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
abi3wheel for CPython 3.9 and newer.
Requirements
- Linux on an AWS instance with one or more EFA devices attached.
libibverbs.so.1andlibefa.so.1at runtime. The AWS EFA installer and currentrdma-coredistributions provide both.- The EFA device nodes available inside the process or container, normally
/dev/infiniband/uverbs*. - A C compiler, Cython, and the
rdma-coredevelopment 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. Set
RDMA4PY_SKIP_HARDWARE_TESTS=1 to force all hardware-dependent tests to skip.
License
BSD-3-Clause. See LICENSE.
1 maintainer
Metadata
1 maintainer
Release files for efa 2026.10.16
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | |
|---|---|---|---|
| efa-2026.10.16.tar.gz | 394.0 kB | Details |
Built distributions (wheels)
| File | Interpreter | ABI | Platform | Reset |
|---|---|---|---|---|
| efa-2026.10.16-cp39-abi3-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl | CPython 3.9 | abi3 | Linux glibc 2.17+ ARM64, Linux glibc 2.28+ ARM64 | Details |
| efa-2026.10.16-cp39-abi3-manylinux1_x86_64.manylinux_2_28_x86_64.manylinux_2_5_x86_64.whl | CPython 3.9 | abi3 | Linux glibc 2.5+ x86-64, Linux glibc 2.28+ x86-64 | Details |
Total release size: 3.0 MB
Release files / efa-2026.10.16.tar.gz
| Download URL | efa-2026.10.16.tar.gz |
|---|---|
| Size | 394.0 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
ad82a52f9a07dbaa1324994f7cdd8d86b9881c8c7d1eb5a76fc2d398151efb86
|
|
BLAKE2b-256 checksum How to use checksums |
763833f584b85e607104458acd14ce21901a64c3e958b893f757c2f9a88d0732
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.14
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Oct 16, 2026.
Transparency logRelease files / efa-2026.10.16-cp39-abi3-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl
| Download URL | efa-2026.10.16-cp39-abi3-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl |
|---|---|
| Size | 1.3 MB |
| Tags | CPython 3.9 Linux glibc 2.17+ ARM64 Linux glibc 2.28+ ARM64 abi3 |
|
SHA-256 checksum How to use checksums |
eee39b6d2c324f76b8bf1c71489fac17644cc3f279c251fc6093dc24f98cd026
|
|
BLAKE2b-256 checksum How to use checksums |
9c06c93eb142d795d52f990758ad2d83134778aea27cdc10d236a81274adafee
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.14
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Oct 16, 2026.
Transparency logRelease files / efa-2026.10.16-cp39-abi3-manylinux1_x86_64.manylinux_2_28_x86_64.manylinux_2_5_x86_64.whl
| Download URL | efa-2026.10.16-cp39-abi3-manylinux1_x86_64.manylinux_2_28_x86_64.manylinux_2_5_x86_64.whl |
|---|---|
| Size | 1.3 MB |
| Tags | CPython 3.9 Linux glibc 2.28+ x86-64 Linux glibc 2.5+ x86-64 abi3 |
|
SHA-256 checksum How to use checksums |
2080d74a41c6adf470baaf606fe9ad8efc78da4e4ac62e257e1dc2bb862326eb
|
|
BLAKE2b-256 checksum How to use checksums |
aa5355be53d53f1ee58219da773bf34099e8d651b0a9cb9e310e586e9aa114f3
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.14
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Oct 16, 2026.
Transparency log
