X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=mygpt.py;h=6a12a5a35c760b41bd80db009a3705942b3dddda;hb=ed60c541ca2225d69df96c2a382bb83c947bfe0e;hp=5ea4668203f0f4eaf9ccbc586f2d58a348fe0a3f;hpb=0633abb4fa80284003f1d6d1b5b9e758c35ed10f;p=picoclvr.git

diff --git a/mygpt.py b/mygpt.py
index 5ea4668..6a12a5a 100755
--- a/mygpt.py
+++ b/mygpt.py
@@ -5,6 +5,11 @@
 
 # Written by Francois Fleuret <francois@fleuret.org>
 
+# This is an implementation from scratch of a "GPT", that is a model
+# composed of several causal self-attention blocks. It is equipped
+# with a caching mechanism for keys and values to avoid a O(N^3) cost
+# for auto-regression.
+
 import math
 
 import torch
@@ -14,19 +19,6 @@ from torch.nn import functional as F
 
 ######################################################################
 
-
-class WithResidual(nn.Module):
-    def __init__(self, *f):
-        super().__init__()
-        self.f = f[0] if len(f) == 1 else nn.Sequential(*f)
-
-    def forward(self, bs):
-        bs.x = bs.x + self.f(bs).x
-        return bs
-
-
-######################################################################
-
 # A BracketedSequence is a BxTx... tensor with a first and a nb time
 # steps to compute.
 
@@ -78,6 +70,19 @@ class CacheWrapper(nn.Module):
 ##############################
 
 
+class WithResidual(nn.Module):
+    def __init__(self, *f):
+        super().__init__()
+        self.f = f[0] if len(f) == 1 else nn.Sequential(*f)
+
+    def forward(self, bs):
+        bs.x = bs.x + self.f(bs).x
+        return bs
+
+
+##############################
+
+
 class AddPositionalEncoding(nn.Module):
     def __init__(self, len_max):
         super().__init__()
@@ -197,7 +202,6 @@ class MyGPT(nn.Module):
         dropout=0.0,
         len_max=1e5,
     ):
-
         super().__init__()
 
         assert dim_model % nb_heads == 0
@@ -258,7 +262,6 @@ class MyGPT(nn.Module):
 ######################################################################
 
 if __name__ == "__main__":
-
     print("Basic check.")
 
     vocabulary_size = 10