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

ai_edge_torch/experimental/__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/__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/examples/__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/examples/gemma/__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/examples/gemma/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.gemma import gemma
+from ai_edge_torch.generative.quantize import quant_recipes
+
+
+def convert_gemma_to_tflite(
+    checkpoint_path: str,
+    prefill_seq_len: int = 512,
+    kv_cache_max_len: int = 1024,
+    quantize: bool = True,
+):
+  """An example method for converting a Gemma 2B 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 = gemma.build_2b_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/gemma_seq{prefill_seq_len}_kv{kv_cache_max_len}.tflite')
+
+
+if __name__ == '__main__':
+  checkpoint_path = os.path.join(Path.home(), 'Downloads/llm_data/gemma-2b')
+  convert_gemma_to_tflite(checkpoint_path)

ai_edge_torch/generative/examples/gemma/gemma.py

@@ -0,0 +1,174 @@
+# 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 Gemma model.
+
+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=None,
+)
+
+
+class Gemma(nn.Module):
+
+  def __init__(self, config: cfg.ModelConfig):
+    super().__init__()
+
+    self.config = config
+    # Construct model layers.
+    self.tok_embedding = nn.Embedding(
+        config.vocab_size, config.embedding_dim, padding_idx=0
+    )
+    self.lm_head = nn.Linear(
+        config.embedding_dim,
+        config.vocab_size,
+        bias=config.lm_head_use_bias,
+    )
+    # Gemma re-uses the embedding as the head projection layer.
+    self.lm_head.weight.data = self.tok_embedding.weight.data
+    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]
+
+    # token embeddings of shape (b, t, n_embd)
+    x = self.tok_embedding(idx)
+    x = x * (self.config.embedding_dim**0.5)
+
+    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_2b(kv_cache_max_len: int = 1024) -> cfg.ModelConfig:
+  attn_config = cfg.AttentionConfig(
+      num_heads=8,
+      num_query_groups=1,
+      rotary_percentage=1.0,
+  )
+  ff_config = cfg.FeedForwardConfig(
+      type=cfg.FeedForwardType.GATED,
+      activation=cfg.ActivationConfig(cfg.ActivationType.GELU_TANH),
+      intermediate_size=16384,
+  )
+  norm_config = cfg.NormalizationConfig(
+      type=cfg.NormalizationType.RMS_NORM,
+      epsilon=1e-6,
+      zero_centered=True,
+  )
+  config = cfg.ModelConfig(
+      vocab_size=256000,
+      num_layers=18,
+      max_seq_len=8192,
+      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,
+      parallel_residual=False,
+      lm_head_use_bias=False,
+      enable_hlfb=True,
+  )
+  return config
+
+
+def get_fake_model_config_2b_for_test() -> cfg.ModelConfig:
+  config = get_model_config_2b()
+  config.num_layers = 2
+  return config
+
+
+def build_2b_model(checkpoint_path, **kwargs) -> nn.Module:
+  config = get_model_config_2b(**kwargs)
+  model = Gemma(config)
+  loader = loading_utils.ModelLoader(checkpoint_path, TENSOR_NAMES)
+  # since embedding and lm-head use the same weight, we need to set strict
+  # to False.
+  loader.load(model, strict=False)
+  return model
+
+
+def define_and_run_2b() -> None:
+  kv_cache_max_len = 1024
+  checkpoint_path = os.path.join(Path.home(), "Downloads/llm_data/gemma-2b")
+  model = build_2b_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_2b()

ai_edge_torch/generative/examples/phi2/__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/examples/phi2/convert_to_tflite.py

@@ -0,0 +1,64 @@
+# 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.phi2 import phi2
+from ai_edge_torch.generative.quantize import quant_recipes
+
+
+def convert_phi2_to_tflite(
+    checkpoint_path: str,
+    prefill_seq_len: int = 512,
+    kv_cache_max_len: int = 1024,
+    quantize: bool = True,
+):
+  """An example method for converting a Phi-2 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 = phi2.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/phi2_seq{prefill_seq_len}_kv{kv_cache_max_len}.tflite')
+
+
+if __name__ == '__main__':
+  checkpoint_path = os.path.join(Path.home(), 'Downloads/llm_data/phi2')
+  convert_phi2_to_tflite(checkpoint_path)

ai_edge_torch/generative/examples/phi2/phi2.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 phi-2 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.fc1",
+    ff_down_proj="model.layers.{}.mlp.fc2",
+    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.dense",
+    pre_attn_norm="model.layers.{}.input_layernorm",
+    embedding="model.embed_tokens",
+    final_norm="model.final_layernorm",
+    lm_head="lm_head",
+)
+
+
+class Phi2(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=32,
+      rotary_percentage=0.4,
+      qkv_use_bias=True,
+      output_proj_use_bias=True,
+  )
+  ff_config = cfg.FeedForwardConfig(
+      type=cfg.FeedForwardType.SEQUENTIAL,
+      activation=cfg.ActivationConfig(cfg.ActivationType.GELU_TANH),
+      intermediate_size=10240,
+      use_bias=True,
+  )
+  norm_config = cfg.NormalizationConfig(type=cfg.NormalizationType.LAYER_NORM)
+  config = cfg.ModelConfig(
+      vocab_size=51200,
+      num_layers=32,
+      max_seq_len=2048,
+      kv_cache_max_len=kv_cache_max_len,
+      embedding_dim=2560,
+      attn_config=attn_config,
+      ff_config=ff_config,
+      pre_attention_norm_config=norm_config,
+      final_norm_config=norm_config,
+      parallel_residual=True,
+      lm_head_use_bias=True,
+      enable_hlfb=True,
+  )
+  return config
+
+
+def get_fake_model_config_for_test() -> cfg.ModelConfig:
+  config = get_model_config()
+  config.num_layers = 2
+  return config
+
+
+def build_model(checkpoint_path, **kwargs) -> nn.Module:
+  config = get_model_config(**kwargs)
+  model = Phi2(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/phi2")
+  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/examples/stable_diffusion/__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/examples/stable_diffusion/attention.py

@@ -0,0 +1,106 @@
+# 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 math
+
+import torch
+from torch import _decomp
+from torch import nn
+from torch._prims_common import mask_tensor
+from torch._prims_common.wrappers import out_wrapper
+from torch.nn import functional as F
+
+
+def triu(a):
+  h, w = a.shape[-2:]
+  mask = (
+      torch.arange(w, device=a.device).unsqueeze(-2)
+      - torch.arange(h, device=a.device).unsqueeze(-1)
+  ) >= 1
+  mask = torch.broadcast_to(mask, a.shape)
+  return torch.ops.aten.logical_and(a, mask).contiguous()
+
+
+# _decomp.decomposition_table[torch.ops.aten.triu.default] = triu
+
+
+class SelfAttention(nn.Module):
+
+  def __init__(self, n_heads, d_embed, in_proj_bias=True, out_proj_bias=True):
+    super().__init__()
+    self.in_proj = nn.Linear(d_embed, 3 * d_embed, bias=in_proj_bias)
+    self.out_proj = nn.Linear(d_embed, d_embed, bias=out_proj_bias)
+    self.n_heads = n_heads
+    self.d_head = d_embed // n_heads
+
+  def forward(self, x, causal_mask=False):
+    input_shape = x.shape
+    batch_size, sequence_length, d_embed = input_shape
+    interim_shape = (batch_size, sequence_length, self.n_heads, self.d_head)
+
+    q, k, v = self.in_proj(x).chunk(3, dim=-1)
+
+    q = q.view(interim_shape).transpose(1, 2)
+    k = k.view(interim_shape).transpose(1, 2)
+    v = v.view(interim_shape).transpose(1, 2)
+
+    weight = q @ k.transpose(-1, -2)
+    if causal_mask:
+      # mask = torch.ones_like(weight, dtype=torch.bool).triu(1)
+      mask = triu(torch.ones_like(weight, dtype=torch.bool))
+      weight.masked_fill_(mask, -torch.inf)
+    weight /= math.sqrt(self.d_head)
+    weight = F.softmax(weight, dim=-1)
+
+    output = weight @ v
+    output = output.transpose(1, 2)
+    output = output.reshape(input_shape)
+    output = self.out_proj(output)
+    return output
+
+
+class CrossAttention(nn.Module):
+
+  def __init__(self, n_heads, d_embed, d_cross, in_proj_bias=True, out_proj_bias=True):
+    super().__init__()
+    self.q_proj = nn.Linear(d_embed, d_embed, bias=in_proj_bias)
+    self.k_proj = nn.Linear(d_cross, d_embed, bias=in_proj_bias)
+    self.v_proj = nn.Linear(d_cross, d_embed, bias=in_proj_bias)
+    self.out_proj = nn.Linear(d_embed, d_embed, bias=out_proj_bias)
+    self.n_heads = n_heads
+    self.d_head = d_embed // n_heads
+
+  def forward(self, x, y):
+    input_shape = x.shape
+    batch_size, sequence_length, d_embed = input_shape
+    interim_shape = (batch_size, -1, self.n_heads, self.d_head)
+
+    q = self.q_proj(x)
+    k = self.k_proj(y)
+    v = self.v_proj(y)
+
+    q = q.view(interim_shape).transpose(1, 2)
+    k = k.view(interim_shape).transpose(1, 2)
+    v = v.view(interim_shape).transpose(1, 2)
+
+    weight = q @ k.transpose(-1, -2)
+    weight /= math.sqrt(self.d_head)
+    weight = F.softmax(weight, dim=-1)
+
+    output = weight @ v
+    output = output.transpose(1, 2).contiguous()
+    output = output.view(input_shape)
+    output = self.out_proj(output)
+    return output