poly-attention 0.0.1__py3-none-any.whl

poly_attention/__init__.py

@@ -0,0 +1 @@
+from poly_attention.poly_attention import PolyAttention

poly_attention/poly_attention.py

@@ -0,0 +1,123 @@
+# Chakrabarti et al. https://arxiv.org/abs/2602.02422
+
+from functools import partial
+
+import torch
+from torch import nn, einsum
+from torch.nn import Module, RMSNorm
+import torch.nn.functional as F
+
+import einx
+from einops import rearrange
+from einops.layers.torch import Rearrange
+
+# constants
+
+LinearNoBias = partial(nn.Linear, bias=False)
+
+# helper functions
+
+def exists(val):
+    return val is not None
+
+def default(val, d):
+    return val if exists(val) else d
+
+def divisible_by(num, den):
+    return (num % den) == 0
+
+def softclamp(x, c):
+    return c * torch.tanh(x / c)
+
+# poly attention
+
+class PolyAttention(Module):
+    def __init__(
+        self,
+        dim,
+        heads = 8,
+        dim_head = 64,
+        causal = False,
+        softclamp_value = 20.,
+        eps = 1e-9
+    ):
+        super().__init__()
+        dim_inner = dim_head * heads
+
+        self.causal = causal
+        self.softclamp_value = softclamp_value
+        self.scale = dim_head ** -0.5
+        self.eps = eps
+
+        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.to_q_v = LinearNoBias(dim, dim_inner * 5)
+        
+        self.q1_norm = RMSNorm(dim_head)
+        self.q2_norm = RMSNorm(dim_head)
+        self.q3_norm = RMSNorm(dim_head)
+
+        self.to_out = nn.Linear(dim_inner, dim)
+
+        self._reset_parameters()
+
+    def _reset_parameters(self):
+        # small initialization is crucial to prevent exp() overflow
+        nn.init.normal_(self.to_q_v.weight, mean = 0.0, std = 0.01)
+        nn.init.normal_(self.to_out.weight, mean = 0.0, std = 0.01)
+        nn.init.zeros_(self.to_out.bias)
+
+    def forward(self, x, mask = None):
+        device = x.device
+
+        q1, q2, q3, v2, v3 = map(self.split_heads, self.to_q_v(x).chunk(5, dim = -1))
+        
+        q1 = self.q1_norm(q1)
+        q2 = self.q2_norm(q2)
+        q3 = self.q3_norm(q3)
+        
+        q_left = torch.stack((q1, q2))
+        q_right = torch.stack((q2, q3))
+
+        # unscaled exp values
+        
+        scores = einsum('... i d, ... j d -> ... i j', q_left, q_right) * self.scale
+        scores = softclamp(scores, self.softclamp_value)
+        
+        exp_scores = scores.exp()
+        
+        # causal masking
+        
+        if self.causal:
+            i, j = exp_scores.shape[-2:]
+            causal_mask = torch.ones((i, j), device = device, dtype = torch.bool).triu(1)
+            exp_scores = exp_scores.masked_fill(causal_mask, 0.)
+        
+        # padding masking
+        
+        if exists(mask):
+            exp_scores = einx.where('b j, c b h i j, -> c b h i j', mask, exp_scores, 0.)
+
+        exp_scores12, exp_scores23 = exp_scores
+
+        # aggregate
+        
+        exp_scores23_v3 = einsum('b h i j, b h j d -> b h i d', exp_scores23, v3)
+        unnormalized_out = einsum('b h i j, b h j d -> b h i d', exp_scores12, exp_scores23_v3)
+        
+        # elementwise multiply the root values (v2) with the aggregated messages from the rest of the tree
+
+        out = v2 * unnormalized_out
+
+        # normalize
+
+        exp_scores23_sum = exp_scores23.sum(dim = -1, keepdim = True)
+
+        denominator = einsum('b h i j, b h j d -> b h i d', exp_scores12, exp_scores23_sum)
+
+        out = unnormalized_out / denominator.clamp(min = self.eps)
+
+        # combine heads
+
+        return self.to_out(self.merge_heads(out))

poly_attention-0.0.1.dist-info/METADATA

@@ -0,0 +1,83 @@
+Metadata-Version: 2.4
+Name: poly-attention
+Version: 0.0.1
+Summary: PolyAttention
+Project-URL: Homepage, https://pypi.org/project/poly-attention/
+Project-URL: Repository, https://github.com/lucidrains/poly-attention
+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,attention mechanism,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: einx>=0.3.0
+Requires-Dist: torch>=2.5
+Requires-Dist: x-transformers>=2.21.4
+Provides-Extra: examples
+Provides-Extra: test
+Requires-Dist: pytest; extra == 'test'
+Description-Content-Type: text/markdown
+
+
+## Poly Attention (wip)
+
+Implementation of <a href="https://arxiv.org/abs/2602.02422">Poly-Attention</a>, a general scheme for higher-order self-attention
+
+## Install
+
+```bash
+$ pip install poly-attention
+```
+
+## Usage
+
+```python
+import torch
+from poly_attention import PolyAttention
+
+attn = PolyAttention(
+    dim = 512,
+    heads = 8,
+    dim_head = 64,
+    causal = False
+)
+
+x = torch.randn(1, 1024, 512)
+out = attn(x) # (1, 1024, 512)
+```
+
+## Citations
+
+```bibtex
+@inproceedings{chakrabarti2026poly,
+    title   = {Poly-attention: a general scheme for higher-order self-attention},
+    author  = {Chakrabarti, Sayak and Pitassi, Toniann and Alman, Josh},
+    booktitle = {International Conference on Learning Representations (ICLR)},
+    year    = {2026}
+}
+```

poly_attention-0.0.1.dist-info/RECORD

@@ -0,0 +1,6 @@
+poly_attention/__init__.py,sha256=IZGqAZPTmP2Td66HzQk0M3a6Kk6lExVM-1ReQ2y7jdg,56
+poly_attention/poly_attention.py,sha256=aeLT6gQWsrma5LqU_zSru_nBqs9kcr-7hQIJgAOWjVs,3405
+poly_attention-0.0.1.dist-info/METADATA,sha256=AvPUwBRDvdVrxq5BCvhBh8zwoqpFYtnprJ06O3qPjCE,2883
+poly_attention-0.0.1.dist-info/WHEEL,sha256=mffPy8wBnZQn2VnJUU5jE99KsxaSfiyMHV9Yt0aLVxs,87
+poly_attention-0.0.1.dist-info/licenses/LICENSE,sha256=e6AOF7Z8EFdK3IdcL0x0fLw4cY7Q0d0kNR0o0TmBewM,1066
+poly_attention-0.0.1.dist-info/RECORD,,

poly_attention-0.0.1.dist-info/WHEEL

@@ -0,0 +1,4 @@
+Wheel-Version: 1.0
+Generator: hatchling 1.30.1
+Root-Is-Purelib: true
+Tag: py3-none-any

poly_attention-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.