vllm.models.beam.nvidia

NVIDIA-specific Beam model implementation.

Modules:

• activation –
• attention –
• attention_ops –

  Fused Q/K normalization, position transforms, and attention output gating.

• attention_ops_cutedsl –

  Bit-exact BF16 output gate in CuTe DSL.

• dflash2 –

  DFlash2 draft model for Beam targets.

• dflash2_context_ops –
• dflash2_conv_ops –
• dflash2_ops –
• layernorm –

  Beam kernels preserving FP32 normalization and affine arithmetic.

• model –

  Inference-only Beam model compatible with HuggingFace weights.

• moe –
• projection –

  Packed attention projections using shared low-latency GEMMs.