x-transformers 2.11.24__py3-none-any.whl → 2.12.1__py3-none-any.whl

x_transformers/x_transformers.py

@@ -779,6 +779,49 @@ def apply_rotary_pos_emb(t, freqs, scale = 1):
 
     return out.type(orig_dtype)
 
+class PolarEmbedding(Module):
+    """ https://arxiv.org/abs/2509.10534 """
+
+    def __init__(
+        self,
+        dim,
+        bias_uniform_init = False,
+        base = 10000,
+    ):
+        super().__init__()
+        inv_freq = 1. / (base ** (arange(0, dim).float() / dim))
+        self.register_buffer('inv_freq', inv_freq)
+
+        self.learned_bias = nn.Parameter(torch.zeros(dim))
+
+        if bias_uniform_init:
+            self.learned_bias.uniform_(-2. * math.pi, 0.)
+
+    @autocast('cuda', enabled = False)
+    def forward(self, t, offset = 0):
+        max_pos = t.max() + 1
+
+        if t.ndim == 1:
+            t = rearrange(t, 'n -> 1 n')
+
+        freqs = torch.einsum('b i , j -> b i j', t.type_as(self.inv_freq), self.inv_freq)
+
+        bias = self.learned_bias.clamp(-2. * math.pi, 0.)
+
+        return freqs, bias
+
+@autocast('cuda', enabled = False)
+def apply_polar_pos_emb(t, freqs):
+    rot_dim, seq_len, orig_dtype = freqs.shape[-1], t.shape[-2], t.dtype
+    freqs = freqs[:, -seq_len:]
+
+    t = t.float()
+
+    t = F.softplus(t)
+    out = cat((t * freqs.cos(), t * freqs.sin()), dim = -1)
+
+    return out.type(orig_dtype)
+
 # norms
 
 class Scale(Module):
@@ -1745,6 +1788,7 @@ class Attention(Module):
         attn_bias = None,
         rotary_pos_emb = None,
         context_rotary_pos_emb = None,
+        polar_pos_emb = None,
         pos = None, # for custom alibi positions
         prev_attn = None,
         mem = None,
@@ -1896,6 +1940,11 @@ class Attention(Module):
                 q = cat((q_rest, q), dim = 1)
                 k = cat((k_rest, k), dim = 1)
 
+        if exists(polar_pos_emb):
+            freqs, bias = polar_pos_emb
+            q = apply_polar_pos_emb(q, freqs)
+            k = apply_polar_pos_emb(k, freqs + bias)
+
         input_mask = context_mask
 
         if not exists(input_mask) and not has_context:
@@ -2174,6 +2223,8 @@ class AttentionLayers(Module):
         rotary_xpos_scale_base = 512,
         rotary_base_rescale_factor = 1.,
         rotate_num_heads = None,
+        polar_pos_emb = False,
+        polar_bias_uniform_init = False,
         weight_tie_layers = False,
         custom_layers: tuple[str, ...] | None = None,
         layers_execute_order: tuple[int, ...] | None = None,
@@ -2250,14 +2301,13 @@ class AttentionLayers(Module):
 
         # LIMe
 
-        hiddens_counter = 0
         self.layer_integrators = ModuleList([])
 
         assert not (qkv_receive_diff_residuals and not (hyper_conn_produce_diff_views or integrate_layers))
 
         # positions related
 
-        self.disable_abs_pos_emb = default(disable_abs_pos_emb, (rel_pos_bias or rotary_pos_emb))
+        self.disable_abs_pos_emb = default(disable_abs_pos_emb, (rel_pos_bias or rotary_pos_emb or polar_pos_emb))
 
         rotary_emb_dim = default(rotary_emb_dim, dim_head // 2)
 
@@ -2266,9 +2316,14 @@ class AttentionLayers(Module):
         if verbose and rotary_emb_dim < 32:
             logger.warning('when training language model, rotary embedding dimension should be at least 32')
 
+        assert at_most_one_of(rotary_pos_emb, polar_pos_emb), f'either rotary positional embedding or polar positional embedding can be turned on'
         assert not (rotary_xpos and not causal), 'rotary xpos is not compatible with bidirectional attention'
         self.rotary_pos_emb = RotaryEmbedding(rotary_emb_dim, use_xpos = rotary_xpos, scale_base = rotary_xpos_scale_base, interpolation_factor = rotary_interpolation_factor, base_rescale_factor = rotary_base_rescale_factor) if rotary_pos_emb else None
 
+        # polar positional embedding (PoPE) - https://arxiv.org/abs/2509.10534
+
+        self.polar_pos_emb = PolarEmbedding(dim_head, polar_bias_uniform_init) if polar_pos_emb else None
+
         assert at_most_one_of(alibi_pos_bias, rel_pos_bias, data_dependent_alibi), 'you can only choose one of Alibi positional bias, data dependent Alibi (forgetting transformers), dynamic tanh, or T5 relative positional bias'
         assert rel_pos_num_buckets <= rel_pos_max_distance, 'number of relative position buckets must be less than the relative position max distance'
 
@@ -2626,6 +2681,7 @@ class AttentionLayers(Module):
         cache_age = 1,
         return_hiddens = False,
         rotary_pos_emb = None,
+        polar_pos_emb = None,
         pos = None,
         context_pos = None,
         attn_bias = None,
@@ -2721,6 +2777,15 @@ class AttentionLayers(Module):
                     context_rotary_pos_emb = context_rotary_pos_emb
                 )
 
+        # polar positions
+
+        if exists(self.polar_pos_emb):
+            if not exists(polar_pos_emb):
+                if not exists(pos):
+                    pos = arange(x.shape[1] + seq_pos_offset, device = x.device)
+
+                polar_pos_emb = self.polar_pos_emb(pos)
+
         # assume cached key / values
 
         prev_cache_length = 0
@@ -2910,7 +2975,7 @@ 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, pos = pos, rotary_pos_emb = rotary_pos_emb, additional_key_values = next(iter_self_attn_kv, None), additional_key_value_mask = additional_kv_mask, prev_attn = prev_attn, cache = next(iter_attn_cache, None), mem = layer_mem, mem_mask = layer_mem_mask, attn_bias = attn_bias, kv_input_residual = next(self_attn_kv_residuals_iter, None), 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, polar_pos_emb = polar_pos_emb, additional_key_values = next(iter_self_attn_kv, None), additional_key_value_mask = additional_kv_mask, prev_attn = prev_attn, cache = next(iter_attn_cache, None), mem = layer_mem, mem_mask = layer_mem_mask, attn_bias = attn_bias, kv_input_residual = next(self_attn_kv_residuals_iter, None), 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), kv_input_residual = next(cross_attn_kv_residuals_iter, None), value_residual = maybe_cross_attn_value_residual, **cross_attn_rotary_pos_emb, return_intermediates = True)
             elif layer_type == 'f':

{x_transformers-2.11.24.dist-info → x_transformers-2.12.1.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: x-transformers
-Version: 2.11.24
+Version: 2.12.1
 Summary: X-Transformers
 Project-URL: Homepage, https://pypi.org/project/x-transformers/
 Project-URL: Repository, https://github.com/lucidrains/x-transformers
@@ -2630,4 +2630,16 @@ ids_out, num_out, is_number_mask = model.generate(start_ids, start_nums, 17)
 }
 ```
 
+```bibtex
+@misc{gopalakrishnan2025decouplingwhatwherepolar,
+    title   = {Decoupling the "What" and "Where" With Polar Coordinate Positional Embeddings}, 
+    author  = {Anand Gopalakrishnan and Robert Csordás and Jürgen Schmidhuber and Michael C. Mozer},
+    year    = {2025},
+    eprint  = {2509.10534},
+    archivePrefix = {arXiv},
+    primaryClass = {cs.LG},
+    url     = {https://arxiv.org/abs/2509.10534}, 
+}
+```
+
 *solve intelligence... then use that to solve everything else.* - Demis Hassabis

{x_transformers-2.11.24.dist-info → x_transformers-2.12.1.dist-info}/RECORD

@@ -11,10 +11,10 @@ x_transformers/multi_input.py,sha256=tCh-fTJDj2ib4SMGtsa-AM8MxKzJAQSwqAXOu3HU2mg
 x_transformers/neo_mlp.py,sha256=XCNnnop9WLarcxap1kGuYc1x8GHvwkZiDRnXOxSl3Po,3452
 x_transformers/nonautoregressive_wrapper.py,sha256=hMQqNimGtchNIe13cR5LZule1V7I1qM5LmY8VQfVdnA,11698
 x_transformers/up_wrapper.py,sha256=YC2LN14_7Xx9Wtiek2rtEJ_qHqdfSmKlh3d7Cgxwd80,7073
-x_transformers/x_transformers.py,sha256=ESupgE2cCteH4_PMWz9tPolAmT0lqBEwITUnT6ZhR8Y,129954
+x_transformers/x_transformers.py,sha256=hW_adsM99N3MGIvsn7zv5w2KRZ3pXgiMsTK333MEIko,132038
 x_transformers/xl_autoregressive_wrapper.py,sha256=CvZMJ6A6PA-Y_bQAhnORwjJBSl6Vjq2IdW5KTdk8NI8,4195
 x_transformers/xval.py,sha256=AwwYUm8yDAtKQyKJDIhYMsiLTJ_skh3scUFMjp5sda8,8597
-x_transformers-2.11.24.dist-info/METADATA,sha256=i318BRxUVm0HCN6Yr63oRFG8ISc04t6-8wT8N6mfFew,96751
-x_transformers-2.11.24.dist-info/WHEEL,sha256=WLgqFyCfm_KASv4WHyYy0P3pM_m7J5L9k2skdKLirC8,87
-x_transformers-2.11.24.dist-info/licenses/LICENSE,sha256=As9u198X-U-vph5noInuUfqsAG2zX_oXPHDmdjwlPPY,1066
-x_transformers-2.11.24.dist-info/RECORD,,
+x_transformers-2.12.1.dist-info/METADATA,sha256=mlqg7VnbIjDvh_Zp7HXHJu1HLzUWB2SB_5qCPlalGhc,97174
+x_transformers-2.12.1.dist-info/WHEEL,sha256=WLgqFyCfm_KASv4WHyYy0P3pM_m7J5L9k2skdKLirC8,87
+x_transformers-2.12.1.dist-info/licenses/LICENSE,sha256=As9u198X-U-vph5noInuUfqsAG2zX_oXPHDmdjwlPPY,1066
+x_transformers-2.12.1.dist-info/RECORD,,