x-transformers 2.11.5__tar.gz → 2.11.8__tar.gz

{x_transformers-2.11.5 → x_transformers-2.11.8}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: x-transformers
-Version: 2.11.5
+Version: 2.11.8
 Summary: X-Transformers
 Project-URL: Homepage, https://pypi.org/project/x-transformers/
 Project-URL: Repository, https://github.com/lucidrains/x-transformers

{x_transformers-2.11.5 → x_transformers-2.11.8}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "x-transformers"
-version = "2.11.5"
+version = "2.11.8"
 description = "X-Transformers"
 authors = [
     { name = "Phil Wang", email = "lucidrains@gmail.com" }

{x_transformers-2.11.5 → x_transformers-2.11.8}/tests/test_x_transformers.py

@@ -1434,3 +1434,20 @@ def test_free(
     loss.backward()
 
     assert aux_loss.numel() == 1
+
+def test_kv_input_residual():
+    attn = Decoder(
+        dim = 256,
+        depth = 2,
+        heads = 4,
+        cross_attend = True
+    )
+
+    tokens = torch.randn(3, 32, 256)
+    context = torch.randn(3, 64, 256)
+
+    condition = torch.randn(2, 3, 64, 256)
+
+    out = attn(tokens, context = context, cross_attn_kv_residuals = condition)
+
+    assert tokens.shape == out.shape

{x_transformers-2.11.5 → x_transformers-2.11.8}/x_transformers/free_transformer.py

@@ -66,7 +66,6 @@ class BinaryMapper(Module):
 
         self.bits = bits
         self.num_codes = 2 ** bits
-        self.kl_loss_threshold = kl_loss_threshold
 
         power_two = 2 ** arange(bits)
         codes = (arange(self.num_codes)[:, None].bitwise_and(power_two) != 0).byte().bool()
@@ -74,13 +73,20 @@ class BinaryMapper(Module):
         self.register_buffer('power_two', power_two, persistent = False)
         self.register_buffer('codes', codes, persistent = False)
 
+        # aux loss
+
+        self.kl_loss_threshold = kl_loss_threshold
+        self.register_buffer('zero', tensor(0.), persistent = False)
+
     def forward(
         self,
         logits,
         temperature = 1.,
-        straight_through = None
+        straight_through = None,
+        calc_aux_loss = None
     ):
         straight_through = default(straight_through, self.training)
+        calc_aux_loss = default(calc_aux_loss, self.training)
 
         assert logits.shape[-1] == self.bits, f'logits must have a last dimension of {self.bits}'
 
@@ -95,26 +101,29 @@ class BinaryMapper(Module):
 
         one_hot = F.one_hot(indices, self.num_codes).float()
 
-        # return hard one hot if not training or overridden
+        # maybe calculate aux loss
 
-        if not straight_through:
-            return one_hot
+        aux_kl_loss = self.zero
 
-        # calculate negative entropy
+        if calc_aux_loss:
+            # calculate negative entropy
 
-        kl_div = self.bits * NAT - binary_entropy(logits)
-        aux_kl_loss = F.relu(kl_div - self.kl_loss_threshold).mean()
+            kl_div = self.bits * NAT - binary_entropy(logits)
+            aux_kl_loss = F.relu(kl_div - self.kl_loss_threshold).mean()
 
-        # get the soft G for the gradients and do a straight through
+        # maybe straight through
 
-        soft_G = (
-            einsum(F.logsigmoid(logits), self.codes.float(), '... bits, codes bits -> ... codes') +
-            einsum(F.logsigmoid(-logits), (~self.codes).float(), '... bits, codes bits -> ... codes')
-        ).exp()
+        if straight_through:
+            # get the soft G for the gradients and do a straight through
 
-        # straight through
+            soft_G = (
+                einsum(F.logsigmoid(logits), self.codes.float(), '... bits, codes bits -> ... codes') +
+                einsum(F.logsigmoid(-logits), (~self.codes).float(), '... bits, codes bits -> ... codes')
+            ).exp()
 
-        one_hot = one_hot + soft_G - soft_G.detach()
+            # straight through
+
+            one_hot = one_hot + soft_G - soft_G.detach()
 
         return one_hot, aux_kl_loss
 
@@ -163,6 +172,7 @@ class FreeTransformer(Module):
             cross_attend = True,
             use_rmsnorm = True,
             rotary_pos_emb = True,
+            pre_norm_has_final_norm = True,
             **kwargs,
             **enc_kwargs
         )
@@ -242,7 +252,7 @@ class FreeTransformer(Module):
 
         bit_logits = self.to_latent_bit_logits(pooled)
 
-        one_hot_latents, kl_loss = self.binary_mapper(bit_logits, straight_through = True)
+        one_hot_latents, kl_loss = self.binary_mapper(bit_logits, calc_aux_loss = return_kl_loss)
 
         if not return_kl_loss:
             return one_hot_latents
@@ -286,10 +296,7 @@ class FreeTransformer(Module):
 
             head_embed = self.decoder_head(tokens)
 
-            if exists(condition):
-                head_embed = head_embed + condition
-
-            tail_embed = self.decoder_tail(head_embed)
+            tail_embed = self.decoder_tail(head_embed, self_attn_kv_residuals = condition)
 
             tail_embed = tail_embed[:, -1]
 
@@ -329,7 +336,7 @@ class FreeTransformer(Module):
 
         # decoder tail
 
-        tokens = self.decoder_tail(tokens + condition)
+        tokens = self.decoder_tail(tokens, self_attn_kv_residuals = condition)
 
         # cross entropy loss
 

{x_transformers-2.11.5 → x_transformers-2.11.8}/x_transformers/x_transformers.py

@@ -1686,6 +1686,7 @@ class Attention(Module):
         value_residual = None,
         additional_key_values: tuple[Tensor, Tensor] | None = None,
         additional_key_value_mask = None,
+        kv_input_residual = None,
     ):
         b, n, h, kv_h, head_scale, num_mem_kv, device, has_context, qkv_receive_diff_residuals, is_multi_latent_attn = x.shape[0], x.shape[1], self.heads, self.kv_heads, self.head_scale, self.num_mem_kv, x.device, exists(context), self.qkv_receive_diff_residuals, self.use_latent_kv
 
@@ -1702,6 +1703,12 @@ class Attention(Module):
             kv_input = default(context, x)
             q_input, k_input, v_input = x, kv_input, kv_input
 
+        # done for free transformer
+
+        if exists(kv_input_residual):
+            k_input = k_input + kv_input_residual
+            v_input = v_input + kv_input_residual
+
         if exists(mem):
             k_input, mem_packed_shape = pack([mem, k_input], 'b * d')
             v_input, _ = pack([mem, v_input], 'b * d')
@@ -2543,7 +2550,9 @@ class AttentionLayers(Module):
         route_additional_kv_to_top = True,
         condition = None,
         in_attn_cond = None, # https://arxiv.org/abs/2105.04090
-        layers_execute_order: tuple[int, ...] | None = None
+        layers_execute_order: tuple[int, ...] | None = None,
+        self_attn_kv_residuals: Tensor | None = None,
+        cross_attn_kv_residuals: Tensor | None = None
     ):
         assert not (self.cross_attend ^ exists(context)), 'context must be passed in if cross_attend is set to True'
         assert not (exists(condition) ^ self.need_condition), 'condition needs to be passed in if using adaptive layernorm or vice versa'
@@ -2721,6 +2730,23 @@ class AttentionLayers(Module):
 
         skip_hiddens = []
 
+        # for residuals to key value inputs for self and cross attention
+
+        self_attn_kv_residuals_iter = iter((None,))
+        cross_attn_kv_residuals_iter = iter((None,))
+
+        if exists(self_attn_kv_residuals):
+            if self_attn_kv_residuals.ndim == 3:
+                self_attn_kv_residuals = rearrange(self_attn_kv_residuals, '... ->  1 ...')
+
+            self_attn_kv_residuals_iter = iter(self_attn_kv_residuals)
+
+        if exists(cross_attn_kv_residuals):
+            if cross_attn_kv_residuals.ndim == 3:
+                cross_attn_kv_residuals = rearrange(cross_attn_kv_residuals, '... ->  1 ...')
+
+            cross_attn_kv_residuals_iter = iter(cross_attn_kv_residuals)
+
         # for value residuals
 
         first_self_attn_inter = None
@@ -2794,9 +2820,9 @@ 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, 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, 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), value_residual = maybe_cross_attn_value_residual, **cross_attn_rotary_pos_emb, return_intermediates = True)
+                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':
                 out = block(x, deep_embed = next(deep_embeds_iter, None))