Add public FlashInfer MoE solutions

#12
by ransiki - opened

Adds 15 Solution records for existing MoE Definitions: five TRT-LLM FP8 block-scale adapters, five TRT-LLM MXFP4 adapters, and five CUTLASS MXFP8 x MXFP4 adapters. Validated on NVIDIA B200 with public flashinfer-python 0.6.16.post1. No Definitions or existing dataset assets are modified. Workloads and traces are intentionally excluded because RandomInput cannot represent valid scaled FP8/FP4 quantized tensors without safetensors.

ransiki changed pull request status to closed

Superseded by consolidated PR #14, which moves Solutions under flashinfer and consolidates shape-specific FP8 Solutions into routing-contract Definitions with multiple Workloads.

Sign up or log in to comment