x-transformers 1.43.5__py3-none-any.whl → 1.44.1__py3-none-any.whl

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.dist-info → x_transformers-1.44.1.dist-info}/METADATA

@@ -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.dist-info → x_transformers-1.44.1.dist-info}/RECORD

@@ -6,11 +6,11 @@ x_transformers/dpo.py,sha256=xt4OuOWhU8pN3OKN2LZAaC2NC8iiEnchqqcrPWVqf0o,3521
 x_transformers/multi_input.py,sha256=tCh-fTJDj2ib4SMGtsa-AM8MxKzJAQSwqAXOu3HU2mg,9252
 x_transformers/neo_mlp.py,sha256=XCNnnop9WLarcxap1kGuYc1x8GHvwkZiDRnXOxSl3Po,3452
 x_transformers/nonautoregressive_wrapper.py,sha256=2NU58hYMgn-4Jzg3mie-mXb0XH_dCN7fjlzd3K1rLUY,10510
-x_transformers/x_transformers.py,sha256=z3RH6jvjcxaAVfZoCS0HWrE0Gy55-eXOKtzRt7rRRIw,100811
+x_transformers/x_transformers.py,sha256=yjtB4kV4N9mzHdliIM9MjyA6SoMtvpzc2Z4iU6R9_Uc,101859
 x_transformers/xl_autoregressive_wrapper.py,sha256=CvZMJ6A6PA-Y_bQAhnORwjJBSl6Vjq2IdW5KTdk8NI8,4195
 x_transformers/xval.py,sha256=7S00kCuab4tWQa-vf-z-XfzADjVj48MoFIr7VSIvttg,8575
-x_transformers-1.43.5.dist-info/LICENSE,sha256=As9u198X-U-vph5noInuUfqsAG2zX_oXPHDmdjwlPPY,1066
-x_transformers-1.43.5.dist-info/METADATA,sha256=crd2xA3NbodKbOz9xY4D1j3XDbTmaY9vwkZZJOGoEw4,738
-x_transformers-1.43.5.dist-info/WHEEL,sha256=PZUExdf71Ui_so67QXpySuHtCi3-J3wvF4ORK6k_S8U,91
-x_transformers-1.43.5.dist-info/top_level.txt,sha256=hO6KGpFuGucRNEtRfme4A_rGcM53AKwGP7RVlRIxS5Q,15
-x_transformers-1.43.5.dist-info/RECORD,,
+x_transformers-1.44.1.dist-info/LICENSE,sha256=As9u198X-U-vph5noInuUfqsAG2zX_oXPHDmdjwlPPY,1066
+x_transformers-1.44.1.dist-info/METADATA,sha256=Zw_Rscb4vNZxlKosWSHSQy4EsICF45U58K0hipxydpQ,738
+x_transformers-1.44.1.dist-info/WHEEL,sha256=A3WOREP4zgxI0fKrHUG8DC8013e3dK3n7a6HDbcEIwE,91
+x_transformers-1.44.1.dist-info/top_level.txt,sha256=hO6KGpFuGucRNEtRfme4A_rGcM53AKwGP7RVlRIxS5Q,15
+x_transformers-1.44.1.dist-info/RECORD,,

{x_transformers-1.43.5.dist-info → x_transformers-1.44.1.dist-info}/WHEEL

@@ -1,5 +1,5 @@
 Wheel-Version: 1.0
-Generator: setuptools (75.6.0)
+Generator: setuptools (75.7.0)
 Root-Is-Purelib: true
 Tag: py3-none-any