vllm.v1.attention.backends

Modules:

• b12x –

  b12x paged causal attention backend for SM12x.

• composite –

  Reusable attention backend composition and routing policies.

• cpu_attn –
• fa_utils –
• flash_attn –

  Attention layer with FlashAttention.

• flash_attn_diffkv –

  Attention layer with FlashAttention.

• flashinfer –

  Attention layer with FlashInfer.

• flex_attention –

  Attention layer with FlexAttention.

• gdn_attn –

  Backend for GatedDeltaNet attention.

• hpc_attn –

  HPC Attention Backend.

• mamba2_attn –
• mamba_attn –
• mla –
• recoverssm_metadata –
• registry –

  Attention backend registry.

• rocm_aiter_fa –

  Attention layer with AiterFlashAttention.

• rocm_aiter_unified_attn –

  Attention layer with PagedAttention and Triton prefix prefill.

• rocm_attn –

  Attention layer with PagedAttention and Triton prefix prefill.

• short_conv_attn –
• triton_attn –

  High-Performance Triton-only Attention layer.

• triton_attn_diffkv –

  Triton attention backend with different K/V head dimensions (DiffKV).

• triton_flash_attn –

  Triton image-mask and FlashAttention causal attention composite.

• triton_flashinfer –

  Triton image-mask and FlashInfer causal attention composite.

• turboquant_attn –

  TurboQuant attention backend for vLLM.

• utils –
• zentorch_sdpa –

  Encoder attention through the zentorch SDPA kernel on Zen CPUs.