gqa fuse attention qkv

paddlenlp/transformers/llama/modeling.py

@@ -588,17 +588,15 @@
         self.head_dim = self.hidden_size // config.num_attention_heads
 
         self.num_key_value_heads = config.num_key_value_heads
+        assert config.num_attention_heads // config.num_key_value_heads
         self.num_key_value_groups = config.num_attention_heads // config.num_key_value_heads
+        self.gqa_or_mqa = config.num_attention_heads != config.num_key_value_heads
 
         self.max_position_embeddings = config.max_position_embeddings
         self.seq_length = config.seq_length
         self.sequence_parallel = config.sequence_parallel
 
         self.fuse_attention_qkv = config.fuse_attention_qkv
-        if self.fuse_attention_qkv and config.num_attention_heads != config.num_key_value_heads:
-            raise ValueError(
-                f"fuse_attention_qkv can't be True when num_attention_heads {config.num_attention_heads}!= num_key_value_heads {config.num_key_value_heads}"
-            )
 
         self.kv_indices = None
         # Note that we will actually perform a recompute only if both enable_recompute and layerwise_recompute are set to True
@@ -615,6 +613,11 @@
             if self.num_key_value_heads % config.tensor_parallel_degree == 0:
                 self.num_key_value_heads = self.num_key_value_heads // config.tensor_parallel_degree
             else:
+                if self.fuse_attention_qkv:
+                    # TODO(Yuang): support fusion for kv when kv heads cannot be divided by mp
+                    raise ValueError(
+                        f"fuse_attention_qkv can't be True when num_key_value_heads {config.num_key_value_heads} % tensor_parallel_degree {config.tensor_parallel_degree} != 0"
+                    )
                 logger.warning(
                     f"Get num_key_value_heads: {self.num_key_value_heads}, can't split to tensor_parallel_degree: {config.tensor_parallel_degree}, so we don't spilt key value weight."
                 )
@@ -644,7 +647,7 @@
             if self.fuse_attention_qkv:
                 self.qkv_proj = ColumnParallelLinear(
                     self.hidden_size,
-                    3 * self.hidden_size,
+                    self.hidden_size + 2 * self.config.num_key_value_heads * self.head_dim,
                     has_bias=False,
                     gather_output=False,
                 )
@@ -684,7 +687,7 @@
             if self.fuse_attention_qkv:
                 self.qkv_proj = nn.Linear(
                     self.hidden_size,
-                    3 * self.hidden_size,
+                    self.hidden_size + 2 * self.config.num_key_value_heads * self.head_dim,
                     bias_attr=False,
                 )
             else:
@@ -771,12 +774,27 @@
 
         if self.fuse_attention_qkv:
             mix_layer = self.qkv_proj(hidden_states)
+            # NOTE for GQA attention fusion (compatible with MHA and MQA):
+            # The weight for qkv_proj is in shape like [hidden_size, hidden_size + 2 * num_kv_heads * head_dim].
+            # After the projection, the mix_layer is in shape like [b, s, hidden_size + 2 * num_kv_heads * head_dim].
+            # Reshape the mix_layer into a shape like [b, s, num_kv_heads, (num_groups + 2) * head_dim],
+            # where num_groups = num_q_heads // num_kv_heads.
+            # Split the mix_layer on the last axis into three sections [num_groups * head_dim, head_dim, head_dim]
+            # to represent the q, k and v respectively.
+            # The q is in the shape like [b, s, num_kv_heads, num_groups * head_dim].
+            # The k and v are in the shape like [b, s, num_kv_heads, head_dim].
+            # Under MHA, the q is ready for the following calculation since num_kv_heads == num_q_heads,
+            # But for the GQA or MQA, q should be reshaped into [b, s, num_q_heads, head_dim].
             if self.reshard_layer is not None:
                 if self.sequence_parallel:
                     assert self.seq_length % self.config.sep_parallel_degree == 0
                     mix_layer = paddle.reshape_(
                         mix_layer,
-                        [-1, self.seq_length // self.config.sep_parallel_degree, 3 * self.num_heads * self.head_dim],
+                        [
+                            -1,
+                            self.seq_length // self.config.sep_parallel_degree,
+                            self.num_heads * self.head_dim + 2 * self.num_key_value_heads * self.head_dim,
+                        ],
                     )
                 # [bs, seq_len / sep, num_head, head_dim] -> [bs, seq_len, num_head / sep, head_dim]
                 mix_layer = self.reshard_layer(
@@ -785,15 +803,26 @@
                     concat_axis=1,
                 )
                 mix_layer = paddle.reshape_(
-                    mix_layer, [0, self.seq_length, -1, 3 * self.head_dim]
+                    mix_layer, [0, self.seq_length, -1, (self.num_key_value_groups + 2) * self.head_dim]
                 )  # [bs, seq_len, num_head/k, 3*head_dim], k is sep degree
             else:
                 if self.sequence_parallel:
-                    target_shape = [-1, self.seq_length, self.num_heads, 3 * self.head_dim]
+                    target_shape = [
+                        -1,
+                        self.seq_length,
+                        self.num_key_value_heads,
+                        (self.num_key_value_groups + 2) * self.head_dim,
+                    ]
                 else:
-                    target_shape = [0, 0, self.num_heads, 3 * self.head_dim]
+                    target_shape = [0, 0, self.num_key_value_heads, (self.num_key_value_groups + 2) * self.head_dim]
                 mix_layer = paddle.reshape_(mix_layer, target_shape)
-            query_states, key_states, value_states = paddle.split(mix_layer, num_or_sections=3, axis=-1)
+            query_states, key_states, value_states = paddle.split(
+                mix_layer,
+                num_or_sections=[self.num_key_value_groups * self.head_dim, self.head_dim, self.head_dim],
+                axis=-1,
+            )
+            if self.gqa_or_mqa:
+                query_states = paddle.reshape_(query_states, [0, 0, self.num_heads, self.head_dim])
         else:
             query_states = self.q_proj(hidden_states)
             key_states = self.k_proj(hidden_states)
@@ -807,11 +836,19 @@
                     )
                     key_states = paddle.reshape(
                         key_states,
-                        [-1, self.seq_length // self.config.sep_parallel_degree, self.num_heads * self.head_dim],
+                        [
+                            -1,
+                            self.seq_length // self.config.sep_parallel_degree,
+                            self.num_key_value_heads * self.head_dim,
+                        ],
                     )
                     value_states = paddle.reshape(
                         value_states,
-                        [-1, self.seq_length // self.config.sep_parallel_degree, self.num_heads * self.head_dim],
+                        [
+                            -1,
+                            self.seq_length // self.config.sep_parallel_degree,
+                            self.num_key_value_heads * self.head_dim,
+                        ],
                     )
                 query_states = self.reshard_layer(
                     query_states,