Recent Posts#
Iteratively Tuning hipBLASLt TensileLite Kernels: A Smaller Search, a Faster Kernel
Learn why a small iterative search beats a big one-shot sweep for hipBLASLt TensileLite kernels, in less time and with no risk of regression.
veRL on AMD: Production-Ready RL Post-Training on ROCm
Run veRL RL post-training on AMD Instinct GPUs with a turnkey ROCm container, AITER-accelerated rollout, and validated accuracy on MI300 and MI355.
Efficiently Serving NVFP4 Models on AMD Instinct™ MI350X/MI355X Accelerators via Online NVFP4 to Quark MXFP4 Requantization
Serve NVFP4 models on MI350X/MI355X via SGLang's online NVFP4 to MXFP4 requantization: no preprocessing, minimal accuracy impact, native throughput.
Enabling DeepSeek-V4-Flash Training on AMD Instinct MI355X GPUs with Primus
DeepSeek-V4-Flash training on AMD Instinct GPUs with Primus: model architecture introduction, performance projection, kernel optimizations, and how to reproduce.
Optimizing ATOM and vLLM-ATOM for High-Interactivity Inference
Cut per-token latency on AMD Instinct MI355X GPUs: see how ATOM, vLLM-ATOM, and AITER strip fixed costs out of the LLM decode path.
4-bit KV Caching in LMCache: Offloading Quantized KV Beyond HBM for Context-Heavy Agents on AMD MI355X
Combining 4-bit KV quantization (TurboQuant) with KV offload (LMCache) on AMD MI355X nearly doubles serving goodput for long-context agents.
A Deep Dive into LDS Optimizations on AMD Instinct MI450 GPUs
Learn how to optimize LDS traffic in Gluon kernels on AMD Instinct MI450 GPUs using transposed loads and partition-conflict-free layouts.
ROCm 10.0: A Decade of Open Compute, Built for the Age of Agentic AI
Explore what's new in ROCm 10.0, headlined by ROCm.AI, the ROCm CLI, AMD Skills, and Hyperloom, alongside the ROCm Core SDK and platform-wide improvements
Enabling Physical AI Agents with Lemonade
Learn to deploy local agents for interactive robot arm manipulation using the Lemonade framework.
Serving 64Mi-Token Contexts on One AMD Instinct™ MI355X Node
Explore how one 8-GPU AMD MI355X node serves Kimi Linear from 1K to 64M tokens under vLLM—and the TTFT and decode throughput behind the run.
DI Series: Scaling GLM-5.1-FP8 to 64 MI300X GPUs
Scale GLM-5.1-FP8 long-context serving beyond a single node on AMD Instinct MI300X with WideEP prefill-decode disaggregation and MoRI.
Exploring XGBoost: A Deep Dive
A single source explanation of XGBoost features and its working principles.