x-transformers 2.5.6__tar.gz → 2.6.1__tar.gz

{x_transformers-2.5.6 → x_transformers-2.6.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: x-transformers
-Version: 2.5.6
+Version: 2.6.1
 Summary: X-Transformers
 Project-URL: Homepage, https://pypi.org/project/x-transformers/
 Project-URL: Repository, https://github.com/lucidrains/x-transformers

{x_transformers-2.5.6 → x_transformers-2.6.1}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "x-transformers"
-version = "2.5.6"
+version = "2.6.1"
 description = "X-Transformers"
 authors = [
     { name = "Phil Wang", email = "lucidrains@gmail.com" }

{x_transformers-2.5.6 → x_transformers-2.6.1}/tests/test_x_transformers.py

@@ -1210,3 +1210,28 @@ def test_prompts_given_as_list_tensor():
     ], 256)
 
     assert sampled.shape == (4, 256)
+
+def test_external_key_values():
+    from x_transformers import AutoregressiveWrapper
+
+    model = TransformerWrapper(
+        num_tokens = 20000,
+        max_seq_len = 1024,
+        attn_layers = Decoder(
+            dim = 512,
+            depth = 2,
+            heads = 8,
+            attn_dim_head = 16
+        )
+    )
+
+    seq = torch.randint(0, 20000, (3, 1024))
+
+    key_values = [
+        (torch.randn(3, 8, 32, 16), torch.randn(3, 8, 32, 16)),
+        (torch.randn(3, 8, 32, 16), torch.randn(3, 8, 32, 16)),
+    ]
+
+    additional_kv_mask = torch.randint(0, 2, (3, 32)).bool()
+
+    logits = model(seq, self_attn_additional_kv = key_values, additional_kv_mask = additional_kv_mask)

{x_transformers-2.5.6 → x_transformers-2.6.1}/x_transformers/x_transformers.py

@@ -1617,7 +1617,9 @@ class Attention(Module):
         mem_mask = None,
         return_intermediates = False,
         cache: Intermediates | None = None,
-        value_residual = None
+        value_residual = None,
+        additional_key_values: tuple[Tensor, Tensor] | None = None,
+        additional_key_value_mask = 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
 
@@ -1787,6 +1789,26 @@ class Attention(Module):
             if exists(input_mask):
                 input_mask = pad_at_dim(input_mask, (self.num_mem_kv, 0), dim = -1, value = True)
 
+        # maybe append additional key / values
+
+        if exists(additional_key_values):
+            seq_len = k.shape[-2]
+
+            added_k, added_v = additional_key_values
+
+            k = cat((added_k, k), dim = -2)
+            v = cat((added_v, v), dim = -2)
+
+            if (exists(input_mask) or exists(additional_key_value_mask)):
+
+                if not exists(additional_key_value_mask):
+                    added_kv_len = added_k.shape[-2]
+                    input_mask = pad_at_dim(input_mask, (added_kv_len, 0), dim = -1, value = True)
+                elif not exists(input_mask):
+                    input_mask = pad_at_dim(additional_key_value_mask, (0, seq_len), dim = -1, value = True)
+                else:
+                    input_mask = cat((additional_key_value_mask, input_mask), dim = -1)
+
         # determine masking
 
         mask_value = max_neg_value(q)
@@ -2411,6 +2433,8 @@ class AttentionLayers(Module):
         context_pos = None,
         attn_bias = None,
         deep_embeds_and_ids: tuple[nn.Parameter, Tensor] | None = None,
+        self_attn_additional_kv: list[tuple[Tensor, Tensor]] | None = None,
+        additional_kv_mask = None,
         condition = None,
         in_attn_cond = None, # https://arxiv.org/abs/2105.04090
         layers_execute_order: tuple[int, ...] | None = None
@@ -2520,6 +2544,10 @@ class AttentionLayers(Module):
 
         iter_attn_cache = iter(attn_cache)
 
+        # additional self attn key / values
+
+        iter_self_attn_kv = iter(default(self_attn_additional_kv, ()))
+
         # handle deep embeds if needed
 
         deep_embeds = []
@@ -2647,7 +2675,7 @@ 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, 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, 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)
             elif layer_type == 'f':