NeuralDBG

Training failed? NeuralDBG tells you why. It hooks into your PyTorch training loop, detects what went wrong (vanishing gradients, exploding gradients, data anomalies), and pinpoints the exact layer and step — so you fix it in seconds, not hours.

Overview

NeuralDBG treats training as a semantic trace of learning dynamics rather than a black box. It extracts meaningful events and provides causal hypotheses about training failures, enabling researchers to:

• Identify gradient health transitions (stable -> vanishing/saturated)
• Detect activation regime shifts (normal -> saturated/dead)
• Detect optimizer instability (loss plateaus, spikes, divergence)
• Catch data anomalies (NaN, Inf, distribution shifts)
• Track propagation of instabilities through network layers
• Generate ranked causal explanations for training failures

Unlike traditional monitoring tools (TensorBoard, Weights & Biases), NeuralDBG focuses on causal inference rather than metric tracking.

Neural Suite

Component	Availability	Role
NeuralDBG	Public (PyPI)	Causal diagnostics in your training loop
Diagnostic Workspace	Private beta	Visual causal graphs and hypothesis explorer
Neural Agent	Private beta	Auto-remediation from causal hypotheses

Request early access: open an issue with label suite-access.

Try it in 60 seconds (Colab)

Or locally: python examples/quickstart_interactive.py

Public benchmark

Reproducible causal accuracy on synthetic failures (details):

python -m benchmark_public.run

Latest results: benchmark_public/results.json

Why NeuralDBG?

Feature	TensorBoard / W&B	NeuralDBG
What it shows	Graphs of loss/accuracy over time	Why the loss spiked or vanished
Diagnosis	Manual inspection of curves	Automated causal hypotheses
Actionable?	You guess the fix	Suggests root causes (LR, Init, Data)
Integration	Separate dashboard	One line of code in your loop
Privacy	Data sent to cloud	100% Local (unless you opt-in)

"TensorBoard tells you when it failed. NeuralDBG tells you why."

Latest: v1.5.0 (July 5, 2026)

• Tier 1 Black-Swan detection: 96% — GNN 88%, MoE 100%, Diffusion 100% (104/108 bugs detected)
• Tier 2 Black-Swan detection: 94% — FlashAttention 100%, Neural ODE 100%, Quantized 83% (102/108)
• 200-architecture validation — 79% global detection across 5 families (MLP/CNN/RNN/Transformer/Hybrid)
• NeuralPrune v0.1 — Non-destructive redundancy diagnostic: dead neurons, low-rank weights, quantization opportunities
• LSTM/GRU support — Forward hooks unwrap RNN output tuples. Per-gate gradient tracking (input/forget/cell/output)
• Architecture fuzzer — 94% crash rate across 50 randomly generated architectures
• Stress test suite — 15/15 tests pass: 10x gradients, NaN/Inf, fp16, 100-layer depth, 1K token attention
• GPU v4 model — Qwen2-0.5B + LoRA trained on 538 examples across 5 architecture families, 92.3% accuracy
• Aquarium web dashboard — Zero-dependency HTML causal viewer. Open Aquarium
• 4 upstream PyTorch PRs — svdvals NaN, F.normalize zero-input, gradient health tests, varlen_attn NaN fix
• 100% detection on DeepMLP (6/6 bugs) | 96% Tier 1 black-swans | 94% Tier 2 black-swans
• 4 upstream PRs submitted to PyTorch | CI benchmark workflow on GitHub Actions
• E2E RNN pipeline — detect→diagnose→fix→validate on LSTM. 2/4 bugs auto-fixed.

Key Features

• Semantic Event Extraction: Detects meaningful transitions in training dynamics
• Causal Compression: Identifies first occurrences and propagation patterns
• Post-Mortem Reasoning: Provides ranked hypotheses about failure causes
• Optimizer Instability Detection: Tracks loss plateaus, spikes, and divergence
• Data Anomaly Detection: Catches NaN, Inf, and distribution shifts in inputs
• Event Collapsing: Merges sequential events into summary traces
• Compiler-Aware: Operates at module boundaries to survive torch.compile
• Non-Invasive: Wraps existing PyTorch training loops without code changes
• Minimal API: Focused on explanations, not raw data dumps
• Diagnostic package export: JSON export for the Neural Suite visualizer (private beta)

Quick Start

Installation

pip install neuraldbg

Basic Usage

import torch
import torch.nn as nn
from neuraldbg import NeuralDbg

# Your existing model and training setup
model = nn.Sequential(nn.Linear(10, 5), nn.ReLU(), nn.Linear(5, 1))
optimizer = torch.optim.SGD(model.parameters(), lr=0.01)
criterion = nn.MSELoss()

# Wrap your training loop
with NeuralDbg(model) as dbg:
    for step, (inputs, targets) in enumerate(dataloader):
        optimizer.zero_grad()

        outputs = model(inputs)
        loss = criterion(outputs, targets)
        loss.backward()
        dbg.record_loss(loss.item())
        optimizer.step()

# After training failure, query for explanations
explanations = dbg.explain_failure()
print(explanations[0])  # "Gradient vanishing originated in layer 'linear1' at step 234..."

Inference API

# Get ranked causal hypotheses for the failure
hypotheses = dbg.get_causal_hypotheses()

# Query specific causal chains
chain = dbg.trace_causal_chain('vanishing_gradients')

# Check for coupled failures
couplings = dbg.detect_coupled_failures()

# Export diagnostic package (JSON) for Neural Suite visualizer
dbg.export_aquarium_package('debug_session.json')

Optimizer Instability Detection

with NeuralDbg(model) as dbg:
    for step in range(num_steps):
        dbg.step = step
        output = model(inputs)
        loss = criterion(output, targets)
        loss.backward()
        dbg.record_loss(loss.item())
        optimizer.step()

# Detect loss plateaus, spikes, or divergence
hypotheses = dbg.explain_failure("optimizer_instability")
for h in hypotheses:
    print(h.description)

Data Anomaly Detection

Data anomalies (NaN, Inf, distribution shifts) are detected automatically from layer inputs during the forward pass:

with NeuralDbg(model) as dbg:
    # ... training loop ...
    pass

hypotheses = dbg.explain_failure("data_anomaly")
for h in hypotheses:
    print(h.description)  # "NaN values detected in input to layer 'linear1'..."

Supported Architectures

Validated on 200 architectures, 1200 bug injections — the most comprehensive ML debugging tool validation to date.

Feed-Forward Architectures — 91% detection (782/852)

Family	Configs	Detection	Status
MLP (deep, residual, bottleneck)	50	93% (280/300)	✅
CNN (Conv2d, varying kernel/depth/norm)	40	91% (219/240)	✅
Transformer (encoder, varying heads/depth)	40	92% (221/240)	✅
Hybrid non-RNN (CNN+MLP, Attention+MLP)	12	86% (62/72)	✅
Feed-Forward Total	142	91%	✅

Recurrent Architectures — 49% detection (178/360)

⚠️ Known limitation: NeuralDBG's per-step gradient hooks are less effective on RNNs due to weight sharing across timesteps. This is inherent to the hook-based approach and affects LSTM, GRU, and RNN-containing hybrids equally.

Family	Configs	Detection	Status
RNN (LSTM/GRU, varying depth/width/bidir)	40	49% (117/240)	⚠️
Hybrid with RNN	18	10% (18/180)	⚠️

By Bug Type (200 configs each)

Bug	Detection	Difficulty
Exploding gradients	91%	Easy
Optimizer divergence	91%	Easy
NaN in data	83%	Moderate
Dead activations (bias)	71%	Moderate
Zero initialization	71%	Moderate
Vanishing gradients (sigmoid)	44%	Hard

Reproduce: python validate_combinatorial.py --full (200 configs, ~10 min on CPU).

Detailed results: combinatorial_results.json |

Supported Failure Types

Failure Type	Description
vanishing_gradients	Root cause + saturation coupling
exploding_gradients	First layer to explode
dead_neurons	Neuron death in activation layers
saturated_activations	Activation saturation patterns
optimizer_instability	Loss plateaus, spikes, divergence
data_anomaly	NaN/Inf/distribution shift in inputs

Architecture

Core Components

• Semantic Event Extractor: Detects meaningful transitions in learning dynamics
• Causal Compressor: Identifies patterns and propagation in training failures
• Post-Mortem Reasoner: Generates ranked hypotheses about failure causes
• Compiler-Aware Monitor: Operates at safe boundaries for optimization compatibility

Event Types

Event Type	Source	Detects
gradient_health_transition	Backward hooks	Vanishing, exploding, saturated gradients
activation_regime_shift	Forward hooks	Dead neurons, saturated activations
optimizer_instability	record_loss()	Loss plateaus, spikes, divergence
data_anomaly	Forward hooks (inputs)	NaN, Inf, distribution shifts

Editions

Edition	Package	License	Features
Core	pip install neuraldbg	MIT	Hooks, events, export JSON, basic heuristics
Engine	pip install neuraldbg-engine	Proprietary	Full causal inference, detailed hypotheses, coupling detection

The Core edition works standalone with basic heuristic fallbacks. Install the Engine for advanced causal reasoning.

Target Users

• ML Researchers seeking causal explanations for training failures
• PhD Students analyzing learning dynamics in novel architectures
• Research Engineers understanding optimization instabilities

Limitations

• PyTorch only
• Focus on semantic events, not tensor inspection

Contributing

1. Fork the repository
2. Create a feature branch
3. Add tests for new functionality
4. Ensure all tests pass
5. Submit a pull request

Developer Setup

make bootstrap
source .venv/bin/activate  # Linux/macOS
# or
.venv\Scripts\activate     # Windows

License

MIT License - see LICENSE.md for details.

Documentation

• Landing page — branding and suite overview
• PyTorch support matrix
• CHANGELOG.md - Version history and notable changes
• logic_graph.md - System architecture and data flow
• docs/PHASE2_DOGFOODING.md - Detailed dogfooding scenarios

Citation

If you use NeuralDBG in your research, please cite:

@misc{neuraldbg2026,
  title={NeuralDBG: A Causal Inference Engine for Deep Learning Training Dynamics},
  author={SENOUVO Jacques-Charles Gad},
  year={2026},
  url={https://github.com/LambdaSection/NeuralDBG}
}