Hi Amir!

Answering inline:

On Thu, Aug 3, 2023 at 10:11 PM Amir E. Aharoni < amir.aharoni@mail.huji.ac.il> wrote:

The email says that "All ML models currently accessible on ORES are also currently accessible on Lift Wing", and if I understand correctly, this means that this feature in Recent Changes will keep working. Do I understand correctly? :)

Definitely yes, we are working on migrating the ORES extension to Lift Wing, without any change required for users. The tracking task is https://phabricator.wikimedia.org/T319170. At the moment all wikis with the ORES extension enabled, except fi/en/wikidata, are already using models from Lift Wing.

In addition, I have some followup questions:

1. The MediaWiki extension that implements the frontend in Recent Changes

is itself named "ORES". It's an internal name that isn't seen much by wiki editors except if they go to Special:Version or to translatewiki. Nevertheless, as the time goes by, seeing the old name may start getting weird. So what's the plan about it? Will this extension remain as is? Will it be renamed? Will it be replaced with a new frontend extension in the foreseeable future?

This is a good question and we don't have a definitive answer at the moment. Our understanding is that renaming extensions in MediaWiki is a long and complicated process, so we'll likely not be able to rename it in the foreseeable future. We would definitely like to add more models to RC Filters, for example Revert Risk (for the curious, see https://meta.wikimedia.org/wiki/Machine_learning_models/Proposed/Language-ag...), but we are not sure yet if it is worth to create a new extension or just to expand the ORES one. We'll get back to this list as soon as we have a better plan :)

2. Back when ORES was originally developed and deployed around 2017,

several wiki editors' communities participated in the development by adapting the product to the needs of their wikis and languages by translating the ORES extension's user interface and, more importantly, by labelling a sample of several thousands of diffs from their wiki using the Wikilabels tool. The communities that did that whole process were, more or less, the communities to which this Recent Changes enhancement was deployed. Will anything like that have to be done again along with the move away from ORES?

The first goal of Lift Wing is to provide a more modern and easy-to-use infrastructure to host models at the WMF, for internal teams and for the community. The focus of the Machine Learning team is to provide infrastructure to run models on, so other teams and the community will be able to propose what to host and we'll vet what is possible and what not (following strict criteria like security of data and PII, model architecture feasibility, etc..). Once a model is deployed on Lift Wing, there will be a team or a community group owning it, namely responsible for its development in terms of features etc.. (more info in https://wikitech.wikimedia.org/wiki/Machine_Learning/LiftWing#Hosting_a_mode... ). To summarize: * All the work done so far with ORES models will be preserved, it is already available on Lift Wing and anybody can use it. We hope that it is now easier to play with model servers and improve them (for WMF and the community), but we are open to any suggestion and feedback about how to improve it. For the curious, more details in the Lift Wing Wikitech page ( https://wikitech.wikimedia.org/wiki/Machine_Learning/LiftWing). * The future work will be split into two main areas (as I see it): ** The ML team will keep working on improving the infrastructure, documentation, performance, etc.. of Lift Wing, to provide better tools and data access for any new idea related to models and their usage. We'll maintain the infrastructure with monitoring/alarms/etc.., so the day-to-day ops will not fall on the model owners (WMF and community), so that they will be able to concentrate themselves only on the models and their future steps. ** Other WMF teams like Research will propose and work on new models that the community needs, but we'll also focus on improving what is currently being used. For example, most of the ORES traffic is for the goodfaith and damaging models that worked very well over the years but they rely on old training data and architectures. The Revert Risk models (for example, https://meta.wikimedia.org/wiki/Machine_learning_models/Proposed/Language-ag...) are an attempt to improve the reliability and performance of the aforementioned models, using a single score instead of multiple ones.

3. Will this change open up the possibility of deploying this Recent

Changes enhancement, or a newer version thereof, to more wikis and languages?

It may be possible in the future to enhance even more the RC Filters, at the moment we are concentrating on migrating the current ones to Lift Wing, but after that we'll start figuring out what is the next step. Any suggestion or advice is really welcome! (see https://wikitech.wikimedia.org/wiki/ORES#Machine_Learning_contacts)

If you think that my questions show a wrong understanding of something,

please let me know—as I said in the beginning, its quite possible :)

Thanks a lot for the questions, I hope I answered your doubts, feel free to follow up if anything is missing!

Luca (on behalf of the ML team)

Our understanding is that renaming extensions in MediaWiki is a long and

complicated process, so we'll likely not be able to rename it in the foreseeable future.

Why?

Renaming is usually a bad thing because it often confuses the hell out of users, but from a technical perspective it is pretty trivial.

-- Bawolff

On Friday, August 4, 2023, Luca Toscano ltoscano@wikimedia.org wrote:

Hi Amir!

Answering inline:

On Thu, Aug 3, 2023 at 10:11 PM Amir E. Aharoni < amir.aharoni@mail.huji.ac.il> wrote:

...

The email says that "All ML models currently accessible on ORES are also currently accessible on Lift Wing", and if I understand correctly, this means that this feature in Recent Changes will keep working. Do I understand correctly? :)

Definitely yes, we are working on migrating the ORES extension to Lift Wing, without any change required for users. The tracking task is https://phabricator.wikimedia.org/T319170. At the moment all wikis with the ORES extension enabled, except fi/en/wikidata, are already using models from Lift Wing.

...

In addition, I have some followup questions:

1. The MediaWiki extension that implements the frontend in Recent Changes

is itself named "ORES". It's an internal name that isn't seen much by wiki editors except if they go to Special:Version or to translatewiki. Nevertheless, as the time goes by, seeing the old name may start getting weird. So what's the plan about it? Will this extension remain as is? Will it be renamed? Will it be replaced with a new frontend extension in the foreseeable future?

This is a good question and we don't have a definitive answer at the moment. Our understanding is that renaming extensions in MediaWiki is a long and complicated process, so we'll likely not be able to rename it in the foreseeable future. We would definitely like to add more models to RC Filters, for example Revert Risk (for the curious, see https://meta.wikimedia.org/wiki/Machine_learning_models/Proposed/Language- agnostic_revert_risk), but we are not sure yet if it is worth to create a new extension or just to expand the ORES one. We'll get back to this list as soon as we have a better plan :)

...

2. Back when ORES was originally developed and deployed around 2017,

several wiki editors' communities participated in the development by adapting the product to the needs of their wikis and languages by translating the ORES extension's user interface and, more importantly, by labelling a sample of several thousands of diffs from their wiki using the Wikilabels tool. The communities that did that whole process were, more or less, the communities to which this Recent Changes enhancement was deployed. Will anything like that have to be done again along with the move away from ORES?

The first goal of Lift Wing is to provide a more modern and easy-to-use infrastructure to host models at the WMF, for internal teams and for the community. The focus of the Machine Learning team is to provide infrastructure to run models on, so other teams and the community will be able to propose what to host and we'll vet what is possible and what not (following strict criteria like security of data and PII, model architecture feasibility, etc..). Once a model is deployed on Lift Wing, there will be a team or a community group owning it, namely responsible for its development in terms of features etc.. (more info in https://wikitech.wikimedia.org/wiki/Machine_Learning/ LiftWing#Hosting_a_model). To summarize:

• All the work done so far with ORES models will be preserved, it is

already available on Lift Wing and anybody can use it. We hope that it is now easier to play with model servers and improve them (for WMF and the community), but we are open to any suggestion and feedback about how to improve it. For the curious, more details in the Lift Wing Wikitech page ( https://wikitech.wikimedia.org/wiki/Machine_Learning/LiftWing).

• The future work will be split into two main areas (as I see it):

** The ML team will keep working on improving the infrastructure, documentation, performance, etc.. of Lift Wing, to provide better tools and data access for any new idea related to models and their usage. We'll maintain the infrastructure with monitoring/alarms/etc.., so the day-to-day ops will not fall on the model owners (WMF and community), so that they will be able to concentrate themselves only on the models and their future steps. ** Other WMF teams like Research will propose and work on new models that the community needs, but we'll also focus on improving what is currently being used. For example, most of the ORES traffic is for the goodfaith and damaging models that worked very well over the years but they rely on old training data and architectures. The Revert Risk models (for example, https://meta.wikimedia.org/wiki/Machine_learning_models/Proposed/ Language-agnostic_revert_risk) are an attempt to improve the reliability and performance of the aforementioned models, using a single score instead of multiple ones.

...

3. Will this change open up the possibility of deploying this Recent

Changes enhancement, or a newer version thereof, to more wikis and languages?

It may be possible in the future to enhance even more the RC Filters, at the moment we are concentrating on migrating the current ones to Lift Wing, but after that we'll start figuring out what is the next step. Any suggestion or advice is really welcome! (see https://wikitech. wikimedia.org/wiki/ORES#Machine_Learning_contacts)

If you think that my questions show a wrong understanding of something,

...

please let me know—as I said in the beginning, its quite possible :)

Thanks a lot for the questions, I hope I answered your doubts, feel free to follow up if anything is missing!

Luca (on behalf of the ML team)

בתאריך יום ו׳, 4 באוג׳ 2023, 13:53, מאת Brian Wolff ‏bawolff@gmail.com:

...

Our understanding is that renaming extensions in MediaWiki is a long

and complicated process, so we'll likely not be able to rename it in the foreseeable future.

Why?

Renaming is usually a bad thing because it often confuses the hell out of users, but from a technical perspective it is pretty trivial.

--

I'm not the biggest expert on MediaWiki, but from the little I do know, the truth is closer to what Bawolff says. It's just a bit of careful searching and replacing. And in this case, it probably doesn't affect the users very much, because, as I've already written above, the name is not seen by most users in the frontend.

Although precisely because of that, it's not the most important thing to do either. It will just become a bit confusing in the long run that the ORES technology is declared as deprecated and the servers are turned off, but the ORES extension is still installed on some big wikis.

If this extension only adds the Recent Changes filtering and highlighting, perhaps it can be given a name that describes its function, such as "MLRevisionLabels" or something like that.

Hi Strainu!

On Fri, Aug 4, 2023 at 3:25 PM Strainu strainu10@gmail.com wrote:

Hi Chris & ML team,

Good to see LiftWing is finally becoming a reality. There are a few things in the documentation that I would like to clarify.

1. In [1], the bot owner is encouraged to move to the revertrisk score.

However, in [2], it's explicitly mentioned that the model should not be used for "Auto-removing edits that a user makes without another editor in the loop". So, should bot owners currently reverting based on goodfaith and damaging scores explore the new models? If so, do you have any suggestions on how to automatically match thresholds between the old and new models?

Diego (from the Research team) answered this bit afaics, but I read it in another Wikitech-l thread (maybe it is my email reader, not sure, but I wanted to point it out in case you missed it). Quoting Diego: """ Sorry for the confusion, we have updated this model card. You can use this model for "automatically reverting content" as you were using ORES. Here you can see the model's performance comparison.

Our current recommendation is to use the Language Agnostic model for this task (patrolling bots). The Multilingual model is performing better for IP Edits, but we are still working on improving its stability. Within the next 3 months we expect to improve Language Agnostic accuracy in anonymous edits, and also Multilingual model stability. """

2. I could not find any reference regarding the ores scores exposed through

other APIs (specifically the RC API [3]). Will those be available going forward? Under which names?

I am very ignorant about RC APIs, but if you want to explore this part more please open a task in Phabricator with the Machine-Learning-team tag, we'll try to research what is possible and get back to you. We'd be also curious to know the use case, to figure out how to best support it.

3. Will it still be possible to (re-)train existing and new model for a

specific wiki? How and when?

So far the ML team concentrated all the efforts in the serving infrastructure (Lift Wing), meanwhile the training part is still to be decided. In [1] we added info about how to request to host a model on Lift Wing, but we didn't provide any automated way to train or retrain the models over time. It is a big effort that we'll tackle in the future, we'll keep this list updated as much as possible. All the models that we host now have been trained on big nodes like the Analytics statistics ones [2], but every re-train is manual and ad-hoc for a specific use case. We are also strongly encouraging people to migrate away from Revscoring models (goodfaith, damaging, etc..) as much as possible, we'd prefer not to to retrain those (where possible) and migrate people to more modern solutions (like Revert Risk). Having said this, if you have any specific request please open a Phabricator task with the Machine-Learning-team tag and we'll evaluate the use case.

Thanks!

Luca

[1]: https://wikitech.wikimedia.org/wiki/Machine_Learning/LiftWing#Hosting_a_mode... [2]: https://wikitech.wikimedia.org/wiki/Analytics/Systems/Clients

Hi Tilman,

Most of the work is still very experimental. We have hosted a few LLMs on Lift Wing already (StarCoder for example) but they were just running on CPU, far too slow for real use cases. But it proves that we can easily host LLMs on Lift Wing. We have been pretty quiet about it while we focus on the ORES migration, but it is our next big project. More soon hopefully!

Where we are now is that we have budget for a big GPU purchase (~10-20 GPUs depending on cost), the question we will try to answer after the ORES migration is complete is: what GPUs should we purchase? We are trying to balance our strong preference to stay open source (i.e. AMD mROC) in a world dominated by a single closed source vendor (i.e. Nvidia). In addition, do we go for a few expensive GPUs better suited to LLMs (A1000, H100, etc) or a mix of big and small? We will need to figure out all this.

I wouldn't characterize WMF's Language Team using CPU as because of AMD, rather at the time we didn't have the budget for GPUs so Lift Wing didn't have any. Since then we have moved two GPUs onto Lift Wing for testing but they are pretty old (2017ish). Once we make the big GPU purchase Lift Wing will gain a lot of functionality for LLM and similar models.

Chris

On Sun, Aug 6, 2023 at 9:57 PM Tilman Bayer haebwiki@gmail.com wrote:

On Thu, Aug 3, 2023 at 7:16 AM Chris Albon calbon@wikimedia.org wrote:

...

Hi everybody,

TL;DR We would like users of ORES models to migrate to our new open source ML infrastructure, Lift Wing, within the next five months. We are available to help you do that, from advice to making code commits. It is important to note: All ML models currently accessible on ORES are also currently accessible on Lift Wing.

As part of the Machine Learning Modernization Project ( https://www.mediawiki.org/wiki/Machine_Learning/Modernization), the Machine Learning team has deployed a Wikimedia’s new machine learning inference infrastructure, called Lift Wing ( https://wikitech.wikimedia.org/wiki/Machine_Learning/LiftWing). Lift Wing brings a lot of new features such as support for GPU-based models, open source LLM hosting, auto-scaling, stability, and ability to host a larger number of models.

This sounds quite exciting! What's the best place to read up on that planned support for GPU-based models and open source LLMs? (I also saw in the recent NYT article[1] that the team is "in the process of adapting A.I. models that are 'off the shelf; — essentially models that have been made available by researchers for anyone to freely customize — so that Wikipedia’s editors can use them for their work.")

I'm aware of the history[2] of not being able to use NVIDIA GPUs due to their CUDA drivers being proprietary. It was mentioned recently in the Wikimedia AI Telegram group that this is still a serious limitation, despite some new explorations with AMD GPUs[3] - to the point that e.g. the WMF's Language team has resorted to using models without GPU support (CPU only).[4] It sounds like there is reasonable hope that this situation could change fairly soon? Would it also mean both at the same time, i.e. open source LLMs running with GPU support (considering that at least some well-known ones appear to require torch.cuda.is_available() == True for that)?

Regards, Tilman

[1] https://www.nytimes.com/2023/07/18/magazine/wikipedia-ai-chatgpt.html [2] https://techblog.wikimedia.org/2020/04/06/saying-no-to-proprietary-code-in-p... [3] https://phabricator.wikimedia.org/T334583 etc. [4] https://diff.wikimedia.org/2023/06/13/mint-supporting-underserved-languages-... or https://thottingal.in/blog/2023/07/21/wikiqa/ (experimental but, I understand, written to be deployable on WMF infrastructure)

...

With the creation of Lift Wing, the team is turning its attention to deprecating the current machine learning infrastructure, ORES. ORES served us really well over the years, it was a successful project but it came before radical changes in technology like Docker, Kubernetes and more recently MLOps. The servers that run ORES are at the end of their planned lifespan and so to save cost we are going to shut them down in early 2024.

We have outlined a deprecation path on Wikitech ( https://wikitech.wikimedia.org/wiki/ORES), please read the page if you are a maintainer of a tool or code that uses the ORES endpoint https://ores.wikimedia.org/). If you have any doubt or if you need assistance in migrating to Lift Wing, feel free to contact the ML team via:

• Email: ml@wikimedia.org
• Phabricator: #Machine-Learning-Team tag
• IRC (Libera): #wikimedia-ml

The Machine Learning team is available to help projects migrate, from offering advice to making code commits. We want to make this as easy as possible for folks.

High Level timeline:

**By September 30th 2023: *Infrastructure powering the ORES API endpoint will be migrated from ORES to Lift Wing. For users, the API endpoint will remain the same, and most users won’t notice any change. Rather just the backend services powering the endpoint will change.

Details: We'd like to add a DNS CNAME that points ores.wikimedia.org to ores-legacy.wikimedia.org, a new endpoint that offers a almost complete replacement of the ORES API calling Lift Wing behind the scenes. In an ideal world we'd migrate all tools to Lift Wing before decommissioning the infrastructure behind ores.wikimedia.org, but it turned out to be really challenging so to avoid disrupting users we chose to implement a transition layer/API.

To summarize, if you don't have time to migrate before September to Lift Wing, your code/tool should work just fine on ores-legacy.wikimedia.org and you'll not have to change a line in your code thanks to the DNS CNAME. The ores-legacy endpoint is not a 100% replacement for ores, we removed some very old and not used features, so we highly recommend at least test the new endpoint for your use case to avoid surprises when we'll make the switch. In case you find anything weird, please report it to us using the aforementioned channels.

**September to January: *We will be reaching out to every user of ORES we can identify and working with them to make the migration process as easy as possible.

**By January 2024: *If all goes well, we would like zero traffic on the ORES API endpoint so we can turn off the ores-legacy API.

If you want more information about Lift Wing, please check https://wikitech.wikimedia.org/wiki/Machine_Learning/LiftWing

Thanks in advance for the patience and the help!

Regards,

The Machine Learning Team _______________________________________________ Wikitech-l mailing list -- wikitech-l@lists.wikimedia.org To unsubscribe send an email to wikitech-l-leave@lists.wikimedia.org

https://lists.wikimedia.org/postorius/lists/wikitech-l.lists.wikimedia.org/

---

Wikitech-l mailing list -- wikitech-l@lists.wikimedia.org To unsubscribe send an email to wikitech-l-leave@lists.wikimedia.org https://lists.wikimedia.org/postorius/lists/wikitech-l.lists.wikimedia.org/

Hi Tilman!

On Tue, Aug 8, 2023 at 5:45 AM Tilman Bayer haebwiki@gmail.com wrote:

Hi Chris,

On Mon, Aug 7, 2023 at 11:51 AM Chris Albon calbon@wikimedia.org wrote:

...

Hi Tilman,

Most of the work is still very experimental. We have hosted a few LLMs on Lift Wing already (StarCoder for example) but they were just running on CPU, far too slow for real use cases. But it proves that we can easily host LLMs on Lift Wing. We have been pretty quiet about it while we focus on the ORES migration, but it is our next big project. More soon hopefully!

Understood. Looking forward to learning more later!

...

Where we are now is that we have budget for a big GPU purchase (~10-20 GPUs depending on cost), the question we will try to answer after the ORES migration is complete is: what GPUs should we purchase? We are trying to balance our strong preference to stay open source (i.e. AMD mROC) in a world dominated by a single closed source vendor (i.e. Nvidia). In addition, do we go for a few expensive GPUs better suited to LLMs (A1000, H100, etc) or a mix of big and small? We will need to figure out all this.

I see. On that matter, what do you folks make of the recent announcements of AMD's partnerships with Hugging Face and Pytorch[5]? (which, I understand, came after the ML team had already launched the aforementioned new AMD explorations)

"Open-source AI: AMD looks to Hugging Face and Meta spinoff PyTorch to take on Nvidia [...] Both partnerships involve AMD’s ROCm AI software stack, the company’s answer to Nvidia’s proprietary CUDA platform and application-programming interface. AMD called ROCm an open and portable AI system with out-of-the-box support that can port to existing AI models. [...B]oth AMD and Hugging Face are dedicating engineering resources to each other and sharing data to ensure that the constantly updated AI models from Hugging Face, which might not otherwise run well on AMD hardware, would be “guaranteed” to work on hardware like the MI300X. [...] AMD said PyTorch will fully upstream the ROCm software stack and “provide immediate ‘day zero’ support for PyTorch 2.0 with ROCm release 5.4.2 on all AMD Instinct accelerators,” which is meant to appeal to those customers looking to switch from Nvidia’s software ecosystem."

In their own announcement, Hugging Face offered further details, including a pretty impressive list of models to be supported:[6]

"We intend to support state-of-the-art transformer architectures for natural language processing, computer vision, and speech, such as BERT, DistilBERT, ROBERTA, Vision Transformer, CLIP, and Wav2Vec2. Of course, generative AI models will be available too (e.g., GPT2, GPT-NeoX, T5, OPT, LLaMA), including our own BLOOM and StarCoder models. Lastly, we will also support more traditional computer vision models, like ResNet and ResNext, and deep learning recommendation models, a first for us. [..] We'll do our best to test and validate these models for PyTorch, TensorFlow, and ONNX Runtime for the above platforms. [...] We will integrate the AMD ROCm SDK seamlessly in our open-source libraries, starting with the transformers library."

Do you think this may promise too much, or could it point to a possible solution of the Foundation's conundrum?

In https://phabricator.wikimedia.org/T334583 we experimented with LLMs and AMD GPUs on Lift Wing, and we confirmed the good results that Pytorch announced, We were able to run bloom-3b, bloom-560m, nllb-200 and falcon-7b on Lift Wing, having issues only with the last one since the GPU VRAM was not enough (16GB are low for Falcon-7b). So we can confirm that AMD ROCm works really well with Pytorch :)

In any case, this seems to be an interesting moment where many in AI are trying to move away from Nvidia's proprietary CUDA platform.

This is my own view, not my team's, so I can't speak up for what the WMF will decide, but I think we should keep going with AMD and avoid Nvidia as much as possible. Our strong stand against proprietary software should hold, even if it means more efforts and work to advance in the ML field. I completely get the frustration when common libraries and tools have more difficulty to run on AMD than Nvidia, but our communities should align (in my opinion) to the most open source solution and contribute (where possible) so that more and more people adopt the same. Adding proprietary software to the WMF infrastructure and practices is also something that is technically difficult for various reasons (from the Linux Kernel maintenance to Debian package upload), meanwhile we already have everything set up and working for AMD (that works nicely with our infrastructure). Moreover Debian upstream has recently created a team to maintain AMD ROCm packages (https://lists.debian.org/debian-ai/), so it will be interesting to see what their direction will be (so far it seems aligned to ours).

Thanks!

Luca

Hello!

As a next step in the deprecation process of ORES https://wikitech.wikimedia.org/wiki/ORES the Machine Learning team will switch the backend of ores.wikimedia.org to ores-legacy, a k8s application meant to provide a compatibility layer between ORES and Lift Wing so users that have not yet migrated to Lift Wing will be transparently migrated. Ores-legacy is an application that has the same API as ORES but in the background makes requests to Lift Wing, allowing us to decommission the ORES servers until all clients have moved.

This change is planned to take place on Monday 25th of September. If you have a client/application that is still using ORES we expect that this switch is going to be transparent for you.

However keep in mind that ores-legacy is not a 100% replacement for ORES as some old and unused features are no longer supported.

If you see anything out of the ordinary, feel free to contact the Machine Learning team:

IRC libera: #wikimedia-ml

Phabricator: Machine-Learning-team tag

Thank you!

On Wed, Aug 9, 2023 at 1:22 PM Chaloemphon Praphuchakang < yoshrakpraphu@gmail.com> wrote:

On Tue, 8 Aug 2023, 10:45 Tilman Bayer, haebwiki@gmail.com wrote:

...

Hi Chris,

On Mon, Aug 7, 2023 at 11:51 AM Chris Albon calbon@wikimedia.org wrote:

...

Hi Tilman,

Most of the work is still very experimental. We have hosted a few LLMs on Lift Wing already (StarCoder for example) but they were just running on CPU, far too slow for real use cases. But it proves that we can easily host LLMs on Lift Wing. We have been pretty quiet about it while we focus on the ORES migration, but it is our next big project. More soon hopefully!

Understood. Looking forward to learning more later!

...

Where we are now is that we have budget for a big GPU purchase (~10-20 GPUs depending on cost), the question we will try to answer after the ORES migration is complete is: what GPUs should we purchase? We are trying to balance our strong preference to stay open source (i.e. AMD mROC) in a world dominated by a single closed source vendor (i.e. Nvidia). In addition, do we go for a few expensive GPUs better suited to LLMs (A1000, H100, etc) or a mix of big and small? We will need to figure out all this.

I see. On that matter, what do you folks make of the recent announcements of AMD's partnerships with Hugging Face and Pytorch[5]? (which, I understand, came after the ML team had already launched the aforementioned new AMD explorations)

"Open-source AI: AMD looks to Hugging Face and Meta spinoff PyTorch to take on Nvidia [...] Both partnerships involve AMD’s ROCm AI software stack, the company’s answer to Nvidia’s proprietary CUDA platform and application-programming interface. AMD called ROCm an open and portable AI system with out-of-the-box support that can port to existing AI models. [...B]oth AMD and Hugging Face are dedicating engineering resources to each other and sharing data to ensure that the constantly updated AI models from Hugging Face, which might not otherwise run well on AMD hardware, would be “guaranteed” to work on hardware like the MI300X. [...] AMD said PyTorch will fully upstream the ROCm software stack and “provide immediate ‘day zero’ support for PyTorch 2.0 with ROCm release 5.4.2 on all AMD Instinct accelerators,” which is meant to appeal to those customers looking to switch from Nvidia’s software ecosystem."

In their own announcement, Hugging Face offered further details, including a pretty impressive list of models to be supported:[6]

"We intend to support state-of-the-art transformer architectures for natural language processing, computer vision, and speech, such as BERT, DistilBERT, ROBERTA, Vision Transformer, CLIP, and Wav2Vec2. Of course, generative AI models will be available too (e.g., GPT2, GPT-NeoX, T5, OPT, LLaMA), including our own BLOOM and StarCoder models. Lastly, we will also support more traditional computer vision models, like ResNet and ResNext, and deep learning recommendation models, a first for us. [..] We'll do our best to test and validate these models for PyTorch, TensorFlow, and ONNX Runtime for the above platforms. [...] We will integrate the AMD ROCm SDK seamlessly in our open-source libraries, starting with the transformers library."

Do you think this may promise too much, or could it point to a possible solution of the Foundation's conundrum? In any case, this seems to be an interesting moment where many in AI are trying to move away from Nvidia's proprietary CUDA platform. Most of them probably more for financial and availability reasons though, given the current GPU shortages[7] (which the ML team is undoubtedly aware of already; mentioning this as context for others on this list. See also Marketwatch's remarks about current margins[5]).

Regards, Tilman

[5] https://archive.ph/2023.06.15-173527/https://www.marketwatch.com/amp/story/o... [6] https://huggingface.co/blog/huggingface-and-amd [7] See e.g. https://gpus.llm-utils.org/nvidia-h100-gpus-supply-and-demand/ (avoid playing the song though. Don't say I didn't warn you)

...

I wouldn't characterize WMF's Language Team using CPU as because of AMD, rather at the time we didn't have the budget for GPUs so Lift Wing didn't have any. Since then we have moved two GPUs onto Lift Wing for testing but they are pretty old (2017ish). Once we make the big GPU purchase Lift Wing will gain a lot of functionality for LLM and similar models.

Chris

On Sun, Aug 6, 2023 at 9:57 PM Tilman Bayer haebwiki@gmail.com wrote:

...

On Thu, Aug 3, 2023 at 7:16 AM Chris Albon calbon@wikimedia.org wrote:

...

Hi everybody,

TL;DR We would like users of ORES models to migrate to our new open source ML infrastructure, Lift Wing, within the next five months. We are available to help you do that, from advice to making code commits. It is important to note: All ML models currently accessible on ORES are also currently accessible on Lift Wing.

As part of the Machine Learning Modernization Project ( https://www.mediawiki.org/wiki/Machine_Learning/Modernization), the Machine Learning team has deployed a Wikimedia’s new machine learning inference infrastructure, called Lift Wing ( https://wikitech.wikimedia.org/wiki/Machine_Learning/LiftWing). Lift Wing brings a lot of new features such as support for GPU-based models, open source LLM hosting, auto-scaling, stability, and ability to host a larger number of models.

This sounds quite exciting! What's the best place to read up on that planned support for GPU-based models and open source LLMs? (I also saw in the recent NYT article[1] that the team is "in the process of adapting A.I. models that are 'off the shelf; — essentially models that have been made available by researchers for anyone to freely customize — so that Wikipedia’s editors can use them for their work.")

I'm aware of the history[2] of not being able to use NVIDIA GPUs due to their CUDA drivers being proprietary. It was mentioned recently in the Wikimedia AI Telegram group that this is still a serious limitation, despite some new explorations with AMD GPUs[3] - to the point that e.g. the WMF's Language team has resorted to using models without GPU support (CPU only).[4] It sounds like there is reasonable hope that this situation could change fairly soon? Would it also mean both at the same time, i.e. open source LLMs running with GPU support (considering that at least some well-known ones appear to require torch.cuda.is_available() == True for that)?

Regards, Tilman

[1] https://www.nytimes.com/2023/07/18/magazine/wikipedia-ai-chatgpt.html [2] https://techblog.wikimedia.org/2020/04/06/saying-no-to-proprietary-code-in-p... [3] https://phabricator.wikimedia.org/T334583 etc. [4] https://diff.wikimedia.org/2023/06/13/mint-supporting-underserved-languages-... or https://thottingal.in/blog/2023/07/21/wikiqa/ (experimental but, I understand, written to be deployable on WMF infrastructure)

...

With the creation of Lift Wing, the team is turning its attention to deprecating the current machine learning infrastructure, ORES. ORES served us really well over the years, it was a successful project but it came before radical changes in technology like Docker, Kubernetes and more recently MLOps. The servers that run ORES are at the end of their planned lifespan and so to save cost we are going to shut them down in early 2024.

We have outlined a deprecation path on Wikitech ( https://wikitech.wikimedia.org/wiki/ORES), please read the page if you are a maintainer of a tool or code that uses the ORES endpoint https://ores.wikimedia.org/). If you have any doubt or if you need assistance in migrating to Lift Wing, feel free to contact the ML team via:

• Email: ml@wikimedia.org
• Phabricator: #Machine-Learning-Team tag
• IRC (Libera): #wikimedia-ml

The Machine Learning team is available to help projects migrate, from offering advice to making code commits. We want to make this as easy as possible for folks.

High Level timeline:

**By September 30th 2023: *Infrastructure powering the ORES API endpoint will be migrated from ORES to Lift Wing. For users, the API endpoint will remain the same, and most users won’t notice any change. Rather just the backend services powering the endpoint will change.

Details: We'd like to add a DNS CNAME that points ores.wikimedia.org to ores-legacy.wikimedia.org, a new endpoint that offers a almost complete replacement of the ORES API calling Lift Wing behind the scenes. In an ideal world we'd migrate all tools to Lift Wing before decommissioning the infrastructure behind ores.wikimedia.org, but it turned out to be really challenging so to avoid disrupting users we chose to implement a transition layer/API.

To summarize, if you don't have time to migrate before September to Lift Wing, your code/tool should work just fine on ores-legacy.wikimedia.org and you'll not have to change a line in your code thanks to the DNS CNAME. The ores-legacy endpoint is not a 100% replacement for ores, we removed some very old and not used features, so we highly recommend at least test the new endpoint for your use case to avoid surprises when we'll make the switch. In case you find anything weird, please report it to us using the aforementioned channels.

**September to January: *We will be reaching out to every user of ORES we can identify and working with them to make the migration process as easy as possible.

**By January 2024: *If all goes well, we would like zero traffic on the ORES API endpoint so we can turn off the ores-legacy API.

If you want more information about Lift Wing, please check https://wikitech.wikimedia.org/wiki/Machine_Learning/LiftWing

Thanks in advance for the patience and the help!

Regards,

The Machine Learning Team _______________________________________________ Wikitech-l mailing list -- wikitech-l@lists.wikimedia.org To unsubscribe send an email to wikitech-l-leave@lists.wikimedia.org

https://lists.wikimedia.org/postorius/lists/wikitech-l.lists.wikimedia.org/

---

Wikitech-l mailing list -- wikitech-l@lists.wikimedia.org To unsubscribe send an email to wikitech-l-leave@lists.wikimedia.org

https://lists.wikimedia.org/postorius/lists/wikitech-l.lists.wikimedia.org/

---

Wikitech-l mailing list -- wikitech-l@lists.wikimedia.org To unsubscribe send an email to wikitech-l-leave@lists.wikimedia.org

https://lists.wikimedia.org/postorius/lists/wikitech-l.lists.wikimedia.org/

---

Wikitech-l mailing list -- wikitech-l@lists.wikimedia.org To unsubscribe send an email to wikitech-l-leave@lists.wikimedia.org

https://lists.wikimedia.org/postorius/lists/wikitech-l.lists.wikimedia.org/

---

Wikitech-l mailing list -- wikitech-l@lists.wikimedia.org To unsubscribe send an email to wikitech-l-leave@lists.wikimedia.org https://lists.wikimedia.org/postorius/lists/wikitech-l.lists.wikimedia.org/

Hi Aaron!

Thanks for following up. The API is almost compatible with what ORES currently does, but there are limitations (like the max number of revisions in a batch etc..). The API clearly states when something is not supported, so you can check its compatibility now making some requests to:

https://ores-legacy.wikimedia.org

If you open a task with a list of systems that you need to migrate we can definitely take a look and help. So far the traffic being served by ORES has been reduced to few clients, and all of them don't run with recognizable UAs (see https://meta.wikimedia.org/wiki/User-Agent_policy) so we'll try our best to support them. The migration to Lift Wing has been widely publicized, a lot of documentation is available to migrate. We'd suggest trying Lift Wing for your systems instead (see https://wikitech.wikimedia.org/wiki/Machine_Learning/LiftWing/Usage).

The Machine Learning plan is to eventually deprecate ores-legacy too, to maintain only one system (namely Lift Wing). There is no final date yet, we'll try to reach out to all remaining users first, so if you plan to keep using ores-legacy please follow up with us first :)

Thanks!

Luca (on behalf of the ML Team)

On Fri, Sep 22, 2023 at 5:10 PM Aaron Halfaker aaron.halfaker@gmail.com wrote:

Does the new ores-legacy support the same feature set. E.g. features output, injection, and threshold optimizations. Or is it just prediction? This will affect some of the systems I need to migrate.

On Fri, Sep 22, 2023, 06:21 Ilias Sarantopoulos < isarantopoulos@wikimedia.org> wrote:

...

Hello!

As a next step in the deprecation process of ORES https://wikitech.wikimedia.org/wiki/ORES the Machine Learning team will switch the backend of ores.wikimedia.org to ores-legacy, a k8s application meant to provide a compatibility layer between ORES and Lift Wing so users that have not yet migrated to Lift Wing will be transparently migrated. Ores-legacy is an application that has the same API as ORES but in the background makes requests to Lift Wing, allowing us to decommission the ORES servers until all clients have moved.

This change is planned to take place on Monday 25th of September. If you have a client/application that is still using ORES we expect that this switch is going to be transparent for you.

However keep in mind that ores-legacy is not a 100% replacement for ORES as some old and unused features are no longer supported.

If you see anything out of the ordinary, feel free to contact the Machine Learning team:

IRC libera: #wikimedia-ml

Phabricator: Machine-Learning-team tag

Thank you!

On Wed, Aug 9, 2023 at 1:22 PM Chaloemphon Praphuchakang < yoshrakpraphu@gmail.com> wrote:

...

On Tue, 8 Aug 2023, 10:45 Tilman Bayer, haebwiki@gmail.com wrote:

...

Hi Chris,

On Mon, Aug 7, 2023 at 11:51 AM Chris Albon calbon@wikimedia.org wrote:

...

Hi Tilman,

Most of the work is still very experimental. We have hosted a few LLMs on Lift Wing already (StarCoder for example) but they were just running on CPU, far too slow for real use cases. But it proves that we can easily host LLMs on Lift Wing. We have been pretty quiet about it while we focus on the ORES migration, but it is our next big project. More soon hopefully!

Understood. Looking forward to learning more later!

...

Where we are now is that we have budget for a big GPU purchase (~10-20 GPUs depending on cost), the question we will try to answer after the ORES migration is complete is: what GPUs should we purchase? We are trying to balance our strong preference to stay open source (i.e. AMD mROC) in a world dominated by a single closed source vendor (i.e. Nvidia). In addition, do we go for a few expensive GPUs better suited to LLMs (A1000, H100, etc) or a mix of big and small? We will need to figure out all this.

I see. On that matter, what do you folks make of the recent announcements of AMD's partnerships with Hugging Face and Pytorch[5]? (which, I understand, came after the ML team had already launched the aforementioned new AMD explorations)

"Open-source AI: AMD looks to Hugging Face and Meta spinoff PyTorch to take on Nvidia [...] Both partnerships involve AMD’s ROCm AI software stack, the company’s answer to Nvidia’s proprietary CUDA platform and application-programming interface. AMD called ROCm an open and portable AI system with out-of-the-box support that can port to existing AI models. [...B]oth AMD and Hugging Face are dedicating engineering resources to each other and sharing data to ensure that the constantly updated AI models from Hugging Face, which might not otherwise run well on AMD hardware, would be “guaranteed” to work on hardware like the MI300X. [...] AMD said PyTorch will fully upstream the ROCm software stack and “provide immediate ‘day zero’ support for PyTorch 2.0 with ROCm release 5.4.2 on all AMD Instinct accelerators,” which is meant to appeal to those customers looking to switch from Nvidia’s software ecosystem."

In their own announcement, Hugging Face offered further details, including a pretty impressive list of models to be supported:[6]

"We intend to support state-of-the-art transformer architectures for natural language processing, computer vision, and speech, such as BERT, DistilBERT, ROBERTA, Vision Transformer, CLIP, and Wav2Vec2. Of course, generative AI models will be available too (e.g., GPT2, GPT-NeoX, T5, OPT, LLaMA), including our own BLOOM and StarCoder models. Lastly, we will also support more traditional computer vision models, like ResNet and ResNext, and deep learning recommendation models, a first for us. [..] We'll do our best to test and validate these models for PyTorch, TensorFlow, and ONNX Runtime for the above platforms. [...] We will integrate the AMD ROCm SDK seamlessly in our open-source libraries, starting with the transformers library."

Do you think this may promise too much, or could it point to a possible solution of the Foundation's conundrum? In any case, this seems to be an interesting moment where many in AI are trying to move away from Nvidia's proprietary CUDA platform. Most of them probably more for financial and availability reasons though, given the current GPU shortages[7] (which the ML team is undoubtedly aware of already; mentioning this as context for others on this list. See also Marketwatch's remarks about current margins[5]).

Regards, Tilman

[5] https://archive.ph/2023.06.15-173527/https://www.marketwatch.com/amp/story/o... [6] https://huggingface.co/blog/huggingface-and-amd [7] See e.g. https://gpus.llm-utils.org/nvidia-h100-gpus-supply-and-demand/ (avoid playing the song though. Don't say I didn't warn you)

...

I wouldn't characterize WMF's Language Team using CPU as because of AMD, rather at the time we didn't have the budget for GPUs so Lift Wing didn't have any. Since then we have moved two GPUs onto Lift Wing for testing but they are pretty old (2017ish). Once we make the big GPU purchase Lift Wing will gain a lot of functionality for LLM and similar models.

Chris

On Sun, Aug 6, 2023 at 9:57 PM Tilman Bayer haebwiki@gmail.com wrote:

...

On Thu, Aug 3, 2023 at 7:16 AM Chris Albon calbon@wikimedia.org wrote:

> Hi everybody, > > TL;DR We would like users of ORES models to migrate to our new open > source ML infrastructure, Lift Wing, within the next five months. We are > available to help you do that, from advice to making code commits. It is > important to note: All ML models currently accessible on ORES are also > currently accessible on Lift Wing. > > As part of the Machine Learning Modernization Project ( > https://www.mediawiki.org/wiki/Machine_Learning/Modernization), the > Machine Learning team has deployed a Wikimedia’s new machine learning > inference infrastructure, called Lift Wing ( > https://wikitech.wikimedia.org/wiki/Machine_Learning/LiftWing). > Lift Wing brings a lot of new features such as support for GPU-based > models, open source LLM hosting, auto-scaling, stability, and ability to > host a larger number of models. >

This sounds quite exciting! What's the best place to read up on that planned support for GPU-based models and open source LLMs? (I also saw in the recent NYT article[1] that the team is "in the process of adapting A.I. models that are 'off the shelf; — essentially models that have been made available by researchers for anyone to freely customize — so that Wikipedia’s editors can use them for their work.")

I'm aware of the history[2] of not being able to use NVIDIA GPUs due to their CUDA drivers being proprietary. It was mentioned recently in the Wikimedia AI Telegram group that this is still a serious limitation, despite some new explorations with AMD GPUs[3] - to the point that e.g. the WMF's Language team has resorted to using models without GPU support (CPU only).[4] It sounds like there is reasonable hope that this situation could change fairly soon? Would it also mean both at the same time, i.e. open source LLMs running with GPU support (considering that at least some well-known ones appear to require torch.cuda.is_available() == True for that)?

Regards, Tilman

[1] https://www.nytimes.com/2023/07/18/magazine/wikipedia-ai-chatgpt.html [2] https://techblog.wikimedia.org/2020/04/06/saying-no-to-proprietary-code-in-p... [3] https://phabricator.wikimedia.org/T334583 etc. [4] https://diff.wikimedia.org/2023/06/13/mint-supporting-underserved-languages-... or https://thottingal.in/blog/2023/07/21/wikiqa/ (experimental but, I understand, written to be deployable on WMF infrastructure)

> > With the creation of Lift Wing, the team is turning its attention to > deprecating the current machine learning infrastructure, ORES. ORES served > us really well over the years, it was a successful project but it came > before radical changes in technology like Docker, Kubernetes and more > recently MLOps. The servers that run ORES are at the end of their planned > lifespan and so to save cost we are going to shut them down in early 2024. > > We have outlined a deprecation path on Wikitech ( > https://wikitech.wikimedia.org/wiki/ORES), please read the page if > you are a maintainer of a tool or code that uses the ORES endpoint > https://ores.wikimedia.org/). If you have any doubt or if you need > assistance in migrating to Lift Wing, feel free to contact the ML team via: > > - Email: ml@wikimedia.org > - Phabricator: #Machine-Learning-Team tag > - IRC (Libera): #wikimedia-ml > > The Machine Learning team is available to help projects migrate, > from offering advice to making code commits. We want to make this as easy > as possible for folks. > > High Level timeline: > > **By September 30th 2023: *Infrastructure powering the ORES API > endpoint will be migrated from ORES to Lift Wing. For users, the API > endpoint will remain the same, and most users won’t notice any change. > Rather just the backend services powering the endpoint will change. > > Details: We'd like to add a DNS CNAME that points ores.wikimedia.org > to ores-legacy.wikimedia.org, a new endpoint that offers a almost > complete replacement of the ORES API calling Lift Wing behind the scenes. > In an ideal world we'd migrate all tools to Lift Wing before > decommissioning the infrastructure behind ores.wikimedia.org, but > it turned out to be really challenging so to avoid disrupting users we > chose to implement a transition layer/API. > > To summarize, if you don't have time to migrate before September to > Lift Wing, your code/tool should work just fine on > ores-legacy.wikimedia.org and you'll not have to change a line in > your code thanks to the DNS CNAME. The ores-legacy endpoint is not a 100% > replacement for ores, we removed some very old and not used features, so we > highly recommend at least test the new endpoint for your use case to avoid > surprises when we'll make the switch. In case you find anything weird, > please report it to us using the aforementioned channels. > > **September to January: *We will be reaching out to every user of > ORES we can identify and working with them to make the migration process as > easy as possible. > > **By January 2024: *If all goes well, we would like zero traffic on > the ORES API endpoint so we can turn off the ores-legacy API. > > If you want more information about Lift Wing, please check > https://wikitech.wikimedia.org/wiki/Machine_Learning/LiftWing > > Thanks in advance for the patience and the help! > > Regards, > > The Machine Learning Team > _______________________________________________ > Wikitech-l mailing list -- wikitech-l@lists.wikimedia.org > To unsubscribe send an email to wikitech-l-leave@lists.wikimedia.org > > https://lists.wikimedia.org/postorius/lists/wikitech-l.lists.wikimedia.org/

---

Wikitech-l mailing list -- wikitech-l@lists.wikimedia.org To unsubscribe send an email to wikitech-l-leave@lists.wikimedia.org

https://lists.wikimedia.org/postorius/lists/wikitech-l.lists.wikimedia.org/

---

Wikitech-l mailing list -- wikitech-l@lists.wikimedia.org To unsubscribe send an email to wikitech-l-leave@lists.wikimedia.org

https://lists.wikimedia.org/postorius/lists/wikitech-l.lists.wikimedia.org/

---

Wikitech-l mailing list -- wikitech-l@lists.wikimedia.org To unsubscribe send an email to wikitech-l-leave@lists.wikimedia.org

https://lists.wikimedia.org/postorius/lists/wikitech-l.lists.wikimedia.org/

---

Wikitech-l mailing list -- wikitech-l@lists.wikimedia.org To unsubscribe send an email to wikitech-l-leave@lists.wikimedia.org

https://lists.wikimedia.org/postorius/lists/wikitech-l.lists.wikimedia.org/

---

Wikitech-l mailing list -- wikitech-l@lists.wikimedia.org To unsubscribe send an email to wikitech-l-leave@lists.wikimedia.org

https://lists.wikimedia.org/postorius/lists/wikitech-l.lists.wikimedia.org/

---

Wikitech-l mailing list -- wikitech-l@lists.wikimedia.org To unsubscribe send an email to wikitech-l-leave@lists.wikimedia.org https://lists.wikimedia.org/postorius/lists/wikitech-l.lists.wikimedia.org/

It looks like model_info is not implemented at all. E.g. https://ores-legacy.wikimedia.org/v3/scores/enwiki?model_info=statistics.thr...

I get {"detail":{"error":{"code":"bad request","message":"model_info query parameter is not supported by this endpoint anymore. For more information please visit https://wikitech.wikimedia.org/wiki/ORES%22%7D%7D%7D

But when I go to that page, nothing discusses model_info. Is there a way to get this from LiftWing?

On Fri, Sep 22, 2023 at 8:53 AM Aaron Halfaker aaron.halfaker@gmail.com wrote:

Do you have a tag for filing bugs against ORES-legacy? I can't seem to find a relevant one in phab.

On Fri, Sep 22, 2023 at 8:39 AM Luca Toscano ltoscano@wikimedia.org wrote:

...

Hi Aaron!

Thanks for following up. The API is almost compatible with what ORES currently does, but there are limitations (like the max number of revisions in a batch etc..). The API clearly states when something is not supported, so you can check its compatibility now making some requests to:

https://ores-legacy.wikimedia.org

If you open a task with a list of systems that you need to migrate we can definitely take a look and help. So far the traffic being served by ORES has been reduced to few clients, and all of them don't run with recognizable UAs (see https://meta.wikimedia.org/wiki/User-Agent_policy) so we'll try our best to support them. The migration to Lift Wing has been widely publicized, a lot of documentation is available to migrate. We'd suggest trying Lift Wing for your systems instead (see https://wikitech.wikimedia.org/wiki/Machine_Learning/LiftWing/Usage).

The Machine Learning plan is to eventually deprecate ores-legacy too, to maintain only one system (namely Lift Wing). There is no final date yet, we'll try to reach out to all remaining users first, so if you plan to keep using ores-legacy please follow up with us first :)

Thanks!

Luca (on behalf of the ML Team)

On Fri, Sep 22, 2023 at 5:10 PM Aaron Halfaker aaron.halfaker@gmail.com wrote:

...

Does the new ores-legacy support the same feature set. E.g. features output, injection, and threshold optimizations. Or is it just prediction? This will affect some of the systems I need to migrate.

On Fri, Sep 22, 2023, 06:21 Ilias Sarantopoulos < isarantopoulos@wikimedia.org> wrote:

...

Hello!

As a next step in the deprecation process of ORES https://wikitech.wikimedia.org/wiki/ORES the Machine Learning team will switch the backend of ores.wikimedia.org to ores-legacy, a k8s application meant to provide a compatibility layer between ORES and Lift Wing so users that have not yet migrated to Lift Wing will be transparently migrated. Ores-legacy is an application that has the same API as ORES but in the background makes requests to Lift Wing, allowing us to decommission the ORES servers until all clients have moved.

This change is planned to take place on Monday 25th of September. If you have a client/application that is still using ORES we expect that this switch is going to be transparent for you.

However keep in mind that ores-legacy is not a 100% replacement for ORES as some old and unused features are no longer supported.

If you see anything out of the ordinary, feel free to contact the Machine Learning team:

IRC libera: #wikimedia-ml

Phabricator: Machine-Learning-team tag

Thank you!

On Wed, Aug 9, 2023 at 1:22 PM Chaloemphon Praphuchakang < yoshrakpraphu@gmail.com> wrote:

...

On Tue, 8 Aug 2023, 10:45 Tilman Bayer, haebwiki@gmail.com wrote:

...

Hi Chris,

On Mon, Aug 7, 2023 at 11:51 AM Chris Albon calbon@wikimedia.org wrote:

> Hi Tilman, > > Most of the work is still very experimental. We have hosted a few > LLMs on Lift Wing already (StarCoder for example) but they were just > running on CPU, far too slow for real use cases. But it proves that we can > easily host LLMs on Lift Wing. We have been pretty quiet about it while we > focus on the ORES migration, but it is our next big project. More soon > hopefully! > Understood. Looking forward to learning more later!

> Where we are now is that we have budget for a big GPU purchase > (~10-20 GPUs depending on cost), the question we will try to answer after > the ORES migration is complete is: what GPUs should we purchase? We are > trying to balance our strong preference to stay open source (i.e. AMD mROC) > in a world dominated by a single closed source vendor (i.e. Nvidia). In > addition, do we go for a few expensive GPUs better suited to LLMs (A1000, > H100, etc) or a mix of big and small? We will need to figure out all this. > I see. On that matter, what do you folks make of the recent announcements of AMD's partnerships with Hugging Face and Pytorch[5]? (which, I understand, came after the ML team had already launched the aforementioned new AMD explorations)

"Open-source AI: AMD looks to Hugging Face and Meta spinoff PyTorch to take on Nvidia [...] Both partnerships involve AMD’s ROCm AI software stack, the company’s answer to Nvidia’s proprietary CUDA platform and application-programming interface. AMD called ROCm an open and portable AI system with out-of-the-box support that can port to existing AI models. [...B]oth AMD and Hugging Face are dedicating engineering resources to each other and sharing data to ensure that the constantly updated AI models from Hugging Face, which might not otherwise run well on AMD hardware, would be “guaranteed” to work on hardware like the MI300X. [...] AMD said PyTorch will fully upstream the ROCm software stack and “provide immediate ‘day zero’ support for PyTorch 2.0 with ROCm release 5.4.2 on all AMD Instinct accelerators,” which is meant to appeal to those customers looking to switch from Nvidia’s software ecosystem."

In their own announcement, Hugging Face offered further details, including a pretty impressive list of models to be supported:[6]

"We intend to support state-of-the-art transformer architectures for natural language processing, computer vision, and speech, such as BERT, DistilBERT, ROBERTA, Vision Transformer, CLIP, and Wav2Vec2. Of course, generative AI models will be available too (e.g., GPT2, GPT-NeoX, T5, OPT, LLaMA), including our own BLOOM and StarCoder models. Lastly, we will also support more traditional computer vision models, like ResNet and ResNext, and deep learning recommendation models, a first for us. [..] We'll do our best to test and validate these models for PyTorch, TensorFlow, and ONNX Runtime for the above platforms. [...] We will integrate the AMD ROCm SDK seamlessly in our open-source libraries, starting with the transformers library."

Do you think this may promise too much, or could it point to a possible solution of the Foundation's conundrum? In any case, this seems to be an interesting moment where many in AI are trying to move away from Nvidia's proprietary CUDA platform. Most of them probably more for financial and availability reasons though, given the current GPU shortages[7] (which the ML team is undoubtedly aware of already; mentioning this as context for others on this list. See also Marketwatch's remarks about current margins[5]).

Regards, Tilman

[5] https://archive.ph/2023.06.15-173527/https://www.marketwatch.com/amp/story/o... [6] https://huggingface.co/blog/huggingface-and-amd [7] See e.g. https://gpus.llm-utils.org/nvidia-h100-gpus-supply-and-demand/ (avoid playing the song though. Don't say I didn't warn you)

> I wouldn't characterize WMF's Language Team using CPU as because of > AMD, rather at the time we didn't have the budget for GPUs so Lift Wing > didn't have any. Since then we have moved two GPUs onto Lift Wing for > testing but they are pretty old (2017ish). Once we make the big GPU > purchase Lift Wing will gain a lot of functionality for LLM and similar > models. > > Chris > > On Sun, Aug 6, 2023 at 9:57 PM Tilman Bayer haebwiki@gmail.com > wrote: > >> On Thu, Aug 3, 2023 at 7:16 AM Chris Albon calbon@wikimedia.org >> wrote: >> >>> Hi everybody, >>> >>> TL;DR We would like users of ORES models to migrate to our new >>> open source ML infrastructure, Lift Wing, within the next five months. We >>> are available to help you do that, from advice to making code commits. It >>> is important to note: All ML models currently accessible on ORES are also >>> currently accessible on Lift Wing. >>> >>> As part of the Machine Learning Modernization Project ( >>> https://www.mediawiki.org/wiki/Machine_Learning/Modernization), >>> the Machine Learning team has deployed a Wikimedia’s new machine learning >>> inference infrastructure, called Lift Wing ( >>> https://wikitech.wikimedia.org/wiki/Machine_Learning/LiftWing). >>> Lift Wing brings a lot of new features such as support for GPU-based >>> models, open source LLM hosting, auto-scaling, stability, and ability to >>> host a larger number of models. >>> >> >> This sounds quite exciting! What's the best place to read up on >> that planned support for GPU-based models and open source LLMs? (I also saw >> in the recent NYT article[1] that the team is "in the process of adapting >> A.I. models that are 'off the shelf; — essentially models that have been >> made available by researchers for anyone to freely customize — so that >> Wikipedia’s editors can use them for their work.") >> >> I'm aware of the history[2] of not being able to use NVIDIA >> GPUs due to their CUDA drivers being proprietary. It was mentioned recently >> in the Wikimedia AI Telegram group that this is still a serious limitation, >> despite some new explorations with AMD GPUs[3] - to the point that e.g. the >> WMF's Language team has resorted to using models without GPU support (CPU >> only).[4] >> It sounds like there is reasonable hope that this situation could >> change fairly soon? Would it also mean both at the same time, i.e. open >> source LLMs running with GPU support (considering that at least some >> well-known ones appear to require torch.cuda.is_available() == True for >> that)? >> >> Regards, Tilman >> >> [1] >> https://www.nytimes.com/2023/07/18/magazine/wikipedia-ai-chatgpt.html >> [2] >> https://techblog.wikimedia.org/2020/04/06/saying-no-to-proprietary-code-in-p... >> [3] https://phabricator.wikimedia.org/T334583 etc. >> [4] >> https://diff.wikimedia.org/2023/06/13/mint-supporting-underserved-languages-... >> or https://thottingal.in/blog/2023/07/21/wikiqa/ (experimental >> but, I understand, written to be deployable on WMF infrastructure) >> >> >>> >>> With the creation of Lift Wing, the team is turning its attention >>> to deprecating the current machine learning infrastructure, ORES. ORES >>> served us really well over the years, it was a successful project but it >>> came before radical changes in technology like Docker, Kubernetes and more >>> recently MLOps. The servers that run ORES are at the end of their planned >>> lifespan and so to save cost we are going to shut them down in early 2024. >>> >>> We have outlined a deprecation path on Wikitech ( >>> https://wikitech.wikimedia.org/wiki/ORES), please read the page >>> if you are a maintainer of a tool or code that uses the ORES endpoint >>> https://ores.wikimedia.org/). If you have any doubt or if you >>> need assistance in migrating to Lift Wing, feel free to contact the ML team >>> via: >>> >>> - Email: ml@wikimedia.org >>> - Phabricator: #Machine-Learning-Team tag >>> - IRC (Libera): #wikimedia-ml >>> >>> The Machine Learning team is available to help projects migrate, >>> from offering advice to making code commits. We want to make this as easy >>> as possible for folks. >>> >>> High Level timeline: >>> >>> **By September 30th 2023: *Infrastructure powering the ORES API >>> endpoint will be migrated from ORES to Lift Wing. For users, the API >>> endpoint will remain the same, and most users won’t notice any change. >>> Rather just the backend services powering the endpoint will change. >>> >>> Details: We'd like to add a DNS CNAME that points >>> ores.wikimedia.org to ores-legacy.wikimedia.org, a new endpoint >>> that offers a almost complete replacement of the ORES API calling Lift Wing >>> behind the scenes. In an ideal world we'd migrate all tools to Lift Wing >>> before decommissioning the infrastructure behind >>> ores.wikimedia.org, but it turned out to be really challenging so >>> to avoid disrupting users we chose to implement a transition layer/API. >>> >>> To summarize, if you don't have time to migrate before September >>> to Lift Wing, your code/tool should work just fine on >>> ores-legacy.wikimedia.org and you'll not have to change a line in >>> your code thanks to the DNS CNAME. The ores-legacy endpoint is not a 100% >>> replacement for ores, we removed some very old and not used features, so we >>> highly recommend at least test the new endpoint for your use case to avoid >>> surprises when we'll make the switch. In case you find anything weird, >>> please report it to us using the aforementioned channels. >>> >>> **September to January: *We will be reaching out to every user of >>> ORES we can identify and working with them to make the migration process as >>> easy as possible. >>> >>> **By January 2024: *If all goes well, we would like zero traffic >>> on the ORES API endpoint so we can turn off the ores-legacy API. >>> >>> If you want more information about Lift Wing, please check >>> https://wikitech.wikimedia.org/wiki/Machine_Learning/LiftWing >>> >>> Thanks in advance for the patience and the help! >>> >>> Regards, >>> >>> The Machine Learning Team >>> _______________________________________________ >>> Wikitech-l mailing list -- wikitech-l@lists.wikimedia.org >>> To unsubscribe send an email to >>> wikitech-l-leave@lists.wikimedia.org >>> >>> https://lists.wikimedia.org/postorius/lists/wikitech-l.lists.wikimedia.org/ >> >> _______________________________________________ >> Wikitech-l mailing list -- wikitech-l@lists.wikimedia.org >> To unsubscribe send an email to >> wikitech-l-leave@lists.wikimedia.org >> >> https://lists.wikimedia.org/postorius/lists/wikitech-l.lists.wikimedia.org/ > > _______________________________________________ > Wikitech-l mailing list -- wikitech-l@lists.wikimedia.org > To unsubscribe send an email to wikitech-l-leave@lists.wikimedia.org > > https://lists.wikimedia.org/postorius/lists/wikitech-l.lists.wikimedia.org/

---

Wikitech-l mailing list -- wikitech-l@lists.wikimedia.org To unsubscribe send an email to wikitech-l-leave@lists.wikimedia.org

https://lists.wikimedia.org/postorius/lists/wikitech-l.lists.wikimedia.org/

---

Wikitech-l mailing list -- wikitech-l@lists.wikimedia.org To unsubscribe send an email to wikitech-l-leave@lists.wikimedia.org

https://lists.wikimedia.org/postorius/lists/wikitech-l.lists.wikimedia.org/

---

Wikitech-l mailing list -- wikitech-l@lists.wikimedia.org To unsubscribe send an email to wikitech-l-leave@lists.wikimedia.org

https://lists.wikimedia.org/postorius/lists/wikitech-l.lists.wikimedia.org/

---

Wikitech-l mailing list -- wikitech-l@lists.wikimedia.org To unsubscribe send an email to wikitech-l-leave@lists.wikimedia.org

https://lists.wikimedia.org/postorius/lists/wikitech-l.lists.wikimedia.org/

---

Wikitech-l mailing list -- wikitech-l@lists.wikimedia.org To unsubscribe send an email to wikitech-l-leave@lists.wikimedia.org

https://lists.wikimedia.org/postorius/lists/wikitech-l.lists.wikimedia.org/

We could definitely file a task. However, it does seem like highlighting the features that will no longer be available is an appropriate topic for a discussion about migration in a technical mailing list.

Is there a good reference for which features have been excluded from ores-legacy? It looks like https://wikitech.wikimedia.org/wiki/ORES covers some of the excluded features/models, but not all of them.

I see now that it looks like the RevertRisk model will be replacing the *damaging *and *goodfaith *models that differentiate intentional damage from unintentional damage. There's a large body of research on why this is valuable and important to the social functioning of the wikis. This literature also discusses why being reverted is not a very good signal for damage/vandalism and can lead to problems when used as a signal for patrolling. Was there a community discussion about this deprecation that I missed? I have some preliminary results (in press) that demonstrate that the RevertRisk model performs significantly worse than the damaging and goodfaith models in English Wikipedia for patrolling work. Do you have documentation for how you evaluated this model and compared it to damaging/goodfaith?

FWIW, from my reading of these announcement threads, I believed that generally functionality and models would be preserved in ores-legacy/LiftWing. This is the first time I've realized the scale of what will become unavailable.

On Fri, Sep 22, 2023 at 9:07 AM Luca Toscano ltoscano@wikimedia.org wrote:

Let's discuss the issue in a Phabricator task, it seems more appropriate than here (so other folks can chime in etc.. more easily).

From our traffic analysis there is no current client using model_info, so we didn't add it to the feature set. We are working on an equivalent solution in Lift Wing for all hosted models, not only revscoring ones, but we don't have anything available now (a sort of "explainer" for the model's metadata basically).

Luca

On Fri, Sep 22, 2023 at 6:01 PM Aaron Halfaker aaron.halfaker@gmail.com wrote:

...

It looks like model_info is not implemented at all. E.g. https://ores-legacy.wikimedia.org/v3/scores/enwiki?model_info=statistics.thr...

I get {"detail":{"error":{"code":"bad request","message":"model_info query parameter is not supported by this endpoint anymore. For more information please visit https://wikitech.wikimedia.org/wiki/ORES%22%7D%7D%7D

But when I go to that page, nothing discusses model_info. Is there a way to get this from LiftWing?

On Fri, Sep 22, 2023 at 8:53 AM Aaron Halfaker aaron.halfaker@gmail.com wrote:

...

Do you have a tag for filing bugs against ORES-legacy? I can't seem to find a relevant one in phab.

On Fri, Sep 22, 2023 at 8:39 AM Luca Toscano ltoscano@wikimedia.org wrote:

...

Hi Aaron!

Thanks for following up. The API is almost compatible with what ORES currently does, but there are limitations (like the max number of revisions in a batch etc..). The API clearly states when something is not supported, so you can check its compatibility now making some requests to:

https://ores-legacy.wikimedia.org

If you open a task with a list of systems that you need to migrate we can definitely take a look and help. So far the traffic being served by ORES has been reduced to few clients, and all of them don't run with recognizable UAs (see https://meta.wikimedia.org/wiki/User-Agent_policy) so we'll try our best to support them. The migration to Lift Wing has been widely publicized, a lot of documentation is available to migrate. We'd suggest trying Lift Wing for your systems instead (see https://wikitech.wikimedia.org/wiki/Machine_Learning/LiftWing/Usage).

The Machine Learning plan is to eventually deprecate ores-legacy too, to maintain only one system (namely Lift Wing). There is no final date yet, we'll try to reach out to all remaining users first, so if you plan to keep using ores-legacy please follow up with us first :)

Thanks!

Luca (on behalf of the ML Team)

On Fri, Sep 22, 2023 at 5:10 PM Aaron Halfaker < aaron.halfaker@gmail.com> wrote:

...

Does the new ores-legacy support the same feature set. E.g. features output, injection, and threshold optimizations. Or is it just prediction? This will affect some of the systems I need to migrate.

On Fri, Sep 22, 2023, 06:21 Ilias Sarantopoulos < isarantopoulos@wikimedia.org> wrote:

...

Hello!

As a next step in the deprecation process of ORES https://wikitech.wikimedia.org/wiki/ORES the Machine Learning team will switch the backend of ores.wikimedia.org to ores-legacy, a k8s application meant to provide a compatibility layer between ORES and Lift Wing so users that have not yet migrated to Lift Wing will be transparently migrated. Ores-legacy is an application that has the same API as ORES but in the background makes requests to Lift Wing, allowing us to decommission the ORES servers until all clients have moved.

This change is planned to take place on Monday 25th of September. If you have a client/application that is still using ORES we expect that this switch is going to be transparent for you.

However keep in mind that ores-legacy is not a 100% replacement for ORES as some old and unused features are no longer supported.

If you see anything out of the ordinary, feel free to contact the Machine Learning team:

IRC libera: #wikimedia-ml

Phabricator: Machine-Learning-team tag

Thank you!

On Wed, Aug 9, 2023 at 1:22 PM Chaloemphon Praphuchakang < yoshrakpraphu@gmail.com> wrote:

> > On Tue, 8 Aug 2023, 10:45 Tilman Bayer, haebwiki@gmail.com wrote: > >> >> Hi Chris, >> >> On Mon, Aug 7, 2023 at 11:51 AM Chris Albon calbon@wikimedia.org >> wrote: >> >>> Hi Tilman, >>> >>> Most of the work is still very experimental. We have hosted a few >>> LLMs on Lift Wing already (StarCoder for example) but they were just >>> running on CPU, far too slow for real use cases. But it proves that we can >>> easily host LLMs on Lift Wing. We have been pretty quiet about it while we >>> focus on the ORES migration, but it is our next big project. More soon >>> hopefully! >>> >> Understood. Looking forward to learning more later! >> >> >>> Where we are now is that we have budget for a big GPU purchase >>> (~10-20 GPUs depending on cost), the question we will try to answer after >>> the ORES migration is complete is: what GPUs should we purchase? We are >>> trying to balance our strong preference to stay open source (i.e. AMD mROC) >>> in a world dominated by a single closed source vendor (i.e. Nvidia). In >>> addition, do we go for a few expensive GPUs better suited to LLMs (A1000, >>> H100, etc) or a mix of big and small? We will need to figure out all this. >>> >> I see. On that matter, what do you folks make of the recent >> announcements of AMD's partnerships with Hugging Face and Pytorch[5]? >> (which, I understand, came after the ML team had already launched the >> aforementioned new AMD explorations) >> >> "Open-source AI: AMD looks to Hugging Face and Meta spinoff PyTorch >> to take on Nvidia [...] >> Both partnerships involve AMD’s ROCm AI software stack, the >> company’s answer to Nvidia’s proprietary CUDA platform and >> application-programming interface. AMD called ROCm an open and portable AI >> system with out-of-the-box support that can port to existing AI models. >> [...B]oth AMD and Hugging Face are dedicating engineering resources to each >> other and sharing data to ensure that the constantly updated AI models from >> Hugging Face, which might not otherwise run well on AMD hardware, would be >> “guaranteed” to work on hardware like the MI300X. [...] AMD said PyTorch >> will fully upstream the ROCm software stack and “provide immediate ‘day >> zero’ support for PyTorch 2.0 with ROCm release 5.4.2 on all AMD Instinct >> accelerators,” which is meant to appeal to those customers looking to >> switch from Nvidia’s software ecosystem." >> >> >> In their own announcement, Hugging Face offered further details, >> including a pretty impressive list of models to be supported:[6] >> >> >> "We intend to support state-of-the-art transformer architectures >> for natural language processing, computer vision, and speech, such as BERT, >> DistilBERT, ROBERTA, Vision Transformer, CLIP, and Wav2Vec2. Of course, >> generative AI models will be available too (e.g., GPT2, GPT-NeoX, T5, OPT, >> LLaMA), including our own BLOOM and StarCoder models. Lastly, we will also >> support more traditional computer vision models, like ResNet and ResNext, >> and deep learning recommendation models, a first for us. [..] We'll do our >> best to test and validate these models for PyTorch, TensorFlow, and ONNX >> Runtime for the above platforms. [...] We will integrate the AMD ROCm SDK >> seamlessly in our open-source libraries, starting with the transformers >> library." >> >> >> Do you think this may promise too much, or could it point to a >> possible solution of the Foundation's conundrum? >> In any case, this seems to be an interesting moment where many in >> AI are trying to move away from Nvidia's proprietary CUDA platform. Most of >> them probably more for financial and availability reasons though, given the >> current GPU shortages[7] (which the ML team is undoubtedly aware of >> already; mentioning this as context for others on this list. See also >> Marketwatch's remarks about current margins[5]). >> >> Regards, Tilman >> >> >> [5] >> https://archive.ph/2023.06.15-173527/https://www.marketwatch.com/amp/story/o... >> [6] https://huggingface.co/blog/huggingface-and-amd >> [7] See e.g. >> https://gpus.llm-utils.org/nvidia-h100-gpus-supply-and-demand/ >> (avoid playing the song though. Don't say I didn't warn you) >> >> >>> I wouldn't characterize WMF's Language Team using CPU as because >>> of AMD, rather at the time we didn't have the budget for GPUs so Lift Wing >>> didn't have any. Since then we have moved two GPUs onto Lift Wing for >>> testing but they are pretty old (2017ish). Once we make the big GPU >>> purchase Lift Wing will gain a lot of functionality for LLM and similar >>> models. >>> >>> Chris >>> >>> On Sun, Aug 6, 2023 at 9:57 PM Tilman Bayer haebwiki@gmail.com >>> wrote: >>> >>>> On Thu, Aug 3, 2023 at 7:16 AM Chris Albon calbon@wikimedia.org >>>> wrote: >>>> >>>>> Hi everybody, >>>>> >>>>> TL;DR We would like users of ORES models to migrate to our new >>>>> open source ML infrastructure, Lift Wing, within the next five months. We >>>>> are available to help you do that, from advice to making code commits. It >>>>> is important to note: All ML models currently accessible on ORES are also >>>>> currently accessible on Lift Wing. >>>>> >>>>> As part of the Machine Learning Modernization Project ( >>>>> https://www.mediawiki.org/wiki/Machine_Learning/Modernization), >>>>> the Machine Learning team has deployed a Wikimedia’s new machine learning >>>>> inference infrastructure, called Lift Wing ( >>>>> https://wikitech.wikimedia.org/wiki/Machine_Learning/LiftWing). >>>>> Lift Wing brings a lot of new features such as support for GPU-based >>>>> models, open source LLM hosting, auto-scaling, stability, and ability to >>>>> host a larger number of models. >>>>> >>>> >>>> This sounds quite exciting! What's the best place to read up on >>>> that planned support for GPU-based models and open source LLMs? (I also saw >>>> in the recent NYT article[1] that the team is "in the process of adapting >>>> A.I. models that are 'off the shelf; — essentially models that have been >>>> made available by researchers for anyone to freely customize — so that >>>> Wikipedia’s editors can use them for their work.") >>>> >>>> I'm aware of the history[2] of not being able to use NVIDIA >>>> GPUs due to their CUDA drivers being proprietary. It was mentioned recently >>>> in the Wikimedia AI Telegram group that this is still a serious limitation, >>>> despite some new explorations with AMD GPUs[3] - to the point that e.g. the >>>> WMF's Language team has resorted to using models without GPU support (CPU >>>> only).[4] >>>> It sounds like there is reasonable hope that this situation could >>>> change fairly soon? Would it also mean both at the same time, i.e. open >>>> source LLMs running with GPU support (considering that at least some >>>> well-known ones appear to require torch.cuda.is_available() == True for >>>> that)? >>>> >>>> Regards, Tilman >>>> >>>> [1] >>>> https://www.nytimes.com/2023/07/18/magazine/wikipedia-ai-chatgpt.html >>>> [2] >>>> https://techblog.wikimedia.org/2020/04/06/saying-no-to-proprietary-code-in-p... >>>> [3] https://phabricator.wikimedia.org/T334583 etc. >>>> [4] >>>> https://diff.wikimedia.org/2023/06/13/mint-supporting-underserved-languages-... >>>> or https://thottingal.in/blog/2023/07/21/wikiqa/ (experimental >>>> but, I understand, written to be deployable on WMF infrastructure) >>>> >>>> >>>>> >>>>> With the creation of Lift Wing, the team is turning its >>>>> attention to deprecating the current machine learning infrastructure, ORES. >>>>> ORES served us really well over the years, it was a successful project but >>>>> it came before radical changes in technology like Docker, Kubernetes and >>>>> more recently MLOps. The servers that run ORES are at the end of their >>>>> planned lifespan and so to save cost we are going to shut them down in >>>>> early 2024. >>>>> >>>>> We have outlined a deprecation path on Wikitech ( >>>>> https://wikitech.wikimedia.org/wiki/ORES), please read the page >>>>> if you are a maintainer of a tool or code that uses the ORES endpoint >>>>> https://ores.wikimedia.org/). If you have any doubt or if you >>>>> need assistance in migrating to Lift Wing, feel free to contact the ML team >>>>> via: >>>>> >>>>> - Email: ml@wikimedia.org >>>>> - Phabricator: #Machine-Learning-Team tag >>>>> - IRC (Libera): #wikimedia-ml >>>>> >>>>> The Machine Learning team is available to help projects migrate, >>>>> from offering advice to making code commits. We want to make this as easy >>>>> as possible for folks. >>>>> >>>>> High Level timeline: >>>>> >>>>> **By September 30th 2023: *Infrastructure powering the ORES API >>>>> endpoint will be migrated from ORES to Lift Wing. For users, the API >>>>> endpoint will remain the same, and most users won’t notice any change. >>>>> Rather just the backend services powering the endpoint will change. >>>>> >>>>> Details: We'd like to add a DNS CNAME that points >>>>> ores.wikimedia.org to ores-legacy.wikimedia.org, a new endpoint >>>>> that offers a almost complete replacement of the ORES API calling Lift Wing >>>>> behind the scenes. In an ideal world we'd migrate all tools to Lift Wing >>>>> before decommissioning the infrastructure behind >>>>> ores.wikimedia.org, but it turned out to be really challenging >>>>> so to avoid disrupting users we chose to implement a transition layer/API. >>>>> >>>>> To summarize, if you don't have time to migrate before September >>>>> to Lift Wing, your code/tool should work just fine on >>>>> ores-legacy.wikimedia.org and you'll not have to change a line >>>>> in your code thanks to the DNS CNAME. The ores-legacy endpoint is not a >>>>> 100% replacement for ores, we removed some very old and not used features, >>>>> so we highly recommend at least test the new endpoint for your use case to >>>>> avoid surprises when we'll make the switch. In case you find anything >>>>> weird, please report it to us using the aforementioned channels. >>>>> >>>>> **September to January: *We will be reaching out to every user >>>>> of ORES we can identify and working with them to make the migration process >>>>> as easy as possible. >>>>> >>>>> **By January 2024: *If all goes well, we would like zero >>>>> traffic on the ORES API endpoint so we can turn off the ores-legacy API. >>>>> >>>>> If you want more information about Lift Wing, please check >>>>> https://wikitech.wikimedia.org/wiki/Machine_Learning/LiftWing >>>>> >>>>> Thanks in advance for the patience and the help! >>>>> >>>>> Regards, >>>>> >>>>> The Machine Learning Team >>>>> _______________________________________________ >>>>> Wikitech-l mailing list -- wikitech-l@lists.wikimedia.org >>>>> To unsubscribe send an email to >>>>> wikitech-l-leave@lists.wikimedia.org >>>>> >>>>> https://lists.wikimedia.org/postorius/lists/wikitech-l.lists.wikimedia.org/ >>>> >>>> _______________________________________________ >>>> Wikitech-l mailing list -- wikitech-l@lists.wikimedia.org >>>> To unsubscribe send an email to >>>> wikitech-l-leave@lists.wikimedia.org >>>> >>>> https://lists.wikimedia.org/postorius/lists/wikitech-l.lists.wikimedia.org/ >>> >>> _______________________________________________ >>> Wikitech-l mailing list -- wikitech-l@lists.wikimedia.org >>> To unsubscribe send an email to >>> wikitech-l-leave@lists.wikimedia.org >>> >>> https://lists.wikimedia.org/postorius/lists/wikitech-l.lists.wikimedia.org/ >> >> _______________________________________________ >> Wikitech-l mailing list -- wikitech-l@lists.wikimedia.org >> To unsubscribe send an email to >> wikitech-l-leave@lists.wikimedia.org >> >> https://lists.wikimedia.org/postorius/lists/wikitech-l.lists.wikimedia.org/ > > _______________________________________________ > Wikitech-l mailing list -- wikitech-l@lists.wikimedia.org > To unsubscribe send an email to wikitech-l-leave@lists.wikimedia.org > > https://lists.wikimedia.org/postorius/lists/wikitech-l.lists.wikimedia.org/

---

Wikitech-l mailing list -- wikitech-l@lists.wikimedia.org To unsubscribe send an email to wikitech-l-leave@lists.wikimedia.org

https://lists.wikimedia.org/postorius/lists/wikitech-l.lists.wikimedia.org/

---

Wikitech-l mailing list -- wikitech-l@lists.wikimedia.org To unsubscribe send an email to wikitech-l-leave@lists.wikimedia.org

https://lists.wikimedia.org/postorius/lists/wikitech-l.lists.wikimedia.org/

---

Wikitech-l mailing list -- wikitech-l@lists.wikimedia.org To unsubscribe send an email to wikitech-l-leave@lists.wikimedia.org

https://lists.wikimedia.org/postorius/lists/wikitech-l.lists.wikimedia.org/

---

Wikitech-l mailing list -- wikitech-l@lists.wikimedia.org To unsubscribe send an email to wikitech-l-leave@lists.wikimedia.org

https://lists.wikimedia.org/postorius/lists/wikitech-l.lists.wikimedia.org/

---

Wikitech-l mailing list -- wikitech-l@lists.wikimedia.org To unsubscribe send an email to wikitech-l-leave@lists.wikimedia.org https://lists.wikimedia.org/postorius/lists/wikitech-l.lists.wikimedia.org/

All fine points. As you can see, I've filed some phab tasks where I saw a clear opportunity to do so.

as mentioned before all the models that currently run on ORES are

available in both ores-legacy and Lift Wing.

I thought I read that damaging and goodfaith models are going to be replaced. Should I instead read that they are likely to remain available for the foreseeable future? When I asked about a community discussion about the transition from damaging/goodfaith to revertrisk, I was imagining that many people who use those predictions might have an opinion about them going away. E.g. people who use the relevant filters in RecentChanges. Maybe I missed the discussions about that.

I haven't seen a mention of the article quality or article topic models in the docs. Are those also going to remain available? I have some user scripts that use these models and are relatively widely used. I didn't notice anyone reaching out. ... So I checked and setting a User-Agent on my user scripts doesn't actually change the User-Agent. I've read that you need to set "Api-User-Agent" instead, but that causes a CORS error when querying ORES. I'll file a bug.

On Fri, Sep 22, 2023 at 1:22 PM Luca Toscano ltoscano@wikimedia.org wrote:

On Fri, Sep 22, 2023 at 8:59 PM Aaron Halfaker aaron.halfaker@gmail.com wrote:

...

We could definitely file a task. However, it does seem like highlighting the features that will no longer be available is an appropriate topic for a discussion about migration in a technical mailing list.

A specific question related to a functionality is the topic for a task, I don't think that we should discuss every detail that differs from the ORES API (Wikitech-l doesn't seem a good medium for it). We are already following up on Phabricator, let's use tasks if possible to keep the conversation as light and targeted as possible.

Is there a good reference for which features have been excluded from

...

ores-legacy? It looks like https://wikitech.wikimedia.org/wiki/ORES covers some of the excluded features/models, but not all of them.

We spent the last months helping the community to migrate away from the ORES API (to use Lift Wing instead), the remaining traffic is only related to few low traffic IPs that we are not able to contact. We didn't add feature injection or threshold optimization to ores-legacy, for example, since there was no indication on our logs that users were relying on it. We have always stated everywhere (including all emails sent in this mailing list) that we are 100% open to add a functionality if it is backed up by a valid use case.

...

I see now that it looks like the RevertRisk model will be replacing the *damaging *and *goodfaith *models that differentiate intentional damage from unintentional damage. There's a large body of research on why this is valuable and important to the social functioning of the wikis. This literature also discusses why being reverted is not a very good signal for damage/vandalism and can lead to problems when used as a signal for patrolling. Was there a community discussion about this deprecation that I missed? I have some preliminary results (in press) that demonstrate that the RevertRisk model performs significantly worse than the damaging and goodfaith models in English Wikipedia for patrolling work. Do you have documentation for how you evaluated this model and compared it to damaging/goodfaith?

We have model cards related to both Revert Risk models, all of them linked in the API portal docs (more info: https://api.wikimedia.org/wiki/Lift_Wing_API). All the community folks that migrated their bots/tools/etc.. to Revert Risk were very happy about the change, and we haven't had any request to switch back since then.

The ML team provides all the models deployed on ORES on Lift Wing, so any damaging and goodfaith variant is available in the new API. We chose to not pursue the development of those models for several reasons:

• We haven't had any indication/request from the community about those

models in almost two years, except few Phabricator updates that we followed up on.

• Managing several hundreds models for goodfaith and damaging is not very

scalable in a modern micro-service architecture like Lift Wing (since we have a model for each supported wiki). We (both Research and ML) are oriented on having fewer models that manage more languages at the same time, and this is the direction that we are following at the moment. It may not be the perfect one but so far it seems a good choice. If you want to chime in and provide your inputs we are 100% available in hearing suggestions/concerns/doubts/recommendations/etc.., please follow up in any of our channels (IRC, mailing lists, Phabricator for example).

• Last but not the least, most of the damaging/goodfaith models have been

trained with data coming from years ago, and never re-trained. The efforts to keep several hundreds models up-to-date with recent data versus doing the same of few models (like revert risk) weights in favor of the latter for a relatively small team of engineers like us.

...

FWIW, from my reading of these announcement threads, I believed that generally functionality and models would be preserved in ores-legacy/LiftWing. This is the first time I've realized the scale of what will become unavailable.

This is the part that I don't get, since as mentioned before all the models that currently run on ORES are available in both ores-legacy and Lift Wing. What changes is that we don't expose anymore functionality that logs clearly show are not used, and that would need to be maintained and improved over time. We are open to improve and add any requirement that the community needs, the only thing that we ask is to provide a valid use case to support it.

I do think that Lift Wing is a great improvement for the community, we have been working with all the folks that reached out to us, without hiding anything (including deprecation plans and path forwards).

Thanks for following up!

Luca _______________________________________________ Wikitech-l mailing list -- wikitech-l@lists.wikimedia.org To unsubscribe send an email to wikitech-l-leave@lists.wikimedia.org https://lists.wikimedia.org/postorius/lists/wikitech-l.lists.wikimedia.org/

Hey SJ!

Is there a reason to think that separate models for each wiki are more

effective than one general model that sees the name of the wiki as part of its context?

Intuitively one model per wiki has a lot of merit. The training data comes from the community that is impacted by the model, etc. However, there are scale and equity issues we wrestled with. One lesson we have learned training ~300+ models for the Add-A-Link https://www.mediawiki.org/wiki/Growth/Personalized_first_day/Structured_tasks/Add_a_link project is that if we continued down that path, Lift Wing would eventually be hosting 3000+ models (i.e. 330 models per new feature) pretty quickly and overwhelm any ability of our small team to maintain, quality control, support, and improve them over their lifespan. Regarding equity, even with a multi-year effort the one model per wiki RevScoring models only covered ~33 out of 330 wikis. The communities we didn't reach didn't get the benefit of those models. But with language agnostic models we can make that model available to all communities. For example, the language agnostic revert risk model will likely be the model selected for the automoderator project https://www.mediawiki.org/wiki/Moderator_Tools/Automoderator, which means hundreds more wikis get access to the tool compared to a one model per wiki approach.

I'd love to read more about the cost of training and updating current

models, how much material they are trained on, and how others w/ their own GPUs can contribute to updates.

The training data information should be available in the model cards https://meta.wikimedia.org/wiki/Machine_learning_models. If it isn't, let me know so we can change it. Regarding GPUs and contributions, we are still working on what a good training environment will be. Our initial idea was a kubeflow-based cluster we called Train Wing, but we've had to put them on hold for resource reasons (i.e. we couldn't build Lift Wing, deprecate ORES, and build Train Wing all at the same time). More on that soon after the Research-ML offsite when we'll have those conversations.

All this said, one thing we do want to support is hosting community created models. So, if a community has a model they want to host, we can load it into Lift Wing and host it for them at scale. We have a lot of details to work out (e.g. community consensus, human rights review as part of being a Very Large Online Platform etc.) as to what that would look like, but that is the goal.

Chris

On Fri, Sep 22, 2023 at 5:42 PM Samuel Klein meta.sj@gmail.com wrote:

Luca writes:

...

Managing several hundreds models for goodfaith and damaging is not very

scalable in a modern micro-service architecture like Lift Wing

...

(since we have a model for each supported wiki). We (both Research and

ML) are oriented on having fewer models that manage more languages at the same time,

Is there a reason to think that separate models for each wiki are more effective than one general model that sees the name of the wiki as part of its context? I'd love to read more about the cost of training and updating current models, how much material they are trained on, and how others w/ their own GPUs can contribute to updates.

Personally I wouldn't mind a single model that can suggest multiple properties of an edit, including goodfaith, damaging, and likelihood of reversion. They are different if related concepts -- the first deals with the intent and predicted further editing history of the editor, the second with article accuracy and quality, and the latter with the size + activity + norms of the other editors...

SJ

On Fri, Sep 22, 2023 at 5:34 PM Aaron Halfaker aaron.halfaker@gmail.com wrote:

...

All fine points. As you can see, I've filed some phab tasks where I saw a clear opportunity to do so.

...

as mentioned before all the models that currently run on ORES are

available in both ores-legacy and Lift Wing.

I thought I read that damaging and goodfaith models are going to be replaced. Should I instead read that they are likely to remain available for the foreseeable future? When I asked about a community discussion about the transition from damaging/goodfaith to revertrisk, I was imagining that many people who use those predictions might have an opinion about them going away. E.g. people who use the relevant filters in RecentChanges. Maybe I missed the discussions about that.

I haven't seen a mention of the article quality or article topic models in the docs. Are those also going to remain available? I have some user scripts that use these models and are relatively widely used. I didn't notice anyone reaching out. ... So I checked and setting a User-Agent on my user scripts doesn't actually change the User-Agent. I've read that you need to set "Api-User-Agent" instead, but that causes a CORS error when querying ORES. I'll file a bug.

On Fri, Sep 22, 2023 at 1:22 PM Luca Toscano ltoscano@wikimedia.org wrote:

...

On Fri, Sep 22, 2023 at 8:59 PM Aaron Halfaker aaron.halfaker@gmail.com wrote:

...

We could definitely file a task. However, it does seem like highlighting the features that will no longer be available is an appropriate topic for a discussion about migration in a technical mailing list.

A specific question related to a functionality is the topic for a task, I don't think that we should discuss every detail that differs from the ORES API (Wikitech-l doesn't seem a good medium for it). We are already following up on Phabricator, let's use tasks if possible to keep the conversation as light and targeted as possible.

Is there a good reference for which features have been excluded from

...

ores-legacy? It looks like https://wikitech.wikimedia.org/wiki/ORES covers some of the excluded features/models, but not all of them.

We spent the last months helping the community to migrate away from the ORES API (to use Lift Wing instead), the remaining traffic is only related to few low traffic IPs that we are not able to contact. We didn't add feature injection or threshold optimization to ores-legacy, for example, since there was no indication on our logs that users were relying on it. We have always stated everywhere (including all emails sent in this mailing list) that we are 100% open to add a functionality if it is backed up by a valid use case.

...

I see now that it looks like the RevertRisk model will be replacing the *damaging *and *goodfaith *models that differentiate intentional damage from unintentional damage. There's a large body of research on why this is valuable and important to the social functioning of the wikis. This literature also discusses why being reverted is not a very good signal for damage/vandalism and can lead to problems when used as a signal for patrolling. Was there a community discussion about this deprecation that I missed? I have some preliminary results (in press) that demonstrate that the RevertRisk model performs significantly worse than the damaging and goodfaith models in English Wikipedia for patrolling work. Do you have documentation for how you evaluated this model and compared it to damaging/goodfaith?

We have model cards related to both Revert Risk models, all of them linked in the API portal docs (more info: https://api.wikimedia.org/wiki/Lift_Wing_API). All the community folks that migrated their bots/tools/etc.. to Revert Risk were very happy about the change, and we haven't had any request to switch back since then.

The ML team provides all the models deployed on ORES on Lift Wing, so any damaging and goodfaith variant is available in the new API. We chose to not pursue the development of those models for several reasons:

• We haven't had any indication/request from the community about those

models in almost two years, except few Phabricator updates that we followed up on.

• Managing several hundreds models for goodfaith and damaging is not

very scalable in a modern micro-service architecture like Lift Wing (since we have a model for each supported wiki). We (both Research and ML) are oriented on having fewer models that manage more languages at the same time, and this is the direction that we are following at the moment. It may not be the perfect one but so far it seems a good choice. If you want to chime in and provide your inputs we are 100% available in hearing suggestions/concerns/doubts/recommendations/etc.., please follow up in any of our channels (IRC, mailing lists, Phabricator for example).

• Last but not the least, most of the damaging/goodfaith models have

been trained with data coming from years ago, and never re-trained. The efforts to keep several hundreds models up-to-date with recent data versus doing the same of few models (like revert risk) weights in favor of the latter for a relatively small team of engineers like us.

...

FWIW, from my reading of these announcement threads, I believed that generally functionality and models would be preserved in ores-legacy/LiftWing. This is the first time I've realized the scale of what will become unavailable.

This is the part that I don't get, since as mentioned before all the models that currently run on ORES are available in both ores-legacy and Lift Wing. What changes is that we don't expose anymore functionality that logs clearly show are not used, and that would need to be maintained and improved over time. We are open to improve and add any requirement that the community needs, the only thing that we ask is to provide a valid use case to support it.

I do think that Lift Wing is a great improvement for the community, we have been working with all the folks that reached out to us, without hiding anything (including deprecation plans and path forwards).

Thanks for following up!

Luca _______________________________________________ Wikitech-l mailing list -- wikitech-l@lists.wikimedia.org To unsubscribe send an email to wikitech-l-leave@lists.wikimedia.org

https://lists.wikimedia.org/postorius/lists/wikitech-l.lists.wikimedia.org/

---

Wikitech-l mailing list -- wikitech-l@lists.wikimedia.org To unsubscribe send an email to wikitech-l-leave@lists.wikimedia.org

https://lists.wikimedia.org/postorius/lists/wikitech-l.lists.wikimedia.org/

-- Samuel Klein @metasj w:user:sj +1 617 529 4266 _______________________________________________ Wikitech-l mailing list -- wikitech-l@lists.wikimedia.org To unsubscribe send an email to wikitech-l-leave@lists.wikimedia.org https://lists.wikimedia.org/postorius/lists/wikitech-l.lists.wikimedia.org/

See https://phabricator.wikimedia.org/T347344

On Mon, Sep 25, 2023 at 12:26 PM Aaron Halfaker aaron.halfaker@gmail.com wrote:

It looks like user-scripts running on Wikipedia can no longer use ORES. I'm getting a CORS error. You can test this by trying to run the following the JS dev console on a Wikimedia page:

...

$.ajax({url: "https://ores.wikimedia.org/v3/scores/

"}).done(function(response){console.log(response)})

This is what I see:

...

Access to XMLHttpRequest at 'https://ores.wikimedia.org/v3/scores/'

from origin 'https://en.wikipedia.org' has been blocked by CORS policy: No 'Access-Control-Allow-Origin' header is present on the requested resource.

...

GET https://ores.wikimedia.org/v3/scores/ net::ERR_FAILED 307

I'll file a bug, but I thought elevating this to the migration thread was a good idea.

On Mon, Sep 25, 2023 at 7:37 AM Chris Albon calbon@wikimedia.org wrote:

...

Hey SJ!

...

Is there a reason to think that separate models for each wiki are more

effective than one general model that sees the name of the wiki as part of its context?

Intuitively one model per wiki has a lot of merit. The training data comes from the community that is impacted by the model, etc. However, there are scale and equity issues we wrestled with. One lesson we have learned training ~300+ models for the Add-A-Link https://www.mediawiki.org/wiki/Growth/Personalized_first_day/Structured_tasks/Add_a_link project is that if we continued down that path, Lift Wing would eventually be hosting 3000+ models (i.e. 330 models per new feature) pretty quickly and overwhelm any ability of our small team to maintain, quality control, support, and improve them over their lifespan. Regarding equity, even with a multi-year effort the one model per wiki RevScoring models only covered ~33 out of 330 wikis. The communities we didn't reach didn't get the benefit of those models. But with language agnostic models we can make that model available to all communities. For example, the language agnostic revert risk model will likely be the model selected for the automoderator project https://www.mediawiki.org/wiki/Moderator_Tools/Automoderator, which means hundreds more wikis get access to the tool compared to a one model per wiki approach.

...

I'd love to read more about the cost of training and updating current

models, how much material they are trained on, and how others w/ their own GPUs can contribute to updates.

The training data information should be available in the model cards https://meta.wikimedia.org/wiki/Machine_learning_models. If it isn't, let me know so we can change it. Regarding GPUs and contributions, we are still working on what a good training environment will be. Our initial idea was a kubeflow-based cluster we called Train Wing, but we've had to put them on hold for resource reasons (i.e. we couldn't build Lift Wing, deprecate ORES, and build Train Wing all at the same time). More on that soon after the Research-ML offsite when we'll have those conversations.

All this said, one thing we do want to support is hosting community created models. So, if a community has a model they want to host, we can load it into Lift Wing and host it for them at scale. We have a lot of details to work out (e.g. community consensus, human rights review as part of being a Very Large Online Platform etc.) as to what that would look like, but that is the goal.

Chris

On Fri, Sep 22, 2023 at 5:42 PM Samuel Klein meta.sj@gmail.com wrote:

...

Luca writes:

...

Managing several hundreds models for goodfaith and damaging is not

very scalable in a modern micro-service architecture like Lift Wing

...

(since we have a model for each supported wiki). We (both Research

and ML) are oriented on having fewer models that manage more languages at the same time,

Is there a reason to think that separate models for each wiki are more effective than one general model that sees the name of the wiki as part of its context? I'd love to read more about the cost of training and updating current models, how much material they are trained on, and how others w/ their own GPUs can contribute to updates.

Personally I wouldn't mind a single model that can suggest multiple properties of an edit, including goodfaith, damaging, and likelihood of reversion. They are different if related concepts -- the first deals with the intent and predicted further editing history of the editor, the second with article accuracy and quality, and the latter with the size + activity + norms of the other editors...

SJ

On Fri, Sep 22, 2023 at 5:34 PM Aaron Halfaker aaron.halfaker@gmail.com wrote:

...

All fine points. As you can see, I've filed some phab tasks where I saw a clear opportunity to do so.

...

as mentioned before all the models that currently run on ORES are

available in both ores-legacy and Lift Wing.

I thought I read that damaging and goodfaith models are going to be replaced. Should I instead read that they are likely to remain available for the foreseeable future? When I asked about a community discussion about the transition from damaging/goodfaith to revertrisk, I was imagining that many people who use those predictions might have an opinion about them going away. E.g. people who use the relevant filters in RecentChanges. Maybe I missed the discussions about that.

I haven't seen a mention of the article quality or article topic models in the docs. Are those also going to remain available? I have some user scripts that use these models and are relatively widely used. I didn't notice anyone reaching out. ... So I checked and setting a User-Agent on my user scripts doesn't actually change the User-Agent. I've read that you need to set "Api-User-Agent" instead, but that causes a CORS error when querying ORES. I'll file a bug.

On Fri, Sep 22, 2023 at 1:22 PM Luca Toscano ltoscano@wikimedia.org wrote:

...

On Fri, Sep 22, 2023 at 8:59 PM Aaron Halfaker < aaron.halfaker@gmail.com> wrote:

...

We could definitely file a task. However, it does seem like highlighting the features that will no longer be available is an appropriate topic for a discussion about migration in a technical mailing list.

A specific question related to a functionality is the topic for a task, I don't think that we should discuss every detail that differs from the ORES API (Wikitech-l doesn't seem a good medium for it). We are already following up on Phabricator, let's use tasks if possible to keep the conversation as light and targeted as possible.

Is there a good reference for which features have been excluded from

...

ores-legacy? It looks like https://wikitech.wikimedia.org/wiki/ORES covers some of the excluded features/models, but not all of them.

We spent the last months helping the community to migrate away from the ORES API (to use Lift Wing instead), the remaining traffic is only related to few low traffic IPs that we are not able to contact. We didn't add feature injection or threshold optimization to ores-legacy, for example, since there was no indication on our logs that users were relying on it. We have always stated everywhere (including all emails sent in this mailing list) that we are 100% open to add a functionality if it is backed up by a valid use case.

...

I see now that it looks like the RevertRisk model will be replacing the *damaging *and *goodfaith *models that differentiate intentional damage from unintentional damage. There's a large body of research on why this is valuable and important to the social functioning of the wikis. This literature also discusses why being reverted is not a very good signal for damage/vandalism and can lead to problems when used as a signal for patrolling. Was there a community discussion about this deprecation that I missed? I have some preliminary results (in press) that demonstrate that the RevertRisk model performs significantly worse than the damaging and goodfaith models in English Wikipedia for patrolling work. Do you have documentation for how you evaluated this model and compared it to damaging/goodfaith?

We have model cards related to both Revert Risk models, all of them linked in the API portal docs (more info: https://api.wikimedia.org/wiki/Lift_Wing_API). All the community folks that migrated their bots/tools/etc.. to Revert Risk were very happy about the change, and we haven't had any request to switch back since then.

The ML team provides all the models deployed on ORES on Lift Wing, so any damaging and goodfaith variant is available in the new API. We chose to not pursue the development of those models for several reasons:

• We haven't had any indication/request from the community about those

models in almost two years, except few Phabricator updates that we followed up on.

• Managing several hundreds models for goodfaith and damaging is not

very scalable in a modern micro-service architecture like Lift Wing (since we have a model for each supported wiki). We (both Research and ML) are oriented on having fewer models that manage more languages at the same time, and this is the direction that we are following at the moment. It may not be the perfect one but so far it seems a good choice. If you want to chime in and provide your inputs we are 100% available in hearing suggestions/concerns/doubts/recommendations/etc.., please follow up in any of our channels (IRC, mailing lists, Phabricator for example).

• Last but not the least, most of the damaging/goodfaith models have

been trained with data coming from years ago, and never re-trained. The efforts to keep several hundreds models up-to-date with recent data versus doing the same of few models (like revert risk) weights in favor of the latter for a relatively small team of engineers like us.

...

FWIW, from my reading of these announcement threads, I believed that generally functionality and models would be preserved in ores-legacy/LiftWing. This is the first time I've realized the scale of what will become unavailable.

This is the part that I don't get, since as mentioned before all the models that currently run on ORES are available in both ores-legacy and Lift Wing. What changes is that we don't expose anymore functionality that logs clearly show are not used, and that would need to be maintained and improved over time. We are open to improve and add any requirement that the community needs, the only thing that we ask is to provide a valid use case to support it.

I do think that Lift Wing is a great improvement for the community, we have been working with all the folks that reached out to us, without hiding anything (including deprecation plans and path forwards).

Thanks for following up!

Luca _______________________________________________ Wikitech-l mailing list -- wikitech-l@lists.wikimedia.org To unsubscribe send an email to wikitech-l-leave@lists.wikimedia.org

https://lists.wikimedia.org/postorius/lists/wikitech-l.lists.wikimedia.org/

---

Wikitech-l mailing list -- wikitech-l@lists.wikimedia.org To unsubscribe send an email to wikitech-l-leave@lists.wikimedia.org

https://lists.wikimedia.org/postorius/lists/wikitech-l.lists.wikimedia.org/

-- Samuel Klein @metasj w:user:sj +1 617 529 4266 _______________________________________________ Wikitech-l mailing list -- wikitech-l@lists.wikimedia.org To unsubscribe send an email to wikitech-l-leave@lists.wikimedia.org

https://lists.wikimedia.org/postorius/lists/wikitech-l.lists.wikimedia.org/

---

Wikitech-l mailing list -- wikitech-l@lists.wikimedia.org To unsubscribe send an email to wikitech-l-leave@lists.wikimedia.org

https://lists.wikimedia.org/postorius/lists/wikitech-l.lists.wikimedia.org/

On Fri, Sep 22, 2023 at 11:34 PM Aaron Halfaker aaron.halfaker@gmail.com wrote:

All fine points. As you can see, I've filed some phab tasks where I saw a clear opportunity to do so.

Thanks a lot! We are going to review them next week and decide the next steps, but we'd like to proceed anyway to migrate ores to ores-legacy on Monday (this will allow us to free some old nodes that need to be decommed etc..). Adding features later on to the models on Lift Wing should be doable, and our goal is to transition away from ores-legacy in a few months (to avoid maintaining too many systems). The timeline is not yet set in stone, we'll update this mailing list when the time comes (and we'll follow up with the remaining users of ores-legacy as well). To summarize: we start with Ores -> Ores Legacy on Monday, and we'll do Ores Legacy -> Lift Wing in a second step.

as mentioned before all the models that currently run on ORES are available in both ores-legacy and Lift Wing.

I thought I read that damaging and goodfaith models are going to be replaced. Should I instead read that they are likely to remain available for the foreseeable future? When I asked about a community discussion about the transition from damaging/goodfaith to revertrisk, I was imagining that many people who use those predictions might have an opinion about them going away. E.g. people who use the relevant filters in RecentChanges. Maybe I missed the discussions about that.

This is a good point, I'll clarify the documentation on Wikitech. Until models are used we'll not remove them from Lift Wing, but we'll propose to use Revert Risk where it is suited since it is a model family on which we decided to invest time and efforts. Basic maintenance will be performed on the goodfaith/damaging/articlequality/etc.. models on Lift Wing, but we don't have (at the moment) any bandwidth to guarantee retraining or more complex workflows on them. This is why we used the term "deprecated" on Wikitech, but we need to specify what we mean to avoid confusion. Thanks for the feedback :)

I haven't seen a mention of the article quality or article topic models in the docs. Are those also going to remain available? I have some user scripts that use these models and are relatively widely used. I didn't notice anyone reaching out. ... So I checked and setting a User-Agent on my user scripts doesn't actually change the User-Agent. I've read that you need to set "Api-User-Agent" instead, but that causes a CORS error when querying ORES. I'll file a bug.

Will update the docs as well, as mentioned above we'll keep the current ORES models available on Lift Wing. Eventually new models will be proposed by Research and other teams (like Revert Risk), and at that point we (as ML team) will decide what recommendation to give. Nothing will be removed from Lift Wing if there are active users on it, but we'll certainly try to reduce the amount of models to maintain (based on common functionality etc..), so some changes will be proposed in the future.

Luca

Hi!

I'll leave the comments related to model architecture and behavior to others (more expert than me), I'd like to comment on the process/infrastructure parts :)

On Sat, Sep 23, 2023 at 9:03 PM Strainu strainu10@gmail.com wrote:

Hi folks,

So glad to see the old and new ML teams have an open discussion about this subject.

I understand that the team might prefer to have several tickets for different issues, but the discussion about the general approach to the different models is of interest to many people and is more easily digested on email. I would suggest to continue discussing the merits of the current strategy (and not necessarily of a model or another) on email.

I proposed Phabricator tasks because I think that they better target different broad subjects, it is easier to involve specific teams/people and to define the goal of the conversations. In this big email thread we started outlining the migration/deprecation of ORES in favor of Lift Wing, and now we are talking about model architectures and strategies to use for various use cases in the future. I really like the conversation, but if we wanted to be strict a new email thread (with a different subject) should be created, instead of mixing multiple subjects. People interested in the Lift Wing migration wouldn't be able to add comments, or if they did it would become difficult to follow all the discussions.

As stated before, I'll clarify the "deprecation" term mentioned in Wikitech for the various revscoring-based models, but it is not something that is related to the Lift Wing migration (since all models present on ORES are also on Lift Wing). It is a long term and wider project that will happen over the upcoming months/years, and that requires a broader discussion.

This is why I propose to discuss models on Phabricator, rather than Wikitech-l :)

On the long run, I believe an unique model good enough can be developed for revert bots. However, it would be great if there were some clear quality criteria that the community can verify and the old models are maintained for a wiki until we are sure the new model passes that criteria on that wiki.

Definitely, I just want to make it clear that the ML team has no intention to force any choice to the community, we are just trying to optimize our infrastructure to serve a wide variety of models and in the process we have to choose the best strategy to follow. On Lift Wing we require that every new model has a model card that explains how it works, how it was trained, best use cases, etc.. For example, these are the API Portal's pages for the two Revert Risk models (they contain the link to model cards): https://api.wikimedia.org/wiki/Lift_Wing_API/Reference/Get_reverted_risk_lan... https://api.wikimedia.org/wiki/Lift_Wing_API/Reference/Get_reverted_risk_mul...

A change in hosting should not be the guiding force in any team's roadmap, but the needs of its users.

I hope that we (as ML) didn't describe our intentions in the wrong way, since our aim is absolutely not to impose anything, but to improve our infrastructure to better serve users in the future (WMF internal use cases and the community). ORES served us well over the years, it was a pioneering project on a topic, ML, that was only discussed in Research papers and some futuristic set of libraries at the time. Big players were already working on it internally, but there was no clear guidance or standards, and over the years stuff like MLOps formed and nowadays they are the de facto standard to operate. We are trying to follow those best practices, because we are convinced that they will surely improve and ease the process to build and publish a model at the WMF. If you are curious, the ML team worked a lot on documentation, see https://wikitech.wikimedia.org/wiki/Machine_Learning/LiftWing. We tried as best as we could to make the transition smooth and to highlight new features and improvements for every user.

To summarize - we, as ML team, have created Lift Wing to serve the community and our internal use cases, and we wouldn't remove support or change dramatically how the community operates without a gradual migration path and proposing new solutions first. During the migration to Lift Wing we asked folks to test Revert Risk models, instead of goodfaith/damanging ones, and the solution seems to have suited a lot of use cases. Maybe in the future we'll have a mixture of specialized models for certain wikis, and more "multi-purpose" ones, but finding the right solution will surely involve community feedback and several tries.

Thanks for the feedback!

Luca