x-transformers 1.42.7__py3-none-any.whl → 1.42.9__py3-none-any.whl

x_transformers/x_transformers.py

@@ -238,6 +238,13 @@ class TokenEmbedding(Module):
         token_emb = self.emb(x.long())
         return l2norm(token_emb) if self.l2norm_embed else token_emb
 
+    def init_(self):
+        if self.l2norm_embed:
+            nn.init.normal_(self.emb.weight, std=1e-5)
+            return
+        nn.init.kaiming_normal_(self.emb.weight)
+
+
 # positional embeddings
 
 class AbsolutePositionalEmbedding(Module):
@@ -445,13 +452,20 @@ class DynamicPositionBias(Module):
         return bias
 
 class AlibiPositionalBias(Module):
-    def __init__(self, heads, total_heads = None, **kwargs):
+    def __init__(
+        self,
+        heads,
+        total_heads = None,
+        slopes: list[int] | None = None,
+        **kwargs
+    ):
         super().__init__()
         self.heads = heads
         self.total_heads = default(total_heads, heads)
 
-        slopes = Tensor(self._get_slopes(heads))
+        slopes = Tensor(default(slopes, self._get_slopes(heads)))
         slopes = rearrange(slopes, 'h -> h 1 1')
+
         self.register_buffer('slopes', slopes, persistent = False)
         self.register_buffer('bias', None, persistent = False)
     
@@ -480,7 +494,10 @@ class AlibiPositionalBias(Module):
         h, device = self.total_heads, self.device
 
         pos_j = default(pos_j, pos_i)
-        bias = -einx.subtract('... j, ... i -> ... 1 i j', pos_j, pos_i).abs()
+        bias = -einx.subtract('... j, ... i -> ... i j', pos_j, pos_i).abs()
+
+        if bias.ndim == 3:
+            bias = rearrange(bias, 'b i j -> b 1 i j')
 
         bias = bias * self.slopes
         num_heads_unalibied = h - bias.shape[-3]
@@ -1524,8 +1541,9 @@ class AttentionLayers(Module):
         use_layerscale = False,
         layerscale_init_value = 0.,
         unet_skips = False,
-        reinject_input = False, # seen first in DEQ paper https://arxiv.org/abs/1909.01377, but later used in a number of papers trying to achieve depthwise generalization https://arxiv.org/abs/2410.03020v1
-        add_value_residual = False, # resformer from Zhou et al - https://arxiv.org/abs/2410.17897v1
+        reinject_input = False,         # seen first in DEQ paper https://arxiv.org/abs/1909.01377, but later used in a number of papers trying to achieve depthwise generalization https://arxiv.org/abs/2410.03020v1
+        add_value_residual = False,     # resformer from Zhou et al - https://arxiv.org/abs/2410.17897v1
+        rel_pos_kwargs: dict = dict(),
         **kwargs
     ):
         super().__init__()
@@ -1566,14 +1584,14 @@ class AttentionLayers(Module):
 
         if rel_pos_bias:
             assert not flash_attn, 'flash attention not compatible with t5 relative positional bias'
-            self.rel_pos = RelativePositionBias(scale = dim_head ** 0.5, causal = causal, heads = heads, num_buckets = rel_pos_num_buckets, max_distance = rel_pos_max_distance)
+            self.rel_pos = RelativePositionBias(scale = dim_head ** 0.5, causal = causal, heads = heads, num_buckets = rel_pos_num_buckets, max_distance = rel_pos_max_distance, **rel_pos_kwargs)
         elif dynamic_pos_bias:
             assert not flash_attn, 'flash attention not compatible with dynamic positional bias'
-            self.rel_pos = DynamicPositionBias(dim = dim // 4, heads = heads, log_distance = dynamic_pos_bias_log_distance, depth = dynamic_pos_bias_mlp_depth, norm = dynamic_pos_bias_norm)
+            self.rel_pos = DynamicPositionBias(dim = dim // 4, heads = heads, log_distance = dynamic_pos_bias_log_distance, depth = dynamic_pos_bias_mlp_depth, norm = dynamic_pos_bias_norm, **rel_pos_kwargs)
         elif alibi_pos_bias:
             alibi_num_heads = default(alibi_num_heads, heads)
             assert alibi_num_heads <= heads, 'number of ALiBi heads must be less than the total number of heads'
-            self.rel_pos = AlibiPositionalBias(heads = alibi_num_heads, total_heads = heads)
+            self.rel_pos = AlibiPositionalBias(heads = alibi_num_heads, total_heads = heads, **rel_pos_kwargs)
 
         assert at_most_one_of(sandwich_norm, resi_dual), 'either sandwich norm or resiDual is selected, but not both'
         assert not (not pre_norm and sandwich_norm), 'sandwich norm cannot be used when not using prenorm'
@@ -2261,7 +2279,8 @@ class TransformerWrapper(Module):
         token_emb: TokenEmbedding | None = None,
         mixture_of_softmax = False,
         mixture_of_softmax_k = 4,
-        sigsoftmax_logits = False
+        sigsoftmax_logits = False,
+        to_logits: Module | None = None,
     ):
         super().__init__()
 
@@ -2363,11 +2382,12 @@ class TransformerWrapper(Module):
         if return_only_embed:
             self.to_logits = None
         elif tie_embedding:
+            assert isinstance(token_emb, TokenEmbedding), 'can only tie embedding if using `TokenEmbedding`'
             self.to_logits = lambda t: t @ self.token_emb.emb.weight.t()
         elif num_output_heads > 1:
             self.to_logits = ModuleList([LinearNoBias(dim, logits_dim) for _ in range(num_output_heads)])
         else:
-            self.to_logits = LinearNoBias(dim, logits_dim)
+            self.to_logits = LinearNoBias(dim, logits_dim) if not exists(to_logits) else to_logits
 
         # memory tokens (like [cls]) from Memory Transformers paper
 
@@ -2388,13 +2408,12 @@ class TransformerWrapper(Module):
         self.can_cache_kv_outside_max_seq_len = no_abs_pos_emb
 
     def init_(self):
+        if hasattr(self.token_emb, 'init_'):
+            self.token_emb.init_()
+
         if self.l2norm_embed:
-            nn.init.normal_(self.token_emb.emb.weight, std = 1e-5)
             if not isinstance(self.pos_emb, always):
                 nn.init.normal_(self.pos_emb.emb.weight, std = 1e-5)
-            return
-
-        nn.init.kaiming_normal_(self.token_emb.emb.weight)
 
     def forward(
         self,
@@ -2417,7 +2436,9 @@ class TransformerWrapper(Module):
         attn_z_loss_weight = 1e-4,
         seq_start_pos = None,
         cache: LayerIntermediates | None = None,
-        **kwargs
+        token_emb_kwargs = dict(),
+        to_logits_kwargs = dict(),
+        **kwargs,
     ):
         b, n, device, num_mems, has_memory_tokens, emb_frac_gradient, orig_mask = x.shape[0], x.shape[1], x.device, self.num_memory_tokens, self.num_memory_tokens > 0, self.emb_frac_gradient, mask
 
@@ -2428,7 +2449,7 @@ class TransformerWrapper(Module):
 
         external_pos_emb = exists(pos) and pos.dtype != torch.long
         pos_emb = self.pos_emb(x, pos = pos, seq_start_pos = seq_start_pos) if not external_pos_emb else pos
-        x = self.token_emb(x) + pos_emb
+        x = self.token_emb(x, **token_emb_kwargs) + pos_emb
 
         # add additional embeddings
 
@@ -2583,9 +2604,9 @@ class TransformerWrapper(Module):
 
         if not return_embeddings:
             if self.has_multiple_heads:
-                logits = tuple(fn(x) for fn in self.to_logits)
+                logits = tuple(fn(x, **to_logits_kwargs) for fn in self.to_logits)
             else:
-                logits = self.to_logits(x)
+                logits = self.to_logits(x, **to_logits_kwargs)
 
         # maybe sig softmax
 

{x_transformers-1.42.7.dist-info → x_transformers-1.42.9.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: x-transformers
-Version: 1.42.7
+Version: 1.42.9
 Summary: X-Transformers - Pytorch
 Home-page: https://github.com/lucidrains/x-transformers
 Author: Phil Wang

{x_transformers-1.42.7.dist-info → x_transformers-1.42.9.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=6jXSMHViCU64gLMbxRJ6C8bgcLrPFbT-m-fhtusqq3g,93117
+x_transformers/x_transformers.py,sha256=VxdA44EYQhVH1Rp7wreJ83I2e0Ea7VN_bFRE-iDXOI8,93833
 x_transformers/xl_autoregressive_wrapper.py,sha256=CvZMJ6A6PA-Y_bQAhnORwjJBSl6Vjq2IdW5KTdk8NI8,4195
 x_transformers/xval.py,sha256=7S00kCuab4tWQa-vf-z-XfzADjVj48MoFIr7VSIvttg,8575
-x_transformers-1.42.7.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,,
+x_transformers-1.42.9.dist-info/LICENSE,sha256=As9u198X-U-vph5noInuUfqsAG2zX_oXPHDmdjwlPPY,1066
+x_transformers-1.42.9.dist-info/METADATA,sha256=k9r-D0b0xnf8gwE-SwwgybnfQpoRwiY0wthOn66xc6Y,689
+x_transformers-1.42.9.dist-info/WHEEL,sha256=P9jw-gEje8ByB7_hXoICnHtVCrEwMQh-630tKvQWehc,91
+x_transformers-1.42.9.dist-info/top_level.txt,sha256=hO6KGpFuGucRNEtRfme4A_rGcM53AKwGP7RVlRIxS5Q,15
+x_transformers-1.42.9.dist-info/RECORD,,