This is actually really neat, I had an agent create some experimental probes and check out whats going on here and it seems like a pretty cool project. I'm gonna need to digest how this might be useful but I think this could be cool for some sort of FPGA uses perhaps.
> this is not going to beat XLA for standard deep learning workloads. XLA has years of hyper-specific optimizations for linear algebra on GPUs and TPUs.
I'm not so convinced about this.
It's actually really easy to write two mathematically equivalent formulations in Python of something basic that have over a 2x performance difference in them after jax jit.
XLA is not that smart. And I'm not talking some niche nonsense I mean simple matrix multiplication graphs and residual connections on the CUDA backend.
Can you figure out that 16 matrix vector multiplications into a concatenate is the same as concatenating first into a matrix matrix operation which can go on the GEMM. This is like the most basic thing you can imagine doing.
Turns out no it doesn't and there is a 23% performance difference by moving the concatenate up in the python code in my specific test.
Not to say that it didn't recognize it. I had an agent look at the XLA and the graph actually does a partial fusion into a sum and stack. But does not realize the whole thing is just a GEMM.
In more complicated examples the differences you can get can be much larger.
Neat. XLA predates MLIR. Interesting stories there. You'd have to stop by the Bay Area LLVM monthly meetup to hear them. :-X
> So if XLA already uses LLVM, why is our approach different?
Uses MLIR, XLA does not.
> So… what is the point?
> Honestly? We aren’t entirely sure yet.
> Let me be perfectly clear: this is not going to beat XLA for standard deep learning workloads. XLA has years of hyper-specific optimizations for linear algebra on GPUs and TPUs. If you are training a massive transformer, stick to standard JAX.
> But what we do think is cool is what happens when you connect JAX directly to the broader LLVM ecosystem and drop the heavy XLA runtime. (Plus, no need to build XLA using Bazel either! You’re welcome.)
Indeed, I am an outgroup member. look at me the dumbfuck. Now how does one become an ingroup member, by learning.
I am an active contributer to MLIR, if you knew my name and you weren't some weirdo, you could know me. This is a pseudonym account of which I have several, since I grew up in an age where you hid your identity. Now how XLA interacts with that I have no clue dude. Maybe help with that and don't be a low effort dick.
brother i'm trying to tell you that the guy you're asking to be clearer is being vague because he's employing a shibboleth - a "wink wink i'm an og and you can only access me via the secret handshake at our club meeting". not you. i pasted the link because you explicitly say "I can't tell what you mean by this comment" so i'm proposing what they mean.
> this is not going to beat XLA for standard deep learning workloads. XLA has years of hyper-specific optimizations for linear algebra on GPUs and TPUs.
I'm not so convinced about this.
It's actually really easy to write two mathematically equivalent formulations in Python of something basic that have over a 2x performance difference in them after jax jit.
XLA is not that smart. And I'm not talking some niche nonsense I mean simple matrix multiplication graphs and residual connections on the CUDA backend.