Skip to content

Fix dgc bug. #16709

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
gongweibao merged 2 commits into from
Apr 9, 2019
Merged

Fix dgc bug. #16709

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
cf2999b
fix test=release/1.4
gongweibao Apr 8, 2019
8372bf4
follow comments test=release/1.4
gongweibao Apr 8, 2019
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
6 changes: 5 additions & 1 deletion paddle/fluid/framework/details/all_reduce_op_handle.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -53,6 +53,10 @@ AllReduceOpHandle::AllReduceOpHandle(ir::Node *node,
this->SetDeviceContext(p, nccl_ctxs_->DevCtx(p));
}
}
// TODO(gongwb) :polish them!
if (is_encoded) {
VLOG(1) << "Use dgc allreduce mode";
}
}
#else
AllReduceOpHandle::AllReduceOpHandle(ir::Node *node,
Expand Down Expand Up @@ -86,7 +90,7 @@ void AllReduceOpHandle::RunImplEncoded() {
paddle::framework::GradOriginalVarName(in_var_handles[i]->name());
auto encode_var_name = original_name + g_dgc_encoded;
auto *in_var = local_scope->FindVar(encode_var_name);
PADDLE_ENFORCE_NOT_NULL(in_var);
PADDLE_ENFORCE_NOT_NULL(in_var, "%s should not be null", encode_var_name);
auto &in = in_var->Get<LoDTensor>();
ins.emplace_back(&in);

Expand Down
2 changes: 1 addition & 1 deletion python/paddle/fluid/optimizer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -752,7 +752,7 @@ def _append_dgc_ops(self, param_and_grads):
force_cpu=True)

for param_var, grad_var in param_and_grads:
var_numel = reduce(lambda x, y: x * y, param_var.shape)
var_numel = abs(reduce(lambda x, y: x * y, param_var.shape))
if var_numel < 16384 or \
param_var.type == core.VarDesc.VarType.SELECTED_ROWS or \
grad_var.type == core.VarDesc.VarType.SELECTED_ROWS or \
Expand Down
10 changes: 8 additions & 2 deletions python/paddle/fluid/parallel_executor.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -104,10 +104,11 @@ def __init__(self,
self._scope = scope if scope is not None else executor.global_scope()

if main_program is not None and main_program._enable_dgc:
assert num_trainers > 1, "dgc is not useful for single trainer training."
assert build_strategy.reduce_strategy == BuildStrategy.ReduceStrategy.AllReduce
assert num_trainers * len(
self._places) > 1, "dgc is not useful for single card training"
assert use_cuda
self._places) > 1, "dgc is not useful for single card training."
assert use_cuda, "dgc only used when cuda is used."

main_program = main_program if main_program is not None \
else framework.default_main_program()
Expand All @@ -123,6 +124,11 @@ def __init__(self,
exec_strategy=exec_strategy,
share_vars_from=share_vars_from._compiled_program
if share_vars_from else None)

# FIXME(gongwb): I will move dgc from dist mode to allreduce mode in next pr.
if main_program._enable_dgc:
self._compiled_program._build_strategy.is_distribution = True

self._place = core.CUDAPlace(0) if use_cuda else core.CPUPlace()
self._exe = executor.Executor(self._place)
self._compiled_program._compile(place=self._place, scope=self._scope)
Expand Down