vllm.v1.attention.ops ¶

Modules:

• chunked_prefill_paged_decode –
• common –
• cp_common –

  Shared symmetric-memory infrastructure for context-parallel attention.

• dcp –

  MLA DCP collective selection and direct symmetric-memory implementations.

• flashmla –
• flydsl_kernels –
• flydsl_turboquant_decode –

  FlyDSL TurboQuant decode launcher (vLLM-side).

• int4_per_token_head –

  Sub-byte packed (INT4) per-token-head KV cache mode.

• merge_attn_states –
• pcp –
• prefix_prefill –
• rocm_aiter_mla_sparse –
• triton_attention_helpers –

  Shared @triton.jit helpers used by the unified attention kernel

• triton_decode_attention –

  Memory-efficient attention for decoding.

• triton_fp8_mqa_logits –

  Temporary gfx942 fallback for AITER's fp8_mqa_logits kernel.

• triton_prefill_attention –

  Memory-efficient attention for prefill.

• triton_reshape_and_cache_flash –
• triton_turboquant_decode –

  Triton fused TurboQuant decode attention.

• triton_turboquant_store –

  Fused Triton kernels for TurboQuant KV store.

• triton_unified_attention –
• triton_unified_attention_diffkv –

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

• turboquant_soa –
• vit_attn_wrappers –

  This file contains ops for ViT attention to be compatible with torch.compile

• xpu_mla_sparse –