Comparing INT4 and NVFP4 Palettes on Real Gradient Tensors
Four-bit training quantizes every number to one of 16 values. NVFP4's menu is {0, ±0.5, ±1, ±1.5, ±2, ±3, ±4, ±6} , with one scale factor per block of 16 elements. Those levels are spaced like a float: fine near zero, coarse at the top. The standard pipeline also applies a random Hadamard rotation before quantizing, which spreads outlier energy across a block and pulls the per-coordinate…
INT4 and NVFP4 quantization palettes were compared in a study examining their effectiveness on gradient tensors. INT4 quantizes numbers into 16 values while NVFP4 offers a menu of 16 values with one scale factor per block of 16 elements. Both palettes utilize a random Hadamard rotation to handle outliers, redistributing energy and aligning distributions toward a bell curve.
Uniform INT4 outperformed NVFP4 on both rotated and unrotated data by a factor of 2.3. An additional analysis on 45 real gradient tensors revealed that uniform INT4 was 22.1% more accurate at the unbiased corner and 17.5% more accurate at the biased corner compared to NVFP4, even without rotation. The rotation, though slightly improving the result, did not enable the lower error as initially suggested.
The outliers in real gradient tensors reside between blocks rather than within them. Quantizers with per-16-element scales focus on one block at a time, influenced by each block's absmax. This emphasizes that the trade-off between unbiased and biased estimators is inherent to having 16 values, not the specific choice of values by NVIDIA.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
