From 493a1d29dcbe43e786752c344292b1de9f2fb74b Mon Sep 17 00:00:00 2001
From: Yih-Dar <2521628+ydshieh@users.noreply.github.com>
Date: Fri, 10 Mar 2023 21:54:38 +0100
Subject: [PATCH] Revert "[GPT2] Propose fix for #21080 (#21853)"

This reverts commit a3fef89b2694fac4dd642a3f77d3e96d0c3df82a.
---
 .../modeling_decision_transformer.py          |  9 +-------
 src/transformers/models/gpt2/modeling_gpt2.py |  9 +-------
 tests/models/gpt2/test_modeling_gpt2.py       | 21 -------------------
 3 files changed, 2 insertions(+), 37 deletions(-)

diff --git a/src/transformers/models/decision_transformer/modeling_decision_transformer.py b/src/transformers/models/decision_transformer/modeling_decision_transformer.py
index d489be4f2d3fd8..0add713d8e878e 100755
--- a/src/transformers/models/decision_transformer/modeling_decision_transformer.py
+++ b/src/transformers/models/decision_transformer/modeling_decision_transformer.py
@@ -553,14 +553,7 @@ def forward(
             past_key_values = tuple([None] * len(self.h))
         else:
             past_length = past_key_values[0][0].size(-2)
-
-        if attention_mask is not None and len(attention_mask.shape) == 2 and position_ids is None:
-            # create position_ids on the fly for batch generation
-            position_ids = attention_mask.long().cumsum(-1) - 1
-            position_ids.masked_fill_(attention_mask == 0, 1)
-            if past_length > 0:
-                position_ids = position_ids[:, past_length : input_shape[-1] + past_length :]
-        elif position_ids is None:
+        if position_ids is None:
             position_ids = torch.arange(past_length, input_shape[-1] + past_length, dtype=torch.long, device=device)
             position_ids = position_ids.unsqueeze(0).view(-1, input_shape[-1])
 
diff --git a/src/transformers/models/gpt2/modeling_gpt2.py b/src/transformers/models/gpt2/modeling_gpt2.py
index 03a1a3c02e0bd5..bf665143148075 100644
--- a/src/transformers/models/gpt2/modeling_gpt2.py
+++ b/src/transformers/models/gpt2/modeling_gpt2.py
@@ -797,14 +797,7 @@ def forward(
             past_key_values = tuple([None] * len(self.h))
         else:
             past_length = past_key_values[0][0].size(-2)
-
-        if attention_mask is not None and len(attention_mask.shape) == 2 and position_ids is None:
-            # create position_ids on the fly for batch generation
-            position_ids = attention_mask.long().cumsum(-1) - 1
-            position_ids.masked_fill_(attention_mask == 0, 1)
-            if past_length > 0:
-                position_ids = position_ids[:, past_length : input_shape[-1] + past_length :]
-        elif position_ids is None:
+        if position_ids is None:
             position_ids = torch.arange(past_length, input_shape[-1] + past_length, dtype=torch.long, device=device)
             position_ids = position_ids.unsqueeze(0).view(-1, input_shape[-1])
 
diff --git a/tests/models/gpt2/test_modeling_gpt2.py b/tests/models/gpt2/test_modeling_gpt2.py
index be5445525b4e9a..09d828fd7f33a9 100644
--- a/tests/models/gpt2/test_modeling_gpt2.py
+++ b/tests/models/gpt2/test_modeling_gpt2.py
@@ -590,27 +590,6 @@ def test_batch_generation(self):
         self.assertTrue(batch_out_sentence_tt != batch_out_sentence)  # token_type_ids should change output
         self.assertListEqual(expected_output_sentence, [non_padded_sentence, padded_sentence])
 
-    @slow
-    def test_batch_forward(self):
-        tokenizer = GPT2Tokenizer.from_pretrained("gpt2")
-        tokenizer.padding_side = "left"
-
-        # This tokenizer has no pad token, so we have to set it in some way
-        # Define PAD Token = EOS Token = 50256
-        tokenizer.pad_token = tokenizer.eos_token
-
-        model = GPT2LMHeadModel.from_pretrained("gpt2", pad_token_id=tokenizer.eos_token_id)
-        sentences = ["Hello, my dog is a little bit of a mess. I'm not sure if he's"]
-        inputs = tokenizer(sentences, padding=True, return_tensors="pt")
-        logits = model(**inputs).logits[:, -1, :]
-        indexes = torch.argmax(logits).item()
-
-        inputs_padded = tokenizer(sentences, padding="max_length", max_length=30, return_tensors="pt")
-        logits_padded = model(**inputs_padded).logits[:, -1, :]
-        indexes_padded = torch.argmax(logits_padded).item()
-
-        self.assertTrue(indexes == indexes_padded)
-
     @slow
     def test_batch_generation_2heads(self):
         model = GPT2DoubleHeadsModel.from_pretrained("gpt2")