Focused on high-performance LLM inference, GPU kernel optimization, and distributed runtime design.
Currently based in Curitiba, Brazil (GMT-3). Open to relocation worldwide.
Full C++/CUDA Runtime
- Transformer inference engine in pure C++/CUDA (no PyTorch execution path).
- 32.4k tok/s peak, achieving 87% of INT8 Tensor Core ceiling on RTX 2070.
- Custom Flash Attention, cuBLASLt INT8, CUDA Graphs, Megatron-style TP.
- Runs TinyStories 110M and TinyLlama 1.1B end-to-end.
PagedAttention + Scheduling Runtime (Single-GPU Study)
- Custom CUDA implementation of block-based PagedAttention (logical → physical KV mapping).
- 2.89M tok/s kernel microbenchmark (small-head configuration, RTX 2070).
- 29× kernel optimization gain (v1 → v3): warp reduction + online softmax + shared memory tiling + half2 vectorization.
- 205–397× vs naive PyTorch attention (non-Flash baseline, attention-layer only).
- 66.7% memory reduction via block-based KV allocation.
- Includes continuous batching, swap-first preemption, and cost-based victim policy.
- Designed as a single-GPU experimental study of memory paging and runtime scheduling.
PTX-Level GEMM Engineering
- Manual PTX instruction emission for Tensor Core GEMM.
- Outperforms cuBLAS on specific 2048×2048×2048 configuration (SM75 tuned).
- Achieves 60–65% of cuBLAS performance at 8192³.
- Auto-tuning over WARPS / STAGE_K / ILP_DEPTH configuration space.
INT4 CUDA Kernel + LLM Integration
- Bit-accurate INT4 weight-only quantization kernel using raw PTX assembly.
- 1.000 cosine similarity vs FP16 baseline (layer-level validation).
- End-to-end Qwen2-0.5B integration: −14.7% VRAM usage with coherent outputs.
- Documents SM75 structural limits (no cp.async, no native INT4 MMA operands).
Compiler MVP with CUDA Codegen
- Full pipeline: IR → CUDA codegen → nvcc → GPU execution.
- 9 Transformer fusion patterns (GEMM+Bias+GeLU, LayerNorm fusion, etc.).
- 10.5 TFLOPS peak on RTX 2070 (37% of SM75 theoretical peak).
- Templated CUDA generation with type/shape specialization.
Distributed LLM Runtime
- Multi-process runtime implementing Megatron-style Tensor Parallelism and Pipeline Parallelism.
- GPipe-style fill-drain micro-batching.
- KV cache sharding across processes.
- Bit-identical validation on Qwen2-0.5B weights (controlled setup).
OpenAI-Compatible Serving Runtime
- Continuous batching, SSE streaming, KV cache management.
- Prometheus metrics, graceful shutdown, health checks.
- 33/33 unit tests + 6/6 end-to-end scenarios passing.
Model Compression Research Implementation
- From-scratch AWQ and GPTQ implementations.
- Validated on Qwen2-0.5B with 26.05 perplexity at 4-bit.
- Explores Hessian-based compensation vs activation-aware weighting.
- flash-attention-sm75-register-fused — Register-fused FlashAttention (5.0 TFLOPS peak, roofline-validated).
- flash-attention-sm75 — FlashAttention v1 forward pass in CUDA (SM75).
- fused-backward-kernels-sm75 — Fused GEMM+Bias+GeLU backward kernels (3.1× vs autograd at M=1024).
- speculative-decoding-runtime — 1.41× speedup via draft-model verification.
- vlm-inference-runtime — Vision-Language deployment wrapper.
- lora-inference-runtime — Multi-tenant adapter serving with hot-swapping.
| Layer | Technologies |
|---|---|
| GPU & Low-Level | CUDA C++, PTX, Tensor Cores (WMMA), CUDA Graphs, Nsight Compute |
| Compilers | IR design, CUDA codegen, MLIR/LLVM concepts |
| AI Frameworks | PyTorch, HuggingFace, PEFT, ONNX |
| Distributed | NCCL, torch.distributed, Tensor Parallelism, Pipeline Parallelism |
| Serving | FastAPI, Continuous Batching, PagedAttention, KV Management |
| Backend | Redis, PostgreSQL, Prometheus |
| DevOps | Docker, GitHub Actions |
| Languages | C, C++17/20, CUDA, PTX, Python, Rust |
| Achievement | Project |
|---|---|
| 32.4k tok/s peak | mini-llm-inference-engine |
| 87% INT8 Tensor Core ceiling | mini-llm-inference-engine |
| 2.89M tok/s kernel microbenchmark | paged-attention-runtime |
| 29× kernel optimization gain | paged-attention-runtime |
| 10.5 TFLOPS compiler-generated | llm-fusion-compiler |
| 1.41× speedup speculative decoding | speculative-decoding-runtime |
| 26.05 PPL at 4-bit (AWQ) | quantization-runtime |
| R² = 0.9996 comm cost model | comm-cost-modeling |
AI Infrastructure Engineer (Contract) — Confidential Leading Chinese LLM Company (Under NDA) | Sep 2025 – Mar 2026
Built Rust-based ingestion pipelines with Tokio async and backpressure-safe systems for production LLM infrastructure.
Software Architecture Lead — ZBrasil | Apr 2025 – Aug 2025
Led backend architecture and GPU infrastructure decisions.
Contributed to LLM Inference Handbook (llm-inference-at-scale) — module on observability signal coverage for LLM serving.
LinkedIn: linkedin.com/in/joaofelipescheuer
Email: [email protected]
WeChat: JohnScheuer7
Location: Curitiba, Brazil — Open to Relocation Worldwide