log-depth-recurrent-modeling 0.0.1__py3-none-any.whl

log_depth_recurrent_modeling/__init__.py

@@ -0,0 +1,7 @@
+from log_depth_recurrent_modeling.ar_grc import (
+    AutoregressiveGatedRecursiveCell
+)
+
+__all__ = [
+    'AutoregressiveGatedRecursiveCell'
+]

log_depth_recurrent_modeling/ar_grc.py

@@ -0,0 +1,167 @@
+from __future__ import annotations
+import math
+
+import torch
+from torch import nn
+from torch.nn import Module, Linear, RMSNorm, LayerNorm
+import torch.nn.functional as F
+
+from einops import einsum, rearrange, repeat
+
+from x_mlps_pytorch import create_mlp
+
+from torch_einops_utils import pad_at_dim_to_multiple, pack_with_inverse
+
+# helper functions
+
+def exists(v):
+    return v is not None
+
+def default(*args):
+    for arg in args:
+        if exists(arg):
+            return arg
+    return None
+
+# classes
+
+class GatedRecursiveCell(Module):
+    def __init__(
+        self,
+        dim,
+        dim_hidden = None,
+        activation = nn.SiLU(),
+        depth = 1,
+        use_rmsnorm = True
+    ):
+        super().__init__()
+        dim_out = dim * 4
+        dim_hidden = default(dim_hidden, dim_out)
+
+        self.mlp = create_mlp(
+            dim_in = 2 * dim,
+            dim = dim_hidden,
+            dim_out = dim_out,
+            activation = activation,
+            depth = depth
+        )
+
+        self.norm = RMSNorm(dim) if use_rmsnorm else LayerNorm(dim, bias = False)
+
+    def forward(self, x, y):
+
+        x_gate, y_gate, c_gate, c = self.mlp((x, y)).chunk(4, dim = -1)
+
+        return self.norm(x * x_gate.sigmoid() + y * y_gate.sigmoid() + c_gate.sigmoid() * c)
+
+# main class
+
+class AutoregressiveGatedRecursiveCell(Module):
+    def __init__(
+        self,
+        *,
+        num_tokens,
+        dim_embed,
+        dim,
+        window_size,
+        separate_grc = False,
+        grc_kwargs: dict = dict()
+    ):
+        super().__init__()
+
+        tree_depth = math.log2(window_size)
+        assert tree_depth.is_integer()
+
+        self.dim = dim
+
+        # embed
+
+        self.token_embed = nn.Embedding(num_tokens, dim_embed)
+
+        # project to and from
+
+        self.embed_to_model = Linear(dim_embed, dim)
+        self.model_to_embed = Linear(dim, dim_embed)
+
+        # recursive tree, up and down (blelloch scan)
+
+        self.window_size = window_size
+
+        self.tree_depth = int(tree_depth)
+
+        self.root_hidden = nn.Parameter(torch.randn(dim) * 1e-2)
+
+        self.separate_grc = separate_grc
+        self.up_grc = GatedRecursiveCell(dim = dim, **grc_kwargs)
+        self.down_grc = GatedRecursiveCell(dim = dim, **grc_kwargs) if separate_grc else self.up_grc
+
+    def forward(
+        self,
+        ids,
+        return_loss = False
+    ):
+        up_grc, down_grc, root_hidden = self.up_grc, self.down_grc, self.root_hidden
+
+        if return_loss:
+            ids, labels = ids[:, :-1], ids[:, 1:]
+
+        embeds = self.token_embed(ids)
+
+        x = self.embed_to_model(embeds)
+
+        x, remove_padding = pad_at_dim_to_multiple(x, multiple = self.window_size, dim = -2)
+
+        # divide into window size
+
+        x = rearrange(x, 'b (w n) d -> b w n d', n = self.window_size)
+        x, inverse_pack_window = pack_with_inverse(x, '* n d')
+
+        # up sweep
+
+        curr = x
+        up_hiddens = [x]
+
+        for _ in range(self.tree_depth - 1):
+
+            left, right = rearrange(curr, 'b (h two) d -> two b h d', two = 2)
+            curr = up_grc(left, right)
+
+            up_hiddens.append(curr)
+
+        # down sweep (blelloch)
+
+        curr = repeat(root_hidden, 'd -> b 1 d', b = x.shape[0])
+
+        for _ in range(self.tree_depth):
+            left_up, _ = rearrange(up_hiddens.pop(), 'b (h two) d -> two b h d', two = 2)
+
+            right_carry = down_grc(curr, left_up)
+
+            curr = rearrange([curr, right_carry], 'two b h d -> b (h two) d')
+
+        # include each leaf
+
+        x = down_grc(curr, x)
+
+        # restore window
+
+        x = inverse_pack_window(x)
+        x = rearrange(x, 'b w n d -> b (w n) d')
+
+        # slice out padding
+
+        x = remove_padding(x)
+
+        embeds = self.model_to_embed(x)
+
+        logits = einsum(embeds, self.token_embed.weight, 'b n d, l d -> b n l')
+
+        if not return_loss:
+            return logits
+
+        return F.cross_entropy(rearrange(logits, 'b n l -> b l n'), labels, ignore_index = -1)
+
+# alias
+
+GRC = GatedRecursiveCell
+ARGRC = AutoregressiveGatedRecursiveCell

log_depth_recurrent_modeling-0.0.1.dist-info/METADATA

@@ -0,0 +1,88 @@
+Metadata-Version: 2.5
+Name: log-depth-recurrent-modeling
+Version: 0.0.1
+Summary: Log Depth Recurrent Modeling
+Project-URL: Homepage, https://pypi.org/project/log-depth-recurrent-modeling/
+Project-URL: Repository, https://github.com/lucidrains/log-depth-recurrent-modeling
+Author-email: Phil Wang <lucidrains@gmail.com>
+License: MIT License
+        
+        Copyright (c) 2026 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,deep learning
+Classifier: Development Status :: 4 - Beta
+Classifier: Intended Audience :: Developers
+Classifier: License :: OSI Approved :: MIT License
+Classifier: Programming Language :: Python :: 3.10
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Requires-Python: >=3.10
+Requires-Dist: einops>=0.8.1
+Requires-Dist: torch-einops-utils>=0.1.30
+Requires-Dist: torch>=2.5
+Requires-Dist: x-mlps-pytorch>=0.6.2
+Provides-Extra: examples
+Provides-Extra: test
+Requires-Dist: pytest; extra == 'test'
+Description-Content-Type: text/markdown
+
+<img src="./ldrlm-fig1.png" width="350px"></img>
+
+## Log Depth Recurrent Modeling (wip)
+
+Explorations into the [Log Depth Recurrent Modeling](https://arxiv.org/abs/2609.28212) proposed by Yiqin Wang of Imperial College London
+
+## Citations
+
+```bibtex
+@misc{wang2026logdepthrecurrentlanguagemodeling,
+    title    = {Log-Depth Recurrent Language Modeling},
+    author   = {Yiqin Wang and Nuri Cingillioglu and Charles Pert},
+    year     = {2026},
+    eprint   = {2609.28212},
+    archivePrefix = {arXiv},
+    primaryClass = {cs.LG},
+    url      = {https://arxiv.org/abs/2609.28212},
+}
+```
+
+```bibtex
+@misc{shen2019orderedmemory,
+    title   = {Ordered Memory},
+    author  = {Yikang Shen and Shawn Tan and Arian Hosseini and Zhouhan Lin and Alessandro Sordoni and Aaron Courville},
+    year    = {2019},
+    eprint  = {1910.13466},
+    archivePrefix = {arXiv},
+    primaryClass = {cs.LG},
+    url     = {https://arxiv.org/abs/1910.13466},
+}
+```
+
+```bibtex
+@misc{pert2026lengthgeneralizationlogdepthrecurrent,
+    title     = {Length Generalization with Log-Depth Recurrent Units},
+    author    = {Charles Pert and Dalal Alrajeh and Alessandra Russo},
+    year      = {2026},
+    eprint    = {2605.26035},
+    archivePrefix = {arXiv},
+    primaryClass = {cs.LG},
+    url       = {https://arxiv.org/abs/2605.26035},
+}
+```

log_depth_recurrent_modeling-0.0.1.dist-info/RECORD

@@ -0,0 +1,6 @@
+log_depth_recurrent_modeling/__init__.py,sha256=2JG1vAchNKPZ3QMhlmL9dLkw61VFhGzai5w-SQQhH2M,143
+log_depth_recurrent_modeling/ar_grc.py,sha256=CesL3qGlA9JIXb3BpixX7YDeFi6Q7ORzH5TFqh2srV0,4085
+log_depth_recurrent_modeling-0.0.1.dist-info/METADATA,sha256=trKFJhn6lb7s6fEwdXDmv0MgBXV6IHzJUQX1_pNdtkw,3490
+log_depth_recurrent_modeling-0.0.1.dist-info/WHEEL,sha256=W3fkpkm7-wf9vBI5Z-7s0eWkeM-spu78I8Neb98DeEg,87
+log_depth_recurrent_modeling-0.0.1.dist-info/licenses/LICENSE,sha256=e6AOF7Z8EFdK3IdcL0x0fLw4cY7Q0d0kNR0o0TmBewM,1066
+log_depth_recurrent_modeling-0.0.1.dist-info/RECORD,,

log_depth_recurrent_modeling-0.0.1.dist-info/WHEEL

@@ -0,0 +1,4 @@
+Wheel-Version: 1.0
+Generator: hatchling 1.32.4
+Root-Is-Purelib: true
+Tag: py3-none-any

log_depth_recurrent_modeling-0.0.1.dist-info/licenses/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 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.