PoPE-pytorch 0.0.6__py3-none-any.whl

PoPE_pytorch/PoPE.py

@@ -0,0 +1,146 @@
+from __future__ import annotations
+from collections import namedtuple
+from math import pi
+
+import torch
+from torch import arange, cat, stack, is_tensor, Tensor
+from torch.nn import Module, Parameter
+from torch.amp import autocast
+
+import torch.nn.functional as F
+
+from einops import einsum, rearrange
+
+from torch_einops_utils import slice_right_at_dim
+
+# constants
+
+PolarEmbedReturn = namedtuple('PolarEmbedReturn', ('freqs', 'bias'))
+
+# helper functions
+
+def exists(v):
+    return v is not None
+
+def default(v, d):
+    return v if exists(v) else d
+
+# applying pope to qk
+
+@autocast('cuda', enabled = False)
+def apply_pope_to_qk(
+    pope: PolarEmbedReturn,
+    q, k,
+    to_magnitude = F.softplus,
+    return_complex = False
+):
+    freqs, bias = pope
+
+    q_len, k_len, qk_dim, rotate_dim = q.shape[-2], k.shape[-2], q.shape[-1], freqs.shape[-1]
+
+    assert q_len <= k_len and rotate_dim <= qk_dim
+
+    is_partial_rotate = rotate_dim < qk_dim
+
+    if is_partial_rotate:
+        q, q_rest = q[..., :rotate_dim], q[..., rotate_dim:]
+        k, k_rest = k[..., :rotate_dim], k[..., rotate_dim:]
+
+        if return_complex:
+            q_rest = torch.polar(q_rest, torch.zeros_like(q_rest))
+            k_rest = torch.polar(k_rest, torch.zeros_like(k_rest))
+
+    if freqs.ndim == 3:
+        freqs = rearrange(freqs, 'b n d -> b 1 n d')
+
+    freqs_with_bias = freqs + rearrange(bias, 'h d -> h 1 d')
+
+    # convert q and k to polar magnitudes with activation
+
+    q, k = to_magnitude(q), to_magnitude(k)
+
+    # apply rotations
+
+    freqs = slice_right_at_dim(freqs, q_len, dim = -2)
+
+    if return_complex:
+        q = torch.polar(q, freqs)
+    else:
+        qcos, qsin = freqs.cos(), freqs.sin()
+        q = rearrange([q * qcos, q * qsin], 'two ... d -> ... (d two)')
+
+    # handle inference
+
+    if return_complex:
+        k = torch.polar(k, freqs_with_bias)
+    else:
+        kcos, ksin = freqs_with_bias.cos(), freqs_with_bias.sin()
+        k = rearrange([k * kcos, k * ksin], 'two ... d -> ... (d two)')
+
+    # concat
+
+    if is_partial_rotate:
+        q = cat((q, q_rest), dim = -1)
+        k = cat((k, k_rest), dim = -1)
+
+    return q, k
+
+# main class
+
+class PoPE(Module):
+    def __init__(
+        self,
+        dim,
+        *,
+        heads,
+        theta = 10000,
+        bias_uniform_init = False
+    ):
+        super().__init__()
+
+        # freqs
+
+        inv_freqs = theta ** -(arange(dim).float() / dim)
+        self.register_buffer('inv_freqs', inv_freqs)
+
+        # the learned bias on the keys
+
+        self.bias = Parameter(torch.zeros(heads, dim))
+
+        if bias_uniform_init:
+            self.bias.uniform_(-2. * pi, 0.)
+
+        # convenience
+
+        self.apply_pope_to_qk = staticmethod(apply_pope_to_qk)
+
+    @property
+    def device(self):
+        return self.inv_freqs.device
+
+    @autocast('cuda', enabled = False)
+    def forward(
+        self,
+        pos_or_seq_len: Tensor | int,
+        offset = 0
+    ):
+
+        # get positions depending on input
+
+        if is_tensor(pos_or_seq_len):
+            pos = pos_or_seq_len
+        else:
+            seq_len = pos_or_seq_len
+            pos = arange(seq_len, device = self.device, dtype = self.inv_freqs.dtype)
+
+        pos = pos + offset
+
+        # freqs
+
+        freqs = einsum(pos, self.inv_freqs, '... i, j -> ... i j')
+
+        # the bias, with clamping
+
+        bias = self.bias.clamp(-2. * pi, 0.)
+
+        return PolarEmbedReturn(freqs, bias)

PoPE_pytorch/__init__.py

@@ -0,0 +1,4 @@
+from PoPE_pytorch.PoPE import (
+    PoPE,
+    apply_pope_to_qk
+)

pope_pytorch-0.0.6.dist-info/METADATA

@@ -0,0 +1,97 @@
+Metadata-Version: 2.4
+Name: PoPE-pytorch
+Version: 0.0.6
+Summary: PoPE
+Project-URL: Homepage, https://pypi.org/project/PoPE-pytorch/
+Project-URL: Repository, https://github.com/lucidrains/PoPE-pytorch
+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,positional embedding
+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>=0.0.14
+Requires-Dist: torch>=2.5
+Provides-Extra: examples
+Provides-Extra: test
+Requires-Dist: pytest; extra == 'test'
+Description-Content-Type: text/markdown
+
+<img src="./pope.png" width="400px"></img>
+
+## PoPE-pytorch (wip)
+
+Efficient implementation (and explorations) into [polar coordinate positional embedding (PoPE)](https://arxiv.org/abs/2509.10534) - from [Gopalakrishnan](https://agopal42.github.io/) et al. under Schmidhuber
+
+## Install
+
+```shell
+$ pip install PoPE-pytorch
+```
+
+## Usage
+
+```python
+import torch
+from PoPE_pytorch import PoPE
+
+# define pope
+
+pope = PoPE(64, heads = 8)
+
+# pass in sequence length
+
+pos_embed = pope(1024)
+
+# queries and keys in attention
+
+q = torch.randn(1, 8, 1024, 64)
+k = torch.randn(1, 8, 1024, 64)
+
+# training
+
+rotated_q, rotated_k = pope.apply_pope_to_qk(pos_embed, q, k)
+
+# inference
+
+rotated_q, rotated_k = pope.apply_pope_to_qk(pos_embed, q[..., -1:, :], k)
+```
+
+## Citations
+
+```bibtex
+@misc{gopalakrishnan2025decouplingwhatwherepolar,
+    title   = {Decoupling the "What" and "Where" With Polar Coordinate Positional Embeddings}, 
+    author  = {Anand Gopalakrishnan and Robert Csordás and Jürgen Schmidhuber and Michael C. Mozer},
+    year    = {2025},
+    eprint  = {2509.10534},
+    archivePrefix = {arXiv},
+    primaryClass = {cs.LG},
+    url     = {https://arxiv.org/abs/2509.10534}, 
+}
+```

pope_pytorch-0.0.6.dist-info/RECORD

@@ -0,0 +1,6 @@
+PoPE_pytorch/PoPE.py,sha256=2QeXhQM3-7SNUJ8EID1S3lpP-yKXQsWP0T2mkIGJaiU,3428
+PoPE_pytorch/__init__.py,sha256=mVk9Y0nAmhMxlEQjZKY0OLmTKsz9oyMSqF7n5LoWcbg,65
+pope_pytorch-0.0.6.dist-info/METADATA,sha256=it21ePxjcQQJvsQkxOkAsSsaPMGk-P7Mo6dwZHdoWSg,3270
+pope_pytorch-0.0.6.dist-info/WHEEL,sha256=WLgqFyCfm_KASv4WHyYy0P3pM_m7J5L9k2skdKLirC8,87
+pope_pytorch-0.0.6.dist-info/licenses/LICENSE,sha256=e6AOF7Z8EFdK3IdcL0x0fLw4cY7Q0d0kNR0o0TmBewM,1066
+pope_pytorch-0.0.6.dist-info/RECORD,,

pope_pytorch-0.0.6.dist-info/WHEEL

@@ -0,0 +1,4 @@
+Wheel-Version: 1.0
+Generator: hatchling 1.28.0
+Root-Is-Purelib: true
+Tag: py3-none-any

pope_pytorch-0.0.6.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.