Skip to content

vllm.models.inkling.amd.ops

Inkling kernels (NVIDIA).

rmsnorm / sconv import eagerly. The SwiGLU kernels and the FA4 relative-attention wrapper are exposed lazily to keep this package's import path lightweight.

Modules:

  • fa4_rel_attention –

    ROCm paged attention with Inkling's query-dependent relative bias.

  • fa4_warmup –

    Compatibility shim for the ROCm relative-attention implementation.

  • gluon –

    GFX950-only Gluon relative-attention kernels for Inkling.

  • lamport –

    Deadlock-free fused RS + short-conv + AG + residual + RMSNorm.

  • mm_towers –

    Fused CUDA kernels for the Inkling vision/audio towers.

  • norm –
  • rel_attention_decode –

    Split-KV decode for Inkling relative attention on ROCm.

  • sconv –

    Inkling short-convolution kernels backed by a paged sliding-window state cache.

  • silu_and_mul –

    SwiGLU kernels for the Inkling MLP layers.