Add MultiHeadAttentionWithCache for self attention with cache fast path (facebookresearch#449)

Summary:
Pull Request resolved: facebookresearch#449

When doing self attention, an optimization is to combine the Q, K, V input projection matrices and do a single matmul, instead of 3. Adding this optimization for MHA with cache in a new module `MultiHeadSelfAttentionWithCache`.

Note: we are primarily using a new module to avoid breaking checkpoint BC with respect to `MultiHeadAttentionWithCache`. In the future, we should consolidate these MHA implementations.

Differential Revision: D48418780

fbshipit-source-id: 5ad930ff27a4b131f8ff1f097a4c9e1548efb587

Author: rohan-varma

tests/modules/layers/test_multi_head_attention.py

@@ -11,6 +11,7 @@
 from torchmultimodal.modules.layers.multi_head_attention import (
     MultiHeadAttentionWithCache,
     MultiHeadSelfAttention,
+    MultiHeadSelfAttentionWithCache,
 )
 
 
@@ -103,6 +104,13 @@ def multi_head_self_attn_use_cache(self, dim_q):
         mha.eval()
         return mha
 
+    @pytest.fixture
+    def multi_head_self_attn_module_with_cache(self, dim_q):
+        mha = MultiHeadSelfAttentionWithCache(dim_q, num_heads=2)
+        init_weights_with_constant(mha)
+        mha.eval()
+        return mha
+
     @pytest.fixture
     def multi_head_cross_attn(self, dim_q, dim_kv):
         mha = MultiHeadAttentionWithCache(dim_q, dim_kv, num_heads=2)
@@ -117,16 +125,7 @@ def multi_head_cross_attn_without_bias(self, dim_q, dim_kv):
         mha.eval()
         return mha
 
-    def test_multi_head_self_attention_use_cache(
-        self,
-        multi_head_self_attn_use_cache,
-        current_key_value,
-        past_key_value,
-        q,
-    ):
-        actual = multi_head_self_attn_use_cache(
-            q, q, q, past_key_value=(past_key_value, past_key_value), use_cache=True
-        )
+    def _assert_mha_self_attn_equal(self, actual, past_key_value, current_key_value):
         expected = torch.tensor(
             [
                 [
@@ -138,6 +137,7 @@ def test_multi_head_self_attention_use_cache(
         )
         assert_expected(actual.attn_output, expected, rtol=0, atol=1e-4)
         # Check that the cache is properly updated
+        torch.cat([past_key_value, current_key_value], dim=2)
         assert_expected(
             actual.past_key_value[0],
             torch.cat([past_key_value, current_key_value], dim=2),
@@ -147,6 +147,59 @@ def test_multi_head_self_attention_use_cache(
             torch.cat([past_key_value, current_key_value], dim=2),
         )
 
+    def test_multi_head_self_attention_use_cache(
+        self,
+        multi_head_self_attn_use_cache,
+        current_key_value,
+        past_key_value,
+        q,
+    ):
+        actual = multi_head_self_attn_use_cache(
+            q, q, q, past_key_value=(past_key_value, past_key_value), use_cache=True
+        )
+        self._assert_mha_self_attn_equal(actual, past_key_value, current_key_value)
+
+    def test_multi_head_self_attn_module_with_cache(
+        self,
+        multi_head_self_attn_module_with_cache,
+        current_key_value,
+        past_key_value,
+        q,
+    ):
+        actual = multi_head_self_attn_module_with_cache(
+            q, past_key_value=(past_key_value, past_key_value), use_cache=True
+        )
+        self._assert_mha_self_attn_equal(actual, past_key_value, current_key_value)
+
+    def test_multi_head_attention_with_cache_modules_equal(
+        self,
+        multi_head_self_attn_use_cache,
+        multi_head_self_attn_module_with_cache,
+        current_key_value,
+        past_key_value,
+        q,
+    ):
+        mha_with_cache_cls_output = multi_head_self_attn_use_cache(
+            q, q, q, past_key_value=(past_key_value, past_key_value), use_cache=True
+        )
+        sa_with_cache_cls_output = multi_head_self_attn_module_with_cache(
+            q, past_key_value=(past_key_value, past_key_value), use_cache=True
+        )
+        assert_expected(
+            mha_with_cache_cls_output.attn_output,
+            sa_with_cache_cls_output.attn_output,
+            rtol=0,
+            atol=1e-4,
+        )
+        assert_expected(
+            mha_with_cache_cls_output.past_key_value[0],
+            mha_with_cache_cls_output.past_key_value[0],
+        )
+        assert_expected(
+            mha_with_cache_cls_output.past_key_value[1],
+            mha_with_cache_cls_output.past_key_value[1],
+        )
+
     def test_multi_head_cross_attention(self, multi_head_cross_attn, q, kv):
         actual = multi_head_cross_attn(q, kv, kv)
         expected = torch.tensor(

torchmultimodal/modules/layers/multi_head_attention.py

@@ -77,6 +77,92 @@ def forward(
         return attn_out
 
 
+class MultiHeadSelfAttentionWithCache(nn.Module):
+    """
+    MultiHeadAttention module for self-attention(SA). Similar to MultiHeadAttentionWithCache,
+    but only supports self attention and uses a fast path where the query, key, and value projections
+    are batched into a single matrix multiplication as opposed to three separate matmuls.
+    This class supports a cache mechanism for decoders to store previous states through
+    "past_key_value".
+
+    Args:
+        embed_dim (int): query, key, value embedding dimension
+        num_heads (int): number of attention heads
+        dropout (float): dropout rate
+        add_bias (bool): if ``True``, adds a learnable bias to query, key, value input projection matrix.
+            Defaults to True.
+    """
+
+    def __init__(
+        self,
+        embed_dim: int,
+        num_heads: int,
+        dropout: float = 0.0,
+        add_bias: bool = True,
+    ) -> None:
+        super().__init__()
+        self.num_heads = num_heads
+        self.input_proj = nn.Linear(embed_dim, 3 * embed_dim, bias=add_bias)
+        self.output_proj = nn.Linear(embed_dim, embed_dim)
+        self.dropout = dropout
+
+    def forward(
+        self,
+        query: Tensor,
+        attn_mask: Optional[Tensor] = None,
+        past_key_value: Optional[Tuple[Tensor, Tensor]] = None,
+        is_causal: bool = False,
+        use_cache: bool = False,
+    ) -> Union[Tensor, MHAWithCacheOutput]:
+        """
+        Args:
+            query (Tensor): input query of shape bsz x target_seq_len x embed_dim
+            attn_mask (optional Tensor): Attention mask of shape bsz x num_heads x target_seq_len x source_seq_len.
+                Note that the num_heads dimension can equal 1 and the mask will be broadcasted to all heads.
+                Two types of masks are supported. A boolean mask where a value of True
+                indicates that the element *should* take part in attention.
+                A float mask of the same type as query, key, value that is added to the attention score.
+            past_key_value (optional tuple of tensors): cached key and value with the same shape of key, value inputs.
+                The size of tuple should be 2, where the first entry is for cached key and second entry is for cached value.
+            is_causal (bool): If true, does causal attention masking, attn_mask should be set to None if this is set to True
+                 is_causal is a hint that the mask is a causal mask, providing incorrect hints can result in incorrect execution.
+            use_cache (bool): whether to use cache for key and value tensors
+
+        Returns:
+            if use_cache is off, return attn_output tensor of shape bsz x seq_len x embed_dim;
+            otherwise return namedtuple with attn_output, cached key and value.
+        """
+        bsz = query.size(0)
+        embed_dim = query.size(-1)
+        head_dim = embed_dim // self.num_heads
+        projected_query = self.input_proj(query)
+        query, key, value = projected_query.chunk(3, dim=-1)
+
+        # bsz x seq_len x embed_dim => bsz x num_heads x seq_len x head_dim
+        query = query.view(bsz, -1, self.num_heads, head_dim).transpose(1, 2)
+        if key.size(0) != bsz or value.size(0) != bsz:
+            raise ValueError("key and value should have the same bsz as query.")
+        key = key.view(bsz, -1, self.num_heads, head_dim).transpose(1, 2)
+        value = value.view(bsz, -1, self.num_heads, head_dim).transpose(1, 2)
+
+        # concat key value with cached values
+        if past_key_value is not None:
+            key = torch.cat([past_key_value[0], key], dim=2)
+            value = torch.cat([past_key_value[1], value], dim=2)
+
+        # turn off causal attention inside scaled_dot_product_attention, we handle it separately with attn_mask.
+        attn = F.scaled_dot_product_attention(
+            query, key, value, attn_mask, self.dropout, is_causal
+        )
+        attn = attn.transpose(1, 2).reshape(bsz, -1, embed_dim)
+
+        # add dense layer after attention
+        attn_output = self.output_proj(attn)
+        if use_cache:
+            return MHAWithCacheOutput(attn_output, (key, value))
+        return attn_output
+
+
 class MultiHeadAttentionWithCache(nn.Module):
     """
     MultiHeadAttention module for both self-attention(SA) and cross-attention(CA).