rocm/primus repository overview

⁠Primus

Primus/Primus-LM is a flexible and high-performance training framework designed for large-scale foundation model training and inference on AMD GPUs. It supports pretraining, post-training, and reinforcement learning workflows with multiple backends including Megatron-LM⁠, TorchTitan⁠ alongside ROCm-optimized components.

Please refer to ROCm documentation here⁠ for more details.

⁠Release Notes

⁠v26.4

• Update ROCm to 7.14 (TheRock based build)

• Upgrade PyTorch to 2.12

• Enabled release wheels for pip install

• Introduce FP4 training example

• Introduce MXFP8 training examples with Primus Turbo enhanced kernels

• Introduce Primus native SFT and LoRA trainer

• Introduce tuning agent - agent based scaling strategy explore and suggestion based on Primus Projection tool

• Release version aligned between docker images and Primus repo release tags

• Known issues:

  • Segfault during process teardown with Megatron-Bridge based workload
  • MI355X Llama-3.1-70B BF16 (TorchTitan) performance regression ~8% due to updated checkpoint recompute behavior of PyTorch
  • Performance regression of zebra-llama-1b model

⁠v26.3

• HipblasLT regression of v26.2 fully fixed on MI355X (Thanks to HIPBLASLt team for the collaobrative support!)
• Upgrade Megatron-LM backend to more recent upstream
• Primus Turbo update to more recent version
  • GEMM/Grouped GEMM Triton backend.
  • AITER asm mxfp4 GEMM integration.
• MoE Package 2.0 enabled, more tuned recipes for large scale MoE training
• Performance highlights:
  • MoE performance improved with MoE Package 2.0 (4-13% improvement on MoE models)
  • Performance optimization on Mamba models and showing close to 2x performance comparing to v26.2
• Known issues:
  • Memory usage increased on legacy LLama models

⁠v26.2

• Upgrade to ROCm 7.2
• Primus
  • Primus Perf Projection Modeling Toolkit
  • Added Primus Finetuning support w/ Megatron-Bridge (functional supporting, performance optimization on-going)
  • Added Hybrid Model Support with Primus (functional supporting, performance optimization on-going)
• Reduced docker size by 50+% for lighter weight deployments

⁠v26.1

• Added default AINIC support
• Primus update
  • Introducing CLI interface
• Updated libraries
  • Transformer Engine
  • HipBlasLT

⁠v25.11

• Primus-LM 0.6.0
  • MoE Package full support aligning with blog post: https://rocm.blogs.amd.com/software-tools-optimization/primus-moe-package/README.html⁠
  • Interactive Auto-benchmarking tool
• Updated MAD to align with Primus Examples
• MoE performance further improved up to 28% from v25.10 with fully enabled MoE package

⁠v25.10

• Merge MI300X and MI355X docker
• Primus TorchTitan backend upgrade to recent version
• Primus Megatron backend upgrade to recent version
• Primus v0.4.0
  • Support MLFlow
  • Improved Megatron-LM profiling
  • MoE Package Feature available:
    • MoE CPU Sync-free
    • 1f1b MoE overlapping
    • Zero-bubble
    • Router Fusion
  • Tools
    • Model based memory projection/modeling tool
• Primus-Turbo v0.2.0
  • FP8 GEMM
  • Grouped GEMM
  • DeepEP support
• Added Primus Torchtitan MoE support with matching performance of PyTorch Conference presentation
  • DeepSeekV3
• Diffusion Benchmark with Video Gen models
  • Hunyuan Video
  • Wan2.1-I2V
  • Mochi-1
• Performance update
  • Torchtitan avg 2.6% uplift
  • Torchtune avg 1.6% uplift
  • Megatron mi300/325 0.42% drop, mi355 4% uplift

⁠v25.9

• ROCm 7.0 support
• Combined PyTorch Docker supports
  • Primus with Megatron-LM and Torchtitan backend - v0.3.0
  • Primus Turbo - v0.1.1
  • Finetuning
    • HF PEFT
    • Torchtune
• Added/updated Diffusion Benchmark support for following models
  • Flux with smaller testing dataset
  • SDXL
• Support MI300X, MI325X, MI350X, MI355X
• Known Issue:
  • Separate images for MI300X/MI325X and MI350X/MI355X due to cross architecture compatibility issues, to be fixed in v25.10 release