#2550 docs: AI_SETUP.md and INSTALL.md offer the sizes setup offers (Swift 1.5 has IQ3_S, not Q2_0; UD-Q4_K_XL runs on AMD)

说明

**What it changes.** Three lines that disagree with what `setup.py` offers:

- `docs/AI_SETUP.md`: Swift 1.5's sizes were given as "Q2_0, IQ2_XS, IQ3_XXS". In `setup.py`'s `MODELS`, Q2_0 is `"families": ("qwen",)` only, and IQ3_S has been a Swift size since #1651. An AI assistant following the page with `--family swift --model Q2_0` gets "Swift 1.5 has no Q2_0 model file". Now: IQ2_XS, IQ3_XXS, IQ3_S; no Q2_0.
- `docs/AI_SETUP.md`: `--model UD-Q4_K_XL` was "experimental: NVIDIA only". No size in `MODELS` is marked `experimental`, and only the NVFP4 packs are `nvidia_only` (#429; the R9700 numbers are in `docs/QUANTS.md`). Now: 4-bit, 111 GB, part of its experts read from the SSD on a 64 GB PC, 7-8.5 tokens/s (with the machine it was measured on, from `docs/UNSLOTH_Q4.md`: RTX 5070, 64 GB, Windows), an NVMe SSD helps.
- `docs/INSTALL.md` (AMD section): "Unsloth's 4-bit model needs an NVIDIA card" is removed for the same reason.

**What it leaves alone.** Every other line, including the 64 GB "recommended" row (setup's recommendation changed in 0.1.44; that is a separate question) and the `--calibrate` remark (setup offers it on NVIDIA only; a run by hand works on Linux AMD, `docs/AMD_HIP.md`).

**How it was checked.** Against `setup.py` (`MODELS`, the family check that fails a size a family does not have, the AMD paths that test `nvidia_only`), `tools/test_setup_choices.py` (#1651's Swift IQ3_S files), `docs/UNSLOTH_Q4.md` and `docs/QUANTS.md`. A sub-agent reviewer checked each claim. No code changes.

🤖 Generated with [Claude Code](https://claude.com/claude-code)