efa 2026.7.18

Low-level Pythonic bindings for AWS EFA (SRD RDMA via libibverbs/libefa), with GPUDirect support

efa

Low-level Python bindings for AWS Elastic Fabric Adapter (EFA), including Scalable Reliable Datagram (SRD), one-sided RDMA, and GPUDirect transfers to and from torch CUDA tensors.

The package wraps libibverbs and EFA's libefa direct-verbs API in Cython. Its data path calls the provider's inline verbs directly, releases the GIL around blocking and posting operations, and does not import torch or link against CUDA.

• No Python runtime dependencies.
• SRD and UD queue pairs with SEND, RDMA read, and RDMA write operations.
• Host buffers, CUDA device pointers, and dma-buf memory registration.
• Classic and extended CQs, sender GID and unsolicited-write metadata, and direct EFA CQ, SQ, RQ, MR, and AH queries.
• One Linux abi3 wheel for CPython 3.9 and newer.

Requirements

• Linux on an AWS instance with one or more EFA devices attached.
• libibverbs.so.1 and libefa.so.1 at runtime. The AWS EFA installer and current rdma-core distributions provide both.
• The EFA device nodes available inside the process or container, normally /dev/infiniband/uverbs*.
• A C compiler, Cython, and the rdma-core development headers only when building from source.

For GPUDirect, the EFA device and instance type must support RDMA read/write, and the NVIDIA driver must support dma-buf export or nvidia_peermem.

Install

pip install efa

To build from this checkout:

pip install "Cython>=3.0" "setuptools>=77" wheel
pip install ./efa

SRD Quickstart

SRD is reliable and connectionless. Each process creates a ready-to-send QP, exchanges a 24-byte EndpointInfo out of band, and resolves the remote GID to an address handle:

import numpy as np
import efa

dev = efa.get_efa_device_list()[0]
ctx = dev.open()
pd = ctx.alloc_pd()
cq = ctx.create_cq(256)
qp = pd.create_qp(
    efa.QPInitAttr(send_cq=cq, recv_cq=cq)
).prepare(qkey=0x1234)

local_info = efa.local_endpoint_info(qp, qkey=0x1234)
# Exchange local_info.to_bytes() with the other process.
remote_info = efa.EndpointInfo.from_bytes(remote_bytes)
peer = remote_info.peer(pd)

buf = np.zeros(4096, dtype=np.uint8)
access = (
    efa.AccessFlags.LOCAL_WRITE
    | efa.AccessFlags.REMOTE_WRITE
    | efa.AccessFlags.REMOTE_READ
)
mr = efa.reg_tensor(pd, buf, access)

qp.post_send(efa.SendWR(
    wr_id=1,
    sg_list=[mr.sge()],
    opcode=efa.WROpcode.SEND,
    send_flags=efa.SendFlags.SIGNALED,
    dest=peer,
))
for wc in cq.poll(16):
    wc.raise_for_status()

Every handle is an idempotent context manager. A QP retains its PD and CQs, an MR retains its PD and backing tensor, and an SGE retains the MR it addresses.

One-sided RDMA

EFA requires the responder to have an address handle for the requester before it accepts RDMA reads or writes. Both processes should therefore resolve the other process's EndpointInfo, even if traffic is currently one-way. A missing reverse AH completes with REM_OP_ERR and EFA vendor status 0x0e (REMOTE_ERROR_UNKNOWN_PEER).

Large buffers can be split at the device's max_rdma_size:

wrs = efa.write_wrs(local_mr, peer, remote_addr, remote_rkey)
qp.post_send(wrs)

read_wrs provides the corresponding RDMA-read operation. Every generated WR is signaled, as required by EFA.

GPUDirect With Torch

The optional efa.cuda module is torch-free and duck-types objects exposing data_ptr(), numel(), and element_size(). For torch's dma-buf path, enable VMM-backed allocations before CUDA initializes:

import os
os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"

import torch
import efa
import efa.cuda

src = torch.arange(1 << 20, dtype=torch.float32, device="cuda:0")
dst = torch.zeros_like(src)

src_mr = efa.cuda.register_tensor(pd, src, access)
dst_mr = efa.cuda.register_tensor(pd, dst, access)

torch.cuda.synchronize(src.device)
qp.post_send(efa.SendWR(
    wr_id=2,
    sg_list=[src_mr.sge()],
    opcode=efa.WROpcode.RDMA_WRITE,
    send_flags=efa.SendFlags.SIGNALED,
    remote_addr=remote_dst_addr,
    rkey=remote_dst_rkey,
    dest=peer,
))

After the receiver observes a completion or protocol-level write notification, it must order inbound NIC writes before CUDA consumes the destination:

with torch.cuda.device(dst.device):
    efa.cuda.flush_gpudirect_writes()

register_tensor first exports a dma-buf fd and calls ibv_reg_dmabuf_mr. If that path is unavailable, it falls back to ibv_reg_mr, which requires nvidia_peermem. The returned GpuMR retains the tensor allocation and keeps the actual CUDA virtual address because ibv_mr.addr is not meaningful for a dma-buf MR.

Direct EFA API

Area	API
Device capabilities	Context.query_efa_device
EFA CQ creation	Context.create_cq_ex
Sender and unsolicited metadata	CQEx.poll, WC.sgid, WC.unsolicited
CQ layout	CQ.query_efa, CQEx.query_efa
SQ/RQ layout	QP.query_wqs
MR interconnect IDs	MR.query_efa
Address-handle number	AH.ahn
SRD QP creation	PD.create_qp with QPType.SRD

When using unsolicited RDMA write-with-immediate completions, create every communicating QP with QPInitAttr(..., unsolicited_write_recv=True) and use an extended CQ created with unsolicited=True. EFA requires peers to negotiate the same QP feature set.

The direct layout queries expose process-local addresses for advanced consumers. They do not transfer ownership of provider memory.

Testing

pip install -e "./efa[test,gpu]"
cd efa
pytest -rs

Tests marked integration exercise real EFA hardware. Tests marked gpu perform torch-verified GPU-to-GPU, GPU-to-host, and host-to-GPU transfers. Unavailable hardware capabilities are skipped explicitly.

License

BSD-3-Clause. See LICENSE.