Skip to content

fix: Implement a patch for gelu schema change in older NGC containers #845

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
Apr 9, 2022
Merged

fix: Implement a patch for gelu schema change in older NGC containers #845

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
34 changes: 33 additions & 1 deletion core/lowering/passes/reduce_gelu.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -8,10 +8,17 @@ namespace passes {

void ReduceGelu(std::shared_ptr<torch::jit::Graph>& graph) {
std::string gelu_pattern = R"IR(
graph(%x):
graph(%x : Tensor):
%out : Tensor = aten::gelu(%x)
return (%out))IR";

// This gelu_approximate_pattern schema exists in 21.11, 21.12, 22.01 containers of pytorch. These container versions use
// an unmerged PR in pytorch : https://github.com/pytorch/pytorch/pull/61439. We reduce this to regular Gelu.
std::string gelu_approximate_pattern = R"IR(
graph(%x : Tensor, %approx):
%out : Tensor = aten::gelu(%x, %approx)
return (%out))IR";

std::string gelu_reduce_pattern = R"IR(
graph(%x.1 : Tensor):
%6 : float = prim::Constant[value=0.044714999999999998]()
Expand All @@ -30,11 +37,36 @@ void ReduceGelu(std::shared_ptr<torch::jit::Graph>& graph) {
%15 : Tensor = aten::mul(%7, %14)
return (%15))IR";

// This is same as gelu_reduce_pattern except for an additional input %approx.
// SubgraphRewriter only works as expected if the number of inputs to gelu_approximate_pattern
// and gelu_reduce_multi_input_pattern are same.
std::string gelu_reduce_multi_input_pattern = R"IR(
graph(%x.1 : Tensor, %approx):
%6 : float = prim::Constant[value=0.044714999999999998]()
%5 : float = prim::Constant[value=0.79788456080000003]()
%4 : float = prim::Constant[value=1.]()
%3 : float = prim::Constant[value=0.5]()
%2 : int = prim::Constant[value=1]()
%7 : Tensor = aten::mul(%x.1, %3)
%8 : Tensor = aten::mul(%x.1, %5)
%9 : Tensor = aten::mul(%x.1, %6)
%10 : Tensor = aten::mul(%9, %x.1)
%11 : Tensor = aten::add(%10, %4, %2)
%12 : Tensor = aten::mul(%8, %11)
%13 : Tensor = aten::tanh(%12)
%14 : Tensor = aten::add(%13, %4, %2)
%15 : Tensor = aten::mul(%7, %14)
return (%15))IR";

// replace aten::gelu with pointwise operations
torch::jit::SubgraphRewriter map_gelu_to_pointwise_ops;
map_gelu_to_pointwise_ops.RegisterRewritePattern(gelu_pattern, gelu_reduce_pattern);
map_gelu_to_pointwise_ops.runOnGraph(graph);

torch::jit::SubgraphRewriter map_gelu_approximate_to_pointwise_ops;
map_gelu_approximate_to_pointwise_ops.RegisterRewritePattern(gelu_approximate_pattern, gelu_reduce_multi_input_pattern);
map_gelu_approximate_to_pointwise_ops.runOnGraph(graph);

LOG_GRAPH("Post lowering of [aten::gelu] -> " << *graph);
}

Expand Down
35 changes: 35 additions & 0 deletions tests/core/lowering/test_reduce_gelu.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -40,3 +40,38 @@ TEST(LoweringPasses, ReduceGeluCorrectly) {

ASSERT_TRUE(!torch::jit::findPatternMatches(*tg, *sg).empty());
}

TEST(LoweringPasses, ReduceGeluApproximateCorrectly) {
std::string source_graph = R"IR(
graph(%x, %approx):
%out : Tensor = aten::gelu(%x, %approx)
return (%out))IR";
std::string target_graph = R"IR(
graph(%x.1 : Tensor, %approx):
%6 : float = prim::Constant[value=0.044714999999999998]()
%5 : float = prim::Constant[value=0.79788456080000003]()
%4 : float = prim::Constant[value=1.]()
%3 : float = prim::Constant[value=0.5]()
%2 : int = prim::Constant[value=1]()
%7 : Tensor = aten::mul(%x.1, %3)
%8 : Tensor = aten::mul(%x.1, %5)
%9 : Tensor = aten::mul(%x.1, %6)
%10 : Tensor = aten::mul(%9, %x.1)
%11 : Tensor = aten::add(%10, %4, %2)
%12 : Tensor = aten::mul(%8, %11)
%13 : Tensor = aten::tanh(%12)
%14 : Tensor = aten::add(%13, %4, %2)
%15 : Tensor = aten::mul(%7, %14)
return (%15))IR";

torch_tensorrt::core::util::logging::get_logger().set_reportable_log_level(
torch_tensorrt::core::util::logging::LogLevel::kGRAPH);
auto sg = std::make_shared<torch::jit::Graph>();
torch::jit::parseIR(source_graph, &*sg);
torch_tensorrt::core::lowering::passes::ReduceGelu(sg);

auto tg = std::make_shared<torch::jit::Graph>();
torch::jit::parseIR(target_graph, &*tg);

ASSERT_TRUE(!torch::jit::findPatternMatches(*tg, *sg).empty());
}