flashinfer.comm.mnnvl_cutedsl_moe_finalize.MNNVLCuteDSLMoEFinalizeAllReduceWorkspace

class flashinfer.comm.mnnvl_cutedsl_moe_finalize.MNNVLCuteDSLMoEFinalizeAllReduceWorkspace(tp_size: int, tp_rank: int, max_token_num: int, hidden_dim: int, top_k: int, *, group: ProcessGroup, enable_pdl: bool = True)

Mailbox, stage state and compiled kernels for one TP group.

Construction is collective over group. Compile every configuration with compile() before CUDA graph capture. Calls must not overlap. Before selecting this path, pass each rank’s local supports_* verdict to check_support() on every rank. Local checks alone are not a rank-consistent fallback decision.

__init__(tp_size: int, tp_rank: int, max_token_num: int, hidden_dim: int, top_k: int, *, group: ProcessGroup, enable_pdl: bool = True) → None

Methods

__init__(tp_size, tp_rank, max_token_num, ...)
check_support(local_supported)	Collectively agree on a local supports_* result before capture.
compile(*, row_stride, weight_dtype, eps[, ...])	Compile one configuration ahead of time, e.g. before graph capture.
destroy()	Release the mailbox once in-flight work on every rank is done.

Attributes

impl