locoformer 0.0.6__tar.gz

locoformer-0.0.6/.github/workflows/python-publish.yml

@@ -0,0 +1,36 @@
+# This workflow will upload a Python Package using Twine when a release is created
+# For more information see: https://help.github.com/en/actions/language-and-framework-guides/using-python-with-github-actions#publishing-to-package-registries
+
+# This workflow uses actions that are not certified by GitHub.
+# They are provided by a third-party and are governed by
+# separate terms of service, privacy policy, and support
+# documentation.
+
+name: Upload Python Package
+
+on:
+  release:
+    types: [published]
+
+jobs:
+  deploy:
+
+    runs-on: ubuntu-latest
+
+    steps:
+    - uses: actions/checkout@v2
+    - name: Set up Python
+      uses: actions/setup-python@v2
+      with:
+        python-version: '3.x'
+    - name: Install dependencies
+      run: |
+        python -m pip install --upgrade pip
+        pip install build
+    - name: Build package
+      run: python -m build
+    - name: Publish package
+      uses: pypa/gh-action-pypi-publish@27b31702a0e7fc50959f5ad993c78deac1bdfc29
+      with:
+        user: __token__
+        password: ${{ secrets.PYPI_API_TOKEN }}

locoformer-0.0.6/.github/workflows/test.yml

@@ -0,0 +1,21 @@
+name: Pytest
+on: [push, pull_request]
+
+jobs:
+  build:
+
+    runs-on: ubuntu-latest
+
+    steps:
+    - uses: actions/checkout@v4
+    - name: Set up Python 3.10
+      uses: actions/setup-python@v5
+      with:
+        python-version: "3.10"
+    - name: Install dependencies
+      run: |
+        python -m pip install --upgrade pip
+        python -m pip install -e .[test]
+    - name: Test with pytest
+      run: |
+        python -m pytest tests/

locoformer-0.0.6/.gitignore

@@ -0,0 +1,208 @@
+
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[codz]
+*$py.class
+
+# C extensions
+*.so
+
+# 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
+
+# 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
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py.cover
+.hypothesis/
+.pytest_cache/
+cover/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+.pybuilder/
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# 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
+
+# UV
+#   Similar to Pipfile.lock, it is generally recommended to include uv.lock in version control.
+#   This is especially recommended for binary packages to ensure reproducibility, and is more
+#   commonly ignored for libraries.
+#uv.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
+#poetry.toml
+
+# pdm
+#   Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control.
+#   pdm recommends including project-wide configuration in pdm.toml, but excluding .pdm-python.
+#   https://pdm-project.org/en/latest/usage/project/#working-with-version-control
+#pdm.lock
+#pdm.toml
+.pdm-python
+.pdm-build/
+
+# pixi
+#   Similar to Pipfile.lock, it is generally recommended to include pixi.lock in version control.
+#pixi.lock
+#   Pixi creates a virtual environment in the .pixi directory, just like venv module creates one
+#   in the .venv directory. It is recommended not to include this directory in version control.
+.pixi
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm
+__pypackages__/
+
+# Celery stuff
+celerybeat-schedule
+celerybeat.pid
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.envrc
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# Pyre type checker
+.pyre/
+
+# pytype static type analyzer
+.pytype/
+
+# Cython debug symbols
+cython_debug/
+
+# 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/
+
+# Abstra
+# Abstra is an AI-powered process automation framework.
+# Ignore directories containing user credentials, local state, and settings.
+# Learn more at https://abstra.io/docs
+.abstra/
+
+# Visual Studio Code
+#  Visual Studio Code specific template is maintained in a separate VisualStudioCode.gitignore 
+#  that can be found at https://github.com/github/gitignore/blob/main/Global/VisualStudioCode.gitignore
+#  and can be added to the global gitignore or merged into this file. However, if you prefer, 
+#  you could uncomment the following to ignore the entire vscode folder
+# .vscode/
+
+# Ruff stuff:
+.ruff_cache/
+
+# PyPI configuration file
+.pypirc
+
+# Cursor
+#  Cursor is an AI-powered code editor. `.cursorignore` specifies files/directories to
+#  exclude from AI features like autocomplete and code analysis. Recommended for sensitive data
+#  refer to https://docs.cursor.com/context/ignore-files
+.cursorignore
+.cursorindexingignore
+
+# Marimo
+marimo/_static/
+marimo/_lsp/
+__marimo__/

locoformer-0.0.6/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2025 Phil Wang
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

locoformer-0.0.6/PKG-INFO

@@ -0,0 +1,71 @@
+Metadata-Version: 2.4
+Name: locoformer
+Version: 0.0.6
+Summary: LocoFormer
+Project-URL: Homepage, https://pypi.org/project/locoformer/
+Project-URL: Repository, https://github.com/lucidrains/locoformer
+Author-email: Phil Wang <lucidrains@gmail.com>
+License: MIT License
+        
+        Copyright (c) 2025 Phil Wang
+        
+        Permission is hereby granted, free of charge, to any person obtaining a copy
+        of this software and associated documentation files (the "Software"), to deal
+        in the Software without restriction, including without limitation the rights
+        to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+        copies of the Software, and to permit persons to whom the Software is
+        furnished to do so, subject to the following conditions:
+        
+        The above copyright notice and this permission notice shall be included in all
+        copies or substantial portions of the Software.
+        
+        THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+        IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+        FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+        AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+        LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+        OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+        SOFTWARE.
+License-File: LICENSE
+Keywords: artificial intelligence,attention mechanism,cross-embodiment,deep learning,robotics,transformer
+Classifier: Development Status :: 4 - Beta
+Classifier: Intended Audience :: Developers
+Classifier: License :: OSI Approved :: MIT License
+Classifier: Programming Language :: Python :: 3.9
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Requires-Python: >=3.9
+Requires-Dist: assoc-scan
+Requires-Dist: einops>=0.8.0
+Requires-Dist: einx>=0.3.0
+Requires-Dist: rotary-embedding-torch
+Requires-Dist: torch>=2.4
+Requires-Dist: x-mlps-pytorch
+Provides-Extra: examples
+Requires-Dist: accelerate; extra == 'examples'
+Requires-Dist: tqdm; extra == 'examples'
+Provides-Extra: test
+Requires-Dist: pytest; extra == 'test'
+Description-Content-Type: text/markdown
+
+<img src="./fig3.png" width="400px"></img>
+
+## LocoFormer (wip)
+
+[LocoFormer - Generalist Locomotion via Long-Context Adaptation](https://generalist-locomotion.github.io/)
+
+The gist is they trained a simple Transformer-XL in simulation on robots with many different bodies (cross-embodiment). When transferring to the real-world, they noticed the robot now gains the ability to adapt to insults. The XL memories span across multiple trials, which allowed the robot to learn in-context adaptation.
+
+## Sponsors
+
+This open sourced work is sponsored by [Safe Sentinel](https://www.safesentinels.com/)
+
+## Citations
+
+```bibtex
+@article{liu2025locoformer,
+    title   = {LocoFormer: Generalist Locomotion via Long-Context Adaptation},
+    author  = {Liu, Min and Pathak, Deepak and Agarwal, Ananye},
+    journal = {Conference on Robot Learning ({CoRL})},
+    year    = {2025}
+}
+```

locoformer-0.0.6/README.md

@@ -0,0 +1,22 @@
+<img src="./fig3.png" width="400px"></img>
+
+## LocoFormer (wip)
+
+[LocoFormer - Generalist Locomotion via Long-Context Adaptation](https://generalist-locomotion.github.io/)
+
+The gist is they trained a simple Transformer-XL in simulation on robots with many different bodies (cross-embodiment). When transferring to the real-world, they noticed the robot now gains the ability to adapt to insults. The XL memories span across multiple trials, which allowed the robot to learn in-context adaptation.
+
+## Sponsors
+
+This open sourced work is sponsored by [Safe Sentinel](https://www.safesentinels.com/)
+
+## Citations
+
+```bibtex
+@article{liu2025locoformer,
+    title   = {LocoFormer: Generalist Locomotion via Long-Context Adaptation},
+    author  = {Liu, Min and Pathak, Deepak and Agarwal, Ananye},
+    journal = {Conference on Robot Learning ({CoRL})},
+    year    = {2025}
+}
+```

locoformer-0.0.6/data/README.md

@@ -0,0 +1,3 @@
+# Data source
+
+The enwik8 data was downloaded from the Hutter prize page: http://prize.hutter1.net/

locoformer-0.0.6/locoformer/__init__.py

@@ -0,0 +1 @@
+from locoformer.locoformer import Locoformer

locoformer-0.0.6/locoformer/locoformer.py

@@ -0,0 +1,483 @@
+from __future__ import annotations
+from functools import partial
+
+import torch
+from torch import nn, cat, stack, arange, is_tensor
+import torch.nn.functional as F
+from torch.nn import Module, ModuleList, Linear, RMSNorm, Identity, Sequential
+from torch.utils._pytree import tree_map
+
+import einx
+from einops import rearrange, einsum
+from einops.layers.torch import Rearrange
+
+from rotary_embedding_torch import RotaryEmbedding
+
+from assoc_scan import AssocScan
+
+LinearNoBias = partial(Linear, bias = False)
+
+# helper functions
+
+def exists(v):
+    return v is not None
+
+def default(v, d):
+    return v if exists(v) else d
+
+def first(arr):
+    return arr[0]
+
+def divisible_by(num, den):
+    return (num % den) == 0
+
+def tree_map_tensor(x, fn):
+    return tree_map(lambda t: t if not is_tensor(t) else fn(t), x)
+
+def detach_all(x):
+    return tree_map_tensor(x, lambda t: t.detach())
+
+def combine_kv_cache(cache1, cache2):
+    combined_cache = []
+
+    for layer_cache1, layer_cache2 in zip(cache1, cache2):
+        next_cache = cat((layer_cache1, layer_cache2), dim = -2)
+        combined_cache.append(next_cache)
+
+    return combined_cache
+
+# generalized advantage estimate
+
+@torch.no_grad()
+def calc_gae(
+    rewards,
+    values,
+    masks,
+    gamma = 0.99,
+    lam = 0.95,
+    use_accelerated = None
+):
+    assert values.shape[-1] == rewards.shape[-1]
+    use_accelerated = default(use_accelerated, rewards.is_cuda)
+
+    values = F.pad(values, (0, 1), value = 0.)
+    values, values_next = values[..., :-1], values[..., 1:]
+
+    delta = rewards + gamma * values_next * masks - values
+    gates = gamma * lam * masks
+
+    scan = AssocScan(reverse = True, use_accelerated = use_accelerated)
+
+    gae = scan(gates, delta)
+
+    returns = gae + values
+
+    return returns
+
+# transformer-xl mask w/ flex attn
+
+flex_attention = None
+
+try:
+    from torch.nn.attention.flex_attention import flex_attention, create_block_mask
+    if torch.cuda.is_available():
+        flex_attention = torch.compile(flex_attention)
+except ImportError:
+    pass
+
+def create_xl_mask(
+    seq_len,
+    kv_seq_len,
+    window_size,
+    episode_ids = None,  # (b n) - in the case that within the same batch there are multiple episodes
+    lookback_blocks = 1, # in transformer-xl, lookback is one window size block, but can be multiple for longer context
+    device = None
+):
+    assert kv_seq_len >= seq_len
+    assert window_size <= seq_len
+
+    offset = kv_seq_len - seq_len
+
+    def create_block_mask_fn(b, __, q, k):
+        offset_q = q + offset
+        block_q = offset_q // window_size
+        block_k = k // window_size
+
+        causal_mask = offset_q >= k
+
+        # in transformer-xl, the previous segment is fully attended to - may just double the segments and make this sliding for ease of inference logic
+
+        block_mask = (block_q >= block_k) & (block_q <= (block_k + lookback_blocks))
+
+        mask = causal_mask & block_mask
+
+        # handle intra-episodic attention if needed
+
+        if exists(episode_ids):
+            q_episode = episodes[b, q + offset]
+            k_episode = episodes[b, k]
+
+            intra_episode_mask = q_episode == k_episode
+            mask = mask & intra_episode_mask
+
+        return mask
+
+    create_kwargs = dict(device = device) if exists(device) else dict()
+    return create_block_mask(create_block_mask_fn, B = None, H = None, Q_LEN = seq_len, KV_LEN = kv_seq_len, _compile = True, **create_kwargs)
+
+def create_sliding_mask(
+    seq_len,
+    kv_seq_len,
+    window_size,
+    device = None
+):
+    assert kv_seq_len >= seq_len
+    offset = kv_seq_len - seq_len
+
+    def sliding_mask(_, __, q, k):
+        offset_q = q + offset
+        distance = offset_q - k
+
+        backward_sliding_mask = distance <= window_size
+        forward_sliding_mask = distance >= 0
+
+        return backward_sliding_mask & forward_sliding_mask
+
+    create_kwargs = dict(device = device) if exists(device) else dict()
+    return create_block_mask(sliding_mask, B = None, H = None, Q_LEN = seq_len, KV_LEN = kv_seq_len, _compile = True, **create_kwargs)
+
+# transformer-xl with ppo
+
+class Attention(Module):
+    def __init__(
+        self,
+        dim,
+        window_size,
+        dim_head = 64,
+        heads = 8,
+        pre_rmsnorm = True,
+        fixed_window_size = False,
+        accept_value_residual = False
+    ):
+        super().__init__()
+        self.scale = dim_head ** -0.5
+
+        self.norm = RMSNorm(dim) if pre_rmsnorm else Identity()
+
+        self.split_heads = Rearrange('b n (h d) -> b h n d', h = heads)
+        self.merge_heads = Rearrange('b h n d -> b n (h d)')
+
+        self.rotary_embed = RotaryEmbedding(dim_head)
+
+        dim_inner = dim_head * heads
+        self.to_q = LinearNoBias(dim, dim_inner)
+        self.to_kv = LinearNoBias(dim, dim_inner * 2)
+        self.to_out = LinearNoBias(dim_inner, dim)
+
+        self.to_v_gates = Sequential(
+            LinearNoBias(dim, heads),
+            Rearrange('b n h -> b h n 1'),
+            nn.Sigmoid()
+        )
+
+        # value residual
+
+        self.accept_value_residual = accept_value_residual
+
+        if accept_value_residual:
+            self.to_value_residual_mix = Sequential(
+                LinearNoBias(dim, heads),
+                Rearrange('b n h -> b h n 1'),
+                nn.Sigmoid()                
+            )
+
+        # fixed window size
+
+        self.fixed_window_size = fixed_window_size
+        self.window_size = window_size
+
+    def forward(
+        self,
+        tokens,
+        value_residual = None,
+        kv_cache = None,
+        return_kv_cache = False,
+    ):
+        seq_len = tokens.shape[-2]
+        assert seq_len <= self.window_size
+
+        device = tokens.device
+
+        tokens = self.norm(tokens)
+
+        q, k, v = (self.to_q(tokens), *self.to_kv(tokens).chunk(2, dim = -1))
+
+        q, k, v = map(self.split_heads, (q, k, v))
+
+        orig_v = v
+
+        q = q * self.scale
+
+        if exists(value_residual):
+            assert self.accept_value_residual
+            mix = self.to_value_residual_mix(tokens)
+            v = v.lerp(value_residual, mix)
+
+        if exists(kv_cache):
+            ck, cv = kv_cache
+            k = cat((ck, k), dim = -2)
+            v = cat((cv, v), dim = -2)
+
+        if return_kv_cache:
+            next_kv_cache = stack((k, v))
+
+        q, k = self.rotary_embed.rotate_queries_with_cached_keys(q, k)
+
+        sim = einsum(q, k, 'b h i d, b h j d -> b h i j')
+
+        i, j = sim.shape[-2:]
+
+        if self.fixed_window_size:
+            i_seq = arange(i, device = device)
+            j_seq = arange(j, device = device) - (j - i)
+            dist = einx.subtract('i, j -> i j', i_seq, j_seq)
+            causal_mask = (dist < 0) | (dist > self.window_size)
+        else:
+            causal_mask = torch.ones((i, j), dtype = torch.bool, device = sim.device).triu(j - i + 1)
+
+        sim = sim.masked_fill(causal_mask, -torch.finfo(sim.dtype).max)
+
+        attn = sim.softmax(dim = -1)
+
+        out = einsum(attn, v, 'b h i j, b h j d -> b h i d')
+
+        out = out * self.to_v_gates(tokens)
+
+        out = self.merge_heads(out)
+
+        out = self.to_out(out)
+
+        if not return_kv_cache:
+            return out
+
+        return out, (next_kv_cache, orig_v)
+
+class FeedForward(Module):
+    def __init__(
+        self,
+        dim,
+        expansion_factor = 4.,
+        pre_rmsnorm = True
+    ):
+        super().__init__()
+        self.norm = RMSNorm(dim) if pre_rmsnorm else Identity()
+
+        dim_inner = int(dim * expansion_factor * 2 / 3)
+
+        self.proj_in = Linear(dim, dim_inner * 2)
+        self.proj_out = Linear(dim_inner, dim)
+
+    def forward(
+        self,
+        x
+    ):
+        x = self.norm(x)
+
+        x, gates = self.proj_in(x).chunk(2, dim = -1)
+
+        x = x * F.gelu(gates)
+
+        return self.proj_out(x)
+
+class TransformerXL(Module):
+    def __init__(
+        self,
+        dim,
+        depth,
+        window_size,
+        dim_head = 64,
+        heads = 8,
+        expansion_factor = 4.,
+        final_norm = True,
+        fixed_window_size = False,
+    ):
+        super().__init__()
+
+        layers = ModuleList([])
+
+        for i in range(depth):
+            is_first = i == 0
+
+            attn = Attention(dim = dim, dim_head = dim_head, heads = heads, fixed_window_size = fixed_window_size, window_size = window_size, accept_value_residual = not is_first)
+
+            ff = FeedForward(dim = dim, expansion_factor = expansion_factor)
+
+            layers.append(ModuleList([
+                attn, ff
+            ]))
+
+        self.layers = layers
+        self.norm = RMSNorm(dim) if final_norm else Identity()
+
+        # fixed window size
+
+        self.fixed_window_size = fixed_window_size
+        self.window_size = window_size
+
+    def forward(
+        self,
+        x,
+        cache = None,
+        return_kv_cache = False
+    ):
+
+        cache = default(cache, (None,) * len(self.layers))
+
+        next_kv_caches = []
+        value_residual = None
+
+        for (attn, ff), kv_cache in zip(self.layers, cache):
+
+            attn_out, (next_kv_cache, values) = attn(x, value_residual = value_residual, kv_cache = kv_cache, return_kv_cache = True)
+
+            x = attn_out + x
+            x = ff(x) + x
+
+            next_kv_caches.append(next_kv_cache)
+            value_residual = default(value_residual, values)
+
+        embed = self.norm(x)
+
+        if not return_kv_cache:
+            return embed
+
+        next_kv_cache = stack(next_kv_caches)
+
+        next_kv_cache = next_kv_cache[..., -self.window_size:, :]
+
+        return embed, next_kv_cache
+
+# class
+
+class Locoformer(Module):
+    def __init__(
+        self,
+        embedder: Module,
+        unembedder: Module,
+        transformer: dict | TransformerXL,
+        value_network: Module | None = None
+    ):
+        super().__init__()
+
+        if isinstance(transformer, dict):
+            transformer = TransformerXL(**transformer)
+
+        self.transformer = transformer
+
+        self.embedder = embedder
+        self.unembedder = unembedder
+
+        self.value_network = value_network
+
+        self.fixed_window_size = transformer.fixed_window_size
+        self.window_size = transformer.window_size
+
+    @property
+    def device(self):
+        return next(self.parameters()).device
+
+    def get_stateful_forward(
+        self,
+        initial_states: Tensor | None = None,
+        inference_mode = False,
+        has_batch_dim = False,
+        **kwargs
+    ):
+        window_size = self.window_size
+
+        cache = None
+
+        def stateful_forward(state: Tensor, **override_kwargs):
+            nonlocal cache
+
+            # handle no batch, for easier time rolling out against envs
+
+            if not has_batch_dim:
+                state = rearrange(state, '... -> 1 ...')
+
+            # forwards
+
+            out, cache = self.forward(state, cache = cache, **{**kwargs, **override_kwargs})
+
+            # handle cache
+
+            cache_len = cache.shape[-2]
+
+            if self.fixed_window_size or divisible_by(cache_len, window_size * 2):
+                cache = cache[..., -window_size:, :]
+
+            # maybe remove batch
+
+            if not has_batch_dim:
+                out = tree_map_tensor(out, lambda t: rearrange(t, '1 ... -> ...'))
+
+            return out
+
+        if inference_mode:
+            stateful_forward = torch.inference_mode()(stateful_forward)
+
+        # handle prompt
+
+        if not exists(initial_states):
+            return stateful_forward
+
+        initial_logits = []
+
+        for state_segments in initial_states.split(self.window_size, dim = -1):
+
+            logits = stateful_forward(state_segments, return_values = False)
+            initial_logits.append(logits)
+
+        initial_logits = cat(initial_logits, dim = -2)
+
+        return stateful_forward, initial_logits
+
+    def forward(
+        self,
+        state: Tensor,
+        cache: Tensor | None = None,
+        detach_cache = False,
+        return_values = False
+    ):
+
+        tokens = self.embedder(state)
+
+        embed, kv_cache = self.transformer(tokens, cache = cache, return_kv_cache = True)
+
+        # unembed to actions - in language models this would be the next state
+
+        action_logits = self.unembedder(embed)
+
+        out = action_logits
+
+        # maybe detach cache
+
+        if detach_cache:
+            kv_cache = detach_all(kv_cache)
+
+        # handle returning of values
+
+        if return_values:
+            assert exists(self.value_network)
+
+            values = self.value_network(embed)
+
+            if values.ndim == 3:
+                assert values.shape[-1] == 1
+                values = rearrange(values, '... 1 -> ...')
+
+            out = (out, values)
+
+        # output and cache
+
+        return out, kv_cache

locoformer-0.0.6/pyproject.toml

@@ -0,0 +1,69 @@
+[project]
+name = "locoformer"
+version = "0.0.6"
+description = "LocoFormer"
+authors = [
+    { name = "Phil Wang", email = "lucidrains@gmail.com" }
+]
+readme = "README.md"
+requires-python = ">= 3.9"
+license = { file = "LICENSE" }
+keywords = [
+    'artificial intelligence',
+    'deep learning',
+    'transformer',
+    'attention mechanism',
+    'robotics',
+    'cross-embodiment',
+]
+
+classifiers=[
+    'Development Status :: 4 - Beta',
+    'Intended Audience :: Developers',
+    'Topic :: Scientific/Engineering :: Artificial Intelligence',
+    'License :: OSI Approved :: MIT License',
+    'Programming Language :: Python :: 3.9',
+]
+
+dependencies = [
+    "assoc-scan",
+    "einx>=0.3.0",
+    "einops>=0.8.0",
+    "rotary-embedding-torch",
+    "torch>=2.4",
+    "x-mlps-pytorch",
+]
+
+[project.urls]
+Homepage = "https://pypi.org/project/locoformer/"
+Repository = "https://github.com/lucidrains/locoformer"
+
+[project.optional-dependencies]
+
+examples = [
+    "accelerate",
+    "tqdm"
+]
+
+test = [
+    "pytest"
+]
+
+[tool.pytest.ini_options]
+pythonpath = [
+  "."
+]
+
+[build-system]
+requires = ["hatchling"]
+build-backend = "hatchling.build"
+
+[tool.rye]
+managed = true
+dev-dependencies = []
+
+[tool.hatch.metadata]
+allow-direct-references = true
+
+[tool.hatch.build.targets.wheel]
+packages = ["locoformer"]

locoformer-0.0.6/tests/test_locoformer.py

@@ -0,0 +1,38 @@
+import pytest
+param = pytest.mark.parametrize
+
+import torch
+from x_mlps_pytorch import MLP
+
+from einops import rearrange
+
+def test_locoformer():
+    from locoformer.locoformer import Locoformer
+    from torch import nn
+    
+    model = Locoformer(
+        embedder = nn.Embedding(256, 128),
+        unembedder = nn.Linear(128, 256, bias = False),
+        value_network = MLP(128, 32, 1),
+        transformer = dict(
+            dim = 128,
+            depth = 1,
+            window_size = 256
+        )
+    )
+
+    seq = torch.randint(0, 256, (3, 512))
+
+    (logits, values), cache = model(seq, return_values = True)
+    (logits, values), cache = model(seq, return_values = True, cache = cache)
+    (logits, values), cache = model(seq, return_values = True, cache = cache)
+
+    assert logits.shape == (3, 512, 256)
+
+    stateful_forward = model.get_stateful_forward(256, has_batch_dim = True, return_values = True, inference_mode = True)
+
+    for state in seq.unbind(dim = -1):
+        state = rearrange(state, 'b -> b 1')
+
+        logits, values = stateful_forward(state)
+        assert logits.shape == (3, 1, 256)

locoformer-0.0.6/train.py

@@ -0,0 +1,191 @@
+# /// script
+# dependencies = [
+#   'accelerate',
+#   'locoformer',
+#   'tqdm'
+# ]
+# ///
+
+import tqdm
+import gzip
+from math import ceil
+import numpy as np
+
+import torch
+from torch import nn
+from torch import from_numpy
+from torch.optim import Adam
+from torch.nn import functional as F
+from torch.utils.data import DataLoader, Dataset
+
+from einops import rearrange
+from accelerate import Accelerator
+
+from locoformer.locoformer import Locoformer
+
+# constants
+
+NUM_BATCHES = int(1e5)
+BATCH_SIZE = 16
+LEARNING_RATE = 2e-4
+VALIDATE_EVERY  = 100
+
+GENERATE_EVERY  = 250
+PRIME_LENGTH    = 32
+GENERATE_LENGTH = 1024
+
+SEQ_LEN = 256
+NUM_SEGMENTS = 4
+FIXED_WINDOW_SIZE = False
+
+# helpers
+
+def cycle(loader):
+    while True:
+        for data in loader:
+            yield data
+
+def divisible_by(num, den):
+    return (num % den) == 0
+
+def decode_token(token):
+    return str(chr(max(32, token)))
+
+def decode_tokens(tokens):
+    return ''.join(list(map(decode_token, tokens)))
+
+# sampling
+
+def log(t, eps = 1e-20):
+    return t.clamp(min = eps).log()
+
+def gumbel_noise(t):
+    return -log(-log(torch.rand_like(t)))
+
+def gumbel_sample(logits, temperature = 1., eps = 1e-6, keepdim = True):
+    noise = gumbel_noise(logits)
+    return ((logits / max(temperature, eps)) + noise).argmax(dim = -1, keepdim = keepdim)
+
+def topk_logits_filter(logits, frac_num_tokens = 0.1):
+    num_tokens = logits.shape[-1]
+    k = ceil(frac_num_tokens * num_tokens)
+
+    val, ind = torch.topk(logits, k)
+    probs = torch.full_like(logits, float('-inf'))
+    probs.scatter_(-1, ind, val)
+    return probs
+
+# instantiate model
+
+dim_model = 512
+
+model = Locoformer(
+    embedder = nn.Embedding(256, dim_model),
+    unembedder = nn.Linear(dim_model, 256, bias = False),
+    transformer = dict(
+        dim = dim_model,
+        depth = 6,
+        fixed_window_size = FIXED_WINDOW_SIZE,
+        window_size = SEQ_LEN
+    )
+)
+
+# prepare enwik8 data
+
+with gzip.open('./data/enwik8.gz') as file:
+    data = np.frombuffer(file.read(int(95e6)), dtype = np.uint8).copy()
+    train_data, valid_data = np.split(data, [int(90e6)])
+    data_train, data_val = from_numpy(train_data), from_numpy(valid_data)
+
+class TextSamplerDataset(Dataset):
+    def __init__(self, data, seq_len, segments):
+        super().__init__()
+        self.data = data
+        self.seq_len = seq_len
+        self.segments = segments
+        self.total_len = seq_len * segments
+
+    def __getitem__(self, index):
+        rand_start = torch.randint(0, self.data.size(0) - self.total_len - 1, (1,))
+        full_seq = self.data[rand_start: rand_start + self.total_len + 1].long()
+        return full_seq
+
+    def __len__(self):
+        return self.data.size(0) // self.total_len
+
+train_dataset = TextSamplerDataset(data_train, SEQ_LEN, NUM_SEGMENTS)
+val_dataset   = TextSamplerDataset(data_val, SEQ_LEN, NUM_SEGMENTS)
+train_loader  = DataLoader(train_dataset, batch_size = BATCH_SIZE)
+val_loader    = DataLoader(val_dataset, batch_size = BATCH_SIZE)
+
+# optimizer
+
+optim = Adam(model.parameters(), lr = LEARNING_RATE)
+
+# prepare accelerate
+
+accelerate = Accelerator()
+
+model, optim, train_loader = accelerate.prepare(model, optim, train_loader)
+
+# training loop
+
+train_loader_iter = cycle(train_loader)
+val_loader_iter = cycle(val_loader)
+
+for i in range(NUM_BATCHES):
+    model.train()
+
+    seq = next(train_loader_iter)
+    seq, labels = seq[:, :-1], seq[:, 1:]
+
+    cache = None
+
+    for segment_seq, segment_labels in zip(seq.chunk(NUM_SEGMENTS, dim = -1), labels.chunk(NUM_SEGMENTS, dim = -1)):
+
+        logits, cache = model(
+            segment_seq,
+            cache = cache,
+            detach_cache = True
+        )
+
+        loss = F.cross_entropy(
+            rearrange(logits, 'b n l -> b l n'),
+            segment_labels
+        )
+
+        accelerate.backward(loss / NUM_SEGMENTS)
+        accelerate.print(f'[{i}] loss: {loss.item():.3f}')
+
+        optim.step()
+        optim.zero_grad()
+
+    if divisible_by(i + 1, GENERATE_EVERY):
+        model.eval()
+
+        val_seq = next(val_loader_iter)
+        prime = val_seq[0, :PRIME_LENGTH]
+
+        prime = prime.to(model.device)
+        out = prime
+
+        stateful_forward, logits = model.get_stateful_forward(has_batch_dim = False, initial_states = prime, inference_mode = True)
+
+        # sample
+
+        while out.shape[-1] < GENERATE_LENGTH:
+            filtered_logits = topk_logits_filter(logits[-1])
+
+            sampled = gumbel_sample(filtered_logits)
+            out = torch.cat((out, sampled), dim = -1)
+
+            logits = stateful_forward(sampled)
+
+        # decoded
+
+        decoded_prime = decode_tokens(prime.cpu())
+        decoded_string = decode_tokens(out[PRIME_LENGTH:].cpu())
+
+        print(f'\n\n[prime]: {decoded_prime}\n\n')
+        print('*' * 100)
+        print(f'\n\n [generated]: {decoded_string}\n\n')