arXiv subjects

Lukas Stepanek

Publications and source records attributed to Lukas Stepanek.

4 recordsLinked to original sources

arXiv2026

Memory-Sovereign Inference: Output-Exact Execution Beyond Full Residency

Storage-backed inference is easy to overclaim: process RSS excludes charged page cache, process-local device readings exclude board-wide use, and successful generation does not establish correct asynchronous reuse. We present a falsifiable certificate separating representation, semantic demand, scheduler requests, and traffic while naming resource authorities, exactness horizons, and tested reuse transitions. At one Qwen3-Next identity, the stock router selects all 48 x 512 managed layer-expert objects during 32K prefill. Their duplicate-free, overlap-free canonical union gives a 43.59375 GiB semantic-demand lower bound, exceeding the declared 34 GiB full-residency envelope and the 11 GiB host-hard plus physical 24 GiB-device envelope. LRU64 execution stays within its host-hard/GPU-audited contract. Against one prespecified zero-cache oracle, all 64 token IDs, every byte of 64 complete 151,936-float logit rows, 3,408 route events, response bytes, and recorded consumer and destination identities are exact. Recurrent-state and upstream-runtime equality are excluded. In a matched source campaign, the buffered path completes exactly but reaches the 11 GiB host ceiling and records 33,481 memory.max events. The blocking one-window direct path and complete eight-window asynchronous component are exact with positive margin and zero limit events. Across six counterbalanced pairs, the complete asynchronous component takes 32.3% of the blocking direct path's wall time at identical physical source bytes per output; the comparison jointly changes queue depth, overlap, and lifecycle implementation. Separately, fourteen prespecified control/fault cells pass across later Qwen3-Next and Gemma 4 binaries, supporting fail-closed behavior only for the named transitions. The principal experiment is one fixed model, workload, runtime, device, and 64-output horizon.

cs.DC↗

arXiv2026

Route-Block Membership Selects Packed-AWQ Arithmetic: A Controlled Single-Fixture Mechanism Study

Mixture-of-experts (MoE) inference first aligns routed tokens into padded expert blocks, then executes packed quantized matrix multiplication over those blocks. This preprocessing is often treated as bookkeeping. In one pre-specified Qwen3-Coder AWQ layer-6 fixture on a pinned vLLM/Marlin build and RTX 3090 runtime, we show that the tested route-block interventions select exact packed arithmetic trajectories. Two fixed preconstruction histories produced distinct native alignments and exact trajectories. Injecting the opposite alignment transferred W13, activation, routed-W2, and final outputs. Permuting two routes within one block preserved each native trajectory, while exchanging two prior-data-selected routes across the boundary between expert-106 blocks 40 and 41 transferred the complete opposite trajectory. Source- and binary-derived schedule geometry maps those blocks to direct/full-K and split/global-reduction classes. Forcing a single-slice 200-block grid made W13 bitwise equal. Stable canonical construction made both histories converge to a third exact trajectory. The confirmatory cohort contains 70 valid cold processes and seven required perturbation rejections. This is a causal mechanism result for one fixture, not a prevalence, allocator, portability, or serving-impact claim.

cs.DC↗

arXiv2026

Fail-Closed Lowering of Resident KV Claims onto LLM Serving Runtimes

LLM serving runtimes increasingly expose KV-cache primitives that resemble future-reuse controls: retention priority, TTL-like duration, host or storage offload, block events, active no-evict scheduling, and KV-aware routing. This paper argues that such primitives are weaker than accepted future-KV obligations. A runtime can expose priority, offload, events, and routing without accepting responsibility for a future reuse claim. We study ResidentClaim lowering: when a runtime primitive, trusted adapter, or patch can be treated as satisfying an accepted claim about future KV reuse. A conformant lowering must bind behavior to accepted claim identity, a materialization predicate, ordered lifecycle events, and claim-scoped outcomes. We contribute a fail-closed lowering relation, checker, descriptor format, and bad-lowering suite that classify runtime/mode mappings as native conformance, adapter-observational evidence, adapter-policy evidence under controlled pressure, approximation substrate, rejected mapping, or unknown evidence. The checker validates manually curated, anchored runtime descriptors against obligation bundles; it does not prove that unaudited runtime behavior is complete. Public TensorRT-LLM, SGLang/HiCache, and Dynamo expose strong substrates and selected adapter positives, but not native ResidentClaim conformance. The positive systems witness is a local patched vLLM connector/scheduler-boundary mechanism: claim metadata flows through real in-process offload/load behavior, and controlled same-claim restoration failure reaches vLLM's invalid-KV-load path and becomes an ordered claim-scoped fail-closed outcome. The result is a calibrated semantics boundary, not a production performance claim or a compatibility survey.

cs.DC↗

arXiv2026

Resident KV Claims: A Conformance Contract for Future Reuse under Active KV Pressure

KV-cache reuse mechanisms increasingly expose priority, duration, offload, routing hints, scheduler modes, and event streams. These mechanisms help preserve reusable prefixes, but they do not by themselves define a portable contract for accepted future-reuse state when resident KV and active live KV cannot both fit. We introduce resident KV claims, a conformance contract that binds future-reuse intent to a materialization predicate, lifecycle state, active/resident feasibility outcome, and claim-level telemetry. In controlled vLLM allocator probes, a 60-block resident claim and a 70-block active prefill exceed an 80-block usable KV pool. Write no-admit prevents the active request from becoming future reusable state, but it still allows active allocation to evict residents from the shared pool. A minimal vLLM prototype shows that hard protected resident claims convert this failure mode into scheduler-visible active refusal with direct blocking-claim attribution. The result is not a production speedup or a new cache-replacement algorithm. It is a runtime contract that turns unreported resident loss into reconstructable active/resident arbitration. A companion MicroRuntime and vLLM litmus suite distinguish ordinary eviction, soft priority, write no-admit, accepted hard claims, materialization failure, demotion, expiry, active refusal, and trace-level outcome reconstruction.

cs.DC↗