[MetaSchedule][M4a] User-API: Tune-TE/TIR/Relay#10079
Conversation
|
Let's wait for the rest 3 rules to be upstreamed before merging this PR :-) |
490dfcc to
512ac20
Compare
Co-authored-by: Junru Shao <junrushao1994@gmail.com> Co-authored-by: Bohan Hou <32121147+spectrometerHBH@users.noreply.github.com> Co-authored-by: Ruihang Lai <lairuihangdongdong@qq.com> Co-authored-by: Hongyi Jin <3231950289@qq.com> Co-authored-by: Wuwei Lin <wuwei@apache.org> Co-authored-by: Siyuan Feng <Hzfengsy@sjtu.edu.cn> Minor fix. Nits. Add back tests.
512ac20 to
e213c9c
Compare
| test_meta_schedule_tune_relay("resnet18", 1, "llvm --num-cores=16") | ||
| test_meta_schedule_tune_relay("resnet18", 1, "nvidia/geforce-rtx-3070") | ||
| test_meta_schedule_tune_relay("mobilenet_v2", 1, "llvm --num-cores=16") | ||
| test_meta_schedule_tune_relay("mobilenet_v2", 1, "nvidia/geforce-rtx-3070") | ||
| test_meta_schedule_tune_relay("bert_base", 1, "llvm --num-cores=16") | ||
| test_meta_schedule_tune_relay("bert_base", 1, "nvidia/geforce-rtx-3070") |
There was a problem hiding this comment.
Wonder how long would these 6 unittests take in total?
* Add tuning scripts for tir, te & relay. Co-authored-by: Junru Shao <junrushao1994@gmail.com> Co-authored-by: Bohan Hou <32121147+spectrometerHBH@users.noreply.github.com> Co-authored-by: Ruihang Lai <lairuihangdongdong@qq.com> Co-authored-by: Hongyi Jin <3231950289@qq.com> Co-authored-by: Wuwei Lin <wuwei@apache.org> Co-authored-by: Siyuan Feng <Hzfengsy@sjtu.edu.cn> Minor fix. Nits. Add back tests. * slightly improve tune.py Co-authored-by: Junru Shao <junrushao1994@gmail.com>
| int running_tasks = tasks.size(); | ||
| for (int task_id; (task_id = NextTaskId()) != -1;) { | ||
| LOG(INFO) << "Scheduler picks Task #" << task_id << ": " << tasks[task_id]->task_name; | ||
| LOG(INFO) << "Scheduler picks Task #" << task_id + 1 << ": " << tasks[task_id]->task_name; |
There was a problem hiding this comment.
Hi Xiyou! I'm curious about this change. After this change the printed ids at several places differ (here with + 1 while others without.)
Perhaps it's better to keep them consistent 👀?
There was a problem hiding this comment.
Hi, thanks for pointing that out. The reason for this change is previously we have task 0/18 to task 17/18 which seems confusing to me. I intend to make sure the log is consistent but I might missed a couple places. After second thoughts I think it might be better to keep the task_id the same way it is so that it would be easier for debugging later. See #10478.
This PR is part of the stage M4a of the meta schedule project (#8473).
The architecture is re-designed by Junru and Xiyou. In this PR we introduced the tuning scripts for tir, te & relay workloads with meta schedule. Unittests are added for each type of tuning, and for relay workloads we added tests of ResNet18, MobileNetV2 and BertBase.
Thanks to all co-authors for contributing!
Co-authored-by: Junru Shao <junrushao1994@gmail.com>
Co-authored-by: Bohan Hou <32121147+spectrometerHBH@users.noreply.github.com>
Co-authored-by: Ruihang Lai <lairuihangdongdong@qq.com>
Co-authored-by: Hongyi Jin <3231950289@qq.com>
Co-authored-by: Wuwei Lin <wuwei@apache.org>
Co-authored-by: Siyuan Feng <Hzfengsy@sjtu.edu.cn>