[EP] Add SM80/A100 build support with SM90 feature guards

[bkpathak]

Description

Please include a summary of the changes and the related issue.

Fixes # (issue)

Type of Change

• Bug fix
• New feature
• Documentation update

How Has This Been Tested?

Include any tests here.

• Unit tests
• Integration tests
• Manual testing

Checklist

• [x ] I have run format.sh to follow the style guidelines.
• I have run build.sh to verify compilation.
• I have removed redundant variables and comments.
• I have updated the documentation.
• I have added tests.

##Test-bed

+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 580.159.03             Driver Version: 580.159.03     CUDA Version: 13.0     |
+-----------------------------------------+------------------------+----------------------+
| GPU  Name                 Persistence-M | Bus-Id          Disp.A | Volatile Uncorr. ECC |
| Fan  Temp   Perf          Pwr:Usage/Cap |           Memory-Usage | GPU-Util  Compute M. |
|                                         |                        |               MIG M. |
|=========================================+========================+======================|
|   0  NVIDIA A100-SXM4-40GB          Off |   00000000:06:00.0 Off |                    0 |
| N/A   29C    P0             47W /  400W |       0MiB /  40960MiB |      0%      Default |
|                                         |                        |             Disabled |
+-----------------------------------------+------------------------+----------------------+

+-----------------------------------------------------------------------------------------+
| Processes:                                                                              |
|  GPU   GI   CI              PID   Type   Process name                        GPU Memory |
|        ID   ID                                                               Usage      |
|=========================================================================================|
|  No running processes found                                                             |
+-----------------------------------------------------------------------------------------+
NVIDIA A100-SXM4-40GB
(8, 0)
nvcc not found
Linux 129-146-53-27 5.15.0-143-generic #153-Ubuntu SMP Fri Jun 13 19:10:45 UTC 2025 x86_64 x86_64 x86_64 GNU/Linux
NAME="Ubuntu"
VERSION="22.04.5 LTS (Jammy Jellyfish)"
Python 3.12.13
PyTorch: 2.5.1+cu121

Build Command

TORCH_CUDA_ARCH_LIST="8.0" \
DISABLE_SM90_FEATURES=1 \
DISABLE_AGGRESSIVE_PTX_INSTRS=1 \
PER_EXPERT_BATCHING=1 \
bash build.sh cu12 ep 3.12

Install Wheel

python3.12 -m pip install wheelhouse-cu12/uccl-0.1.1-cp312-abi3-manylinux_2_35_x86_64.whl --no-deps

Test

 python3.12 -c "
import torch
import uccl.ep

print('=== Hardware ===')
print('Device:', torch.cuda.get_device_name(0))
print('Capability:', torch.cuda.get_device_capability(0))
print('CUDA available:', torch.cuda.is_available())

print()
print('=== UCCL EP ===')
print('SM90 compiled:', uccl.ep.is_sm90_compiled())
print('Buffer:', uccl.ep.Buffer)
print('Config:', uccl.ep.Config)
print('APIs:', [x for x in dir(uccl.ep) if not x.startswith('_')])

print()
print('success')
"
=== Hardware ===
Device: NVIDIA A100-SXM4-40GB
Capability: (8, 0)
CUDA available: True

=== UCCL EP ===
SM90 compiled: False
Buffer: <class 'uccl.ep.Buffer'>
Config: <class 'uccl.ep.Config'>
APIs: ['Bench', 'BenchFifo', 'Buffer', 'Config', 'EnvInfo', 'EventHandle', 'EventOverlap', 'Fifo', 'FifoProxy', 'Proxy', 'Stats', 'alloc_cmd_ring', 'can_register_rdma_gpu_buffer', 'check_stream', 'connect_atomic_buffer', 'device_reset', 'free_cmd_ring', 'get_device', 'get_low_latency_rdma_size_hint', 'get_num_proxy_threads', 'get_oob_ip', 'get_rdma_buffer', 'has_proxy', 'is_sm90_compiled', 'launch_gpu_issue_kernel', 'rdma_buffer_should_use_host_alloc', 'register_proxies', 'register_proxy', 'set_device', 'stop_all_registered_proxies', 'stream_query', 'sync_stream', 'unregister_proxy']

success

[YangZhou1997]

@manojgop does this fix the issues you mentioned in https://uccl-dev.slack.com/archives/C0B2FQHQNTT/p1779093594675089?thread_ts=1778975124.159549&cid=C0B2FQHQNTT? I do not have A100 machines in hand, thus cannot verify.

But I have tested on NV H100 + EFA and AMD MI325x + Broadcom, both working.

@MaoZiming @zhenhuang12, can any of you also review it?

[manojgop]

@manojgop does this fix the issues you mentioned in https://uccl-dev.slack.com/archives/C0B2FQHQNTT/p1779093594675089?thread_ts=1778975124.159549&cid=C0B2FQHQNTT? I do not have A100 machines in hand, thus cannot verify.

But I have tested on NV H100 + EFA and AMD MI325x + Broadcom, both working.

@MaoZiming @zhenhuang12, can any of you also review it?

@bkpathak I've fixed the build for cu13 and added additional changes to verify A100 with low latency mode for inter node and intra node in a two node setup with two A100 GPU per node. You may take this commit and update the PR #963 --> manojgop@f9dab2e. Note: I've tested this with Intel RDMA NIC

[bkpathak]

Sure, will send the updated PR.

[manojgop]

@YangZhou1997 @MaoZiming - Could you please review this PR?. @bkpathak This PR needs to be rebased with main branch

[MaoZiming]

The __threadfence(); here is not a acquire fence? Is it needed

[MaoZiming]

@bkpathak Thank you for the contribution! would you also like to share some performance results on A100?

[bkpathak]

Sure, will add the test results.