Technical papers produced from real lab work at Dielabs. Each paper documents original findings from hands-on experimentation with LLM inference systems.
Degenerative decoding loops in vLLM: diagnosis, three-plane framework, and cross-model comparison (Qwen2.5-7B vs Mistral-7B). Identifies sampling entropy collapse and model-specific coherence thresholds.
Experimental investigation of KV cache offloading mechanisms in vLLM 0.15.1 on consumer hardware. Reveals three distinct mechanisms commonly confused under one term. On this setup, tuning the base scheduler parameters gave a much larger gain than any offloading mechanism.
Benchmark of SN / TP-2 / DP-2 architectures on Dell PowerEdge R730 with vLLM. Shows the same single-node, TP and DP patterns seen on GPUs, suggesting they are general principles rather than GPU artifacts. Identifies TP as anti-pattern on slow interconnects.
Three benchmarks on a single RTX 4070 Super with Qwen3-8B-AWQ via vLLM. Shows how the binding constraint moves from KV capacity (FP16, prompt-heavy) to a higher KV ceiling (FP8) to memory bandwidth (FP8, chat-like). A methodology for inference sizing that starts from workload and SLO.
Local AI on a consumer GPU: a 35B MoE model (Qwen3.6-35B-A3B) served by llama.cpp on a 12 GB RTX 4070 Super, with experts offloaded to RAM. Shows that the first 3.5 GB of VRAM are worth more than the next 7, that prefill is bound by PCIe and decode by GPU-CPU handovers, and how to spend VRAM on the metric you care about.
Coding Benchmark @Dielabs: 17 LLM configurations on 50 Aider Polyglot exercises, local (llama.cpp on a 12 GB GPU) and remote (pinned providers via OpenRouter). Quality, cost and time per solved exercise; what model size, reasoning, sampling and quantization really change, and how to choose a model for code.
Measured best practices for an eight-GPU, air-cooled PCIe cloud server: discover the effective topology, test contention, place communicating workers within fast domains, and qualify RAM paths and whole-node thermals before the LLM benchmark.
All content is original Dielabs work by Diego Bardella.