X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=blobdiff_plain;f=mygpt.py;h=6a12a5a35c760b41bd80db009a3705942b3dddda;hb=f91736e6e56152746b3c44342748b70ad1c89888;hp=0ed7eb0b15be4130cc7b7feb733be14d8c368860;hpb=943a440a83b98de60bad767a9ad09f63b5088514;p=picoclvr.git

diff --git a/mygpt.py b/mygpt.py
index 0ed7eb0..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__()
@@ -128,17 +133,15 @@ class QKVAttention(nn.Module):
         self.w_v = randw(nb_heads, dim_v, dim_in)
         self.w_o = randw(dim_v * nb_heads, dim_in)
 
-    def forward(self, bs_q, x_kv=None):
+    def forward(self, bs_q):
         x_q = bs_q.x
-        if x_kv is None:
-            x_kv = x_q
 
         if bs_q.first == 0:
             self.cache_k = x_q.new_zeros(
-                x_q.size(0), self.w_k.size(0), x_kv.size(1), self.w_k.size(1)
+                x_q.size(0), self.w_k.size(0), x_q.size(1), self.w_k.size(1)
             )
             self.cache_v = x_q.new_zeros(
-                x_q.size(0), self.w_v.size(0), x_kv.size(1), self.w_v.size(1)
+                x_q.size(0), self.w_v.size(0), x_q.size(1), self.w_v.size(1)
             )
             self.cache_y = x_q.new_zeros(x_q.size(0), x_q.size(1), self.w_o.size(1))
 
@@ -146,10 +149,10 @@ class QKVAttention(nn.Module):
             "ntc,hdc->nhtd", x_q[:, bs_q.first : bs_q.first + bs_q.nb], self.w_q
         )
         self.cache_k[:, :, bs_q.first : bs_q.first + bs_q.nb] = torch.einsum(
-            "ntc,hdc->nhtd", x_kv[:, bs_q.first : bs_q.first + bs_q.nb], self.w_k
+            "ntc,hdc->nhtd", x_q[:, bs_q.first : bs_q.first + bs_q.nb], self.w_k
         )
         self.cache_v[:, :, bs_q.first : bs_q.first + bs_q.nb] = torch.einsum(
-            "ntc,hdc->nhtd", x_kv[:, bs_q.first : bs_q.first + bs_q.nb], self.w_v
+            "ntc,hdc->nhtd", x_q[:, bs_q.first : bs_q.first + bs_q.nb], self.w_v
         )
 
         a = torch.einsum(
@@ -160,7 +163,7 @@ class QKVAttention(nn.Module):
             if bs_q.first == 0:
                 self.cache_attzero = (
                     torch.arange(x_q.size(1), device=q.device)[None, None, :, None]
-                    < torch.arange(x_kv.size(1), device=q.device)[None, None, None, :]
+                    < torch.arange(x_q.size(1), device=q.device)[None, None, None, :]
                 )
             a = a.masked_fill(
                 self.cache_attzero[
@@ -199,7 +202,6 @@ class MyGPT(nn.Module):
         dropout=0.0,
         len_max=1e5,
     ):
-
         super().__init__()
 
         assert dim_model % nb_heads == 0
@@ -260,7 +262,6 @@ class MyGPT(nn.Module):
 ######################################################################
 
 if __name__ == "__main__":
-
     print("Basic check.")
 
     vocabulary_size = 10