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.Modulecodegen. 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)¶
- rasbt/LLM-architecture-gallery: Apache-2.0, 1.5k stars, active. Figures + structured data for every major open architecture (GQA/MLA/MoE). Best reference catalog for our block library and presets.
- Devisri-B/LLM-Architectures: React Flow diagrams of GPT/Llama, viz only.
- Liears/llm-architecture-atlas: static catalog from
models.yml. - llm-arch-reviewer: block diagrams with vLLM/SGLang profile overlays.
- KV/memory calculators: inference-estimator, LMCache kvcache-view.
- Educational: Transformer Explainer, LLM Visualizer.
- Searches for "visual LLM builder" are dominated by agent/pipeline builders (Flowise, ADK Visual Agent Builder). Different problem.
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;
isValidConnectionhook for typed ports; MiniMap, Controls, Background, NodeToolbar, NodeResizer. - Sub-flows:
parentId,extent: 'parent',expandParent,groupnode type. Parents must precede children in thenodesarray. 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.