Here is everything I've worked on so far, ordered and dated.
I recently achieved 84.4% Memory Bandwidth Efficiency (~541 GB/s) on my Radeon 9070XT (RDNA 4) using wave shuffles, LDS, and a grid-stride loop. You can see the code and profiling writeup for the kernel here.
My primary focus is contributing to vLLM's ROCm inference stack. My secondary focus is further digging into ROCm and HIP, focusing on vectorization and quantization.
| Project | PR | Description | Status |
|---|---|---|---|
| vllm-project/vllm | #40827 | Rename LLMM1 → vecMatMul, refactor, fix two RDNA4 bugs |
Open |
| vllm-project/vllm | #41187 | Fix reduction_smem layout (stacks on #40827) |
Open |
| vllm-project/vllm | #35173 | Move TORCH_CHECK assertions in wvSplitK to fire before values are consumed |
Open |
| vllm-project/vllm | #35672 | Identified and removed dead production code — a utility function that was silently unreachable because AIter had taken over its dispatch path | Merged |
| foundation-model-stack/fastsafetensors | #67 | Remove hipify-perl build dependency, enable ROCm manylinux wheel builds |
Merged |
| foundation-model-stack/fastsafetensors | #78 | Universal CUDA/ROCm wheel via runtime dlopen detection — single build works on both platforms |
Merged |
| vllm-project/vllm | #43625 | Bump fastsafetensors to v0.3.2 from PyPI, remove ROCm git source build | Merged |
| vllm-project/vllm | #48688 | Enable the fp32 head_dtype torch.mm fast path on ROCm |
Merged |
| Project | Issue | Description | Status |
|---|---|---|---|
| pytorch/pytorch | #185074 | RFC: torch.compile + CUDA graph capture — Inductor JIT has no capture-state awareness, causes CPU→GPU copy error on first call |
Closed (completed) |
Hi, I'm Turner Jabbour. I've been a software engineer for ~7 years, primarily working in Node and React. Around September 2025, I became deeply interested in GPU programming, ROCm, and the broader world of low-level performance engineering.
This repository is my space to learn in public as I delve into GPU kernel engineering and inference systems work.
There are three important directories:
- kernels - I explore different kernels and include a write up of what I learned and how it relates to inference.
- papers - I summarize and discuss different papers.
- topics - I dive deep into some specific topic.
My long-term goal is to build strong competency in HIP, Triton, and AMD's GPU software stack, with a focus on high-performance inference.