x-transformers 2.5.2__py3-none-any.whl → 2.5.4__py3-none-any.whl

x_transformers/autoregressive_wrapper.py

@@ -4,9 +4,10 @@ from math import ceil, log
 from typing import Tuple, Callable
 
 import torch
-from torch import nn, Tensor
+from torch import nn, tensor, Tensor
 from torch.nn import Module
 import torch.nn.functional as F
+from torch.nn.utils.rnn import pad_sequence
 
 from einops import rearrange, repeat, pack, unpack
 
@@ -347,7 +348,7 @@ class AutoregressiveWrapper(Module):
     @eval_decorator
     def generate(
         self,
-        prompts,
+        prompts: list[Tensor] | Tensor,
         seq_len,
         eos_token = None,
         temperature = 1.,
@@ -363,11 +364,23 @@ class AutoregressiveWrapper(Module):
         cache_kv = True,
         **kwargs
     ):
-        max_seq_len, greedy, device = self.max_seq_len, temperature == 0., prompts.device
+        max_seq_len, greedy = self.max_seq_len, temperature == 0.
+
+        # handle prompts given as list of variable lengthed token ids
+
+        if isinstance(prompts, list):
+            assert len(prompts) > 0, 'prompts cannot be empty list'
+            assert not exists(prompt_lens), '`prompt_len` will be auto derived if prompts are passed in as list of Tensors'
+
+            prompt_lens = tensor([t.shape[0] for t in prompts], device = prompts[0].device)
+
+            prompts = pad_sequence(prompts, batch_first = True)
+
+        # pack maybe no batch
 
         prompts, ps = pack([prompts], '* n')
 
-        b, t = prompts.shape
+        b, t, device = *prompts.shape, prompts.device
 
         # handle filter logits fn given as string
 
@@ -380,6 +393,7 @@ class AutoregressiveWrapper(Module):
 
         seq_start_pos = None
         if exists(prompt_lens):
+            print('prompt lens')
             prompts = align_right(prompts, prompt_lens, pad_id = self.pad_value)
             seq_start_pos = t - prompt_lens
 

x_transformers/x_transformers.py

@@ -2787,6 +2787,7 @@ class AttentionPool(Module):
             self.pooler = Attention(dim = dim, dim_context = dim_context, heads = heads, dim_head = dim_head, **attn_kwargs)
 
         self.add_residual = add_residual
+        self.squeeze_output = squeeze_output
 
     def forward(self, context, mask = None):
         batch = context.shape[0]
@@ -2798,6 +2799,9 @@ class AttentionPool(Module):
         if self.add_residual:
             pooled = pooled + queries
 
+        if self.squeeze_output:
+            pooled = rearrange(pooled, 'b 1 d -> b d')
+
         return pooled
 
 class ViTransformerWrapper(Module):

{x_transformers-2.5.2.dist-info → x_transformers-2.5.4.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: x-transformers
-Version: 2.5.2
+Version: 2.5.4
 Summary: X-Transformers
 Project-URL: Homepage, https://pypi.org/project/x-transformers/
 Project-URL: Repository, https://github.com/lucidrains/x-transformers

{x_transformers-2.5.2.dist-info → x_transformers-2.5.4.dist-info}/RECORD

@@ -1,6 +1,6 @@
 x_transformers/__init__.py,sha256=aVuhUU0572TJHW88BVc4yA2tla0Zb8l3NH7W4RZ1AEs,1005
 x_transformers/attend.py,sha256=Ax34Rw56xXAWfFPqtZ_B8iKEW2EfQdbVoc9uFjfeNjA,17404
-x_transformers/autoregressive_wrapper.py,sha256=y798kS9_VvPOY_5Ilits_64aXNqYvGuilsky1y07ryE,17834
+x_transformers/autoregressive_wrapper.py,sha256=wu7yJOf2XL5QI0vJtRTSsyOI-JpCuwE7YXGEumhtQYQ,18434
 x_transformers/belief_state_wrapper.py,sha256=YLUMk6t2MhFBEw5lHDDHJHcoCxTIkHvxTNY__GGZEKU,13374
 x_transformers/continuous.py,sha256=hpb1sSbt3k2LNzzjrjSd8F5xOIbKj7IluV9MBEAFLkw,13031
 x_transformers/dpo.py,sha256=xt4OuOWhU8pN3OKN2LZAaC2NC8iiEnchqqcrPWVqf0o,3521
@@ -9,10 +9,10 @@ x_transformers/multi_input.py,sha256=tCh-fTJDj2ib4SMGtsa-AM8MxKzJAQSwqAXOu3HU2mg
 x_transformers/neo_mlp.py,sha256=XCNnnop9WLarcxap1kGuYc1x8GHvwkZiDRnXOxSl3Po,3452
 x_transformers/nonautoregressive_wrapper.py,sha256=2NU58hYMgn-4Jzg3mie-mXb0XH_dCN7fjlzd3K1rLUY,10510
 x_transformers/up_wrapper.py,sha256=YC2LN14_7Xx9Wtiek2rtEJ_qHqdfSmKlh3d7Cgxwd80,7073
-x_transformers/x_transformers.py,sha256=vmMrHP3hAQ9iAJlRN1pKmXOn7pD3mfh_ndtaR7LMPzU,119860
+x_transformers/x_transformers.py,sha256=fW-AoomNCw4n2JFbZN9rZV3lKQvz_Tl6L4txUvac_9o,119993
 x_transformers/xl_autoregressive_wrapper.py,sha256=CvZMJ6A6PA-Y_bQAhnORwjJBSl6Vjq2IdW5KTdk8NI8,4195
 x_transformers/xval.py,sha256=AwwYUm8yDAtKQyKJDIhYMsiLTJ_skh3scUFMjp5sda8,8597
-x_transformers-2.5.2.dist-info/METADATA,sha256=yeferX_PJIv0Lxs36vZSV7Z2w9ol4udiUAON95hP_bY,90223
-x_transformers-2.5.2.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
-x_transformers-2.5.2.dist-info/licenses/LICENSE,sha256=As9u198X-U-vph5noInuUfqsAG2zX_oXPHDmdjwlPPY,1066
-x_transformers-2.5.2.dist-info/RECORD,,
+x_transformers-2.5.4.dist-info/METADATA,sha256=sDkhsAQnYbflFctIdoHdtrXV6doTSO_jNLcnFus0CEY,90223
+x_transformers-2.5.4.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
+x_transformers-2.5.4.dist-info/licenses/LICENSE,sha256=As9u198X-U-vph5noInuUfqsAG2zX_oXPHDmdjwlPPY,1066
+x_transformers-2.5.4.dist-info/RECORD,,