[GitHub] [tvm] yelite commented on a diff in pull request #12914: [MetaSchedule] Add Script for TorchBench Model Tuning & Benchmarking

[GitBox <gi...@apache.org>]

yelite commented on code in PR #12914:
URL: https://github.com/apache/tvm/pull/12914#discussion_r982847452


##########
python/tvm/meta_schedule/testing/torchbench/run.py:
##########
@@ -0,0 +1,591 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements.  See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership.  The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License.  You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied.  See the License for the
+# specific language governing permissions and limitations
+# under the License.
+"""
+This script is for benchmarking TVM performance on models from TorchBench.
+It uses the TorchDynamo as the frontend to ingest models into TVM, and it also
+leverages the benchmark util from TorchDynamo.
+
+TorchDynamo (https://github.com/pytorch/torchdynamo) and TorchBench
+(https://github.com/pytorch/benchmark) need to be in the parent directory of TVM.
+We need a local clone of these repos because torchbench and the benchmark runner
+in TorchDynamo isn't designed to be used as a Python package.
+
+To setup the environment, run the following commands in the parent directory of TVM and with 
+the appropriate Python environment:
+```bash
+# torchdynamo requires nightly pytorch. If it fails to find the specified version, try
+# installing the latest nightly pytorch.
+pip3 install --pre \
+    --extra-index-url https://download.pytorch.org/whl/nightly/cu116 \
+    torch==1.13.0.dev20220926 \
+    torchvision==0.14.0.dev20220926 \
+    torchtext==0.14.0.dev20220926
+
+git clone https://github.com/pytorch/torchdynamo
+pushd torchdynamo
+git checkout c537639f9712621dc04ca09908796dbbe86c354b
+pip install -e .
+popd
+
+sudo apt install git-lfs  # git lfs is used for TorchBench
+git clone https://github.com/pytorch/benchmark
+pushd benchmark
+python install.py --continue_on_fail  # fambench_xlmr might fail to install
+popd
+```
+
+To run a benchmark, the script can be run under 'tune' mode by
+```bash
+python python/tvm/meta_schedule/testing/torchbench/run.py \
+    --mode tune \
+    --model resnet50 \
+    --target "nvidia/geforce-rtx-3070" \
+    --work-dir ../workdir \
+    --num-trials 20000 \
+    --rpc-host <rpc tracker host for tuning> \
+    --rpc-port <rpc tracker port for tuning> \
+    --rpc-key <rpc key> \
+```
+
+All available target tags (like nvidia/geforce-rtx-3070) can be found at
+https://github.com/apache/tvm/blob/main/src/target/tag.cc
+
+Then the script can be run under 'eval' mode to actual benchmark the performance,
+using the tuning database under the work directory. This can be executed on a different
+machine than the one executes tuning (the database json files need to be inside
+of the work directory).
+```bash
+python python/tvm/meta_schedule/testing/torchbench/run.py \
+    --mode eval \
+    --model resnet50 \
+    --target "nvidia/geforce-rtx-3070" \
+    --work-dir ../workdir \
+    --num-trials 0 
+```
+
+Alternatively, both tuning and evaluation can be done in a single run on the same machine,
+by
+```bash
+python python/tvm/meta_schedule/testing/torchbench/run.py \
+    --mode all \
+    --model resnet50 \
+    --target "llvm -num-cores 6" \
+    --work-dir ../workdir \
+    --num-trials 0
+```
+"""
+
+import argparse
+import functools
+import logging
+import warnings
+from enum import Enum
+from typing import Callable, List, Tuple
+
+import numpy as np  # type: ignore

Review Comment:
   They are for suppressing errors like `Cannot find implementation or library stub for module named "torch"`(https://ci.tlcpack.ai/blue/organizations/jenkins/tvm/detail/PR-12914/7/pipeline#step-97-log-74)



-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: commits-unsubscribe@tvm.apache.org

For queries about this service, please contact Infrastructure at:
users@infra.apache.org