enable a QATWrapper for non-parameterized matmuls in BERT self attention

src/transformers/models/bert/modeling_bert.py

@@ -210,6 +210,22 @@ def forward(
         return embeddings
 
 
+class QATMatMul(nn.Module):
+    def __init__(self):
+        super().__init__()
+
+        # behaves like normal torch.matmul unless a SparseML QuantizationModifier
+        # is initialized
+        self.wrap_qat = True
+        self.qat_wrapper_kwargs = {
+            "num_inputs": 2,
+            "input_qconfigs": ["asymmetric", "symmetric"],
+        }
+
+    def forward(self, a: torch.Tensor, b: torch.Tensor):
+        return torch.matmul(a, b)
+
+
 class BertSelfAttention(nn.Module):
     def __init__(self, config):
         super().__init__()
@@ -227,6 +243,11 @@ def __init__(self, config):
         self.key = nn.Linear(config.hidden_size, self.all_head_size)
         self.value = nn.Linear(config.hidden_size, self.all_head_size)
 
+        # non-parameterized matmuls will behave as normal torch.matmul ops unless
+        # Quantization-Aware-Training is invoked
+        self.attention_scores_matmul = QATMatMul()
+        self.context_layer_matmul = QATMatMul()
+
         self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
         self.position_embedding_type = getattr(config, "position_embedding_type", "absolute")
         if self.position_embedding_type == "relative_key" or self.position_embedding_type == "relative_key_query":
@@ -288,7 +309,7 @@ def forward(
             past_key_value = (key_layer, value_layer)
 
         # Take the dot product between "query" and "key" to get the raw attention scores.
-        attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
+        attention_scores = self.attention_scores_matmul(query_layer, key_layer.transpose(-1, -2))
 
         if self.position_embedding_type == "relative_key" or self.position_embedding_type == "relative_key_query":
             seq_length = hidden_states.size()[1]
@@ -322,7 +343,7 @@ def forward(
         if head_mask is not None:
             attention_probs = attention_probs * head_mask
 
-        context_layer = torch.matmul(attention_probs, value_layer)
+        context_layer = self.context_layer_matmul(attention_probs, value_layer)
 
         context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
         new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)