Quantization Diff
Comparing a quantized model against its full-precision source is a special case of weight diff with a different question behind it: not "did this change," but "is the expected quantization error the whole story, or did something else change too."
Why this needs its own view
Every element of a quantized tensor differs from its full-precision source by construction - that's what quantization is. A plain weight diff between the two is therefore never "clean," which makes an ordinary diff heatmap nearly useless here: everything lights up, whether or not anything is actually wrong.
Quantization diff instead compares each element's actual error against the error expected from its quantization scale and granularity - an element whose error is in line with what its scale would predict is quantization noise; an element whose error is much larger than expected is a real signal worth investigating, even though both look identical in a raw diff.
Reading a quantization diff
The Matrix/Heatmap view's quantization-diff mode shades by this normalized error rather than by raw difference, so expected noise reads as uniform low-level color and genuinely anomalous regions stand out the way an outlier would in any other statistical view - the same visual language, applied to error-relative-to-expected-error instead of value-relative-to-distribution.
Common findings
- Uniform low error - quantization is behaving as expected; nothing further to investigate.
- Concentrated high-error regions - often a granularity mismatch (a per-tensor scale applied where per-channel would have been more appropriate), visible as a cluster rather than noise scattered evenly.
- High error correlated with sparsity - small values rounding to exactly zero (see Sparsity), which is a real behavioral change even though it's a predictable consequence of low-bit quantization.
Related resources
- Dtype and Quantization - the scale/zero-point/granularity model a quantization diff is normalized against.
- Weight Diff - the general-purpose numeric diff this view specializes.
- Quantized Models - the applied, format-level context for reading a quantized checkpoint at all.