sft

The Swiss army knife for .safetensors files.

sft is a single-binary CLI for inspecting, editing, and diffing .safetensors files — and an interactive terminal browser for poking around large checkpoints. Most commands read the file header only, so multi-gigabyte models open in milliseconds.

It also ships a skill that teaches AI coding agents (Claude Code, Cursor, Codex CLI) when to reach for it and how to parse the output.

Install

uv tool install sft-cli                 # recommended
pip install sft-cli                     # or pip
uv tool install 'sft-cli[torch]'        # + .pt/.pth conversion

🚀 Quick start

sft model.safetensors                   # open the interactive browser
sft info model.safetensors              # one-shot summary
sft info model.safetensors --json       # machine-readable

The bare sft <file> form is a shortcut for sft browse <file>. Inside the browser: ↑↓ to navigate, / to filter, L on a LoRA file for LoRA Mode, D to diff against another file, q to quit.

What it does

Inspect a file without loading it:

sft info  model.safetensors             # size, tensor count, dtypes, metadata
sft ls    model.safetensors             # flat list, sort/filter friendly
sft tree  model.safetensors --depth=2   # hierarchical view
sft stat  model.safetensors             # per-tensor mean/std/min/max/sparsity
sft check model.safetensors             # corruption + NaN/Inf scan

Compare two checkpoints:

sft diff base.safetensors finetuned.safetensors --delta \
    --include='**.self_attn.**'         # cosine, L2, max-abs per tensor

Edit without writing Python:

sft slice  big.safetensors --include='**.weight' -o weights-only.safetensors
sft strip  big.safetensors --exclude='*lora_*'
sft cast   model.safetensors --dtype fp16
sft cat    a.safetensors b.safetensors -o merged.safetensors
sft rename model.safetensors --sub 'model\.' 'backbone.'
sft split  model.safetensors --max-size 4GB
sft convert pytorch_model.bin           # → safetensors

Every write command supports --dry-run and never overwrites the input — outputs default to {stem}.{suffix}.safetensors.

Adapter workflows (PEFT and Kohya):

sft lora info    adapter.safetensors            # rank, alpha, target modules
sft lora svd     adapter.safetensors            # singular-value spectrum
sft lora compat  base.safetensors adapter.safetensors
sft lora extract base.safetensors ft.safetensors --rank 16
sft lora resize  adapter.safetensors --rank auto    # per-pair adaptive rank
sft lora stack   a.safetensors b.safetensors -a 0.7 -b 0.3
sft lora merge   base.safetensors adapter.safetensors
sft lora convert adapter.safetensors --to peft      # Kohya ↔ PEFT

--rank auto picks each pair's output rank from its singular-value spectrum (ceil(stable_rank) + 1), so over-parameterized pairs compress harder than rich ones. auto+N adds a safety margin.

The browser

Press a key, get a result.

🤖 AI agents

sft skill install                       # auto-detects Claude / Cursor / Codex
sft skill status
sft skill uninstall

The installer symlinks sft's skill into your agent's well-known skills directory (~/.claude/skills/sft, ~/.cursor/skills/sft, etc.), so it stays in sync when you uv tool upgrade sft-cli. Pass --mode copy for a frozen snapshot.

Every command supports --json for clean parsing:

sft info model.safetensors --json | jq '.tensors'
sft lora info adapter.safetensors --json | jq '.rank'
sft stat model.safetensors --json --include='**.q_proj.*'

License

MIT — see LICENSE.