TorchDynamo Update 8: TorchDynamo passed correctness check on 7k+ github models

smth · June 30, 2022, 5:22am

I was wondering how we generated valid input shapes for all these models.

Apparently @jansel wrote a fairly nice deducer with a funny search that reads the error messages and tries to prune the space down.

Had a lot of fun reading through the code.

github.com

jansel/pytorch-jit-paritybench/blob/master/paritybench/deduce_parameters.py

import ast
import inspect
import itertools
import logging
import operator
import os
import re
import sys
import time
import traceback
from functools import reduce
from typing import Callable, List

import torch

from ._paritybench_helpers import _mock_layer

log = logging.getLogger(__name__)

This file has been truncated. show original

Topic		Replies	Views
TorchDynamo Update 5: Improved Capture & Bigger Graphs compiler	4	4230	April 14, 2022
TorchDynamo: An Experiment in Dynamic Python Bytecode Transformation compiler	10	18321	December 12, 2025
TorchDynamo Update 3: GPU Inference Edition compiler	12	6943	February 2, 2023
TorchDynamo Update: 1.48x geomean speedup on TorchBench CPU Inference compiler	0	5849	November 12, 2021
TorchDynamo Update 4: LazyTensor & nvFuser Experiments compiler	4	4963	February 9, 2024

TorchDynamo Update 8: TorchDynamo passed correctness check on 7k+ github models

Related topics