deepseek-ai/DeepEP

View on GitHub

DeepEP: an efficient expert-parallel communication library

Stars: 10.1k
Forks: 1.4k
Open beginner issues: 0
Indexed issues: 202
Avg merge: 4d 1h
Merged PRs (30d): 2

Dominant language: Cuda
License: MIT
Last GitHub push: Aug 4, 2026
Latest indexed: Sep 15, 2026
Contributing guide: No contributing guide
Code of conduct: No code of conduct
Beginner labels: No beginner labels indexed

Open All

202 open issues indexed

Ada Lovelace support Open

deepseek-ai/DeepEP#7 · 8 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
When testing test_internode.py, what is the effect of setting NVSHMEM_DISABLE_P2P=1? Will NVLink be disabled? Open

deepseek-ai/DeepEP#30 · 5 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
Why is a mechanism like PXN in NCCL not used in low-latency kernel? It should be able to achieve even lower latency. Open

deepseek-ai/DeepEP#31 · 3 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
Is setting IBGDA necessary for test_internode.py? Open

deepseek-ai/DeepEP#36 · 17 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
Intra-node compute-communication overlap without SM cores Open

deepseek-ai/DeepEP#37 · 7 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
Test test_low_latency.py failed on H100 with ROCE Open

deepseek-ai/DeepEP#38 · 38 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
Easier Potential Overall Design Open

deepseek-ai/DeepEP#39 · 0 comments · 44 reactions · 0 assignees · Indexed Sep 15, 2026
About num_qps_per_rank Open

deepseek-ai/DeepEP#46 · 6 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
How to apply Moe MLP to recv_x returned by dispatch when num_expert_per_rank > 1 Open

deepseek-ai/DeepEP#48 · 2 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
Assertion failed: internode_ll.cu:259, condition: num_recv_tokens != 0 Open

deepseek-ai/DeepEP#49 · 5 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
Why NUM_MAX_NVL_PEERS must be 8? Open

deepseek-ai/DeepEP#50 · 5 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
Confuse about the estimated bandwidth on tests/test_internode.py Open

deepseek-ai/DeepEP#51 · 7 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
Why is it possible to disable congestion control? Open

deepseek-ai/DeepEP#52 · 3 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
test_low_latency failed Open

deepseek-ai/DeepEP#55 · 31 comments · 1 reaction · 0 assignees · Indexed Sep 15, 2026
output tokens in intranode::dispatch for each expert isn't packed? Open

deepseek-ai/DeepEP#58 · 1 comment · 0 reactions · 0 assignees · Indexed Sep 15, 2026
AllToAll performance not improving Open

deepseek-ai/DeepEP#59 · 5 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
Questions about Compute-Communication Overlap Open

deepseek-ai/DeepEP#61 · 4 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
Confues about the Low-latency kernels data issue Open

deepseek-ai/DeepEP#71 · 1 comment · 0 reactions · 0 assignees · Indexed Sep 15, 2026
Question about interface of low_latency_dispatch Open

deepseek-ai/DeepEP#72 · 3 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
curious about the difference between latency and throughput Open

deepseek-ai/DeepEP#75 · 4 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
Could normal kernel and ll kernel be executed in same process? Open

deepseek-ai/DeepEP#76 · 6 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
About the number of messages chunked in IBGDA Open

deepseek-ai/DeepEP#80 · 2 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
Benchmark test over RoCE network Open

deepseek-ai/DeepEP#82 · 17 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
socketStartConnect: exceeded timeouts (3)socketStartConnect: exceeded timeouts (3) Open

deepseek-ai/DeepEP#86 · 1 comment · 0 reactions · 0 assignees · Indexed Sep 15, 2026
"RuntimeError: Failed: CUDA error /opt/DeepEP/csrc/kernels/internode.cu:133 ' named symbol not found" in container when running test_intranode.py Open

deepseek-ai/DeepEP#87 · 3 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
Will the Normal Kernel (dispatch kernel) exclusively occupy the SM of the GPU? Open

deepseek-ai/DeepEP#92 · 1 comment · 0 reactions · 0 assignees · Indexed Sep 15, 2026
Question about the global synchronization/barrier in internode notify_dispatch Open

deepseek-ai/DeepEP#94 · 7 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
Suggestions for Improving the Readability of DeepEP Code Open

deepseek-ai/DeepEP#98 · 12 comments · 2 reactions · 0 assignees · Indexed Sep 15, 2026
`python test_low_latency.py` blocked in H20 Open

deepseek-ai/DeepEP#103 · 1 comment · 0 reactions · 0 assignees · Indexed Sep 15, 2026
deepEP run two nodes, each nodes equipped 1 A100+1CX6, hit assert. Open

deepseek-ai/DeepEP#104 · 5 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
how can I run the test across multi nodes？ Open

deepseek-ai/DeepEP#106 · 4 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
some questions in ibgda Open

deepseek-ai/DeepEP#113 · 2 comments · 1 reaction · 0 assignees · Indexed Sep 15, 2026
Crash when test DeepEP over 16 H100 Servers Open

deepseek-ai/DeepEP#114 · 2 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
test_inter_node.py failed: an illegal memory access was encountered Open

deepseek-ai/DeepEP#126 · 3 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
[intranode] when num_tokens change to 65536, then check failed Open

deepseek-ai/DeepEP#127 · 2 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
Ways to do overlapping of computation and commucation. Open

deepseek-ai/DeepEP#131 · 1 comment · 0 reactions · 0 assignees · Indexed Sep 15, 2026
error: nvshmem API called before nvshmem_init Open

deepseek-ai/DeepEP#133 · 6 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
Assertion failed: /DeepEP/csrc/kernels/internode.cu:483, condition: ibgda_get_state()->num_rc_per_pe >= num_channels Open

deepseek-ai/DeepEP#134 · 2 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
`L1::no_allocate` is not safe to load volatile data Open

deepseek-ai/DeepEP#136 · 3 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
[test_internode.py] failed on multi-QP: dispatch timeout on ROCE network with testing 2*H20 nodes Open

deepseek-ai/DeepEP#137 · 39 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
when i installed nvshmem_src, nccl get error, how to fix? Open

deepseek-ai/DeepEP#143 · 3 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
Profiling Results on H20 with DSv3 Open

deepseek-ai/DeepEP#144 · 1 comment · 1 reaction · 0 assignees · Indexed Sep 15, 2026
Caught signal 11 (Segmentation fault: address not mapped to object at address 0x8000008f) Open

deepseek-ai/DeepEP#145 · 1 comment · 0 reactions · 0 assignees · Indexed Sep 15, 2026
low latency mode bandwidth Open

deepseek-ai/DeepEP#148 · 5 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
illegal memory on low_latency_dispatch when test dataset Open

deepseek-ai/DeepEP#152 · 6 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
test_low_latency.py Failed on 2 node H800 Open

deepseek-ai/DeepEP#157 · 5 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
[Bug] test_internode.py timeout on 4 * 8 H20 Open

deepseek-ai/DeepEP#158 · 1 comment · 0 reactions · 0 assignees · Indexed Sep 15, 2026
question about the modification of "pybind11::gil_scoped_release release" Open

deepseek-ai/DeepEP#159 · 2 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
Internode low latency test dispatch & combine finished with error Open

deepseek-ai/DeepEP#160 · 5 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026
Setting deepep to low_latency mode fails to start model serving Open

deepseek-ai/DeepEP#165 · 5 comments · 0 reactions · 0 assignees · Indexed Sep 15, 2026

deepseek-ai/DeepEP

Get new issues in your inbox