[backport] feat/bench_y reuse-budget fixes onto v1.3.0rc11 (issue #13318)

[yifjiang]

Draft PR — not for upstream merge. This is a downstream-only branch on the yifjiang fork that captures the cherry-pick state used by the dynamo-trtllm v4 production image overlay (head-health-cuda-v4-benchy-rc11). PR base is v1.3.0rc11-base (a fork-side branch pointing exactly at the upstream v1.3.0rc11 tag commit 4e69c14f7), so the diff shown is exactly the three cherry-picks.

Cherry-picks (in order)

Commit	Upstream PR	Files	Purpose
c0d22ad6c (was d9038d3b90)	NVIDIA#12682	4	Introduces _estimate_post_reuse_compute() + Python-side remaining_budget re-validation in prepare_resources (catches C++ scheduler over-admission before _prepare_tp_inputs). Also patches microBatchScheduler.cpp for the C++ admission math; we keep the C++ change in source for fidelity but the production overlay only ships the Python parts.
9faa1dd27 (was 01c4947990)	NVIDIA#12806	1	Pre-subtract non-first-chunk context costs from remaining_budget before first-chunk checks. Patches a bug introduced inside NVIDIA#12682's gate.
e11cedca5 (was f24dcb33e1)	NVIDIA#12909	1	Disagg: pass the filtered (post-prepare_resources) batch to _recv_disagg_gen_cache, fixing CacheReceiver::requestSync unordered_map::at when a request is skipped by the budget gate. Required for our disagg deployments (Qwen3-Coder-480B).

Authors preserved (Liao Lanyu, Jin Li, Lizhi Zhou). Committer rewritten to satisfy GitHub email-privacy on this fork; signed-off-by trailers from the original commits are intact.

Why this exists

Issue NVIDIA#13318 — the C++ scheduler under-counts last-chunk reuse cost, leading to assert total_num_tokens <= self.max_num_tokens killing the event loop. The proper upstream fix is on feat/bench_y (16 commits ahead of main, 125 commits behind v1.3.0rc11). We need the fix on v1.3.0rc11 now for the Qwen3-Coder-480B deployment that crashed three times on 2026-04-22. Cherry-picking the three relevant commits onto v1.3.0rc11 is the cleanest available path until feat/bench_y lands on main and rolls into a release wheel.

Why three commits and not the whole feat/bench_y

feat/bench_y has 16 unique commits; the other 13 are DSA perf work, prefix-reuse refactor, and tests that are unrelated to this bug or are scoped to features we do not use. Importing them would broaden the patch surface for no reason.

Cherry-pick base verification

v1.3.0rc11 tag commit:        4e69c14f732a6e6afce4f71616db5b5cd2b10530
merge-base HEAD v1.3.0rc11:   4e69c14f732a6e6afce4f71616db5b5cd2b10530

Branch tip is "v1.3.0rc11 + the three cherry-picks, nothing else."

Cherry-pick result

All three commits applied via 3-way auto-merge with no conflicts. Files changed (vs v1.3.0rc11):

cpp/tensorrt_llm/batch_manager/microBatchScheduler.cpp  +43 -32
tensorrt_llm/_torch/pyexecutor/model_engine.py          +18  -0
tensorrt_llm/_torch/pyexecutor/py_executor.py          +135 -15
tensorrt_llm/_torch/pyexecutor/resource_manager.py      +84  -4
                                              total:   +280 -51

Downstream consumer

This branch is consumed by docs/build_dynamo_trtllm_head_health_cuda_v4_benchy_rc11.md in the claude-local-orchestrator repo. The dynamo-trtllm Docker build clones this branch, copies the three patched Python files (the C++ change is in source for fidelity but is not built into the image — has_trtllm_context: "0"), and overlays them on the installed tensorrt-llm==1.3.0rc11 wheel.

[yifjiang]

Superseded by upstream draft PR NVIDIA#13359 — same branch, same content, opened against NVIDIA:main for CI signal and maintainer visibility. Branch v1.3.0rc11-benchy-subset remains on this fork as the canonical cherry-pick state.