Parallelism Explained

Parallelism Explained

Technique	Used in this project?	Meaning
Data / batch parallel inference	Yes	Independent input batches go to independent full model replicas
multiprocessing with spawn	Yes	Worker process per chosen device
Dynamic task queue	Yes	Available worker picks the next batch
DistributedDataParallel (DDP) training	No	Training workers synchronize gradients across replicas
Tensor / model parallelism	No	Slices a single model's weights/compute across GPUs
Pipeline parallelism	No	Splits model layers/stages across accelerators
HTTP serving / vLLM	No	Exposes a service with a scheduler and request endpoints

An unrelated uploaded notebook in legacy/parallel_llm_inference_colab.ipynb does demonstrate DDP training. It is retained as original material but does not implement the batch inference described here.