#2142 Community benchmark: mixed dual V100 + RTX PRO 4000 Blackwell on Windows

Description

This results-only report measures IQ2_XS with 262144 context, INT8 KV, MTP and Vision on Windows 10, EPYC 7A23 and 64 GB RAM, using the existing CUDA 12.9 engine 0.1.40.

It compares the dual-V100 baseline, both three-stage GPU orders, and a dual-V100 language engine with Vision on the RTX PRO 4000 Blackwell. Each configuration has three fresh-prompt runs at 4096/32768/128000/258000 input tokens, 256 output tokens and zero prefix reuse (48 trials). With PRO first, 258000-token median TTFT changed from 158.86 to 87.24 seconds and decode from 82.2 to 95.9 tok/s. Nine long-context recall checks and API/image/tool checks passed; all four coding scores, including failed edge cases, are disclosed.

The report includes sanitized configs, measured allocations, all per-run data/outputs, timing excerpts and small reproduction scripts. Limitations include different reserve settings, auto-split changes, fixed test order, background desktop activity, thermal/power limits, an older installed binary, and ambiguous coding-task boundary wording. This does not claim universal CUDA 12/Blackwell compatibility or broad Agent quality. No engine changes, binaries, model files or private transcripts are included.