x-transformers 1.37.5__py3-none-any.whl → 1.37.7__py3-none-any.whl

x_transformers/attend.py

@@ -64,15 +64,15 @@ print_once = once(print)
 
 # alternative distance functions
 
-def qk_l2_distance(q, k):
+def qk_l2_dist_squared(q, k):
     if k.ndim == 3:
         k = repeat(k, 'b j d -> b h j d', h = q.shape[1])
 
     q, packed_shape = pack_one(q, '* i d')
     k, _ = pack_one(k, '* j d')
 
-    distance = torch.cdist(q, k)
-    return unpack_one(distance, packed_shape, '* i j')
+    l2_dist_squared = torch.cdist(q, k) ** 2
+    return unpack_one(l2_dist_squared, packed_shape, '* i j')
 
 # functions for creating causal mask
 # need a special one for onnx cpu (no support for .triu)
@@ -145,7 +145,6 @@ class Attend(Module):
 
         # l2 distance attention
 
-        assert not (flash and l2_distance), 'l2 distance attention does not work with flash attention just yet'
         self.l2_distance = l2_distance
 
         # add a key / value token composed of zeros
@@ -208,6 +207,17 @@ class Attend(Module):
         if v.ndim == 3:
             v = repeat(v, 'b ... -> b h ...', h = q.shape[1])
 
+        # handle maybe l2 distance
+
+        if self.l2_distance:
+            k_norm_sq = k.norm(dim = -1, keepdim = True) ** 2
+            k = F.pad(k, (0, 1), value = 1.)
+            k = torch.cat((k, -k_norm_sq), dim = -1)
+
+            q_norm_sq = q.norm(dim = -1, keepdim = True) ** 2
+            q = torch.cat((2 * q, -q_norm_sq), dim = -1)
+            q = F.pad(q, (0, 1), value = 1.)
+
         # handle scale - by default they scale by dim_head ** -0.5, but need to take care if using cosine sim attention
 
         if exists(self.scale):
@@ -353,7 +363,7 @@ class Attend(Module):
         if not self.l2_distance:
             sim = einsum(f'b h i d, {kv_einsum_eq} -> b h i j', q, k)
         else:
-            sim = -qk_l2_distance(q, k)
+            sim = -qk_l2_dist_squared(q, k)
 
         sim = sim * scale
 

{x_transformers-1.37.5.dist-info → x_transformers-1.37.7.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: x-transformers
-Version: 1.37.5
+Version: 1.37.7
 Summary: X-Transformers - Pytorch
 Home-page: https://github.com/lucidrains/x-transformers
 Author: Phil Wang

{x_transformers-1.37.5.dist-info → x_transformers-1.37.7.dist-info}/RECORD

@@ -1,5 +1,5 @@
 x_transformers/__init__.py,sha256=-MkQrSc37cTVDX7AOykxunYnqVtFlQ7lb0Cse5dsGWU,793
-x_transformers/attend.py,sha256=4RnX1yhWZIf8holucqnYXTIP7U1m40UpP58RZNT_2sM,13128
+x_transformers/attend.py,sha256=BOnMjgV5O5DyAM_bUz1rI6n1j_eLXu8GIEljT-MMnWU,13434
 x_transformers/autoregressive_wrapper.py,sha256=DOJJCMMDOqDYKWy_IaG5IyKsXD3AW6amzfUgdAADOLY,10500
 x_transformers/continuous.py,sha256=cIVEdhfei258__ziV7kQBrJMxCel54bExBTDrO9rfCI,6450
 x_transformers/dpo.py,sha256=LjvWgCkqTl-UuehrzQ8nkX5guLr4whYwsmm7SKSwdls,3450
@@ -8,8 +8,8 @@ x_transformers/nonautoregressive_wrapper.py,sha256=2NU58hYMgn-4Jzg3mie-mXb0XH_dC
 x_transformers/x_transformers.py,sha256=-2fj6QcDSfMI5lJA_fzOW2mdzdS1C1LD6jMBtGQY48E,83752
 x_transformers/xl_autoregressive_wrapper.py,sha256=CvZMJ6A6PA-Y_bQAhnORwjJBSl6Vjq2IdW5KTdk8NI8,4195
 x_transformers/xval.py,sha256=QE1ltYZTR_eGgIHPP2BrMWVWVLqMW-OpDZh87BSmQEg,8563
-x_transformers-1.37.5.dist-info/LICENSE,sha256=As9u198X-U-vph5noInuUfqsAG2zX_oXPHDmdjwlPPY,1066
-x_transformers-1.37.5.dist-info/METADATA,sha256=zHUhvP1bQjFbMtxnVO9iDESgXpGOQxuBCsm4b6K1w44,661
-x_transformers-1.37.5.dist-info/WHEEL,sha256=GV9aMThwP_4oNCtvEC2ec3qUYutgWeAzklro_0m4WJQ,91
-x_transformers-1.37.5.dist-info/top_level.txt,sha256=hO6KGpFuGucRNEtRfme4A_rGcM53AKwGP7RVlRIxS5Q,15
-x_transformers-1.37.5.dist-info/RECORD,,
+x_transformers-1.37.7.dist-info/LICENSE,sha256=As9u198X-U-vph5noInuUfqsAG2zX_oXPHDmdjwlPPY,1066
+x_transformers-1.37.7.dist-info/METADATA,sha256=svdldk1hpiBN4xxKkOaQuMEXAxvaj6fT7Ri9NXwZJCU,661
+x_transformers-1.37.7.dist-info/WHEEL,sha256=GV9aMThwP_4oNCtvEC2ec3qUYutgWeAzklro_0m4WJQ,91
+x_transformers-1.37.7.dist-info/top_level.txt,sha256=hO6KGpFuGucRNEtRfme4A_rGcM53AKwGP7RVlRIxS5Q,15
+x_transformers-1.37.7.dist-info/RECORD,,