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

[DRR] Add FusedLinearGeluPattern and FusedLinearReluPattern in fused_gemm_e… #58897

Merged
merged 6 commits into from
Nov 15, 2023
Merged

[DRR] Add FusedLinearGeluPattern and FusedLinearReluPattern in fused_gemm_e… #58897

Show file tree
Hide file tree
Changes from 5 commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
f72c329
add FusedLinearGeluPattern and FusedLinearReluPattern in fused_gemm_e…
gongshaotian Nov 10, 2023
ff56f25
Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…
gongshaotian Nov 10, 2023
e9fb452
update gpt_with_pir.py
gongshaotian Nov 10, 2023
18d5a9b
fix bug
gongshaotian Nov 12, 2023
76d45d7
fix FusedLinearReluGradPattern match bug
gongshaotian Nov 13, 2023
98e9047
Explicitly declare the output tensor reserve_space
gongshaotian Nov 14, 2023
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
125 changes: 107 additions & 18 deletions paddle/fluid/pir/transforms/fusion/fused_gemm_epilogue_pass.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -108,6 +108,83 @@ class FusedLinearGradPattern
}
};

class FusedLinearGeluPattern
: public pir::drr::DrrPatternBase<FusedLinearGeluPattern> {
public:
void operator()(pir::drr::DrrPatternContext *ctx) const override {
pir::drr::SourcePattern pat = ctx->SourcePattern();
// Source pattern
const auto &fused_gemm_epilogue =
pat.Op(paddle::dialect::FusedGemmEpilogueOp::name(),
{{{"trans_x", pat.Attr("trans_x")},
{"trans_y", pat.Attr("trans_y")},
{"activation", pat.Attr("act")}}});
const auto &gelu = pat.Op(paddle::dialect::GeluOp::name());
fused_gemm_epilogue(
{&pat.Tensor("x"), &pat.Tensor("w"), &pat.Tensor("bias")},
{&pat.Tensor("fuse_out"), &pat.Tensor("reserve_space")});
pat.Tensor("out") = gelu(pat.Tensor("fuse_out"));

// Constrains the activation is none
pat.RequireNativeCall([&](const pir::drr::MatchContext &match_ctx) {
return (match_ctx.Attr<std::string>("act") == "none");
});

// Result pattern
pir::drr::ResultPattern res = pat.ResultPattern();
const auto &act_attr =
res.Attr([](const pir::drr::MatchContext &match_ctx) -> std::any {
return "gelu";
});
const auto &fused_gemm_epilogue_gelu =
res.Op(paddle::dialect::FusedGemmEpilogueOp::name(),
{{{"trans_x", pat.Attr("trans_x")},
{"trans_y", pat.Attr("trans_y")},
{"activation", act_attr}}});
fused_gemm_epilogue_gelu(
{&res.Tensor("x"), &res.Tensor("w"), &res.Tensor("bias")},
{&res.Tensor("out")});
}
};
class FusedLinearReluPattern
: public pir::drr::DrrPatternBase<FusedLinearReluPattern> {
public:
void operator()(pir::drr::DrrPatternContext *ctx) const override {
pir::drr::SourcePattern pat = ctx->SourcePattern();
// Source pattern
const auto &fused_gemm_epilogue =
pat.Op(paddle::dialect::FusedGemmEpilogueOp::name(),
{{{"trans_x", pat.Attr("trans_x")},
{"trans_y", pat.Attr("trans_y")},
{"activation", pat.Attr("act")}}});
const auto &relu = pat.Op(paddle::dialect::ReluOp::name());
fused_gemm_epilogue(
{&pat.Tensor("x"), &pat.Tensor("w"), &pat.Tensor("bias")},
{&pat.Tensor("fuse_out"), &pat.Tensor("reserve_space")});
pat.Tensor("out") = relu(pat.Tensor("fuse_out"));

// Constrains the activation is none
pat.RequireNativeCall([&](const pir::drr::MatchContext &match_ctx) {
return (match_ctx.Attr<std::string>("act") == "none");
});

// Result pattern
pir::drr::ResultPattern res = pat.ResultPattern();
const auto &act_attr =
res.Attr([](const pir::drr::MatchContext &match_ctx) -> std::any {
return "relu";
});
const auto &fused_gemm_epilogue_relu =
res.Op(paddle::dialect::FusedGemmEpilogueOp::name(),
{{{"trans_x", pat.Attr("trans_x")},
{"trans_y", pat.Attr("trans_y")},
{"activation", act_attr}}});
fused_gemm_epilogue_relu(
{&res.Tensor("x"), &res.Tensor("w"), &res.Tensor("bias")},
{&res.Tensor("out")});
}
};

class FusedLinearGeluGradPattern
: public pir::drr::DrrPatternBase<FusedLinearGeluGradPattern> {
public:
Expand Down Expand Up @@ -196,20 +273,20 @@ class FusedLinearReluGradPattern
{{{"trans_x", pat.Attr("trans_x3")},
{"trans_y", pat.Attr("trans_y3")},
{"activation_grad", pat.Attr("act3")}}});

fused_gemm_epilogue(
{&pat.Tensor("x"), &pat.Tensor("w"), &pat.Tensor("bias")},
{&pat.Tensor("fuse_out"), &pat.Tensor("reserve_space")});
pat.Tensor("out") = pat.Op("pd_op.relu")(pat.Tensor("fuse_out"));

fused_gemm_epilogue_grad1({&pat.Tensor("x1"),
&pat.Tensor("w1"),
&pat.Tensor("reserve_space2"),
&pat.Tensor("out_grad")},
{&pat.Tensor("x1_grad"),
&pat.Tensor("w1_grad"),
&pat.Tensor("bias1_grad")});
pat.Tensor("relu_dx") =
pat.Op("pd_op.relu_grad")(pat.Tensor("x1"), pat.Tensor("x1_grad"));

pat.Tensor("relu_dx") = pat.Op(paddle::dialect::ReluGradOp::name())(
pat.Tensor("x1"), pat.Tensor("x1_grad"));
fused_gemm_epilogue_grad({&pat.Tensor("x"),
&pat.Tensor("w"),
&pat.Tensor("reserve_space1"),
Expand All @@ -219,39 +296,49 @@ class FusedLinearReluGradPattern
&pat.Tensor("bias_grad")});

pat.RequireNativeCall([&](const pir::drr::MatchContext &match_ctx) {
return match_ctx.Attr<std::string>("act1") == "none" &&
return match_ctx.Attr<std::string>("act1") == "relu" &&
match_ctx.Attr<std::string>("act3") == "none";
});

pir::drr::ResultPattern res = pat.ResultPattern();
const auto &act_attr =
res.Attr([](const pir::drr::MatchContext &match_ctx) -> std::any {
return "relu";
});
const auto &fused_gemm_epilogue_new =
const auto &res_fused_gemm_epilogue =
res.Op(paddle::dialect::FusedGemmEpilogueOp::name(),
{{{"trans_x", pat.Attr("trans_x1")},
{"trans_y", pat.Attr("trans_y1")},
{"activation", act_attr}}});
{"activation", pat.Attr("act1")}}});
const auto &res_fused_gemm_epilogue_grad =
res.Op(paddle::dialect::FusedGemmEpilogueGradOp::name(),
{{{"trans_x", pat.Attr("trans_x2")},
{"trans_y", pat.Attr("trans_y2")},
{"activation_grad", pat.Attr("act2")}}});
const auto &act_grad_attr =
res.Attr([](const pir::drr::MatchContext &match_ctx) -> std::any {
return "relu_grad";
});
const auto &fused_gemm_epilogue_grad1_new =
const auto &res_fused_gemm_epilogue_grad1 =
res.Op(paddle::dialect::FusedGemmEpilogueGradOp::name(),
{{{"trans_x", pat.Attr("trans_x2")},
{"trans_y", pat.Attr("trans_y2")},
{{{"trans_x", pat.Attr("trans_x3")},
{"trans_y", pat.Attr("trans_y3")},
{"activation_grad", act_grad_attr}}});
fused_gemm_epilogue_new(

res_fused_gemm_epilogue(
{&res.Tensor("x"), &res.Tensor("w"), &res.Tensor("bias")},
{&res.Tensor("out"), &res.Tensor("reserve_space3")});
fused_gemm_epilogue_grad1_new({&res.Tensor("x1"),
{&res.Tensor("fuse_out"), &res.Tensor("reserve_space")});
res_fused_gemm_epilogue_grad1({&res.Tensor("x1"),
&res.Tensor("w1"),
&res.Tensor("reserve_space3"),
&res.Tensor("reserve_space"),
&res.Tensor("out_grad")},
{&res.Tensor("relu_dx"),
&res.Tensor("w1_grad"),
&res.Tensor("bias1_grad")});

res_fused_gemm_epilogue_grad({&res.Tensor("x"),
&res.Tensor("w"),
&res.Tensor("reserve_space1"),
&res.Tensor("relu_dx")},
{&res.Tensor("x_grad"),
&res.Tensor("w_grad"),
&res.Tensor("bias_grad")});
}
};

Expand All @@ -263,6 +350,8 @@ class FusedGemmEpiloguePass : public pir::Pass {
pir::RewritePatternSet ps(context);
ps.Add(FusedLinearGradPattern().Build(context));
ps.Add(FusedLinearPattern().Build(context));
ps.Add(FusedLinearGeluPattern().Build(context));
ps.Add(FusedLinearReluPattern().Build(context));
ps.Add(FusedLinearGeluGradPattern().Build(context));
ps.Add(FusedLinearReluGradPattern().Build(context));

Expand Down
5 changes: 4 additions & 1 deletion paddle/pir/pattern_rewrite/pattern_match.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -128,7 +128,10 @@ void RewriterBase::ReplaceOp(Operation* op,
}

void RewriterBase::EraseOp(Operation* op) {
IR_ENFORCE(op->use_empty(), "expected 'op' to have no uses");
IR_ENFORCE(
op->use_empty(),
"Erase op failed. op(%s) is used, the expectation is that it is not used",
op->name());
NotifyOperationRemoved(op);
op->Erase();
}
Expand Down
17 changes: 3 additions & 14 deletions test/auto_parallel/gpt_with_pir.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -170,21 +170,10 @@ def test_dp_with_fused_linear(self):
out_dp_ir = engine_dp_ir.fit(
self.dataset, 3, batch_size=self.batch_size, log_freq=1
)
# TODO(zhiqiu): fix accuracy problem and use array_equal to check it
np.testing.assert_allclose(
out_dp_prog.history["loss"][0],
out_dp_ir.history["loss"][0],
rtol=1e-5,
err_msg='pass {} has wrong results!, \nu={}\nv={}\ndiff={}'.format(
__class__,
out_dp_prog.history["loss"][0],
out_dp_ir.history["loss"][0],
out_dp_prog.history["loss"][0] - out_dp_ir.history["loss"][0],
),

self.check_results(
out_dp_prog.history["loss"][0], out_dp_ir.history["loss"][0]
)
# self.check_results(
# out_dp_prog.history["loss"][0], out_dp_ir.history["loss"][0]
# )

def test_mp(self):
self.enable_pir(False)
Expand Down