x-transformers 2.11.23__tar.gz → 2.12.0__tar.gz

{x_transformers-2.11.23 → x_transformers-2.12.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: x-transformers
-Version: 2.11.23
+Version: 2.12.0
 Summary: X-Transformers
 Project-URL: Homepage, https://pypi.org/project/x-transformers/
 Project-URL: Repository, https://github.com/lucidrains/x-transformers
@@ -2618,4 +2618,28 @@ ids_out, num_out, is_number_mask = model.generate(start_ids, start_nums, 17)
 }
 ```
 
+```bibtex
+@misc{chen2025strongernormalizationfreetransformers,
+    title   = {Stronger Normalization-Free Transformers}, 
+    author  = {Mingzhi Chen and Taiming Lu and Jiachen Zhu and Mingjie Sun and Zhuang Liu},
+    year    = {2025},
+    eprint  = {2512.10938},
+    archivePrefix = {arXiv},
+    primaryClass = {cs.LG},
+    url     = {https://arxiv.org/abs/2512.10938}, 
+}
+```
+
+```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-2.11.23 → x_transformers-2.12.0}/README.md

@@ -2569,4 +2569,28 @@ ids_out, num_out, is_number_mask = model.generate(start_ids, start_nums, 17)
 }
 ```
 
+```bibtex
+@misc{chen2025strongernormalizationfreetransformers,
+    title   = {Stronger Normalization-Free Transformers}, 
+    author  = {Mingzhi Chen and Taiming Lu and Jiachen Zhu and Mingjie Sun and Zhuang Liu},
+    year    = {2025},
+    eprint  = {2512.10938},
+    archivePrefix = {arXiv},
+    primaryClass = {cs.LG},
+    url     = {https://arxiv.org/abs/2512.10938}, 
+}
+```
+
+```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-2.11.23 → x_transformers-2.12.0}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "x-transformers"
-version = "2.11.23"
+version = "2.12.0"
 description = "X-Transformers"
 authors = [
     { name = "Phil Wang", email = "lucidrains@gmail.com" }

{x_transformers-2.11.23 → x_transformers-2.12.0}/tests/test_x_transformers.py

@@ -1488,3 +1488,41 @@ def test_belief_attn(
     x = torch.randint(0, 256, (1, 10))
 
     logits = model(x)
+
+def test_derf():
+    from x_transformers import TransformerWrapper, Decoder
+
+    model = TransformerWrapper(
+        num_tokens = 256,
+        max_seq_len = 1024,
+        attn_layers = Decoder(
+            dim = 512,
+            depth = 6,
+            heads = 8,
+            attn_kv_heads = 4,
+            rotary_pos_emb = True,
+            use_derf = True
+        )
+    )
+
+    x = torch.randint(0, 256, (1, 10))
+
+    logits = model(x)
+
+def test_pope():
+    from x_transformers import TransformerWrapper, Decoder
+
+    model = TransformerWrapper(
+        num_tokens = 256,
+        max_seq_len = 1024,
+        attn_layers = Decoder(
+            dim = 512,
+            depth = 6,
+            heads = 8,
+            polar_pos_emb = True,
+        )
+    )
+
+    x = torch.randint(0, 256, (1, 10))
+
+    logits = model(x)

{x_transformers-2.11.23 → x_transformers-2.12.0}/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):
@@ -941,6 +984,31 @@ class DynamicTanh(Module):
         gamma = self.gamma + self.gamma_offset
         return (x * pre_tanh_scale).tanh() * gamma + self.beta
 
+class Derf(Module):
+    """ https://arxiv.org/abs/2512.10938 """
+    def __init__(
+        self,
+        dim,
+        init_alpha = 0.5,
+        init_bias = 0.,
+        unit_offset = False
+    ):
+        super().__init__()
+        scale_offset = 1. if unit_offset else 0.
+
+        self.alpha = nn.Parameter(tensor(init_alpha) - scale_offset)
+        self.s = nn.Parameter(tensor(init_bias))
+
+        self.gamma = nn.Parameter(torch.ones(dim) - scale_offset)
+        self.beta = nn.Parameter(torch.zeros(dim))
+
+        self.scale_offset = scale_offset
+
+    def forward(self, x):
+        x = x * (self.alpha + self.scale_offset) + self.s
+        activated = torch.erf(x)
+        return activated * (self.gamma + self.scale_offset) + self.beta
+
 # residual and residual gates
 
 class Residual(Module):
@@ -1720,6 +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,
@@ -1871,6 +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:
@@ -2123,6 +2197,7 @@ class AttentionLayers(Module):
         use_scalenorm = False,
         use_rmsnorm = False,
         use_dynamic_tanh = False,
+        use_derf = False,
         dynamic_tanh_init_alpha = 1.,
         use_simple_rmsnorm = False,
         use_adaptive_layernorm = False,
@@ -2148,6 +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,
@@ -2240,9 +2317,14 @@ 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'
 
@@ -2277,7 +2359,7 @@ class AttentionLayers(Module):
 
         # determine norm
 
-        assert at_most_one_of(use_scalenorm, use_rmsnorm, use_dynamic_tanh, use_simple_rmsnorm, use_adaptive_layernorm, use_adaptive_rmsnorm), 'you can only use either scalenorm, rmsnorm, adaptive layernorm, adaptive rmsnorm, or simple rmsnorm'
+        assert at_most_one_of(use_scalenorm, use_rmsnorm, use_dynamic_tanh, use_derf, use_simple_rmsnorm, use_adaptive_layernorm, use_adaptive_rmsnorm), 'you can only use either scalenorm, rmsnorm, adaptive layernorm, adaptive rmsnorm, or simple rmsnorm'
 
         norm_need_condition = False
         dim_condition = default(dim_condition, dim)
@@ -2295,6 +2377,8 @@ class AttentionLayers(Module):
         elif use_dynamic_tanh:
             assert pre_norm, 'dynamic tanh norm only tested for pre-norm'
             norm_class = partial(DynamicTanh, init_alpha = dynamic_tanh_init_alpha)
+        elif use_derf:
+            norm_class = Derf
         elif use_adaptive_layernorm:
             norm_need_condition = True
             norm_class = partial(AdaptiveLayerNorm, dim_condition = dim_condition * dim_condition_mult)
@@ -2598,6 +2682,7 @@ 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,
@@ -2693,6 +2778,15 @@ 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
@@ -2882,7 +2976,7 @@ 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':