flashinfer.comm.mnnvl.MnnvlMemory¶

class flashinfer.comm.mnnvl.MnnvlMemory(mapping: Mapping, size: int)¶

__init__(mapping: Mapping, size: int)¶

Methods

`__init__`(mapping, size)
`as_torch_strided_tensor`(dtype)
`close_mnnvl_memory`(ptr)
`get_allocation_granularity`(dev_id)	Return the recommended granularity of both FABRIC and POSIX file descriptor allocations, so an allocation keeps its size when it falls back from one handle type to the other.
`get_allocation_prop`(dev_id, handle_type)
`get_comm`(mapping)
`initialize`()
`new_mnnvl_memory_address`(mapping, size)
`open_mnnvl_memory`(mapping, size)
`set_comm_from_config`(mapping[, config])
`support_nvlink`([need_all_up])
`supports_mnnvl`()
`uses_fabric_handles`(comm, dev_id)	Return True if MNNVL memory shared over `comm` starts with FABRIC handles.

Attributes

`address_refcnt`
`allocated_map`
`allocation_granularity`
`comm`
`config`
`current_mem_offset`
`current_rank_stride`
`current_start_address`
`dev_id`
`fabric_handles`	True if the allocation is shared through FABRIC handles, False for POSIX file descriptors.
`fabric_page_size`
`initialized`
`local_mem_handle`	This rank's `CUmemGenericAllocationHandle` backing the allocation.
`mapped`