[BlivionIaG]

Change

First bounded optimization pass after #146: submit missed experts' six persistent SoA payloads through one native C++ call, using HIP 7.14 hipMemcpyBatchAsync when available and a native per-copy loop otherwise. Update the existing pinned int32 slot map through a persistent zero-copy NumPy view to remove scalar Tensor dispatch. Preserve payload addresses, pool budgets, LFU/LRU semantics, leases and consumer/DMA fences.

Fix a reproducible GDN Q32 squared-sum overflow with fixed-order FP32 tile partials, without another kernel launch. Reject invalid accepted-state indices before reads, and extend the post-step fail-stop to the offloaded-expert workspace. Device diagnostics and quiescent inspection controls remain development-only.

Tensor copies remain the default. Enable the tested native path explicitly with rocm_mono_decode.expert_cache_copy="native". Keep the K2 gate at 8; wide-kernel work belongs to #110. This follows the merged core in #146 and does not duplicate either PR. Prefetch, slot rebalancing and further profiling remain later passes in the roadmap.

Validation

• HIP 7.14 / gfx1030: 147 focused host/GPU tests passed, including real CT packing, overflow and cross-stream parity for both backends. Client failure/cold-prefix gates: 28 passed. Failure-monitor CPU contracts: 14 passed. GDN diagnostic suite: 17 passed; default GDN regression/parity: 49 passed, four diagnostics-only skips. The tested native binary hash is retained.
• Full TP2/EP2 Flash-Next AWQ text serving, MTP2, segmented FULL_AND_PIECEWISE graphs, GPUs 6–7: 16 ABBA arms, 72 exact-length requests and 55,296 output tokens. Both backends passed greedy, perplexity (native 4.82 / tensor 4.83), prefix reuse and 48 normal-EOS concurrency checks each, with zero hard failures. Post-load greedy passed; no numeric or new GPU fault markers.
• Eager native-copy fallback (enforce_eager=True, graph mode NONE): greedy, PPL 4.84, prefix reuse, 48 normal-EOS concurrency checks with zero hard failures, exact-length 1k/512 and 16k/1k c1 workloads, and post-load greedy passed without numeric or new GPU fault markers.
• Local commit hooks passed. Hosted checks were skipped by the repository workflows; they are not represented as GPU validation.

Reproduction: GPU=7 bash tools/rdna/mono/test_offload.sh; .venv/bin/python -m pytest --noconftest tests/models/qwen4_exp/test_offload_long_probe.py; the exact graph/eager launch and load-hook scripts, raw results, pool/copy counters and source hashes are under bench_results/2026-10-29-moe-offload-first-pass/.

Measured scope and limits

• Matched actual CPU cache control-path benchmark, ten active experts: 83.42 → 37.73 µs median (30 ABBA samples per arm). This excludes GPU event/DMA time. Matched native-copy byte/order tests and submission/completion timings are retained in the earlier receipts.
• Cold 16k/c1 TTFT: tensor 17.10–17.66 s, native 16.06–16.12 s. Cold 16k/c8 aggregate output: tensor 24.17–25.10 tok/s, native 31.91–32.63 tok/s. Short c1 decode was mixed. All four cells' prefill/decode/TTFT/ITL ranges are in graph-performance-summary.txt and JSON.
• The A/B now resets both GPU and external CPU prefixes and rejects measured hits or missing hit counters. Expert histories, transferred bytes and MTP acceptance still differ; these model observations are not a causal or universal backend speedup claim. Defaults and memory budgets stay unchanged.
• An earlier pre-fix run failed at 810/1,024 tokens. The isolated GDN overflow is fixed, but that earlier serving failure has not been causally attributed. Its receipt remains preserved; later diagnostic replay and final graph/eager qualification passed without recurrence. Older-HIP fallback has no older-runtime qualification.

AI assistance was used. The change is limited to this first pass; the remaining optimization backlog is not a merge requirement for it.