ugh

vall_e.cpp/vall_e.cpp

@@ -14,7 +14,7 @@
 #include <unordered_map>
 #include <iostream>
 
-#define LLAMA_CPP_EXTENDED 1 // whether the underlying llama.cpp has some extra functions
+#define LLAMA_CPP_EXTENDED 0 // whether the underlying llama.cpp has some extra functions
 #define LLAMA_CPP_USE_VALL_E_ARCH 0 // whether the underlying llama.cpp is to use the VALL_E arch
 
 #if !LLAMA_CPP_EXTENDED

vall_e/models/base.py

@@ -1436,6 +1436,9 @@ class Base(nn.Module):
 					continue
 
 				# offset to flattened vocab ranges
+				if self.classifier is not None:
+					compute_acc = False
+				"""
 				if self.classifier is not None:
 					offsets = _get_offsets()
 
@@ -1454,6 +1457,7 @@ class Base(nn.Module):
 							if t == self.ignore_index:
 								continue
 							token[i] += start
+				"""
 
 				if token.is_floating_point():
 					ignored = True
@@ -1679,13 +1683,6 @@ class Base(nn.Module):
 			for i, state in enumerate( hidden_states ):
 				hidden_states[i] = [ hi[:li] for hi, li in zip(hidden_states[i], map(len, x_list)) ]
 		
-		if not training:
-			loss = None
-			stats = None
-
-			self.loss = None
-			self.stats = None
-
 		# de-offset if needed
 		if self.classifier is not None:
 			offsets = _get_offsets()
@@ -1704,6 +1701,12 @@ class Base(nn.Module):
 
 				logits[batch_index] = logits[batch_index][:, start:start+end]
 
+		if not training:
+			loss = None
+			stats = None
+
+			self.loss = None
+			self.stats = None
 		# compute loss if the target is given
 		else:
 			loss, stats = self.calc_loss( inputs=inputs, logits=logits, quant_levels=quant_levels )