update

heneil · web-flow · commit e8b4821e4219 · 2025-12-10T14:19:29.000-06:00
diff --git a/helm/utils/train_util.py b/helm/utils/train_util.py
@@ -121,7 +121,7 @@ def is_no_decay(name):
             name.endswith(".bias")
             or name.endswith(".weight") and "norm" in name.lower()
             or name.endswith(".gate.bias")
-            or name.endswith(".scale")          # if you exposed trainable scales
+            or name.endswith(".scale")          
             or "w_x" in name.lower()
         )
     
@@ -180,4 +180,4 @@ def is_no_decay(name):
             num_processes = num_processes
         )
         decoder, scheduler_euc, optimizer = accelerator.prepare(decoder, scheduler_euc, opt_euc)
-        return train_dataloader, decoder, scheduler_euc, optimizer
+        return train_dataloader, decoder, scheduler_euc, optimizer

Original file line number	Diff line number	Diff line change
`@@ -121,7 +121,7 @@ def is_no_decay(name):`
`121`	`121`	`name.endswith(".bias")`
`122`	`122`	`or name.endswith(".weight") and "norm" in name.lower()`
`123`	`123`	`or name.endswith(".gate.bias")`
`124`		`- or name.endswith(".scale") # if you exposed trainable scales`
	`124`	`+ or name.endswith(".scale")`
`125`	`125`	`or "w_x" in name.lower()`
`126`	`126`	`)`
`127`	`127`
`@@ -180,4 +180,4 @@ def is_no_decay(name):`
`180`	`180`	`num_processes = num_processes`
`181`	`181`	`)`
`182`	`182`	`decoder, scheduler_euc, optimizer = accelerator.prepare(decoder, scheduler_euc, opt_euc)`
`183`		`- return train_dataloader, decoder, scheduler_euc, optimizer`
	`183`	`+ return train_dataloader, decoder, scheduler_euc, optimizer`