"1001 Ways to Write CUDA Kernels in Python"
https://www.youtube.com/watch?v=_XW6Yu6VBQE
"The CUDA Python Developer’s Toolbox"
https://www.nvidia.com/en-us/on-demand/session/gtc25-S72448/
"Accelerated Python: The Community and Ecosystem"
https://www.youtube.com/watch?v=6IcvKPfNXUw
"Tensor Core Programming in Python with CUTLASS 4.0"
https://www.linkedin.com/posts/nvidia-ai_python-cutlass-acti...
There is also Julia, as the black swan many outside Python community have moved into, with much more mature tooling, and first tier Windows support, for those researchers that for whatever reason have Windows issued work laptops.
https://info.juliahub.com/industries/case-studies
Mojo as programming language seems interesting as language nerd, but I think the judge is still out there if this is going to be another Swift, or Swift for Tensorflow, in regards to market adoption, given the existing contenders.
So going after people who need to build low latency high-throughput inference systems.
Also as someone else pointed out, they also target all kinds of hardware, not just NVidia.
The primary advantage of mojo seems to be Gil-free syntax that is as close to Python as possible.
They also miss CPUs on Windows, unless using WSL.
This may appeal to people wanting to run their code on different hardware brands fro various reasons.
And for those that care, Julia is available today on different hardware brands, as there are other Python DSL JITs as well.
I agree they will get there, now the question is will they get there fast enough to matter, versus what the mainstream market cares about.
The competitor is Triton, not CUDA or Julia...
He also discussed open sourcing Mojo and where the company expects to make its money.
FAQ:
> Why not make Julia better? > We think Julia is a great language and it has a wonderful community, but Mojo is completely different. While Julia and Mojo might share some goals and look similar as an easy-to-use and high-performance alternative to Python, we’re taking a completely different approach to building Mojo. Notably, Mojo is Python-first and doesn't require existing Python developers to learn a new syntax.
https://docs.modular.com/mojo/faq/#why-not-make-julia-better
Now :
>We oversold Mojo as a Python superset too early and realized that we should focus on what Mojo can do for people TODAY, not what it will grow into. As such, we currently explain Mojo as a language that's great for making stuff go fast on CPUs and GPUs.
Here is a link to the episode listing for that podcast, which might help.
I'm logged into youtube, if you're not something with that perhaps?
I liked mojo as a python superset. Wanted to be able to run arbitrary python through it and selectively change parts to use the new stuff.
A "pythonic language" sounds like that goal has been dropped, at which point the value prop is much less clear to me.
Yep, that's right. Int behaving like a machine integer is very important for systems performance. Leaving the "int" namespace untouched allows us to have a object-based bigint in the future for compatibility with python. As others have mentioned above, it is still a goal to be compatible with python in time, just not a short term priority
It was highly aspirational goal, and practically speaking it's better right now to take inspiration from Python and have stronger integration hooks into the language (full disclosure, I work at Modular). We've specifically stopped using the "superset of Python" language to be more accurate about what the language is meant for right now.
Embrace. Extend. Extinguish.
I'm just nervous how much VC funding they've raised and what kind of impacts that could have on their business model as they mature.
https://docs.modular.com/mojo/manual/get-started/
curl -fsSL https://pixi.sh/install.sh | sh
pixi init life \
-c https://conda.modular.com/max-nightly/ -c conda-forge \
&& cd life
We've run a few events this year (disclosing again that I work for Modular, to put my comments into context), and we've had some great feedback from people who have never done GPU programming about how easy it was to get started using Mojo.
C++ is a powerful and mature language, but it also has a steep learning curve. There's a lot of space for making GPU (and other high performance computing) easier, and platforms/languages like Triton and Julia are also exploring the space alongside Mojo. There's a huge opportunity to make GPU programming easier, and a bit part of that opportunity is in abstracting away as much of the device-specific coding as you can.
I was a HPC C++ programmer for a long time, and I always found recompiling for new devices to be one of the most painful things about working in C++ (for example, the often necessary nightmare of cmake). Mojo offers a lot of affordances that improve on the programming experience at the language, compiler, and runtime levels.
It's unclear to me from this comment, what exactly your pushback is.
I expect that opinion will not be universally agreed upon.
The space that I am interested in is execution time compiled programs. A usecase of this is to generate a perfect hash data structure. Say you have a config file that lists out the keywords that you want to find, and then dynamically generate the perfect hash data structure compiled as if those keywords are compile time values (because they are).
Or, if the number of keywords is too small, fallback to a linear search method. All done in compile time without the cost of dynamic dispatch.
Of course, I am talking about numba. But I think it is cursed by the fact that the host language is Python. Imagine if Python is stronger typed, it would open up a whole new scale of optimization.
Sadly the contenders on the corner get largely ignored, so we need to contend with special cased JIT DSLs, or writing native extensions, as in many cases CPython is only implementation that is available.
Jitting like you mentioned is supported by the MAX graph API: https://docs.modular.com/max/tutorials/build-custom-ops. It could have a nicer syntax though to be more like Numba, I think you have an interesting idea there.
> The space that I am interested in is execution time compiled programs. A usecase of this is to generate a perfect hash data structure. Say you have a config file that lists out the keywords that you want to find, and then dynamically generate the perfect hash data structure compiled as if those keywords are compile time values (because they are).
I'm not sure I understand you correctly, but these two seem connected. If I were to do what you want to do here in Python I'd create a zig build-lib and use it with ctypes.
``` python program.py --config <change this> ```
It is basically a recompilation of the whole program at every execution taking into account the config/machine combination.
So if the config contains no keyword for lookup, then the program should be able to be compiled into a noop. Or if the config contains keyword that permits a simple perfect hash algorithm, then it should recompile itself to use that mechanism.
I dont think any of the typical systems programming allows this.
Is there a zero-copy interface for larger objects? How do object lifetimes work in that case? Especially if this is to be used for ML, you need to haul over huge matrices. And the GIL stuff is also a thing.
I wonder how Mojo handles all that.
That said, importing into Python this easily is a pretty big deal. I can see a lot of teams who just want to get unblocked by some performance thing, finding this insanely helpful!
This is not really true. Even though Mojo is adopting Python's syntax, it is a drastically different language under the hood. Mojo is innovating in many directions (eg: mlir integration, ownership model, comptime, etc). The creators didn't feel the need to innovate on syntax in addition to all that.
I didn't mean to undermine the ambitious goals the project has. I still wish it was a little bolder on syntax though, Python is a large and complex language as is, so a superset of Python is inherently going to be a very complicated language.
So, the message is that it is possible to create nice Python bindings from Mojo code, but only if your Mojo code makes the effort to create an interface that uses PythonObject.
Useful, but I don’t see how that’s different from C code coding the same, as bindings go.
Both make it easier to gradually move Python code over to a compiled language.
Mojo presumably will have the advantage that porting from Python to Mojo is much closer to a copy paste job than porting Python to C is.
In all fairness, their website now reads: "Mojo is a pythonic language for blazing-fast CPU+GPU execution without CUDA. Optionally use it with MAX for insanely fast AI inference."
So I suppose now is just a compiled language with superficially similar syntax and completely different semantics to Python?
That said, the upside is huge. If they can get to a point where Python programmers that need to add speed learn Mojo, because it feels more familiar and interops more easily, rather than C/CPP that would be huge. And it's a much lower bar than superset of python.
I'd argue that I am not sure what kind of Python programmer is capable of learning things like comptime, borrow checking, generics but would struggle with different looking syntax. So to me this seemed like a deliberate misrepresentation of the actual challenges to generate hype and marketing.
Which fair enough, I suppose this is how things work. But it should be _fair_ to point out the obvious too.
What else was proclaimed just to raise VC money?
I feel like that depends quite a lot on what exactly is in the non-subset part of the language. Being able to use a library from the superset in the subset requires being able to translate the features into something that can run in the subset, so if the superset is doing a lot of interesting things at runtime, that isn't necessarily going to be trivial.
(I have no idea exactly what features Mojo provides beyond what's already in Python, so maybe it's not much of an achievement in this case, but my point is that this has less to do with just being a superset but about what exactly the extra stuff is, so I'm not sure I buy the argument that the marketing you mention of enough to conclude that this isn't much of an achievement.)
And so his improvements in mojo and now calling mojo code from python just make a lot more net positive to the community than being, some other Ai infrastructure company.
So I do wish a lot of good luck to mojo. I have heard that mojo isn't open source but it has plans to do so. I'd like to try it once if its as fast / even a little slower than rust and comparable to understanding as python.
> Further, we decided that the right long-term goal for Mojo is to adopt the syntax of Python (that is, to make Mojo compatible with existing Python programs) and to embrace the CPython implementation for long-tail ecosystem support
Which I don't think has changed.
’’’ 3628800 Time taken: 3.0279159545898438e-05 seconds for mojo
3628800 Time taken: 5.0067901611328125e-06 seconds for python ’’’
[1] https://github.com/python/cpython/blob/0d9d48959e050b66cb37a...
[2] https://github.com/python/cpython/blob/0d9d48959e050b66cb37a...
But asking people to learn another language and migrate is a tall barrier.
Why not write python and transpile to mojo?
As part of this it has a stronger type/lifecycle/memory model than python.
Maybe you could write, some level of transpiler, but so much of the optimizations rely on things that python does not expose (types), and there are things that python can do that are not supported.
Py2many has had a mojo backend for a year now.
If they can make calling Mojo from Python smooth it would be a great replacement for Cython. You also then get easy access to your GPU etc.
I’m always disappointed when I hear anything about mojo. I can’t even fully use it, to say nothing of examine it.
We all need money, and like to have our incubators, but the LLVM guy thinks like Jonathan Blow with jai?
I don’t see the benefit of joining an exclusive club to learn exclusively-useful things. That sounds more like a religion or MLM than anything RMS ever said :p
I would not compare Chris Lattner with Jonathan Blow. Lattner is a person with a reputation for delivering programming languages and infrastructure; whereas for Blow, it seems like an ego issue. He's built a nice little cult of personality around his unreleased language, and releasing it will kill a lot of that magic (running a language project with actual users that has to make good on promises is much different than running a language project with followers and acolytes that can promise anything and never deliver on it).
Lattner has a record of actually delivering dev products people can download and use. Mojo is closed source to make raising money easier, but at least you can actually use it. Jai isn't even available for people to use, and after a decade of Blow dangling it in front of people, it's not clear it'll ever be available, because I'm not sure he wants it to be available.
Lol wut. For the life of me I cannot fathom what design decision in their cconv/ABI leads to this.
add_function(PythonObject)
add_function(PythonObject, PythonObject)
add_function(PythonObject, PythonObject, PythonObject)
There was a similar pattern in the Guava library years ago, where ImmutableList.of(…) would only support up to 20 arguments because there were 20 different instances of the method for each possible argument count.
:)))
> As of February 2025, the Mojo compiler is closed source
And that's where the story begins and ends for me.
> Will Mojo be open-sourced?
> We have committed to open-sourcing Mojo in 2026. Mojo is still young, so we will continue to incubate it within Modular until more of its internal architecture is fleshed out.
Soon it might be though.
