I haven't kept up with python development lately, so thanks for pointing that out. I looked this up and it does seem like some sort of property access is being optimized. What's not clear is what the hidden runtime type model is being used for this, or if it's a lightweight tracking that captures some subset of property accesses.
Python's object semantics are broadly similar to Javascript, bar the prototypical inheritance chain, but with the extra complexity of potentially needing to handle dynamic `__*attr__` overrides. Optimizing for this would require a tree of hidden map objects much like JS JIT engines use. I dug around the codebase a while back looking for something like that but didn't find it. `OrderedDict` seemed promising but after more investigation that turned out to just be adding property ordering to dictionaries.
The thing is, you can build that entire hidden type structure out, AND optimize against it, without ever having to mess with things like bytecode replacement. Once you have the hidden type model, you can modify the property lookup algorithm in the VM itself (as in cpython c code) to use some annotation structure hanging off of the bytecode to look up inline caches of type-guarded fastpaths. It would be an optimized fastpath that integrated directly into the C code that interprets the GETATTR opcode.
> Optimizing for this would require a tree of hidden map objects much like JS JIT engines use.
I don’t really think that’s true or even that relevant , at least in context of classes (especially with slots - the class, by definition, _is_ the “hidden map”) and not generic dicts.
> The thing is, you can build that entire hidden type structure out, AND optimize against it, without ever having to mess with things like bytecode replacement.
Pretty sure the motivation was exactly the opposite - that a simple “if a LOAD_ATTR encounters the same class with slots N times in a row, swap it out for LOAD_ATTR_SLOT with inline cached slot index, and swap back to generic one if something changes” is way easier (both to implement and reason about) than any alternative that’d interact with the object model.
(Also it means you skip having to branch between a slow and fast path at all)
> you can modify the property lookup algorithm in the VM itself (as in cpython c code) to use some annotation structure hanging off of the bytecode to look up inline caches of type-guarded fastpaths. It would be an optimized fastpath that integrated directly into the C code that interprets the GETATTR opcode.
Sorry, that sounds a bit vague and not too different from what CPython already is doing.
Ah, the `__slots__` feature makes it make a lot more sense.
> Sorry, that sounds a bit vague and not too different from what CPython already is doing.
With slots being available, I can see where I was misled. I went looking for a type structure that modeled the dynamic properties on an object which would be used as a typeguard, but the class itself functions as that if you're restricting the optimization to slotted properties.
(Note that this optimization can also trigger for plain classes (that’s LOAD_ATTR_INSTANCE_VALUE), it just needs a bit more guards and is only useful if you don’t abuse instances being dynamic objects. I just used slots as the most obvious example, but there are lots of different op specializations since like 3.11)
This has been very informative. With object property/method lookup optimized the next big leap in JIT-style perf is actually a hard pull. The next big perf win lies behind local type inference, and a method-jit with your standard bevy of SSA optimizations to hoist the typeguards.
I suppose the various JIT projects are trying to come at that from different angles incrementally, but that's a genuinely hard problem as you don't see real benefits until multiple complex pieces of infrastructure are in place and with good coverage over the language.
Unfortunate. I've been following the work on it occasionally.
On one hand, it seemed like JIT work was really progressing in high gear until SC requested a new PEP. Also unfortunate, as it seemed like part of the hope for the JIT was that it'd counter the performance penalty from free threading.
On the other, it did feel (to me) like with all the fancy (and definitely cool) optimizations it got (automatic refcount operation removal, new type-based specializations, inplace math operations), the macrobenchmark benefits weren't _that_ exciting for a JIT. It definitely has potential for more, maybe with even fancier optimizations like unboxing ints, but the complexity just keeps going up.
> It said that the PEP was "overly optimistic" about user-visible differences with the JIT enabled.
The JIT is technically generated from the same source as the interpreter loop itself, so originally the differences were extremely unlikely, but I imagine the risk just kept going up as new JIT-only paths and optimizations were introduced.
And politically, I feel like SC’s reasoning was just “this side project has landed in main branch almost by accident and has been stewing without oversight ever since, with no stabilization in sight”; and the latest PEP was probably still too vague, with a lot of “here’s more things we can add to the stew” - so I can understand if SC wasn’t satisfied.
I think this is a major issue with requiring a PEP for what's a performance optimization.
JITs are big nasty things that by their nature have to do a lot of things. You are basically saying "Give us a PEP for a new python interpreter and compiler".
What they do is fuzzy by their nature because optimizations are very rarely universally beneficial. It's a game of measuring, looking for regressions, and tackling the biggest problem next for whatever sample applications you have available. You always run into the risk of overfitting on top of all that (You may have made django faster, but you've made flask slower).
This is the sort of thing that, IMO, should not go through the PEP process. It's an implementation detail for the runtime.
Imagine, for example, if the JVM needed a bunch of JEPs for JIT improvements. They'd never get done. Or if any javascript engine needed ECMA acceptance before they could do a new JIT optimization.
Does not have the same level of requirements being asked of the Python JIT. Further, it's well understood that this JEP will "live on main" for quite some time in preview before it's accepted and merged.
What the JEP does, instead, is simply describe the feature and potential risks.
The vast majority of JVM optimizations do not get a JEP.
Fair, I don’t argue with that, I’m not knowledgeable about Java’s processes. Just one thing:
> The vast majority of JVM optimizations do not get a JEP.
In Python too. Still, the JIT really is just a really big thing. I don’t know if there could be a better way to do it, as either 1. You develop it off-branch and… make an enormous PR when you’re done, with risks of getting rejected anyway? Or 2, you allow more experiments on main like Rust, but then you end up with hundreds of minor half-maintained nightly-only features, possibly conflicting with each other, and eventually go through bureaucracy to remove them anyway.
So, I'd grant that it'd probably be a JEP for java (the initial JIT) but not that every change to the JIT would result in a new JEP. Java's JIT, hotspot, was integrated right before they started the JEP process. The JEP for the JIT would likely be pretty boring like "It's a JIT, it does this, in some applications we see this sort of speed up". Much like the AOT JEP https://openjdk.org/jeps/514
The Java governance allows for a lot more open ended projects with follow ups. Not that it's reckless, Valhalla is proof of that, but rather that once something is good enough they are pretty open to just shipping it as a preview feature and refining until it is worthy of being stable.
That's why, for example, we saw ZGC get merged, and then JDK after JDK seeing continuing development and improvement on ZGC (Or G1 for that matter). These GCs were fairly rough on their first release and now they are pretty much best in class.
I see a Python JIT as the same thing. Delivering the infrastructure for a JIT would need a PIP, but I wouldn't put in things like "acceptance criteria" beyond "doesn't break older code". I would expect such a project to get continual updates without PEPs.
> I would expect such a project to get continual updates without PEPs.
Just for completeness: the original JIT PR was merged quite enthusiastically
and only after the fact it was pointed out that it should have went through a PEP process [1]. The following PEP 744 was of the “informational” type, and the just-withdrawn PEP 836 was the first “standards track” PEP, which means that (in my understanding) it’s technically the first actual PEP that the SC would formally comment on. So in a way it’s not “can we update the JIT” PEP, but the first real “does SC want this JIT as proposed” PEP.
That said, I do agree with most of what you said, especially the acceptance criteria part.
The issue I have and find it hard to resolve is that since the code was already on main, the feature was already in “sunk cost” mode. It’s hard when the code is already on main and another 3 optimization PRs are always there behind the corner, but it needs to be possible for SC to say “we decided it’s not actually worth all this complexity for <some small number>% speed improvement” (I’m not commenting whether in this particular case it’d have been a right call), and the PEP (if not withdrawn) was the way to exercise that control.
Oh, it's not just a performance optimization. But you've said it yourself: "a new python interpreter and compiler". A lot of Python code out there lives with assumptions of CPython interpreter built into it. Pretty sure pytest would've been affected, should JIT replace the existing interpreter, for example.
Anyways. I don't believe that Python needs a JIT compiler. I don't believe that CPython can realistically replace its interpreter with JIT compiler without causing major problems with the community packages.
The way I see it: Python users today desperately want Java. They want it called Python because there's some vague memory still around of Python being cool and Java not being cool.
In Russian there's a children tale about a hungry soldier returning from a war and a greedy old woman who offered him a bed for the night. The old woman then pretends to have no food for dinner, and a soldier promises her a magic trick of making a porridge out of a hatchet. And so he tells the woman to put a pot on a stove with a hatchet in it, and step by step he makes her add all the ingredients necessary to make a porridge to the pot. The story gave birth to the idiom "hatchet porridge", which is used to describe anything that was created in a similar way: by deception, when another goal was allegedly being pursued.
So, to me, the current state of Python is this "hatchet porridge". None of its users want the actual Python (the hatchet is still sticking out of the pot), and they keep adding Java ingredients to the pot in hopes to cook up Java. And so was JIT. It just turned out to be too difficult to throw away the hatchet. An the porridge still has some chipped wood in it as well as a metallic taste.
Aren’t you confusing internals and API? I don’t think nobody cares about HOW the optimisation is done, but SC wanted API with the rest of the interpreter be defined and stable. Two different things. I think.
I am confused. Why can't they have an experimental, optionally-enabled feature for a few years while they iron out all the issues ? What is wrong with that ? The JDK keeps doing this all the while. They write a new JEP if there are changes across previews.
> On the other, it did feel (to me) like with all the fancy (and definitely cool) optimizations it got (automatic refcount operation removal, new type-based specializations, inplace math operations), the macrobenchmark benefits weren't _that_ exciting for a JIT. It definitely has potential for more, maybe with even fancier optimizations like unboxing ints, but the complexity just keeps going up.
Zero chance of shoehorning that into a standards-compliant Python runtime. Some things are too fundamental to change without affecting the user-visible language: JITting, lazy imports, threading, copying gc, ref counting, etc.
I mean, I have had great success with prototyping stuff in python, then letting an llm translate it to C. My 3090 with Gemma4:31b does this very well with a few guidelines.
Dunno how much investment is needed into language optimization anymore.
That works well for greenfield code, but if working on compiler optimisations has taught me anything, it's that there is a lot of legacy code out there in the world that would benefit from re-compilation with optimisations - but would be very difficult to re-write for speed (regardless of whether the work is done by a human, AI, or other machine tool). JIT compilation fits very nicely in this niche, as it means that existing Python projects can get (theoretically) better performance without re-writing.
You can use static type analysers and runtime type profilers to add type annotations to a legacy codebase, but the interesting bit would be if you're built on top of a large untyped codebase you don't control like, say, Django, how far it can go.
In my experiments with this, maybe, sorta, it'll do some AST stuff then it'll go "dang this shit is dynamic AF" then suggest a static type inference tool, then a runtime type profiler.
People have been calling Python a "prototyping" language for decades. I'd be interested to see how many Python projects ever actually get rewritten in a more performant language. My guess is, up until the last five years that was almost never done. Even with LLMs to help, is that the best use of your tokens and humans' time? It doesn't add a feature. It could conceivably cut costs after enough time, but only if done correctly.
> My guess is, up until the last five years that was almost never done.
Because having just the hot loop in the performant language is often deemed good enough. And there's tons of Python code out there where the "hot loop" is already automatically in C (the various operations implemented in C by the runtime) or where you can get close enough e.g. with Numpy. Consider that things like Pygame can exist; for a lot of the games people want to make, the only thing that really needs to happen anywhere near "bare-metal" speed is blitting.
Doing it currently, Python to Go. Early indications are a 50-98% increase in speed (e.g. one testcase went from 26hrs to 30mins) for network/CPU heavy workload. Conversion via Qwem, lots of testing needed but compute savings will be worth it.
Did you mean “50-98% decrease in running time”? Just calculating from your 26hrs to 30 mins seems to imply so. In terms of speed that translates to 100% to 5100% increase in speed.
Where you using standard optimized libraries (e.g., NumPy for numerical computation) or did you do everything in Python? This kind of increase of speed seems to indicate extremely sloppy Python programming.
It's the same level of testing for us for either Python2->3 or Go (it's a legacy code base, low automated coverage).
Conversion is now (somewhat) easier, the testing will be the slow part. For the speed improvements, which we also need for a migration from on-prem to cloud, it's a no brainer vs the other options at our disposal (like GraalPy).
The big reddit thread on/r/cpp might actually be the _most_ prominent public example of how political it is (and how the committee operates in general), with several people involved with WG21 arguing in the open.
A random number I checked feels kinda optimistic, or there's something wrong with my setup; Gemma 26B-A4B on a 16GB 5060Ti is reported to get 21-88tok/s with Q4_K_M, meanwhile I can just barely eke out 70tok/s with Q3 (which fully fits in vram), and 40tok/s with Q4 (with recently-merged MoE cache on llama-cpp). I guess the average fits, but the peak feels implausible.
No clue about details, but Sebastian Raschka's site reports its kv cache usage at 210kB per token, which is like 6x more than this site claims ("0.273 GB per 8k"). Also, the site claims qwen 27b to use 2x more memory per token than gemma, which is the opposite from what I know.
Maybe I'm missing something, but why couldn't it be generated by the model? In older classification tasks with transformers like BERT, you could absolutely obtain a confidence score.
That's the thing, though. Those modal offer popups have been in macOS and iOS for a really long time, and does cheapen the experience like Microsoft and Google products do.
> Ternary Bonsai 2 27B uses ternary {−1, 0, +1} weights with FP16 group-wise scaling, for 1.76 effective bits per weight
If I recall correctly, a recent post [1] has shown that Q2 quants (with like 2.6 bpw) of the same base Qwen model sit at the edge between "noticeably worse" and Q1's "useless". I took a quick glance at Bonsai's blog posts, and don't really see them comparing themselves to "typical" quants or explaining what's the special sauce that makes them better?
I wonder how well it performs in practice, because I can't help seriously doubting those benchmarks. That would put this 6GB model in Opus 4.6+ ballpark. Granted, that's mostly to Qwen 3.8's credit, but it's hard to believe that Qwen's already unbelievable capability density can still be compressed this much more.
I think the general idea is naive quantization falls apart below 4bpw but you can go lower with more sophisticated QAT-adjacent methods. Bonsai's quantization method is proprietary though.
That's correct. I think there can certainly be issues even with 4bpw with naive quantization (IE you'll notice far better results from a QAT 4bpw vs a naive 4bpw).
One such method that I've been meaning to look into further is Tencent's AngelSlim QAT/PTQ approach. They did a Hy4 preview release thats an STQ_1_0 at 2.38 bpw:
Of course, it's still 213g of VRAM I'd need so it's somewhat out of the range of what I can run locally. In contrast, this new Bonsai is nice because the original was already exciting for making use of low VRAM devices. Could breath new life into some of the older GPUs that were previously close to top of the line just quite VRAM constrained by modern standards and still quite cost effective for now.
1.76 bpw number is kinda misleading if you compare it directly to IQ2/Q2. The encoding is ternary, but the quantization procedure is way more sophisticated than "round Qwen weights to {-1,0,+1}."
They rotate the weights into a quantization-friendly basis first, then ternarize with per-group scales and error compensation.
> something I was doing with my "strict MVP" build of Anubis' WASM wasn't in fact sticking to just the MVP features of WebAssembly
Fun fact, this _used_ to be the case - wasm32-unknown-unknown had extra non-mvp features added later, which in my eyes has been a breaking change on stable. You're not the first person to have been bitten by this, there's [1] and a similar story in Ruffle [2].
Weirdly, the response from the rustc side (IIRC we also talked with one of the devs of wasm component on discord a bit around 2024) has contained a lot of (paraphrasing) "in retrospect we think wasm32-unknown-unknown target was designed badly*, so we'd prefer if you either dealt with it or switched to another wasm target <with its different set of downsides> than us fixing it". (*I acknowledge that the way std was implemented on that target was unusual and "not clean", but it works _really well_ for a lot of people.)
> I looked into the process involved for rebuilding the standard library twice: once with only MVP wasm features enabled and once with an "all yes config" like usual. Based on some research I did this seemed like a massive pain.
That's what Ruffle does [3], it's a bit unfortunate but wouldn't call it a massive pain.
> Weirdly, the response from the rustc side (IIRC we also talked with one of the devs of wasm component on discord a bit) has contained a lot of (paraphrasing) "in retrospect we think wasm32-unknown-unknown target was designed badly, so we'd prefer if you either dealt with it or switched to another wasm target <with its different set of downsides> than us fixing it". (I acknowledge that the way std was implemented on that target was unusual and "not clean", but it works _really well_ for a lot of people.)
That's about right. Specifically, wasm32-unknown-unknown is serving multiple targets, which aren't very compatible, and it implements std by panicking in all the bits that can't work.
One target is "you truly have absolutely nothing, and don't really have much of `std` at all". Another target is "you're in a browser, and can do anything a typical browser can provide" (ideally targeting the future where wasm has full browser APIs without going through javascript). Another target is "hosted environment that isn't a browser, and which capabilities you have depends on the environment". Those are three very different things.
I would argue that we need at least two new targets, possibly three:
- `wasm32-none` or similar (by analogy with `x86_64-unknown-none`), a target where you don't have anything except for what you bring with you. The equivalent of writing kernel code.
- `wasm32-wasi-hosted` (or pick another name), where you the hoster of the wasm can supply whatever capabilities you have available.
- `wasm32-wasi-browser`.
It's possible that `none` could be a special case of `hosted`, where the services you provide are "none".
It’s trading one property for another, isn’t it? Standard beziers with symmetric handles are C1 continuous but not necessarily G2 continuous; author’s scheme is the other way around, which makes them better for some use cases and worse for others.
Small pushback here: it (and bluesky too) is absolutely atrocious for art, at least as a consumer. Once in a while I stumble upon an artist who only posts their drawings on twitter/bsky, and it’s pretty much impossible to browse their historical works because the „media” tab is just an infinite scroll without any organization; and that’s assuming they don’t also post 10 photos of their cat or other life updates in between their art.
It’s obviously still popular for posting art for many other reasons (I just dislike putting „X”, „art” and „does well” in the same sentence). So IMO the replacements - and lots of these use cases already gave replacements! - struggle simply because of network effects, not their inherent quality.
Python already does it in a lot of cases, see opcode specializations like LOAD_ATTR_INSTANCE_VALUE, LOAD_ATTR_SLOT.
reply