x-transformers 2.6.7__tar.gz → 2.7.1__tar.gz

{x_transformers-2.6.7 → x_transformers-2.7.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: x-transformers
-Version: 2.6.7
+Version: 2.7.1
 Summary: X-Transformers
 Project-URL: Homepage, https://pypi.org/project/x-transformers/
 Project-URL: Repository, https://github.com/lucidrains/x-transformers
@@ -2509,11 +2509,22 @@ ids_out, num_out, is_number_mask = model.generate(start_ids, start_nums, 17)
 
 ```bibtex
 @misc{openai_gpt_oss,
-  author       = {OpenAI},
-  title        = {Introducing gpt-oss},
-  howpublished = {https://openai.com/index/introducing-gpt-oss},
-  month        = {August},
-  year         = {2025}
+    author       = {OpenAI},
+    title        = {Introducing gpt-oss},
+    howpublished = {https://openai.com/index/introducing-gpt-oss},
+    month        = {August},
+    year         = {2025}
+}
+```
+
+```bibtex
+@article{Sahoo2024SimpleAE,
+    title   = {Simple and Effective Masked Diffusion Language Models},
+    author  = {Subham Sekhar Sahoo and Marianne Arriola and Yair Schiff and Aaron Gokaslan and Edgar Marroquin and Justin T Chiu and Alexander Rush and Volodymyr Kuleshov},
+    journal = {ArXiv},
+    year    = {2024},
+    volume  = {abs/2406.07524},
+    url     = {https://api.semanticscholar.org/CorpusID:270380319}
 }
 ```
 

{x_transformers-2.6.7 → x_transformers-2.7.1}/README.md

@@ -2461,11 +2461,22 @@ ids_out, num_out, is_number_mask = model.generate(start_ids, start_nums, 17)
 
 ```bibtex
 @misc{openai_gpt_oss,
-  author       = {OpenAI},
-  title        = {Introducing gpt-oss},
-  howpublished = {https://openai.com/index/introducing-gpt-oss},
-  month        = {August},
-  year         = {2025}
+    author       = {OpenAI},
+    title        = {Introducing gpt-oss},
+    howpublished = {https://openai.com/index/introducing-gpt-oss},
+    month        = {August},
+    year         = {2025}
+}
+```
+
+```bibtex
+@article{Sahoo2024SimpleAE,
+    title   = {Simple and Effective Masked Diffusion Language Models},
+    author  = {Subham Sekhar Sahoo and Marianne Arriola and Yair Schiff and Aaron Gokaslan and Edgar Marroquin and Justin T Chiu and Alexander Rush and Volodymyr Kuleshov},
+    journal = {ArXiv},
+    year    = {2024},
+    volume  = {abs/2406.07524},
+    url     = {https://api.semanticscholar.org/CorpusID:270380319}
 }
 ```
 

{x_transformers-2.6.7 → x_transformers-2.7.1}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "x-transformers"
-version = "2.6.7"
+version = "2.7.1"
 description = "X-Transformers"
 authors = [
     { name = "Phil Wang", email = "lucidrains@gmail.com" }

{x_transformers-2.6.7 → x_transformers-2.7.1}/tests/test_x_transformers.py

@@ -1287,3 +1287,30 @@ def test_accept_layer_intermediates():
     )
 
     assert embeds.shape == (3, 32, 512)
+
+@pytest.mark.parametrize('use_loss_weight', (False, True))
+def test_simple_mdlm(
+    use_loss_weight
+):
+    from x_transformers.nonautoregressive_wrapper import NonAutoregressiveWrapper
+
+    model = TransformerWrapper(
+        num_tokens = 256 + 1,
+        max_seq_len = 1024,
+        attn_layers = Encoder(
+            dim = 512,
+            depth = 4,
+            rotary_pos_emb = True
+        )
+    )
+
+    nar = NonAutoregressiveWrapper(
+        model,
+        mask_id = 256,
+        use_simple_mdlm_loss_weight = use_loss_weight
+    )
+
+    seq = torch.randint(0, 256, (1, 1024))
+
+    loss = nar(seq)
+    loss.loss.backward()

{x_transformers-2.6.7 → x_transformers-2.7.1}/x_transformers/continuous.py

@@ -241,6 +241,7 @@ class ContinuousAutoregressiveWrapper(Module):
         self,
         net: ContinuousTransformerWrapper,
         loss_fn: Module | None = None,
+        use_l1_loss = False,
         equal_loss_weight_batch = False,  # setting this to True, if the mask is passed in and sequences are variable in length, each sequence will be weighted the same (as opposed to each token)
     ):
         super().__init__()
@@ -250,7 +251,15 @@ class ContinuousAutoregressiveWrapper(Module):
         probabilistic = net.probabilistic
         self.probabilistic = probabilistic
 
-        loss_fn = default(loss_fn, nn.MSELoss(reduction = 'none') if not probabilistic else GaussianNLL())
+        # default loss function
+
+        if not exists(loss_fn):
+            if probabilistic:
+                loss_fn = GaussianNLL()
+            elif use_l1_loss:
+                loss_fn = nn.L1Loss(reduction = 'none')
+            else:
+                loss_fn = nn.MSELoss(reduction = 'none')
 
         self.loss_fn = loss_fn
         self.equal_loss_weight_batch = equal_loss_weight_batch

{x_transformers-2.6.7 → x_transformers-2.7.1}/x_transformers/nonautoregressive_wrapper.py

@@ -1,16 +1,20 @@
+from __future__ import annotations
+
 import math
 from random import random
 from contextlib import nullcontext
 from collections import namedtuple
 
 import torch
+from torch import nn, pi
+from torch.nn import Module
+from torch.func import grad_and_value, vmap
 import torch.nn.functional as F
-from torch import nn
 
+import einx
 from einops import rearrange, repeat, pack, unpack
 
 from x_transformers.x_transformers import TransformerWrapper
-from typing import Optional
 
 # constants
 
@@ -75,12 +79,12 @@ def linear_schedule(t):
 
 def cosine_schedule(t):
     """ https://arxiv.org/abs/2202.04200 """
-    return torch.cos(t * math.pi / 2)
+    return torch.cos(t * pi / 2)
 
 # self token critic
 # inspired by Nijkamp et al. - https://aclanthology.org/2021.naacl-main.409/
 
-class SelfCritic(nn.Module):
+class SelfCritic(Module):
     def __init__(self, net):
         super().__init__()
         self.net = net
@@ -92,7 +96,7 @@ class SelfCritic(nn.Module):
         embed = self.net(x, return_embeddings = True)
         return self.to_logits(embed)
 
-class NonAutoregressiveWrapper(nn.Module):
+class NonAutoregressiveWrapper(Module):
     """
     https://arxiv.org/abs/1904.09324
     https://arxiv.org/abs/2202.04200
@@ -110,9 +114,10 @@ class NonAutoregressiveWrapper(nn.Module):
         random_token_prob = 0.1,         # which percentage of tokens to be replaced with random token, done in original MLM paper
         schedule = 'linear',
         can_mask_prev_unmasked = False,  # when unmasking, whether it can remask previously unmasked
-        token_critic: Optional[TransformerWrapper] = None,
+        token_critic: TransformerWrapper | None = None,
         self_token_critic = False,
-        critic_loss_weight = 1.
+        critic_loss_weight = 1.,
+        use_simple_mdlm_loss_weight = True # Sahoo et al. https://arxiv.org/abs/2406.07524
     ):
         super().__init__()
         assert not (self_token_critic and exists(token_critic))
@@ -143,6 +148,23 @@ class NonAutoregressiveWrapper(nn.Module):
         else:
             raise ValueError(f'invalid schedule {schedule}')
 
+        # whether to use the loss weighting proposed in simple diffusion lm paper
+
+        self.loss_weight_fn = None
+
+        if use_simple_mdlm_loss_weight:
+            grad_and_value_schedule_fn = vmap(grad_and_value(self.schedule_fn))
+
+            # eq (10)
+
+            def loss_weight_fn(times):
+                grad, value = grad_and_value_schedule_fn(times)
+                return grad / (1. - value)
+
+            self.loss_weight_fn = loss_weight_fn
+
+        # whether to mask previous - in the simple mdlm paper, they chose not to
+
         self.can_mask_prev_unmasked = can_mask_prev_unmasked
 
         # self conditioning
@@ -311,12 +333,27 @@ class NonAutoregressiveWrapper(nn.Module):
 
         loss_fn = F.cross_entropy if not self.net.output_is_log_prob else F.nll_loss
 
-        # cross entropy loss
+        # loss
 
-        loss = loss_fn(
-            logits[mask],
-            orig_seq[mask]
-        )
+        if exists(self.loss_weight_fn):
+            # using simple mdlm loss weighting
+
+            loss = loss_fn(
+                rearrange(logits, 'b n l -> b l n'),
+                orig_seq,
+                reduction = 'none'
+            )
+
+            loss_weights = self.loss_weight_fn(rand_times)     # calculate loss weight
+            loss = einx.multiply('b n, b', loss, loss_weights) # apply loss weights
+
+            loss = loss[mask].mean()
+
+        else:
+            loss = loss_fn(
+                logits[mask],
+                orig_seq[mask],
+            )
 
         if not exists(self.token_critic) or only_train_generator:
             return Losses(loss, loss, None)