improve logprob masking and shift in trainer

src/axolotl/integrations/kd/chat_template.py

@@ -53,6 +53,10 @@ class ChatTemplateStrategyWithKD(ChatTemplateStrategy):
         )
 
     def transform_logprobs(self, sample):
+        """
+        Transform logprobs to target format for KD training
+        """
+
         logprobs = sample.pop(self.logprobs_field)
         target_seq_len = len(logprobs)
         input_seq_len = len(sample["input_ids"])
@@ -62,9 +66,20 @@ class ChatTemplateStrategyWithKD(ChatTemplateStrategy):
         target_token_ids = []
         target_mask = []
 
+        if input_padding_len < 0:
+            # logprobs is longer than target_seq_len,
+            # so we need to slice from the left/beginning of logprobs
+            logprobs = logprobs[:-input_seq_len]
+            input_padding_len = 0
+            target_seq_len = input_seq_len
+
         # fill with -inf for padding_len tokens for top_k tokens
         # extend target_logprobs with a padding_len x top_k 2D list filled with -inf
-        for _ in range(1, input_padding_len):  # start at 1 since this is causal
+
+        # for causal models, if we start the range at 1, then we don't need to shift in the trainer
+        # otherwise, we need to shift in the trainer
+        shift = 0
+        for _ in range(shift, input_padding_len):
             target_logprobs.append([-float("inf")] * top_k)
             target_token_ids.append(list(range(top_k)))
             target_mask.append([0] * top_k)
@@ -73,6 +88,12 @@ class ChatTemplateStrategyWithKD(ChatTemplateStrategy):
             # TODO also check against sample["labels"]
             target_mask.append([1] * top_k)
 
+        for position in range(input_padding_len, input_seq_len):
+            if sample["labels"][position] == -100:
+                target_mask.append([0] * top_k)
+            else:
+                target_mask.append([1] * top_k)
+
         for _, token_pos_logprobs in enumerate(logprobs):
             # Initialize collections for logprobs and token_ids
             position_logprobs = []
@@ -120,10 +141,11 @@ class ChatTemplateStrategyWithKD(ChatTemplateStrategy):
             target_logprobs.append(position_logprobs_scaled)
             target_token_ids.append(position_token_ids)
 
-        # since we started at index 1 for causal, we need one more padding token
-        target_logprobs.append([-float("inf")] * top_k)
-        target_token_ids.append(list(range(top_k)))
-        target_mask.append([0] * top_k)
+        if shift == 1:
+            # since we started at index 1 for causal, we need one more padding token
+            target_logprobs.append([-float("inf")] * top_k)
+            target_token_ids.append(list(range(top_k)))
+            target_mask.append([0] * top_k)
 
         # Update sample with transformed logprobs
         sample["target_logprobs"] = target_logprobs

src/axolotl/integrations/kd/trainer.py

@@ -45,7 +45,7 @@ class AxolotlKDTrainer(AxolotlTrainer):
         inputs,
         return_outputs=False,
         num_items_in_batch=None,
-        shift_targets=False,
+        shift_targets=True,
     ):
         """
         How the loss is computed by Trainer. By default, all models return the loss in the first element.