neuron-tai/.scratch/distributed-gguf-runtime/issues/066-implement-and-benchmark-v4-mtp.md

DGR-066: Implement and benchmark V4 MTP

• Status / triage: specification only; ready-for-agent; passes: false
• Execution mode: AFK
• Milestone: M4
• Dependencies: DGR-054, DGR-060, DGR-064, DGR-065
• Blocks (derived): DGR-067, DGR-070
• Labels: area:deepseek-v4, area:mtp, type:runtime, priority:p1, ready-for-agent
• Evidence class: real-hardware
• Hardware: required
• Model: deepseek-v4-flash
• Upstream: yes

Objective / description

Fresh Ralph session: read .scratch/distributed-gguf-runtime/RALPH-CONTEXT.md, source issue .scratch/distributed-gguf-runtime/issues/066-implement-and-benchmark-v4-mtp.md, and evidence READMEs for dependencies (DGR-054, DGR-060, DGR-064, DGR-065) before changing code. Inspect live source/tests rather than trusting legacy pass states. Objective: Implement the locked MTP contract around upstream llama.cpp before beta, without making MTP an alpha prerequisite.

Acceptance criteria

• Implement the exact DGR-065 MTP ownership and typed state contract against the pinned upstream V4 graph while preserving shard-local caches.
• Pass whole-model versus ranged MTP correctness, accepted-token, rejection, rollback/truncation, and session-isolation tests.
• Benchmark MTP-off versus MTP-on quality, accepted-token rate, TTFT, decode speed, seam bytes, and memory on the locked lane.
• Distinguish implemented, certified, disabled, and unsupported MTP recipes in fail-closed capability advertisement.
• Applicable shared quality gates in prd.json pass, and the evidence handoff records exact commands/results, changed files, limitations, and dependency handoff.

Shared quality gates

• Targeted deterministic tests pass; Python changes also pass python -m compileall packages tests.
• git diff --check passes.
• Default tests are model-download-free, API-credit-free, and GPU-free.
• Evidence README records exact changed files, commands/results, limitations, and dependency handoff; no fabricated evidence or inherited completion credit.
• Native changes pass focused out-of-tree CMake build and CTest; patch changes verify clean apply/check/reverse against the exact llama.cpp pin.
• Runs are opt-in and record exact artifact/split hashes, runtime/upstream pin, backend/driver, hardware, network, commands, and raw metrics. Model artifacts use configured mounted-drive storage and never /home.
• Preserve existing Transformers behavior and backend-agnostic Tracker routing/load balancing/billing/relay semantics unless an explicit versioned contract says otherwise. One scoped story commit is expected during execution, but this specification-materialization change is not committed.

Evidence handoff

Write and verify .scratch/distributed-gguf-runtime/evidence/DGR-066/README.md. Until every criterion and applicable gate has real evidence, this story remains passes: false. Legacy evidence is provenance only, not completion credit.