TL/MLX5: a2a part 1 -- coll init

[samnordmann]

What

a2a part 1 -- coll init

[manjugv]

Depends on #784

[Sergei-Lebedev]

@samnordmann please rebase

[samnordmann]

@edgargabriel we are getting an error with Linter-ROCM
any idea ?
Example:

#warning "hip_version.h has moved to /opt/rocm-5.6.0/include/hip and package include paths have changed. Provide include path as /opt/rocm-5.6.0/include when using cmake packages."

[shimmybalsam]

@edgargabriel we are getting an error with Linter-ROCM any idea ? Example:

#warning "hip_version.h has moved to /opt/rocm-5.6.0/include/hip and package include paths have changed. Provide include path as /opt/rocm-5.6.0/include when using cmake packages."

Hi @edgargabriel I am getting this error in PR #801 as well.

[edgargabriel]

@edgargabriel we are getting an error with Linter-ROCM any idea ? Example:

#warning "hip_version.h has moved to /opt/rocm-5.6.0/include/hip and package include paths have changed. Provide include path as /opt/rocm-5.6.0/include when using cmake packages."

Hi @edgargabriel I am getting this error in PR #801 as well.

I will have a look

[edgargabriel]

not entirely sure why that is happening, ucc compiles for me with rocm 5.6 without issues, probably a minor difference in the environment. I will try to remove an include path that is there for historic reasons but is not really required anymore, I think that is what is causing the issue.

[edgargabriel]

I think this is the cause of the issue, the version check fails for whatever reason and because of that it adds in the include path to the deprecated directories (which causes the issue). This test doesn't fail on my test systems

checking if ROCm version is 5.0 or above... no