[arathi-hlab]

Motivation

#46138 added a module-level from sglang.srt.distributed.device_communicators.custom_all_reduce_v2 import CustomAllReduceV2 to dp_attention.py. custom_all_reduce_v2 imports sglang/kernels/ops/communication/all_reduce.py, which does import tvm_ffi at module level. tvm_ffi is not installed in the XPU image (intel/sglang-dev:latest), and dp_attention.py is imported almost everywhere, so importing sglang on XPU now fails. For example, the XPU diffusion job dies before running any test:

File ".../sglang/srt/layers/dp_attention.py", line 20, in <module>
  from sglang.srt.distributed.device_communicators.custom_all_reduce_v2 import (
File ".../sglang/kernels/ops/communication/all_reduce.py", line 7, in <module>
  import tvm_ffi
ModuleNotFoundError: No module named 'tvm_ffi'

https://github.com/sgl-project/sglang/actions/runs/40179531469/job/121127481365

Modifications

On XPU only, skip the import and bind CustomAllReduceV2 = (). The class's only use is isinstance(comm, CustomAllReduceV2) in _decode_sum_len_fits_custom_all_reduce, and isinstance(x, ()) is always False. That is correct on XPU, which cannot create a v2 custom all-reduce.

• CUDA / ROCm / other platforms: unchanged, the module-level import is kept as is.
• XPU: sglang imports again; the SUM_LEN shortcut is never taken, as before [Fix] dp-attention: Align prefill graph buckets and decode padding across DP ranks #46138.

Checklist

• pre-commit passes
• XPU CI green

---

CI States

Latest PR Test (Base): ✅ Run #40180912773
Latest PR Test (Extra): ❌ Run #40180912700
Latest PR Test (AMD ROCm 10): ❌ Run #40180912717