fix: Fix evaluation loss in KD trainer (#3271)

* fix: Fix evaluation loss in KD trainer

* Fix v2 strategy super() call

* fix: Add safety check for total_tokens in log method

* fix: simplified num items and outputs return handling

* fix: add missing model forward pass in compute_loss

* refactor: Use Template Method pattern for chat template strategies

* refactor: use pop(None) and remove v2 override

* chore: lint

---------

Co-authored-by: NanoCode012 <nano@axolotl.ai>
Co-authored-by: Wing Lian <wing@axolotl.ai>

src/axolotl/core/trainers/base.py

@@ -631,7 +631,11 @@ class AxolotlTrainer(
             logs["tokens_per_second_per_gpu"] = round(
                 self.state.last_tokens_per_second.item() / self.args.logging_steps, 2
             )
-            logs["total_tokens"] = int(self.state.total_tokens.item())
+            if (
+                hasattr(self.state, "total_tokens")
+                and self.state.total_tokens is not None
+            ):
+                logs["total_tokens"] = int(self.state.total_tokens.item())
 
         del self._stored_metrics[train_eval]
 

src/axolotl/integrations/kd/chat_template.py

@@ -179,8 +179,17 @@ class ChatTemplateStrategyWithKD(ChatTemplateStrategy):
         logprobs = prompt.pop(self.logprobs_field)
         tokenized_prompt = super()._tokenize_single_prompt(prompt)
         tokenized_prompt[self.logprobs_field] = logprobs
-        tokenized_prompt = self.transform_logprobs(tokenized_prompt)
 
+        # let subclasses add fields before transform
+        tokenized_prompt = self._prepare_kd_fields(tokenized_prompt, prompt)
+
+        tokenized_prompt = self.transform_logprobs(tokenized_prompt)
+        return tokenized_prompt
+
+    def _prepare_kd_fields(self, tokenized_prompt, original_prompt):
+        """
+        Hook for subclasses to prepare additional KD fields before transform
+        """
         return tokenized_prompt
 
 
@@ -283,14 +292,13 @@ class ChatTemplateStrategyWithKDv2(ChatTemplateStrategyWithKD):
 
         return sample
 
-    def _tokenize_single_prompt(self, prompt):
-        target_token_ids = prompt.get("target_token_ids", None)
-
-        tokenized_prompt = super()._tokenize_single_prompt(prompt)
-
+    def _prepare_kd_fields(self, tokenized_prompt, original_prompt):
+        """
+        Add pre-tokenized target_token_ids for v2 format
+        """
+        target_token_ids = original_prompt.pop("target_token_ids", None)
         if target_token_ids is not None:
             tokenized_prompt["target_token_ids"] = target_token_ids
-
         return tokenized_prompt
 
 

src/axolotl/integrations/kd/trainer.py

@@ -16,6 +16,8 @@
 KD trainer
 """
 
+from typing_extensions import override
+
 from axolotl.core.trainers.base import AxolotlTrainer
 
 from .kernels.liger import LigerFusedLinearKLTopKLogprobLoss
@@ -60,6 +62,7 @@ class AxolotlKDTrainer(AxolotlTrainer):
             if columns_to_add:
                 self._signature_columns += columns_to_add
 
+    @override
     def compute_loss(
         self,
         model,
@@ -79,10 +82,22 @@ class AxolotlKDTrainer(AxolotlTrainer):
         ):
             del inputs["attention_mask"]
 
+        if num_items_in_batch is None and "labels" in inputs:
+            num_items_in_batch = (inputs["labels"] != -100).sum().item()
+
         if self.model_accepts_loss_kwargs:
             loss_kwargs = {}
             if num_items_in_batch is not None:
                 loss_kwargs["num_items_in_batch"] = num_items_in_batch
             inputs = {**inputs, **loss_kwargs}
+
         outputs = model(**inputs)
-        return outputs[0]
+
+        if isinstance(outputs, dict):
+            loss = outputs["loss"]
+        elif isinstance(outputs, tuple):
+            loss = outputs[0]
+        else:
+            loss = outputs.loss if hasattr(outputs, "loss") else outputs
+
+        return (loss, outputs) if return_outputs else loss