Steven S. Lyubomirsky

icon indicating a website link https://homes.cs.washington.edu/~sslyu/ icon indicating an email [email protected]

icon company @NVIDIA icon location New York City icon location Senior Software Engineer, @NVIDIA. Alumnus of @uwplse and @uwsampl. Previously at @octoml

Results 113 comments of Steven S. Lyubomirsky

[IR] First-class StructInfo

I'd like to clarify something about the examples in the discussion section: Do we want to permit Relax variables to appear in shape annotations? I wrote the draft spec on...

[IR] First-class StructInfo

Okay, we might need to define how that should work.

[IR] First-class StructInfo

Since I won't be able to be at the community meeting tomorrow, I'll give some of my thoughts in advance in writing (would be happy to discuss further based on...

[Op] Implement several high-priority tensor ops: dense, conv2d, max_pool2d, relu, softmax, flatten

For testing, there is an easy route available for gathering more test cases: You can port over some of the Relay tests for these ops, of which there are quite...

[Op] Implement several high-priority tensor ops: dense, conv2d, max_pool2d, relu, softmax, flatten

>it seems like Relay does not have comprehensive op testing neither? Seems pretty much similar with what this PR has. It's not perfect, but it's a start haha. Not every...

[Bug][VM] Memory corruption over RPC when simply returning a function's argument

Wow! edit: Maybe it's somehow related to object lifetimes in Python? Maybe having the NDArray inline can result in its being freed somehow? Certainly seems like a bug.

[UX][IR] Cleanup global_symbol attachment and linkage

Could you please provide more context for the advantage specifying privacy would provide? How would it manifest in compilation?

[UX][IR] Cleanup global_symbol attachment and linkage

Thank you for the explanation. To my understanding, this is mainly a matter of the external interface for the `IRModule` and I can appreciate the utility of marking functions as...

[Discuss][Parsing][IR] Parsing `_` shape dimensions as `tir::Any`

That's an interesting example. I would argue that if it's important for two variables to match, they should probably have explicit names--a wildcard signifies that the dimension shouldn't be matched...

[Discuss][Parsing][IR] Parsing `_` shape dimensions as `tir::Any`

You could prove that the shapes are equal by adding a `MatchShape` binding, but that may be unnecessary complexity. I think the simplicity of using fresh vars is a good...