x-transformers 1.42.5__py3-none-any.whl → 1.42.7__py3-none-any.whl

x_transformers/x_transformers.py

@@ -1055,7 +1055,12 @@ class Attention(Module):
         logit_softclamp_value = 50.,
         neutreno_value_residual = False, # Nguyen et al. https://arxiv.org/abs/2312.00751
         neutreno_alpha = 0.4,
-        onnxable = False
+        onnxable = False,
+        attend_sdp_kwargs: dict = dict(
+            enable_flash = True,
+            enable_math = True,
+            enable_mem_efficient = True
+        )
     ):
         super().__init__()
         dim_kv = default(dim_context, dim)
@@ -1188,7 +1193,8 @@ class Attention(Module):
             softclamp_logits = softclamp_logits,
             logit_softclamp_value = logit_softclamp_value,
             cope = cope,
-            onnxable = onnxable
+            onnxable = onnxable,
+            sdp_kwargs = attend_sdp_kwargs
         )
 
         # head scaling
@@ -1240,6 +1246,7 @@ class Attention(Module):
         rel_pos = None,
         attn_bias = None,
         rotary_pos_emb = None,
+        pos = None, # for custom alibi positions
         prev_attn = None,
         mem = None,
         mem_mask = None,
@@ -1386,7 +1393,14 @@ class Attention(Module):
 
         if exists(rel_pos):
             assert not exists(attn_bias)
-            attn_bias = rel_pos(i, j)
+
+            if exists(pos):
+                assert isinstance(rel_pos, AlibiPositionalBias), 'only alibi allowed for custom positions at the moment'
+                # allow for custom positions to be passed in
+                attn_bias = rel_pos.forward_custom_pos(pos)
+            else:
+                attn_bias = rel_pos(i, j)
+
             attn_bias = pad_at_dim(attn_bias, (num_mem_kv, 0), value = 0.) # handle memory key / values
 
         # prepare data dependent alibi from forgetting transformers paper, if needed
@@ -1837,6 +1851,7 @@ class AttentionLayers(Module):
         cache_age = 1,
         return_hiddens = False,
         rotary_pos_emb = None,
+        pos = None,
         attn_bias = None,
         condition = None,
         in_attn_cond = None, # https://arxiv.org/abs/2105.04090
@@ -1900,7 +1915,9 @@ class AttentionLayers(Module):
             maybe_mem = mems[0] # todo - handle edge case where different layers get different memory lengths. don't think this will ever come up but who knows
             mem_len = maybe_mem.shape[1] if exists(maybe_mem) else 0
 
-            pos = torch.arange(x.shape[1] + mem_len, device = x.device) - mem_len
+            if not exists(pos):
+                pos = torch.arange(x.shape[1] + mem_len, device = x.device) - mem_len
+
             rotary_pos_emb = self.rotary_pos_emb(pos)
 
         # assume cached key / values
@@ -2024,7 +2041,7 @@ class AttentionLayers(Module):
             # forward depending on layer type
 
             if layer_type == 'a':
-                out, inter = block(x, mask = mask, context_mask = self_attn_kv_mask, attn_mask = attn_mask, rel_pos = self.rel_pos, rotary_pos_emb = rotary_pos_emb, prev_attn = prev_attn, cache = next(iter_attn_cache, None), mem = layer_mem, mem_mask = layer_mem_mask, attn_bias = attn_bias, value_residual = maybe_self_attn_value_residual, return_intermediates = True)
+                out, inter = block(x, mask = mask, context_mask = self_attn_kv_mask, attn_mask = attn_mask, rel_pos = self.rel_pos, pos = pos, rotary_pos_emb = rotary_pos_emb, prev_attn = prev_attn, cache = next(iter_attn_cache, None), mem = layer_mem, mem_mask = layer_mem_mask, attn_bias = attn_bias, value_residual = maybe_self_attn_value_residual, return_intermediates = True)
             elif layer_type == 'c':
                 out, inter = block(x, context = context, mask = mask, context_mask = context_mask, prev_attn = prev_cross_attn, cache = next(iter_attn_cache, None), value_residual = maybe_cross_attn_value_residual, return_intermediates = True)
             elif layer_type == 'f':

{x_transformers-1.42.5.dist-info → x_transformers-1.42.7.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: x-transformers
-Version: 1.42.5
+Version: 1.42.7
 Summary: X-Transformers - Pytorch
 Home-page: https://github.com/lucidrains/x-transformers
 Author: Phil Wang

{x_transformers-1.42.5.dist-info → x_transformers-1.42.7.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=KIR7efx59xl0BVshU1e6RO0YKgz7zeYBXITDNYWJ4mQ,92506
+x_transformers/x_transformers.py,sha256=6jXSMHViCU64gLMbxRJ6C8bgcLrPFbT-m-fhtusqq3g,93117
 x_transformers/xl_autoregressive_wrapper.py,sha256=CvZMJ6A6PA-Y_bQAhnORwjJBSl6Vjq2IdW5KTdk8NI8,4195
 x_transformers/xval.py,sha256=7S00kCuab4tWQa-vf-z-XfzADjVj48MoFIr7VSIvttg,8575
-x_transformers-1.42.5.dist-info/LICENSE,sha256=As9u198X-U-vph5noInuUfqsAG2zX_oXPHDmdjwlPPY,1066
-x_transformers-1.42.5.dist-info/METADATA,sha256=cLnay5nt6F6GKdghsaqHiZQsVmJ9dS5l-IDozZIs3ec,689
-x_transformers-1.42.5.dist-info/WHEEL,sha256=P9jw-gEje8ByB7_hXoICnHtVCrEwMQh-630tKvQWehc,91
-x_transformers-1.42.5.dist-info/top_level.txt,sha256=hO6KGpFuGucRNEtRfme4A_rGcM53AKwGP7RVlRIxS5Q,15
-x_transformers-1.42.5.dist-info/RECORD,,
+x_transformers-1.42.7.dist-info/LICENSE,sha256=As9u198X-U-vph5noInuUfqsAG2zX_oXPHDmdjwlPPY,1066
+x_transformers-1.42.7.dist-info/METADATA,sha256=tM7s2gIMFH8hy_YZY84BhZ-yUoH6PTyjusK0dMOpTN8,689
+x_transformers-1.42.7.dist-info/WHEEL,sha256=P9jw-gEje8ByB7_hXoICnHtVCrEwMQh-630tKvQWehc,91
+x_transformers-1.42.7.dist-info/top_level.txt,sha256=hO6KGpFuGucRNEtRfme4A_rGcM53AKwGP7RVlRIxS5Q,15
+x_transformers-1.42.7.dist-info/RECORD,,