ai-edge-torch-nightly 0.3.0.dev20240910__py3-none-any.whl → 0.3.0.dev20240911__py3-none-any.whl

ai_edge_torch/generative/examples/test_models/toy_model_with_kv_cache.py

@@ -12,14 +12,17 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 # ==============================================================================
-# A toy example which has basic transformer block (w/ KV-Cache).
+
+"""A toy example which has basic transformer block (w/ externalized KV-Cache)."""
+
 from typing import Tuple
 
 import ai_edge_torch
 from ai_edge_torch import lowertools
-from ai_edge_torch.generative.layers.attention import TransformerBlock
+from ai_edge_torch.generative.layers import attention
+from ai_edge_torch.generative.layers import builder
+from ai_edge_torch.generative.layers import kv_cache as kv_utils
 import ai_edge_torch.generative.layers.attention_utils as attn_utils
-import ai_edge_torch.generative.layers.builder as builder
 import ai_edge_torch.generative.layers.model_config as cfg
 import torch
 import torch.nn as nn
@@ -27,7 +30,7 @@ import torch.nn as nn
 RoPECache = Tuple[torch.Tensor, torch.Tensor]
 
 
-class ToyModelWithKV(torch.nn.Module):
+class ToyModelWithKVCache(torch.nn.Module):
 
   def __init__(self, config: cfg.ModelConfig) -> None:
     super().__init__()
@@ -36,7 +39,7 @@ class ToyModelWithKV(torch.nn.Module):
     )
     self.tok_embedding = nn.Embedding(config.vocab_size, config.embedding_dim)
     self.transformer_blocks = nn.ModuleList(
-        TransformerBlock(config) for _ in range(config.num_layers)
+        attention.TransformerBlock(config) for _ in range(config.num_layers)
     )
     self.final_norm = builder.build_norm(
         config.embedding_dim,
@@ -57,18 +60,29 @@ class ToyModelWithKV(torch.nn.Module):
     )
     self.config = config
 
-  @torch.inference_mode
-  def forward(self, idx: torch.Tensor, input_pos: torch.Tensor) -> torch.Tensor:
-    x = self.tok_embedding(idx)
+  def forward(
+      self,
+      tokens: torch.Tensor,
+      input_pos: torch.Tensor,
+      kv_cache: kv_utils.KVCache,
+  ) -> Tuple[torch.Tensor, kv_utils.KVCache]:
+    x = self.tok_embedding(tokens)
     cos, sin = self.rope_cache
     cos = cos.index_select(0, input_pos)
     sin = sin.index_select(0, input_pos)
     mask = self.mask_cache.index_select(2, input_pos)
     mask = mask[:, :, :, : self.config.max_seq_len]
+
+    updated_kv_entires = []
     for i, block in enumerate(self.transformer_blocks):
-      x = block(x, (cos, sin), mask, input_pos)
+      kv_entry = kv_cache.caches[i] if kv_cache else None
+      x, kv_entry = block(x, (cos, sin), mask, input_pos, kv_entry)
+      if kv_entry:
+        updated_kv_entires.append(kv_entry)
+
     x = self.final_norm(x)
-    return self.lm_head(x)
+    updated_kv_cache = kv_utils.KVCache(tuple(updated_kv_entires))
+    return {'logits': self.lm_head(x), 'kv_cache': updated_kv_cache}
 
 
 def _export_stablehlo_mlir(model, args):
@@ -89,7 +103,7 @@ def get_model_config() -> cfg.ModelConfig:
   config = cfg.ModelConfig(
       vocab_size=150,
       num_layers=2,
-      max_seq_len=500,
+      max_seq_len=100,
       embedding_dim=128,
       attn_config=attn_config,
       ff_config=ff_config,
@@ -102,40 +116,59 @@ def get_model_config() -> cfg.ModelConfig:
 
 
 def get_sample_prefill_inputs() -> Tuple[torch.Tensor, torch.Tensor]:
-  idx = torch.unsqueeze(torch.arange(0, 100), 0)
+  tokens = torch.unsqueeze(torch.arange(0, 100), 0)
   input_pos = torch.arange(0, 100)
-  return idx, input_pos
+  return tokens, input_pos
 
 
 def get_sample_decode_inputs() -> Tuple[torch.Tensor, torch.Tensor]:
-  idx = torch.tensor([[1]], dtype=torch.long)
-  input_pos = torch.tensor([10], dtype=torch.int64)
-  return idx, input_pos
+  tokens = torch.tensor([[1]], dtype=torch.long)
+  input_pos = torch.tensor([10])
+  return tokens, input_pos
 
 
 def define_and_run() -> None:
   dump_mlir = False
 
   config = get_model_config()
-  model = ToyModelWithKV(config)
+  model = ToyModelWithExternalKV(config)
+  model.eval()
   print('running an inference')
-  idx, input_pos = get_sample_prefill_inputs()
-  decode_idx, decode_input_pos = get_sample_decode_inputs()
-  print(model.forward(idx, input_pos))
+  kv = kv_utils.KVCache.from_model_config(config)
+
+  tokens, input_pos = get_sample_prefill_inputs()
+  decode_token, decode_input_pos = get_sample_decode_inputs()
+  print(model.forward(tokens, input_pos, kv))
 
   if dump_mlir:
-    mlir_text = _export_stablehlo_mlir(model, (idx, input_pos))
-    with open('/tmp/toy_model_with_kv.stablehlo.mlir', 'w') as f:
+    mlir_text = _export_stablehlo_mlir(model, (tokens, input_pos, kv))
+    with open('/tmp/toy_model_with_external_kv.stablehlo.mlir', 'w') as f:
       f.write(mlir_text)
 
   # Convert model to tflite with 2 signatures (prefill + decode).
   print('converting toy model to tflite with 2 signatures (prefill + decode)')
   edge_model = (
-      ai_edge_torch.signature('prefill', model, (idx, input_pos))
-      .signature('decode', model, (decode_idx, decode_input_pos))
+      ai_edge_torch.signature(
+          'prefill',
+          model,
+          sample_kwargs={
+              'tokens': tokens,
+              'input_pos': input_pos,
+              'kv_cache': kv,
+          },
+      )
+      .signature(
+          'decode',
+          model,
+          sample_kwargs={
+              'tokens': decode_token,
+              'input_pos': decode_input_pos,
+              'kv_cache': kv,
+          },
+      )
       .convert()
   )
-  edge_model.export('/tmp/toy_kv_cache.tflite')
+  edge_model.export('/tmp/toy_external_kv_cache.tflite')
 
 
 if __name__ == '__main__':

ai_edge_torch/generative/examples/tiny_llama/convert_to_tflite.py

@@ -13,11 +13,14 @@
 # limitations under the License.
 # ==============================================================================
 
+"""Example of converting TinyLlama model to multi-signature tflite model."""
+
 import os
-from pathlib import Path
+import pathlib
 
 import ai_edge_torch
 from ai_edge_torch.generative.examples.tiny_llama import tiny_llama
+from ai_edge_torch.generative.layers import kv_cache as kv_utils
 from ai_edge_torch.generative.quantize import quant_recipes
 import torch
 
@@ -48,20 +51,36 @@ def convert_tiny_llama_to_tflite(
   prefill_input_pos = torch.arange(0, prefill_seq_len)
   decode_token = torch.tensor([[0]], dtype=torch.long)
   decode_input_pos = torch.tensor([0], dtype=torch.int64)
+  kv = kv_utils.KVCache.from_model_config(pytorch_model.config)
 
   quant_config = quant_recipes.full_int8_dynamic_recipe() if quantize else None
   edge_model = (
       ai_edge_torch.signature(
-          'prefill', pytorch_model, (prefill_tokens, prefill_input_pos)
+          'prefill',
+          pytorch_model,
+          sample_kwargs={
+              'tokens': prefill_tokens,
+              'input_pos': prefill_input_pos,
+              'kv_cache': kv,
+          },
+      )
+      .signature(
+          'decode',
+          pytorch_model,
+          sample_kwargs={
+              'tokens': decode_token,
+              'input_pos': decode_input_pos,
+              'kv_cache': kv,
+          },
       )
-      .signature('decode', pytorch_model, (decode_token, decode_input_pos))
       .convert(quant_config=quant_config)
   )
+  quant_suffix = 'q8' if quantize else 'f32'
   edge_model.export(
-      f'/tmp/tiny_llama_seq{prefill_seq_len}_kv{kv_cache_max_len}.tflite'
+      f'/tmp/tiny_llama_{quant_suffix}_seq{prefill_seq_len}_ekv{kv_cache_max_len}.tflite'
   )
 
 
 if __name__ == '__main__':
-  checkpoint_path = os.path.join(Path.home(), 'Downloads/llm_data/tiny_llama')
-  convert_tiny_llama_to_tflite(checkpoint_path)
+  path = os.path.join(pathlib.Path.home(), 'Downloads/llm_data/tiny_llama')
+  convert_tiny_llama_to_tflite(path)

ai_edge_torch/generative/examples/tiny_llama/tiny_llama.py

@@ -12,13 +12,15 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 # ==============================================================================
-# Example of building a TinyLlama model from the Edge Generative API layers.
+
+"""Example of building a TinyLlama model."""
 
 import os
-from pathlib import Path
+import pathlib
 
 from ai_edge_torch.generative.layers import attention
 from ai_edge_torch.generative.layers import builder
+from ai_edge_torch.generative.layers import kv_cache as kv_utils
 import ai_edge_torch.generative.layers.attention_utils as attn_utils
 import ai_edge_torch.generative.layers.model_config as cfg
 import ai_edge_torch.generative.utilities.loader as loading_utils
@@ -80,16 +82,22 @@ class TinyLLamma(nn.Module):
     )
     self.config = config
 
-  # The model's forward function takes in additional k/v cache tensors
-  # and returns the updated k/v cache tensors to the caller.
-  # This can be eliminated if we handle k/v cache updates inside the model itself.
   @torch.inference_mode
-  def forward(self, idx: torch.Tensor, input_pos: torch.Tensor) -> torch.Tensor:
-    _, seq_len = idx.size()
+  def forward(
+      self,
+      tokens: torch.Tensor,
+      input_pos: torch.Tensor,
+      kv_cache: kv_utils.KVCache,
+  ) -> dict[torch.Tensor, kv_utils.KVCache]:
+    _, seq_len = tokens.size()
     assert self.config.max_seq_len >= seq_len, (
         f"Cannot forward sequence of length {seq_len}, max seq length is only"
         f" {self.config.max_seq_len}"
     )
+    assert len(self.transformer_blocks) == len(kv_cache.caches), (
+        "The number of transformer blocks and the number of KV cache entries"
+        " must be the same."
+    )
 
     cos, sin = self.rope_cache
     cos = cos.index_select(0, input_pos)
@@ -97,16 +105,20 @@ class TinyLLamma(nn.Module):
     mask = self.mask_cache.index_select(2, input_pos)
     mask = mask[:, :, :, : self.config.kv_cache_max]
 
-    # forward the model itself
-    x = self.tok_embedding(idx)  # token embeddings of shape (b, t, n_embd)
+    # token embeddings of shape (b, t, n_embd)
+    x = self.tok_embedding(tokens)
 
-    for _, block in enumerate(self.transformer_blocks):
-      x = block(x, (cos, sin), mask, input_pos)
+    updated_kv_entires = []
+    for i, block in enumerate(self.transformer_blocks):
+      kv_entry = kv_cache.caches[i] if kv_cache else None
+      x, kv_entry = block(x, (cos, sin), mask, input_pos, kv_entry)
+      if kv_entry:
+        updated_kv_entires.append(kv_entry)
+    updated_kv_cache = kv_utils.KVCache(tuple(updated_kv_entires))
 
     x = self.final_norm(x)
-
-    res = self.lm_head(x)  # (b, t, vocab_size)
-    return res
+    logits = self.lm_head(x)  # (b, t, vocab_size)
+    return {"logits": logits, "kv_cache": updated_kv_cache}
 
 
 def get_model_config(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
@@ -147,8 +159,8 @@ def get_model_config(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
   return config
 
 
-def get_fake_model_config() -> cfg.ModelConfig:
-  config = get_model_config()
+def get_fake_model_config(**kwargs) -> cfg.ModelConfig:
+  config = get_model_config(**kwargs)
   config.vocab_size = 128
   config.num_layers = 2
   config.ff_config.intermediate_size = 64
@@ -160,26 +172,30 @@ def build_model(checkpoint_path: str, **kwargs) -> nn.Module:
   model = TinyLLamma(config)
   loader = loading_utils.ModelLoader(checkpoint_path, TENSOR_NAMES)
   loader.load(model)
+  model.eval()
   return model
 
 
-def define_and_run() -> None:
+def define_and_run(checkpoint_path: str) -> None:
   """Instantiates and runs a TinyLlama model."""
 
-  current_dir = Path(__file__).parent.resolve()
+  current_dir = pathlib.Path(__file__).parent.resolve()
   tiny_llama_goldens = torch.load(current_dir / "tiny_llama_lm_logits.pt")
   kv_cache_max_len = 1024
-  checkpoint_path = os.path.join(Path.home(), "Downloads/llm_data/tiny_llama")
   model = build_model(checkpoint_path, kv_cache_max_len=kv_cache_max_len)
   idx = torch.from_numpy(np.array([[1, 2, 3, 4]]))
   tokens = torch.full((1, kv_cache_max_len), 0, dtype=torch.long, device="cpu")
   tokens[0, :4] = idx
   input_pos = torch.arange(0, kv_cache_max_len)
-  lm_logits = model.forward(tokens, input_pos)
+  kv = kv_utils.KVCache.from_model_config(model.config)
+  output = model.forward(tokens, input_pos, kv)
   assert torch.allclose(
-      tiny_llama_goldens, lm_logits[0, idx.shape[1] - 1, :], atol=1e-05
+      tiny_llama_goldens, output["logits"][0, idx.shape[1] - 1, :], atol=1e-02
   )
 
 
 if __name__ == "__main__":
-  define_and_run()
+  input_checkpoint_path = os.path.join(
+      pathlib.Path.home(), "Downloads/llm_data/tiny_llama"
+  )
+  define_and_run(input_checkpoint_path)

ai_edge_torch/generative/layers/attention.py

@@ -12,16 +12,16 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 # ==============================================================================
-# Common building blocks for Attention layer.
 
-from typing import Optional, Tuple
+"""Common building blocks for Attention layer."""
 
-import ai_edge_torch.generative.layers.builder as builder
-from ai_edge_torch.generative.layers.kv_cache import KVCache
+from typing import Optional, Tuple, Union
+
+from ai_edge_torch.generative.layers import builder
+from ai_edge_torch.generative.layers import kv_cache as kv_utils
+from ai_edge_torch.generative.layers import scaled_dot_product_attention as sdpa
 import ai_edge_torch.generative.layers.model_config as cfg
 import ai_edge_torch.generative.layers.rotary_position_embedding as rotary_pos_emb
-from ai_edge_torch.generative.layers.scaled_dot_product_attention import scaled_dot_product_attention  # NOQA
-from ai_edge_torch.generative.layers.scaled_dot_product_attention import scaled_dot_product_attention_with_hlfb  # NOQA
 import torch
 from torch import nn
 
@@ -62,7 +62,6 @@ class TransformerBlock(nn.Module):
       config (cfg.ModelConfig): the configuration object for this transformer
         block.
     """
-
     super().__init__()
     self.pre_atten_norm = builder.build_norm(
         config.embedding_dim, config.pre_attention_norm_config
@@ -71,7 +70,6 @@ class TransformerBlock(nn.Module):
         config.batch_size,
         config.embedding_dim,
         config.attn_config,
-        config.kv_cache_max,
         config.enable_hlfb,
     )
     self.post_atten_norm = builder.build_norm(
@@ -86,7 +84,8 @@ class TransformerBlock(nn.Module):
       rope: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
       mask: Optional[torch.Tensor] = None,
       input_pos: Optional[torch.Tensor] = None,
-  ) -> torch.Tensor:
+      kv_cache: kv_utils.KVCacheEntry = None,
+  ) -> Union[torch.Tensor, Tuple[torch.Tensor, kv_utils.KVCacheEntry]]:
     """Forward function of the TransformerBlock.
 
     Args:
@@ -94,24 +93,34 @@ class TransformerBlock(nn.Module):
       rope (Tuple[torch.Tensor, torch.Tensor]): the input rope tensor.
       mask (torch.Tensor): the optional mask tensor.
       input_pos (torch.Tensor): the optional input position tensor.
+      kv_cache (KVCacheEntry): the optional kv cache entry.
 
     Returns:
-      output activation from this transformer block.
+      output activation from this transformer block, and updated kv cache (if
+      passed in).
     """
-
+    kv = None
     if self.config.parallel_residual:
       x_norm = self.pre_atten_norm(x)
-      attn_out = self.atten_func(x_norm, rope, mask, input_pos)
+      atten_func_out = self.atten_func(x_norm, rope, mask, input_pos, kv_cache)
+      if kv_cache is None:
+        attn_out = atten_func_out
+      else:
+        attn_out, kv = atten_func_out
       ff_out = self.ff(x_norm)
       output = x + attn_out + ff_out
     else:
       x_norm = self.pre_atten_norm(x)
-      attn_out = self.atten_func(x_norm, rope, mask, input_pos)
+      atten_func_out = self.atten_func(x_norm, rope, mask, input_pos, kv_cache)
+      if kv_cache is None:
+        attn_out = atten_func_out
+      else:
+        attn_out, kv = atten_func_out
       x = x + attn_out
       x_norm = self.post_atten_norm(x)
       output = x + self.ff(x_norm)
 
-    return output
+    return output if kv is None else (output, kv)
 
 
 class CausalSelfAttention(nn.Module):
@@ -121,7 +130,6 @@ class CausalSelfAttention(nn.Module):
       batch_size: int,
       dim: int,
       config: cfg.AttentionConfig,
-      kv_cache_max: int,
       enable_hlfb: bool,
   ) -> None:
     """Initialize an instance of CausalSelfAttention.
@@ -130,8 +138,6 @@ class CausalSelfAttention(nn.Module):
       batch_size (int): batch size of the input tensor.
       dim (int): causal attention's input/output dimmension.
       config (cfg.AttentionConfig): attention specific configurations.
-      kv_cache_max (int): determines the size of the KV Cache buffer, if
-        enabled.
       enable_hlfb (bool): whether hlfb is enabled or not.
     """
     super().__init__()
@@ -147,21 +153,13 @@ class CausalSelfAttention(nn.Module):
     self.output_projection = nn.Linear(
         output_shape, dim, bias=config.output_proj_use_bias
     )
-
-    # Build a k/v cache with size (batch_size, kv_cache_max, n_heads, head_dim).
-    if config.enable_kv_cache:
-      self.kv_cache = KVCache(
-          batch_size,
-          kv_cache_max,
-          config.num_query_groups,
-          config.head_dim,
-          enable_hlfb,
-      )
-
-    if enable_hlfb:
-      self.sdpa_func = scaled_dot_product_attention_with_hlfb
-    else:
-      self.sdpa_func = scaled_dot_product_attention
+    self.config = config
+    self.enable_hlfb = enable_hlfb
+    self.sdpa_func = (
+        sdpa.scaled_dot_product_attention_with_hlfb
+        if enable_hlfb
+        else sdpa.scaled_dot_product_attention
+    )
 
   def forward(
       self,
@@ -169,7 +167,8 @@ class CausalSelfAttention(nn.Module):
       rope: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
       mask: Optional[torch.Tensor] = None,
       input_pos: Optional[torch.Tensor] = None,
-  ) -> torch.Tensor:
+      kv_cache: Optional[kv_utils.KVCacheEntry] = None,
+  ) -> Union[torch.Tensor, Tuple[torch.Tensor, kv_utils.KVCacheEntry]]:
     """Forward function of the CausalSelfAttention layer, which can support
 
        MQA, GQA and MHA.
@@ -179,9 +178,11 @@ class CausalSelfAttention(nn.Module):
       rope (Tuple[torch.Tensor, torch.Tensor]): the input rope tensor.
       mask (torch.Tensor): the optional mask tensor.
       input_pos (torch.Tensor): the optional input position tensor.
+      kv_cache (KVCacheEntry): The KV cache entry corresponding to this module.
 
     Returns:
-      output activation from this self attention layer.
+      output activation from this self attention layer, and the updated
+        KV Cach Entry (if passed in).
     """
     # Batch size, sequence length, embedding dimensionality.
     B, T, E = x.size()
@@ -224,9 +225,11 @@ class CausalSelfAttention(nn.Module):
     n_elem = int(self.config.rotary_percentage * self.config.head_dim)
     q, k = _embed_rope(q, k, n_elem, rope)
 
-    if self.kv_cache is not None:
-      # TODO(haoliang): Handle when execeeding max sequence length.
-      k, v = self.kv_cache.update_cache(input_pos, k, v)
+    if kv_cache is not None:
+      kv_cache = kv_utils.update(
+          kv_cache, input_pos, k, v, enable_hlfb=self.enable_hlfb
+      )
+      k, v = kv_cache.k_cache, kv_cache.v_cache
 
     y = self.sdpa_func(
         q,
@@ -240,7 +243,7 @@ class CausalSelfAttention(nn.Module):
 
     # Compute the output projection.
     y = self.output_projection(y)
-    return y
+    return y if kv_cache is None else (y, kv_cache)
 
 
 class SelfAttention(CausalSelfAttention):
@@ -251,16 +254,19 @@ class SelfAttention(CausalSelfAttention):
       x: torch.Tensor,
       rope: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
       input_pos: Optional[torch.Tensor] = None,
-  ) -> torch.Tensor:
+      kv_cache: Optional[kv_utils.KVCacheEntry] = None,
+  ) -> Union[torch.Tensor, Tuple[torch.Tensor, kv_utils.KVCacheEntry]]:
     """Forward function of the SelfAttention layer, which can support MQA, GQA and MHA.
 
     Args:
       x (torch.Tensor): the input tensor.
       rope (Tuple[torch.Tensor, torch.Tensor]): the input rope tensor.
       input_pos (torch.Tensor): the optional input position tensor.
+      kv_cache (KVCacheEntry): The KV cache entry corresponding to this module.
 
     Returns:
-      output activation from this self attention layer.
+      output activation from this self attention layer, and the updated
+        KV Cach Entry (if passed in).
     """
     B, T, _ = x.size()
     return super().forward(
@@ -279,9 +285,8 @@ class CrossAttention(nn.Module):
       query_dim: int,
       cross_dim: int,
       config: cfg.AttentionConfig,
-      kv_cache_max: int,
       enable_hlfb: bool,
-  ) -> None:
+  ):
     """Initialize an instance of CrossAttention.
 
     Args:
@@ -289,8 +294,6 @@ class CrossAttention(nn.Module):
       query_dim (int): query tensor's dimension.
       cross_dim (int): cross attention's dimensions, for key and value tensors.
       config (cfg.AttentionConfig): attention specific configurations.
-      kv_cache_max (int): determines the size of the KV Cache buffer, if
-        enabled.
       enable_hlfb (bool): whether hlfb is enabled or not.
     """
     super().__init__()
@@ -309,21 +312,11 @@ class CrossAttention(nn.Module):
         query_dim, query_dim, bias=config.output_proj_use_bias
     )
 
-    self.kv_cache = None
-    # Build a k/v cache with size (batch_size, kv_cache_max, n_heads, head_dim).
-    if config.enable_kv_cache:
-      self.kv_cache = KVCache(
-          batch_size,
-          kv_cache_max,
-          config.num_query_groups,
-          self.config.head_dim,
-          enable_hlfb,
-      )
-
-    if enable_hlfb:
-      self.sdpa_func = scaled_dot_product_attention_with_hlfb
-    else:
-      self.sdpa_func = scaled_dot_product_attention
+    self.sdpa_func = (
+        sdpa.scaled_dot_product_attention_with_hlfb
+        if enable_hlfb
+        else sdpa.scaled_dot_product_attention
+    )
 
   def forward(
       self,
@@ -332,6 +325,7 @@ class CrossAttention(nn.Module):
       rope: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
       mask: Optional[torch.Tensor] = None,
       input_pos: Optional[torch.Tensor] = None,
+      kv_cache: Optional[kv_utils.KVCacheEntry] = None,
   ):
     """Forward function of the CrossAttention layer.
 
@@ -342,6 +336,7 @@ class CrossAttention(nn.Module):
       mask (torch.Tensor): the optional mask tensor can be broadcaseted to shape
         [B, n_heads, target_seq_len, source_seq_len].
       input_pos (torch.Tensor): the optional input position tensor.
+      kv_cache (KVCacheEntry): The KV cache entry corresponding to this module.
 
     Returns:
       output activation from this cross attention layer.
@@ -363,9 +358,11 @@ class CrossAttention(nn.Module):
     n_elem = int(self.config.rotary_percentage * self.config.head_dim)
     q, k = _embed_rope(q, k, n_elem, rope)
 
-    if self.kv_cache is not None:
-      # TODO(haoliang): Handle when execeeding max sequence length.
-      k, v = self.kv_cache.update_cache(input_pos, k, v)
+    if kv_cache is not None:
+      kv_cache = kv_utils.update(
+          kv_cache, input_pos, k, v, enable_hlfb=self.enable_hlfb
+      )
+      k, v = kv_cache.k_cache, kv_cache.v_cache
     if mask is None:
       mask = torch.zeros(
           (batch_size, 1, target_seq_len, source_seq_len), dtype=torch.float32
@@ -375,4 +372,4 @@ class CrossAttention(nn.Module):
 
     # Compute the output projection.
     y = self.output_projection(y)
-    return y
+    return y if kv_cache is None else (y, kv_cache)