flashinfer.fused_moe.BGMVMoEPortablePlan

class flashinfer.fused_moe.BGMVMoEPortablePlan(*, y_accum: Tensor, shrink_out: Tensor, x: Tensor, lora_a_weights: List[Tensor], lora_b_weights: List[Tensor], sorted_token_ids: Tensor, expert_ids: Tensor, lora_indices: Tensor, topk_weights: Tensor, num_experts: int, fallback_reason: str)

Prepared plan that runs the portable bgmv_moe_shrink + bgmv_moe_expand path.

Returned by prepare_bgmv_moe() when the inputs are outside the generated Cake support set and fallback=True. Same interface as BGMVMoECakePlan (run zeroes the workspaces, runs shrink and expand, and returns the FP32 accumulator; first eager call captures a CUDA Graph). The portable expand accumulates with atomics, so replays are not guaranteed bitwise identical. backend_used is "portable" and schedule_id is None.

__init__(*, y_accum: Tensor, shrink_out: Tensor, x: Tensor, lora_a_weights: List[Tensor], lora_b_weights: List[Tensor], sorted_token_ids: Tensor, expert_ids: Tensor, lora_indices: Tensor, topk_weights: Tensor, num_experts: int, fallback_reason: str) → None

Methods

__init__(*, y_accum, shrink_out, x, ...)
close()	Release graph resources after pending replay work completes.
run()	Run or replay the prepared zero+shrink+expand pipeline.

Attributes

backend_used