Upload docs/5090D-acceptance-checklist.md with huggingface_hub
Browse files
docs/5090D-acceptance-checklist.md
ADDED
|
@@ -0,0 +1,14 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# 5090D acceptance checklist
|
| 2 |
+
|
| 3 |
+
Run the k=2 v2 profile on a single RTX 5090D before publishing performance numbers.
|
| 4 |
+
|
| 5 |
+
- Use `context_length=262144` and test `261632` input tokens plus `56` output tokens.
|
| 6 |
+
- Keep main KV at NVFP4 and MTP draft KV at FP8 E4M3.
|
| 7 |
+
- 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.
|
| 8 |
+
- Sample NVML memory at 20 ms or faster from process start through request completion.
|
| 9 |
+
- Require peak device usage at or below 32268 MiB (32 GiB minus 500 MiB).
|
| 10 |
+
- Require at least 500 MiB free relative to a 32 GiB device budget.
|
| 11 |
+
- Check prompt token count, cached tokens, all three retrieval facts, MTP acceptance, and decode throughput.
|
| 12 |
+
- If k=2 exceeds the memory gate, rerun the unchanged k=1 fallback before changing weight precision.
|
| 13 |
+
|
| 14 |
+
The current k=2 v2 evidence is from an RTX PRO 6000 and cannot substitute for this hardware check.
|