Covers the CLAUDE.md tiering (user/project/local/one-off task) for
splitting instructions between the local coder model and Claude Code,
plus how Claude Code reads from self-hosted Gitea (plain git vs. the
official gitea/gitea-mcp server for PR/issue-level integration).
"Tesla" was NVIDIA's datacenter-card brand (retired after Volta), spanning
four unrelated architecture generations (Kepler/Maxwell/Pascal/Volta) - a
different, much older thing than the Tesla microarchitecture itself, which
actually predates Fermi/Kepler/Maxwell/Pascal/Volta. Clarify to avoid
reading the power-connector note as claiming Tesla == Volta.
Confirmed via NVIDIA's own release notes: R580 is the last branch
adding Volta/Pascal support (through ~June 2028). Also flag the
proprietary-vs-open kernel module split, since -open only covers
Turing and newer and autoinstall could pick it on a newer distro.
CUDA 13 already dropped Pascal/Volta (not a future warning anymore) -
call out pinning the driver branch instead. Also flag the 8-pin
CPU/EPS12V power connector and passive cooling on Tesla PCIe cards,
since it trips up anyone repurposing one in a tower case.
MoE VRAM tracks total params but compute tracks active params, so a
model like qwen3-coder:30b-a3b (30B total, ~3.3B active) runs light
enough that old GPUs' weak/missing tensor cores barely matter.
None of local-ai-setup.sh's tier-selected models (CHAT_MODEL/CODE_MODEL/
EMBED_MODEL) can read an image — there was no way to get vision support out
of this stack at all before now. Added a numbered pick-list to the
generated pull-models.sh, right after the existing DeepSeek-R1 optional
pull, matching that same read -rp pattern:
1) moondream ~1.7 GB by Moondream AI — tiny, built for
CPU-only or weak/old-GPU hardware
2) llava:7b ~4.7 GB general-purpose vision
3) qwen2.5vl:7b ~6 GB stronger accuracy, more RAM/VRAM
4) llama3.2-vision:11b ~7.9 GB heaviest of the four
moondream is the recommended default — sized for exactly the "6 vCPU, 8GB
RAM, no GPU" case this was asked for, unlike the other three which assume
real GPU/RAM headroom.
Verified by actually running the heredoc that generates pull-models.sh
(with EMBED_MODEL/CHAT_MODEL/CODE_MODEL stood in) and syntax-checking the
resulting output script, not just the source — the outer heredoc is
unquoted so $-escaping mistakes wouldn't show up as a bash -n failure on
local-ai-setup.sh itself, only on what it generates.
services/ai-stack.md gets a matching "Vision models" section (sizes, the
manual pull command, and how to point an app's OPENAI_MODEL at one).
laptop_full_setup.sh's separate, non-interactive pull-models.sh generator
is untouched — it's not invoked anywhere in this repo's own install flow
(only local-ai-setup.sh is, from install_ai-stack()), so it's out of
scope here.
Applies the new write_readme companion-doc convention to ai-stack,
paintplus, and kyber-launcher: install-time-invariant content (usage
walkthroughs, service tables, troubleshooting) moves out of the .sh
heredocs into sibling services/<name>.md files, leaving only genuinely
install-specific content inline.
- services/paintplus.md: config/cloud/GPU/ai-stack-backend/update/Caddy
sections, picked up automatically via write_readme's companion-doc
support.
- services/ai-stack.md: roles, GPU switcher, service URLs, cloud LLM
provider setup, update, Caddy. ai-stack.sh can't use write_readme
directly (its POST-INSTALL-NOTES.md filename deliberately avoids
colliding with the vendored app's own README.md in the same
directory), so it appends the companion file manually.
- services/kyber-launcher.md: the full SWBF2 (2017) + Kyber walkthrough,
moved out of the root README's "Gaming scripts" section (which now
just points here). kyber-launcher.sh now calls write_readme to deploy
it to ~/.local/share/kyber/README.md, fixing a stale in-script pointer
to a README section that no longer exists.
Claude