make it work

src/axolotl/core/trainers/kd.py

@@ -13,46 +13,52 @@ def kd_loss_function(
     student_logits,
     target_token_ids,
     target_logprobs,
+    target_mask,
     num_items_in_batch: Optional[int] = None,
-    **kwargs,  # pylint: disable=unused-argument
 ):
-    # student_logits: [B, seq_len, vocab_size] from the student's forward pass
+    # teacher_mask: [B, teacher_seq_len, K], where 1 indicates a valid token and 0 indicates padding
-    # target_token_ids: [B, teacher_seq_len, K] top-K token IDs from teacher
-    # target_logprobs: [B, teacher_seq_len, K] teacher logprobs for these top-K tokens
 
+    # Determine the teacher sequence length
     teacher_seq_len = target_token_ids.shape[1]
 
-    # Slice the student logits to match the teacher-provided seq length
+    # Slice student logits to match the teacher-provided sequence length
     student_logits_for_kd = student_logits[
         :, -teacher_seq_len:, :
-    ]  # Now [B, teacher_seq_len, vocab_size]
+    ]  # [B, teacher_seq_len, vocab_size]
 
     # Gather student logits for teacher's top-K tokens
     student_logits_topk = torch.gather(
         student_logits_for_kd, dim=-1, index=target_token_ids
     )  # [B, teacher_seq_len, K]
 
-    # Convert student top-K logits to logprobs
+    # Convert student top-k logits to logprobs
     student_logprobs_topk = student_logits_topk - torch.logsumexp(
         student_logits_topk, dim=-1, keepdim=True
-    )
+    )  # [B, seq_len, K]
 
-    # teacher_probs are simply exp of teacher_logprobs (already scaled)
+    # Convert teacher_mask to boolean for indexing
+    valid_mask = target_mask.bool()
+
+    # Prune tensors to only keep valid tokens
+    # This will result in 1D arrays of only valid positions
+    student_logprobs_topk = student_logprobs_topk[valid_mask]  # [N_valid_tokens]
+    target_logprobs = target_logprobs[valid_mask]  # [N_valid_tokens]
+
+    # Since teacher_logprobs are already normalized, just exponentiate to get probabilities
     teacher_probs = target_logprobs.exp()
 
-    # Compute forward KL
+    # Compute forward KL:
-    # L_kl = sum_k p^T_k (log p^T_k - log p^S_k)
+    # KL = sum p^T_k (log p^T_k - log p^S_k), summed over all valid tokens.
-    kd_loss_per_position = (
+    kd_loss_per_token = teacher_probs * (target_logprobs - student_logprobs_topk)
-        teacher_probs * (target_logprobs - student_logprobs_topk)
+    kd_loss = kd_loss_per_token.sum()
-    ).sum(
-        dim=-1
-    )  # [B, teacher_seq_len]
 
-    # gradient accumulation fixes
+    # Normalize by number of items or mean over valid tokens
-    if num_items_in_batch:
+    if num_items_in_batch is not None:
-        kd_loss = kd_loss_per_position.sum() / num_items_in_batch  # Scalar
+        # If you know how many items should be considered in the batch
+        kd_loss = kd_loss / num_items_in_batch
     else:
-        kd_loss = kd_loss_per_position.mean()  # Scalar
+        # Otherwise, just average over all valid tokens
+        kd_loss = kd_loss / kd_loss_per_token.size(0)
 
     return kd_loss
 
@@ -70,6 +76,8 @@ class AxolotlKDTrainer(AxolotlTrainer):
                 columns_to_add.append("target_logprobs")
             if "target_token_ids" not in self._signature_columns:
                 columns_to_add.append("target_token_ids")
+            if "target_mask" not in self._signature_columns:
+                columns_to_add.append("target_mask")
             if columns_to_add:
                 self._signature_columns += columns_to_add
 
@@ -83,6 +91,7 @@ class AxolotlKDTrainer(AxolotlTrainer):
         """
         target_logprobs = inputs.pop("target_logprobs")
         target_token_ids = inputs.pop("target_token_ids")
+        target_mask = inputs.pop("target_mask")
 
         if self.model_accepts_loss_kwargs:
             loss_kwargs = {}
@@ -96,6 +105,7 @@ class AxolotlKDTrainer(AxolotlTrainer):
             student_logits,
             target_token_ids,
             target_logprobs,
+            target_mask,
             num_items_in_batch=num_items_in_batch,
         )
 

src/axolotl/prompt_strategies/chat_template.py

@@ -490,8 +490,23 @@ class ChatTemplateStrategyWithKD(ChatTemplateStrategy):
 
     def transform_logprobs(self, sample):
         logprobs = sample.pop(self.logprobs_field)
+        target_seq_len = len(logprobs)
+        input_seq_len = len(sample["input_ids"])
+        padding_len = input_seq_len - target_seq_len
+        top_k = len(logprobs[0])
         target_logprobs = []
         target_token_ids = []
+        target_mask = []
+
+        # 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(padding_len):
+            target_logprobs.append([-float("inf")] * top_k)
+            target_token_ids.append(list(range(top_k)))
+            target_mask.append([0] * top_k)
+
+        for _ in range(target_seq_len):
+            target_mask.append([1] * top_k)
 
         for _, token_pos_logprobs in enumerate(logprobs):
             # Initialize collections for logprobs and token_ids
@@ -519,6 +534,7 @@ class ChatTemplateStrategyWithKD(ChatTemplateStrategy):
             # Apply temperature scaling at data load time
             # log p_k^(T) = (log p_k / T) - logsumexp(log p_j / T)
             position_logprobs_tensor = position_logprobs_tensor / self.temperature
+            # normalize to probabilities so they sum up to 1
             position_logprobs_tensor = position_logprobs_tensor - torch.logsumexp(
                 position_logprobs_tensor, dim=0, keepdim=True
             )
@@ -531,6 +547,7 @@ class ChatTemplateStrategyWithKD(ChatTemplateStrategy):
         # Update sample with transformed logprobs
         sample["target_logprobs"] = target_logprobs
         sample["target_token_ids"] = target_token_ids
+        sample["target_mask"] = target_mask
 
         return sample
 
@@ -538,7 +555,9 @@ class ChatTemplateStrategyWithKD(ChatTemplateStrategy):
         logprobs = prompt.pop(self.logprobs_field)
         tokenized_prompt = super().tokenize_prompt(prompt)
         tokenized_prompt[self.logprobs_field] = logprobs
-        return self.transform_logprobs(tokenized_prompt)
+        tokenized_prompt = self.transform_logprobs(tokenized_prompt)
+
+        return tokenized_prompt
 
 
 def load(tokenizer, cfg, ds_cfg: Optional[Dict[str, Any]] = None, processor=None):

src/axolotl/utils/collators/kd.py

@@ -1,5 +1,6 @@
 """
-DataCollator for axolotl to handle KD fields
+DataCollator for axolotl to handle KD fields without using -inf for padding,
+and with a teacher_mask to identify padded positions.
 """
 
 from dataclasses import dataclass
@@ -17,6 +18,9 @@ from axolotl.utils.collators.batching import DataCollatorForSeq2Seq
 class DataCollatorForKD(DataCollatorForSeq2Seq):
     """
     Data collator for KD, including handling KD-specific fields.
+
+    This version avoids using -inf and instead uses a large negative value for padding
+    target_logprobs. It also creates a teacher_mask to indicate which entries are valid.
     """
 
     tokenizer: PreTrainedTokenizerBase
@@ -32,7 +36,9 @@ class DataCollatorForKD(DataCollatorForSeq2Seq):
         if return_tensors is None:
             return_tensors = self.return_tensors
 
-        # Extract labels and position_ids first (as in original code)
+        padding_side = self.tokenizer.padding_side
+
+        # Pad labels and position_ids first
         for feature_name, pad_token_id in [
             ("labels", self.label_pad_token_id),
             ("position_ids", self.position_pad_token_id),
@@ -46,7 +52,6 @@ class DataCollatorForKD(DataCollatorForSeq2Seq):
                         // self.pad_to_multiple_of
                     ) * self.pad_to_multiple_of
 
-                padding_side = self.tokenizer.padding_side
                 for f in features:  # pylint: disable=invalid-name
                     remainder = [pad_token_id] * (max_len - len(f[feature_name]))
                     if isinstance(f[feature_name], list):
@@ -69,63 +74,77 @@ class DataCollatorForKD(DataCollatorForSeq2Seq):
         # Handle target_logprobs and target_token_ids manually
         target_logprobs_list = []
         target_token_ids_list = []
+        target_mask_list = []
         has_teacher_data = ("target_logprobs" in features[0]) and (
             "target_token_ids" in features[0]
         )
 
         if has_teacher_data:
-            # Extract these fields
+            # Extract and remove from features
             for f in features:  # pylint: disable=invalid-name
                 target_logprobs_list.append(f.pop("target_logprobs"))
                 target_token_ids_list.append(f.pop("target_token_ids"))
+                target_mask_list.append(f.pop("target_mask"))
 
-            # Determine max lengths to pad
+            # Determine max lengths
             max_teacher_seq_len = max(len(seq) for seq in target_logprobs_list)
             max_k = max(len(seq_k) for seq in target_logprobs_list for seq_k in seq)
 
-            # Pad target_logprobs and target_token_ids
             padded_target_logprobs = []
             padded_target_token_ids = []
-            for t_logprobs, t_ids in zip(target_logprobs_list, target_token_ids_list):
+            padded_teacher_mask_list = []
-                # Pad seq dimension
+
+            for t_logprobs, t_ids, t_mask in zip(
+                target_logprobs_list, target_token_ids_list, target_mask_list
+            ):
                 t_logprobs_padded = []
                 t_ids_padded = []
-                for i in range(  # pylint: disable=consider-using-enumerate
+                t_mask_padded = []
-                    len(t_logprobs)
+
+                for lp, ids, mask in zip(  # pylint: disable=invalid-name
+                    t_logprobs, t_ids, t_mask
                 ):
-                    lp = t_logprobs[i]  # pylint: disable=invalid-name
-                    ids = t_ids[i]
-                    # Pad K dimension
                     lp_len = len(lp)
                     if lp_len < max_k:
-                        lp = lp + [-float("inf")] * (  # pylint: disable=invalid-name
+                        # Use -1e9 for padding logprobs and 0 for token_ids
-                            max_k - lp_len
+                        pad_len = max_k - lp_len
-                        )  # or some pad value that won't break exp()
+                        lp = lp + [-1e9] * pad_len  # pylint: disable=invalid-name
-                        ids = ids + [0] * (max_k - lp_len)
+                        ids = ids + [0] * pad_len
+                        mask = mask + [0] * pad_len
+                    else:
+                        lp = lp[:max_k]  # pylint: disable=invalid-name
+                        ids = ids[:max_k]
+                        mask = mask[:max_k]
+
                     t_logprobs_padded.append(lp)
                     t_ids_padded.append(ids)
+                    t_mask_padded.append(mask)
 
-                # If sequence is shorter than max_teacher_seq_len
                 seq_len_diff = max_teacher_seq_len - len(t_logprobs_padded)
                 if seq_len_diff > 0:
+                    # Pad sequences fully if needed
                     t_logprobs_padded.extend(
-                        [[-float("inf")] * max_k for _ in range(seq_len_diff)]
+                        [[-1e9] * max_k for _ in range(seq_len_diff)]
                     )
                     t_ids_padded.extend([[0] * max_k for _ in range(seq_len_diff)])
+                    t_mask_padded.extend([[0] * max_k for _ in range(seq_len_diff)])
 
                 padded_target_logprobs.append(t_logprobs_padded)
                 padded_target_token_ids.append(t_ids_padded)
+                padded_teacher_mask_list.append(t_mask_padded)
 
             # Convert to tensors
             padded_target_logprobs = torch.tensor(
                 padded_target_logprobs, dtype=torch.float
             )
-            # We can store token_ids as long tensor
             padded_target_token_ids = torch.tensor(
                 padded_target_token_ids, dtype=torch.long
             )
+            padded_teacher_mask_list = torch.tensor(
+                padded_teacher_mask_list, dtype=torch.int
+            )
 
-        # Now pad using tokenizer for the remaining fields (input_ids, attention_mask, etc.)
+        # Pad using tokenizer for regular fields
         features = self.tokenizer.pad(
             features,
             padding=self.padding,
@@ -134,12 +153,13 @@ class DataCollatorForKD(DataCollatorForSeq2Seq):
             return_tensors=return_tensors,
         )
 
-        # Add back the teacher data if it exists
+        # Add back teacher data if present
         if has_teacher_data:
             features["target_logprobs"] = padded_target_logprobs
             features["target_token_ids"] = padded_target_token_ids
+            features["target_mask"] = padded_teacher_mask_list
 
-        # Prepare decoder_input_ids if applicable
+        # Prepare decoder_input_ids if the model supports it
         if (
             "labels" in features
             and self.model is not None