You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
g_idx is a quantization parameter used when applying activation ordering using the GPTQ algorithm. This parameter is initialized to -1s. However, when GPTQ is applied, the parameter's value is set, and it is a bug to read the value before the value is calibrated by GPTQ (in the same way that it is a bug to read scales before they are calibrated).
Purpose
Reduce runtime when using activation ordering
Clean codebase of extra cases
Proposed Changes
Remove all instances and cases of using g_idx full of negative ones from CT and LC
Background
g_idxis a quantization parameter used when applying activation ordering using the GPTQ algorithm. This parameter is initialized to-1s. However, when GPTQ is applied, the parameter's value is set, and it is a bug to read the value before the value is calibrated by GPTQ (in the same way that it is a bug to read scales before they are calibrated).Purpose
Proposed Changes