Add Rotary Positional Embeddings

.gitignore

@@ -0,0 +1,179 @@
+# Derived from basic .gitignore template for python projects:
+#   https://github.com/github/gitignore/blob/main/Python.gitignore
+# Please maintain the alphabetic order of the section titles
+# To debug why a file is being ignored, use the command:
+#    git check-ignore -v $my_ignored_file
+
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+
+# C extensions
+*.so
+
+# Cython debug symbols
+cython_debug/
+
+# Celery stuff
+celerybeat-schedule
+celerybeat.pid
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# Django stuff
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Environments
+.env
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Flask stuff
+instance/
+.webassets-cache
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# IPython
+profile_default/
+ipython_config.py
+
+# Jupyter Notebook
+*.ipynb_checkpoints
+
+# mkdocs documentation
+/site
+
+# Model saving / checkpointing
+*.pt
+*.pth
+*.ckpt
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# PyBuilder
+.pybuilder/
+target/
+
+# PyCharm
+#  JetBrains specific template is maintained in a separate JetBrains.gitignore that can
+#  be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore
+#  and can be added to the global gitignore or merged into this file.  For a more nuclear
+#  option (not recommended) you can uncomment the following to ignore the entire idea folder.
+#.idea/
+
+# PyInstaller
+#  Usually these files are written by a python script from a template
+#  before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# pyenv
+#   For a library or package, you might want to ignore these files since the code is
+#   intended to run in multiple environments; otherwise, check them in:
+# .python-version
+
+# pipenv
+#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+#   However, in case of collaboration, if having platform-specific dependencies or dependencies
+#   having no cross-platform support, pipenv may install dependencies that don't work, or not
+#   install all needed dependencies.
+# Pipfile.lock
+
+# poetry
+#   Similar to Pipfile.lock, it is generally recommended to include poetry.lock in version control.
+#   This is especially recommended for binary packages to ensure reproducibility, and is more
+#   commonly ignored for libraries.
+#   https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control
+# poetry.lock
+
+# PEP 582: https://peps.python.org/pep-0582/
+#   This PEP proposes to add to Python a mechanism to automatically recognize a __pypackages__
+#   directory and prefer importing packages installed in this location over user or global site-packages.
+#   This will avoid the steps to create, activate or deactivate virtual environments. Python will use
+#   the __pypackages__ from the base directory of the script when present.
+__pypackages__/
+
+# Pyre type checker
+.pyre/
+
+# pytype static type analyzer
+.pytype/
+
+# Rope project settings
+.ropeproject
+
+# SageMath parsed files
+*.sage.py
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# System / program generated files
+*.err
+*.log
+*.swp
+.DS_Store
+
+# Translations
+*.mo
+*.pot
+
+# TorchX
+*.torchxconfig
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py,cover
+.hypothesis/
+.pytest_cache/
+cover/
+
+# VSCode
+.vscode/

llm/llama2/__init__.py

@@ -0,0 +1,5 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.

llm/llama2/position_embeddings.py

@@ -0,0 +1,85 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+
+from torch import nn
+
+
+class RotaryPositionalEmbeddings(nn.Module):
+    """
+    This class implements Rotary Positional Embeddings (RoPE)
+    proposed in: https://arxiv.org/abs/2104.09864
+
+    Reference implementation (used for correctness verfication)
+    can be found here:
+    https://github.com/facebookresearch/llama/blob/main/llama/model.py#L450
+
+    Attributes:
+        dim (int): Embedding dimension for each head, computed as:
+            embed_size //  num_heads
+        max_seq_len (int): Maximum expected sequence length for the
+            model, if exceeded the cached freqs will be recomputed
+        base (int): The base for the geometric progression used to compute
+            the rotation angles
+
+    Args:
+        x (tensor): input tensor to which rope is applied
+
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        max_seq_len: int = 4096,
+        base: int = 10_000,
+    ) -> None:
+        super().__init__()
+        self.dim = dim
+
+        theta = 1.0 / (
+            base ** (torch.arange(0, dim, 2)[: (dim // 2)].float() / dim)
+        )
+        self.register_buffer("theta", theta)
+        self.build_rope_cache(max_seq_len)
+
+    def build_rope_cache(self, max_seq_len: int = 4096) -> None:
+        # Create position indexes `[0, 1, ..., max_seq_len - 1]`
+        seq_idx = torch.arange(
+            max_seq_len, dtype=self.theta.dtype, device=self.theta.device
+        )
+
+        # Outer product of theta and position index
+        idx_theta = torch.einsum("i, j -> ij", seq_idx, self.theta).float()
+
+        cache = torch.stack(
+            [torch.cos(idx_theta), torch.sin(idx_theta)], dim=-1
+        )
+        self.register_buffer("cache", cache)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        TODO: The implementation below can be made more efficient
+        for inference.
+        """
+        seq_len = x.size(1)
+        rope_cache = self.cache[:seq_len]
+
+        # cast because the reference does
+        xshaped = x.float().reshape(*x.shape[:-1], -1, 2)
+        rope_cache = rope_cache.view(1, xshaped.size(1), 1, xshaped.size(3), 2)
+        x_out2 = torch.stack(
+            [
+                xshaped[..., 0] * rope_cache[..., 0]
+                - xshaped[..., 1] * rope_cache[..., 1],
+                xshaped[..., 1] * rope_cache[..., 0]
+                + xshaped[..., 0] * rope_cache[..., 1],
+            ],
+            -1,
+        )
+
+        x_out2 = x_out2.flatten(3)
+        return x_out2.type_as(x)

tests/__init__.py

@@ -0,0 +1,5 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.

tests/llm/__init__.py

@@ -0,0 +1,5 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.

tests/llm/llama2/__init__.py

@@ -0,0 +1,5 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.

tests/llm/llama2/test_position_embeddings.py

@@ -0,0 +1,50 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+import pytest
+
+import torch
+from llm.llama2.position_embeddings import RotaryPositionalEmbeddings
+from tests.test_utils import assert_expected, set_rng_seed
+
+
+@pytest.fixture(autouse=True)
+def random():
+    set_rng_seed(0)
+
+
+class TestRotaryPositionEmbedding:
+    """
+    Class for testing our Rotary Positional Embeddings (RoPE)
+    implementation. The expected tensors are computed from the
+    reference implementation here:
+    https://github.com/facebookresearch/llama/blob/main/llama/model.py#L450
+    """
+
+    @pytest.fixture
+    def input_params(self):
+        bsz = 4
+        num_heads = 32
+        embed_dim = 4096
+        head_dim = embed_dim // num_heads
+        seq_len = 4096
+        return bsz, num_heads, head_dim, seq_len
+
+    @pytest.fixture
+    def input(self, input_params):
+        bsz, num_heads, head_dim, seq_len = input_params
+        return torch.randn(bsz, seq_len, num_heads, head_dim)
+
+    @pytest.fixture
+    def rope(self, input_params):
+        _, _, head_dim, seq_len = input_params
+        return RotaryPositionalEmbeddings(dim=head_dim, max_seq_len=seq_len)
+
+    def test_forward(self, input, rope):
+        x_out = rope(input)
+        assert_expected(x_out.mean(), torch.tensor(-4.3060e-05), rtol=1e-05, atol=1e-8)
+        assert_expected(x_out.sum(), torch.tensor(-2889.6804), rtol=1e-05, atol=1e-8)
+        assert_expected(x_out.max(), torch.tensor(5.6446), rtol=1e-05, atol=1e-8)

tests/test_utils.py

@@ -0,0 +1,33 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+import random
+
+import torch
+from typing import Any, Optional
+
+
+def set_rng_seed(seed):
+    """Sets the seed for random number generators"""
+    torch.manual_seed(seed)
+    random.seed(seed)
+
+
+def assert_expected(
+    actual: Any,
+    expected: Any,
+    rtol: float = 1e-5,
+    atol: float = 1e-8,
+    check_device=True,
+):
+    torch.testing.assert_close(
+        actual,
+        expected,
+        rtol=rtol,
+        atol=atol,
+        check_device=check_device,
+        msg=f"actual: {actual}, expected: {expected}",
+    )