flux-comm
Collectives that know the topology.
Communication library implementing all-reduce, all-gather and point-to-point over HyperLink, RDMA Ethernet and PCIe, choosing algorithms from the discovered topology rather than from environment variables.
Specifications
| Collectives | All-reduce, all-gather, reduce-scatter, A2A |
|---|---|
| Transports | HyperLink, RoCE, IB, PCIe, shared mem |
| Topology | Auto-discovered, per-job plans |
| In-network | Offload to Cascade switches |
| Family | Core Libraries |
| Software stack | FLUX — source-compatible across the whole catalog |
Disclosure
flux-comm is invented. The numbers above were written to be internally consistent and plausible, not measured. Treat every figure on this site as placeholder content for layout, search and API testing.
Related
Often deployed together
FLUX
The parallel programming model everything else is built on.
- Languages
- C++23, Python, Fortran, Rust (community)
- Targets
- All AXION accelerators, x86 + Arm hosts
FLUX Toolkit
Compiler, libraries, headers and samples in one install.
- Install size
- 6.2 GB full / 1.1 GB runtime
- Package managers
- apt, dnf, conda, pip, brew
flux-dnn
Hand-tuned primitives for neural network layers.
- Primitives
- Attention, conv, norm, GEMM epilogues
- Precisions
- FP4, FP6, FP8, BF16, TF32, FP64
flux-blas
Dense linear algebra, at the achievable roofline.
- Coverage
- BLAS 1/2/3 + extensions
- Peak efficiency
- 94% of dense roofline