This document is relevant for: Trn2, Trn3

Accuracy

Design documentation for the accuracy debugging tools — how they are built internally. Read these only if you want to understand, extend, or contribute to the tool implementations.

To use the accuracy tools, see:

• Debugging accuracy issues — methodology

• Accuracy debugger tools — how to run them

• Accuracy examples — simplified examples for accuracy tools

How the tools are built:

Topic	What it explains
Accuracy debugger design	Accuracy debugger design and report explanation
Dataset evaluation	How task-level lm_eval dataset runners work
Logit validation	How teacher-forced logit comparison works
KV cache analysis	How KV cache extraction and comparison works
Tensor capture	How intermediate tensors are captured from compiled graphs
Tensor compare	Two-way and three-way tensor comparison algorithms
Tensor replacement	How tensors can be injected into compiled graphs for isolation
Input snapshot	How NRT-boundary input tensors are captured for off-chip replay
Module test guidelines	Writing per-module accuracy tests with appropriate thresholds

This document is relevant for: Trn2, Trn3