add ttnn parser and ttnn ops md files [#26] #35
Conversation
|
Sorry I'm not writing Input and output shapes correctly. Will fix this |
ddilbazTT
force-pushed
the
ddilbaz/parse_ttnn
branch
from
394ca68 to
301b3d3
Compare
ddilbazTT
force-pushed
the
ddilbaz/parse_ttnn
branch
from
24a88cf to
17a6d39
Compare
There was a problem hiding this comment.
Nice job @ddilbazTT!
@AleksKnezevic @nsmithtt it'll be good to think about have similar type of logic on TTNN level (directly in MLIR).
More precisely, from the Forge-FE use case, this approach is okay if ops are mapped 1:1. However, that often isn't the case as many ops will be decomposed into multiple ones. Therefore, mapping our ops to single TTNN ops will not be feasible.
However, if we have something on the TTNN level it'll be a great help for each frontend. More precisely, when a model is compiled through MLIR, we're getting TTNN graphs (flatbuffer binary). If we have the mechanism to split that TTNN into unique ops, and run each of them separately and generate these markdowns, we'll have a solution for each frontend down the like. Therefore, if you have time please think about this approach as well :))
That said, from Forge-FE perspective I still need e2e approach comingfrom Forge-FE, as I can hit many issues before we even generate TTIR and then TTNN. So my main priority is to stabilize our model analysis that will focus on framework ops, rather than pure TTNN ops. However, if anyone from Forge team is available, it'll be good to explore this approach as well in parallel while we're pushing custom solutions through each frontend.
Let me know your thoughts.
No description provided.