How a request flows: production system vs OSS
Pipeline table, official architecture image, production metrics with attribution, OSS gaps.
Prefill vs Decode: same engine, two phases
Same-engine table, OSS vs production gaps, Q&A parallelism ranges with caveats, Pareto charts.
Hardware gate: SM100 & Ascend 950 today; 910C & H20 only RFCs
Support matrix, capacity/blocker tables for RFC #46/#47, serving gaps, skip checklist.
DeepInfer OSS scope: what ships, what does not
Ascend day-0 in the listed matrix; KV/PD/DP LB out; no official Docker.