Skip to content
View sbeierle's full-sized avatar

Block or report sbeierle

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
sbeierle/README.md

Stefan Beierle // Systems Architecture Banner

Stefan Beierle

Systems Architect | Mechanistic Interpretability & AI Safety Specialist | High-Throughput Inference Engineering

Independent researcher and systems architect focusing on sovereign runtime resilience, local bare-metal inference, and mechanistic interpretability. My work bridges low-level accelerator telemetry (CUDA/ROCm, virtual memory management, distributed fabrics) with latent activation steering and critical infrastructure AI safety.


🔬 Core Architecture & Research Domains

  • Inference Engine Telemetry & Scaling: Cross-layer root-cause forensics for vLLM, SGLang, and FlashInfer (KV-cache & split-reduction mismatches, CUDA Graph lifecycle & module-eviction bugs, multi-process CUDA driver context taxes, distributed NCCL/MNNVL topologies).
  • Mechanistic Interpretability & Latent Control: Neural weight patching (down_proj, lm_head), activation steering, and deterministic safety override design for sovereign deployment.
  • Bare-Metal & Accelerated Infrastructure: Custom ROCm builds, NVIDIA Blackwell / MNNVL fabrics, kernel-level dispatch auditing, and memory-bandwidth bounded roofline optimization.
  • Adversarial Resilience & Evaluation: White-box safety verification, constraint leakage analysis, and alignment stress-testing.

📦 Key Projects & Audits

  • inference-fail-lab – Systematic root-cause forensics, bare-metal verification matrices, and cross-layer dispatch audits across vLLM, SGLang, and FlashInfer on NVIDIA Blackwell (GB10 / GB300 NVL, SM103/SM121) and AMD CDNA/RDNA (gfx90a, gfx950/MI355X, gfx1100, gfx1201). Isolates Python scheduler-to-C++/ASM boundary failures and CUDA Graph replay hazards (vllm#67110 / PR #66444, vllm#62697 / PR #63735, sglang#51278, flashinfer#9670, flashinfer#6558 / #7883).
  • BlackSky-Havarie / Venom – Open-source frameworks for mechanistic interpretability, latent control steering, and runtime emergency override.
  • mistral-downproj-rlhf-patch – Targeted activation steering and surgical representation patching via projection subgraphs.
  • Upstream Inference Telemetry – Root-cause isolation for high-worker VMM initialization bottlenecks on next-gen multi-node architectures (sglang#45345 / PR #45384).

🛠️ Technology Stack & Telemetry

  • Engines & Kernels: vLLM, SGLang, FlashInfer, llama.cpp, Triton, CuTeDSL, PyTorch
  • Hardware & Drivers: AMD ROCm (RDNA3 / RDNA4 / CDNA2 / CDNA3.5), NVIDIA CUDA / Blackwell / MNNVL, Linux Bare-Metal, PCIe/Fabric Diagnostics
  • Languages: Python, C/C++, Bash

🌐 Presence & Publications

  • Locations: Karlsruhe, Germany | Makkah, Saudi Arabia
  • ORCID: 0009-0005-8512-3839
  • Languages: 🇩🇪 German (Native) | 🇺🇸 English (Fluent) | 🇸🇦 Arabic (Professional)

ستيفان بيرله

مهندس معماري للأنظمة | باحث في أمن الذكاء الاصطناعي والتفسير الميكانيكي | متخصص في كفاءة الاستدلال عالي الأداء

مهندس وباحث مستقل يركز على استقلالية تشغيل النماذج الكبيرة (LLMs)، وتحليل الآليات العصبية الداخلية (Mechanistic Interpretability)، والتحقق من موثوقية البنى التحتية الحساسة. يرتكز عملي على تحسين أداء المعالجات ومراقبة بروتوكولات الاتصال (CUDA / ROCm / MNNVL) وتوجيه التنشيط العصبي في النماذج محلياً.

🔧 مجالات التركيز التقنية:

  • مختبر تحليل أعطال الاستدلال (inference-fail-lab): تحليل جذري للأعطال عبر طبقات البرمجيات (Python / C++ / CUDA / ROCm) في محركات vLLM و SGLang و FlashInfer على معماريات NVIDIA Blackwell و AMD MI355X.
  • تحليل أعطال بيئات الاستدلال: تشخيص مشاكل الذاكرة وجدولة المهام في محركات الاستدلال المتقدمة (vLLM / SGLang).
  • التفسير الميكانيكي وتعديل الأوزان: توجيه التنشيط العصبي الداخلي وهندسة بروتوكولات الأمان دون الاعتماد على السحابة.
  • البنى التحتية الصلبة (Bare-Metal): بناء بيئات مخصصة لـ ROCm و CUDA، وتحليل كفاءة الذاكرة وعرض النطاق الترددي للشرائح الحوسبية.

Pinned Loading

  1. inference-fail-lab inference-fail-lab Public

    Systematic edge-case verification, failure telemetry, and kernel invariants in LLM serving engines (vLLM, SGLang, TensorRT-LLM)

    Python

  2. kolibri1-praxistest kolibri1-praxistest Public

    Empirical evaluation of Aleph Alpha Kolibri-1 (78B MoE) on NVIDIA H200 using vLLM: throughput, prompt injection, and German NIS2 regulatory compliance.

    Python 1

  3. fame-pytorch-kit fame-pytorch-kit Public

    Custom-built PyTorch ROCm Kit for AMD GPUs – compile, install, and test with full control. 🔥

    Shell