Expand from write nodes inside reductions #395
Conversation
GMNGeoffrey
force-pushed
the
leaf-expansion
branch
from
555ed06 to
0e646d2
Compare
GMNGeoffrey
changed the title
|
I actually signed the commits with my ssh key, but it looks like it's not accepting that. |
I'm surprised the vscode workspace wasn't already in there, as I think that's the typical place to put it. I think it's generally useful to have a place to put output files (and the MLIR dumps should probably go there by default). I can drop these though. Signed-off-by: Geoffrey Martin-Noble <[email protected]>
Right now, it only looks at write nodes in the root graph and misses ones in reductions. Instead we walk all the regions for `Write` nodes and `GetResult` nodes without users. `get_leaf_node_types` also appeared to always return `[Write]` as it sets `leaf_nodes` to that and then immediately checks if it's empty, which it never will be, so I deleted it. Signed-off-by: Geoffrey Martin-Noble <[email protected]>
GMNGeoffrey
force-pushed
the
leaf-expansion
branch
from
0e646d2 to
4b133fd
Compare
|
@harsh-nod PTAL |
DCO allows signing (instead of sign-off) for organization members. I've sent you an invite to iree-org + the iree-triage team. That will also let checks run without approval. IREE docs: https://iree.dev/developers/general/contributing/#developer-certificate-of-origin |
|
Added tests. PTAL |
Confirmed this fails on main and succeeds with my change.
The thing I was doing here actually failed because apparently _node_list isn't an iterator, so switched to looking at `_root.prev`, althoug I'm a bit leery of using a private member here.
GMNGeoffrey
force-pushed
the
leaf-expansion
branch
from
02881c4 to
c6532d6
Compare
|
The failing test doesn't seem related to my change and passes on my local mi210. It's also a numerical diff on a check that already has rtol set to 0.8, which suggests to me that it's already broken and should probably be xfailed or skipped... |
Yes we have flaky attention tests that need to be debugged, but you can just retrigger for now. |
Right now, it only looks at write nodes in the root graph and misses ones in reductions. Instead we walk all the regions for
Writenodes andGetResultnodes without users.get_leaf_node_typesalso appeared to always return[Write]as it setsleaf_nodesto that and then immediately checks if it's empty, which it never will be, so I deleted it.Also includes some .gitignore additions that I can drop if you don't like them.
Fixes #374