Conversation
| prop->maxThreadsDim[2] = max_work_item_sizes[2]; | ||
|
|
||
| // Max grid size - not directly available in SYCL, use reasonable defaults | ||
| prop->maxGridSize[0] = 2147483647; |
There was a problem hiding this comment.
then, should we submit request to sycl?
There was a problem hiding this comment.
In the current NCCL implementation, getDeviceProperties isn’t actually used anywhere; it merely fills in some values to maintain consistency with NCCL’s format. In fact, within PyTorch, the DeviceProperties functions for XPU and CUDA are entirely different. Once this function becomes actively used in the future, I’ll coordinate with guangye to discuss how to modify it.
| int device = ::c10::xpu::current_device(); | ||
| sycl::device& dev = ::c10::xpu::get_raw_device(device); | ||
|
|
||
| *devPtr = sycl::malloc_device(size, dev, ctx); |
There was a problem hiding this comment.
So it does its own memory management? Is there any cache strategy?
There was a problem hiding this comment.
Just a wrapper of sycl malloc api, used by barrier(create and cache tiny buffer allreduce)
|
|
||
| xpu_result_t DefaultXpuApi::eventCreate(xpuEvent_t& event) { | ||
| try { | ||
| event = ::at::xpu::XPUEvent(false); // No timing |
There was a problem hiding this comment.
We can expect timing enable in basekit 2025.2, right?
There was a problem hiding this comment.
which in next fun eventCreateWithFlags
| class TorchCommWindowXCCL; | ||
|
|
||
| // Forward declaration for test class | ||
| namespace test { |
There was a problem hiding this comment.
I don’t think it’s necessary, but since other backends have reserved this interface, it’s likely intended for future use.
| // | ||
| // NOTE: finalize MUST return without holding references to any work object, | ||
| // otherwise it may leak object and cause side effects. | ||
| stream_work_queues_.clear(); |
There was a problem hiding this comment.
Store the works queued on each stream to allow the background cleanup thread to traverse the queue of every stream.
| } | ||
|
|
||
| onecclResult_t DefaultXcclApi::commDeregister(onecclComm_t comm, void *handle) { | ||
| // return onecclCommDeregister(comm, handle); |
There was a problem hiding this comment.
Please check if we have submit feature request to oneCCL. If not, then a ticket is needed.
|
I don't get where is the profiler enabled. |
| checkAndAbortIfTimedOutOrError(); | ||
| ensureTensorContiguous(tensor); | ||
|
|
||
| tracing_->recordEventWithInputOutput("all_reduce", rank_, {tensor}, {tensor}); |
There was a problem hiding this comment.
@zhangxiaoli73 here is profiler register. https://github.com/meta-pytorch/torchcomms/blob/f0a411256447066c9b384727ad785ce8dc8e3ca0/comms/torchcomms/TorchCommTracing.cpp#L31 call RECORD_PARAM_COMMS_DATA, same profiler helper with processgroup
Summary: Pull Request resolved: meta-pytorch#326 add test building/running support Test Plan: ``` ctest --build-and-test ./ ./build --build-generator "Ninja" --build-options -DBUILD_TESTS=ON --test-command ctest --output-on-failure Start 1: TorchCommOptionsTest.EnvToValueImplBool 1/3 Test #1: TorchCommOptionsTest.EnvToValueImplBool ... Passed 0.31 sec Start 2: TorchCommOptionsTest.StringToBool 2/3 Test #2: TorchCommOptionsTest.StringToBool ......... Passed 0.32 sec Start 3: TorchCommFactoryTest 3/3 Test #3: TorchCommFactoryTest ...................... Passed 0.32 sec 100% tests passed, 0 tests failed out of 3 Total Test time (real) = 0.95 sec ``` imported-using-ghimport Reviewed By: d4l3k Differential Revision: D91159066 Pulled By: dolpm fbshipit-source-id: 77758d1d2d82535001e03cdfd13be299adece2aa
Summary: Pull Request resolved: meta-pytorch#326 add test building/running support Test Plan: ``` ctest --build-and-test ./ ./build --build-generator "Ninja" --build-options -DBUILD_TESTS=ON --test-command ctest --output-on-failure Start 1: TorchCommOptionsTest.EnvToValueImplBool 1/3 Test #1: TorchCommOptionsTest.EnvToValueImplBool ... Passed 0.31 sec Start 2: TorchCommOptionsTest.StringToBool 2/3 Test #2: TorchCommOptionsTest.StringToBool ......... Passed 0.32 sec Start 3: TorchCommFactoryTest 3/3 Test #3: TorchCommFactoryTest ...................... Passed 0.32 sec 100% tests passed, 0 tests failed out of 3 Total Test time (real) = 0.95 sec ``` imported-using-ghimport Reviewed By: d4l3k Differential Revision: D91159066 Pulled By: dolpm fbshipit-source-id: 77758d1d2d82535001e03cdfd13be299adece2aa
2 participants
RFC: #2
Build
export USE_XCCL=ON
export USE_NCCLX=OFF
export USE_NCCL=OFF
export USE_GLOO=OFF
export USE_SYSTEM_LIBS=1
conda install fmt -y
source oneCCL2/env/setvars.sh // direct use oneCCL2 C API
python setup.py install