Skip to content
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

[BugFix]Fix bug in pfp16 in DataParallel #38378

Merged
merged 3 commits into from
Dec 27, 2021
Merged

[BugFix]Fix bug in pfp16 in DataParallel #38378

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
bb923e5
fix bug in pfp16
ForFishes Dec 23, 2021
26a0bb3
fix hip
ForFishes Dec 23, 2021
581b1a2
fix hip
ForFishes Dec 23, 2021
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
30 changes: 30 additions & 0 deletions paddle/fluid/framework/data_type.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -89,6 +89,22 @@ struct DataTypeTrait<void> {
_ForEachDataTypeHelper_(callback, int, INT32); \
_ForEachDataTypeHelper_(callback, int64_t, INT64);

// It's only for DataParallel in HIP, bf16 not support in HIP.
#define _ForEachDataTypeForHIP_(callback) \
_ForEachDataTypeHelper_(callback, float, FP32); \
_ForEachDataTypeHelper_(callback, ::paddle::platform::float16, FP16); \
_ForEachDataTypeHelper_(callback, double, FP64); \
_ForEachDataTypeHelper_(callback, int, INT32); \
_ForEachDataTypeHelper_(callback, int64_t, INT64); \
_ForEachDataTypeHelper_(callback, bool, BOOL); \
_ForEachDataTypeHelper_(callback, uint8_t, UINT8); \
_ForEachDataTypeHelper_(callback, int16_t, INT16); \
_ForEachDataTypeHelper_(callback, int8_t, INT8); \
_ForEachDataTypeHelper_(callback, ::paddle::platform::complex<float>, \
COMPLEX64); \
_ForEachDataTypeHelper_(callback, ::paddle::platform::complex<double>, \
COMPLEX128);

#define DefineDataTypeTrait(cpp_type, proto_type) \
template <> \
struct DataTypeTrait<cpp_type> { \
Expand Down Expand Up @@ -147,6 +163,20 @@ inline void VisitDataTypeTiny(proto::VarType::Type type, Visitor visitor) {
#undef VisitDataTypeCallbackTiny
}

template <typename Visitor>
inline void VisitDataTypeForHIP(proto::VarType::Type type, Visitor visitor) {
#define VisitDataTypeCallbackHIP(cpp_type, proto_type) \
do { \
if (type == proto_type) { \
visitor.template apply<cpp_type>(); \
return; \
} \
} while (0)

_ForEachDataTypeForHIP_(VisitDataTypeCallbackHIP);
#undef VisitDataTypeCallbackHIP
}

extern std::string DataTypeToString(const proto::VarType::Type type);
extern size_t SizeOfType(proto::VarType::Type type);
inline std::ostream& operator<<(std::ostream& out,
Expand Down
12 changes: 11 additions & 1 deletion paddle/fluid/imperative/reducer.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -48,9 +48,19 @@ void Group::DivNRanks(const platform::DeviceContext &context, int64_t nranks) {
} else if (platform::is_cpu_place(tensor->place())) {
VLOG(4) << "before div 2" << *tensor;
VLOG(4) << "NDiv for cpu devices : rank = " << nranks;
framework::VisitDataTypeSmall(
#ifdef PADDLE_WITH_HIP
if (dtype_ == paddle::framework::proto::VarType_Type_BF16) {
PADDLE_THROW(paddle::platform::errors::Fatal(
"Unsupport BF16 in DataParallel for now"));
}
framework::VisitDataTypeForHIP(
dtype_, DivNRanksForAllReduce<platform::CPUDeviceContext>(
tensor, nranks, context));
#else
framework::VisitDataType(dtype_,
DivNRanksForAllReduce<platform::CPUDeviceContext>(
tensor, nranks, context));
#endif
VLOG(4) << "after div 2" << *tensor;
} else if (platform::is_xpu_place(tensor->place())) {
#ifdef PADDLE_WITH_XPU_BKCL
Expand Down
12 changes: 11 additions & 1 deletion paddle/fluid/imperative/reducer.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -20,9 +20,19 @@ namespace imperative {
#if defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_RCCL)
void Group::DivNRanks(framework::Tensor *tensor, int64_t nranks,
const platform::DeviceContext &context) {
framework::VisitDataTypeSmall(
#ifdef PADDLE_WITH_HIP
if (dtype_ == paddle::framework::proto::VarType_Type_BF16) {
PADDLE_THROW(paddle::platform::errors::Fatal(
"Unsupport BF16 in DataParallel for now"));
}
framework::VisitDataTypeForHIP(
dtype_, DivNRanksForAllReduce<platform::CUDADeviceContext>(tensor, nranks,
context));
#else
framework::VisitDataType(
dtype_, DivNRanksForAllReduce<platform::CUDADeviceContext>(tensor, nranks,
context));
#endif
}
#endif

Expand Down