ai-edge-torch-nightly 0.1.dev202405131930__py3-none-any.whl

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.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/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 +1,288 @@
+# 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.
+
+import math
+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.hlfb import StableHLOCompositeBuilder
+
+
+def scaled_dot_product_attention(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    v: torch.Tensor,
+    head_size: int,
+    mask: Optional[torch.Tensor] = None,
+    scale: Optional[float] = None,
+):
+  """Scaled dot product attention.
+
+  Args:
+    q (torch.Tensor): Query tensor, with shape [B, T, N, H].
+    k (torch.Tensor): Key tensor, with shape [B, T, KV_LEN, H].
+    v (torch.Tensor): Value tensor, with shape [B, T, KV_LEN, H].
+    head_size (int): head dimension.
+    mask (torch.Tensor): the optional mask tensor.
+
+  Returns:
+    The output tensor of scaled_dot_product_attention.
+  """
+
+  if scale is None:
+    scale = 1.0 / math.sqrt(head_size)
+
+  q = q.transpose(1, 2)
+  k = k.transpose(1, 2)
+  v = v.transpose(1, 2)
+  if q.size() != k.size():
+    # Handle the GQA case, where q.shape[1] % k.shape[1] == 0.
+    k = k.repeat_interleave(q.shape[1] // k.shape[1], dim=1)
+    v = v.repeat_interleave(q.shape[1] // v.shape[1], dim=1)
+  y = F.scaled_dot_product_attention(
+      q,
+      k,
+      v,
+      attn_mask=mask,
+      dropout_p=0.0,
+      is_causal=mask is None,
+      scale=scale,
+  )
+  return y.transpose(1, 2)
+
+
+def scaled_dot_product_attention_with_hlfb(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    v: torch.Tensor,
+    head_size: int,
+    mask: Optional[torch.Tensor] = None,
+    scale: Optional[float] = None,
+):
+  """Scaled dot product attention with high-level function boundary enabled.
+
+  Args:
+    q (torch.Tensor): Query tensor, with shape [B, T, N, H].
+    k (torch.Tensor): Key tensor, with shape [B, T, KV_LEN, H].
+    v (torch.Tensor): Value tensor, with shape [B, T, KV_LEN, H].
+    head_size (int): head dimension.
+    mask (torch.Tensor): the optional mask tensor.
+
+  Returns:
+    The output tensor of scaled_dot_product_attention.
+  """
+
+  if scale is None:
+    scale = 1.0 / math.sqrt(head_size)
+
+  builder = StableHLOCompositeBuilder(
+      name="odml.scaled_dot_product_attention", attr={"scale": scale}
+  )
+  q, k, v, mask = builder.mark_inputs(q, k, v, mask)
+
+  q = q.transpose(1, 2)
+  k = k.transpose(1, 2)
+  v = v.transpose(1, 2)
+  if q.size() != k.size():
+    # Handle the GQA case, where q.shape[1] % k.shape[1] == 0.
+    k = k.repeat_interleave(q.shape[1] // k.shape[1], dim=1)
+    v = v.repeat_interleave(q.shape[1] // v.shape[1], dim=1)
+  y = F.scaled_dot_product_attention(
+      q,
+      k,
+      v,
+      attn_mask=mask,
+      dropout_p=0.0,
+      is_causal=mask is None,
+      scale=scale,
+  )
+
+  result = y.transpose(1, 2)
+  result = builder.mark_outputs(result)
+  return result
+
+
+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.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: Tuple[torch.Tensor, torch.Tensor],
+      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
+
+
+# CausalSelfAttention which can support MHQ, MQA or GQA.
+class CausalSelfAttention(nn.Module):
+
+  def __init__(
+      self,
+      dim: int,
+      config: cfg.AttentionConfig,
+      kv_cache_max: int,
+      enable_hlfb: bool,
+  ) -> None:
+    """Initialize an instance of CausalSelfAttention.
+
+    Args:
+      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
+
+    # Build a k/v cache with size (batch_size, kv_cache_max, n_heads, head_dim).
+    # Now only supports batch_size of 1.
+    # TODO(haoliang): support batch_size greater than 1.
+    if config.enable_kv_cache:
+      self.kv_cache = KVCache(
+          1,
+          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: Tuple[torch.Tensor, torch.Tensor],
+      mask: Optional[torch.Tensor] = None,
+      input_pos: Optional[torch.Tensor] = None,
+  ) -> torch.Tensor:
+    """Forward function of the CausalSelfAttention layer.
+
+    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 == 1, "Currently only batch_size = 1 is supported."
+
+    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
+    total_qkv = q_per_kv + 2  # Each group has >=1 queries, 1 key, and 1 value.
+    qkv = qkv.view(
+        B, T, self.config.num_query_groups, total_qkv, self.head_dim
+    )  # (B, T, num_query_groups, total_qkv, head_dim)
+
+    # Split batched computation into three.
+    q, k, v = qkv.split((q_per_kv, 1, 1), dim=-2)
+
+    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)
+    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)
+
+    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

ai_edge_torch/generative/layers/attention_utils.py

@@ -0,0 +1,169 @@
+# 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 utility functions used with attention module.
+
+import math
+from typing import Tuple
+
+import torch
+
+
+def build_rope_cache(
+    size: int,
+    dim: int,
+    base: int = 10000,
+    condense_ratio: int = 1,
+    dtype: torch.dtype = torch.float32,
+    device: torch.device = None,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+  """Precompute Rotary Positional Embedding Sin and Cos values for quick lookups
+  during the inference.
+
+  Args:
+      size (int): The size of the built cache.
+      dim (int): Each sequence's dimmension.
+      base (int, optional): Rope base value. Defaults to 10000.
+      condense_ratio (int, optional): The ratio by which sequence indicies are
+        condensed. Defaults to 1.
+      dtype (torch.dtype, optional): Output tensor's data type. Defaults to
+        torch.float32.
+      device (torch.device, optional): Output tensor's data type. Defaults to
+        None in which case "cpu" is used.
+
+  Returns:
+      Tuple[torch.Tensor, torch.Tensor]: Rope's Cosine and Sine waves.
+  """
+  if device is None:
+    device = torch.device('cpu')
+  theta = 1.0 / (base ** (torch.arange(0, dim, 2).float() / dim))
+  seq_idx = torch.arange(size) / condense_ratio
+  idx_theta = torch.outer(seq_idx, theta)
+  cos = torch.cos(idx_theta).to(dtype=dtype, device=device)
+  sin = torch.sin(idx_theta).to(dtype=dtype, device=device)
+  return cos, sin
+
+
+def build_causal_mask_cache(
+    size: int,
+    dtype: torch.dtype = torch.float32,
+    device: torch.device = None,
+) -> torch.Tensor:
+  """Build a cache for causal attention mask.
+
+  Args:
+      size (int): The size of the built mask cache.
+      dtype (torch.dtype, optional): Output tensor's data type. Defaults to
+        torch.float32.
+      device (torch.device, optional): Output tensor's data type. Defaults to
+        None in which case "cpu" is used.
+
+  Returns:
+      torch.Tensor: Causal attention mask.
+  """
+  if device is None:
+    device = torch.device('cpu')
+  mask = torch.full((size, size), float('-inf'), dtype=dtype, device=device)
+  return torch.triu(mask, diagonal=1).unsqueeze(0).unsqueeze(0)
+
+
+def relative_position_bucket(
+    relative_position: torch.Tensor,
+    bidirectional: bool,
+    num_buckets: int,
+    max_distance: int,
+) -> torch.Tensor:
+  """
+  Adapted from Mesh Tensorflow:
+  https://github.com/tensorflow/mesh/blob/0cb87fe07da627bf0b7e60475d59f95ed6b5be3d/mesh_tensorflow/transformer/transformer_layers.py#L593
+
+  Translate relative position to a bucket number for relative attention. The relative position is defined as
+  memory_position - query_position, i.e. the distance in tokens from the attending position to the attended-to
+  position. If bidirectional=False, then positive relative positions are invalid. We use smaller buckets for
+  small absolute relative_position and larger buckets for larger absolute relative_positions. All relative
+  positions >=max_distance map to the same bucket. All relative positions <=-max_distance map to the same bucket.
+  This should allow for more graceful generalization to longer sequences than the model has been trained on
+
+  Args:
+      relative_position: an int32 Tensor
+      bidirectional: a boolean - whether the attention is bidirectional
+      num_buckets: an integer for number of buckets.
+      max_distance: an integer for max distance.
+
+  Returns:
+      a Tensor with the same shape as relative_position, containing int32 values in the range [0, num_buckets)
+  """
+  relative_buckets = 0
+  if bidirectional:
+    num_buckets //= 2
+    relative_buckets += (relative_position > 0).to(torch.long) * num_buckets
+    relative_position = torch.abs(relative_position)
+  else:
+    relative_position = -torch.min(
+        relative_position, torch.zeros_like(relative_position)
+    )
+  # now relative_position is in the range [0, inf)
+
+  # half of the buckets are for exact increments in positions
+  max_exact = num_buckets // 2
+  is_small = relative_position < max_exact
+
+  # The other half of the buckets are for logarithmically bigger bins in positions up to max_distance
+  relative_position_if_large = max_exact + (
+      torch.log(relative_position.float() / max_exact)
+      / math.log(max_distance / max_exact)
+      * (num_buckets - max_exact)
+  ).to(torch.long)
+  relative_position_if_large = torch.min(
+      relative_position_if_large,
+      torch.full_like(relative_position_if_large, num_buckets - 1),
+  )
+
+  relative_buckets += torch.where(
+      is_small, relative_position, relative_position_if_large
+  )
+  return relative_buckets
+
+
+def build_relative_position_buckets(
+    query_length: int,
+    key_length: int,
+    bidirectional: bool = True,
+    num_buckets: int = 32,
+    max_distance: int = 128,
+) -> torch.Tensor:
+  """Relative position buckets for computing bias.
+
+  Args:
+    query_length: an integer of length of current query tensor.
+    key_length: an integer of length of current key tensor.
+    bidirectional: a boolean - whether the attention is bidirectional, default is True.
+    num_buckets: an integer for number of buckets, default is 32.
+    max_distance: an integer for max distance, default is 128.
+
+  Returns:
+    A torch.Tensor of computed relative position buckets.
+  """
+  context_position = torch.arange(query_length, dtype=torch.long)[:, None]
+  memory_position = torch.arange(key_length, dtype=torch.long)[None, :]
+  relative_position = (
+      memory_position - context_position
+  )  # shape (query_length, key_length)
+  rel_pos_bucket = relative_position_bucket(
+      relative_position,  # shape (query_length, key_length)
+      bidirectional=bidirectional,
+      num_buckets=num_buckets,
+      max_distance=max_distance,
+  )
+  return rel_pos_bucket.unsqueeze(0).unsqueeze(0)