File tree Expand file tree Collapse file tree
Expand file tree Collapse file tree Original file line number Diff line number Diff line change @@ -201,10 +201,6 @@ def get_device_config(adaptor_flag):
201201 library_dirs += ["/opt/tops/lib" , pytorch_library_path ]
202202 libs += ["topsrt" , "torch_gcu" ]
203203 elif adaptor_flag == "-DUSE_SUNRISE_ADAPTOR" :
204- # All torchpt:: symbols (PTPUStream/PTPUEvent/PTPUDeviceGuardImpl)
205- # live only in torch_ptpu/_C.cpython-*.so. The torch_ptpu/lib/ folder
206- # holds kernel libs with zero torchpt:: symbols, and TorchPTPUConfig.cmake
207- # itself targets _C*.so, so we must link against _C*.so by exact name.
208204 import torch_ptpu
209205 torch_ptpu_dir = os .path .dirname (os .path .abspath (torch_ptpu .__file__ ))
210206 c_so_basename = os .path .basename (torch_ptpu ._C .__file__ )
Original file line number Diff line number Diff line change @@ -488,15 +488,10 @@ c10::intrusive_ptr<Work> flagcxBackend::endCoalescing() {
488488 " endCoalescing called without a matching startCoalescing on "
489489 " the PTPU backend" );
490490
491- // Sort by peer asc so every rank touches pair sub-comms in canonical
492- // (min,max) order. getOrInitPtpuPairComm blocks on a peer handshake;
493- // canonical ordering turns that into a total order on pair keys and
494- // eliminates the ring-of-pairs deadlock for >=3 ranks.
495-
496- // No flagcxGroupStart/End: PCCL group brackets are process-global and
497- // pair-comm send/recv are already async per-stream, so serial issue
498- // still satisfies batch_isend_irecv's "enqueue-then-wait" semantics
499- // and avoids the comm mismatch that caused the SIGSEGV.
491+ // Sort by peer asc to issue pair sub-comms in canonical (min,max) order,
492+ // avoiding the ring-of-pairs deadlock from getOrInitPtpuPairComm handshake.
493+ // No flagcxGroupStart/End: pair-comm send/recv are already async per-stream,
494+ // so serial issue still meets batch_isend_irecv's enqueue-then-wait semantics.
500495 std::stable_sort (
501496 ptpuCoalesce_.pendingOps .begin (), ptpuCoalesce_.pendingOps .end (),
502497 [](const auto &a, const auto &b) { return a.first < b.first ; });
You can’t perform that action at this time.
0 commit comments