flashinfer.fused_moe.CakeWarpDecodeConfig

class flashinfer.fused_moe.CakeWarpDecodeConfig(backend: Literal['cake'] = 'cake')

Explicit Cake NVFP4 warp-decode backend for exact SM103.

This backend is intentionally narrow: it accepts only the two calibrated expert geometries documented by CakeWarpDecodeRunner, 1–32 tokens, unpacked precomputed routing, and the default SwiGLU semantics. It is never part of the default backend list; users opt in with CakeWarpDecodeConfig(backend="cake").

The physical weight and activation layouts are exactly those produced by TrtllmFp4Config for QuantVariant.NVFP4. This keeps one quantized representation usable by both runners.

__init__(backend: Literal['cake'] = 'cake') → None

Methods

__init__([backend])
prepare_activations(hidden_states_bf16, *[, ...])	Build the shared TRTLLM NVFP4 packed activation view.
prepare_weights(w1_bf16, w2_bf16, *[, ...])	Build the shared TRTLLM NVFP4 physical weight view.
supported(arch)

Attributes

backend