Quantization
Outlier channels contain values whose magnitudes are much larger than most peer channels. When one tensor scale preserves those extremes, fewer useful levels remain for ordinary values.
Outliers explain why one scale for an entire tensor can lose important resolution. Channelwise scales, protected channels, or wider precision can isolate their effect.
Quantizers inspect activation and weight distributions before choosing scale granularity. Teams verify that outlier handling improves task quality without adding unsupported layouts or excessive kernel work.
Outlier channels contain values whose magnitudes are much larger than most peer channels. When one tensor scale preserves those extremes, fewer useful levels remain for ordinary values. Quantizers inspect activation and weight distributions before choosing scale granularity. Teams verify that outlier handling improves task quality without adding unsupported layouts or excessive kernel work.
Outliers explain why one scale for an entire tensor can lose important resolution. Channelwise scales, protected channels, or wider precision can isolate their effect.
RunInfraby RightNow© 2026 RunInfra. All rights reserved.
