Add DeepSeek W4A8 MoE trace coverage
#13
by ransiki - opened
Adds a model-sourced DeepSeek-V3/R1 W4A8 Definition with MXFP4 weights, MXFP8 activations, grouped top-k routing, and an EP32 local expert slice. Includes three seq_len workloads (1, 16, 64), public TRT-LLM and CUTLASS Solutions, shared quantized safetensors, and six B200 PASSED traces. Validated with flashinfer-python 0.6.16.post1; standalone-vs-FIB 5% gates passed with max ratios 0.9937 (TRT-LLM) and 1.0193 (CUTLASS). Llama4 is out of scope.
ransiki changed pull request status to closed