[nativert] triton runtime implementation (#161798)

Summary:
att
Test Plan:
ci
Rollback Plan:

Reviewed By: minjang

Differential Revision: D80828148

Pull Request resolved: https://github.com/pytorch/pytorch/pull/161798
Approved by: https://github.com/minjang, https://github.com/SherlockNoMad
This commit is contained in:
dolpm
2025-09-04 19:00:11 +00:00
committed by PyTorch MergeBot
parent 1f51056bd6
commit 3dde5d7f9b
12 changed files with 578 additions and 1 deletions

View File

@ -635,6 +635,12 @@ libtorch_nativert_sources = [
"torch/nativert/graph/passes/pass_manager/GraphPasses.cpp",
"torch/nativert/graph/passes/pass_manager/PassManager.cpp",
"torch/nativert/kernels/KernelHandlerRegistry.cpp",
"torch/nativert/kernels/TritonKernel.cpp",
"torch/nativert/executor/triton/CpuTritonKernelManager.cpp",
]
libtorch_nativert_cuda_sources = [
"torch/nativert/executor/triton/CudaTritonKernelManager.cpp",
]
torch_mobile_tracer_sources = [
@ -770,7 +776,7 @@ libtorch_cuda_distributed_sources = libtorch_cuda_distributed_base_sources + lib
libtorch_cuda_sources = libtorch_cuda_core_sources + libtorch_cuda_distributed_sources + [
"torch/csrc/cuda/nccl.cpp",
]
] + libtorch_nativert_cuda_sources
torch_cpp_srcs = [
"torch/csrc/api/src/cuda.cpp", # this just forwards stuff, no real CUDA