ai-edge-torch-nightly 0.2.0.dev20240714__py3-none-any.whl

ai_edge_torch/generative/examples/tiny_llama/convert_to_tflite.py

@@ -0,0 +1,66 @@
+# Copyright 2024 The AI Edge Torch Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+
+import os
+from pathlib import Path
+
+import torch
+
+import ai_edge_torch
+from ai_edge_torch.generative.examples.tiny_llama import tiny_llama
+from ai_edge_torch.generative.quantize import quant_recipes
+
+
+def convert_tiny_llama_to_tflite(
+    checkpoint_path: str,
+    prefill_seq_len: int = 512,
+    kv_cache_max_len: int = 1024,
+    quantize: bool = True,
+):
+  """An example method for converting TinyLlama model to multi-signature
+  tflite model.
+
+  Args:
+      checkpoint_path (str): The filepath to the model checkpoint, or directory holding the checkpoint.
+      prefill_seq_len (int, optional): The maximum size of prefill input tensor.
+        Defaults to 512.
+      kv_cache_max_len (int, optional): The maximum size of KV cache buffer,
+        including both prefill and decode. Defaults to 1024.
+      quantize (bool, optional): Whether the model should be quanized.
+        Defaults to True.
+  """
+  pytorch_model = tiny_llama.build_model(
+      checkpoint_path, kv_cache_max_len=kv_cache_max_len
+  )
+  # Tensors used to trace the model graph during conversion.
+  prefill_tokens = torch.full((1, prefill_seq_len), 0, dtype=torch.long)
+  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)
+
+  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)
+      )
+      .signature('decode', pytorch_model, (decode_token, decode_input_pos))
+      .convert(quant_config=quant_config)
+  )
+  edge_model.export(f'/tmp/tiny_llama_seq{prefill_seq_len}_kv{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)

ai_edge_torch/generative/examples/tiny_llama/tiny_llama.py

@@ -0,0 +1,164 @@
+# Copyright 2024 The AI Edge Torch Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# 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.
+
+import os
+from pathlib import Path
+
+import numpy as np
+import torch
+import torch.nn as nn
+
+from ai_edge_torch.generative.layers.attention import TransformerBlock
+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 ai_edge_torch.generative.utilities.loader as loading_utils
+
+TENSOR_NAMES = loading_utils.ModelLoader.TensorNames(
+    ff_up_proj="model.layers.{}.mlp.up_proj",
+    ff_down_proj="model.layers.{}.mlp.down_proj",
+    ff_gate_proj="model.layers.{}.mlp.gate_proj",
+    attn_query_proj="model.layers.{}.self_attn.q_proj",
+    attn_key_proj="model.layers.{}.self_attn.k_proj",
+    attn_value_proj="model.layers.{}.self_attn.v_proj",
+    attn_output_proj="model.layers.{}.self_attn.o_proj",
+    pre_attn_norm="model.layers.{}.input_layernorm",
+    pre_ff_norm="model.layers.{}.post_attention_layernorm",
+    embedding="model.embed_tokens",
+    final_norm="model.norm",
+    lm_head="lm_head",
+)
+
+
+class TinyLLamma(nn.Module):
+
+  def __init__(self, config: cfg.ModelConfig):
+    super().__init__()
+
+    self.config = config
+    # Construct model layers.
+    self.lm_head = nn.Linear(
+        config.embedding_dim, config.vocab_size, bias=config.lm_head_use_bias
+    )
+    self.tok_embedding = nn.Embedding(
+        config.vocab_size, config.embedding_dim, padding_idx=0
+    )
+    self.transformer_blocks = nn.ModuleList(
+        TransformerBlock(config) for _ in range(config.num_layers)
+    )
+    self.final_norm = builder.build_norm(
+        config.embedding_dim,
+        config.final_norm_config,
+    )
+    self.rope_cache = attn_utils.build_rope_cache(
+        size=config.kv_cache_max,
+        dim=int(config.attn_config.rotary_percentage * config.head_dim),
+        base=10_000,
+        condense_ratio=1,
+        dtype=torch.float32,
+        device=torch.device("cpu"),
+    )
+    self.mask_cache = attn_utils.build_causal_mask_cache(
+        size=config.kv_cache_max, dtype=torch.float32, device=torch.device("cpu")
+    )
+    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:
+    B, T = idx.size()
+    assert (
+        self.config.max_seq_len >= T
+    ), f"Cannot forward sequence of length {T}, max seq length is only {self.config.max_seq_len}"
+
+    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.kv_cache_max]
+
+    # forward the model itself
+    x = self.tok_embedding(idx)  # token embeddings of shape (b, t, n_embd)
+
+    for i, block in enumerate(self.transformer_blocks):
+      x = block(x, (cos, sin), mask, input_pos)
+
+    x = self.final_norm(x)
+
+    res = self.lm_head(x)  # (b, t, vocab_size)
+    return res
+
+
+def get_model_config(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
+  attn_config = cfg.AttentionConfig(
+      num_heads=32,
+      num_query_groups=4,
+      rotary_percentage=1.0,
+  )
+  ff_config = cfg.FeedForwardConfig(
+      type=cfg.FeedForwardType.GATED,
+      activation=cfg.ActivationConfig(cfg.ActivationType.SILU),
+      intermediate_size=5632,
+  )
+  norm_config = cfg.NormalizationConfig(type=cfg.NormalizationType.RMS_NORM)
+  config = cfg.ModelConfig(
+      vocab_size=32000,
+      num_layers=22,
+      max_seq_len=2048,
+      embedding_dim=2048,
+      kv_cache_max_len=kv_cache_max_len,
+      attn_config=attn_config,
+      ff_config=ff_config,
+      pre_attention_norm_config=norm_config,
+      pre_ff_norm_config=norm_config,
+      final_norm_config=norm_config,
+      enable_hlfb=True,
+  )
+  return config
+
+
+def get_fake_model_config_for_test() -> cfg.ModelConfig:
+  config = get_model_config()
+  config.vocab_size = 128
+  config.num_layers = 2
+  config.ff_config.intermediate_size = 256
+  return config
+
+
+def build_model(checkpoint_path, **kwargs) -> nn.Module:
+  config = get_model_config(**kwargs)
+  model = TinyLLamma(config)
+  loader = loading_utils.ModelLoader(checkpoint_path, TENSOR_NAMES)
+  loader.load(model)
+  return model
+
+
+def define_and_run() -> None:
+  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)
+  print("running an inference")
+  print(model.forward(tokens, input_pos))
+
+
+if __name__ == "__main__":
+  define_and_run()

ai_edge_torch/generative/fx_passes/__init__.py

@@ -0,0 +1,31 @@
+# Copyright 2024 The AI Edge Torch Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+import torch
+
+from ai_edge_torch.convert.fx_passes import CanonicalizePass
+from ai_edge_torch.convert.fx_passes import run_passes
+from ai_edge_torch.generative.fx_passes.remove_sdpa_zero_mask_pass import RemoveSDPACompositeZeroMaskPass  # NOQA
+
+
+def run_generative_passes(
+    exported_program: torch.export.ExportedProgram,
+) -> torch.export.ExportedProgram:
+  return run_passes(
+      exported_program,
+      [
+          RemoveSDPACompositeZeroMaskPass(),
+          CanonicalizePass(),
+      ],
+  )

ai_edge_torch/generative/fx_passes/remove_sdpa_zero_mask_pass.py

@@ -0,0 +1,47 @@
+# Copyright 2024 The AI Edge Torch Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+import torch
+
+from ai_edge_torch.convert.fx_passes._pass_base import ExportedProgramPassBase
+from ai_edge_torch.convert.fx_passes._pass_base import ExportedProgramPassResult  # NOQA
+
+
+class RemoveSDPACompositeZeroMaskPass(ExportedProgramPassBase):
+
+  def is_zero_tensor_node(self, node: torch.fx.Node):
+    return node.target == torch.ops.aten.zeros.default
+
+  def call(self, exported_program: torch.export.ExportedProgram):
+    graph = exported_program.graph_module.graph
+    for node in graph.nodes:
+      if not (
+          node.op == "call_function"
+          and node.target == torch.ops.xla.mark_tensor.default
+      ):
+        continue
+
+      source, name, io_position, id, is_input = node.args[:5]
+      # Composite info:
+      # - name: odml.scaled_dot_product_attention
+      # - inputs: q, k, v, mask
+      if name == "odml.scaled_dot_product_attention" and is_input and io_position == 3:
+        if self.is_zero_tensor_node(source):
+          # Remove the mark_tensor call on the mask input by
+          # replacing the target with an identity function.
+          node.target = lambda *args, **kwargs: torch.zeros_like(args[0])
+
+    exported_program.graph_module.graph.lint()
+    exported_program.graph_module.recompile()
+    return ExportedProgramPassResult(exported_program, True)

ai_edge_torch/generative/layers/__init__.py

@@ -0,0 +1,14 @@
+# Copyright 2024 The AI Edge Torch Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================

ai_edge_torch/generative/layers/attention.py

@@ -0,0 +1,354 @@
+# Copyright 2024 The AI Edge Torch Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# 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
+
+import torch
+from torch import nn
+import torch.nn.functional as F
+
+import ai_edge_torch.generative.layers.builder as builder
+from ai_edge_torch.generative.layers.kv_cache import KVCache
+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
+
+
+def _embed_rope(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    n_elem: int,
+    rope: Tuple[torch.Tensor, torch.Tensor],
+) -> Tuple[torch.Tensor, torch.Tensor]:
+  """Embed rotary positional embedding for query and key.
+
+  Args:
+    q (torch.Tensor): query tensor.
+    k (torch.Tensor): key tensor.
+    n_elem (int): number of elements to embed rotarty positional embedding.
+    rope (Tuple[torch.Tensor, torch.Tensor]): the input rope tensor.
+  """
+  if n_elem > 0:
+    cos, sin = rope
+    q_roped = rotary_pos_emb.apply_rope(
+        q[..., :n_elem], cos.repeat(1, 2), sin.repeat(1, 2)
+    )
+    k_roped = rotary_pos_emb.apply_rope(
+        k[..., :n_elem], cos.repeat(1, 2), sin.repeat(1, 2)
+    )
+    q = torch.cat((q_roped, q[..., n_elem:]), dim=-1)
+    k = torch.cat((k_roped, k[..., n_elem:]), dim=-1)
+  return q, k
+
+
+class TransformerBlock(nn.Module):
+
+  def __init__(self, config: cfg.ModelConfig) -> None:
+    """Initialize an instance of the TransformerBlock.
+
+    Args:
+      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
+    )
+    self.atten_func = CausalSelfAttention(
+        config.batch_size,
+        config.embedding_dim,
+        config.attn_config,
+        config.kv_cache_max,
+        config.enable_hlfb,
+    )
+    self.pre_ff_norm = builder.build_norm(
+        config.embedding_dim, config.pre_ff_norm_config
+    )
+    self.ff = builder.build_ff(config.embedding_dim, config.ff_config)
+    self.config = config
+
+  def forward(
+      self,
+      x: torch.Tensor,
+      rope: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+      mask: Optional[torch.Tensor] = None,
+      input_pos: Optional[torch.Tensor] = None,
+  ) -> torch.Tensor:
+    """Forward function of the TransformerBlock.
+
+    Args:
+      x (torch.Tensor): the input tensor.
+      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.
+
+    Returns:
+      output activation from this transformer block.
+    """
+
+    if self.config.parallel_residual:
+      x_norm = self.pre_atten_norm(x)
+      attn_out = self.atten_func(x_norm, rope, mask, input_pos)
+      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)
+      x = x + attn_out
+      x_norm = self.pre_ff_norm(x)
+      output = x + self.ff(x_norm)
+
+    return output
+
+
+class CausalSelfAttention(nn.Module):
+
+  def __init__(
+      self,
+      batch_size: int,
+      dim: int,
+      config: cfg.AttentionConfig,
+      kv_cache_max: int,
+      enable_hlfb: bool,
+  ) -> None:
+    """Initialize an instance of CausalSelfAttention.
+
+    Args:
+      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__()
+    self.head_dim = dim // config.num_heads
+    shape = (config.num_heads + 2 * config.num_query_groups) * self.head_dim
+    # Key, query, value projections for all heads.
+    self.qkv_projection = nn.Linear(dim, shape, bias=config.qkv_use_bias)
+    self.output_projection = nn.Linear(dim, dim, bias=config.output_proj_use_bias)
+    self.config = config
+    self.kv_cache = None
+    self.batch_size = batch_size
+
+    # 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.head_dim,
+          enable_hlfb,
+      )
+
+    if enable_hlfb:
+      self.sdpa_func = scaled_dot_product_attention_with_hlfb
+    else:
+      self.sdpa_func = scaled_dot_product_attention
+
+  def forward(
+      self,
+      x: torch.Tensor,
+      rope: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+      mask: Optional[torch.Tensor] = None,
+      input_pos: Optional[torch.Tensor] = None,
+  ) -> torch.Tensor:
+    """Forward function of the CausalSelfAttention 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.
+      mask (torch.Tensor): the optional mask tensor.
+      input_pos (torch.Tensor): the optional input position tensor.
+
+    Returns:
+      output activation from this self attention layer.
+    """
+    # Batch size, sequence length, embedding dimensionality.
+    B, T, E = x.size()
+    assert (
+        B == self.batch_size
+    ), "batch size of input tensor must match with the batch size specified in the model configuration."
+
+    qkv = self.qkv_projection(x)
+
+    # Assemble into a number of query groups to support MHA, MQA and GQA.
+    q_per_kv = self.config.num_heads // self.config.num_query_groups
+    # Each group has >=1 queries, 1 key, and 1 value.
+    if self.config.qkv_transpose_before_split:
+      qkv = qkv.view(B, T, -1, self.head_dim)
+      q, k, v = qkv.split(
+          (
+              q_per_kv * self.config.num_query_groups,
+              self.config.num_query_groups,
+              self.config.num_query_groups,
+          ),
+          dim=-2,
+      )
+    else:
+      qkv = qkv.view(B, T, self.config.num_query_groups, -1)
+      q, k, v = qkv.split(
+          (q_per_kv * self.head_dim, self.head_dim, self.head_dim), dim=-1
+      )
+
+    q = q.reshape(B, T, -1, self.head_dim)
+    k = k.reshape(B, T, -1, self.head_dim)
+    v = v.reshape(B, T, -1, self.head_dim)
+
+    # Compute rotary positional embedding for query and key.
+    n_elem = int(self.config.rotary_percentage * self.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)
+
+    y = self.sdpa_func(q, k, v, self.head_dim, mask=mask)
+    y = y.reshape(B, T, E)
+
+    # Compute the output projection.
+    y = self.output_projection(y)
+    return y
+
+
+class SelfAttention(CausalSelfAttention):
+  """Non-causal Self Attention module, which is equivalent to CausalSelfAttention without mask."""
+
+  def forward(
+      self,
+      x: torch.Tensor,
+      rope: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+      input_pos: Optional[torch.Tensor] = None,
+  ) -> torch.Tensor:
+    """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.
+
+    Returns:
+      output activation from this self attention layer.
+    """
+    B, T, _ = x.size()
+    return super().forward(
+        x,
+        rope=rope,
+        mask=torch.zeros((B, 1, T, T), dtype=torch.float32),
+        input_pos=input_pos,
+    )
+
+
+class CrossAttention(nn.Module):
+
+  def __init__(
+      self,
+      batch_size: int,
+      query_dim: int,
+      cross_dim: int,
+      config: cfg.AttentionConfig,
+      kv_cache_max: int,
+      enable_hlfb: bool,
+  ) -> None:
+    """Initialize an instance of CrossAttention.
+
+    Args:
+      batch_size (int): batch size of the input tensor.
+      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__()
+    self.config = config
+    self.head_dim = query_dim // config.num_heads
+    self.n_heads = config.num_heads
+    self.q_projection = nn.Linear(query_dim, query_dim, bias=config.qkv_use_bias)
+    self.k_projection = nn.Linear(cross_dim, query_dim, bias=config.qkv_use_bias)
+    self.v_projection = nn.Linear(cross_dim, query_dim, bias=config.qkv_use_bias)
+    self.output_projection = nn.Linear(
+        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.head_dim,
+          enable_hlfb,
+      )
+
+    if enable_hlfb:
+      self.sdpa_func = scaled_dot_product_attention_with_hlfb
+    else:
+      self.sdpa_func = scaled_dot_product_attention
+
+  def forward(
+      self,
+      x: torch.Tensor,
+      y: torch.Tensor,
+      rope: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+      mask: Optional[torch.Tensor] = None,
+      input_pos: Optional[torch.Tensor] = None,
+  ):
+    """Forward function of the CrossAttention layer.
+
+    Args:
+      x (torch.Tensor): the target tensor, with shape [B, target_seq_len, ...].
+      y (torch.Tensor): the source tensor, with shape [B, source_seq_len, ...].
+      rope (Tuple[torch.Tensor, torch.Tensor]): the optional input rope tensor.
+      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.
+
+    Returns:
+      output activation from this cross attention layer.
+    """
+    batch_size = x.size()[0]
+    target_seq_len = x.size()[1]
+    source_seq_len = y.size()[1]
+
+    q = self.q_projection(x)
+    k = self.k_projection(y)
+    v = self.v_projection(y)
+
+    interim_shape = (batch_size, -1, self.n_heads, self.head_dim)
+    q = q.view(interim_shape)
+    k = k.view(interim_shape)
+    v = v.view(interim_shape)
+
+    # Compute rotary positional embedding for query and key.
+    n_elem = int(self.config.rotary_percentage * self.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 mask is None:
+      mask = torch.zeros(
+          (batch_size, 1, target_seq_len, source_seq_len), dtype=torch.float32
+      )
+    y = self.sdpa_func(q, k, v, self.head_dim, mask=mask)
+    y = y.reshape(batch_size, target_seq_len, -1)
+
+    # Compute the output projection.
+    y = self.output_projection(y)
+    return y