v0.5.0 — Transformers and LLMs

New features

• HuggingFace parser. It reads config.json and the safetensors header. No torch needed.
• Header-only parsing. A 27B checkpoint opens in about six seconds. The tool never reads the full file.
• INT4 checkpoints work. Packed weights show their true sizes.
• BF16 and FP16 tensors get honest histograms. Small tensors only, with a size cap.
• Transformer blocks draw as a repeated sequence along one axis.
• Attention heads draw as small nodes under the output projection. Each head receives q, k, and v. Each head emits one attends edge.
• Path view. Click a neuron, then press T. One unit lights up through all open layers.
• Layer-of-detail. Models above 20 000 neurons open with layers folded. Double-click a layer to open it.
• Top bar. Fold all, unfold all, jump to any layer, and stretch the horizontal axis.
• Load from disk. Type a model folder path in the page. The server reads the disk itself: no upload and no size limit.
• Context UI. Run shows only for a runnable PyTorch model. The image input shows only for conv models.

Install

pip install -e .
nnviz model.pth          # PyTorch or ONNX file
nnviz llama_folder/      # HuggingFace folder: config.json + safetensors
nnviz serve              # drag & drop, or type a folder path in the page
nnviz live model.pth     # live mode: run the model, watch activations

Design note

The parser never loads a full checkpoint. It reads the safetensors header, then maps single tensors, under a size cap, for histograms. Quantized tensors carry no histogram: packed integers give no honest distribution. Zero new dependencies. The parser needs no torch.

Fixed

• A big model crashed the viewer on load. The collapse message ran before the toast element existed.

Known limits

• Drag & drop rejects files above 2 GiB. Type the folder path in the page instead. Or run nnviz load <folder>.
• A bare safetensors upload shows no heads. The head count lives in config.json, which the upload does not carry.
• Live mode needs a full PyTorch nn.Module. HuggingFace graphs load and draw, but they cannot run.
• Vision towers (VLMs) are skipped. Branch layout is the v0.6 roadmap.
• Head counts above 64 draw as attributes, not nodes.
• Every layer and head is its own draw call. A fluidity pass is planned for v0.8.