vllm.models.flint.ops

Modules:

• cute_primitives –

  CuTe DSL device primitives for Flint's small-M BF16 x NVFP4/BF16 GEMV.

• decode_attention –

  Split-KV paged decode attention for Flint (GQA 8, head_dim 128, bf16 KV).

• linear –

  Flint dense linears with a small-M CuTe GEMV fast path.

• moe –

  Fused NVFP4 W4A16 routed-expert MoE for Flint (plus decode kernels for the

• moe_cute –

  FlashInfer's SM100 NVFP4-weight / BF16-activation grouped MoE GEMMs.

• moe_gemv –

  Expert-indexed small-M NVFP4 GEMV kernels (CuTe DSL) for the low-latency

• norm –

  Flint's RMSNorm, fused residual-add RMSNorm and SiLU-and-mul (Triton).

• nvfp4_gemv –

  Small-M BF16 x NVFP4 (W4A16) and BF16 x BF16 GEMM for SM100 in CuTe DSL.

• qk_norm_rope –

  Fused per-head q/k RMSNorm + NeoX RoPE + paged KV-cache write.

• qkv_fused –

  Fused decode qkv for Flint at M <= 8 tokens (CuTe DSL), two kernels: