vllm.model_executor.layers.fused_moe.experts

Modules:

• aiter_mxfp4_w4a16_moe –
• aiter_mxfp4_w4a8_moe –
• aiter_mxfp8_moe –

  MXFP8 (1x32 block, E8M0) MoE via AITER's FlyDSL two-stage grouped GEMM

• batched_deep_gemm_moe –
• cpu_int4_moe –

  CPU INT4 W4A8 dynamic quantized fused MoE experts.

• cpu_moe –

  CPU fused MoE experts.

• cutlass_moe –

  CUTLASS based Fused MoE kernels.

• deep_gemm_moe –
• fallback –
• flashinfer_b12x_moe –
• flashinfer_cutedsl_batched_moe –
• flashinfer_cutedsl_moe –
• flashinfer_cutlass_moe –
• flashinfer_moe_ep –

  FlashInfer MoE-EP megakernels as a modular-kernel experts implementation.

• fused_batched_moe –

  Fused batched MoE kernel.

• fused_humming_moe –

  Fused MoE utilities for Humming.

• gpt_oss_triton_kernels_moe –
• int4_emulation_moe –

  Int4 weight-only quantization emulation for MoE.

• lora_context –
• lora_experts_mixin –
• marlin_moe –

  Fused MoE utilities for GPTQ.

• moonep_experts –

  MoonEP experts: grouped GEMM over MoonEP's expert-grouped activations.

• mxfp8_emulation_moe –

  MXFP8 (1x32 block, E8M0 scale) MoE experts on Triton.

• mxfp8_native_moe –

  Native MXFP8 (1x32 block, E8M0 scale) MoE for AMD CDNA4 (gfx950) via Triton

• nvfp4_emulation_moe –

  NVFP4 quantization emulation for MoE.

• ocp_mx_emulation_moe –

  OCP MX quantization emulation for MoE.

• rdna3_moe –

  Fused MoE W4A16 experts on the RDNA3 (gfx1100) HIP kernel.

• rocm_aiter_moe –
• triton_cutlass_moe –
• triton_deep_gemm_moe –
• triton_moe –

  Triton-based MoE expert implementations.

• trtllm_bf16_moe –
• trtllm_fp8_moe –
• trtllm_lora_moe –

  LoRA-aware FlashInfer TRT-LLM MoE experts (BF16).

• trtllm_mxfp4_moe –
• trtllm_mxint4_moe –
• trtllm_nvfp4_moe –
• xpu_moe –
• zentorch_moe –

  Zentorch fused MoE experts for AMD Zen CPUs.