►
From YouTube: Apache TVM Community Meeting, September 23 2021
Description
Featured Agenda:
* Roadmap preRFC
* 0.8 Release News - Junru Shao
Join the TVM Community at http://tvm.apache.org/community
A
All
right
so
again
welcome
to
the
september
edition
of
the
tvm
community
meeting.
We
have
a
pretty
big
agenda
today.
We
there
are
there's
a.
A
There
are
a
bunch
of
things
regarding
related
to
the
road
map
and
the
release
of
tvm
that
we're
going
to
be
covering
today
and
so,
even
though
it's
just
a
couple
of
primary
agenda
items,
I
think
it's
going
to
be
a
lot
to
fit
in
during
the
45
minutes,
but,
as
usual,
we
we
start
the
meetings
with
with
introductions
for
for
people
who
are
maybe
new
to
the
meetings.
A
A
All
right
with
with
with
with
nobody
stepping
up,
I
still
want
to
say,
welcome
to
everyone,
and
we
also
have
a
few
announcements.
Over
the
last
month,
the
tvm
community
has
continued
to
grow
and,
as
as
a
lot
of
its,
its
contributors
have
have
kind
of
you
know,
made
significant
contributions
and
they've
and
and
we'd
like
to
recognize
them
for
that
by
you
know,
we
either
either
promoting
them
to
committer
or
reviewer
status.
So
gsrp
rossini
is
a
new
committer.
A
Congratulations,
committer
is
somebody
who
is
able
to
not
who
has
made
significant
contributions
to
tvm
and
is
trusted
to
merge
code
into
the
repository,
and
so
this
is
a.
This
is
a
big
step.
Cu
and
fang
has
also
been
promoted
to
committer.
A
Matt
welch
has
been
promoted
as
a
new
reviewer
for
the
project.
Manupa
karunante
is
a
new
committer,
as
well
as
eric
lunderberg
as
a
new
reviewer.
So
and
and
reviewers
are
folks
who
have.
A
You
know
made
a
lot
of
contributions,
but
have
also
demonstrated
a
commitment
to
reviewing
other
people's
patches
so-
and
this
is
just
a
reminder
that
open
source
projects
depend
upon
a
community
coming
together
and
collaborating,
and
so
that
means
not
just
sending
up
pull
requests
but
also
participating
in
the
review
process,
and-
and
so
we
want
to
thank
everybody
who
has
been
contributing
to
tvm
and
who
has
you
know,
really
shown
their
commitment
to
the
project
and
and
making
it
better.
So
a
big
round
of
congratulations
to
everyone
on
the
list.
A
Okay,
another
announcement
coming
up
is
the
cfp
for
tvm
con
is
now
open.
You
can
you
can
find
a
link
to
it
at
tvmconf.org,
we're
also
working
on
a
new
tvmcon
website
right
now.
A
So
you
should
keep
your
eyes
open
that
for
the
next
few
weeks
and
we're
looking
for
talks,
whether
they
be
short,
lightning,
talks
to
larger
feature,
presentations
to
case
studies
that
are
not
just
related
to
tvm,
but
also
to
machine
learning
acceleration,
and
so,
if
they're,
any
work
that
you're
doing
on
tvm,
but
also
you
know
adjacent
projects
that
you
know
you
think
would
be
of
general
interest
to
the
tvm
community.
Please
submit
a
cfp,
we're
going
to
be
closing
in
october
and
and
and
and
building
a
fantastic
schedule
this
year.
A
Once
again,
the
event
is
going
to
be
entirely
virtual.
We
have
also
moved
the
event
to
december
15-17
to
avoid
some
conflicts
with
other
events
and
also
give
a
little
more
space
for
us
to
prepare
so
mark
the
dates
on
your
calendar
for
the
15th
to
the
17th.
A
It's
going
to
be
an
entirely
free
conference
again
and
but,
if
you're
interested
in
sponsoring,
please
feel
free
to
reach
out
to
me
at
events
at
octoml.ai,
because
we're
also
trying
to
go
we're
trying
to
make
the
event
larger,
bring
in
some
more
bring
in
some
more
speakers,
make
it
a
more
polished
experience
and
and
and
so,
if
you're
interested
in
in
participating
and
helping
realize
that
vision.
We
would
love
to
hear
from
you
and
again
that's
events
at
octoaml.ai.
A
So
thank
you,
everyone
and,
if
there's
any
questions
about
tvm
con
or
we
can
or
we
can
turn
it
over
to
denise,
to
talk
about
the
roadmap
pre-rfc
that
she's
been
working
on.
B
B
B
I'll
talk
a
little
bit
more
about
that
later.
But
what
I'm
trying
to
say
here
is
that
there's
some
really
fine-grained
planning
like
the
individual
task
tracking
on
github
and
then
there's
some
stronger
technical
visions
like
some
of
the
things
that
tnt
posts
or
some
of
the
rfcs
that
all
of
you
in
the
community
post
and
we're
figuring
out
how
to
unify
all
of
those
into
an
easily
readable
context.
B
And
then
the
next
one
is
to
provide
everyone
with
visibility
into
all
of
tbm's
projects.
So
this
is
very
useful
for
people
who
are
newer
to
the
project
or
people
who
focus
in
one
particular
area
of
tbm
and
want
to
discover
more
areas
of
tbm.
So
we're
really
aiming
to
make
it
easy
for
everyone
to
discover
different
levels
of
the
stack
and
to
discover
what
we're
doing
on
the
community
side,
etc,
and
then
the
last
one
is
to
encourage
collaborative
and
open
planning.
B
There
are
so
many
contributors
to
the
open
source
community
and
really
we
want
to
make
sure
that
everyone
is
working
in
a
synergistic
way
and
that
we're
all
working
to
solve
the
same
problems
and
we're
working
together.
So
I
think
that
having
a
roadmap
and
having
clear
ownership
will
really
help
with
that.
B
B
One
of
them
is
the
task
tracking
that
I
mentioned
a
little
bit
earlier,
so
I
have
an
example
on
the
very
left
side
of
a
tracking
issue
that
jen
roo
created
for
the
meta
scheduler,
and
so
this
is
one
of
the
finest
grain
task
tracking
that
exists
in
tbm
right
now,
and
it's
really
useful
for
someone
who's
read
the
rfc
to
take
a
look
at
where
the
project
is
at.
B
It
shows
multiple
features,
getting
added
into
versions
of
tvm
and,
lastly,
we
have,
on
the
rightmost
side
a
technical
vision,
and
this
is
where
we
see
the
project
going
and
tnc
has
posted
these
for
the
past
couple
of
releases,
basically
for
each
level
of
the
stack
there's
a
short
overview
of
features
that
are
to
be
created
here,
and
so
basically,
all
of
these
are
great,
and
the
road
map
is
not
meant
to
replace
any
one
of
these.
It's
meant
to
unify
these
and
make
it
more
readable,
more
discoverable
for
the
community.
B
B
So
there
are
four
primary
areas
of
this
roadmap
that
I'd
like
to
walk
you
through
today.
The
first
one
is
the
background
and
motivations
here,
and
so
this
is
really
helpful
for
someone
who's,
maybe
a
little
bit
newer
to
this
particular
sub
project
or
this
particular
piece
of
tvm.
So
this
is
meant
to
give
like
an
overview
of
key
themes
for
whatever
project
or
subproject,
that
the
roadmap
is
covering
in
this
case,
for
ci
and
testing.
B
So
these
are
really
meant
to
unify
the
features
and
the
task
tracking
and
rfcs
into
a
single
road
map
area
and
also
to
show
it
over
a
longer
term
timeline
and
more
of
a
fixed
timeline,
because
releases
in
the
past
for
tvm
have
been
not
like
on
a
specific
time
basis,
but
having
a
roadmap
that
can
specifically
talk
quarter
by
quarter
about
what
everyone
in
the
community
is
working
on
is
really
important.
B
Maybe
they
see
something
that
they
would
like
to
change
in
tbm,
and
this
is
a
higher
visibility
way
of
showing
it
than
simply
filing
a
github
issue
and
it'll
be
really
nice
when
we
have
roadmaps
for
each
of
the
sub-projects
in
tbm
to
be
able
to
have
a
backlog
for
each
of
those
projects,
and
so
that
people
who
are
new
to
the
project
can
look
at
this
context.
They
can
understand
exactly
the
impacts
that
they're
having
to
the
project
as
they're,
contributing
to
the
backlog
and
moving
those
items
onto
our
roadmap.
B
Okay,
so
going
into
some
of
the
design
considerations,
I
first
wanted
to
talk
about
content
content
is
really
key
and
I
want
to
make
sure
that
the
content
here
is
really
useful
for
people
to
digest.
So
I
really
want
to
emphasize
that
we
want
to
give
people
the
background
information
that
they
need
to
be
able
to
synthesize
any
roadmap,
whether
or
not
they're
an
expert.
B
Each
of
the
road
maps
should
contain
themes
which
encapsulate
key
focus
areas
of
the
roadmap.
You
saw
that
in
the
demo
that
I
just
presented
and
then
the
roadmap
should
contain
items
which
describe
either
features
or
milestones
that
are
to
be
implemented
within
the
timeline
of
that
roadmap,
and
so
within
the
items.
What
I
didn't
show
you
is:
what
actually
is
the
content
within
the
items
so
I'll
describe
it
here?
Briefly,
so
each
item
should
have
clearly
defined
motivations,
end
goals,
deliverables
and
scope.
B
This
is
really
important
context,
not
just
for
yourself
but
for
the
entire
community
and
also
like
smaller
tasks,
can
be
combined
into
a
larger
milestone
and,
for
instance,
if
we
have
a
bunch
of
flaky
tests
in
tvm,
we
can
combine
them
into
a
larger
milestone,
like
fixing
all
of
the
flaky
tests,
and
then
we
can
put
that
onto
the
road
map
and
then
larger
features
such
as
auto
tir
and
meta
scheduling
that
could
be
split
into
a
bunch
of
smaller
milestones,
to
show
like
continued
progress
towards
a
much
larger
feature,
and
so
in
terms
of
some
of
the
items
that
I
mentioned
in
the
background
information,
I
wanted
to
touch
on
schedule
and
prioritization
a
little
bit.
B
This
is
really
important
because,
in
my
mind,
the
priority
of
any
given
feature
should
be
determined
by
community
dis
consensus,
and
then
it
should
be
displayed
prominently
to
the
community
so
that
the
community
can
understand
what
is
important
to
review
and
what
is
important
to
try
and
push
into
the
code
base.
B
Based
on
the
consensus
that
everyone
in
the
community
has
come
to,
and
then
dependencies
and
blockers
should
be
easily
traceable
across
roadmaps
and
then
the
delivery
schedule
quarter
by
quarter
should
be
estimated
based
on
those
prioritizations
dependencies
and
blockers,
and
it
should
be
updated,
as
our
confidence
increases,
whether
that's
with
prs,
merging,
rfcs,
merging
or
anything
else,
and
then
the
last
one
I
wanted
to
cover
is
ownership.
This
is
especially
important
in
a
community
with
so
many
collaborators.
B
B
Okay
and
then
another
design,
consideration
that
we
had
in
this
roadmap
is
overall
usability
and
maintainability,
and
so
really
there's
a
few
key
themes
here.
We
want
to
encourage,
sharing
and
collaboration
in
the
early
stages
of
planning,
and
really
I
see
the
road
map
as
something
that
we
can
start
using
as
a
pre-rfc
incubator.
And
what
I
mean
by
this
is
that
a
road
map
can
describe
what
changes
are
desired.
B
You
can
describe
a
problem
that
you
think
needs
fixing
in
tbm,
or
you
could
describe
something
that
you
think
would
be
an
improvement
to
tbm
without
proposing
an
immediate
solution,
and
then
the
rfc
goes
into
a
little
bit
more
detail
and
describes
how
changes
will
be
implemented,
and
this
can
be
a
virtuous
cycle
where
things
are
initially
defined
in
the
roadmap.
Like
I
may
want
to
work
on
this
in
the
future.
B
I'll
put
it
in
the
backlog,
then
you
create
an
rfc
and
once
that's
merged,
you
can
move
it
back
into
the
roadmap
and
then
move
the
item
into
the
particular
quarter
where
it's
delivered
and
then
the
next
key
principle
is
leveraging
our
existing
task
tracking
mechanisms,
the
ones
shown
in
the
prior
art
slide
that
I
showed
earlier.
Really.
This
makes
it
so
that
there's
easier
maintenance
for
everyday
contributors
and
so
that
we
can
show
updated
progress
towards
feature
completion
automatically
as
people
merge
prs.
B
B
And
so,
lastly,
I
wanted
to
talk
about
some
of
the
upstreaming
timelines.
I
consider
this
to
be
overall
milestone
one
and
so
the
overall
goal
of
milestone
one
is
to
get
the
overall
tvm
roadmap,
as
well
as
the
rest
of
the
sub
component
roadmaps,
upstreamed
and
to
get
all
opens
in
the
rfcs
addressed
and
to
have
documentation
and
public
like
pr
type
things
for
this
roadmap,
and
so
I
split
that
into
three
phases.
B
So
this
one
m1a
is
going
to
be
the
first
one
that
I'll
make
an
rfc
on
in
a
little
bit,
and
basically,
I
want
to
like
basically
upstream
the
initial
prototype
that
I
showed
earlier
in
this
demo
and
then
I'll
have
some
documentation
as
well
as
some
scoping
for
the
future
milestones
and
in
the
future,
milestones
I'll
start
upstreaming,
additional
content
that
I've
created
in
collaboration
with
the
community
and
then
by
the
time
milestone.
One
is
complete.
B
All
of
the
road
maps
will
be
upstreamed,
like
I
said,
and
all
of
the
rfcs
will
be
addressed,
and
if
you
have
any
ideas
for
how
we
can
improve
the
road
mapping
on
its
own,
we
can
incubate
these
in
the
tbm
community
roadmap
until
milestone.
One
is
complete.
They'll
just
enter
the
backlog
and
we'll
work
on
them.
Eventually,.
B
Okay,
so
with
that,
thank
you
so
much
for
your
time
and
listening
to
me
for
10-ish
minutes.
All
feedback
is
now
welcome
and
I
just
wanted
to
especially
thank
certain
folks
on
this
page,
just
for
all
their
feedback
and
support
as
I've
worked
on
this
roadmap
in
terms
of
helping
me
with
content
and
giving
me
feedback,
it's
been
really
helpful.
B
C
Thanks
denise,
it's
really
great
for
presentation,
and
so
one
thing
I'm
thinking
about
this
as
a
distributed
community.
We
are
always
working
together
across
different
companies,
so
I
I
would
say
I
like
to
ask
if
it
is
how
the
procedure
will
look
like
under
a
current
roadmap.
If,
if
like
during
a
release
process,
a
company
may
come
may
come
up
and
say
we
want
to
upstream
those
kind
of
features
to
support
our
own
chips
and
to
apache
tvm.
C
And
so
what
is
what
does
the
procedure
look
like
if
we
want
to
put
that
a
community
developer
want
to
put
something
onto
the
roadmap.
B
Yeah,
this
isn't
a
great
question
and
I
think
this
process
is
really
in
its
early
stages
and
we're
going
to
be
looking
for
feedback
from
the
community.
As
this
happens,
I
think
really
it's
going
to
be
not
too
much
different
from
the
current
rfc
process
in
that
you
will
still
make
an
rfc
in
the
process
of
trying
to
merge
your
code,
especially
if
it's
a
big
feature,
change.
Adding
it
to
the
roadmap
is
simply
a
way
for
more
visibility,
as
well
as
a
place
for
earlier
discussion
prior
to
the
rfc.
B
So,
for
instance,
anyone
should
be
able
to
file
an
issue
to
the
roadmap
repository
and
describe
something
that
they're
interested
in
working
on.
For
instance,
if
you
were
trying
to
describe
like
auto
tir
at
a
very
high
level,
you
might
say
I
want
to
improve
the
scheduling
infrastructure
to
use
both
template,
free
and
template-based
scheduling,
and
you
wouldn't
go
into
the
details
about
how.
B
But
you
would
create
that
and
you
might
go
into
some
of
the
motivations
of
that,
and
that
would
be
a
roadmap
item
and
as
you
go
through
the
typical
rfc
and
community
consensus
process,
you
can
add
those
additional
details
to
the
roadmap
and
then
add
it
to
a
delivery
schedule.
As
you
build
more
confidence
overall
within
the
community
that
you're
going
to
deliver
something
towards
a
given
timeline.
Does
that
answer
your
question.
C
Yeah,
that
sounds
great
sounds
really
really
great.
So
if,
if
there
is
a
rfc
because
we
have
a,
we
have
a
mechanism
that
after
apps
is
accepted,
we
create
a
tracking
issue
for
the
nfc.
So
do
we
include
the
tracking
issue
in
the
roadmap
like
automatically
or
like
that.
B
Yes,
yes,
that
is
the
plan
to
reuse
the
tracking
issues
that
are
created
and
to
make
the
roadmap
items
so
that
they
are
unified
so
that
you
get
the
motivations
from
the
start.
Then
some
of
the
details
from
the
rfc
and
then
obviously
the
tracking
from
the
tracking
issue,
is
also
really
important.
Key
piece
to
unify
it
all
into
a
single
place
got
it.
Thank
you
very
much.
D
I
overall,
I
agree
with
what
was
already
said.
I
think
a
lot
of
stuff
looks
good.
I
would
encourage
people
outside
of
those
of
us
who
work
closely
together
every
day
to,
if
not
in
this
session
kind
of,
provide
feedback
we're
trying
to
improve
a
lot
of
the
planning
process.
You
know
in
in
the
meta
for
the
project
to
give
some
more
transparency.
D
That's
like
one
of
my
and
a
few
other
people's
goals
this
year.
So
if
people
from
other
organizations
kind
of
have
feedback,
we
really
appreciate
it.
You
know
people
are,
are
scared
to
speak
up
right
now
or
don't
have
thoughts
yet
feel
free
to
asynchronously.
A
Yeah,
I
think
I
think
this
is
a
really
exciting
and
well
thought
out
process
for
to
to
kind
of
encourage
larger
community
collaboration
in
the
project,
and
I'm
I'm
really
excited
to
see
it
move
forward
and
how
it's
going
to
turn
out.
So
thank
you,
denise
for
all
of
the
work
that
you've
put
into
this.
I
think
it's
it's
going
to
be
an
important
and
lasting
contribution
to
the
tvm
community.
A
Okay,
well,
once
again,
if
you
have
any
more
questions,
please
please
reach
out
either
on
the
discuss
forum
or
in
the
the
discourse
channel.
You
know,
I
think
that
we're
we're
open
to
a
lot
of
feedback
on
this
and
denise
is
gonna,
take
questions
and,
and-
and
you
know,
suggestions
from
the
community
seriously.
So
we
want
this
to
be
like
a
huge
community
process
now,
along
the
lines
of
of
the
road
map
and
and
what's
happening
with
tvm.
A
We
haven't
had
a
major
release
of
tvm
in
about
in
about
a
year
now
and
and
we're
working
towards
a
new
0.8
release,
which
is
hopefully
going
to
lead
into
a
1.0
release
soon,
and
so
to
talk
about
that,
a
junior
has
prepared
some
slides
about
what's
going
into
the
0.8
release
process
and
and
what
we
can
expect
there
and
going
forward.
C
C
It's
a
so
some
background,
so
the
previous
release
was
last
year
in
october,
so
it's
been
like
a
year
that
we
haven't
released
anything
yet,
and
the
0.80
is
exactly
really
released
like
in
october
this
year,
and
the
goal
of
this
release
is
to
first
introduce
many
useful
major
features
as
experimental
features
and
stabilization
of
previously
experimental
features.
For
example
like
auto,
auto
scheduler
things
like
that,
and
the
release
process
is
pretty
simple.
First,
we
do
release
planning,
so
we
want
to.
C
We
had
a
thread
to
discuss
with
the
community
about
like
what
kind
of
feature
do
we
want
to
include,
for
example,
we
might
want
to
include
some
mascara
features
that
that
is
not
yet
upstream,
but
will
be
up
soon
really
soon.
So
anyone
who
is,
who
is
who
is
willing
to
get
a
specific
commit
included,
can
comment
on
that
redis
planning
threat
and
to
like
make
sure
that
the
committee
is
included,
and
we
create
proposals
for
highlight,
highlights
in
the
release.
C
For
example,
we
we
we
need
to
like
go
over.
Those
commits
in
the
previous
year
and
find
what
are
the
major
changes
in
release
and
decide
the
release
cut
date.
That
is
the
first
stage
and
on
the
second
stage
after
we
kind
of
really
release
candidate,
it
will
produce
a
like
tarball
as
a
artifact
of
the
release
and
the
turbo
will
be
tested
by
pmc
and
the
community.
C
If
there
is
any
any
problem
with
that
hardware
file,
we
cut
another,
we
fix
the
issue
and
cut
another
release
candidate
and
until
there
is
no
problem
with
a
release,
artifact
and
also
we
draft
through
these
notes,
with
the
highlights
proposed
by
the
community
with
the
changes
and
the
last
last
stage
is
just
to
ask
the
pmc
members
to
sign
off
the
release.
C
So
our
expectation
is,
we
cut
the
release
by
the
we
get
released
down
by
the
end
of
october,
but
we
are
happy
to
care
about
the
community's
opinions
on
what
kind
of
features
should
be
included?
What
kind
of
pr
we
need
to
wait
for
and
make
change,
adapt
new
changes
that
quickly.
C
Major
unique
new
features
are
cancel
r,
so
in
this
release
as
a
primary,
a
new
version
of
tensor
r
will
be
included,
is
schedulable,
loop-based
ir,
and
we
can
use
casing
our
schedule
to
write
new
operators
to
generate
hyphens
operators
and
methods,
part
of
the
meta
schedule
that
allows
us
to
do
like,
for
example,
of
matrix
modification
or
tuning
on
cpu,
so
metascale
uses
a
next-generation,
major
system
and
yes,
an
aot
executor,
and
in
the
v0.8
new
front
we
have
the
file
analysis
infra,
which
is
used
in
particular
models
as
well
as
tesla
r,
and
we
have
a
brand
new
welcome
back
end,
but
welcome
to
purpose
new
highlights
for
the
new
features
and
enhancements.
C
C
We
have
we
introduced
a
lot
of
new
features
in
micro,
tvm
and
arm
and
hexagon
backhand
and
operator
coverage,
method,
back-end
and
documentation
tutorials.
So
that
is
a
proposal
for
enhancements
and
also
we
have.
We
have
the
previous
features
stabilized.
The
major
feature
is
auto
scheduler.
It
was
an
experimental
feature
in
the
last
release
and
after
one
year
of
testing
and
bug
fixes,
we
get
is
finally
like
production,
ready
and
rust
binding.
C
Jerry
spent
a
lot
of
time
like
getting
rest
danny,
like
diamondback,
fixed
and
byoc,
for
tensor
rt
and
for
ezio
and
other
libraries
is
downloads,
and
we
got
a
lot
of
new
features
in
tvmc
as
well.
C
So
right
now,
pvm
v0.8
is
still
in
planning
it's
still
september,
so
we
have
almost
roughly
like
a
month.
So
please
share
any
proposals,
inputs
and
questions.
That's
it.
Thank
you
guys.
E
A
small
question
because
now,
obviously
from
0.7
to
0.9,
this
was
a
rather
long
time.
Did
you
ever
plan
to
have
something
like
sub
releases
like
0.71,
0.72.
C
I
think
it
is
a
great
question.
So,
yes,
it
is.
It's
been
like
almost
a
year
since
the
previous
release,
and
we
are
sort
of
we
are
sorting
out
those
commits
which
are
like
65
pages
of
commits
and
into
different
categories
like
in
the
future.
It
is
certainly
possible
to
release
minor
releases
like
every.
C
C
This
is
a
great
question
too,
so
over
the
past
year
our
contributors
really
got
a
nightly
condo
and
a
nightly
condo
release
on
tlc
pack
channel,
and
so,
if
you
guys
are
interested
in
like
trying
out
new
features
without
building
pvm,
you
can
kind
of
install
those
libraries
like
easily.
A
Yeah
and
and
with
regards
to
the
like
why
tlc
pack
versus
a
tvm
branded
release
tvm,
especially
if
you're,
if
you're
running
on
on
gpus,
depends
heavily
on
proprietary
drivers
and
and
and
so
releasing
open
source
software.
That's
packaged
with
proprietary
drivers.
All
can
always
lead
to
to
licensing
problems
and
and
other
issues
so
to
to
to
help
this
to
help
identify
the
software
that
is
has
has
been
linked.
Like
say,
in
video
drivers
we
use
the
tlc
pack,
which
is
a
which
is
a
almost
like
a
a
branch
of
tvm.
A
That's
that's
been
built
so
right
now
it
is
possible
to
install
both
releases
and
the
nightlys
of
of
tvm
through
tlc
pack
and
and
personally,
I
think
that
it's
important
like
that
that
we
have
like
the
tip
installable
releases
down
the
down
the
road,
but
I
I
could
also
see
that
there
there
could
be
a
potential
for
for
releases
that
include
things
only
like
this,
like
the
the
c
plus
plus
runtime
or
or
more
slim
binaries,
that
are
going
to
be
applicable.
A
F
And
definitely
same
for
a
micro
tvm,
you
know
you
would
want
something
even
lighter
separating
the
development
you
know
from
the
runtime.
F
I
have
another
question
about
onnx
support.
F
I
haven't
looked
at
the
latest
contribution,
but
I
see
as
I'm
looking
at
you
know
it
used
to
be
not
supporting
too
many
nodes.
There
were
a
lot
of
them.
You
know
that
we're
missing,
so
is
it
going
to
be
compliant
with,
let's
say,
offset
level
13,
including
also
some
extensions
from
microsoft
for
quantized
nodes?
D
H
That
is
active
work.
We
are
tracking
the
onyx
quantized
nodes,
we've
implemented,
probably
70
of
them.
Now
we
also
come
the
offset
13
tests
against
the
importer
in
ci
and
there
are,
with
a
few
pr's
they're
currently
up
there.
Maybe
a
hundred
unit
tests
out
of
about
a
thousand
that
onyx
ships
that
we're
still
failing
and
we're
actively
working
on
improving
that
support.
So
if
you
want
to
jump
in
and
add
any
hops
we'd
love
to
help.
F
D
Thank
you
just
to
chime
in
on
that.
I
think
those
guys
have
some
tracking
issues
open.
So
if
they're,
you
know,
if
you
want
to
synchronize
some
efforts,
if
there
are
some
critical
ones
you
want,
you
know,
feel
free
to
reach
out
to
like
matt
or
andrew
and
there's
a
couple.
Other
people
who've
touched
him
at
optimal.
Happy
help.
I
think.
F
Right
yeah,
the
the
use
case
that
I
was
looking
in
particular
was
you
know
when
you
use
onnx
runtime
and
you
got
a
generation
of
the
quantized
onnx.
Can
we
import
directly
this
graph
inside
tvm
for
further
optimization?
F
D
D
Just
to
chime
in
and
one
last
thing,
I
think,
that's
kind
of
the
goal
of
what
we're
doing
with
denise
just
to
tie
everything
back
together
is
that
that
that
that
that
information
can
be
more
self-serve
for
everyone,
and
people
can
understand
and
help
vote
on
the
priorities
and
kind
of
build
that
map
based
on
what
the
community
needs,
because
I
think
sometimes
these
days,
it
requires
a
lot
of
talking
to
people
to
get
the
information
out
of
them,
and
ideally
it's
much
easier
to
figure
out
where
everything
is
at.
A
D
D
D
You
know
I
I
was
around
when
we
were
working
on
ross
and
we
went
1.0
and
a
lot
of
that
process
for
the
first
six.
Eight
nine
months
beforehand
was
just
really
triaging
and
figuring
out
what
needed
to
go
in
and
then
maybe
holding
back
stuff
or
marketing.
Things
is
experimental
and
figuring
out
both
the
backwards
compatibility
stories
and
stuff
there,
and
obviously
I
don't
think
we
have
to
be
as
rigid
as
they
are.
They've
got
really
really
strong
backwards
compatibility
guarantees.
D
I
think
that
would
be
the
kind
of
thing
that
we
would
be
like
hey.
We
need
this
from
a
1.0
release.
I
see
each
other
smiling.
I
think
that
would
be
the
kind
of
thing
that
we
would
want
people
to
kind
of
vote
on
in
the
sense
of
put
their.
D
You
know,
put
their
hat
in
the
ring
and
say
this
is
the
thing
that's
really
important
for
me
for
the
project
and
kind
of
understand
those
constraints,
because
I
I
know
at
every
org
who's
using
tvm,
the
constraints
are
different,
especially
in
different
environments
say
the
people
working
on
cloud
gpu
might
have
very
different
requirements
when
people
work
on
their
own
accelerators
or
the
micro
tvm
sort
of
embedded
folks,
and
so
I
think,
just
understanding
the
the
grand
set
of
all
those
requirements
from
different
folks
and
then
figuring
out,
which
ones
we
can
I
see
at
albert
and
sending
even
more
messages.
A
Okay,
so
moving
on
to
the
last
bit
of
the
meeting,
we
also
when,
when
we
have
time
available,
we
like
to
open
up
the
floor.
If
there
are
any
other
last-minute
agenda
items
that
people
would
like
to
to
to
add
or
ask
questions
about,
so
is
there?
Is
there
anything
that
that
folks
have
been
curious
about
and
would
like
to
take
this
opportunity
to
kind
of
have
a
discussion
with
the
larger
community
about.
A
Okay,
I'll
take
the
silence
as
as
as
a
no,
but
I
want
to
say
thank
you
again
for
everybody
for
joining.
Thank
you
to
denise
and
junior
for
presenting
these
are
incredibly
important
topics
and
I'm
really
excited
to
see
how
they're
moving
forward.
G
Hi
hi
chris
sorry,
sorry
for
missing
that
so
hi
this
is
gaurav.
I
have
a
very
generic
question
and
this
is
regarding
past
infrastructure.
If
I
may
ask
so.
G
Yeah,
so
I
went
through
the
documentation
and
what
to
the
code
base
also,
so
what
pass
manager
pass
infrastructure
has
to
offer.
So
currently,
it
is
only
having
timing
thing
time,
profiling
for
the
individual
passes,
but
if
I
want
to
test
out
different
passes
and
combinations
like
what
we
used
to
have,
what
we
have
in
user,
gcc
and
llbm
like
that,
so
are
we
planning
to
extend
it
for
the
memory
profiling?
Also,
I,
like
I'm
kind
of
interested
in
this
area,
so
I'm
just
asking.
D
So,
just
to
clarify
your
question:
we're
talking
about
memory
of
the
passes
or
memory
of
the
generated
code.
G
D
D
I
I
one
of
the
things
that
tristan
did
earlier
this
year
is
try
to
unify
a
lot
of
the
timing
profiling
code
because
before
it
was
kind
of
per
target-
and
there
was
less
sort
of
gen
like
generic
unified
support,
like
the
graph
runtime
profiler,
I
think,
gives
some
estimates,
but
I
think
improving
that
kind
of
stuff
and
and
also
improving
it
for
the
generated
kernels
would
be
something
that's
definitely
worth
doing.
And
if
you
have
ideas,
I
think
everyone's
happy
to
hear
them.
H
Yeah,
so
I
think
a
lot
of
people
would
like
to
get
like
memory
usage
information
from
the
from
the
colonels
in
profiling,
but
it's
a
little
tough
to
do,
especially
in
a
cross-platform
way.
I've
been
looking
at
various
approaches,
but
it's
not
really
on
my
roadmap
and
I
I
haven't
found
a
like
really
good
good
way
to
do.
It.
G
Okay,
thanks,
like
so
same
here
like
even
I'm
interested
kind
of
in
this,
but
again
not
having
much
insight.
How
can
I
do
that?
So
he
is.
That
is
one
thing.
Apart
from
that
to
measure
say
power,
consumption
and
all
those
things
we
prefer
to
use
any
external
profiling
tool
right.
Am
I
correct
on
that.
H
You
can
use
the
internal
profiling
tool
that
we
have
with
the
built-in
performance
counters
and
you
might
be
able
to
get
some
good
energy
information
there
in
general.
If
you've
got
more
questions
about
this,
you
can
ask
me
on
the
discuss
forum
or.
D
D
Yeah,
I
think
the
thing
is,
if
you,
if
one
knows
how
to
get
the
profiling
data
they
care
about
for
their
platform,
I
think
we're
happy
to
advise
on
how
to
hook
that
in
in
a
way
that
will
be
generic
with
a
standard
api
for
everyone,
and
then
I
think
iterate,
because
I
think
that's
where
tris
and
I
were
last
time
we
were
looking
at
this
stuff
was
just
you
know.
Every
gpu
is
slightly
different.
You
know,
every
accelerator
is
different.
D
You
know,
even
cpu
stories
sometimes
are
different,
so
I
think
you
know
understanding
exactly
what
the
best
way
to
do.
It
is
another
thing
that
might
be
useful
is,
if
you
find
yourself
using
external
tools,
as
maybe
contributing
back,
even
whatever,
for
doing
this,
with
tvm
generated
code
or
understanding
if
there
are
weaknesses
or
something
that
we
need
to
help
address
like.
D
I
know,
there's
some
like
missing
debug
information
that
someone
at
facebook
tried
to
add
at
one
point-
and
I
don't
remember
ever
got
emerged
in
but
like
doing
things
like
that,
are
also
appreciated
and
welcome.
The
community.
G
D
So
there
is
arm64
support.
I
guess,
are
you
asking
about
like
just
optimizing
or
increasing.
D
I
don't
know
if
there's
anyone
explicitly
working
on
it
from
octomell
right
now.
I
don't
the
arm.
Folks
are
doing
a
lot
of
stuff,
but
it's
mostly
ethos.
U
focus,
and
maybe
someone
else
has
more
data
on
this.
I
mean
one
thing:
is
it
all
the
tools
there
to
start
iterating
on
it
and
just
trying
it
out
and
seeing
what
weaknesses
are
there
are
there
in
the
terms
of
the
code
generation,
at
least
as
a
one
year
ago,
compared
to
the
competing
solutions
like
the
arm
performance
and
tvm
was
very
good.
D
If
you
auto
tune
like
we
found
that
other
libraries
were
not
very
robust
in
terms
of
amount
of
coverage
and
then
also
performance
for
the
for
varying
workloads.
So
like,
at
least
when
we
wrote
up
the
vm
paper,
which
I
we
again,
we
can
talk
about
in
the
discuss
forum.
We
had
very
good
arm
results
on
like
on
a
few
on
a
few
important
models
like
bert
and
a
few
other
things,
so
I
wouldn't
expect
you
to
have
much
difficulty
there.
D
A
G
So
I
posted
one
question
yesterday
and
like
today
early
morning
I
was
able
to
fix
it.
So
basically
relay
was
giving
me
segmentation
fault
and
then
I
could
fix
it
because
so
for
fujitsu.
Basically
they
provide
custom,
co-design
tensorflow,
and
that
was
having
some
environment
issues
internally.
So
yeah
it
was
a
tough
bit.
I
could
fix
it
but
later
on,
so
I
will
close
that
thing
on
this
course,
but
yeah.
These
are
the
few
things
that
I'm
facing.
G
D
I
think
in
that
case,
I
think
the
discussion
forum
is
probably
the
best
unless
you
isolate
it
and
then
we're
happy
to
you
know,
answer
questions
and
work
with
you
in
terms
of
getting
it
working
on
arm.
I
mean
people
are
running
arm
workloads
at
octomell
like
every
single
day,
so.
G
A
Okay,
well,
thank
you
again,
everybody
and
we'll
see
you
in
the
discussion
forums
on
the
on
the
on
the
discourse
board
and
at
the
meeting
next
week.
So
everyone
have
a
have
a
great
day
and
we'll
see
you
soon,
thanks.