Skip to content

Add SdpaFwdOp::evaluate on meta device #5086

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 8 commits into from
Sep 2, 2025
Merged

Add SdpaFwdOp::evaluate on meta device #5086

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
8 commits
Select commit Hold shift + click to select a range
927f3df
Add SdpaFwdOp::evaluate on meta device
zasdfgbnm Aug 28, 2025
9ebc03b
save
zasdfgbnm Aug 29, 2025
87d8b28
save
zasdfgbnm Aug 29, 2025
5512558
save
zasdfgbnm Aug 29, 2025
6cb326f
save
zasdfgbnm Aug 29, 2025
fcc6753
save
zasdfgbnm Aug 29, 2025
5394e1c
save
zasdfgbnm Aug 29, 2025
9c4e1cb
cleanup
zasdfgbnm Aug 29, 2025
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
52 changes: 43 additions & 9 deletions csrc/ir/nodes.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4890,6 +4890,39 @@ std::string SdpaFwdOp::toInlineString(int indent_size) const {
NVF_CHECK(false, "Tensor op can not be printed inline");
}

namespace spda_meta {

std::tuple<
at::Tensor,
at::Tensor,
at::Tensor,
at::Tensor,
c10::SymInt,
c10::SymInt,
at::Tensor,
at::Tensor,
at::Tensor>
_scaled_dot_product_flash_attention_meta(const at::Tensor& query) {
const auto sizes = query.sizes();
const int batch_size = sizes[0];
int num_heads = sizes[1];
int seqlen_q = sizes[2];
auto logsumexp = at::empty(
{batch_size, num_heads, seqlen_q}, query.options().dtype(at::kFloat));
return std::make_tuple(
query,
logsumexp,
at::Tensor(),
at::Tensor(),
c10::SymInt(seqlen_q),
c10::SymInt(seqlen_q),
at::Tensor(),
at::Tensor(),
at::Tensor());
}

} // namespace spda_meta

std::vector<PolymorphicValue> SdpaFwdOp::evaluate(
const ExpressionEvaluator& ee,
const std::vector<PolymorphicValue>& inputs) const {
Expand Down Expand Up @@ -4961,15 +4994,16 @@ std::vector<PolymorphicValue> SdpaFwdOp::evaluate(
key_seq_len,
philox_seed,
philox_offset,
debug_attn_mask] =
at::_scaled_dot_product_flash_attention(
query,
key,
value,
dropout_p,
is_causal,
/*return_debug_mask=*/false,
scale);
debug_attn_mask] = query.is_meta()
? spda_meta::_scaled_dot_product_flash_attention_meta(query)
: at::_scaled_dot_product_flash_attention(
query,
key,
value,
dropout_p,
is_causal,
/*return_debug_mask=*/false,
scale);

// If the inputs were padded, slice the output to restore the original
// size
Expand Down
82 changes: 76 additions & 6 deletions tests/cpp/test_sdpa_node.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -46,8 +46,12 @@ using AtenSdpaOut = std::tuple<
at::Tensor,
at::Tensor,
at::Tensor>;

using MetaSdpaOut = std::tuple<at::Tensor, at::Tensor>;

auto validateSdpaFwdOutputs = [](KernelArgumentHolder nvf_out,
AtenSdpaOut aten_out) {
AtenSdpaOut aten_out,
MetaSdpaOut aten_out_meta) {
auto
[attn,
log_sumexp,
Expand All @@ -64,6 +68,12 @@ auto validateSdpaFwdOutputs = [](KernelArgumentHolder nvf_out,
// garbage values for this case, so we skip validating those values.
NVF_CHECK(at::allclose(nvf_out[0].as<at::Tensor>(), attn));
NVF_CHECK(at::allclose(nvf_out[1].as<at::Tensor>(), log_sumexp));

auto [attn_meta, log_sumexp_meta] = aten_out_meta;
EXPECT_EQ(attn.sizes(), attn_meta.sizes());
EXPECT_EQ(log_sumexp.sizes(), log_sumexp_meta.sizes());
EXPECT_EQ(attn.strides(), attn_meta.strides());
EXPECT_EQ(log_sumexp.strides(), log_sumexp_meta.strides());
};

// Check SDPAFwdOp mapping in IdModel and ComputeAtMap.
Expand Down Expand Up @@ -258,7 +268,17 @@ TEST_F(SDPATest, NonCausalAttnConcrete) {

FusionExecutorCache executor_cache(std::move(fusion));
auto nvf_out = executor_cache.runFusionWithInputs({q, k, v});
validateSdpaFwdOutputs(nvf_out, aten_out);

ExpressionEvaluator ee;
ee.bind(executor_cache.fusion()->inputs().at(0), q.to(at::kMeta));
ee.bind(executor_cache.fusion()->inputs().at(1), k.to(at::kMeta));
ee.bind(executor_cache.fusion()->inputs().at(2), v.to(at::kMeta));
auto a = ee.evaluate(executor_cache.fusion()->outputs().at(0));
MetaSdpaOut aten_out_meta = {
ee.evaluate(executor_cache.fusion()->outputs().at(0)).as<at::Tensor>(),
ee.evaluate(executor_cache.fusion()->outputs().at(1)).as<at::Tensor>(),
};
validateSdpaFwdOutputs(nvf_out, aten_out, aten_out_meta);
}

TEST_F(SDPATest, NonCausalAttnSymbolic) {
Expand Down Expand Up @@ -305,7 +325,16 @@ TEST_F(SDPATest, NonCausalAttnSymbolic) {

FusionExecutorCache executor_cache(std::move(fusion));
auto nvf_out = executor_cache.runFusionWithInputs({q, k, v});
validateSdpaFwdOutputs(nvf_out, aten_out);

ExpressionEvaluator ee;
ee.bind(executor_cache.fusion()->inputs().at(0), q.to(at::kMeta));
ee.bind(executor_cache.fusion()->inputs().at(1), k.to(at::kMeta));
ee.bind(executor_cache.fusion()->inputs().at(2), v.to(at::kMeta));
MetaSdpaOut aten_out_meta = {
ee.evaluate(executor_cache.fusion()->outputs().at(0)).as<at::Tensor>(),
ee.evaluate(executor_cache.fusion()->outputs().at(1)).as<at::Tensor>(),
};
validateSdpaFwdOutputs(nvf_out, aten_out, aten_out_meta);
}

TEST_F(SDPATest, CausalAttn) {
Expand Down Expand Up @@ -351,7 +380,16 @@ TEST_F(SDPATest, CausalAttn) {

FusionExecutorCache executor_cache(std::move(fusion));
auto nvf_out = executor_cache.runFusionWithInputs({q, k, v});
validateSdpaFwdOutputs(nvf_out, aten_out);

ExpressionEvaluator ee;
ee.bind(executor_cache.fusion()->inputs().at(0), q.to(at::kMeta));
ee.bind(executor_cache.fusion()->inputs().at(1), k.to(at::kMeta));
ee.bind(executor_cache.fusion()->inputs().at(2), v.to(at::kMeta));
MetaSdpaOut aten_out_meta = {
ee.evaluate(executor_cache.fusion()->outputs().at(0)).as<at::Tensor>(),
ee.evaluate(executor_cache.fusion()->outputs().at(1)).as<at::Tensor>(),
};
validateSdpaFwdOutputs(nvf_out, aten_out, aten_out_meta);
}

TEST_F(SDPATest, PairwiseLogicalDomainMap) {
Expand Down Expand Up @@ -828,7 +866,23 @@ TEST_F(SDPATest, Sharded_SdpaFwd) {
FusionExecutorCache executor_cache(std::move(fusion));
auto nvf_out = executor_cache.runFusionWithInputs(
{q.unsqueeze(0), k.unsqueeze(0), v.unsqueeze(0)});
validateSdpaFwdOutputs(nvf_out, aten_out);

ExpressionEvaluator ee;
ee.bind(
executor_cache.fusion()->inputs().at(0), q.to(at::kMeta).unsqueeze(0));
ee.bind(
executor_cache.fusion()->inputs().at(1), k.to(at::kMeta).unsqueeze(0));
ee.bind(
executor_cache.fusion()->inputs().at(2), v.to(at::kMeta).unsqueeze(0));
MetaSdpaOut aten_out_meta = {
ee.evaluate(executor_cache.fusion()->outputs().at(0))
.as<at::Tensor>()
.squeeze(0),
ee.evaluate(executor_cache.fusion()->outputs().at(1))
.as<at::Tensor>()
.squeeze(0),
};
validateSdpaFwdOutputs(nvf_out, aten_out, aten_out_meta);
}

// TODO: Remove/update when https://github.com/NVIDIA/Fuser/issues/2563 is
Expand Down Expand Up @@ -1019,7 +1073,23 @@ TEST_F(SDPATest, ComputeAt) {
FusionExecutorCache executor_cache(std::move(fusion));
auto nvf_out = executor_cache.runFusionWithInputs(
{q.unsqueeze(0), k.unsqueeze(0), v.unsqueeze(0)});
validateSdpaFwdOutputs(nvf_out, aten_out);

ExpressionEvaluator ee;
ee.bind(
executor_cache.fusion()->inputs().at(0), q.to(at::kMeta).unsqueeze(0));
ee.bind(
executor_cache.fusion()->inputs().at(1), k.to(at::kMeta).unsqueeze(0));
ee.bind(
executor_cache.fusion()->inputs().at(2), v.to(at::kMeta).unsqueeze(0));
MetaSdpaOut aten_out_meta = {
ee.evaluate(executor_cache.fusion()->outputs().at(0))
.as<at::Tensor>()
.squeeze(0),
ee.evaluate(executor_cache.fusion()->outputs().at(1))
.as<at::Tensor>()
.squeeze(0),
};
validateSdpaFwdOutputs(nvf_out, aten_out, aten_out_meta);
}

} // namespace nvfuser