I'm a Principal Consultant with DesignMind, specializing in AI and Data solutions. I post videos here relating AI technology, especially as it relates to using AI in data analytics and software development.
This channel is primarily used to host video content included in my blog, so please visit the main blog site at robkerr.ai too!
First thing to look at when you get a 404 is whether the API Version parameter is still valid. The video I think uses a 2022 API version. Check what the current one is and use that in the parameters. Likely that will be the root cause.
Thanks for this.... although had one problem that I finally found a solution to. Currently you need to speicify pytorch=2.2.0 otherwise you get an error due to new torch libraries not being compatible. ( github.com/unslothai/unsloth/issues/73#issuecomment-1998826713 )
Hi Rob, thank you very much for the tutorial! I followed every single of you steps. I am stuck in the inference part. I am getting the following KeyError. line: outputs = model.generate(input_ids = inputs, max_new_tokens = 64, use_cache = True) KeyError: 'Cache only has 0 layers, attempted to access layer with index 0'
Super helpful and straightforward, thank you! I'm just getting started with Fabric - would love to see some content on managing libraries across environments (dev,test,prod) and also your own style of splitting a Fabric platform over multiple workspaces and into various lakehouses and warehouses!
You could, but you would need to deploy that LLM somewhere that you can reach it from Python. Azure OpenAI Services is hosting the model compute (and you pay per token). You can also pay per token by deploying models in Azure ML. You can also host the LLM(s) in other services on other cloud platforms and send them requests from a notebook in Fabric.
is it possible for the user to trigger the run of the Fabric notebook directly from the PowerBI report and then see the output of the notebook (for example, a feature importance chart) again in the PowerBI report?
This really helped. I couldn't find a good video that showed me how to do this all weekend, but I would recommend not cutting out scenes because people are looking at the video and following and literally a glace left and right, and you already did another step that I didn't even notice. Had to watch it like 4 times super slow to catch all the steps. Also I dont know linux so skipping the part of making the code/finetune folders had me thinking i missed something again and that had me turned around for awhile. Also when people stop it, the red bar and play, pause, buttons, subtitles etc cover up the bottom of the screen so we cant even see if you do something quick at the bottom because its covered up when we pause it. Thanks again.
Thanks for the feedback! Sorry you got tripped up in some places, I'll work harder to make sure the flow is clean and doesn't get interrupted in the future.
Thanks Rob! I have 2 questions: 1. Can I create a ticketing system in Copilot Studio? 2. When I publish via Teams to whole office of 60+ colleagues, they dont have to pay extra, right? Many Thanks.
Possibly! It depends on the complexity you need. I'd suggest trying to prototype the functionality you need to see how it can fit. Regarding cost, when using copilot studio, it's billed based on the number of messages that flow through the copilot app you create, so I don't think you would see per user charges for Teams users. Take a look at: www.microsoft.com/en-us/microsoft-copilot/microsoft-copilot-studio#Pricing
Hey Rob, how much time it takes to spin up the Spark cluster when running the new job? Or did you already have a cluster running and when you ran the job, the job simply requested resources from that running Spark Cluster?
If you use the default pool (I almost always do) a cluster will start processing Fabric jobs in 4-5 seconds since Fabric has cluster nodes already running and waiting to be assigned.
@@robkerrai I never had a chance to with Fabric but by default pool, do you mean that a cluster is always up and running in your environment and your job simply takes the resources from there. Or do you mean that default pool is Serverless and it will only charge me for the time I will use it and also, it will spin up within 4-5 seconds? That would be extremely helpful if true. We are working with HD Insight and need to have our nodes running all the time only because we can't afford a time of 10-15 minutes for a node/cluster to spin up when a new user starts his Notebook.
Yes to both questions. Fabric is SAAS (not PAAS), so it works differently in that you pay for a "capacity" that you choose (your peak usage/second), and is a relatively fixed cost rather than clusters you manage yourself billed as a variable cost. The Default pool is a large group of always running nodes waiting to be assigned to your work as needed--you don't pay for compute per-se, you pay for the maximum nodes you can use at any given time. Default pool probably would solve the low latency of starting clusters issue you have. It is possible in Fabric to design your own cluster configuration, which would probably have the same startup delay in a PAAS environment, but needing to do that is the exception not the rule. The default pool meets most general computing needs.
Nice Share. Is there a way to decrypt the wheel file so we can see .py scripts in locally that is obtained from elsewhere. Would you have some resources on this?
Hey Rob, this was really helpful. I was looking to apply information from a form like this onto a prebuilt data entry tool with certain sections for each of the matching sections on the form. Would this be possible?Could you possibly make a video on this? Thanks
It's a good question. I see the benefit with this feature is to view Power BI output within a notebook during development. For example, if the data in a Direct Lake model is being updated in the notebook, the visual is available within the notebook context.
Hi Rob, thanks for this video. How would you go about separate lakehouses for dev and prod, in relation to the notebook you sync? I tried doing exactly what you did here, but then in prod the notebook is trying to talk to the dev lakehouse. It seems the default lakehouse for the notebook is not put in git?
Microsoft's vision as presented at FabCon 24 was that different branches in a CI/CD environment are connected to different workspaces. The git features as I write this is pretty basic, but the design set out at FabCon is moving toward a more automated approach where creating a branch creates a workspace with all artifacts. In that scenario you probably would have a lakehouse in the Dev workspace, another lakehouse (with the same name) in a Prod lakehouse, etc.
This video has been fantastic, but I have a possibly stupid question: When you ran the code, you automatically had a job created; what did you do to get that? When I run mine that does not happen.
Hi Rob, great video about how to use packages on Fabric. In fact this video make think, that we can create a new Workspace and Lakehouse called SourceCode o SharedCode as the example. So we will have the code isolated from the data Bronze, Silver and Gold.
I think you could accomplish this using Fabric environments. Each environment can have different .whl files attached to it, so when you choose an environment to run your code, that code will use whatever .whl files are pre-installed in the environment.
thanks for sharing your video. is it possible to share the Retail Analysis Power BI dataset along with the measures so we can replicate the same at our side.
Thanks, this has been really useful. One issue I found was that the model was showing as ‘resource not found’ on the initial post request. I had to change the ‘project settings’ to ‘endpoint and key’ and then create a new model. Not sure if I had done something wrong but maybe other people have been puzzled by this too! Thanks again for the video, really helpful!
Thanks for this demo. It was very useful and clear. I think that the only bad point on azureml endpoint, is that you can add only one model on a single deployment, and it become very expensive in terms of money, because you need a single VM for each model. Do you know some alternatives on azureml? Thank again Gabriele