[OPT] Run test in lower precision on GPU (huggingface#17353)

* [OPT] Run test only in half precision

* up

* up

* up

* up

* finish

* fix on GPU

* Update tests/models/opt/test_modeling_opt.py

src/transformers/models/opt/modeling_opt.py

@@ -17,9 +17,8 @@
 from typing import List, Optional, Tuple, Union
 
 import torch
-import torch.nn.functional as F
 import torch.utils.checkpoint
-from torch import Tensor, nn
+from torch import nn
 from torch.nn import CrossEntropyLoss
 
 from ...activations import ACT2FN
@@ -86,52 +85,28 @@ def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int]
     return inverted_mask.masked_fill(inverted_mask.bool(), torch.finfo(dtype).min)
 
 
-def make_positions(mask, padding_idx: int):
-    """Replace non-padding symbols with their position numbers.
-
-    Position numbers begin at padding_idx+1. Padding symbols are ignored.
-    """
-    # The series of casts and type-conversions here are carefully
-    # balanced to both work with ONNX export and XLA. In particular XLA
-    # prefers ints, cumsum defaults to output longs, and ONNX doesn't know
-    # how to handle the dtype kwarg in cumsum.
-    positions = (torch.cumsum(mask, dim=1).type_as(mask) * mask).long() + padding_idx
-    return positions
-
-
 class OPTLearnedPositionalEmbedding(nn.Embedding):
     """
-    This module learns positional embeddings up to a fixed maximum size. Padding ids are ignored by either offsetting
-    based on padding_idx or by setting padding_idx to None and ensuring that the appropriate position ids are passed to
-    the forward function.
+    This module learns positional embeddings up to a fixed maximum size.
     """
 
-    def __init__(self, num_embeddings: int, embedding_dim: int, padding_idx: int = 1):
-        super().__init__(num_embeddings, embedding_dim, padding_idx)
-        self.onnx_trace = False
-        if self.padding_idx is not None:
-            self.max_positions = self.num_embeddings - self.padding_idx - 1
-        else:
-            self.max_positions = self.num_embeddings
-
-    def forward(self, attention_mask: Tensor, positions: Optional[Tensor] = None):
-        # attention_masks is expected to be of size [batch_size x seq_len].
-        if not ((positions is None) or (self.padding_idx is None)):
-            raise ValueError("If positions is pre-computed then padding_idx should not be set.")
-
-        if positions is None:
-            attention_mask = attention_mask.long()
-            positions = make_positions(attention_mask, self.padding_idx)
-
-        return F.embedding(
-            positions,
-            self.weight,
-            self.padding_idx,
-            self.max_norm,
-            self.norm_type,
-            self.scale_grad_by_freq,
-            self.sparse,
-        )
+    def __init__(self, num_embeddings: int, embedding_dim: int):
+        # OPT is set up so that if padding_idx is specified then offset the embedding ids by 2
+        # and adjust num_embeddings appropriately. Other models don't have this hack
+        self.offset = 2
+        super().__init__(num_embeddings + self.offset, embedding_dim)
+
+    def forward(self, attention_mask: torch.LongTensor, past_key_values_length: int = 0):
+        """`input_ids_shape` is expected to be [bsz x seqlen]."""
+        attention_mask = attention_mask.long()
+
+        # create positions depending on attention_mask
+        positions = (torch.cumsum(attention_mask, dim=1).type_as(attention_mask) * attention_mask).long() - 1
+
+        # cut positions if `past_key_values_length` is > 0
+        positions = positions[:, past_key_values_length:]
+
+        return super().forward(positions + self.offset)
 
 
 # Copied from transformers.models.bart.modeling_bart.BartAttention with Bart->OPT
@@ -504,12 +479,7 @@ def __init__(self, config: OPTConfig):
         self.vocab_size = config.vocab_size
 
         self.embed_tokens = nn.Embedding(config.vocab_size, config.word_embed_proj_dim, self.padding_idx)
-
-        # OPT is set up so that if padding_idx is specified then offset the embedding ids by 2
-        if self.padding_idx is not None:
-            num_embeddings = config.max_position_embeddings + 2
-
-        self.embed_positions = OPTLearnedPositionalEmbedding(num_embeddings, config.hidden_size, self.padding_idx)
+        self.embed_positions = OPTLearnedPositionalEmbedding(config.max_position_embeddings, config.hidden_size)
 
         if config.word_embed_proj_dim != config.hidden_size:
             self.project_out = nn.Linear(config.hidden_size, config.word_embed_proj_dim, bias=False)
@@ -639,8 +609,7 @@ def forward(
         # embed positions
         if attention_mask is None:
             attention_mask = torch.ones(inputs_embeds.shape[:2], dtype=torch.bool, device=inputs_embeds.device)
-
-        positions = self.embed_positions(attention_mask)[:, past_key_values_length:, :]
+        pos_embeds = self.embed_positions(attention_mask, past_key_values_length)
 
         attention_mask = self._prepare_decoder_attention_mask(
             attention_mask, input_shape, inputs_embeds, past_key_values_length
@@ -649,8 +618,7 @@ def forward(
         if self.project_in is not None:
             inputs_embeds = self.project_in(inputs_embeds)
 
-        hidden_states = inputs_embeds + positions
-
+        hidden_states = inputs_embeds + pos_embeds
         hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
 
         # decoder layers
@@ -671,6 +639,7 @@ def forward(
             # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
             if output_hidden_states:
                 all_hidden_states += (hidden_states,)
+
             dropout_probability = random.uniform(0, 1)
             if self.training and (dropout_probability < self.layerdrop):
                 continue

tests/models/opt/test_modeling_opt.py

@@ -271,14 +271,25 @@ class OPTModelIntegrationTests(unittest.TestCase):
     def test_inference_no_head(self):
         model = OPTModel.from_pretrained("facebook/opt-350m").to(torch_device)
         input_ids = _long_tensor([[0, 31414, 232, 328, 740, 1140, 12695, 69, 46078, 1588, 2]])
+
         with torch.no_grad():
             output = model(input_ids=input_ids).last_hidden_state
+
         expected_shape = torch.Size((1, 11, 512))
         self.assertEqual(output.shape, expected_shape)
         expected_slice = torch.tensor(
-            [[-0.2867, -1.9256, -0.3062], [-1.2711, -0.1337, -0.1897], [0.4109, 0.1187, -1.3142]], device=torch_device
+            [[-0.2873, -1.9242, -0.3059], [-1.2738, -0.1333, -0.1877], [0.4116, 0.1192, -1.3107]],
+            device=torch_device,
         )
-        self.assertTrue(torch.allclose(output[:, :3, :3], expected_slice, atol=1e-3))
+        # Getting different logits results on GPU depending on PyTorch version (1.10+cu11.0 vs. 1.11+cu11.4)
+        # and results also differ between CPU and GPU. Only on CPU it seems to be deterministic.
+
+        # It's not because the weights are saved & loaded in FP16
+        # checked that the same happens when weights are stored in fp32 and loaded in fp32.
+        # The differences start to creep in in the first linear projection matrix project_in_dim
+        # It however also happens for BART (maybe related to training model in fp16?)
+        atol = 1e-2 if torch_device != "cpu" else 1e-3
+        assert_tensors_close(output[0, :3, :3], expected_slice, atol=atol)
 
 
 @require_torch