flashinfer.fused_moe.CuTileFp8PerTensorConfig

class flashinfer.fused_moe.CuTileFp8PerTensorConfig

cuTile per-tensor E4M3 weights and activations.

Inputs are BF16; both GEMM inputs are dynamically quantized to E4M3. Expert parallelism and fused shared experts are not supported.

__init__() → None

Methods

__init__()

prepare_weights(w1_fp8, w1_scale, w2_fp8, ...)

Prepare E4M3 weights with one FP32 dequantization scale per expert.

supported(arch)