Update KV Cache to use num_kv_heads instead of num_heads (#1961)

Author: mirceamironenco

tests/torchtune/models/llama2/scripts/compare_fused_attention.py

@@ -256,7 +256,6 @@ def compare_attn(
     max_seq_len: int,
     use_kv_cache: bool,
 ):
-
     torch.manual_seed(16)
     inputs = torch.randn(4, 2048, 4096)
 
@@ -269,8 +268,9 @@ def compare_attn(
         kv_cache = KVCache(
             batch_size=4,
             max_seq_len=max_seq_len,
-            n_kv_heads=num_heads,
+            num_kv_heads=num_kv_heads,
             head_dim=head_dim,
+            dtype=inputs.dtype,
         )
     else:
         kv_cache = None
@@ -330,7 +330,6 @@ def compare_attn(
 
 
 if __name__ == "__main__":
-
     # compare mha
     mha = {
         "num_heads": 32,

tests/torchtune/models/llama2/scripts/compare_lora_attention.py

@@ -33,7 +33,6 @@ def compare_lora_attention(
     lora_rank: int,
     lora_alpha: float,
 ) -> None:
-
     # make sure we have the right seed for generating outputs
     # this should match up the seed value set in the corresponding
     # unit test
@@ -68,8 +67,9 @@ def compare_lora_attention(
         KVCache(
             batch_size=batch_size,
             max_seq_len=max_seq_len,
-            n_kv_heads=num_heads,
+            num_kv_heads=num_kv_heads,
             head_dim=head_dim,
+            dtype=x.dtype,
         )
         if batch_size is not None
         else None

tests/torchtune/modules/test_attention.py

@@ -123,7 +123,7 @@ def gqa_kv_cache(
         kv_cache = KVCache(
             batch_size=4,
             max_seq_len=max_seq_len,
-            num_heads=num_heads,
+            num_kv_heads=num_kv_heads,
             head_dim=head_dim,
             dtype=torch.float32,
         )
@@ -178,7 +178,7 @@ def mha_kv_cache(
         kv_cache = KVCache(
             batch_size=4,
             max_seq_len=max_seq_len,
-            num_heads=num_heads,
+            num_kv_heads=num_kv_heads,
             head_dim=head_dim,
             dtype=torch.float32,
         )
@@ -233,7 +233,7 @@ def mqa_kv_cache(
         kv_cache = KVCache(
             batch_size=4,
             max_seq_len=max_seq_len,
-            num_heads=num_heads,
+            num_kv_heads=num_kv_heads,
             head_dim=head_dim,
             dtype=torch.float32,
         )
@@ -267,7 +267,6 @@ def test_forward_gqa(self, input: torch.Tensor, gqa: MultiHeadAttention) -> None
     def test_forward_gqa_kv_cache(
         self, input: torch.Tensor, gqa_kv_cache: MultiHeadAttention, attn_params_gqa
     ) -> None:
-
         _, _, _, max_seq_len = attn_params_gqa
         _, seq_len, _ = input.shape
 
@@ -293,7 +292,6 @@ def test_forward_mha(self, input: torch.Tensor, mha: MultiHeadAttention) -> None
     def test_forward_mha_kv_cache(
         self, input: torch.Tensor, mha_kv_cache: MultiHeadAttention, attn_params_mha
     ) -> None:
-
         _, _, _, max_seq_len = attn_params_mha
         _, seq_len, _ = input.shape
 

torchtune/models/gemma2/_attention.py

@@ -149,7 +149,7 @@ def setup_cache(
             self.kv_cache = KVCache(
                 batch_size=batch_size,
                 max_seq_len=max_seq_len,
-                num_heads=self.num_heads,
+                num_kv_heads=self.num_heads,
                 head_dim=self.head_dim,
                 dtype=dtype,
             )
@@ -211,9 +211,9 @@ def forward(
             - h_d: head dim
         """
         #  until flex attention implementation exists, we do not accept block masks
-        if (mask is not None) and (type(mask) != torch.Tensor()):
+        if mask is not None and (not isinstance(mask, torch.Tensor)):
             raise NotImplementedError(
-                "Block masks are not implemeted yet, use packed=False"
+                "Block masks are not implemeted yet, use packed=False."
             )
 
         # x has shape [b, s_x, d]

torchtune/modules/attention.py

@@ -164,7 +164,7 @@ def setup_cache(
             self.kv_cache = KVCache(
                 batch_size=batch_size,
                 max_seq_len=max_seq_len,
-                num_heads=self.num_heads,
+                num_kv_heads=self.num_kv_heads,
                 head_dim=self.head_dim,
                 dtype=dtype,
             )
@@ -258,47 +258,37 @@ def forward(
         else:
             # Update k and v shape, positional embeddings, and normalization
 
-            # k has shape [b, s_y, num_kv_heads * head_dim]
-            # v has shape [b, s_y, num_kv_heads * head_dim]
+            # k,v shape [b, s_y, num_kv_heads * head_dim]
             k = self.k_proj(y)
             v = self.v_proj(y)
 
             # Apply positional embeddings
-            # k: [b, s_y, n_kv, h_d]
+            # k,v shape: [b, s_y, n_kv, h_d]
             k = k.view(b, s_y, -1, self.head_dim)
+            v = v.view(b, s_y, -1, self.head_dim)
             if self.pos_embeddings is not None:
                 k = self.pos_embeddings(k, input_pos=input_pos)
 
-            # View + expand + reshape bring num_kv_heads to num_heads for k and v
-            # to match q.
+            # k,v shape: [b, n_kv, s_y, h_d]
+            k = k.transpose(1, 2)
+            v = v.transpose(1, 2)
 
-            # k: [b, s_y, n_kv, 1, h_d]
-            # v: [b, s_y, n_kv, 1, h_d]
-            k = k.view(b, s_y, self.num_kv_heads, 1, self.head_dim)
-            v = v.view(b, s_y, self.num_kv_heads, 1, self.head_dim)
+            # Update key-value cache
+            if self.kv_cache is not None and self.cache_enabled:
+                k, v = self.kv_cache.update(k, v)
 
             # If needed, expand the key and value tensors to have the same shape
             # as the query tensor by copying values across the relevant dim
+            # k,v shape: [b, n_h, s, h_d]
             if self.num_heads != self.num_kv_heads:
-                k = k.expand(b, s_y, self.num_kv_heads, q_per_kv, self.head_dim)
-                v = v.expand(b, s_y, self.num_kv_heads, q_per_kv, self.head_dim)
-
-            # [b, s, n_h, h_d]
-            k = k.reshape(b, s_y, -1, self.head_dim)
-            v = v.reshape(b, s_y, -1, self.head_dim)
-
-            # [b, n_h, s, h_d]
-            k = k.transpose(1, 2)
-            v = v.transpose(1, 2)
+                expand_shape = (-1, -1, q_per_kv, -1, -1)
+                k = k.unsqueeze(2).expand(expand_shape).flatten(1, 2)
+                v = v.unsqueeze(2).expand(expand_shape).flatten(1, 2)
 
             # Normalize k
             if self.k_norm is not None:
                 k = self.k_norm(k)
 
-            # Update key-value cache
-            if self.kv_cache is not None and self.cache_enabled:
-                k, v = self.kv_cache.update(k, v)
-
         output = self._attention_call(
             q,
             k,

torchtune/modules/kv_cache.py

@@ -17,9 +17,7 @@ class KVCache(nn.Module):
     Args:
         batch_size (int): batch size model will be run with
         max_seq_len (int): maximum sequence length model will be run with
-        num_heads (int): number of heads. We take num_heads instead of num_kv_heads because
-            the cache is created after we've expanded the key and value tensors to have the
-            same shape as the query tensor. See attention.py for more details
+        num_kv_heads (int): number of key/value heads.
         head_dim (int): per-attention head embedding dimension
         dtype (torch.dtype): dtype for the caches
     """
@@ -28,12 +26,12 @@ def __init__(
         self,
         batch_size: int,
         max_seq_len: int,
-        num_heads: int,
+        num_kv_heads: int,
         head_dim: int,
         dtype: torch.dtype,
     ) -> None:
         super().__init__()
-        cache_shape = (batch_size, num_heads, max_seq_len, head_dim)
+        cache_shape = (batch_size, num_kv_heads, max_seq_len, head_dim)
         self.register_buffer(
             "k_cache", torch.zeros(cache_shape, dtype=dtype), persistent=False
         )
@@ -66,7 +64,7 @@ def update(
             already been filled, use ``.reset()``, which will reset the cache to the zero-th position.
 
         Example:
-            >>> cache = KVCache(batch_size=2, max_seq_len=16, num_heads=4, head_dim=32, dtype=torch.bfloat16)
+            >>> cache = KVCache(batch_size=2, max_seq_len=16, num_kv_heads=4, head_dim=32, dtype=torch.bfloat16)
             >>> keys, values = torch.ones((2, 4, 8, 32)), torch.ones((2, 4, 8, 32))
             >>> cache.update(keys, values)
             >>> # now positions 0 through 7 are filled