Skip to content

gloo/cuda: use torch dtype bf16 #441

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 1 commit into from
May 15, 2025
Merged

gloo/cuda: use torch dtype bf16 #441

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
1 change: 1 addition & 0 deletions CMakeLists.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -42,6 +42,7 @@ if(${USE_TCP_OPENSSL_LINK} AND ${USE_TCP_OPENSSL_LOAD})
endif()
option(USE_CUDA "Build with CUDA support" OFF)
option(GLOO_USE_CUDA_TOOLKIT "Build CUDA with FindCUDATookit.cmake and enable_language(CUDA)" OFF)
option(GLOO_USE_TORCH_DTYPES "Build CUDA kernels with pytorch dtypes" 0)

if(MSVC)
message(STATUS "MSVC detected")
Expand Down
10 changes: 10 additions & 0 deletions gloo/CMakeLists.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -171,9 +171,19 @@ target_link_libraries(gloo PRIVATE ${gloo_DEPENDENCY_LIBS})
target_include_directories(gloo INTERFACE $<INSTALL_INTERFACE:include>)
if(USE_CUDA)
target_include_directories(gloo_cuda INTERFACE $<INSTALL_INTERFACE:include>)

message(STATUS "GLOO_USE_TORCH_DTYPES : ${GLOO_USE_TORCH_DTYPES} ${GLOO_TORCH_DIR}")
if(GLOO_USE_TORCH_DTYPES)
target_include_directories(gloo_cuda PRIVATE ${GLOO_TORCH_DIR})
endif()
endif()
if(USE_ROCM)
target_include_directories(gloo_hip INTERFACE $<INSTALL_INTERFACE:include>)

message(STATUS "GLOO_USE_TORCH_DTYPES : ${GLOO_USE_TORCH_DTYPES} ${GLOO_TORCH_DIR}")
if(GLOO_USE_TORCH_DTYPES)
target_include_directories(gloo_hip PRIVATE ${GLOO_TORCH_DIR})
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

interesting... Gloo also supports RCOM?

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yup -- it's a bit of a mess but yes. It uses HIP the same way pytorch does

endif()
endif()

# Install if necessary.
Expand Down
2 changes: 2 additions & 0 deletions gloo/config.h.in
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -37,3 +37,5 @@ static_assert(
#cmakedefine01 GLOO_HAVE_TRANSPORT_TCP_TLS
#cmakedefine01 GLOO_HAVE_TRANSPORT_IBVERBS
#cmakedefine01 GLOO_HAVE_TRANSPORT_UV

#cmakedefine01 GLOO_USE_TORCH_DTYPES
9 changes: 9 additions & 0 deletions gloo/cuda.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -391,4 +391,13 @@ DELEGATE_SIMPLE_CUDA_BINARY_COMPARE(double, cudaMax, >);
DELEGATE_HALF_PRECISION_CUDA_BINARY_COMPARE(cudaMin, <);
DELEGATE_HALF_PRECISION_CUDA_BINARY_COMPARE(cudaMax, >);

#if GLOO_USE_TORCH_DTYPES
using BFloat16 = c10::BFloat16;
INSTANTIATE_COPY_ASYNC(BFloat16);
DELEGATE_SIMPLE_CUDA_BINARY_OPERATOR(BFloat16, cudaSum, +);
DELEGATE_SIMPLE_CUDA_BINARY_OPERATOR(BFloat16, cudaProduct, *);
DELEGATE_SIMPLE_CUDA_BINARY_COMPARE(BFloat16, cudaMin, <);
DELEGATE_SIMPLE_CUDA_BINARY_COMPARE(BFloat16, cudaMax, >);
#endif

} // namespace gloo
4 changes: 4 additions & 0 deletions gloo/cuda_allreduce_bcube.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -514,4 +514,8 @@ INSTANTIATE_TEMPLATE(float);
INSTANTIATE_TEMPLATE(double);
INSTANTIATE_TEMPLATE(float16);

#if GLOO_USE_TORCH_DTYPES
INSTANTIATE_TEMPLATE(c10::BFloat16);
#endif

} // namespace gloo
4 changes: 4 additions & 0 deletions gloo/cuda_allreduce_halving_doubling.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -657,4 +657,8 @@ INSTANTIATE_TEMPLATE(float);
INSTANTIATE_TEMPLATE(double);
INSTANTIATE_TEMPLATE(float16);

#if GLOO_USE_TORCH_DTYPES
INSTANTIATE_TEMPLATE(c10::BFloat16);
#endif

} // namespace gloo
4 changes: 4 additions & 0 deletions gloo/cuda_allreduce_local.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -76,4 +76,8 @@ INSTANTIATE_TEMPLATE(float);
INSTANTIATE_TEMPLATE(double);
INSTANTIATE_TEMPLATE(float16);

#if GLOO_USE_TORCH_DTYPES
INSTANTIATE_TEMPLATE(c10::BFloat16);
#endif

} // namespace gloo
4 changes: 4 additions & 0 deletions gloo/cuda_allreduce_ring.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -188,4 +188,8 @@ INSTANTIATE_TEMPLATE(float);
INSTANTIATE_TEMPLATE(double);
INSTANTIATE_TEMPLATE(float16);

#if GLOO_USE_TORCH_DTYPES
INSTANTIATE_TEMPLATE(c10::BFloat16);
#endif

} // namespace gloo
4 changes: 4 additions & 0 deletions gloo/cuda_allreduce_ring_chunked.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -365,4 +365,8 @@ INSTANTIATE_TEMPLATE(float);
INSTANTIATE_TEMPLATE(double);
INSTANTIATE_TEMPLATE(float16);

#if GLOO_USE_TORCH_DTYPES
INSTANTIATE_TEMPLATE(c10::BFloat16);
#endif

} // namespace gloo
4 changes: 4 additions & 0 deletions gloo/cuda_broadcast_one_to_all.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -197,4 +197,8 @@ INSTANTIATE_TEMPLATE(float);
INSTANTIATE_TEMPLATE(double);
INSTANTIATE_TEMPLATE(float16);

#if GLOO_USE_TORCH_DTYPES
INSTANTIATE_TEMPLATE(c10::BFloat16);
#endif

} // namespace gloo
4 changes: 4 additions & 0 deletions gloo/cuda_private.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -20,6 +20,10 @@
#include "gloo/cuda.h"
#include "gloo/transport/device.h"

#if GLOO_USE_TORCH_DTYPES
#include <c10/util/BFloat16.h>
#endif

namespace gloo {

#define CUDA_CHECK(condition) \
Expand Down
Loading