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

[Distributed]Add check for c_embedding (#55621) #55623

Closed
wants to merge 1 commit into from
Closed

[Distributed]Add check for c_embedding (#55621) #55623

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
4 changes: 4 additions & 0 deletions paddle/fluid/operators/collective/c_embedding_op.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -87,6 +87,10 @@ class CEmbeddingOpMaker : public framework::OpProtoAndCheckerMaker {
"(int64, default 0), The starting index is indeed, "
"and the out-of-bounds will be set to 0 ")
.SetDefault(0);
AddAttr<int64_t>("vocab_size",
"(int64, default -1), The total vocabulary size to check"
"the out-of-bounds ids. If it is -1, no check will be ")
.SetDefault(-1);
AddComment(R"DOC(
c_embedding Operator.

Expand Down
28 changes: 18 additions & 10 deletions paddle/fluid/operators/collective/c_embedding_op.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -42,21 +42,25 @@ __global__ void CEmbedding(T *out,
const int64_t N,
const int64_t start_idx,
const int64_t end_idx,
const int64_t limit) {
const int64_t limit,
const int64_t vocab_size) {
CUDA_KERNEL_LOOP(i, limit) {
size_t row = i / columns;
size_t col = i % columns;
auto id = ids[row];

PADDLE_ENFORCE(
id >= 0 && (vocab_size < 0 || id < vocab_size),
"The index is out of bounds, "
"please check whether the dimensions of index and "
"input meet the requirements. It should "
"be less than [%d] and greater than or equal to 0, but received [%d]",
vocab_size,
id);

if (id >= start_idx && id < end_idx) {
auto real_idx = id - start_idx;
PADDLE_ENFORCE(real_idx < N,
"The index is out of bounds, "
"please check whether the dimensions of index and "
"input meet the requirements. It should "
"be less than [%d], but received [%d]",
N,
real_idx);

out[i] = table[real_idx * columns + col];
} else {
out[i] = static_cast<T>(0);
Expand Down Expand Up @@ -95,6 +99,8 @@ class CEmbeddingCUDAKernel : public framework::OpKernel<T> {

const auto &dev_ctx = context.template device_context<phi::GPUContext>();
const int64_t start_idx = context.Attr<int64_t>("start_index");
const int64_t vocab_size = context.Attr<int64_t>("vocab_size");

size_t N = table_t->dims()[0];
size_t D = table_t->dims()[1];
size_t K = ids_t->numel();
Expand All @@ -119,7 +125,8 @@ class CEmbeddingCUDAKernel : public framework::OpKernel<T> {
N,
start_idx,
end_idx,
limit);
limit,
vocab_size);

} else if (index_type == framework::proto::VarType::INT64) {
CEmbedding<T, int64_t>
Expand All @@ -131,7 +138,8 @@ class CEmbeddingCUDAKernel : public framework::OpKernel<T> {
N,
start_idx,
end_idx,
limit);
limit,
vocab_size);
} else {
PADDLE_THROW(platform::errors::Unavailable(
"GPU c_embedding ids only support int32 or int64."));
Expand Down
2 changes: 2 additions & 0 deletions python/paddle/distributed/fleet/layers/mpu/mp_layers.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -124,6 +124,7 @@ def __init__(
self._size = [per_part_size, embedding_dim]
self._weight_attr = weight_attr
self._name = name
self.num_embeddings = num_embeddings

if self.is_mp and paddle.in_dynamic_mode():
with get_rng_state_tracker().rng_state():
Expand Down Expand Up @@ -151,6 +152,7 @@ def forward(self, x):
self.weight,
x,
start_index=self.vocab_start_index,
vocab_size=self.num_embeddings,
name=self._name,
)
output = mp_ops._mp_allreduce(
Expand Down
12 changes: 8 additions & 4 deletions python/paddle/distributed/fleet/layers/mpu/mp_ops.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -294,7 +294,7 @@ def backward(ctx, dy):
return out


def _c_lookup_table(table, index, start_index=0, name=None):
def _c_lookup_table(table, index, start_index=0, vocab_size=-1, name=None):
"""
Lookup table according to index.

Expand All @@ -310,7 +310,7 @@ def _c_lookup_table(table, index, start_index=0, name=None):
"""
if in_dynamic_mode():
return _legacy_C_ops.c_embedding(
table, index, "start_index", start_index
table, index, "start_index", start_index, "vocab_size", vocab_size
)
else:
op_type = 'c_embedding'
Expand All @@ -322,7 +322,7 @@ def _c_lookup_table(table, index, start_index=0, name=None):
type='c_embedding',
inputs={'Ids': index, 'W': table},
outputs={'Out': tmp},
attrs={"start_index": start_index},
attrs={"start_index": start_index, "vocab_size": vocab_size},
)
return tmp

Expand Down Expand Up @@ -650,7 +650,11 @@ def _parallel_embedding(
main_block.vars[weight.name].is_distributed = True

output_parallel = _c_lookup_table(
weight, x, start_index=vocab_start_index, name=name
weight,
x,
start_index=vocab_start_index,
vocab_size=origin_size[0],
name=name,
)
out = _mp_allreduce(
output_parallel,
Expand Down