[BUG] Inconsistent inference and evaluation results of the XLNET-CLM even on the training set! about transformers4rec HOT 2 OPEN

def apply_mask_to_inputs_CLM(
        self,
        inputs: torch.Tensor,
        mask_schema: torch.Tensor,
        training: bool = False,
        testing: bool = False,
    ) -> torch.Tensor:
        if not training and not testing:
            # Replacing the inputs corresponding to padded items with a trainable embedding
            # To mimic training and evaluation masking strategy
            inputs = torch.where(
                mask_schema.unsqueeze(-1).bool(),
                inputs,
                self.masked_item_embedding.to(inputs.dtype),
            )
            return inputs
        
        # # shift sequence of interaction embeddings
        # pos_emb_inp = inputs[:, :-1]
        # # Adding a masked item in the sequence to return to the initial sequence.
        # pos_emb_inp = torch.cat(  # type: ignore
        #     [
        #         pos_emb_inp,
        #         torch.zeros(
        #             (pos_emb_inp.shape[0], 1, pos_emb_inp.shape[2]),
        #             dtype=pos_emb_inp.dtype,
        #         ).to(inputs.device),
        #     ],
        #     axis=1,
        # )

        pos_emb_inp = inputs
        pos_emb_inp_new = pos_emb_inp.clone()
        # Iterate over each row in the boolean tensor
        for i in range(mask_schema.shape[0]):
            # Find the index of the last True value in the row
            # If there's no True value, idx will be -1
            idx = (mask_schema[i].nonzero(as_tuple=True)[0]).max().item() if mask_schema[i].any() else -1
            # Replace corresponding item in feature tensor with a zero matrix
            if idx != -1:
                pos_emb_inp_new[i, idx] = torch.zeros(pos_emb_inp.shape[2], dtype=pos_emb_inp.dtype).to(inputs.device)

        pos_emb_inp = pos_emb_inp_new
        # Replacing the inputs corresponding to padded items with a trainable embedding
        pos_emb_inp = torch.where(
            mask_schema.unsqueeze(-1).bool(),
            pos_emb_inp,
            self.masked_item_embedding.to(pos_emb_inp.dtype),
        )
        return pos_emb_inp

# same inputs as before!
=========inference===============
PredictionOutput(predictions=(array([[ 4,  3,  7,  5,  6],
       [ 5,  7, 30,  3, 22],
       [17,  7,  6, 26, 11],
       [ 3,  4, 25,  7, 18],
       [71, 26, 24,  4,  3],
       [ 4,  5,  3,  7, 22]]), array([[ 8.456369 ,  5.8312187,  5.6498675,  5.1875997,  4.956415 ],
       [10.017425 ,  6.9653053,  6.9261403,  6.617892 ,  6.447962 ],
       [ 9.492199 ,  8.0066   ,  6.5381365,  6.097307 ,  5.9652367],
       [ 6.56517  ,  6.2221594,  5.269926 ,  5.1904283,  5.092497 ],
       [ 6.0176606,  5.6072693,  5.431714 ,  5.273334 ,  5.039026 ],
       [ 8.518238 ,  6.537841 ,  5.880717 ,  5.6688414,  5.0900187]],
      dtype=float32)), label_ids=None, metrics={'test_runtime': 2.0936, 'test_samples_per_second': 2.866, 'test_steps_per_second': 1.433})
100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [00:00<00:00, 14.08it/s]
=========evaluation===============
PredictionOutput(predictions=(array([[ 4,  7,  3,  6, 30],
       [ 5,  7, 30,  3,  4],
       [17,  7,  6, 26,  4],
       [ 3,  4, 25, 39,  7],
       [ 4, 24, 71, 26, 14],
       [ 4, 30,  5,  3,  7]]), array([[7.4311004, 4.9409456, 4.876472 , 4.848051 , 4.7532825],
       [7.4449983, 6.3823347, 6.3242016, 6.160487 , 5.6924496],
       [7.9830294, 7.4659915, 6.0589595, 5.7209487, 5.57537  ],
       [6.721603 , 5.4623756, 5.102635 , 4.9805446, 4.9469085],
       [4.9136906, 4.843173 , 4.796401 , 4.7077065, 4.5187464],
       [7.7905493, 5.1441355, 5.0995426, 5.0602183, 5.031133 ]],
      dtype=float32)), label_ids=array([ 4,  9,  7, 43, 23, 91]), metrics={'eval_/next-item/ndcg_at_5': 0.27182161808013916, 'eval_/next-item/ndcg_at_10': 0.27182161808013916, 'eval_/next-item/recall_at_5': 0.3333333432674408, 'eval_/next-item/recall_at_10': 0.3333333432674408, 'eval_/next-item/avg_precision_at_5': 0.25, 'eval_/next-item/avg_precision_at_10': 0.25, 'eval_/loss': 4.302186489105225, 'eval_runtime': 0.6978, 'eval_samples_per_second': 8.599, 'eval_steps_per_second': 4.299})