Primus technical blogs

Every published article about Primus, newest first. The last column points at the maintained documentation or code for the same topic: a blog is frozen at its publication date, so where the two disagree, the documentation wins. The top-level README lists only the most recent handful.

Published

Date	Title	What it covers	In this repo
2026/09/03	Enabling DeepSeek-V4-Flash Training on AMD Instinct MI355X GPUs with Primus	DeepSeek-V4-Flash architecture, performance projection, kernel optimizations, and how to reproduce the runs	examples/deepseek-v4
2026/08/12	Using ODC to Accelerate AMD SFT Training	On-demand point-to-point communication replacing the FSDP all-gather / reduce-scatter, over rocSHMEM and MORI	ODC FSDP2 patches
2026/07/06	Primus Tuning Agent: Closing the Configuration-Search Loop	LLM-driven search over parallelism, pipeline layout, and recompute sets, scored by projection instead of cluster time	Tuning agent
2026/06/10	Dropless MoE Training in JAX with Primus-Turbo	Grouped GEMM brought into MaxText through JAX FFI and custom_vjp, with the fan-out / fan-in correctness details	MaxText parameters
2026/04/24	Primus Projection: Estimate Memory and Performance Before You Train	Analytical memory estimation and benchmark-anchored throughput projection for multi-node runs	Projection
2026/02/23	Primus-Pipeline: A More Flexible and Scalable Pipeline Parallelism Implementation	Zero-bubble schedules (zerobubble / zbv / v-half / v-min) alongside 1F1B and interleaved, in the Megatron-LM backend	Parallelism strategies
2026/01/15	Deep Dive into Primus: High-Performance Training for Large Language Models	How the unified CLI and tuned backend presets reach peak dense-LLM throughput with minimal manual tuning	Megatron-LM training
2025/12/16	MoE Training Best Practices on AMD GPUs	The foundational MoE optimizations: DeepEP dispatch, sync-free MoE, 1F1B all-to-all overlap, selective recompute	MoE training deep-dive
2025/08/22	Primus: A Lightweight, Unified Training Framework for Large Models on AMD GPUs	The framework itself: YAML-driven configuration, multi-backend design, preflight validation, structured logging	Project overview

Awaiting publication

Date	Title	What it covers	In this repo
2026/08/21	MoE Training Optimization with Primus	MegaMoE megakernel, FP8 / MXFP8 grouped GEMM, training-at-scale tuning	MegaMoE fused MoE layer

Adding an article

Article sources are maintained in the ROCm Blogs repository, not here. Copies were vendored into this repository once before and drifted out of sync with the published versions, down to stale author lists, so this directory holds the index only. When an article goes live, add a row above and link to rocm.blogs.amd.com rather than adding a copy. Add it to the top-level README too if it belongs in the recent handful, dropping the oldest entry there to keep that list short.