--- title: "Debugging Gradient Flow in Training - Air Market" description: "Sometimes a model trains but parts of it never seem to learn, and it is hard to tell which piece of the loss is doing the work. This walks through the training setup and shows which parts of the loss…" canonical: "https://airmarket.ai/app/debug-gradient-flow" --- [Air Market](/) / [Data & Analytics](/category/data) / Debugging Gradient Flow in Training Skill by [DeepModeling](/@deepmodeling) # Debugging Gradient Flow in Training Find out where training stops learning and why ## What this does Sometimes a model trains but parts of it never seem to learn, and it is hard to tell which piece of the loss is doing the work. This walks through the training setup and shows which parts of the loss reach which parameters, and where the connection stops. Useful when a compiled model behaves differently from the plain one. ### Top 3 things it does for you - Trace which loss terms reach which parameters - Pin down where the gradient chain breaks - Check gradients after compiling a model ## Information **Debugging Gradient Flow in Training at a glance** | | | | --- | --- | | Kind | Skill | | Publisher | [DeepModeling](/@deepmodeling) | | Category | [Data & Analytics](/category/data) | | Works with | Agent Skills - an open standard supported by 40+ agents | | Version | 1.0 | | Size | 17 KB | | Licence | LGPL-3.0 | | Repository | [deepmodeling/deepmd-kit](https://github.com/deepmodeling/deepmd-kit/tree/01bcf476c5e73d009e77c7cb94c513d008c8c1f1/.github/skills/debug-gradient-flow) | | GitHub stars | 2,017 (as of Nov 2026) | | Installs | 3, counted by skills.sh (Nov 2026) | ## Security report **Verified.** AIR scanned this exact version of Debugging Gradient Flow in Training and flagged nothing. AIR's verdict: **safe**. **What Debugging Gradient Flow in Training's check was run against** | | | | --- | --- | | Commit | 01bcf476c5e73d009e77c7cb94c513d008c8c1f1 | | Content hash | 31d2fbb6a7236b444cc79ac6a4e644f504c3f0b3ffdf1cdfc5fdc78c91d362bf | | Files read | [Browse the files at this exact commit](https://github.com/deepmodeling/deepmd-kit/blob/01bcf476c5e73d009e77c7cb94c513d008c8c1f1/.github/skills/debug-gradient-flow/SKILL.md) | A scan is named by the exact content it read, not by a version number. A different version is a different artifact, and nothing here applies to a version AIR has not scanned. [Read the full security report for Debugging Gradient Flow in Training](/security/debug-gradient-flow) ## About Debugging Gradient Flow in Training When training does not behave, the question is usually which part of the loss is actually reaching which parameters. This works through that systematically, separating the energy, force and virial contributions and showing where each one lands. It is aimed in particular at models that have been put through compilation, where the gradient path is harder to follow by reading the code. The result is a picture of where the gradient chain breaks. ## What's inside - SKILL.md - instructions - references/gradient-probe-script.md - reference ## Questions ### Is Debugging Gradient Flow in Training safe to install? AIR scanned this exact version of Debugging Gradient Flow in Training and flagged nothing. AIR's verdict: safe. The verdict is named by the content that was read, at commit 01bcf476c5e7, so it cannot drift onto a version nobody scanned. The full report is at https://airmarket.ai/security/debug-gradient-flow. ### How do I install Debugging Gradient Flow in Training? Open https://airmarket.ai/app/debug-gradient-flow on Air Market and use the install action: Air Market hands Claude a prompt naming this exact skill, and Claude sets it up. Source, pinned to the exact commit this listing describes: https://github.com/deepmodeling/deepmd-kit/tree/01bcf476c5e73d009e77c7cb94c513d008c8c1f1/.github/skills/debug-gradient-flow. Install from that link, not the branch. ## Related on Air Market - [More Data & Analytics listings](/category/data) - [Everything by DeepModeling](/@deepmodeling) --- Canonical URL: