[MengqingCao]

Summary

Split out of #3173 on review: these are P-side metrics and do not depend on decode admission. The PR sits on top of the stack only for ordering; it touches none of the admission code.

Every scheduler exports two gauges, summed across DP ranks:

• atom:requests_deferred_free: finished or aborted requests kept allocated as a whole until every connector releases them.
• atom:kv_cache_blocks_deferred_free: the physical KV blocks those requests hold, shared prefix blocks counted once.

On P/D prefill this is the source KV decode has not pulled yet. An offload save that defers the whole request is counted the same way. Early-release offload leases are not included: those blocks return once every rank has finished reading them, before the store completes.

Both are computed in EngineUtilityHandler.collect_metrics at snapshot time, not per step. physical_block_count marks a boolean mask over the pool from each block table's buffer instead of building a Python set of every id. collect_metrics reads scheduler.deferred_free_blocks directly; Scheduler.__init__ always sets it, and RapidServe's PrefillScheduler has no block manager, so it never reaches this branch.

Both gauges are listed under "Engine aggregate state" in docs/metrics_guide.md.

Test plan

• CPU, at this head: tests/test_pd*.py, tests/test_mooncake*.py, tests/test_kv_connector*.py, tests/test_disagg*.py, tests/test_model_runner*.py, tests/test_scheduler*.py, tests/test_engine_core*.py, tests/entrypoints/test_metrics.py, tests/test_gpu_metrics.py, tests/test_observability_report.py: 1320 passed, 22 skipped. New here: test_deferred_free_counts_shared_blocks_once[list] and [new_block_table], the second over the array('i') a real block table is.
• Black and Ruff on the changed Python files.

---

_{Stack created with GitHub Stacks CLI • Give Feedback 💬}

[github-actions]

🏷️ CI Guide

Runs automatically on every eligible PR before approval:

• ✅ Pre Checkin: Black, Ruff, catalog schema validation, non-GPU unit tests

Heavy model tests:

• ✅ Run after the PR is approved and Pre Checkin passes
• ✅ Run immediately when an approval review is submitted
• ✅ Can be requested before approval with labels

Label	Tests
ci:full	Run all heavy PR model tests: native ATOM, vLLM, and SGLang
ci:atom	Run native ATOM model accuracy tests
ci:vllm	Run ATOM vLLM OOT model accuracy tests
ci:sglang	Run ATOM SGLang model accuracy tests

Heavy jobs are skipped when the PR is not approved and no matching ci:* label is present.
Add labels via the sidebar or gh pr edit 3238 --add-label <label>

[MengqingCao]

@valar-cop

[valar-cop]

@MengqingCao LGTM, reading each table into one mask over the pool is the right call instead of a python set. one thing: did you see it move on a real pd pair, prefill holding source kv until the send completes, or only the cpu tests? hold before pre-checkin and atom test pass.

[valar-cop]

LGTM

ci green.