Skip to content

feat(ascend): enable InfiniOps flash attention #558

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
voltjia merged 1 commit into from
Sep 9, 2026
Merged

feat(ascend): enable InfiniOps flash attention #558

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
5 changes: 3 additions & 2 deletions csrc/engine/infer_engine.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -76,9 +76,10 @@ InferEngine::InferEngine(
&& device_type != infinicore::Device::Type::kMoore
&& device_type != infinicore::Device::Type::kCambricon
&& device_type != infinicore::Device::Type::kIluvatar
&& device_type != infinicore::Device::Type::kHygon) {
&& device_type != infinicore::Device::Type::kHygon
&& device_type != infinicore::Device::Type::kAscend) {
throw std::invalid_argument(
"flash-attn is only available on NVIDIA, MetaX, Moore, Cambricon, Iluvatar, and Hygon devices");
"flash-attn is only available on NVIDIA, MetaX, Moore, Cambricon, Iluvatar, Hygon, and Ascend devices");
}
const auto *paged_cache_config = dynamic_cast<const cache::PagedKVCacheConfig *>(cache_config);
if (paged_cache_config == nullptr) {
Expand Down
9 changes: 7 additions & 2 deletions csrc/infinicore/src/ops/mha_kvcache/mha_kvcache_infiniops.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -18,7 +18,8 @@ using TensorMeta = ::infinicore::op::infiniops::TensorMeta;
// InfiniOps provides device-aware default selection for these operators.
std::size_t implementation_index_for_device(
infini::ops::Device::Type device_type) {
if (device_type == infini::ops::Device::Type::kIluvatar) {
if (device_type == infini::ops::Device::Type::kIluvatar
|| device_type == infini::ops::Device::Type::kAscend) {
return 0;
}
if (device_type == infini::ops::Device::Type::kMoore) {
Expand All @@ -41,7 +42,8 @@ bool is_supported(const Tensor &out,
&& device_type != Device::Type::kMoore
&& device_type != Device::Type::kCambricon
&& device_type != Device::Type::kIluvatar
&& device_type != Device::Type::kHygon)
&& device_type != Device::Type::kHygon
&& device_type != Device::Type::kAscend)
|| q->ndim() != 4
|| out->ndim() != 4
|| k_cache->ndim() != 4
Expand Down Expand Up @@ -214,6 +216,9 @@ static bool registered = []() {
MhaKVCache::plan_dispatcher().registerDevice(Device::Type::kHygon, &plan);
MhaKVCache::run_dispatcher().registerDevice(Device::Type::kHygon, &run);
MhaKVCache::cleanup_dispatcher().registerDevice(Device::Type::kHygon, &cleanup);
MhaKVCache::plan_dispatcher().registerDevice(Device::Type::kAscend, &plan);
MhaKVCache::run_dispatcher().registerDevice(Device::Type::kAscend, &run);
MhaKVCache::cleanup_dispatcher().registerDevice(Device::Type::kAscend, &cleanup);
return true;
}();

Expand Down
12 changes: 10 additions & 2 deletions csrc/infinicore/src/ops/multi_head_attention_varlen/mha_varlen_infiniops.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -18,7 +18,8 @@ using TensorMeta = ::infinicore::op::infiniops::TensorMeta;
// InfiniOps provides device-aware default selection for these operators.
std::size_t implementation_index_for_device(
infini::ops::Device::Type device_type) {
if (device_type == infini::ops::Device::Type::kIluvatar) {
if (device_type == infini::ops::Device::Type::kIluvatar
|| device_type == infini::ops::Device::Type::kAscend) {
return 0;
}
if (device_type == infini::ops::Device::Type::kMoore) {
Expand All @@ -45,7 +46,8 @@ bool is_supported(const Tensor &out,
&& device_type != Device::Type::kMoore
&& device_type != Device::Type::kCambricon
&& device_type != Device::Type::kIluvatar
&& device_type != Device::Type::kHygon)
&& device_type != Device::Type::kHygon
&& device_type != Device::Type::kAscend)
|| q->ndim() != 3
|| out->ndim() != 3
|| ((paged && (k->ndim() != 4 || v->ndim() != 4))
Expand Down Expand Up @@ -258,6 +260,12 @@ static bool registered = []() {
Device::Type::kHygon, &run);
MultiheadAttentionVarlen::cleanup_dispatcher().registerDevice(
Device::Type::kHygon, &cleanup);
MultiheadAttentionVarlen::plan_dispatcher().registerDevice(
Device::Type::kAscend, &plan);
MultiheadAttentionVarlen::run_dispatcher().registerDevice(
Device::Type::kAscend, &run);
MultiheadAttentionVarlen::cleanup_dispatcher().registerDevice(
Device::Type::kAscend, &cleanup);
return true;
}();

Expand Down
5 changes: 3 additions & 2 deletions test/static/test_infinicore_runtime_contracts.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -52,9 +52,9 @@ def test_exact_infiniops_adapter_set_is_registered(self) -> None:
"gelu/gelu_infiniops.cc": 1,
"gelutanh/gelutanh_infiniops.cc": 1,
"gemm/gemm_infiniops.cc": 3,
"mha_kvcache/mha_kvcache_infiniops.cc": 18,
"mha_kvcache/mha_kvcache_infiniops.cc": 21,
"mul/mul_infiniops.cc": 3,
"multi_head_attention_varlen/mha_varlen_infiniops.cc": 18,
"multi_head_attention_varlen/mha_varlen_infiniops.cc": 21,
"ones/ones_infiniops.cc": 3,
"paged_caching/paged_caching_infiniops.cc": 3,
"random_sample/random_sample_infiniops.cc": 1,
Expand Down Expand Up @@ -324,6 +324,7 @@ def test_unavailable_configurable_surfaces_are_rejected(self) -> None:
self.assertIn(
"device_type != infinicore::Device::Type::kCambricon", infer_engine
)
self.assertIn("device_type != infinicore::Device::Type::kAscend", infer_engine)
self.assertIn("paged_cache_config->num_blocks() == 0", infer_engine)
self.assertIn("paged_cache_config->block_size() == 0", infer_engine)
self.assertIn("paged_cache_config->block_size() % 256 != 0", infer_engine)
Expand Down
11 changes: 10 additions & 1 deletion test/static/test_modern_infinicore_compatibility.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -210,6 +210,7 @@ def test_current_infiniops_backend_bridges_are_registered(self) -> None:
"kCambricon",
"kIluvatar",
"kHygon",
"kAscend",
):
with self.subTest(adapter=relative_path, device=device):
self.assertIn(f"device_type != Device::Type::{device}", attention)
Expand All @@ -219,6 +220,10 @@ def test_current_infiniops_backend_bridges_are_registered(self) -> None:
)
self.assertIn("configForImplementation<", attention)
self.assertIn("implementation_index_for_device(device_type)", attention)
self.assertIn(
"device_type == infini::ops::Device::Type::kAscend",
attention,
)
self.assertIn("device_type == Device::Type::kMoore", attention)
self.assertIn("device_type == Device::Type::kIluvatar", attention)
self.assertIn("return 0;", attention)
Expand Down Expand Up @@ -258,7 +263,11 @@ def test_current_infiniops_backend_bridges_are_registered(self) -> None:
engine,
)
self.assertIn(
"flash-attn is only available on NVIDIA, MetaX, Moore, Cambricon, Iluvatar, and Hygon devices",
"device_type != infinicore::Device::Type::kAscend",
engine,
)
self.assertIn(
"flash-attn is only available on NVIDIA, MetaX, Moore, Cambricon, Iluvatar, Hygon, and Ascend devices",
engine,
)
self.assertIn(
Expand Down
Loading