Weightless

Inspect models without loading the weights.

Try it in your browser → — no install; paste a Hub id or drop a local file. Also live as a 🤗 Hugging Face Space. Or get the extension: VS Code Marketplace · Open VSX.

A VSCode extension that opens .safetensors and .gguf checkpoints — including multi-GB, multi-shard ones — instantly, by reading only the file header. It can even inspect models straight from the Hugging Face Hub without downloading them: a 500 GB sharded model costs a few hundred KB of traffic.

Double-click any .safetensors / .gguf file → a rich viewer opens instead of "binary file not shown."

Features

Architecture

• Flowchart view — the model as a branch-and-merge diagram: each input branch (vision tower, audio encoder, …) in its own column with its input type (image / pixels, token ids · vocab 262144) derived from the weights, meeting at a merge junction, flowing through the trunk, and splitting out to each head with its output type (logits · 262144, value (scalar)). Double-click to drill into any module; repeated blocks collapse to ×N (click to expand all N).
• Tree view — the classic collapsible module hierarchy.
• Code view — the selected scope as a reconstructed PyTorch-style definition (Linear(in_features=640, out_features=2048, bias=False)), GitHub-style with line numbers, syntax colors and one-click copy. Gaps in Sequential indices are called out as parameter-free modules (activations).

Understanding

• Backbone identification — names the architecture family from config.json (model_type, architectures, sub-configs), module naming, or structural fingerprints (≈ Gemma 3 from its norm scheme + vocab; ≈ SigLIP So400m from its widths) when nothing else is recorded. Plus exact provenance (base: google/gemma-3-270m) when a sidecar file stored it.
• Composition — actual layer counts (Linear / Attention / MLP / Norm / Embedding / Conv) per scope, updating as you click around the graph.
• I/O & training footprint — image input resolution (from patch + position embeddings), context length, vocab, tied/untied lm_head, LoRA rank + targets, weights size at bf16/fp32/int8/int4, full-finetune (AdamW) and LoRA memory rules-of-thumb, KV-cache per 1k tokens (GQA-aware).

Inventory

• Tensor table — searchable, with a layer column (Linear 640→2048) so mlp.gate_proj reads as what it is; shapes, dtypes, params, bytes, shard.
• Summary — params, tensor count, file size, shards, dtype distribution.
• Config / Training / Generation / Adapter / Tokenizer / Metadata tabs — adjacent sidecar JSONs auto-discovered and pretty-printed with syntax highlighting.

Formats & sources

• .safetensors — single-file and sharded (model.safetensors.index.json).
• .gguf — llama.cpp / quantized ecosystem: quant dtypes (Q4_K, Q6_K, …), exact per-tensor bytes, metadata mapped onto config facts (layers/heads/GQA/context).
• Hugging Face Hub — Weightless: Open Model from Hugging Face Hub command; type a model id, get the full viewer via HTTP Range requests (header-only). Supports HF_TOKEN for gated repos.

How it works (and its one honest limit)

These formats store weights, not code: a header describing every tensor's name, shape, dtype and byte-offsets, followed by raw bytes. Weightless parses only the header and derives everything that is actually derivable: the module hierarchy (dotted names encode it exactly), repeated structure, layer types and feature dims (from weight shapes), input modalities (from embeddings), and head output types.

What no tool can get from weights alone is the true forward-pass graph (execution order, activations, skip connections) — that lives in the model's code. Where Weightless shows inferred wiring, the edges are dashed and labeled as such.

Roadmap

• Per-tensor stats — NaN/Inf detection, min/max/mean/std via bounded ranged reads (the "is my checkpoint corrupt?" check), still never loading the full file.
• Checkpoint diff — compare two files: changed/added/removed tensors, param deltas.
• Tier 2 — true data-flow graph (opt-in): a Python companion tracing a forward pass with torch.fx to render real edges. Requires the model code; strictly optional.

Install / develop

git clone <this repo> && cd weightless
npm test          # header parsers (safetensors + GGUF) — no build step, plain JS
# Press F5 in VSCode to launch an Extension Development Host, then open any
# .safetensors / .gguf file, or run "Weightless: Open Model from Hugging Face Hub".
# to package a .vsix:
npx @vscode/vsce package

License

MIT — see LICENSE.