x-transformers 1.32.14__tar.gz → 1.34.0__tar.gz

{x_transformers-1.32.14/x_transformers.egg-info → x_transformers-1.34.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: x-transformers
-Version: 1.32.14
+Version: 1.34.0
 Summary: X-Transformers - Pytorch
 Home-page: https://github.com/lucidrains/x-transformers
 Author: Phil Wang

{x_transformers-1.32.14 → x_transformers-1.34.0}/README.md

@@ -2187,4 +2187,15 @@ ids_out, num_out, is_number_mask = model.generate(start_ids, start_nums, 17)
 }
 ```
 
+```bibtex
+@article{Nguyen2024MinPS,
+    title   = {Min P Sampling: Balancing Creativity and Coherence at High Temperature},
+    author  = {Minh Nguyen and Andrew Baker and Andreas Kirsch and Clement Neo},
+    journal = {ArXiv},
+    year    = {2024},
+    volume  = {abs/2407.01082},
+    url     = {https://api.semanticscholar.org/CorpusID:270870613}
+}
+```
+
 *solve intelligence... then use that to solve everything else.* - Demis Hassabis

{x_transformers-1.32.14 → x_transformers-1.34.0}/setup.py

@@ -3,7 +3,7 @@ from setuptools import setup, find_packages
 setup(
   name = 'x-transformers',
   packages = find_packages(exclude=['examples']),
-  version = '1.32.14',
+  version = '1.34.0',
   license='MIT',
   description = 'X-Transformers - Pytorch',
   author = 'Phil Wang',

{x_transformers-1.32.14 → x_transformers-1.34.0}/x_transformers/autoregressive_wrapper.py

@@ -19,6 +19,9 @@ def default(val, d):
 def identity(t, *args, **kwargs):
     return t
 
+def join(arr, delimiter = ', '):
+    return delimiter.join(arr)
+
 def cast_tuple(t, length = 1):
     return t if isinstance(t, tuple) else (t,) * length
 
@@ -79,11 +82,29 @@ def top_k(logits, frac_num_tokens = 0.1, k = None):
 # top_a
 
 def top_a(logits, min_p_pow = 2.0, min_p_ratio = 0.02):
-    probs = F.softmax(logits, dim = -1)
-    max_probs = torch.amax(probs, dim = -1, keepdim = True)
+    probs = logits.softmax(dim = -1)
+    max_probs = probs.amax(dim = -1, keepdim = True)
     limit = torch.pow(max_probs, min_p_pow) * min_p_ratio
     return torch.where(probs < limit, float('-inf'), logits)
 
+# min_p
+# https://arxiv.org/abs/2407.01082
+
+def min_p(logits, min_p = 0.1):
+    probs = logits.softmax(dim = -1)
+    max_probs = probs.amax(dim = -1, keepdim = True)
+    limit = min_p * max_probs
+    return torch.where(probs < limit, float('-inf'), logits)
+
+# filter logits functions dict[str -> Callable]
+
+FILTER_LOGITS_FN = dict(
+    top_p = top_p,
+    top_k = top_k,
+    top_a = top_a,
+    min_p = min_p
+)
+
 # contrastive decoding function
 
 def contrastive_decode_fn(
@@ -136,7 +157,7 @@ class AutoregressiveWrapper(Module):
         eos_token = None,
         temperature = 1.,
         prompt_lens: Tensor | None = None,
-        filter_logits_fn: Callable = top_k,
+        filter_logits_fn: str | Callable = top_k,
         restrict_to_max_seq_len = True,
         amateur_model: Module | Tuple[Module] | None = None,
         filter_kwargs: dict = dict(),
@@ -153,6 +174,13 @@ class AutoregressiveWrapper(Module):
 
         b, t = prompts.shape
 
+        # handle filter logits fn given as string
+
+        if isinstance(filter_logits_fn, str):
+            assert filter_logits_fn in FILTER_LOGITS_FN, f"only {join(FILTER_LOGITS_FN.keys())} are available"
+
+            filter_logits_fn = FILTER_LOGITS_FN[filter_logits_fn]
+
         # handle variable lengthed prompts (prefixes)
 
         seq_start_pos = None

{x_transformers-1.32.14 → x_transformers-1.34.0}/x_transformers/x_transformers.py

@@ -8,7 +8,7 @@ import torch
 import torch.nn.functional as F
 from torch import nn, einsum, Tensor
 from torch.nn import Module, ModuleList, ModuleDict
-from torch.cuda.amp import autocast
+from torch.amp import autocast
 
 from functools import partial, wraps
 from collections import namedtuple
@@ -521,7 +521,7 @@ class RotaryEmbedding(Module):
         t = torch.arange(seq_len, device = device)
         return self.forward(t)
 
-    @autocast(enabled = False)
+    @autocast('cuda', enabled = False)
     def forward(self, t):
         max_pos = t.max() + 1
 
@@ -545,7 +545,7 @@ def rotate_half(x):
     x = torch.stack((-x2, x1), dim = -1)
     return rearrange(x, '... d r -> ... (d r)')
 
-@autocast(enabled = False)
+@autocast('cuda', enabled = False)
 def apply_rotary_pos_emb(t, freqs, scale = 1):
     rot_dim, seq_len, orig_dtype = freqs.shape[-1], t.shape[-2], t.dtype
 

{x_transformers-1.32.14 → x_transformers-1.34.0/x_transformers.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: x-transformers
-Version: 1.32.14
+Version: 1.34.0
 Summary: X-Transformers - Pytorch
 Home-page: https://github.com/lucidrains/x-transformers
 Author: Phil Wang