Skip to content

fix(plugin): expose backend ops required by TP comm overlap #95

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

Open
lihongyang1990 wants to merge 1 commit into
base: main
Choose a base branch
from
Open

fix(plugin): expose backend ops required by TP comm overlap #95

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
12 changes: 12 additions & 0 deletions transformer_engine/plugin/core/backends/vendor/cuda/cuda.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1992,6 +1992,7 @@ def create_comm_overlap_p2p(
aggregate: bool = False,
) -> "CommOverlapP2P":
tex = self._get_tex()
comm_type = tex.CommOverlapType(int(comm_type)) if comm_type is not None else None
return tex.CommOverlapP2P(
buffer_shape,
buffer_dtype,
Expand All @@ -2008,3 +2009,14 @@ def create_comm_overlap_p2p(
use_ce,
aggregate,
)

def device_supports_multicast(self, device_id=-1):
return self._get_tex().device_supports_multicast(device_id)

def ubuf_built_with_mpi(self):
tex = self._get_tex()
return tex.ubuf_built_with_mpi()

def get_stream_priority_range(self, device_id=-1):
tex = self._get_tex()
return tex.get_stream_priority_range(device_id)
25 changes: 25 additions & 0 deletions transformer_engine/plugin/core/backends/vendor/cuda/register_ops.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1150,6 +1150,31 @@ def register_builtins(registry) -> None:
vendor="CUDA",
priority=100,
),
OpImpl(
op_name="device_supports_multicast",
impl_id="vendor.cuda",
kind=BackendImplKind.VENDOR,
fn=_bind_is_available(backend.device_supports_multicast, is_avail),
vendor="CUDA",
priority=100,
),

OpImpl(
op_name="ubuf_built_with_mpi",
impl_id="vendor.cuda",
kind=BackendImplKind.VENDOR,
fn=_bind_is_available(backend.ubuf_built_with_mpi, is_avail),
vendor="CUDA",
priority=100,
),
OpImpl(
op_name="get_stream_priority_range",
impl_id="vendor.cuda",
kind=BackendImplKind.VENDOR,
fn=_bind_is_available(backend.get_stream_priority_range, is_avail),
vendor="CUDA",
priority=100,
),
# FlashAttention class getter
OpImpl(
op_name="get_flash_attention_class",
Expand Down
12 changes: 12 additions & 0 deletions transformer_engine/plugin/core/backends/vendor/hygon/hygon.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1954,6 +1954,7 @@ def create_comm_overlap_p2p(
aggregate: bool = False,
) -> "CommOverlapP2P":
tex = self._get_tex()
comm_type = tex.CommOverlapType(int(comm_type)) if comm_type is not None else None
return tex.CommOverlapP2P(
buffer_shape,
buffer_dtype,
Expand All @@ -1970,3 +1971,14 @@ def create_comm_overlap_p2p(
use_ce,
aggregate,
)

def device_supports_multicast(self, device_id=-1):
return self._get_tex().device_supports_multicast(device_id)

def ubuf_built_with_mpi(self):
tex = self._get_tex()
return tex.ubuf_built_with_mpi()

def get_stream_priority_range(self, device_id=-1):
tex = self._get_tex()
return tex.get_stream_priority_range(device_id)
24 changes: 24 additions & 0 deletions transformer_engine/plugin/core/backends/vendor/hygon/register_ops.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1084,6 +1084,30 @@ def register_builtins(registry) -> None:
vendor="HYGON",
priority=100,
),
OpImpl(
op_name="device_supports_multicast",
impl_id="vendor.hygon",
kind=BackendImplKind.VENDOR,
fn=_bind_is_available(backend.device_supports_multicast, is_avail),
vendor="HYGON",
priority=100,
),
OpImpl(
op_name="ubuf_built_with_mpi",
impl_id="vendor.hygon",
kind=BackendImplKind.VENDOR,
fn=_bind_is_available(backend.ubuf_built_with_mpi, is_avail),
vendor="HYGON",
priority=100,
),
OpImpl(
op_name="get_stream_priority_range",
impl_id="vendor.hygon",
kind=BackendImplKind.VENDOR,
fn=_bind_is_available(backend.get_stream_priority_range, is_avail),
vendor="HYGON",
priority=100,
),
# FlashAttention class getter
OpImpl(
op_name="get_flash_attention_class",
Expand Down
Loading