Fix quantization attribute assignment

carmocca · web-flow · commit cc6795d62947 · 2023-06-01T18:51:09.000+02:00
diff --git a/lit_llama/quantization.py b/lit_llama/quantization.py
@@ -574,7 +574,7 @@ def quantize(self):
                             W[:, (i1 + i) : (i1 + i + self.groupsize)]
                         )
                         self.scales[:, (i1 + i) // self.groupsize] = scale
-                        self.zeros[:, (i1 + i) // self.groupsize] = zeros
+                        self.zeros[:, (i1 + i) // self.groupsize] = zero
 
                 q = self.quantize_weight(w.unsqueeze(1), scale, zero, self.maxq)
                 q = q.squeeze(1)

Original file line number	Diff line number	Diff line change
`@@ -574,7 +574,7 @@ def quantize(self):`
`574`	`574`	`W[:, (i1 + i) : (i1 + i + self.groupsize)]`
`575`	`575`	`)`
`576`	`576`	`self.scales[:, (i1 + i) // self.groupsize] = scale`
`577`		`- self.zeros[:, (i1 + i) // self.groupsize] = zeros`
	`577`	`+ self.zeros[:, (i1 + i) // self.groupsize] = zero`
`578`	`578`
`579`	`579`	`q = self.quantize_weight(w.unsqueeze(1), scale, zero, self.maxq)`
`580`	`580`	`q = q.squeeze(1)`