docs/5090D-acceptance-checklist.md

5090D acceptance checklist

Run the k=2 v2 profile on a single RTX 5090D before publishing performance numbers.

• Use context_length=262144 and test 261632 input tokens plus 56 output tokens.
• Keep main KV at NVFP4 and MTP draft KV at FP8 E4M3.
• Keep the validated FlashInfer workspace at 144 MiB; increase only if the backend reports a required buffer larger than this, then repeat the memory gate.
• Sample NVML memory at 20 ms or faster from process start through request completion.
• Require peak device usage at or below 32268 MiB (32 GiB minus 500 MiB).
• Require at least 500 MiB free relative to a 32 GiB device budget.
• Check prompt token count, cached tokens, all three retrieval facts, MTP acceptance, and decode throughput.
• If k=2 exceeds the memory gate, rerun the unchanged k=1 fallback before changing weight precision.

The current k=2 v2 evidence is from an RTX PRO 6000 and cannot substitute for this hardware check.