Triton MoE

Self-contained Triton implementation of the HF/OpenAI-style expert MoE layer.

Exports CUDA inference layers:

• triton_moe.layers.MoE: functional benchmark signature.
• triton_moe.layers.OpenaiExperts: HF OpenaiExperts-compatible stateless layer signature.

The implementation uses Triton route grouping, ragged grouped GEMMs, fused routing-weight application, and token-wise top-k reduction. It does not implement backward.

Benchmarking against comparable kernels

Definitions for terms in the following benchmark results:

1. HF Kernel Hub -> https://huggingface.co/kernels/kernels-community/triton-moe
2. VLLM -> Runner code (https://github.com/vllm-project/vllm/blob/main/vllm/model_executor/models/gpt_oss.py) with kernel source code (https://github.com/vllm-project/vllm/blob/main/vllm/model_executor/layers/fused_moe/fused_moe.py)

We ran comparisons for decode (low batch size) and prefill (large batch size) scenarios. The configuration was for 128 experts with top-k 8.

Numbers in brackets are (-x) relative to (HF Kernel Hub | VLLM) respectively.