Skip to content
This repository has been archived by the owner on Sep 15, 2021. It is now read-only.

feat: add (scatter, gather, scatter_reduce) and all inplace version communication primitives #37

Merged
merged 2 commits into from
Jul 15, 2021
Merged

feat: add (scatter, gather, scatter_reduce) and all inplace version communication primitives #37

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
5b0bcd4
feat: add (scatter, gather, scatter_reduce) communication primitives
liuhatry Jul 14, 2021
a7b806c
add inplace version communication
liuhatry Jul 15, 2021
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
21 changes: 5 additions & 16 deletions bagua-core-internal/src/comm_ops/centralized_full_precision_synchronous.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -29,32 +29,21 @@ impl CommOpTrait for CentralizedFullPrecisionSynchronous {
true,
&mut |c, t| {
tracing::debug!("internode communication started");
let temp_buf = CUDA_DEVICE_MEMORY_POOL[t.raw.device_id]
.try_pull(t.raw.num_elem_allocated * t.raw.dtype.bytes())
.expect("cannot allocate cuda memory");
let mut temp_tensor = BaguaTensorRaw {
ptr: temp_buf.ptr,
num_elem_allocated: t.raw.num_elem_allocated,
dtype: t.raw.dtype.clone(),
num_elem: t.raw.num_elem,
device_id: t.raw.device_id,
pool_allocations: vec![Arc::new(temp_buf)],
};
if self.scattergather {
tracing::debug!("start alltoall");
c.alltoall(&t.raw, &mut temp_tensor);
c.alltoall_inplace(&mut t.raw);
tracing::debug!("start reduce_sum");
if self.average {
temp_tensor.reduce_mean_inplace(c.nranks, c.rank, c.stream_ptr);
t.raw.reduce_mean_inplace(c.nranks, c.rank, c.stream_ptr);
} else {
temp_tensor.reduce_sum_inplace(c.nranks, c.rank, c.stream_ptr);
t.raw.reduce_sum_inplace(c.nranks, c.rank, c.stream_ptr);
}
tracing::debug!("start allgather");
c.allgather(&temp_tensor, &mut t.raw);
c.allgather_inplace(&mut t.raw);
tracing::debug!("internode communication done")
} else {
tracing::debug!("start allreduce");
c.allreduce(&mut t.raw, BaguaReductionOp::SUM);
c.allreduce_inplace(&mut t.raw, BaguaReductionOp::SUM);
tracing::debug!("internode communication done");
if self.average {
t.raw.divide_inplace(stream_ptr, c.nranks as f32);
Expand Down
26 changes: 6 additions & 20 deletions bagua-core-internal/src/comm_ops/centralized_low_precision_synchronous.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -29,31 +29,17 @@ impl CommOpTrait for CentralizedLowPrecisionSynchronous {
true,
&mut |c, t| {
tracing::debug!("start compress");
let compressed_tensor = t
let mut compressed_tensor = t
.raw
.compress(&self.compression_method, c.nranks, c.stream_ptr, -1)
.expect("cannot compress tensor");
let temp_buf = CUDA_DEVICE_MEMORY_POOL[t.raw.device_id]
.try_pull(
compressed_tensor.num_elements_allocated()
* compressed_tensor.dtype().bytes(),
)
.expect("cannot allocate cuda memory");
let mut temp_tensor = BaguaTensorRaw {
ptr: temp_buf.ptr,
num_elem_allocated: compressed_tensor.num_elements_allocated(),
dtype: compressed_tensor.dtype().clone(),
num_elem: compressed_tensor.num_elements(),
device_id: compressed_tensor.device_id(),
pool_allocations: vec![Arc::new(temp_buf)],
};
tracing::debug!("start alltoall");
c.alltoall(compressed_tensor.as_ref(), &mut temp_tensor);
c.alltoall_inplace(compressed_tensor.as_mut());
tracing::debug!("start decompress");
t.raw.decompress_from(
&self.compression_method,
c.nranks,
&temp_tensor,
compressed_tensor.as_ref(),
c.stream_ptr,
);
tracing::debug!("start reduce_sum");
Expand All @@ -63,7 +49,7 @@ impl CommOpTrait for CentralizedLowPrecisionSynchronous {
t.raw.reduce_sum_inplace(c.nranks, c.rank, c.stream_ptr);
}
tracing::debug!("start compress");
let compressed_tensor = t
let mut compressed_tensor = t
.raw
.compress(
&self.compression_method,
Expand All @@ -73,12 +59,12 @@ impl CommOpTrait for CentralizedLowPrecisionSynchronous {
)
.expect("cannot compress tensor");
tracing::debug!("start allgather");
c.allgather(compressed_tensor.as_ref(), &mut temp_tensor);
c.allgather_inplace(compressed_tensor.as_mut());
tracing::debug!("start decompress");
t.raw.decompress_from(
&self.compression_method,
c.nranks,
&temp_tensor,
compressed_tensor.as_ref(),
c.stream_ptr,
);
tracing::debug!("internode communication done");
Expand Down
2 changes: 1 addition & 1 deletion bagua-core-internal/src/comm_ops/decentralized_full_precision_synchronous.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -74,7 +74,7 @@ impl CommOpTrait for DecentralizedFullPrecisionSynchronous {
if step % comm_interval == 0 {
peer_tensor.clone_from(&t.raw, c.stream_ptr);
let _guard = NCCLGroupGuard::new();
c.allreduce(&mut peer_tensor, BaguaReductionOp::SUM);
c.allreduce_inplace(&mut peer_tensor, BaguaReductionOp::SUM);
peer_tensor.divide_inplace(stream_ptr, c.nranks as f32);
}
}
Expand Down
Loading