deepseek-ai/DeepEP-Ascend

DeepEP-Ascend

（中文介绍）DeepEP-Ascend 是面向华为昇腾 NPU 的高性能机器学习训练与推理通信库，提供 MoE dispatch/combine 的专家并行（EP）all-to-all 操作，支持 FP8 dispatch 和延迟 epilogue。此外还提供流水线并行（PP）、面向上下文并行和数据并行（CP/DP）的 Bucket 集合通信，以及 Engram 远端内存访问等通信原语（开发中）。其公开 buffer API 与 NVIDIA 版 DeepEP 对齐。Ascend C 内核使用 HCCL/HCOMM、UBMEM 和 URMA 完成通信，并通过 DeepJIT 在运行时编译。

(English introduction) DeepEP-Ascend is a high-performance communication library for machine learning training and inference on Huawei Ascend NPUs. It provides expert-parallel (EP) all-to-all operations for MoE dispatch and combine, including FP8 dispatch and deferred epilogues. It also offers communication primitives (work in progress) for pipeline parallelism (PP), context and data parallelism through Bucket collectives (CP/DP), and remote memory access through Engram. Its public buffer APIs are aligned with the NVIDIA version of DeepEP. Ascend C kernels use HCCL/HCOMM, UBMEM and URMA, and are compiled at runtime through DeepJIT.

Performance

Measured on Ascend 950DT NPUs with CANN 9.2.0 and the manually configured PoC HDK described under Recommended HDK and firmware. All measurements use netlayer 1, the supernode's external Clos network.

The following runs use 16,384 tokens/rank capacity, hidden size 7168, top-6 routing over 256 experts, 32 AI cores / 64 AIVs, expert alignment 128 and no bias. Dispatch uses FP8 with row-major scales; combine uses BF16. Each rank uses 10 warmups and 50 samples with cache flushing.

| EP size | Dispatch (GB/s) | Combine (GB/s) | | --- | ---: | ---: | | EP8 | 373–375 | 345–347 | | EP16 | 348–352 | 338–341 | | EP32 | 335–340 | 320–324 | | EP64 | 323–327 | 294–298 | | EP128 | 313–320 | 272–278 |

Bandwidth ranges cover all ranks, using timings that include issue and drain but exclude final epilogues. Sustained dispatch reaches roughly 90–95% of the physical payload bandwidth limit for EP sizes up to 32. Larger EP sizes and combine remain under optimization; combine has additional local reduction overhead and HBM contention with URMA. Huawei firmware improvements are also planned to reduce this contention.

Quick start

Requirements

• Linux on an Ascend host with supported NPU drivers and firmware.
• Ascend 950 (A5) with UBMEM connectivity and UBC_CTP/URMA channels between the participating ranks.
• An even number of ranks for multi-rank communication.
• CANN with Ascend C, the Bisheng compiler, HCCL and HCOMM headers/libraries.
• Python 3.10 or newer, a matching PyTorch/torch_npu installation, and NumPy for the tests.
• A C++20 compiler and standard library supporting std::format.

The validated stack is Ascend 950DT, CANN 9.2.0, Python 3.12, PyTorch 2.13.0+cpu and torch_npu 2.13.0rc1, with the PoC HDK configuration described below. Kernel support on other Ascend generations or CANN versions has not been established by these measurements.

Installation builds the host extension; device kernels are compiled on first use. Keep the CANN toolkit and Bisheng available at runtime. Set ASCEND_HOME_PATH to the active CANN installation, or use ASCEND_TOOLKIT_HOME.

Recommended HDK and firmware

Use Huawei's Q3 commercial HDK release for Atlas 850E when it becomes publicly available. Huawei has advised that this release includes the configuration required for full-bandwidth operation. Public availability is currently planned for mid-October 2026, around October 15, through the Huawei Atlas 850E software download page, where users can apply for access when the release becomes available. The date is a vendor release plan; availability is subject to Huawei's publication schedule.

The performance results in this README were collected using a PoC HDK supplied to DeepSeek, with additional manual configuration. That configuration is not a publicly distributed release. Users on earlier PoC versions may observe lower bandwidth and should first check their HDK version and configuration. Use the Q3 commercial release, once available, as the recommended public deployment baseline; these measurements are not results from that unreleased commercial version.

Installation

Activate the CANN environment and install the matching PyTorch/torch_npu packages first, then:

git clone https://github.com/deepseek-ai/DeepEP-Ascend.git
cd DeepEP-Ascend
Fetch the required DeepJIT submodule over HTTPS.
git submodule update --init --recursive third-party/deep_jit
python -m pip install --no-build-isolation .

For development, use bash develop.sh to build and link the extension into the source tree. The clangd-ascend submodule and CMake configuration are for IDE/debugging support.

Features and API compatibility

Training, inference prefill and decoding share the same EPBuffer interface and the deep_ep Python package name. The implementation follows DeepEP's EPBuffer-based V2.5 APIs; supported modes and stream behavior are specific to Ascend.

| Interface | Ascend support | | --- | --- | | EPBuffer | Expanded BF16/FP8 dispatch, BF16 combine, cached handles, deterministic layouts, expert padding and deferred epilogues | | PPBuffer | Send/receive between adjacent pipeline ranks | | EngramBuffer | NPU-backed BF16/FP8 tables and per-layer fetch hooks | | BucketBuffer | Batched all-gather, registered storage, multiple groups and sessions for ordinary tensors | | BufferAllocator, BufferBase, deep_ep.comm | Allocation planning, buffer lifecycle, communicator queries, barriers and stream access |

PP, Engram and Bucket are experimental; remaining work is listed under Ongoing. EP requires do_expand=True and allow_multiple_reduction=True; do_handle_copy must remain False.

Routing indices use deep_ep.topk_idx_t, currently fixed to torch.int64. Explicit RDMA service levels and QP counts are unsupported; leave sl_idx=None and QP arguments at their defaults. The NV-compatible num_sms argument selects AI cores on Ascend EP, with 0 selecting the default; cached dispatch and combine reuse the count saved in the handle.

Ongoing

The Ascend backend is under active development:

• Bucket collectives: batched all-gather is available; Ascend reduce-scatter and all-reduce kernels are still being built.
• Expert load balancing: EPBuffer.lb_prefetch_weights and EPBuffer.lb_reduce_grads expose the aligned APIs, but their Ascend communication kernels are not implemented yet.
• PP and Engram: these interfaces remain experimental, with further correctness and performance validation ongoing.

Hybrid communication, CPU-backed Engram storage and graph capture remain unsupported.

EP usage in training and inference

Buffer initialization

Create one long-lived buffer per EP group and reuse it across layers or microbatches. All ranks must agree on the token capacity. Complete outstanding operations before resizing, reusing or destroying a buffer.

import torch
import torch.distributed as dist
import torch_npu
from deep_ep import EPBuffer
First select the local NPU and initialize torch.distributed with backend='hccl'.
ep_group = dist.group.WORLD
max_tokens_per_rank = 4096
hidden = 7168
num_topk = 6
num_experts = 256
buffer = EPBuffer(
    ep_group,
    num_max_tokens_per_rank=max_tokens_per_rank,
    hidden=hidden,
    num_topk=num_topk,
    use_fp8_dispatch=False,  # Reserve enough storage for BF16 and FP8 dispatch.
    explicitly_destroy=True,
)

Use EPBuffer.get_buffer_size_hint() to check capacity when managing a buffer pool. After all uses and their waits have finished, call buffer.destroy() before tearing down the process group.

Dispatch and combine

Training, prefill and decoding use the same EPBuffer dispatch and combine APIs. The following helpers use the expanded expert layout and defer epilogues with defer_epilogue=True and async_with_compute_stream=True.

def dispatch_forward(buffer, x, topk_idx, topk_weights, num_experts,
                     expert_alignment=1, do_cpu_sync=True):
    """Route BF16 or FP8 inputs to their selected experts."""
    return buffer.dispatch(
        x,
        topk_idx=topk_idx,
        topk_weights=topk_weights,
        num_experts=num_experts,
        expert_alignment=expert_alignment,  # Match the grouped GEMM's token alignment.
        do_cpu_sync=do_cpu_sync,  # False when expert kernels use device-side counts.
        do_expand=True,
        do_zero_padding=True,
        async_with_compute_stream=True,
        defer_epilogue=True,
    )
def dispatch_backward(buffer, grad_recv_x, grad_recv_weights, handle, bias=None):
    """The backward of dispatch is a combine."""
    return buffer.combine(
        grad_recv_x,
        handle=handle,
        topk_weights=grad_recv_weights,
        bias=bias,
        async_with_compute_stream=True,
        defer_epilogue=True,
    )
def combine_forward(buffer, expert_output, handle, bias=None):
    """Reduce BF16 expert outputs back to their original ranks."""
    return buffer.combine(
        expert_output,
        handle=handle,
        bias=bias,
        async_with_compute_stream=True,
        defer_epilogue=True,
    )
def combine_backward(buffer, grad_output, handle):
    """The backward of combine reuses the forward handle for dispatch."""
    return buffer.dispatch(
        grad_output,
        handle=handle,
        do_expand=True,
        do_zero_padding=True,
        async_with_compute_stream=True,
        defer_epilogue=True,
    )

Launch communication first, run independent computation, then call .wait() when its result is needed. EP kernels and their epilogues run on the current NPU stream:

pending = dispatch_forward(
    buffer, x, topk_idx, topk_weights, num_experts,
    expert_alignment=expert_alignment,
    do_cpu_sync=need_cpu_expert_counts,
)
Overlap application computation with communication.
shared_output = shared_experts(x)
recv_x, _, recv_weights, handle = pending.wait()
Expert GEMMs apply recv_weights once and produce BF16 outputs.
expert_output, expert_ctx = local_experts_forward(recv_x, recv_weights, handle)
pending = combine_forward(buffer, expert_output, handle, bias=shared_output)
... other independent computation ...
output, _ = pending.wait()

Training saves handle for combine_backward and dispatch_backward; wait on their returned events in the same way. See EP tests for complete examples.

Other buffer interfaces

Bucket all-gather

Bucket communication runs on the comm stream; its epilogue runs on the caller's current stream. Set EP_AVOID_RECORD_STREAM=1 before use. A session permits ordinary contiguous NPU inputs:

from deep_ep import BucketBuffer, get_num_allocation_alignment
shard = torch.full((1024,), ep_group.rank(), dtype=torch.float32, device='npu')
alignment = get_num_allocation_alignment()
gathered_bytes = ep_group.size() * shard.nbytes
num_bytes = (gathered_bytes + alignment - 1) // alignment * alignment
bucket = BucketBuffer(
    ep_group, num_bytes, explicitly_destroy=True)
with bucket.session():
    pending = bucket.all_gather(shard)
    gathered = pending.wait()  # Flat, rank-ordered view of session storage.
    # Consume or copy gathered before the session storage is reused.
bucket.destroy()

All ranks must supply equally sized shards, and the bucket must have enough space for the gathered output; sessions do not grow its storage. Outside a session, output storage must belong to the bucket; without explicit dsts, each input must be the local rank's shard of a registered gathered tensor. Use BufferAllocator to plan persistent registered storage. Batches support up to 64 tensors, and multi-group buffers select a group with group=. Session exit adds no barrier. See all-gather tests and session tests.

Pipeline parallelism and Engram

PPBuffer reserves send/receive storage with num_max_tensor_bytes and num_max_inflight_tensors, then provides send(x, dst_rank_idx) and recv(x, src_rank_idx) for adjacent pipeline ranks. See PP tests.

EngramBuffer uses get_storage_size_hint(), set_config() and write() to prepare NPU-backed tables. fetch(indices) returns per-layer completion hooks; invoke each hook before consuming its result. See Engram tests, including FP8 mode. PP and Engram run entirely on the current stream.

Environment and tests

Set environment variables consistently on every rank, before importing the package or constructing buffers.

| Variable | Use | | --- | --- | | EP_AVOID_RECORD_STREAM | Required for Bucket all-gather; EP retains URMA sources through epilogue captures independently of this flag | | EP_BUFFER_DEBUG | Print EP buffer initialization diagnostics; unset to disable | | EP_DISABLE_BARRIER_PROFILING | Disable the extra per-iteration synchronization used by benchmark helpers |

DeepJIT settings use EP_JIT_* for DeepEP, falling back to the corresponding DJ_JIT_* process-wide setting and then the built-in default. For example, EP_JIT_CACHE_DIR takes precedence over DJ_JIT_CACHE_DIR. The Ascend backend supports:

| Variable | Default | Use | | --- | --- | --- | | EP_JIT_CACHE_DIR | $HOME/.dj | Cache root or colon-separated roots; search all roots in order and compile misses into the first | | EP_JIT_DEBUG | 0 | Print compiler commands, binary paths and load timings | | EP_JIT_DUMP_ASM | 0 | Save Ascend assembly under the cache entry's asm/ directory on a cache miss | | EP_JIT_KERNEL_DEBUG_INFO | 0 | Include kernel source and line information for debugging | | EP_JIT_LAUNCH_TIMEOUT | 300 | Kernel launch timeout in seconds; 0 disables it | | EP_JIT_PRINT_COMPILER_COMMAND | 0 | Print Bisheng and linker commands | | EP_JIT_PRINT_LOAD_TIME | 0 | Print kernel binary load timings |

See DeepJIT's environment reference for details.

To reproduce the default EP performance cases after activating the CANN environment:

export EP_AVOID_RECORD_STREAM=1
export TASK_QUEUE_ENABLE=0
export HCCL_IF_BASE_PORT=48361
export ASCEND_GLOBAL_LOG_LEVEL=1
export ASCEND_PROCESS_LOG_PATH=/tmp/deepep-ascend-logs
mkdir -p "$ASCEND_PROCESS_LOG_PATH"
python tests/ep/test_ep.py --num-processes 8 --num-ai-cores 32 \
    --num-tokens 16384 --hidden 7168 --num-topk 6 --num-experts 256 \
    --dispatch-dtype bf16 --test-first-only
python tests/ep/test_ep.py --num-processes 8 --num-ai-cores 32 \
    --num-tokens 16384 --hidden 7168 --num-topk 6 --num-experts 256 \
    --dispatch-dtype fp8 --test-first-only

For multiple nodes, launch the same command on each node with a shared MASTER_ADDR and MASTER_PORT, setting WORLD_SIZE to the number of nodes and RANK to the node index. Omit --test-first-only and the dtype filter for the full mode matrix. Additional checks include:

python tests/buffer/test_allocation.py
python tests/comm/test_stream.py
python tests/comm/test_barrier.py
python tests/bucket/test_all_gather.py
python tests/bucket/test_session.py
python tests/pp/test_pp.py
python tests/engram/test_engram.py
python tests/engram/test_engram.py --use-fp8
python tests/utils/test_gate.py

The test helpers in envs.py interpret WORLD_SIZE and RANK as node count and node index, and spawn --num-processes local workers. Multi-node invocations also require a shared MASTER_ADDR and MASTER_PORT; these launch conventions do not imply that every topology is supported by the current backend.

Acknowledgement

DeepEP-Ascend is built on top of Huawei's HCCL/HCOMM, UBMEM and URMA communication stack. Thanks to the asc-comm team for their support!

Citation

@misc{deepepascend2026,
    title = {{DeepEP-Ascend}: an efficient expert-parallel communication library for {Ascend} {NPUs}},
    author = {Chenggang Zhao and Shangyan Zhou and Kexing Zhou and Rui Tian and Chenqi Zhao and Chenhao Xu and Yizhi Wang and Kuai Yu},
    year = {2026},
    publisher = {GitHub},
    howpublished = {\url{https://github.com/deepseek-ai/DeepEP-Ascend}},
}