Pull requests / #1874

#1874 docs/UNSLOTH_Q4: the first NVIDIA numbers for UD-IQ4_XS (2x2080 Ti, 46 GB RAM)

open · @go-farther-and-farther · 0 コメント · GitHub で見る

BenchmarksMulti-GPUNVIDIA / CUDAModels & quantsDocumentation

本文

Issue: Refs #1866

## Summary
The UD-IQ4_XS section asks for NVIDIA reports ("please report what you see"). This adds the first one: 2x RTX 2080 Ti 22 GB with 46 GiB of RAM, engine 0.1.41, both modes measured on the same machine.

## What changed
docs/UNSLOTH_Q4.md, the "UD-IQ4_XS" section only:
- replaces "has not been measured on NVIDIA yet" with the measured paragraph and a small table
- covers the tier math on 46 GiB (10.82 GiB single-GPU cache + 37 GiB budget vs 55.43 GiB of experts), the resulting SSD reads, the freeze without a budget on two GPUs, and what RAM size keeps the SSD out
- same-box IQ3_XXS numbers as the comparison the section's readers will ask for

## Extra Notes
All numbers from one machine; bench script and engine logs available on request. The quality checks passed (retrieval at 262K, arithmetic). The ask in #1866 (the budget on a layer split) is referenced, not claimed.

関連リンク

インストール・モデル・リリースへの站内リンク。