canifinetune 0.1.1

Docs-only patch release — first release published via PyPI trusted publishing.

• README masthead: SVG hero banner (also rendered on the PyPI project page), centered badges
• Architecture diagram now uses an absolute URL so it renders on PyPI
• No code changes since 0.1.0

Install: pip install canifinetune

v0.1.0 — first release

First release. Estimate, benchmark, and generate fine-tuning recipes for
LLMs on consumer GPUs.

What's included

• `canifinetune doctor` — environment + GPU summary.
• `canifinetune estimate` — pure-math feasibility with full memory
  breakdown (weights, trainable params, gradients, optimizer states,
  activations, CUDA / fragmentation, safety margin).
• `canifinetune recommend` — auto-search a feasible config for your card.
• `canifinetune bench` — real PyTorch + PEFT + bitsandbytes + TRL
  training step on your GPU. Captures peak reserved VRAM, tok/s, step time.
• `canifinetune calibrate` — closes the predict ↔ measure gap with a
  multiplicative correction fit to the activations term.
• `canifinetune recipe` — generates a ready-to-run `training.py` for the
  chosen config.
• `canifinetune report` / `canifinetune compare` — markdown tables across
  runs.

RTX 4080 baselines

Real measurements committed in `docs/rtx4080_baselines.md`:

model	method	seq_len	measured peak	tok/sec
`sshleifer/tiny-gpt2` (smoke)	lora	128	0.12 GB	1735
`Qwen/Qwen2.5-0.5B-Instruct`	qlora	1024	3.30 GB	1995
`Qwen/Qwen2.5-1.5B-Instruct`	qlora	1024	4.36 GB	1352
`Qwen/Qwen2.5-1.5B-Instruct`	qlora	2048	7.10 GB	1470
`Qwen/Qwen2.5-3B-Instruct`	qlora	1024	5.54 GB	1158

Tested on

• Python 3.10 / 3.11 / 3.12 (CI matrix)
• Linux + Windows (WSL not required for the math layer; bench needs CUDA)