steeve

This one does, we use it with great success: https://github.com/cerisier/toolchains_llvm_bootstrapped

Yes: https://github.com/cerisier/toolchains_llvm_bootstrapped

This sandboxed LLVM toolchain completely solves that problem: https://github.com/cerisier/toolchains_llvm_bootstrapped

The downvote are actually proving my point

Hard to put into words how those people act like the devil incarnate and a lot of the scene pretends that this doesn't matter.

Like what?

You can also directly call LLVM intrinsics in case this doesn’t work

According to https://huggingface.co/openai/gpt-oss-120b/blob/main/config....

That’s 2880 values (so multiply by dtype)

This is USSR / North Korea stuff

Because of who ?

It absolutely is and we (ZML) are using it with great success. That said, Andrew said he would absolutely would break compat if it meant things go in the right direction. Yes, it can be painful sometimes, yes I do not always agree with his choices, but it has never been a blocker nor a significant time sink.

And in the end, things do improve significantly.

In this case, I think the new IO stuff is incredible.

Musk doing Musk things

Very cool that this runs as a MCP server, very cool demo

No, just because you’re hashing every ms doesn’t mean we all do.

Actually, _two_ fighter jets accidentally dropped bombs. One is bad enough, but two is worrying.

Sir this is not X

Impeccable job guys. Move fast, break things, nuclear edition.

That has absolutely nothing to do with RAII or safety…

We (ZML) have AMD MI300X working just fine, in fact, faster than H100

Look at what we (ZML) are doing: https://github.com/zml/zml

Any model, any hardware, zero compromise. TT support is on the roadmap and we talk to them.

I mean, we (zml) clocked MI300X ($20k) at +30% than H100 ($30k).

So…

We do use ROCm and CUDA. Only we sandbox it with the model and download only the needed parts which are about 1/10th of the size.

You can look us up at https://github.com/zml/zml, we fix that.

We are in line with ROCm 6.2 support. We actually just opened a PR to bump to 6.2.2: https://github.com/zml/zml/pull/39

Hi, we (ZML), fix that: https://github.com/zml/zml

We (ZML) measured MI300X at 30% faster than H100. These are great chips!

pretty easy, usually the hardest part is figuring out what the python code is doing

The last one is so very true

Bazel is amazing and doing C++ with anything other is like going back to the stone age.

The Bazel team has done an amazing job, the VM is embedded and trimmed. It’s as easy and download and run.

And worst case you can invest in Buck2.

Yes, that’s how it works (pipeline parallelism)

It’s outrage and bots.

Fuck Musk.