Skip to content

Research 01 — Landscape: visual NN designers, node editors, config/codegen targets

Researched 2026-09-18.

1. Existing tools

Legacy visual NN builders (all dead or dormant)

Tool What it does Status Shape check / params / codegen License
Sony Neural Network Console Windows GUI, drag-drop layers, NAS, training EOL April 3 2025; repo archived "Network statistics: computational complexity and memory usage"; export NNP/ONNX/TF Apache-2.0
ENNUI (live) Browser drag-drop (dense/conv/pool/norm/dropout/concat/sum), trains in-browser, "Export to Python" Last commit 2023-01. Dead Keras-style codegen; no attention layers; no param/FLOP stats MIT
Deep Cognition / DLS Drag-drop Keras builder (2017–2020) Company pivoted; product gone Keras codegen Proprietary
NVIDIA DIGITS Web UI for Caffe/TF CNN training (form-based) Archived 2025-01 No BSD-3
PerceptiLabs TF visual modeler generating TF code Shut down; last push 2021-08 Codegen yes Proprietary
Cloud-CV Fabrik Browser drag-drop, import/export Caffe/Keras/TF Last push 2020-12 Codegen yes GPL-3.0
Microsoft Lobe No-code image classifier Deprecated ~2023 n/a Proprietary
TorchStudio PyTorch IDE; graph view shows tensor size changes for a coded model Last release 2024-04. Dormant Shape display; no codegen MIT
KAIBER NN Editor Browser editor, real-time output shapes, exports PyTorch Beta 5, Sept 2022. Dormant Shapes + codegen; no stats; no transformer blocks Closed beta

No official Keras/TensorFlow visual builder exists; ML.NET "Model Builder" is AutoML, not architecture design.

Visualization-only (maintained, not editors)

  • Netron: MIT, very active. Renders ONNX/PyTorch graphs with shapes. No editing, no codegen.
  • NN-SVG, PlotNeuralNet: publication schematics only.
  • Google Visual Blocks: Apache-2.0, active; composes ML pipelines (models are opaque nodes); no layer-level design.

2024–2026 projects closest to the idea

  • BuildANeuralNet: React 18 + React Flow + FastAPI; drag-drop layers, real-time connection validation, PyTorch nn.Module codegen. GPL-3.0, 1 star, MLP/CNN only; transformer blocks on TODO.
  • easy-torchcraft: Lovable-generated React app, PyTorch export. No license, 28 stars.
  • v0-pytorch-neural-network-designer: v0-generated toy.
  • NeuralFlows.ai: claims transformer design + in-browser training + PyTorch export. Unverifiable (site unreachable).

LLM-specific visualizers/calculators (none are editors)

Conclusion

No polished "CAD for LLM architectures" exists as of Sept 2026. Every drag-drop layer builder is dead or an MLP/CNN toy. None offer MoE/GQA/MLA/SSM blocks, param/FLOP/KV-cache accounting, and PyTorch codegen together. The niche is open.

2. Node editor libraries

React Flow / @xyflow/react

  • Current: 12.11.6 (2026-09-01). MIT. React 17+ (React 19 works). 38k stars, weekly releases.
  • Custom nodes are React components; handles with ids; isValidConnection hook for typed ports; MiniMap, Controls, Background, NodeToolbar, NodeResizer.
  • Sub-flows: parentId, extent: 'parent', expandParent, group node type. Parents must precede children in the nodes array. Multi-level nesting works in practice.
  • Layouting guide: dagre (simple/fast), elkjs (most configurable, sub-flows + edge routing, ~1 MB, run in a worker), d3-hierarchy, d3-force.
  • Licensing: core is MIT "forever"; Pro buys examples/support, not required.
  • Siblings: Svelte Flow 1.6.x (Svelte 5), Vue Flow 1.48.x.

Alternatives

  • Rete.js 2.x: framework-agnostic, typed sockets, DataflowEngine, scopes plugin for nesting. More assembly, smaller ecosystem.
  • LiteGraph.js: original dormant. Comfy-Org TypeScript fork published as @comfyorg/litegraph (npm MIT; monorepo GPL-3.0, check before vendoring). Canvas2D, imperative, fast with big graphs.
  • Drawflow: dormant, no typed ports.
  • Baklava.js: Vue 3, typed interfaces, active. Vue only.
  • Blueprint-style: Flume (low activity), node-blueprint (Vue). Nothing mainstream.
  • Layout: @dagrejs/dagre 3.x (MIT), elkjs 0.12 (EPL-2.0 OR GPL-3.0), d3-dag (MIT, Sugiyama).

ComfyUI as reference

ComfyUI shipped subgraphs in Aug 2025 (v0.3.51): collapse selection to a super-node with typed I/O, nestable, breadcrumb navigation. Its workflow JSON is its own format (nodes[], links[], groups, extra), not the editor library's. Lesson: persist our own IR; treat editor state as a view.

Recommendation

React Flow (@xyflow/react 12.x) + elkjs (dagre for simple cases). Typed handles + isValidConnection map onto shape checks; sub-flows give the "Transformer Block ×N" container; nodes are React components so shape badges and live counts are trivial; MIT; largest community. Caveat: DOM-based, so collapse repeated blocks into a single ×N node (needed for the math anyway).

3. Declarative configs and codegen: targets and import formats

Project Form Modern blocks Use
HF transformers (v5.x) config.json per model; modular transformers generates modeling files; unified AttentionInterface GQA, MLA, MoE, Mamba hybrids Primary IMPORT (config.json is the de facto standard). Secondary TARGET for interop.
TorchTitan (BSD-3) ModelArgs dataclasses + flavor dicts, TOML selects flavor; DeepSeek-V3 args include num_experts, q_lora_rank, kv_lora_rank Llama 3/4, DeepSeek-V3, Qwen3, GPT-OSS TARGET for scale-out training; args are a checklist of what a design must specify.
litgpt (Apache-2.0) One Config dataclass (~100 fields: n_query_groups, mlp_class_name, norm_class_name, rope params), name_to_config registry GQA, MoE, RoPE variants; no SSM Cheapest TARGET: emit a Config, get pretrain/finetune/serve free. Also IMPORT.
MaxText (JAX) YAML configs, decoder_block selects layer type MLA, shared/routed experts, YaRN Reference/IMPORT only.
nanoGPT / modded-nanogpt (MIT) GPTConfig; modded-nanogpt is a single hand-tuned train_gpt.py Dense GPT + speedrun tricks Minimal TARGET template; reference for modern tricks.
Megatron-Core ModuleSpec + TransformerLayerSubmodules: declarative submodule tree MoE, GQA, MLA, hybrids Closest existing "architecture-as-spec" in production. Heavy; custom license.
nanotron, Lingua YAML / dataclass configs Dense + some MoE Secondary targets.
flash-linear-attention (MIT), mamba (Apache-2.0) Importable nn.Module blocks (GLA, DeltaNet, GatedDeltaNet, Mamba/Mamba2) SSM/linear attention Generated PyTorch should import these for SSM nodes rather than re-implement kernels.
torch.fx Graph IR with Python codegen op-level Possible backend but loses block structure; templating nn.Module is simpler.
torch.export ExportedProgram (ATen IR, shape-aware) op-level Verification target: export generated model, cross-check shapes/params vs editor estimates.
ONNX Op graph poor for MoE/custom kernels Optional export for Netron viewing.

Suggested pipeline: own JSON IR (with a "block ×N" container) → codegen to (a) plain PyTorch nn.Module importing fla/mamba_ssm for exotic blocks, (b) HF config.json + modeling file, (c) litgpt Config where representable; import HF config.json and litgpt Config; validate with torch.export shape propagation.

4. Small-scale validation references (for the "test bench" idea)

  • modded-nanogpt: record #91 (Aug 2026) reaches 3.28 val loss on FineWeb in 1.126 min on 8×H100. Tricks in the record: Muon, value embeddings, U-net skips, FlashAttention 3 + sliding window, FP8, logit softcap, multi-token prediction, bigram hash embeddings, MUDD skips, learnable XSA gates. No single-consumer-GPU track exists; we would define our own tiny track.
  • METR note on NanoGPT progress (2026-04): 31× speedup May 2024 → Mar 2026; early gains were architecture, later gains optimizer and kernels. Imported techniques contributed 6.7×, adapted 3.0×, invented 1.6×. Small-scale-only tricks often do not transfer; "scale-complementary" ones (e.g. Muon) did.
  • Parameter Golf (arXiv 2607.01517): 84 techniques catalogued across 1,430 submissions; individual techniques rarely exceed 1% and their gains shrink when re-measured among competitive submissions. Lesson: an ablation harness must control for the surrounding stack, and report confidence intervals.
  • μP / μTransfer: microsoft/mup, Cerebras practitioner's guide, How To Scale. Relevant for a "scale ladder" feature: tune at small width, transfer to large width.

Unverified

NeuralFlows.ai claims; Sony NNC PyTorch engine support; React Flow practical nesting depth; Megatron-LM license terms.