ended8월 22일· 1 sources

Comparing INT4 and NVFP4 Palettes on Real Gradient Tensors

Why it matters

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...

1
Sources
+0
24h
Growth
29d
Active

Sources

Related Issues