Scala and Spark in the New Deep Learning Era

on YouTube ↗40m

" summary

Demonstration converts Hugging Face Roberta models (base and a 400M-parameter variant from Barcelona Supercomputing Center) into ONNX from Hugging Face and registers them as Spark Estimators, fine-tunes with BigDL on a Spark-on-Kubernetes cluster using Scala 3.3 + Spark 3.2 and JVM-native BLAS; the pipeline runs but demands large heaps (example: batch 128 needs ~300 GB across three nodes), uses synchronous weight sync, and shows LoRA reduces trainable params to ~1%, code and experiments are public as a working proof of concept.

AI-written summary · about

" chapters (10)

1. 0:00Introduction to AI trends
2. 2:30Integration challenges
3. 4:48Core AI concepts
4. 9:31Transformer architecture
5. 12:33Implementation strategy
6. 17:02Model conversion and setup
7. 20:43BigDL and infrastructure
8. 23:48Coding the graph
9. 30:37Drawbacks and optimization
10. 34:29Advanced optimization techniques

" related talks

• 04mYuhao Yang InterviewAlexy Khrabrov, Yuhao Yang · Scale By the Bay 2017
• 22mBringing Deep Learning into the Big Data AnalyticsYuhao Yang · Scale By the Bay 2017
• 39mMachine Learning with Scala on SparkJose Quesada · Scala Days Berlin 2016
• 23mRethinking scalable machine learning with Spark ecosystemAdi Polak · Scale By the Bay 2021
• 54mDeep Learning with Apache SparkFavio Vázquez · ODSC India 2018
• 40mScaling From Research to Production with Skymind DL4J and ScalNetAri Kamlani · Scale By the Bay 2017
• 45mDeep Learning and NLP with SparkMelanie Warrick, Andy Petrella · Scala Days Berlin 2016
• 47mScala: The Unpredicted Lingua Franca for Data ScienceDean Wampler, Andy Petrella · Scala Days Berlin 2016