flashinfer.fused_moe.PrimsTsConfig¶
- class flashinfer.fused_moe.PrimsTsConfig¶
Explicit Prims-TS backend for SM100 and SM103.
Opt-in only: this config is never part of the default backend list. NVFP4×NVFP4, MXFP4×MXFP8, and MXFP4×BF16 reuse
TrtllmFp4Config. BF16×BF16 reusesTrtllmBf16Config. MXFP8×MXFP8 and FP8PerTensor×FP8PerTensor reuseTrtllmFp8BlockConfigandTrtllmFp8PerTensorConfig. DeepSeekFp8×DeepSeekFp8 reuses the TRT-LLM scales and activations, then shuffles the weight payloads (epilogue tile 64); do not register that dictionary under"trtllm_fp8_block".The GEMM middle stage is Prims-TS; routing and finalize stay on the TRT-LLM Gen path.
- __init__() None¶
Methods
__init__()prepare_activations(hidden_states_bf16, *, quant)Build activations for the matching physical layout.
prepare_weights(w1_bf16, w2_bf16, *, quant, ...)Build the Prims-TS weight view for one supported quant pair.
supported(arch)