Technology Choices & Exhaustive Comparisons

Serving Engine Architecture: Custom Scheduler vs vLLM vs TGI vs TensorRT-LLM

Factor	Custom Scheduler [This Project]	vLLM	HuggingFace TGI	TensorRT-LLM
KV Cache Management	Paged KV-Cache (Virtual Memory)	PagedAttention	Paged Attention	Paged KV Cache
Batching Mechanism	Iteration-Level Continuous	Iteration-Level Continuous	Continuous Batching	In-Flight Batching
Inspectability & Control	100% Transparent Python Control Plane	Complex C++/CUDA Kernels	Rust Core + Python	Highly Proprietary C++
Portability	Runs on CPU & GPU (Simulated/Real)	GPU Only (CUDA/ROCm)	GPU Centric	NVIDIA CUDA Only

• Why This Architecture: Designed as a pure, inspectable systems control plane that models and executes the complex state machines of continuous batching, memory page allocation, and multi-replica routing with 99 verified automated unit tests. [CONFIRM WITH DEVELOPER]