x-transformers 1.40.0__py3-none-any.whl → 1.40.2__py3-none-any.whl

x_transformers/attend.py

@@ -119,7 +119,12 @@ def one_hot_straight_through(logits, temperature = 1.):
 # sparse topk attention - only keep topk attn logits for softmax
 # optional straight through with masked out logits by setting `attn_sparse_topk_straight_through = True`
 
-def sparse_topk_attn(logits, sparse_topk, temperature = 1., straight_through = False):
+def sparse_topk_attn(
+    logits,
+    sparse_topk,
+    temperature = 1.,
+    straight_through = False
+):
     orig_logits = logits
 
     mask_value = -torch.finfo(logits.dtype).max
@@ -132,7 +137,7 @@ def sparse_topk_attn(logits, sparse_topk, temperature = 1., straight_through = F
         return topk_attn
 
     soft_attn = (orig_logits / temperature).softmax(dim = -1)
-    return topk_attn + soft_attn - soft_attn.detach()
+    return topk_attn.detach() + soft_attn - soft_attn.detach()
 
 # functions for creating causal mask
 # need a special one for onnx cpu (no support for .triu)

x_transformers/x_transformers.py

@@ -1107,6 +1107,7 @@ class Attention(Module):
         context_mask = None,
         attn_mask = None,
         rel_pos = None,
+        attn_bias = None,
         rotary_pos_emb = None,
         prev_attn = None,
         mem = None,
@@ -1237,8 +1238,8 @@ class Attention(Module):
 
         # prepare relative positional bias, if needed
 
-        attn_bias = None
         if exists(rel_pos):
+            assert not exists(attn_bias)
             attn_bias = rel_pos(i, j)
             attn_bias = pad_at_dim(attn_bias, (num_mem_kv, 0), value = 0.) # handle memory key / values
 

{x_transformers-1.40.0.dist-info → x_transformers-1.40.2.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: x-transformers
-Version: 1.40.0
+Version: 1.40.2
 Summary: X-Transformers - Pytorch
 Home-page: https://github.com/lucidrains/x-transformers
 Author: Phil Wang

{x_transformers-1.40.0.dist-info → x_transformers-1.40.2.dist-info}/RECORD

@@ -1,15 +1,15 @@
 x_transformers/__init__.py,sha256=-MkQrSc37cTVDX7AOykxunYnqVtFlQ7lb0Cse5dsGWU,793
-x_transformers/attend.py,sha256=eoBEK0HdDCWaJgxwGZPeO36ydBt1NbB-gpij_Jkj4Mw,17212
+x_transformers/attend.py,sha256=VbB0fi-ETgAF4dc2a_Meaqvt14LMaRVIjZ8NexUX8F0,17239
 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
 x_transformers/multi_input.py,sha256=tCh-fTJDj2ib4SMGtsa-AM8MxKzJAQSwqAXOu3HU2mg,9252
 x_transformers/nonautoregressive_wrapper.py,sha256=2NU58hYMgn-4Jzg3mie-mXb0XH_dCN7fjlzd3K1rLUY,10510
-x_transformers/x_transformers.py,sha256=TGXJZXCWR5BiMkS5Kx-JhFQ85AxkiJabLiHnrCTC874,84562
+x_transformers/x_transformers.py,sha256=SfM0ql3wK7t8KzBXRNnGTdcyq3tQVmHB4VIcfg5sSv4,84604
 x_transformers/xl_autoregressive_wrapper.py,sha256=CvZMJ6A6PA-Y_bQAhnORwjJBSl6Vjq2IdW5KTdk8NI8,4195
 x_transformers/xval.py,sha256=QE1ltYZTR_eGgIHPP2BrMWVWVLqMW-OpDZh87BSmQEg,8563
-x_transformers-1.40.0.dist-info/LICENSE,sha256=As9u198X-U-vph5noInuUfqsAG2zX_oXPHDmdjwlPPY,1066
-x_transformers-1.40.0.dist-info/METADATA,sha256=WxBpjG7F8utkdJN1AF9vZMAfbltT0lmpdgtUErBbYQY,661
-x_transformers-1.40.0.dist-info/WHEEL,sha256=OVMc5UfuAQiSplgO0_WdW7vXVGAt9Hdd6qtN4HotdyA,91
-x_transformers-1.40.0.dist-info/top_level.txt,sha256=hO6KGpFuGucRNEtRfme4A_rGcM53AKwGP7RVlRIxS5Q,15
-x_transformers-1.40.0.dist-info/RECORD,,
+x_transformers-1.40.2.dist-info/LICENSE,sha256=As9u198X-U-vph5noInuUfqsAG2zX_oXPHDmdjwlPPY,1066
+x_transformers-1.40.2.dist-info/METADATA,sha256=G1LWuKpy25e1rXV7MFRT5r2F4bHjvyMUrgxTgJIQLic,661
+x_transformers-1.40.2.dist-info/WHEEL,sha256=OVMc5UfuAQiSplgO0_WdW7vXVGAt9Hdd6qtN4HotdyA,91
+x_transformers-1.40.2.dist-info/top_level.txt,sha256=hO6KGpFuGucRNEtRfme4A_rGcM53AKwGP7RVlRIxS5Q,15
+x_transformers-1.40.2.dist-info/RECORD,,