Enable flag to not pass PAD tokens in ffwd (#775)

* Changing how attention_mask is being passed around

* adding in option to toggle flags for padding

* moving to flash attn import

* removing unused import

* Removing excess stuff from pyproject

* refactor

* Update llmfoundry/models/layers/blocks.py

Co-authored-by: Mihir Patel <[email protected]>

* Update llmfoundry/models/layers/blocks.py

Co-authored-by: Mihir Patel <[email protected]>

* Changing some naming conventions, moving where tests are done, adding numerics test

* Updating import

* trying to fix tests

* trying to fix tests

* updating tests

* updating tests

* Update llmfoundry/models/mpt/configuration_mpt.py

Co-authored-by: Daniel King <[email protected]>

* Changing gating in tests to check for flash attn

---------

Co-authored-by: Mihir Patel <[email protected]>
Co-authored-by: Daniel King <[email protected]>

llmfoundry/models/layers/blocks.py

@@ -12,6 +12,11 @@
 from llmfoundry.models.layers.ffn import FFN_CLASS_REGISTRY, build_ffn
 from llmfoundry.models.layers.norm import NORM_CLASS_REGISTRY
 
+try:
+    from flash_attn.bert_padding import unpad_input, pad_input  # type: ignore # yapf: disable # isort: skip
+except:
+    unpad_input, pad_input = None, None
+
 attn_config_defaults: Dict = {
     'attn_type': 'multihead_attention',
     'attn_pdrop': 0.0,
@@ -53,6 +58,7 @@ def __init__(
         fc_type: str = 'torch',
         device: Optional[str] = None,
         no_bias: bool = False,
+        use_pad_tok_in_ffn: bool = True,
         **kwargs: Any,
     ):
         if attn_config is None:
@@ -105,6 +111,8 @@ def __init__(
         self.resid_attn_dropout = nn.Dropout(resid_pdrop)
         self.resid_ffn_dropout = nn.Dropout(resid_pdrop)
 
+        self.use_pad_tok_in_ffn = use_pad_tok_in_ffn
+
     def forward(
         self,
         x: torch.Tensor,
@@ -132,6 +140,14 @@ def forward(
         m = x
         if self.norm_2 is not None:
             m = self.norm_2(x)
+        batch_size, seq_len = m.size()[:2]
+        indices = None
+        if not self.use_pad_tok_in_ffn:
+            assert unpad_input is not None
+            m, indices, _, _ = unpad_input(m, attention_mask)
         n = self.ffn(m)
+        if not self.use_pad_tok_in_ffn:
+            assert pad_input is not None
+            n = pad_input(n, indices, batch_size, seq_len)
         x = x + self.resid_ffn_dropout(n)
         return x, attn_weights, past_key_value

llmfoundry/models/mpt/configuration_mpt.py

@@ -60,6 +60,7 @@ def __init__(
         init_config: Dict = init_config_defaults,
         fc_type: str = 'torch',
         tie_word_embeddings: bool = True,
+        use_pad_tok_in_ffn: bool = True,
         verbose: Optional[int] = None,
         **kwargs: Any,
     ):
@@ -131,6 +132,7 @@ def __init__(
                 See llmfoundry.models.utils.param_init_fns.py for info on other param init config options
             fc_type (str): choose fc layer implementation. Options: torch and te. te layers support fp8 when using H100 GPUs.
             tie_word_embeddings (bool): Whether to tie the input embedding and output layers.
+            use_pad_tok_in_ffn (bool): Whether to forward the pad token in the feedforward networks.
         """
         self.d_model = d_model
         self.n_heads = n_heads
@@ -151,6 +153,7 @@ def __init__(
         self.use_cache = use_cache
         self.init_config = init_config
         self.fc_type = fc_type
+        self.use_pad_tok_in_ffn = use_pad_tok_in_ffn
         if verbose is not None:
             warnings.warn(
                 DeprecationWarning(
@@ -292,3 +295,10 @@ def _validate_config(self) -> None:
             self.ffn_config['fc_type'] = self.fc_type
         elif self.ffn_config['ffn_type'] == 'te_ln_mlp':
             self.ffn_config['bias'] = not self.no_bias
+        if not self.use_pad_tok_in_ffn:
+            try:
+                from flash_attn.bert_padding import unpad_input, pad_input  # type: ignore # yapf: disable # isort: skip
+            except:
+                raise ImportError(
+                    'In order to set `use_pad_tok_in_ffn=False`, please install flash-attn==1.0.9 or flash-attn==2.3.2'
+                )

llmfoundry/models/mpt/modeling_mpt.py

@@ -419,7 +419,7 @@ def _attn_bias(
             attn_bias = attn_bias.masked_fill(
                 ~attention_mask.view(-1, 1, 1, s_k), min_val)
 
-        return attn_bias, None
+        return attn_bias, attention_mask
 
     def _apply_prefix_mask(self, attn_bias: torch.Tensor,
                            prefix_mask: torch.Tensor) -> torch.Tensor:

tests/models/test_model.py

@@ -831,6 +831,45 @@ def test_forward_with_padding(attention_impl: str, pos_emb_config: dict,
                 batched_output[1, :],
                 atol=1e-6 if attention_impl == 'torch' else 1e-8)
 
+        try:
+            from flash_attn.bert_padding import unpad_input, pad_input  # type: ignore # yapf: disable # isort: skip
+        except:
+            unpad_input, pad_input = None, None
+
+        if unpad_input is not None and pad_input is not None:
+            # Checking numerical precision with pad_token ffn
+            for block in mpt.transformer.blocks:
+                # Flip the padding usage in the model
+                block.use_pad_tok_in_ffn = not block.use_pad_tok_in_ffn
+
+            right_padding_output_pad_flipped = mpt(
+                right_padding_input_ids,
+                attention_mask=right_padding_attention_mask).logits
+            middle_padding_output_pad_flipped = mpt(
+                middle_padding_input_ids,
+                attention_mask=middle_padding_attention_mask).logits
+            left_padding_output_pad_flipped = mpt(
+                left_padding_input_ids,
+                attention_mask=left_padding_attention_mask).logits
+
+            pad_vs_unpad_rtol = 1e-5
+            pad_vs_unpad_atol = 1e-6
+            assert torch.allclose(right_padding_output[0, :3],
+                                  right_padding_output_pad_flipped[0, :3],
+                                  rtol=pad_vs_unpad_rtol,
+                                  atol=pad_vs_unpad_atol)
+
+            assert torch.allclose(middle_padding_output[0, [0, 1, 5]],
+                                  middle_padding_output_pad_flipped[0,
+                                                                    [0, 1, 5]],
+                                  rtol=pad_vs_unpad_rtol,
+                                  atol=pad_vs_unpad_atol)
+
+            assert torch.allclose(left_padding_output[0, 3:],
+                                  left_padding_output_pad_flipped[0, 3:],
+                                  rtol=pad_vs_unpad_rtol,
+                                  atol=pad_vs_unpad_atol)
+
 
 @pytest.mark.parametrize('attention_impl', ['torch', 'triton'])
 def test_advanced_mask_building(attention_impl: str):