x-transformers 1.26.4__py3-none-any.whl → 1.26.5__py3-none-any.whl

x_transformers/autoregressive_wrapper.py

@@ -188,6 +188,10 @@ class AutoregressiveWrapper(Module):
         for _ in range(seq_len):
 
             if restrict_to_max_seq_len:
+                max_len_exceeded = out.shape[-1] > max_seq_len
+
+                assert not (cache_kv and max_len_exceeded and self.net.can_cache_kv_outside_max_seq_len), 'the network cannot use cached key values when decoding outside the max sequence length. most likely because you are using absolute positional embeeding. you can switch to rotary embeddings to resolve this issue'
+
                 x = out[:, -max_seq_len:]
 
                 if exists(cache):

x_transformers/x_transformers.py

@@ -1495,7 +1495,9 @@ class TransformerWrapper(nn.Module):
         self.l2norm_embed = l2norm_embed
         self.token_emb = TokenEmbedding(emb_dim, num_tokens, l2norm_embed = l2norm_embed)
 
-        if max_seq_len == 0 or not (use_abs_pos_emb and not attn_layers.has_pos_emb):
+        no_abs_pos_emb = max_seq_len == 0 or not (use_abs_pos_emb and not attn_layers.has_pos_emb)
+
+        if no_abs_pos_emb:
             self.pos_emb = always(0)
         elif scaled_sinu_pos_emb:
             self.pos_emb = ScaledSinusoidalEmbedding(emb_dim)
@@ -1536,6 +1538,7 @@ class TransformerWrapper(nn.Module):
         # whether can do cached kv decoding
 
         self.can_cache_kv = self.num_memory_tokens == 0
+        self.can_cache_kv_outside_max_seq_len = not no_abs_pos_emb
 
     def init_(self):
         if self.l2norm_embed:

{x_transformers-1.26.4.dist-info → x_transformers-1.26.5.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: x-transformers
-Version: 1.26.4
+Version: 1.26.5
 Summary: X-Transformers - Pytorch
 Home-page: https://github.com/lucidrains/x-transformers
 Author: Phil Wang

{x_transformers-1.26.4.dist-info → x_transformers-1.26.5.dist-info}/RECORD

@@ -1,13 +1,13 @@
 x_transformers/__init__.py,sha256=pXc_U4M3ONUQcpNgZySDIlCF1rp7u4FFmcOYjc4WuXw,629
 x_transformers/attend.py,sha256=MFl_FbgPsm9mziZPTi_s8QbxASETwbGeciMH8sUIwT8,10188
-x_transformers/autoregressive_wrapper.py,sha256=xqr-CnB98N89y7K0z2SLEy8If8pE_3MP7C9WFTwjefs,9230
+x_transformers/autoregressive_wrapper.py,sha256=47sc7HAMNBJUGZRtZX-cO-yML0YFcw4PF6E-7pp1E0A,9614
 x_transformers/continuous.py,sha256=ixfgi2_zpGN03SX_STXFkNYEOAkgwVIxuS53QgDCx-g,6026
 x_transformers/nonautoregressive_wrapper.py,sha256=AQLE4rA_Kh8VNoe9OzpwyeWson34sRkhks4dn4seNjI,10414
-x_transformers/x_transformers.py,sha256=gXq_IpWswCeM06r0VCxvZGTVndsxD69_uvKSJFC1544,61632
+x_transformers/x_transformers.py,sha256=8n8R_huY0KwKDGTUlLLhleAqNR5M1YI_95KRmhrP_Eg,61740
 x_transformers/xl_autoregressive_wrapper.py,sha256=DCx4n0_c1tFai4nOqaWVnqx2p9eutsZsDMiMP1ckxNU,4117
 x_transformers/xval.py,sha256=lS9W_E_RskPQAqVZkPiUzbByoW1Ajsw_phsikA3JDAg,8139
-x_transformers-1.26.4.dist-info/LICENSE,sha256=As9u198X-U-vph5noInuUfqsAG2zX_oXPHDmdjwlPPY,1066
-x_transformers-1.26.4.dist-info/METADATA,sha256=fX_dEQcxjW00biyG6pOz9J88HGiS9qVS05KP1Z94H44,661
-x_transformers-1.26.4.dist-info/WHEEL,sha256=oiQVh_5PnQM0E3gPdiz09WCNmwiHDMaGer_elqB3coM,92
-x_transformers-1.26.4.dist-info/top_level.txt,sha256=hO6KGpFuGucRNEtRfme4A_rGcM53AKwGP7RVlRIxS5Q,15
-x_transformers-1.26.4.dist-info/RECORD,,
+x_transformers-1.26.5.dist-info/LICENSE,sha256=As9u198X-U-vph5noInuUfqsAG2zX_oXPHDmdjwlPPY,1066
+x_transformers-1.26.5.dist-info/METADATA,sha256=GcEy7CtmuqOpAapRxh7Et5kfPOBiV2EIa6GjN2U-eFM,661
+x_transformers-1.26.5.dist-info/WHEEL,sha256=oiQVh_5PnQM0E3gPdiz09WCNmwiHDMaGer_elqB3coM,92
+x_transformers-1.26.5.dist-info/top_level.txt,sha256=hO6KGpFuGucRNEtRfme4A_rGcM53AKwGP7RVlRIxS5Q,15
+x_transformers-1.26.5.dist-info/RECORD,,