[Roadmap] Primus-Turbo Roadmap H2 2025

[xiaobochen-amd]

This roadmap is the H2 2025 development plan of Primus-Turbo.

Note: The roadmap is flexible and will be updated over time based on project needs and community input.

Release Overview

Version	Framework	Status	Date
v0.1.0	PyTorch + ROCm6.4	✅ Released	2025-09-11
v0.1.1	PyTorch + ROCm7.0	✅ Released	2025-10-15
v0.2.0	PyTorch + ROCm7.x	🚧 In Progress	2025-11 (est.)
v0.3.0	Planning	Planning	TBD
v0.4.0	Planning	Planning	TBD

Detailed Plans

v0.1.0 (Released)

Focus

• Build the foundational framework of Primus-Turbo.
• Provide core operators.

Features

• GEMM: Support FP16/BF16.
• FlashAttention: Support FP16/BF16.
• GroupedGEMM: Support FP16/BF16.

Famework

• Provide PyTorch APIs
• Support ROCm 6.4

v0.2.0 (In Progress)

Focus

• Introduce FP8 foundational support.
• Enable communication primitives with FP8, focusing on DeepEP.

Features

• GEMM: Support FP8 (E4M3/E5M2).
  • Support Tensorwise.
  • Support Rowwise.
  • Support Blockwise.
  • Support MX
• FlashAttention: Support FP8 (E4M3/E5M2).
  • Support Blockwise.
• GroupedGEMM: Support FP8 (E4M3/E5M2).
  • Support Tensorwise.
  • Support Rowwise.
  • Support Blockwise.
  • Support MX
• All2All: FP8 support.
  • Support Tensorwise.
• DeepEP:
  • Intra-Node Normal Kernel.
  • Inter-Node Normal Kernel.
  • Support NICs.
    • ConnectX-7
    • Thor2
    • Pensando
  • Support internode_dispatch GPU-CPU no sync.
  • Support torch.compile
• TokenDispatcher:
  • Integrate Permute/Unpermute
  • Support Sync-Free DeepEPTokenDispatcher
  • Support MoE Fused Activations.

v0.3.0 (Planning)

Focus

• Performance optimization for FP16/BF16 and FP8.

Features

...

v0.4.0 (Planning)

Focus

• Explore ultra-low precision training and inference (FP4 / FP6).

Features

...