Unc Q4_K_M MTP draft head: export, speculative-rollback A/B, and the 200K validation incident

Plain-language summary: FINDINGS.md — what won, why MTP did not ship, what broke at midnight.

• Date: 2026-09-07 → 2026-09-08 (export + A/B + scheduled validation run)
• Host: 12700 — 3x RTX 3090 (72 GB aggregate VRAM), 20 CPUs, Ubuntu Linux
• Target model: orcarouter/Qwen3.8-Flash-Next-Uncensored-FP8 (abliterated 125B-A6B MoE) → production quant AD-4.27bpw Q4_K_M-M64-agentic-v2 (see sibling experiment)
• Draft model: MTP (NextN) head exported from the same checkpoint → mtp-Qwen38-Unc-Q8_0.gguf (4.1 GB, 34 tensors) and mtp-Qwen38-Unc-Q4_K_M.gguf (2.78 GB)
• Inference stack: llama.cpp flash-next fork at WORKSTATION/llama.cpp-flash-next (master 9723942 + local uncommitted patch), --repack --op-offload, -sm layer -ts 1,1,1
• Router: qwen36-multi-proxy on :12434, family qwen38-flash-next-unc-q4km-agentic-v2, backend :12702

Two questions answered: (1) can the Qwen3.8 “qwen4exp” MTP/NextN draft head be exported to GGUF and made to run in llama.cpp, and (2) does draft-MTP speculative decoding beat plain decode on this 3x3090 stack at production sampling. A third section documents what happened when the 200K-context validation was scheduled to run unattended at midnight.

1. Draft-head export (upstream PR #27836 + local fixes)

Upstream PR ggml-org/llama.cpp#27836 (“qwen4exp: add NextN/MTP draft head”) applied via git apply, plus fixes the PR was missing (all working-tree only, never committed):

Pipeline: selective 30-shard/32 GB download → convert_hf_to_gguf.py --mtp --outtype q8_0 (exit 0, 4.12 GiB; MTP-head tensors are BF16 in the FP8 repo, so no dequant-order hazard for the fused eh_proj) → llama-quantize --allow-requantize → Q4_K_M 2.78 GB. Smoke at ctx 32768: draft acceptance 0.615 (64/104), mean accepted len 2.83, coherent reply — draft genuinely active.

GGUF facts: arch qwen4exp, block_count 49, nextn_predict_layers=1, 512 experts, n_embd 2560; 34 tensors = token_embd + output + one full-attn + 512-expert MoE block under blk.48 + nextn.*.

2. Track 1: ported upstream #28123 (conv-state rollback) — the A/B that mattered

The first smoke ran at 21 t/s (2.7x slower than baseline). Root cause: recurrent-state rollback was missing for qwen4exp — every draft rejection discarded the whole sequence. Ported merged upstream #28123 (llm_arch_supports_rs_rollback + slotted conv-state store, the fork’s existing TAG_RECURRENT_ROLLBACK_SPLITS pattern). No-op when n_rs_seq=0; production no-draft path provably untouched (n_layer_nextn=0 → flags=0 everywhere).

A/B on 3x3090, ctx 32768, taskset-pinned, draft on CUDA1, median-of-5 (scripts/ab_*.sh, data/ab_*.json):

• Rollback proven engaged (“bounded partial sequence removal” at startup); stability 5/5 identical-prompt chats with prompt-cache restore — the exact path that aborts with on-device checkpoints (#28118 approach, rejected).
• Acceptance ~0.53–0.62 at both temps (upstream reports 0.79 at p-min 0.7 on the base model) — the Unc fine-tune / base-head mismatch caps gains.
• Box-noise lesson: temp-0.7 runs swing ±20% under CPU contention from other sessions; temp-0 runs are rock-stable (<1% spread). Medians mandatory; never compare across core sets.
• Earlier no-rollback comparison for reference (data/bench_v4_65k*.json): v4+MTP lost 20–48% vs v3 at ctx 65536 despite 60–75% acceptance.

3. Decision: keep v3 (no draft) on :12702

At the production sampling preset (temp 0.7) MTP is parity — no rollout on speed alone. At temp 0 (greedy/code/agentic) n=3 gives a stable +24% → rollout candidate only if greedy workloads matter. Both user presets are acceptance-hostile anyway: thinking runs at temp 1.0 (maximum entropy), and instruct’s presence 1.5 distorts the target distribution away from draftable argmax paths. data/DECIDE.json ranking (warmed mean gen t/s): v3_65k 56.9 > v3_200k 55.6 (shipped) > mtp_v4_65k 39.2 > other MTP variants 35.8–38.5.

4. The 200K validation that never ran — midnight incident post-mortem

out/PLAN_CTX200K.md specified a 4-phase validation (fit ladder 204800→65536 with a 1.5 GiB/GPU floor, speed A/B under the two user presets, cyber suite, 198K qualification) with trap-guaranteed production restore, scheduled one-shot via cron 0 0 8 9 * (2026-09-08 00:00 NZST). The cron fired at 00:00:01 exactly; the failure was environmental, and the runner caught it correctly:

Post-mortem findings (runner hardening lessons):

1. Glob-based MODEL_PATH is fragile — a directory with 2 GGUF candidates (weights + mmproj) must resolve to the entrypoint explicitly, else fail fast, not after 30 min of blind waiting.
2. Health waits should poll with early-exit on dead child processes, not sleep the full window.
3. Preflight was the hero: validating artifacts before stopping production is what prevented a 2.5 h wasted downtime window — but the draft-file check ran before checking production was already down; ordering preflight checks by blast radius matters.
4. Duplicate log lines (tee + cron redirect) are cosmetic but double every diagnostic.
5. Shared-box reality: another actor’s disk cleanup can invalidate a scheduled job’s inputs hours before it fires. Scheduled jobs should re-verify inputs and re-check assumptions at fire time.

Resolution: production was restored externally the next day — :12702 healthy on the consolidated single-file GGUF + mmproj, with the sharp chat template (--chat-template-file), pid verified 2026-09-09. The Q4_K_M draft is gone (Q8_0 draft survives, 3.9 GB; speed-equivalent per DECIDE.json); re-export would need the 32 GB selective download again. The cron line remains (fires next on 2027-09-08; remove for hygiene).

5. Router deployment notes (same arc, worth keeping)

• Family-only exposure works via env on the systemd unit: MODEL_FAMILY_ALLOWLIST + DISABLED_AUXILIARY_MODEL_IDS — but the drop-in must sort last (zz-*.conf) because the stock laguna-env.conf loads /etc/laguna-s21-proxy.env afterwards and overwrites DISABLED_AUXILIARY_MODEL_IDS.
• The family JSON uses llama.cpp’s key name repeat_penalty (not repetition_penalty).
• Sharp template on the backend: CHAT_TEMPLATE_FILE=WORKSTATION/chat_template.jinja through the v3 launch script — required because the baked template ignores enable_thinking on this abliterated model (template finding #1 of qwen-next-flash).
• Presets deployed (verified live via /v1/presets): thinking 1.0/0.95/20/min_p 0/pres 0/rep 1.0; instruct 0.7/0.80/20/min_p 0/pres 1.5/rep 1.0.

6. Honest caveats

• All MTP numbers are ctx 32768 single-node medians; production 200K+MTP fit was never measured (prior attempt OOMs, and the validation that would have settled it never ran).
• The draft head v1 attends densely in the MTP block (trunk QSA-indexer pruning not wired into graph_mtp); numerically a superset, matters only for long-ctx draft cost.
• Q4_K_M draft came from q8_0 requant (--allow-requantize double quant); from-BF16 would be marginally better.
• Acceptance capped ~0.6 suggests the draft head mismatches the abliterated fine-tune; p-min tuning untested.
• Patches are uncommitted in WORKSTATION/llama.cpp-flash-next by instruction; the output_hc_norm mtp_only fix is worth upstreaming to #27836.

Reproduction

1. Export: apply #27836 + §1 fixes, then selective download → convert_hf_to_gguf.py --mtp --outtype q8_0 → llama-quantize --allow-requantize (job log: quant-jobs/mtp-unc-export/logs/).
2. Smoke: scripts/smoke_draft_mtp.sh (target + draft, port 12788).
3. A/B: scripts/ab_baseline.sh / scripts/ab_draft.sh (taskset-pinned, median-of-5).
4. Scheduled validation (as designed): scripts/cron_wrapper_ctx200k.sh → scripts/run_ctx200k_validation.sh; probes scripts/speed_ab_200k.py, scripts/qual_200k.py, suite scripts/bench_ctx200k_mtp.py.
5. Data: data/ab_*.json (Track 1), data/decide_*.json + data/DECIDE.json (ranking), data/bench_v3*.json / data/bench_v4*.json (no-rollback comparison).