vllm.models.inkling.nvidia ¶

Modules:

• attention –
• layernorm –

  Inkling RMSNorm (no bias, weight-scaled), backed by the vendored Triton kernel.

• logits_processor –

  Inkling logits processor with muP output scaling.

• mlp –

  Inkling dense SwiGLU MLP (also used as the MoE shared expert).

• model –

  Inkling model implementation for NVIDIA GPUs.

• moe –

  Inkling mixture-of-experts on vLLM's FusedMoE abstraction.

• ops –

  Inkling kernels (NVIDIA).

• sconv_swa_attn –

  Inkling short-conv state managed as a sliding-window KV cache.

• short_conv –

  Inkling short convolution: depthwise causal conv1d (+ residual) over a paged