Fix ONNX export for causal LM sequence classifiers by removing reverse indexing (#28144)

* normalize reverse indexing for causal lm sequence classifiers

* normalize reverse indexing for causal lm sequence classifiers

* normalize reverse indexing for causal lm sequence classifiers

* use modulo instead

* unify modulo-based sequence lengths

src/transformers/models/bloom/modeling_bloom.py

@@ -1011,7 +1011,10 @@ class BloomForSequenceClassification(BloomPreTrainedModel):
             sequence_lengths = -1
         else:
             if input_ids is not None:
-                sequence_lengths = (torch.ne(input_ids, self.config.pad_token_id).sum(-1) - 1).to(logits.device)
+                # if no pad token found, use modulo instead of reverse indexing for ONNX compatibility
+                sequence_lengths = torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1
+                sequence_lengths = sequence_lengths % input_ids.shape[-1]
+                sequence_lengths = sequence_lengths.to(logits.device)
             else:
                 sequence_lengths = -1
                 logger.warning(

src/transformers/models/ctrl/modeling_ctrl.py

@@ -796,9 +796,10 @@ class CTRLForSequenceClassification(CTRLPreTrainedModel):
             sequence_lengths = -1
         else:
             if input_ids is not None:
-                sequence_lengths = (torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1).to(
-                    logits.device
-                )
+                # if no pad token found, use modulo instead of reverse indexing for ONNX compatibility
+                sequence_lengths = torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1
+                sequence_lengths = sequence_lengths % input_ids.shape[-1]
+                sequence_lengths = sequence_lengths.to(logits.device)
             else:
                 sequence_lengths = -1
                 logger.warning(

src/transformers/models/deprecated/open_llama/modeling_open_llama.py

@@ -923,9 +923,10 @@ class OpenLlamaForSequenceClassification(OpenLlamaPreTrainedModel):
             sequence_lengths = -1
         else:
             if input_ids is not None:
-                sequence_lengths = (torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1).to(
-                    logits.device
-                )
+                # if no pad token found, use modulo instead of reverse indexing for ONNX compatibility
+                sequence_lengths = torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1
+                sequence_lengths = sequence_lengths % input_ids.shape[-1]
+                sequence_lengths = sequence_lengths.to(logits.device)
             else:
                 sequence_lengths = -1
 

src/transformers/models/deprecated/transfo_xl/modeling_transfo_xl.py

@@ -1247,9 +1247,10 @@ class TransfoXLForSequenceClassification(TransfoXLPreTrainedModel):
             sequence_lengths = -1
         else:
             if input_ids is not None:
-                sequence_lengths = (torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1).to(
-                    logits.device
-                )
+                # if no pad token found, use modulo instead of reverse indexing for ONNX compatibility
+                sequence_lengths = torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1
+                sequence_lengths = sequence_lengths % input_ids.shape[-1]
+                sequence_lengths = sequence_lengths.to(logits.device)
             else:
                 sequence_lengths = -1
                 logger.warning(

src/transformers/models/falcon/modeling_falcon.py

@@ -1432,7 +1432,10 @@ class FalconForSequenceClassification(FalconPreTrainedModel):
             sequence_lengths = -1
         else:
             if input_ids is not None:
-                sequence_lengths = (torch.ne(input_ids, self.config.pad_token_id).sum(dim=-1) - 1).to(logits.device)
+                # if no pad token found, use modulo instead of reverse indexing for ONNX compatibility
+                sequence_lengths = torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1
+                sequence_lengths = sequence_lengths % input_ids.shape[-1]
+                sequence_lengths = sequence_lengths.to(logits.device)
             else:
                 sequence_lengths = -1
                 logger.warning(

src/transformers/models/gpt2/modeling_gpt2.py

@@ -1451,9 +1451,10 @@ class GPT2ForSequenceClassification(GPT2PreTrainedModel):
             sequence_lengths = -1
         else:
             if input_ids is not None:
-                sequence_lengths = (torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1).to(
-                    logits.device
-                )
+                # if no pad token found, use modulo instead of reverse indexing for ONNX compatibility
+                sequence_lengths = torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1
+                sequence_lengths = sequence_lengths % input_ids.shape[-1]
+                sequence_lengths = sequence_lengths.to(logits.device)
             else:
                 sequence_lengths = -1
                 logger.warning(

src/transformers/models/gpt_bigcode/modeling_gpt_bigcode.py

@@ -1384,9 +1384,10 @@ class GPTBigCodeForSequenceClassification(GPTBigCodePreTrainedModel):
             sequence_lengths = -1
         else:
             if input_ids is not None:
-                sequence_lengths = (torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1).to(
-                    logits.device
-                )
+                # if no pad token found, use modulo instead of reverse indexing for ONNX compatibility
+                sequence_lengths = torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1
+                sequence_lengths = sequence_lengths % input_ids.shape[-1]
+                sequence_lengths = sequence_lengths.to(logits.device)
             else:
                 sequence_lengths = -1
                 logger.warning(

src/transformers/models/gpt_neo/modeling_gpt_neo.py

@@ -1113,9 +1113,10 @@ class GPTNeoForSequenceClassification(GPTNeoPreTrainedModel):
             sequence_lengths = -1
         else:
             if input_ids is not None:
-                sequence_lengths = (torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1).to(
-                    logits.device
-                )
+                # if no pad token found, use modulo instead of reverse indexing for ONNX compatibility
+                sequence_lengths = torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1
+                sequence_lengths = sequence_lengths % input_ids.shape[-1]
+                sequence_lengths = sequence_lengths.to(logits.device)
             else:
                 sequence_lengths = -1
                 logger.warning(

src/transformers/models/gpt_neox/modeling_gpt_neox.py

@@ -1200,9 +1200,10 @@ class GPTNeoXForSequenceClassification(GPTNeoXPreTrainedModel):
             sequence_lengths = -1
         else:
             if input_ids is not None:
-                sequence_lengths = (torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1).to(
-                    logits.device
-                )
+                # if no pad token found, use modulo instead of reverse indexing for ONNX compatibility
+                sequence_lengths = torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1
+                sequence_lengths = sequence_lengths % input_ids.shape[-1]
+                sequence_lengths = sequence_lengths.to(logits.device)
             else:
                 sequence_lengths = -1
                 logger.warning(

src/transformers/models/gptj/modeling_gptj.py

@@ -1001,9 +1001,10 @@ class GPTJForSequenceClassification(GPTJPreTrainedModel):
             sequence_lengths = -1
         else:
             if input_ids is not None:
-                sequence_lengths = (torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1).to(
-                    logits.device
-                )
+                # if no pad token found, use modulo instead of reverse indexing for ONNX compatibility
+                sequence_lengths = torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1
+                sequence_lengths = sequence_lengths % input_ids.shape[-1]
+                sequence_lengths = sequence_lengths.to(logits.device)
             else:
                 sequence_lengths = -1
                 logger.warning(

src/transformers/models/llama/modeling_llama.py

@@ -1370,9 +1370,10 @@ class LlamaForSequenceClassification(LlamaPreTrainedModel):
             sequence_lengths = -1
         else:
             if input_ids is not None:
-                sequence_lengths = (torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1).to(
-                    logits.device
-                )
+                # if no pad token found, use modulo instead of reverse indexing for ONNX compatibility
+                sequence_lengths = torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1
+                sequence_lengths = sequence_lengths % input_ids.shape[-1]
+                sequence_lengths = sequence_lengths.to(logits.device)
             else:
                 sequence_lengths = -1
 

src/transformers/models/mistral/modeling_mistral.py

@@ -1338,9 +1338,10 @@ class MistralForSequenceClassification(MistralPreTrainedModel):
             sequence_lengths = -1
         else:
             if input_ids is not None:
-                sequence_lengths = (torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1).to(
-                    logits.device
-                )
+                # if no pad token found, use modulo instead of reverse indexing for ONNX compatibility
+                sequence_lengths = torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1
+                sequence_lengths = sequence_lengths % input_ids.shape[-1]
+                sequence_lengths = sequence_lengths.to(logits.device)
             else:
                 sequence_lengths = -1
 

src/transformers/models/mixtral/modeling_mixtral.py

@@ -1518,9 +1518,10 @@ class MixtralForSequenceClassification(MixtralPreTrainedModel):
             sequence_lengths = -1
         else:
             if input_ids is not None:
-                sequence_lengths = (torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1).to(
-                    logits.device
-                )
+                # if no pad token found, use modulo instead of reverse indexing for ONNX compatibility
+                sequence_lengths = torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1
+                sequence_lengths = sequence_lengths % input_ids.shape[-1]
+                sequence_lengths = sequence_lengths.to(logits.device)
             else:
                 sequence_lengths = -1
 

src/transformers/models/mpt/modeling_mpt.py

@@ -729,7 +729,10 @@ class MptForSequenceClassification(MptPreTrainedModel):
             sequence_lengths = -1
         else:
             if input_ids is not None:
-                sequence_lengths = (torch.ne(input_ids, self.config.pad_token_id).sum(-1) - 1).to(logits.device)
+                # if no pad token found, use modulo instead of reverse indexing for ONNX compatibility
+                sequence_lengths = torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1
+                sequence_lengths = sequence_lengths % input_ids.shape[-1]
+                sequence_lengths = sequence_lengths.to(logits.device)
             else:
                 sequence_lengths = -1
                 logger.warning(

src/transformers/models/openai/modeling_openai.py

@@ -814,9 +814,10 @@ class OpenAIGPTForSequenceClassification(OpenAIGPTPreTrainedModel):
             sequence_lengths = -1
         else:
             if input_ids is not None:
-                sequence_lengths = (torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1).to(
-                    logits.device
-                )
+                # if no pad token found, use modulo instead of reverse indexing for ONNX compatibility
+                sequence_lengths = torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1
+                sequence_lengths = sequence_lengths % input_ids.shape[-1]
+                sequence_lengths = sequence_lengths.to(logits.device)
             else:
                 sequence_lengths = -1
                 logger.warning(

src/transformers/models/opt/modeling_opt.py

@@ -1294,9 +1294,10 @@ class OPTForSequenceClassification(OPTPreTrainedModel):
             sequence_lengths = -1
         else:
             if input_ids is not None:
-                sequence_lengths = (torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1).to(
-                    logits.device
-                )
+                # if no pad token found, use modulo instead of reverse indexing for ONNX compatibility
+                sequence_lengths = torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1
+                sequence_lengths = sequence_lengths % input_ids.shape[-1]
+                sequence_lengths = sequence_lengths.to(logits.device)
             else:
                 sequence_lengths = -1
                 logger.warning(

src/transformers/models/persimmon/modeling_persimmon.py

@@ -969,9 +969,10 @@ class PersimmonForSequenceClassification(PersimmonPreTrainedModel):
             sequence_lengths = -1
         else:
             if input_ids is not None:
-                sequence_lengths = (torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1).to(
-                    logits.device
-                )
+                # if no pad token found, use modulo instead of reverse indexing for ONNX compatibility
+                sequence_lengths = torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1
+                sequence_lengths = sequence_lengths % input_ids.shape[-1]
+                sequence_lengths = sequence_lengths.to(logits.device)
             else:
                 sequence_lengths = -1
 

src/transformers/models/phi/modeling_phi.py

@@ -1225,9 +1225,10 @@ class PhiForSequenceClassification(PhiPreTrainedModel):
             sequence_lengths = -1
         else:
             if input_ids is not None:
-                sequence_lengths = (torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1).to(
-                    logits.device
-                )
+                # if no pad token found, use modulo instead of reverse indexing for ONNX compatibility
+                sequence_lengths = torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1
+                sequence_lengths = sequence_lengths % input_ids.shape[-1]
+                sequence_lengths = sequence_lengths.to(logits.device)
             else:
                 sequence_lengths = -1