ram-coffers
LLM infrastructure cost reduction via NUMA-aware weight banking: 147 t/s (8.8x stock llama.cpp) on refurbished enterprise POWER8. Self-hosted inference, no cloud APIs. Part of the Proof of Physical AI stack.
- 170
- Python
- AGPL-3.0
- 2 d ago
LLM infrastructure cost reduction via NUMA-aware weight banking: 147 t/s (8.8x stock llama.cpp) on refurbished enterprise POWER8. Self-hosted inference, no cloud APIs. Part of the Proof of Physical AI stack.
XTU = X Transformers Unity — hybrid (CPU-expert + GPU) inference acceleration for vLLM: mainline plugin + CPU MoE engine (AVX512-VNNI/BF16 MXFP4/FP8), long-prefill GPU streaming, SM80 (A100) port.
A LLaMA2-7b chatbot with memory running on CPU, and optimized using smooth quantization, 4-bit quantization or Intel® Extension For PyTorch with bfloat16.
Linux kernel and systems fast path for LLM inference: eBPF tracing, runtime hints, cgroups, NUMA/GPU locality, KV-cache memory policies, TTFT boost, and experimental kernel primitives.