update all-reduce to enable basic n300 support

[SeanNijjar]

Ticket

Link to Github Issue

Problem description

N300 support wasn't enabled with all-gather so there were some basic issues with it.

What's changed

This commit resolves the basic issue which was that for n300, all-reduce was
a) trying to invoke a ring on an n300 (which is really a line)
b) using the reduce-scatter + all-gather implementation (which is buggy on line topology due to buggy line reduce scatter implementation) -> all-reduce will invoke the naive all-gather + local reduce implementation in this case

Note this is old CCL code that will be deprecated soon. I'm hoping to put as few man hours into it as possible.

Checklist

• Post commit CI: https://github.com/tenstorrent/tt-metal/actions/runs/12207721173
• T3k frequence, nightly, model perf: https://github.com/tenstorrent/tt-metal/actions/runs/12207718435
• Blackhole Post commit (if applicable)
• Model regression CI testing passes (if applicable)
• Device performance regression CI testing passes (if applicable)
• New/Existing tests provide coverage for changes

[jvegaTT]

Probably here because of debug but delete this before merging

[SeanNijjar]

@jvegaTT - can you please clarify your concern here? Is it specifically the comment or the entire conditional block? If the latter, then this change is required for correctness. I kept it intentionally separate from the later

    if (topology == ttnn::ccl::Topology::Linear) {
        // reduce scatter doesn't reliably support line topology yet
        optimized_version = false;
    }

because they are applying those modifications for different reasons:

This block in this comment is related to topology. If we are on a 2 device system, it implies line (atleast today - in the future we will need to generalize to make the op infra more thoroughly detect line vs ring be checking ethernet link connectivity or querying fabric).

The second block is temporarily there because reduce scatter isn't stable on line topology, and it's used for the optimized version of all-reduce, so we disable the optimized version if we are on a line.

[jvegaTT]

OK, it was just the // 2 devices == n300 == linear topology comment. But on second look I think that it is good to clarify this so let's keep it.

[jvegaTT]

I left a few very minor comments but it overall looks good so just accepting in advance.