Accelerating LLMs with TornadoVM: From GPU Kernels to Model Inference

tornadovm as a Java parallel framework for accelerating data parallelization on GPUs and other hardware, first GPU experiences with ELSA Winner and Voodoo cards, explanation of TornadoVM as a plugin to existing JDKs that uses Graal as a library, TornadoVM's programming model with @parallel and @reduce annotations for parallelizable code, introduction of kernel API for lower-level GPU programming, TornadoVM's ability to dynamically reconfigure and select the best hardware for workloads, implementation of LLM inference acceleration with TornadoVM, challenges in accelerating Llama models on GPUs, introduction of tensor types in TornadoVM to support FP8 and FP16 operations, shared buffer capabilities for GPU memory management, comparison of Java Vector API performance versus GPU acceleration, discussion of model quantization as a potential use case for TornadoVM, exploration of Deep Java Library (DJL) and its ND array implementation, potential standardization of tensor types in Java, integration possibilities with Project Babylon and its Code Reflection capabilities, TornadoVM's execution plans and task graphs for defining accelerated workloads, ability to run on multiple GPUs with different backends simultaneously, potential enterprise applications for LLMs in Java including model distillation for domain-specific models, discussion of Foreign Function & Memory API integration in TornadoVM, performance comparison between different GPU backends like OpenCL and CUDA, collaboration with Intel Level Zero oneAPI and integrated graphics support, future plans for RISC-V support in TornadoVM

Juan Fumero on twitter: @snatverk