---
id: nvdlfw-inspect
version: "0.2.2"
license: Apache2
license_treatment: permissive
maintenance: aging
---

# nvdlfw-inspect — Facilitates debugging convergence issues and testing new algorithms/recipes for training LLMs using Nvidia libraries.

License: permissive · Maintenance: aging · Downloads: 147.4K/mo

## What it is and what it does

nvdlfw-inspect is a debugging toolkit for NVIDIA's deep learning framework ecosystem, designed to help diagnose convergence issues and validate new training algorithms when using Transformer Engine, Megatron-LM, NeMo, or plain PyTorch models. It works by attaching instrumentation at the layer level—you define which layers to monitor via regex patterns in a YAML config file, then selectively enable debug features (like tensor statistics collection) only for those layers. This targeted approach avoids the overhead of instrumenting an entire model.

The package provides both generic APIs for framework-agnostic tensor inspection and namespace-scoped APIs for framework-specific behavior. You initialize it once in your training script, configure which features and layers you want to monitor, and then call APIs to log statistics like mean, standard deviation, and norms on activations, weights, and gradients. It's built for multi-GPU training and expects initialization on every rank.

Use it for:

- Diagnosing why an LLM training run is not converging by inspecting weight and activation statistics across selected layers.
- Validating a new training algorithm or recipe by comparing tensor statistics before and after changes.
- Monitoring specific transformer layers (e.g., attention heads, feed-forward blocks) during distributed training without full-model instrumentation overhead.
- Collecting gradient flow statistics to detect vanishing or exploding gradients in deep models.
- Prototyping custom debug features by loading framework-specific feature directories alongside generic ones.

## Worth the install?

AI-flagged interpretation of the facts above — verify before relying.

Provides debugging and instrumentation APIs for LLM training workflows using NVIDIA libraries (Transformer Engine, Megatron-LM, NeMo) and PyTorch models, with configurable layer-level feature selection and tensor statistics collection.

Yes, if you are actively debugging LLM training convergence issues with NVIDIA frameworks and can tolerate the aging maintenance status. The low install friction and permissive license make it a low-risk addition to a training pipeline. However, the limited repository activity (21 stars, last commit 2025-09-17) and aging status suggest this is not a heavily supported tool—verify compatibility with your specific framework versions before relying on it in production.

## Install

pip install nvdlfw-inspect
uv add nvdlfw-inspect
poetry add nvdlfw-inspect

## Installing nvdlfw-inspect

Before you install: Low install friction with a pure-Python wheel. Maintenance status is aging—last commit was 2025-09-17 and the repository has only 21 stars, suggesting limited adoption and uncertain long-term support.

License in practice: Licensed under Apache2 (permissive), so you can use, modify, and distribute the package freely in both open and proprietary projects without viral obligations.

Quickstart:

pip install nvdlfw-inspect

import nvdlfw_inspect.api as debug_api
debug_api.initialize(config_file="debug_config.yaml")
debug_api.log_tensor_stats(layer_name, tensor=weight, tensor_name="weight")

Requires PyTorch and PyYAML as runtime dependencies; intended for multi-GPU training workflows where initialization should occur once per rank.

Verify before relying:

- Whether the package is actively maintained beyond the initial release cycle given the aging status and limited repository activity.
- Real-world performance and stability when used with large-scale distributed training setups.
- Compatibility guarantees with specific versions of Transformer Engine, Megatron-LM, and NeMo.

## Package facts

- License: Apache2 (permissive)
- Python support: supports_current
- Install friction: low
- Maintenance: aging
- Downloads: 147.4K/month (top 15,000 on PyPI)
- Known vulnerabilities: none known

## Tags

llm training debugging, nvidia megatron nemo debug, tensor statistics logging, convergence issue diagnosis, pytorch model instrumentation, transformer engine debugging, layer-level monitoring, llm-debugging, distributed-training, nvidia-ecosystem

[View on SkillFed](https://skillfed.io/packages/nvdlfw-inspect) · [View on PyPI](https://pypi.org/project/nvdlfw-inspect/)