[Apache TVM Discuss] [Development/pre-RFC] [RFC] Visualizing Relay program as graph

2022-02-15 Thread Mingli Sun via Apache TVM Discuss


any way to visualize optimized relay IR? this seems not working to visualize 
optimized relay IR.
```
opt_level = 4
target = tvm.target.Target('cuda')
with tvm.transform.PassContext(opt_level=opt_level):
mod,params = tvm.relay.optimize(mod,target,params=params)
# visualize mod['main']
```





---
[Visit 
Topic](https://discuss.tvm.apache.org/t/rfc-visualizing-relay-program-as-graph/4825/24)
 to respond.

You are receiving this because you enabled mailing list mode.

To unsubscribe from these emails, [click 
here](https://discuss.tvm.apache.org/email/unsubscribe/09536a755864481a285e395fe09b3b26743c82e14a91cc06d804fae10943828f).


[Apache TVM Discuss] [microTVM] No microTVM Community Meeting Feb 16

2022-02-15 Thread Andrew Reusch via Apache TVM Discuss


Hi all,

We're a bit busy this week and haven't gotten the microTVM Community Meeting 
announcement out with enough time to collect ideas, so we'll skip this meeting 
and resume on Mar 2.

Thanks,
Andrew





---
[Visit 
Topic](https://discuss.tvm.apache.org/t/no-microtvm-community-meeting-feb-16/12094/1)
 to respond.

You are receiving this because you enabled mailing list mode.

To unsubscribe from these emails, [click 
here](https://discuss.tvm.apache.org/email/unsubscribe/f9f92f9266c2986260ed557aef00a7cb81cdda15f612116da5e4b2aac90fccd1).


[Apache TVM Discuss] [Development/pre-RFC] [RFC] Remove CODEOWNERS

2022-02-15 Thread driazati via Apache TVM Discuss


# **Summary**

Move `.github/CODEOWNERS` to `.github/CODEOWNERSHIP` to avoid triggering 
GitHub’s [automatic review 
requests](https://docs.github.com/en/repositories/managing-your-repositorys-settings-and-features/customizing-your-repository/about-code-owners#about-code-owners).
 Add GitHub Actions automation to a) populate Review Requests based on PR 
thread traffic and b) ping languishing PRs.

# **Motivation**

Languishing PRs are a relatively common occurrence in TVM today. In order to 
maintain a vibrant open source community, we should work to reduce or eliminate 
these occurrences. PRs languish for a variety of reasons, but a common problem 
new contributors have is finding a reviewer for their PR.

In [June 2021](https://github.com/apache/tvm/pull/8500), TVM attempted to solve 
this problem by introducing `CODEOWNERS`. The goal was to make it easy for 
contributors to find a reviewer for their PR. This approach was problematic 
because the layout of TVM code meant that a file-based approach to sharding 
code ownership was incompatible with most PRs. The average PR spans so many 
`CODEOWNERS` directories that the average PR triggered code-owner requests to 
half of the TVM committers, diluting the responsibility of each reviewer and in 
many cases generating spam for reviewers who won’t end up reviewing the PR at 
all.

It would be nice to rely on automation by tuning the `CODEOWNERS` file. 
However, the organization of TVM is such that the scope of a directory includes 
many different efforts. For instance, a change in `src/relay/backend` may 
affect the core compiler but may also affect automation and runtime. Tuning 
`CODEOWNERS` could well amount to adding file-level ownership, and maintaining 
that is intractable.

We also attempted to tune `CODEOWNERS` by switching to round-robin review style 
in ‣, but this approach ran into problems, summarized by @areusch in [this 
comment](https://github.com/apache/tvm/issues/9057#issuecomment-931579113). 

## Guide-level explanation

Many reviewers find it difficult to sort through review traffic and determine 
which PRs they are on the hook for. GitHub provides solutions for this in the 
form of Review Requests and Assignee fields—reviewers can list the PRs which 
mention them there. Since `CODEOWNERS` worked by auto-populates the Review 
Requests field, removing `CODEOWNERS` from the repo in turn allows us to reuse 
these fields to better track who is on the hook for reviewing a PR. We should 
take this opportunity in spam reduction to attempt to make GitHub PR traffic 
more relevant to the community as a whole. A great way to do this is to develop 
a better system for populating Review Requests and Assignee fields to leverage 
the GitHub PR review system.

However, absent `CODEOWNERS`, these fields need to be manually filled. One 
could imagine that TVM Committers and Triagers could triage new PRs and 
populate those fields. However, there are some limitations on this system 
imposed by GitHub and the Apache Software Foundation which make this difficult. 
Specifically, anyone mentioned in a PR must either be a TVM committer or 
actively participating (e.g. by replying) in the PR in order to be placed in 
the Review Requests field. This means that a committer must continuously 
monitor a PR thread to keep those fields as accurate as possible.

Since manual processes often lead to inconsistencies, and the conditions above 
are somewhat adversarial, some automation is desirable here to attempt to 
standardize on one system for tagging PRs with assigned reviewers. To address 
that need, this RFC proposes two additional changes:

1. Automatically assigning reviewers based on cc tags in PR messages
2. Periodic automated ping messages for participants in PRs to prevent PRs from 
languishing

Committers are responsible for monitoring and triaging new PRs and issues to 
the relevant parties, and this RFC doesn’t change that. It assists by reducing 
notification spam so that each notification a committer gets is now something 
that needs to be addressed.

# **Reference-level explanation**

Removing `CODEOWNERS` solves the spam issue by stopping the deluge of 
notifications to committers, but introduces a new issue in that PR authors 
still need to be able to assign reviewers. The combination of 
https://github.com/apache/tvm/pull/9934 and 
https://github.com/apache/tvm/pull/9973 are meant to address this. Since many 
TVM contributors don’t have permissions to add reviewers themselves, anyone who 
is a committer that is addressed in a PR body with `cc @username` will be added 
as a reviewer.

PRs should not stay open forever and should get reviewed in a timely manner. 
The second PR linked above addresses this by periodically (currently set to 
wait 7 days) pinging PRs that have not had recent activity.

These two tools should make it so the TVM community is still able to maintain a 
good velocity on PRs while avoiding spamming committers with notifications.


Re: [apache/tvm-rfcs] [RFC][BYOC] Marvell ML/AI Accelerator Integration (PR #48)

2022-02-15 Thread Andrew Reusch
@ccjoechou sorry for the delay--i've gotten pretty busy with something and will 
hopefully have some bandwidth towards the end of the week.

cc @jroesch @mbs-octoml in case they have cycles

-- 
Reply to this email directly or view it on GitHub:
https://github.com/apache/tvm-rfcs/pull/48#issuecomment-1040919020
You are receiving this because you are subscribed to this thread.

Message ID: 

Re: [apache/tvm-rfcs] [RFC][BYOC] Marvell ML/AI Accelerator Integration (PR #48)

2022-02-15 Thread Joe Chou
@areusch: No worries. I saw lots TVM emails coming from you & others working on 
other also important stuffs. We will wait for your feedback.

-- 
Reply to this email directly or view it on GitHub:
https://github.com/apache/tvm-rfcs/pull/48#issuecomment-1040923708
You are receiving this because you are subscribed to this thread.

Message ID: