Python at 1.09× CPython, torch at 4.2× PyTorch

Python on the command line runs at 1.09x CPython 3.13 with fifteen programs faster than CPython, and torch training and tensor programs at 4.2x PyTorch's warm time, from 17x.

A store-only __init__ now runs without its frame (0c0fb3b5), method lookups and construction sites answer inline for the class they last saw (8519dc28), sort keys of one form sort by columns (93fac141), a subscript reads a list's, tuple's or dict's record once (e7153c9f), raise has a compiled helper (a3476e76), and loops are planned only where the header dominates the back edge (c5374b88). Over the 36 programs Python runs at 1.09× CPython 3.13's time, from 1.31× at v0.0.23; the 30 small programs are level with CPython, and fifteen programs beat it. 05704850 fixed a zero-argument super() beside a lambda or inside a comprehension, and 79759baa won back 5% on every interpreted Python call after a dispatch-loop change had cost it.

Torch draws its random numbers natively (ce8e1494) and uniform_ now rounds as PyTorch does (832c3756); add, sub, mul, div and the common matrix products are one engine call (4c2c8ed2, 69471503), and powers are computed the way PyTorch's kernel computes them (10d8465e). On one thread, warm, eight training and tensor programs take 4.2× PyTorch 2.11's time, where v0.0.23 took 17×, and a program's first run is faster than PyTorch's on four of them.