Minutes_2020_09_15
stuartarchibald edited this page Sep 16, 2020
·
2 revisions
Attendees: Graham, Guilherme, Hameer, Stuart, Todd, Val, Siu, Juan
- extra repo, numba-extras/addons, how to implement practically
- Q. Should there be more "managers" of this repo? Especially for review and merge.
- A. Could add code-owners to help target review/maintenance/ownership.
- Q. How to get granularity of merge right? like per-repo control
- A. looks like there is per repo access
- Q. Code Safety concern
- (Hameer) it should be pure python
- (Stuart) concern security problem to buildfarm. Building on the cloud and mitigate that.
- Limitation of cloud CI is the lack of free GPU.
- Manual trigger of GPU test on buildfarm should be fine.
- Q. Concern of Cloud CI load
- A. perhaps try Github CI (likely base on Azure but w/ different queue)
- Q. Standard setup the same? CoC, README etc.
- A. Yes.
- Q. What is permitted?
- Pure python, no compiled code
- Just use of extension API.
- Dependencies. Keeps same as Numba. Optional runtime deps ok, hard deps not.
- (Siu) does this help bootstrap taking targets out of the numba core?
- The extras repo will teach us what are part of the extension API
- Day One readiness: fully plumbed CI, CoC, Contrib, Readme, infrastructure
- Name:
- (Hameer) "numba-unstable"
- (Graham) "numba-universe" / "numba-multiverse"
- Not to imply any category to the features
- (Stuart) "numba-addons" "numba-extras"
- (Val) "numba-experimental" "numba-contrib" "numba-boltons" "numba-xtra"
- (Hameer) "numba-all" vs "numba-extras" for e.g. pip
numba[all]
vsnumba[extras]
- Todd - likes boltons but happy with extras.
- Ask for feedback on the names (non-binding etc.)
- Stale issue bot (probot stale)
- (Hameer) stale-bot is negative for users. maybe a postive bot to balance things out, e.g. high-fives, extracting CI error messages, assigning reviewers,
- github discussions... can questions/discussions be moved there.
- Can stable bot label stale and then infinite (e.g. 3 month) timeout be added and a human (the triager) can then make an assessment? Helps remove the potentially negative experience for users?
- #6242 - Invalid incref pruning when passing meminfo to C++
- #6240 - AttributeError: module 'numba' has no attribute 'core'
-
#6238 - literally does not behave like in the doc example (and the test)
- fix docs
- #6237 - allow numba cuda vectorize to transmit scalar arguments as scalars, rather than creating temporary array
- #6236 - memory leak when using map()
-
#6235 - ctypes.Structure support for carray dtype in npm
- could work if
_ctypes.CField
were mapped cf. CFFI
- could work if
-
#6233 - CUDA:
mapped_array
accepts a stream but does nothing with it - #6232 - Incorrect behavior when mutating the captured variable in nopython mode
- #6228 - Some numpy broadcasting operations fail in numba 0.51
-
#6239 - Invalid use of getiter in tuple of arrays
- Do new error message talk about literal_unroll. May be make it implicit.
- #6230 - NUMBA - Building wheel for numba (setup.py) ... error
- #6226 - Confusing error message when cudatoolkit version is wrong
- #6223 - CUDA get_arch_option TypeError in formatting if CUDA install broken
- #6243 - Improve error message for NumPy alias type used as dtype in ArrayNdCtors
- #6241 - [Don't merge] Typing types/abstract.py
- #6234 - [WIP] Ensure DUfunc.ufunc objects are picklable
- #6231 - Remove compatibility mode
- #6225 - adding stale bot config file
- #6224 - Add support for tuple concatenation to array analysis. (#5396 continued)
- #6222 - Add type annotations to parts of numba/core
- #6229 - Test new python 38 builds.
- #6227 - Raise exception when no supported architectures are found
- Requests for 0.52
- Fast(er)
typed.List
/typed.Dict
? Doesn't have to be in 0.52, next 3-5 months is fine.
- Fast(er)