Stop Renting Intelligence: The Train-to-Deploy Loop for Specialized AI — Fireworks AI

AI Engineer · 19 min · 139 sentences · from YouTube's caption track

Each timecode opens YouTube at the start of that sentence. Line anchors (#s42) are the cue ids in the WebVTT, and every line carries its start and end seconds. All transcripts has every talk, and the whole corpus as one file.

00:12Hi everyone.
00:13Uh before we get started, I actually have a quick question for all of you.
00:18How many of you actually today use a closed source model like chat GPD or Anthropic?
00:24Okay.
00:25How many of you feel like it's getting really expensive really quickly?
00:30Okay, that's that's a good percentage of you all.
00:34And uh just like all of you might be thinking of switching from closed source model to open source models for cost reasons, performance reasons, there uh other customers and users that are going through the same journey.
00:49And today I'm going to talk about where these open source models shine and how to make the performance or quality of these open source models match the closed source models itself.
01:02And before I get into that quick introduction of myself uh I work at fireworks.
01:07I lead part of the applied machine learning team.
01:10My name is Jatashri but feel free to call me Jet if that's easier.
01:14Um let's talk about the whole train to deploy loop that that the open source models are taking today when users are switching from closed source to open source models.
01:25But even before we get into that let's talk about some fundamental truths that are true today in the AI ecosystem.
01:33So the three fundamental truths that are true uh are one open source models for the most part are production ready.
01:40If any of you have used GLM 5.2 two, you know that that is true especially for the coding use cases.
01:48Two, if open source models don't like beat or match the quality of close source models, the differentiating factor uh comes from how you customize, train and improve these open source models to match the performance and quality of closed source models.
02:04And three, the AI stack itself for most of our customers and users is shifting from let's just serve a model once once and that's done to a more continuous
02:17loop that is improving from training to inferencing uh the same open source model itself.
02:24So that's where fireworks kind of comes in.
02:27It makes the whole process of you having to go through the continuous journey of training these open source models as well as hosting these open source models at scale much more easier and also much more cost effective as well.
02:39And I'll get into how we actually do the training itself and different flavors of training that we offer towards the end of the presentation.
02:46But I did also want to quickly talk about inference before we go to the training portion of it which has become really popular among our customers and user base.
02:55And before we also talk about inference and training, I just wanted to give you a quick overview of how uh inference looks like especially at scales and what are the kind of production impact that we have seen among some of our customers and user base itself.
03:12So today at in production at scale uh we serve about 30 trillion tokens and it's increasing.
03:19Um this number was actually preGLM 5.2 release.
03:22And I'm sure this has pretty much kind of increased by about at least 5 trillion tokens a day.
03:28Now with the GLM 5.2 release, we support more than 10,000 companies globally and process about 200k requests a particular second itself.
03:39And the kind of production impact that we are seeing within our customer and user base is very significant.
03:46So cursor uh if for those of you who already use cursor and cursor models u the composer 2 model for cursor was trained and also hosted on our platform today
03:57and um it is able to achieve 13x faster inference because of the fire attention engine that we have and I'll speak a little bit about that in the next slide as well.
04:08Uh Versell is also another big customers uh of ours.
04:12We help them actually fine-tune and also host the model on our platform and they see a 40x latency improvement.
04:20And another big use case study is actually genpark.
04:23Um genpark lovables are in very similar field where they actually let you design systems with just prompts itself.
04:31And uh quality is a big bottleneck or was a big bottleneck for them when they were trying to switch from closed source to open source models.
04:40and we helped them fine-tune.
04:41We did uh RFD for them.
04:43We helped them clean the data set, define evaluator functions and at the end of the RFT training job itself, we saw that there was 12% quality improvement for the Gen Spark use case.
04:55Um also we are a trusted platform among all our other customers and user base as well and we have raised around 327 million CDC uh a while ago and we are about 200 people working at fireworks.
05:11That brings me to the next portion of it.
05:13I know in the previous slide I did mention that we made latency faster for a couple of our customers such as versell and how we do it is quite simple
05:23at a at a higher level.
05:25Um at fireworks we have our own inference engine called fire attention.
05:30So you have a model you have GPUs and the way the model sits on top of the GPUs is through a translation layer called CUDA kernels and as part of fire attention we build our own CUDA kernel.
05:41So when a customer comes to us uh with a workload and a performance target in mind then we can optimize this particular inference engine and the whole AI stack including that CUDA kernel for that specific use case.
05:54So most of the time customer like versel in the previous slide come to us and tell us uh they have x input tokens y output tokens uh they have zprompt to uh prompt cache hit rate and they want to meet so and so latency targets
06:08and we help them optimize this whole inference stack including that cuda kernel so that we can serve this model uh at a cheaper price but also at a that can meet their performance target and for their specific workload itself.
06:22uh as part of fire attention and inference optimization, we do also implement the very well-known techniques in the AI field which is things like speculative decoding, disagregation, quantization.
06:36In terms of speculative decoding itself, we also allow customers to train their draft models on our platform.
06:43Uh so they can even upload the data set very similar to how they do for fine-tuning for uh training a draft model.
06:50uh which helps them uh make the model the main model uh perform much faster uh meet latency requirements also gives you better throughput if the draft model is trained uh very well
07:04and um so once people actually people actually get to this stage mainly once they're happy with model quality but when customers are switching from closed source to open source models
07:16the biggest bottleneck or the gap is quality and the The way we we try to solve for it is through our training product and our training product itself comes in three different flavors for different segments
07:30uh within our customer and user base.
07:32One being the training agent.
07:35This is very popular among our product managers and builders that get that got into this AI inferencing and training field very recently.
07:45So the agent can be accessed through our UI and uh you can you you just need to link it the data set that you have uploaded uh on the fireworks platform as well as description on what you're trying to do as part of this training run itself.
08:00Uh it will even let you choose the model.
08:02It will clean the data set for you.
08:04It will also help you do the hyperparameter uh fine-tuning itself because it's not always easy to get the right uh hyperparameters itself.
08:14um when you are doing training for your specific workload.
08:18The next flavor also one of our most popular flavors among our customer and user base is our managed training and this is what most of our customers use today.
08:29They can use our managed training platform one through API and second through the UI itself.
08:37It is as simple as you inputting the hyperparameters that you think will be good for your model.
08:43choosing a model, uploading the data set itself and then clicking a button.
08:47That's how simple we make it on uh fireworks for the manage training itself.
08:52And uh I know one challenge a lot of people have is observability once the training run itself kicks in.
08:59So we also have a good UI dashboard that shows you how the training uh is happening, how well your model is uh learning at each epoch, what does the loss function look like.
09:09Additionally, you can also plug in u your weights and biases API key into the manage training service itself uh and export some of the observability metrics to platforms like weights and biases if you're already using that.
09:24And the third and the kind of the most advanced flavor of training that we offer is the training API itself.
09:32And the training API is very popular among some uh of the ML engineers within uh different organizations or within our customer base.
09:41And in this in this particular this particular offering of the product lets uh ML engineers define their own training loop and loss functions.
09:52And this becomes important when you want to customize your training job according to your needs.
09:57that is that goes well above beyond the managed training service that we offer itself and these these different uh flavors or the products within the training ecosystem that we offered
10:11are offered for supervised finetuning reinforcement fine-tuning and even DPO as well.
10:18Um nowadays I think RFT has become extremely extremely popular and if you ever need help in helping define a good evaluator function then we have worked with customers like Genpark
10:30that I was talking to you about in the previous slide to help them define an evaluator function clean the data set um and also choose the right hyperparameters especially for RFD.
10:41Unfortunately the training agent itself does not support RF.
10:45So it does not uh help you clean the data set or even do the hyperparameter tweaking but we are more than happy to help you do it if you have a big use case in mind.
10:55So now that we know all about the different um training offerings that we have I also wanted to show you a quick demo of each of these the training agent manage training and the training API itself.
11:07Uh but before we get into that I just wanted to close off by saying how the data flywheel looks like today on our platform.
11:15I know in the first slide I was talking to you about the realities of AI today where most people prefer doing a continuous loop where they train host train host as they get new and new data and feedback from their users
11:28and that is what we are trying to make more seamless on our platform and that's also why we have different training products different offerings as well as making it easier
11:38uh for you to host and make it also cheaper and faster for you to host those models at scale itself.
11:44And the way this flywheel works is kind of very simple.
11:48Once these uh models are trained, I will show you uh a demo in the next slide where the models get trained.
11:55It's it's you just are one click away from deploying these models into production.
12:00And once you've deployed, you can uh monitor.
12:02We have u monitoring metrics for the number of tokens, cash it rate, how your speculator is working, if you've trained a speculator.
12:12uh we help you collect gather uh feedback from all the all the metrics that you have collected and if you choose to also retrain on the same platform from the end from the checkpoint that you were using to a new checkpoint itself.
12:26So some of some capabilities on the platform that makes these data flywheel possible is one definitely having a host of uh GPUs.
12:35We do have we do partner with different cloud and neocloud providers which makes it easier for us to provide compute and also host these models at scale.
12:44Another thing is we offer is hotloading and this has become really popular especially when we were training models for cursor and this was actually a feature we implemented specifically for cursor and then we thought would be useful for general public while they were training models as well.
13:00uh what hot loading essentially means is when you have a new checkpoint uh you you can directly go to live deployment without any downtime itself.
13:09So if you have a deployment and you've gone through this continuous loop and trained a new checkpoint then you can just click a button and swap out the weights itself that you were using
13:19uh uh previously with the new endpoints.
13:22Another thing that makes it uh easier to train on our platform is the numerical fidelity.
13:28So a lot of the time customers use platform A for training and host on platform B. But what happens when you do that is you'll have issues in quality of the model.
13:38And the reason you have those issues is humix is not always compatible when you train on a different platform and host on a different platform.
13:46So that's why we offer training and inferencing on the same platform and ensure that uh you won't find any numerical differences after you finish training and once you decide to host this particular model on the platform itself.
14:00And um one other thing that that makes it more seamless is because training and inference is on the same platform.
14:06You're just one click away from uh deploying this model at production scale.
14:12Now that we have learned all about uh the inference and training, I just wanted to show you quick demo of different training products and how they work on our platform.
14:22The first portion of the demo is showing you how you can use the training agent itself.
14:27Over here you can uh input where your training data set lives within the fireworks what it's called.
14:34Uh you can also uh give it the eval data set itself.
14:40uh I know people have the train and eval or a test data set separately.
14:45So you give the mo you give the agent both the data sets and then you ask it uh to fine-tune and even select the hyperparameters and the model for you.
14:54The the agent then goes on and tells you the step that it will it will take to accomplish the task that you have given it.
15:02So it it details the SFT plan.
15:05It tells you why it selected the base model.
15:07In this case, it's selected quen 38B.
15:11And it'll also give you some information on the hyperparameters and what the whole training process can look like at a high level.
15:19And once you're done uh once you are uh you feel okay about the output that the agent has given, you can approve whatever the agent has uh the a the agent has outlined.
15:28And once you approve it, it will actually go through different steps and keep you posted on if it finishes step two, what does uh it look like?
15:37If it successfully finishes step three, what does that look like now?
15:40And at the end of the whole training run, it will also give you a summary of uh the config, the base model um and uh different hyperparameters and how that impacts it as well.
15:59And once it's once the whole model training is done, it actually gives you metrics such as accuracy and F1 score.
16:06Uh and it also summarizes kind of resummarizes and gives you an overview of everything that uh that it did.
16:13And one cool feature is it also gives you recommended next step as part of u the training run itself.
16:20Now the next portion of uh this demo actually talks about how we can use our uh manage training platform.
16:28I did mention that you just have to select the model data set and the hyperparameters and then you're just a button click away from actually training a model.
16:37Say you've uploaded uh the data set, chosen a model.
16:41Uh all you'll have to do is then uh go to the go to the u evaluator settings, choose the evaluator data set and also you get to choose the hyperparameter
16:52as part of optional settings if you decide to have a more fine grained hyperparameters for your uh training run itself.
17:00And once you're happy with all the configurations including the model and the data set itself, you just need to create uh you just need to click on the button that says create.
17:11Um and as soon as you click on the button that says create, a training job then kick kicks out.
17:18And you can see that uh the UI looks similar to this when a training job actually gets completed.
17:26And if you click on that particular training job itself, it'll give you information on the data set it used, the hyperparameters it used, what the model loss itself looks like when the model was training.
17:38Um, and then you can even clone these training runs and reproduce it with different settings and configurations itself.
17:47And once you're happy with the fine-tuned model, you can go to the deployment section, choose the custom model that just got uh trained on our platform, and then click on deploy once you're happy with the deployment configurations.
18:01Or you can choose a set of deployment configurations that we have such as throughput, fast, minimal.
18:09uh and this is uh kind of a template that that works well for generic workloads that we have optimized based on throughput latency uh time to first token and other such metrics
18:23and uh I unfortunately don't have an entire demo for our training API because the setup itself differs from customer to customer but we do have a lot of cookbooks on our websites that you can see and follow especially if you're interested in
18:39customizing your own training loops and loss functions itself.
18:44Uh and that's all I have for you for today.
18:47I just wanted to give you a brief overview of how the whole training to inferencing loop looks like.
18:52Thank you very much.
18:53And if you have any questions, I'm happy to take them in the next minute or so that we have.