x-transformers 1.43.5__tar.gz → 1.44.1__tar.gz

{x_transformers-1.43.5/x_transformers.egg-info → x_transformers-1.44.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: x-transformers
-Version: 1.43.5
+Version: 1.44.1
 Summary: X-Transformers - Pytorch
 Home-page: https://github.com/lucidrains/x-transformers
 Author: Phil Wang

{x_transformers-1.43.5 → x_transformers-1.44.1}/README.md

@@ -317,7 +317,9 @@ model = TransformerWrapper(
 
 Update: MetaAI researchers <a href="https://arxiv.org/abs/2309.16588">have found</a> that adding memory tokens (they call them register tokens), alleviates outliers (which is suspected now to be a pathology of attention networks unable to <a href="https://arxiv.org/abs/2306.12929">attend to nothing</a>).
 
-Update 2: a hybrid architecture out of Nvidia named <a href="https://openreview.net/forum?id=A1ztozypga">Hymba</a> used memory tokens successfully in the autoregressive case, termed meta tokens in their paper
+Update 2: a hybrid architecture out of Nvidia named <a href="https://openreview.net/forum?id=A1ztozypga">Hymba</a> used memory tokens successfully in the autoregressive case, termed meta tokens in their paper.
+
+Update 3: further corroborated by <a href="https://arxiv.org/abs/2501.00663">a paper</a> trying to extend memory in attention networks, termed persistent memory
 
 ### Transformers Without Tears
 
@@ -2374,4 +2376,15 @@ ids_out, num_out, is_number_mask = model.generate(start_ids, start_nums, 17)
 }
 ```
 
+```bibtex
+@inproceedings{anonymous2024hymba,
+    title   = {Hymba: A Hybrid-head Architecture for Small Language Models},
+    author  = {Anonymous},
+    booktitle = {Submitted to The Thirteenth International Conference on Learning Representations},
+    year    = {2024},
+    url     = {https://openreview.net/forum?id=A1ztozypga},
+    note    = {under review}
+}
+```
+
 *solve intelligence... then use that to solve everything else.* - Demis Hassabis

{x_transformers-1.43.5 → x_transformers-1.44.1}/setup.py

@@ -3,7 +3,7 @@ from setuptools import setup, find_packages
 setup(
   name = 'x-transformers',
   packages = find_packages(exclude=['examples']),
-  version = '1.43.5',
+  version = '1.44.1',
   license='MIT',
   description = 'X-Transformers - Pytorch',
   author = 'Phil Wang',

{x_transformers-1.43.5 → x_transformers-1.44.1}/tests/test_x_transformers.py

@@ -613,3 +613,37 @@ def test_hyper_connections(tanh):
     x = torch.randint(0, 20000, (2, 1024))
 
     model(x)
+
+def test_hybrid():
+    from torch.nn import GRU
+
+    dec = TransformerWrapper(
+        num_tokens = 20000,
+        max_seq_len = 1024,
+        attn_layers = Decoder(
+            dim = 128,
+            depth = 6,
+            heads = 8,
+            attn_dim_head = 64,
+            attn_hybrid_module = GRU(128, 64 * 8, batch_first = True)
+        )
+    )
+
+    x = torch.randint(0, 20000, (2, 1024))
+
+    embed = dec(x)
+
+    enc = TransformerWrapper(
+        num_tokens = 20000,
+        max_seq_len = 1024,
+        attn_layers = Encoder(
+            dim = 128,
+            depth = 6,
+            heads = 8,
+            attn_dim_head = 64,
+            attn_hybrid_module = GRU(128, 64 * 4, batch_first = True, bidirectional = True)
+        )
+    )
+
+    mask = torch.randint(0, 2, (2, 1024)).bool()
+    embed = enc(x, mask = mask)

{x_transformers-1.43.5 → x_transformers-1.44.1}/x_transformers/x_transformers.py

@@ -2,14 +2,16 @@ from __future__ import annotations
 from typing import Callable
 
 import math
+from copy import deepcopy
 from random import random, randrange
 from packaging import version
 
 import torch
+from torch.amp import autocast
 import torch.nn.functional as F
 from torch import nn, einsum, Tensor
+from torch.utils._pytree import tree_flatten
 from torch.nn import Module, ModuleList, ModuleDict
-from torch.amp import autocast
 
 from functools import partial, wraps
 from collections import namedtuple
@@ -1136,6 +1138,8 @@ class Attention(Module):
         sigmoid = False,
         selective = False,
         custom_attn_fn: Callable | None = None,
+        hybrid_module: Module | None = None,
+        hybrid_mask_kwarg: str | None = None,
         one_kv_head = False,
         kv_heads = None,
         shared_kv = False,
@@ -1335,6 +1339,12 @@ class Attention(Module):
 
         self.attn_on_attn = on_attn
 
+        # hybrid module, in same vein as hymba https://www.arxiv.org/abs/2411.13676
+
+        self.hybrid_module = deepcopy(hybrid_module) if exists(hybrid_module) else None
+
+        self.hybrid_mask_kwarg = hybrid_mask_kwarg # for bidirectional, can forward `mask` into the hybrid module and let it handle variable lengths
+
         # output dimension by default same as input, but can be overridden
 
         dim_out = default(dim_out, dim)
@@ -1407,6 +1417,16 @@ class Attention(Module):
                 value_residual_mix = self.to_value_residual_mix(q_input)
                 v = v * value_residual_mix + value_residual * (1. - value_residual_mix)
 
+        # qk normalization
+
+        if self.qk_norm:
+            qk_l2norm = partial(l2norm, groups = self.qk_norm_groups)
+            q, k = map(qk_l2norm, (q, k))
+            scale = self.qk_norm_scale
+
+            q = q * self.qk_norm_q_scale
+            k = k * self.qk_norm_k_scale
+
         # take care of caching
 
         if exists(cache):
@@ -1427,14 +1447,6 @@ class Attention(Module):
             mem_len = mem.shape[-2] if exists(mem) else 0
             cached_kv = (k[..., mem_len:, :], v[..., mem_len:, :])
 
-        if self.qk_norm:
-            qk_l2norm = partial(l2norm, groups = self.qk_norm_groups)
-            q, k = map(qk_l2norm, (q, k))
-            scale = self.qk_norm_scale
-
-            q = q * self.qk_norm_q_scale
-            k = k * self.qk_norm_k_scale
-
         if exists(rotary_pos_emb):
             freqs, xpos_scale = rotary_pos_emb
             q_xpos_scale, k_xpos_scale = (xpos_scale, xpos_scale ** -1.) if exists(xpos_scale) else (1., 1.)
@@ -1581,6 +1593,26 @@ class Attention(Module):
 
         out = rearrange(out, 'b h n d -> b n (h d)')
 
+        # hybrid module
+
+        if exists(self.hybrid_module):
+
+            # hybrid input
+
+            hybrid_forward_kwargs = dict()
+
+            if not self.causal and exists(self.hybrid_mask_kwarg):
+                hybrid_forward_kwargs = {self.hybrid_mask_kwarg: mask}
+
+            # hybrid forward
+
+            hybrid_outputs = self.hybrid_module(x, **hybrid_forward_kwargs)
+
+            # handle hybrid out
+
+            (hybrid_out, *rest_hybrid_outs), _ = tree_flatten(hybrid_outs)
+            out = 0.5 * (out + hybrid_out)
+
         # alphafold2 styled gating of the values
 
         if exists(self.to_v_gate):
@@ -2003,7 +2035,7 @@ class AttentionLayers(Module):
 
         # determine whether can cache kv
 
-        self.can_cache_kv = all([module.can_cache_kv for module in self.modules() if isinstance(module, Attention) ])
+        self.can_cache_kv = all([module.can_cache_kv for module in self.modules() if isinstance(module, Attention)])
 
     def forward(
         self,

{x_transformers-1.43.5 → x_transformers-1.44.1/x_transformers.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: x-transformers
-Version: 1.43.5
+Version: 1.44.1
 Summary: X-Transformers - Pytorch
 Home-page: https://github.com/lucidrains/x-transformers
 Author: Phil Wang