling

Layer / File(s)	Summary
Scheduled logical token timing lib/mocker/src/scheduler/vllm/core.rs, lib/mocker/src/common/perf_model.rs	Regular and speculative decode timing now use scheduled sequence lengths capped by total KV capacity. Documentation defines these values as logical KV token counts.
Reservation API cleanup lib/mocker/src/kv_manager/g1_manager.rs, lib/mocker/src/kv_manager/vllm_backend.rs	Removed crate-visible len accessors from decode block reservation types.

Estimated code review effort: 2 (Simple) | ~10 minutes

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Title check	✅ Passed	The title clearly and concisely describes the main change: using logical KV tokens for Mocker decode timing.
Description check	✅ Passed	The description explains the change, affected files, review starting points, related-issue status, and validation results in sufficient detail.

---

_{Comment @coderabbitai help to get the list of available commands.}