x-transformers 2.8.2__tar.gz → 2.8.4__tar.gz

{x_transformers-2.8.2 → x_transformers-2.8.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: x-transformers
-Version: 2.8.2
+Version: 2.8.4
 Summary: X-Transformers
 Project-URL: Homepage, https://pypi.org/project/x-transformers/
 Project-URL: Repository, https://github.com/lucidrains/x-transformers
@@ -2552,4 +2552,25 @@ ids_out, num_out, is_number_mask = model.generate(start_ids, start_nums, 17)
 }
 ```
 
+```bibtex
+@misc{jordan2024muon,
+  author    = {Keller Jordan and Yuchen Jin and Vlado Boza and Jiacheng You and Franz Cesista and Laker Newhouse and Jeremy Bernstein},
+  title     = {Muon: An optimizer for hidden layers in neural networks},
+  year      = {2024},
+  url       = {https://kellerjordan.github.io/posts/muon/}
+}
+```
+
+```bibtex
+@misc{wang2025muonoutperformsadamtailend,
+    title   = {Muon Outperforms Adam in Tail-End Associative Memory Learning}, 
+    author  = {Shuche Wang and Fengzhuo Zhang and Jiaxiang Li and Cunxiao Du and Chao Du and Tianyu Pang and Zhuoran Yang and Mingyi Hong and Vincent Y. F. Tan},
+    year    = {2025},
+    eprint  = {2509.26030},
+    archivePrefix = {arXiv},
+    primaryClass = {cs.LG},
+    url     = {https://arxiv.org/abs/2509.26030}, 
+}
+```
+
 *solve intelligence... then use that to solve everything else.* - Demis Hassabis

{x_transformers-2.8.2 → x_transformers-2.8.4}/README.md

@@ -2504,4 +2504,25 @@ ids_out, num_out, is_number_mask = model.generate(start_ids, start_nums, 17)
 }
 ```
 
+```bibtex
+@misc{jordan2024muon,
+  author    = {Keller Jordan and Yuchen Jin and Vlado Boza and Jiacheng You and Franz Cesista and Laker Newhouse and Jeremy Bernstein},
+  title     = {Muon: An optimizer for hidden layers in neural networks},
+  year      = {2024},
+  url       = {https://kellerjordan.github.io/posts/muon/}
+}
+```
+
+```bibtex
+@misc{wang2025muonoutperformsadamtailend,
+    title   = {Muon Outperforms Adam in Tail-End Associative Memory Learning}, 
+    author  = {Shuche Wang and Fengzhuo Zhang and Jiaxiang Li and Cunxiao Du and Chao Du and Tianyu Pang and Zhuoran Yang and Mingyi Hong and Vincent Y. F. Tan},
+    year    = {2025},
+    eprint  = {2509.26030},
+    archivePrefix = {arXiv},
+    primaryClass = {cs.LG},
+    url     = {https://arxiv.org/abs/2509.26030}, 
+}
+```
+
 *solve intelligence... then use that to solve everything else.* - Demis Hassabis

{x_transformers-2.8.2 → x_transformers-2.8.4}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "x-transformers"
-version = "2.8.2"
+version = "2.8.4"
 description = "X-Transformers"
 authors = [
     { name = "Phil Wang", email = "lucidrains@gmail.com" }

{x_transformers-2.8.2 → x_transformers-2.8.4}/tests/test_x_transformers.py

@@ -1360,3 +1360,16 @@ def test_vae():
     style = torch.randint(0, 256, (1, 1024))
 
     out = model.generate(seq[:, :512], 512, seq_for_latents = style)
+
+def test_muon_params():
+    from x_transformers import Attention, FeedForward, Encoder
+
+    attn = Attention(dim = 512, dim_out = 384)
+    assert len(list(attn.muon_parameters())) == 2
+
+    ff = FeedForward(dim = 512)
+
+    assert len(list(ff.muon_parameters())) == 2
+
+    enc = Encoder(dim = 512, depth = 2)
+    assert len(enc.muon_parameters()) > 0

{x_transformers-2.8.2 → x_transformers-2.8.4}/x_transformers/x_transformers.py

@@ -4,6 +4,11 @@ from typing import Callable
 import math
 from copy import deepcopy
 from random import random, randrange
+from functools import partial, wraps
+from itertools import chain
+from collections import namedtuple
+from contextlib import nullcontext
+from dataclasses import dataclass
 from packaging import version
 
 import torch
@@ -13,11 +18,6 @@ from torch import nn, einsum, tensor, Tensor, cat, stack, arange, is_tensor
 from torch.utils._pytree import tree_flatten, tree_unflatten, tree_map
 from torch.nn import Module, ModuleList, ModuleDict
 
-from functools import partial, wraps
-from collections import namedtuple
-from contextlib import nullcontext
-from dataclasses import dataclass
-
 from loguru import logger
 
 from x_transformers.attend import Attend, Intermediates
@@ -1279,6 +1279,17 @@ class FeedForward(Module):
         if zero_init_output:
             init_zero_(proj_out)
 
+    def muon_parameters(self):
+        weights = []
+
+        for m in self.modules():
+            if not isinstance(m, nn.Linear):
+                continue
+
+            weights.append(m.weight)
+
+        return weights
+
     def forward(
         self,
         x,
@@ -1644,6 +1655,9 @@ class Attention(Module):
         q_weight.mul_(qk_weight_scale)
         k_weight.mul_(qk_weight_scale)
 
+    def muon_parameters(self):
+        return chain(self.to_v.parameters(), self.to_out.parameters())
+
     def forward(
         self,
         x,
@@ -2479,6 +2493,17 @@ class AttentionLayers(Module):
         for attn_layer, attn_inter in zip(attn_layers, attn_intermeds):
             attn_layer.qk_clip_(attn_inter, tau = tau)
 
+    def muon_parameters(self):
+        params = []
+
+        for m in self.modules():
+            if not isinstance(m, (Attention, FeedForward)):
+                continue
+
+            params.extend(list(m.muon_parameters()))
+
+        return params
+
     def forward(
         self,
         x,
@@ -3216,6 +3241,9 @@ class TransformerWrapper(Module):
     ):
         self.attn_layers.attn_qk_clip_(intermediates, tau = tau)
 
+    def muon_parameters(self):
+        return self.attn_layers.muon_parameters()
+
     def forward(
         self,
         x,