diff options
Diffstat (limited to 'gnu/packages/patches/python-pytorch-without-kineto.patch')
| -rw-r--r-- | gnu/packages/patches/python-pytorch-without-kineto.patch | 60 |
1 files changed, 60 insertions, 0 deletions
diff --git a/gnu/packages/patches/python-pytorch-without-kineto.patch b/gnu/packages/patches/python-pytorch-without-kineto.patch new file mode 100644 index 0000000000..f956316866 --- /dev/null +++ b/gnu/packages/patches/python-pytorch-without-kineto.patch @@ -0,0 +1,60 @@ +Even when building without Kineto, the <ActivityType.h> header is still +imported and the ActivityType type is used. This patch was copied from +https://github.com/pytorch/pytorch/pull/111048. + +diff --git a/torch/csrc/profiler/kineto_shim.h b/torch/csrc/profiler/kineto_shim.h +index e92cbf00..68985ab7 100644 +--- a/torch/csrc/profiler/kineto_shim.h ++++ b/torch/csrc/profiler/kineto_shim.h +@@ -12,7 +12,51 @@ + #undef USE_KINETO + #endif + ++#ifdef USE_KINETO + #include <ActivityType.h> ++#else ++namespace libkineto { ++// copied from header ++/* ++ * Copyright (c) Meta Platforms, Inc. and affiliates. ++ * All rights reserved. ++ * ++ * This source code is licensed under the BSD-style license found in the ++ * LICENSE file in the root directory of this source tree. ++ */ ++ ++// Note : All activity types are not enabled by default. Please add them ++// at correct position in the enum ++enum class ActivityType { ++ // Activity types enabled by default ++ CPU_OP = 0, // cpu side ops ++ USER_ANNOTATION, ++ GPU_USER_ANNOTATION, ++ GPU_MEMCPY, ++ GPU_MEMSET, ++ CONCURRENT_KERNEL, // on-device kernels ++ EXTERNAL_CORRELATION, ++ CUDA_RUNTIME, // host side cuda runtime events ++ CUDA_DRIVER, // host side cuda driver events ++ CPU_INSTANT_EVENT, // host side point-like events ++ PYTHON_FUNCTION, ++ OVERHEAD, // CUPTI induced overhead events sampled from its overhead API. ++ ++ // Optional Activity types ++ CUDA_SYNC, // synchronization events between runtime and kernels ++ GLOW_RUNTIME, // host side glow runtime events ++ MTIA_RUNTIME, // host side MTIA runtime events ++ CUDA_PROFILER_RANGE, // CUPTI Profiler range for performance metrics ++ MTIA_CCP_EVENTS, // MTIA ondevice CCP events ++ HPU_OP, // HPU host side runtime event ++ XPU_RUNTIME, // host side xpu runtime events ++ ++ ENUM_COUNT, // This is to add buffer and not used for any profiling logic. Add your new type before it. ++ OPTIONAL_ACTIVITY_TYPE_START = CUDA_SYNC, ++}; ++} ++ ++#endif + + #include <torch/csrc/Export.h> + #include <torch/csrc/profiler/api.h> |