Skip to content
cllm

← use cases · Kernel & perf engineers

Kernel & performance research

A small, GDB-attachable ring-0 codebase where you can experiment with inference optimizations without negotiating with a host kernel.

The problem

Applying low-level optimizations to ML serving usually means fighting a general-purpose kernel: its allocator, its scheduler, its I/O paths. Researchers want a clean substrate where the only code running is the code they are studying.

Why cllm fits

  • make run-debug pauses QEMU on :1234 for a GDB attach — step through ring-0 inference code directly.
  • A few thousand lines total: the boot path, allocator, driver, HTTP server, and API are all readable.
  • The roadmap explicitly targets porting vLLM-style optimizations (continuous batching, kernel fusion) into the bare-metal path.
  • No host kernel between your changes and the hardware.

Honest caveats

  • This is research-grade infrastructure, not a production serving stack.
  • The heavy optimizations (PagedAttention, tensor parallelism) are out of scope; the goal is a substrate to prototype on.
  • You are building on a skeleton — the inference engine is the thing being integrated, not a given.