deepseek-ai/DeepEP
1
0
Fork 0
mirror of https://github.com/deepseek-ai/DeepEP.git synced 2026-11-05 00:47:32 +00:00
Code Issues 308 Projects Packages Wiki Activity

[GH-ISSUE #590] Poor HybridEP performance on dual-node 16x B300 GPUs - IB bandwidth significantly lower than expected #361

New issue
Open
opened 2026-10-26 09:59:09 +00:00 by bsakaguchi · 2 comments
bsakaguchi commented 2026-10-26 09:59:09 +00:00
Owner
Copy link

Originally created by @sssssakuraaaaa on GitHub (Mar 12, 2026).
Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/590

I'm experiencing poor performance with HybridEP on a dual-node setup with 16 B300 GPUs (8 GPUs per node). The Roce bandwidth is significantly lower than expected, causing performance degradation in cross-node communication.

Environment
Hardware: Dual-node setup, 16x NVIDIA B300 GPUs (8 GPUs per node)
Network: RoCE + NVLink (NVL)
Test Configuration：
Processes per node: 8
Token size: 4096
Local experts: 8
Hidden size: 7168
Node num: 2
Nranks: 16
Performance Data

BF16 performance

Operation NVL Bandwidth IB Bandwidth Ratio (IB/NVL)
Dispatch (torch API) ~13.6 GB/s ~4.98 GB/s 36.6%
Dispatch+permute ~34.5 GB/s ~12.6 GB/s 36.5%
Combine (torch API) ~133 GB/s ~49 GB/s 36.8%
Combine+unpermute ~196 GB/s ~71.8 GB/s 36.6%
Dispatch kernel ~160-282 GB/s ~58-103 GB/s ~36%
Combine kernel ~234-245 GB/s ~86-89 GB/s ~37%

FP8 performance

Operation NVL Bandwidth IB Bandwidth Ratio (IB/NVL)
Dispatch (torch API) ~66-69 GB/s ~25 GB/s ~37%
Dispatch+permute ~65-69 GB/s ~24.9 GB/s ~37%
Combine (torch API) ~201-212 GB/s ~76.4 GB/s ~37%
Combine+unpermute ~101-107 GB/s ~38.5 GB/s ~37%
Dispatch kernel ~78-133 GB/s ~28-48 GB/s ~37%
Combine kernel ~264-286 GB/s ~97-103 GB/s ~37%

The HybridEP performance for EP 16 is much worse than DeepEP。During HybridEP running, there is no log except performance log。 how can I debug this issue ?

Originally created by @sssssakuraaaaa on GitHub (Mar 12, 2026). Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/590 I'm experiencing poor performance with HybridEP on a dual-node setup with 16 B300 GPUs (8 GPUs per node). The Roce bandwidth is significantly lower than expected, causing performance degradation in cross-node communication. Environment Hardware: Dual-node setup, 16x NVIDIA B300 GPUs (8 GPUs per node) Network: RoCE + NVLink (NVL) Test Configuration： Processes per node: 8 Token size: 4096 Local experts: 8 Hidden size: 7168 Node num: 2 Nranks: 16 Performance Data BF16 performance Operation | NVL Bandwidth | IB Bandwidth | Ratio (IB/NVL) -- | -- | -- | -- Dispatch (torch API) | ~13.6 GB/s | ~4.98 GB/s | 36.6% Dispatch+permute | ~34.5 GB/s | ~12.6 GB/s | 36.5% Combine (torch API) | ~133 GB/s | ~49 GB/s | 36.8% Combine+unpermute | ~196 GB/s | ~71.8 GB/s | 36.6% Dispatch kernel | ~160-282 GB/s | ~58-103 GB/s | ~36% Combine kernel | ~234-245 GB/s | ~86-89 GB/s | ~37% FP8 performance Operation | NVL Bandwidth | IB Bandwidth | Ratio (IB/NVL) -- | -- | -- | -- Dispatch (torch API) | ~66-69 GB/s | ~25 GB/s | ~37% Dispatch+permute | ~65-69 GB/s | ~24.9 GB/s | ~37% Combine (torch API) | ~201-212 GB/s | ~76.4 GB/s | ~37% Combine+unpermute | ~101-107 GB/s | ~38.5 GB/s | ~37% Dispatch kernel | ~78-133 GB/s | ~28-48 GB/s | ~37% Combine kernel | ~264-286 GB/s | ~97-103 GB/s | ~37% The HybridEP performance for EP 16 is much worse than DeepEP。During HybridEP running, there is no log except performance log。 how can I debug this issue ?
bsakaguchi commented 2026-10-26 09:59:11 +00:00
Author
Owner
Copy link

@alpha-baby commented on GitHub (Apr 17, 2026):

EP 16 hybrid-ep performance

@Autumn1998 Could you please advise if the performance from my testing meets expectations? I tuned the configuration parameters, and this set currently appears to deliver the best performance.

env ：

"NUM_OF_TOKENS_PER_CHUNK_DISPATCH_API=64"
"NUM_OF_TOKENS_PER_CHUNK_COMBINE_API=64"
"NUM_OF_TOKENS_PER_CHUNK_PREPROCESSING_API=64"
"NUM_SMS_DISPATCH=24"
"NUM_SMS_COMBINE=24"
"NUM_BLOCKS_PERMUTE=124"
"NUM_BLOCKS_UNPERMUTE=124"
"CUDA_HOME=/usr/local/cuda"
"HIDDEN_DIM=7168"
"MAX_NUM_OF_TOKENS_PER_RANK=4096"
"NUM_TOKENS_PER_RANK=4096"
"NUM_LOCAL_EXPERTS=8"
"TOPK=8"
"NVSHMEM_DEBUG=WARN"
"NCCL_DEBUG=WARN"


MASTER_ADDR=xxxx MASTER_PORT=29500 WORLD_SIZE=2 RANK=0 NUM_OF_TOKENS_PER_CHUNK_DISPATCH_API=64 NUM_OF_TOKENS_PER_CHUNK_COMBINE_API=64 NUM_OF_TOKENS_PER_CHUNK_PREPROCESSING_API=64 NUM_SMS_DISPATCH=24 NUM_SMS_COMBINE=24 NUM_BLOCKS_PERMUTE=124 NUM_BLOCKS_UNPERMUTE=124 CUDA_HOME=/usr/local/cuda HIDDEN_DIM=7168 MAX_NUM_OF_TOKENS_PER_RANK=4096 NUM_TOKENS_PER_RANK=4096 NUM_LOCAL_EXPERTS=8 TOPK=8 NVSHMEM_DEBUG=WARN NCCL_DEBUG=WARN nohup python3 /root/hybrid-ep/tests/test_hybrid_ep.py --num-processes 8  > /tmp/bench_mark.txt 2>&1 &

dispatch kernel (BF16)(NVL):                     149.24 GB/s, avg_t=2638.9 us [min=2609.0, max=2671.0]
combine kernel(NVL):                             135.98 GB/s, avg_t=2896.4 us [min=2862.0, max=3111.0]
dispatch kernel (BF16)(RDMA):                    44.38 GB/s, avg_t=2638.9 us [min=2609.0, max=2671.0]
combine kernel(RDMA):                            40.43 GB/s, avg_t=2896.4 us [min=2862.0, max=3111.0]
dispatch (BF16):                                 136.54 GB/s (NVL), t: 2884.3 us [min=2884.1, max=2884.6], nvl_recv_bytes: 393.84 MB
dispatch (BF16):                                 40.60 GB/s (RDMA), t: 2884.3 us [min=2884.1, max=2884.6], rdma_send_bytes: 117.11 MB
combine:                                         128.92 GB/s (NVL), t: 3054.8 us [min=3054.6, max=3055.0], combine_send_bytes: 393.84 MB
combine:                                         38.34 GB/s (RDMA), t: 3054.8 us [min=3054.6, max=3055.0], rdma_recv_bytes: 117.11 MB
dispatch+permute (BF16):                         126.49 GB/s (NVL), t: 3113.6 us [min=3113.3, max=3113.8], nvl_recv_bytes: 393.84 MB
dispatch+permute (BF16):                         37.61 GB/s (RDMA), t: 3113.6 us [min=3113.3, max=3113.8], rdma_send_bytes: 117.11 MB
combine+unpermute:                               106.80 GB/s (NVL), t: 3687.6 us [min=3686.8, max=3687.9], combine_send_bytes: 393.84 MB
combine+unpermute:                               31.76 GB/s (RDMA), t: 3687.6 us [min=3686.8, max=3687.9], rdma_recv_bytes: 117.11 MB
fused dispatch+permute (BF16):                   123.63 GB/s (NVL), t: 3185.7 us [min=3185.5, max=3186.0], nvl_recv_bytes: 393.84 MB
fused dispatch+permute (BF16):                   36.76 GB/s (RDMA), t: 3185.7 us [min=3185.5, max=3186.0], rdma_send_bytes: 117.11 MB
fused combine+unpermute:                         128.32 GB/s (NVL), t: 3069.1 us [min=3068.8, max=3069.5], combine_send_bytes: 393.84 MB
fused combine+unpermute:                         38.16 GB/s (RDMA), t: 3069.1 us [min=3068.8, max=3069.5], rdma_recv_bytes: 117.11 MB

=== Kernel Benchmark (BF16, 16 ranks) ===
  Non-fused:  dispatch_kernel only  |  combine_kernel only
  Fused:      fused_permute_dispatch_kernel only  |  fused_combine_unpermute_kernel only
dispatch kernel (BF16)(NVL):                     149.24 GB/s, avg_t=2638.9 us [min=2609.0, max=2671.0]
combine kernel(NVL):                             135.98 GB/s, avg_t=2896.4 us [min=2862.0, max=3111.0]
dispatch kernel (BF16)(RDMA):                    44.38 GB/s, avg_t=2638.9 us [min=2609.0, max=2671.0]
combine kernel(RDMA):                            40.43 GB/s, avg_t=2896.4 us [min=2862.0, max=3111.0]
fused dispatch+permute kernel (BF16)(NVL):       147.96 GB/s, avg_t=2661.8 us [min=2649.0, max=2673.0]
fused combine+unpermute kernel(NVL):             130.83 GB/s, avg_t=3010.2 us [min=2943.0, max=3168.0]
fused dispatch+permute kernel (BF16)(RDMA):      44.00 GB/s, avg_t=2661.8 us [min=2649.0, max=2673.0]
fused combine+unpermute kernel(RDMA):            38.90 GB/s, avg_t=3010.2 us [min=2943.0, max=3168.0]

=== Correctness Check (FP8, 16 ranks) ===
  dispatch+combine API: PASS
  dispatch_with_permute + combine_with_unpermute API (non-fused): PASS
  dispatch_with_permute + combine_with_unpermute API (fused): PASS

=== Torch API Benchmark (FP8, 16 ranks) ===
  Non-permute:  dispatch = dispatch_kernel + d2d + misc
                combine  = d2d + combine_kernel + misc
  Permute:      dispatch = dispatch_kernel + permute_kernel + misc
                combine  = unpermute_kernel + combine_kernel + misc
  Fused:        dispatch = fused_permute_dispatch_kernel + misc
                combine  = fused_combine_unpermute_kernel + misc
  (misc = device_sync, update_flag, etc.)
dispatch (FP8):                                  119.64 GB/s (NVL), t: 1707.3 us [min=1707.1, max=1707.5], nvl_recv_bytes: 204.26 MB
dispatch (FP8):                                  35.38 GB/s (RDMA), t: 1707.3 us [min=1707.1, max=1707.5], rdma_send_bytes: 60.40 MB
combine:                                         119.84 GB/s (NVL), t: 3305.4 us [min=3305.2, max=3305.6], combine_send_bytes: 396.13 MB
combine:                                         18.27 GB/s (RDMA), t: 3305.4 us [min=3305.2, max=3305.6], rdma_recv_bytes: 60.40 MB
dispatch+permute (FP8):                          105.03 GB/s (NVL), t: 1944.7 us [min=1944.4, max=1945.4], nvl_recv_bytes: 204.26 MB
dispatch+permute (FP8):                          31.06 GB/s (RDMA), t: 1944.7 us [min=1944.4, max=1945.4], rdma_send_bytes: 60.40 MB
combine+unpermute:                               110.22 GB/s (NVL), t: 3594.0 us [min=3593.4, max=3594.3], combine_send_bytes: 396.13 MB
combine+unpermute:                               16.81 GB/s (RDMA), t: 3594.0 us [min=3593.4, max=3594.3], rdma_recv_bytes: 60.40 MB
fused dispatch+permute (FP8):                    119.36 GB/s (NVL), t: 1711.2 us [min=1711.0, max=1711.5], nvl_recv_bytes: 204.26 MB
fused dispatch+permute (FP8):                    35.30 GB/s (RDMA), t: 1711.2 us [min=1711.0, max=1711.5], rdma_send_bytes: 60.40 MB
fused combine+unpermute:                         121.03 GB/s (NVL), t: 3273.1 us [min=3272.8, max=3273.4], combine_send_bytes: 396.13 MB
fused combine+unpermute:                         18.45 GB/s (RDMA), t: 3273.1 us [min=3272.8, max=3273.4], rdma_recv_bytes: 60.40 MB

=== Kernel Benchmark (FP8, 16 ranks) ===
  Non-fused:  dispatch_kernel only  |  combine_kernel only
  Fused:      fused_permute_dispatch_kernel only  |  fused_combine_unpermute_kernel only
dispatch kernel (FP8)(NVL):                      143.47 GB/s, avg_t=1423.7 us [min=1420.0, max=1430.0]
combine kernel(NVL):                             136.18 GB/s, avg_t=2908.9 us [min=2853.0, max=3075.0]
dispatch kernel (FP8)(RDMA):                     42.43 GB/s, avg_t=1423.7 us [min=1420.0, max=1430.0]
combine kernel(RDMA):                            20.76 GB/s, avg_t=2908.9 us [min=2853.0, max=3075.0]
fused dispatch+permute kernel (FP8)(NVL):        139.93 GB/s, avg_t=1459.7 us [min=1453.0, max=1466.0]
fused combine+unpermute kernel(NVL):             131.87 GB/s, avg_t=3003.9 us [min=2908.0, max=3098.0]
fused dispatch+permute kernel (FP8)(RDMA):       41.38 GB/s, avg_t=1459.7 us [min=1453.0, max=1466.0]
fused combine+unpermute kernel(RDMA):            20.11 GB/s, avg_t=3003.9 us [min=2908.0, max=3098.0]

my machine

B300 + CX8 (800Gb/s)

root@gpulingjun010006136067:/tmp# nvidia-smi topo -m
        GPU0    GPU1    GPU2    GPU3    GPU4    GPU5    GPU6    GPU7    NIC0    NIC1    NIC2    NIC3    NIC4    NIC5    NIC6    NIC7    CPU Affinity    NUMA Affinity   GPU NUMA ID
GPU0     X      NV18    NV18    NV18    NV18    NV18    NV18    NV18    PXB     NODE    NODE    NODE    SYS     SYS     SYS     SYS     0-63,128-191    0               N/A
GPU1    NV18     X      NV18    NV18    NV18    NV18    NV18    NV18    NODE    PXB     NODE    NODE    SYS     SYS     SYS     SYS     0-63,128-191    0               N/A
GPU2    NV18    NV18     X      NV18    NV18    NV18    NV18    NV18    NODE    NODE    PXB     NODE    SYS     SYS     SYS     SYS     0-63,128-191    0               N/A
GPU3    NV18    NV18    NV18     X      NV18    NV18    NV18    NV18    NODE    NODE    NODE    PXB     SYS     SYS     SYS     SYS     0-63,128-191    0               N/A
GPU4    NV18    NV18    NV18    NV18     X      NV18    NV18    NV18    SYS     SYS     SYS     SYS     PXB     NODE    NODE    NODE    64-127,192-255  1               N/A
GPU5    NV18    NV18    NV18    NV18    NV18     X      NV18    NV18    SYS     SYS     SYS     SYS     NODE    PXB     NODE    NODE    64-127,192-255  1               N/A
GPU6    NV18    NV18    NV18    NV18    NV18    NV18     X      NV18    SYS     SYS     SYS     SYS     NODE    NODE    PXB     NODE    64-127,192-255  1               N/A
GPU7    NV18    NV18    NV18    NV18    NV18    NV18    NV18     X      SYS     SYS     SYS     SYS     NODE    NODE    NODE    PXB     64-127,192-255  1               N/A
NIC0    PXB     NODE    NODE    NODE    SYS     SYS     SYS     SYS      X      NODE    NODE    NODE    SYS     SYS     SYS     SYS
NIC1    NODE    PXB     NODE    NODE    SYS     SYS     SYS     SYS     NODE     X      NODE    NODE    SYS     SYS     SYS     SYS
NIC2    NODE    NODE    PXB     NODE    SYS     SYS     SYS     SYS     NODE    NODE     X      NODE    SYS     SYS     SYS     SYS
NIC3    NODE    NODE    NODE    PXB     SYS     SYS     SYS     SYS     NODE    NODE    NODE     X      SYS     SYS     SYS     SYS
NIC4    SYS     SYS     SYS     SYS     PXB     NODE    NODE    NODE    SYS     SYS     SYS     SYS      X      NODE    NODE    NODE
NIC5    SYS     SYS     SYS     SYS     NODE    PXB     NODE    NODE    SYS     SYS     SYS     SYS     NODE     X      NODE    NODE
NIC6    SYS     SYS     SYS     SYS     NODE    NODE    PXB     NODE    SYS     SYS     SYS     SYS     NODE    NODE     X      NODE
NIC7    SYS     SYS     SYS     SYS     NODE    NODE    NODE    PXB     SYS     SYS     SYS     SYS     NODE    NODE    NODE     X 

Legend:

  X    = Self
  SYS  = Connection traversing PCIe as well as the SMP interconnect between NUMA nodes (e.g., QPI/UPI)
  NODE = Connection traversing PCIe as well as the interconnect between PCIe Host Bridges within a NUMA node
  PHB  = Connection traversing PCIe as well as a PCIe Host Bridge (typically the CPU)
  PXB  = Connection traversing multiple PCIe bridges (without traversing the PCIe Host Bridge)
  PIX  = Connection traversing at most a single PCIe bridge
  NV#  = Connection traversing a bonded set of # NVLinks

NIC Legend:

  NIC0: mlx5_bond_0
  NIC1: mlx5_bond_1
  NIC2: mlx5_bond_2
  NIC3: mlx5_bond_3
  NIC4: mlx5_bond_4
  NIC5: mlx5_bond_5
  NIC6: mlx5_bond_6
  NIC7: mlx5_bond_7

hybrid-ep version:

Image
<!-- gh-comment-id:4267334116 --> @alpha-baby commented on GitHub (Apr 17, 2026): # EP 16 hybrid-ep performance @Autumn1998 Could you please advise if the performance from my testing meets expectations? I tuned the configuration parameters, and this set currently appears to deliver the best performance. env ： ``` "NUM_OF_TOKENS_PER_CHUNK_DISPATCH_API=64" "NUM_OF_TOKENS_PER_CHUNK_COMBINE_API=64" "NUM_OF_TOKENS_PER_CHUNK_PREPROCESSING_API=64" "NUM_SMS_DISPATCH=24" "NUM_SMS_COMBINE=24" "NUM_BLOCKS_PERMUTE=124" "NUM_BLOCKS_UNPERMUTE=124" "CUDA_HOME=/usr/local/cuda" "HIDDEN_DIM=7168" "MAX_NUM_OF_TOKENS_PER_RANK=4096" "NUM_TOKENS_PER_RANK=4096" "NUM_LOCAL_EXPERTS=8" "TOPK=8" "NVSHMEM_DEBUG=WARN" "NCCL_DEBUG=WARN" ``` ``` MASTER_ADDR=xxxx MASTER_PORT=29500 WORLD_SIZE=2 RANK=0 NUM_OF_TOKENS_PER_CHUNK_DISPATCH_API=64 NUM_OF_TOKENS_PER_CHUNK_COMBINE_API=64 NUM_OF_TOKENS_PER_CHUNK_PREPROCESSING_API=64 NUM_SMS_DISPATCH=24 NUM_SMS_COMBINE=24 NUM_BLOCKS_PERMUTE=124 NUM_BLOCKS_UNPERMUTE=124 CUDA_HOME=/usr/local/cuda HIDDEN_DIM=7168 MAX_NUM_OF_TOKENS_PER_RANK=4096 NUM_TOKENS_PER_RANK=4096 NUM_LOCAL_EXPERTS=8 TOPK=8 NVSHMEM_DEBUG=WARN NCCL_DEBUG=WARN nohup python3 /root/hybrid-ep/tests/test_hybrid_ep.py --num-processes 8 > /tmp/bench_mark.txt 2>&1 & ``` ``` dispatch kernel (BF16)(NVL): 149.24 GB/s, avg_t=2638.9 us [min=2609.0, max=2671.0] combine kernel(NVL): 135.98 GB/s, avg_t=2896.4 us [min=2862.0, max=3111.0] dispatch kernel (BF16)(RDMA): 44.38 GB/s, avg_t=2638.9 us [min=2609.0, max=2671.0] combine kernel(RDMA): 40.43 GB/s, avg_t=2896.4 us [min=2862.0, max=3111.0] dispatch (BF16): 136.54 GB/s (NVL), t: 2884.3 us [min=2884.1, max=2884.6], nvl_recv_bytes: 393.84 MB dispatch (BF16): 40.60 GB/s (RDMA), t: 2884.3 us [min=2884.1, max=2884.6], rdma_send_bytes: 117.11 MB combine: 128.92 GB/s (NVL), t: 3054.8 us [min=3054.6, max=3055.0], combine_send_bytes: 393.84 MB combine: 38.34 GB/s (RDMA), t: 3054.8 us [min=3054.6, max=3055.0], rdma_recv_bytes: 117.11 MB dispatch+permute (BF16): 126.49 GB/s (NVL), t: 3113.6 us [min=3113.3, max=3113.8], nvl_recv_bytes: 393.84 MB dispatch+permute (BF16): 37.61 GB/s (RDMA), t: 3113.6 us [min=3113.3, max=3113.8], rdma_send_bytes: 117.11 MB combine+unpermute: 106.80 GB/s (NVL), t: 3687.6 us [min=3686.8, max=3687.9], combine_send_bytes: 393.84 MB combine+unpermute: 31.76 GB/s (RDMA), t: 3687.6 us [min=3686.8, max=3687.9], rdma_recv_bytes: 117.11 MB fused dispatch+permute (BF16): 123.63 GB/s (NVL), t: 3185.7 us [min=3185.5, max=3186.0], nvl_recv_bytes: 393.84 MB fused dispatch+permute (BF16): 36.76 GB/s (RDMA), t: 3185.7 us [min=3185.5, max=3186.0], rdma_send_bytes: 117.11 MB fused combine+unpermute: 128.32 GB/s (NVL), t: 3069.1 us [min=3068.8, max=3069.5], combine_send_bytes: 393.84 MB fused combine+unpermute: 38.16 GB/s (RDMA), t: 3069.1 us [min=3068.8, max=3069.5], rdma_recv_bytes: 117.11 MB === Kernel Benchmark (BF16, 16 ranks) === Non-fused: dispatch_kernel only | combine_kernel only Fused: fused_permute_dispatch_kernel only | fused_combine_unpermute_kernel only dispatch kernel (BF16)(NVL): 149.24 GB/s, avg_t=2638.9 us [min=2609.0, max=2671.0] combine kernel(NVL): 135.98 GB/s, avg_t=2896.4 us [min=2862.0, max=3111.0] dispatch kernel (BF16)(RDMA): 44.38 GB/s, avg_t=2638.9 us [min=2609.0, max=2671.0] combine kernel(RDMA): 40.43 GB/s, avg_t=2896.4 us [min=2862.0, max=3111.0] fused dispatch+permute kernel (BF16)(NVL): 147.96 GB/s, avg_t=2661.8 us [min=2649.0, max=2673.0] fused combine+unpermute kernel(NVL): 130.83 GB/s, avg_t=3010.2 us [min=2943.0, max=3168.0] fused dispatch+permute kernel (BF16)(RDMA): 44.00 GB/s, avg_t=2661.8 us [min=2649.0, max=2673.0] fused combine+unpermute kernel(RDMA): 38.90 GB/s, avg_t=3010.2 us [min=2943.0, max=3168.0] === Correctness Check (FP8, 16 ranks) === dispatch+combine API: PASS dispatch_with_permute + combine_with_unpermute API (non-fused): PASS dispatch_with_permute + combine_with_unpermute API (fused): PASS === Torch API Benchmark (FP8, 16 ranks) === Non-permute: dispatch = dispatch_kernel + d2d + misc combine = d2d + combine_kernel + misc Permute: dispatch = dispatch_kernel + permute_kernel + misc combine = unpermute_kernel + combine_kernel + misc Fused: dispatch = fused_permute_dispatch_kernel + misc combine = fused_combine_unpermute_kernel + misc (misc = device_sync, update_flag, etc.) dispatch (FP8): 119.64 GB/s (NVL), t: 1707.3 us [min=1707.1, max=1707.5], nvl_recv_bytes: 204.26 MB dispatch (FP8): 35.38 GB/s (RDMA), t: 1707.3 us [min=1707.1, max=1707.5], rdma_send_bytes: 60.40 MB combine: 119.84 GB/s (NVL), t: 3305.4 us [min=3305.2, max=3305.6], combine_send_bytes: 396.13 MB combine: 18.27 GB/s (RDMA), t: 3305.4 us [min=3305.2, max=3305.6], rdma_recv_bytes: 60.40 MB dispatch+permute (FP8): 105.03 GB/s (NVL), t: 1944.7 us [min=1944.4, max=1945.4], nvl_recv_bytes: 204.26 MB dispatch+permute (FP8): 31.06 GB/s (RDMA), t: 1944.7 us [min=1944.4, max=1945.4], rdma_send_bytes: 60.40 MB combine+unpermute: 110.22 GB/s (NVL), t: 3594.0 us [min=3593.4, max=3594.3], combine_send_bytes: 396.13 MB combine+unpermute: 16.81 GB/s (RDMA), t: 3594.0 us [min=3593.4, max=3594.3], rdma_recv_bytes: 60.40 MB fused dispatch+permute (FP8): 119.36 GB/s (NVL), t: 1711.2 us [min=1711.0, max=1711.5], nvl_recv_bytes: 204.26 MB fused dispatch+permute (FP8): 35.30 GB/s (RDMA), t: 1711.2 us [min=1711.0, max=1711.5], rdma_send_bytes: 60.40 MB fused combine+unpermute: 121.03 GB/s (NVL), t: 3273.1 us [min=3272.8, max=3273.4], combine_send_bytes: 396.13 MB fused combine+unpermute: 18.45 GB/s (RDMA), t: 3273.1 us [min=3272.8, max=3273.4], rdma_recv_bytes: 60.40 MB === Kernel Benchmark (FP8, 16 ranks) === Non-fused: dispatch_kernel only | combine_kernel only Fused: fused_permute_dispatch_kernel only | fused_combine_unpermute_kernel only dispatch kernel (FP8)(NVL): 143.47 GB/s, avg_t=1423.7 us [min=1420.0, max=1430.0] combine kernel(NVL): 136.18 GB/s, avg_t=2908.9 us [min=2853.0, max=3075.0] dispatch kernel (FP8)(RDMA): 42.43 GB/s, avg_t=1423.7 us [min=1420.0, max=1430.0] combine kernel(RDMA): 20.76 GB/s, avg_t=2908.9 us [min=2853.0, max=3075.0] fused dispatch+permute kernel (FP8)(NVL): 139.93 GB/s, avg_t=1459.7 us [min=1453.0, max=1466.0] fused combine+unpermute kernel(NVL): 131.87 GB/s, avg_t=3003.9 us [min=2908.0, max=3098.0] fused dispatch+permute kernel (FP8)(RDMA): 41.38 GB/s, avg_t=1459.7 us [min=1453.0, max=1466.0] fused combine+unpermute kernel(RDMA): 20.11 GB/s, avg_t=3003.9 us [min=2908.0, max=3098.0] ``` # my machine B300 + CX8 (800Gb/s) ``` root@gpulingjun010006136067:/tmp# nvidia-smi topo -m GPU0 GPU1 GPU2 GPU3 GPU4 GPU5 GPU6 GPU7 NIC0 NIC1 NIC2 NIC3 NIC4 NIC5 NIC6 NIC7 CPU Affinity NUMA Affinity GPU NUMA ID GPU0 X NV18 NV18 NV18 NV18 NV18 NV18 NV18 PXB NODE NODE NODE SYS SYS SYS SYS 0-63,128-191 0 N/A GPU1 NV18 X NV18 NV18 NV18 NV18 NV18 NV18 NODE PXB NODE NODE SYS SYS SYS SYS 0-63,128-191 0 N/A GPU2 NV18 NV18 X NV18 NV18 NV18 NV18 NV18 NODE NODE PXB NODE SYS SYS SYS SYS 0-63,128-191 0 N/A GPU3 NV18 NV18 NV18 X NV18 NV18 NV18 NV18 NODE NODE NODE PXB SYS SYS SYS SYS 0-63,128-191 0 N/A GPU4 NV18 NV18 NV18 NV18 X NV18 NV18 NV18 SYS SYS SYS SYS PXB NODE NODE NODE 64-127,192-255 1 N/A GPU5 NV18 NV18 NV18 NV18 NV18 X NV18 NV18 SYS SYS SYS SYS NODE PXB NODE NODE 64-127,192-255 1 N/A GPU6 NV18 NV18 NV18 NV18 NV18 NV18 X NV18 SYS SYS SYS SYS NODE NODE PXB NODE 64-127,192-255 1 N/A GPU7 NV18 NV18 NV18 NV18 NV18 NV18 NV18 X SYS SYS SYS SYS NODE NODE NODE PXB 64-127,192-255 1 N/A NIC0 PXB NODE NODE NODE SYS SYS SYS SYS X NODE NODE NODE SYS SYS SYS SYS NIC1 NODE PXB NODE NODE SYS SYS SYS SYS NODE X NODE NODE SYS SYS SYS SYS NIC2 NODE NODE PXB NODE SYS SYS SYS SYS NODE NODE X NODE SYS SYS SYS SYS NIC3 NODE NODE NODE PXB SYS SYS SYS SYS NODE NODE NODE X SYS SYS SYS SYS NIC4 SYS SYS SYS SYS PXB NODE NODE NODE SYS SYS SYS SYS X NODE NODE NODE NIC5 SYS SYS SYS SYS NODE PXB NODE NODE SYS SYS SYS SYS NODE X NODE NODE NIC6 SYS SYS SYS SYS NODE NODE PXB NODE SYS SYS SYS SYS NODE NODE X NODE NIC7 SYS SYS SYS SYS NODE NODE NODE PXB SYS SYS SYS SYS NODE NODE NODE X Legend: X = Self SYS = Connection traversing PCIe as well as the SMP interconnect between NUMA nodes (e.g., QPI/UPI) NODE = Connection traversing PCIe as well as the interconnect between PCIe Host Bridges within a NUMA node PHB = Connection traversing PCIe as well as a PCIe Host Bridge (typically the CPU) PXB = Connection traversing multiple PCIe bridges (without traversing the PCIe Host Bridge) PIX = Connection traversing at most a single PCIe bridge NV# = Connection traversing a bonded set of # NVLinks NIC Legend: NIC0: mlx5_bond_0 NIC1: mlx5_bond_1 NIC2: mlx5_bond_2 NIC3: mlx5_bond_3 NIC4: mlx5_bond_4 NIC5: mlx5_bond_5 NIC6: mlx5_bond_6 NIC7: mlx5_bond_7 ``` hybrid-ep version: <img width="2702" height="2208" alt="Image" src="https://github.com/user-attachments/assets/f36b8bce-603d-4439-8ef2-632be75067bf" />
bsakaguchi commented 2026-10-26 09:59:11 +00:00
Author
Owner
Copy link

@Autumn1998 commented on GitHub (Apr 28, 2026):

The original hybrid EP implementation was not adapted for dual-port devices like CX8, so there will be performance issues. I believe your configuration is fine — this is purely a design issue with HEP. For this reason we introduced the nixl backend. We will fix these issues as soon as possible, and we sincerely apologize for any inconvenience caused

<!-- gh-comment-id:4334164990 --> @Autumn1998 commented on GitHub (Apr 28, 2026): The original hybrid EP implementation was not adapted for dual-port devices like CX8, so there will be performance issues. I believe your configuration is fine — this is purely a design issue with HEP. For this reason we introduced the nixl backend. We will fix these issues as soon as possible, and we sincerely apologize for any inconvenience caused
Sign in to join this conversation.
No Branch/Tag specified
Branches Tags
main
hybrid-ep
chore/remove-clang-format-yapf
v2.5
nvDev
try-fix-cu128
mori-ep
antgroup-opt
compat-old-torch
add-linter
eager-rdma
v1.2.1
Labels
Clear labels   
pull-request
Mirrored from GitHub Pull Request

No labels
pull-request
Milestone
Clear milestone
No items
No milestone
Projects
Clear projects
No items
No project
Assignees
Clear assignees
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
deepseek-ai/DeepEP#361
No description provided.