Requirements#

Supported Hardwares#

Vendor

Device

Backend

Status

NVIDIA

CUDA GPUs

torch + FlagCX

Full support

Huawei Ascend

NPU

torch_npu + CANN

Full support

MetaX

C500/C550 series

torch_maca + MACA

Validated

Moore Threads

MUSA S-series GPUs

torch_musa + FlagCX

Validated (heterogeneous)

CPU

x86/ARM

torch (CPU)

Basic support

Operating System#

Linux (official)

Software#

  • Python >= 3.10

  • PyTorch >= 2.4

  • For CUDA: CUDA >= 12.1

  • For MetaX: torch_maca, MACA toolkit

  • For MUSA: torch_musa, MUSA toolkit

  • For NPU: torch_npu, CANN toolkit

FlagOS Dependencies#

  • FlagCX (required) — Unified cross-vendor communication backend

  • vllm-plugin-FL (optional) — vLLM-based rollout engine

  • TransformerEngine-FL (optional) — TE-FL based FSDP training engine

  • Megatron-LM-FL (optional) — Megatron-based training engine

  • FlagGems (optional) — Triton-based operator library