Skip to content

enable VBE for data parallel sharding #2093

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

Closed
wants to merge 1 commit into from
Closed

enable VBE for data parallel sharding #2093

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
6 changes: 5 additions & 1 deletion torchrec/distributed/batched_embedding_kernel.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -776,7 +776,11 @@ def forward(self, features: KeyedJaggedTensor) -> torch.Tensor:
if weights is not None and not torch.is_floating_point(weights):
weights = None
if features.variable_stride_per_key() and isinstance(
self.emb_module, SplitTableBatchedEmbeddingBagsCodegen
self.emb_module,
(
SplitTableBatchedEmbeddingBagsCodegen,
DenseTableBatchedEmbeddingBagsCodegen,
),
):
return self.emb_module(
indices=features.values().long(),
Expand Down
4 changes: 0 additions & 4 deletions torchrec/distributed/sharding/dp_sharding.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -153,10 +153,6 @@ def forward(
Awaitable[Awaitable[SparseFeatures]]: awaitable of awaitable of SparseFeatures.
"""

if sparse_features.variable_stride_per_key():
raise ValueError(
"Dense TBE kernel does not support variable batch per feature"
)
return NoWait(cast(Awaitable[KeyedJaggedTensor], NoWait(sparse_features)))


Expand Down
20 changes: 11 additions & 9 deletions torchrec/distributed/test_utils/test_model_parallel.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -245,15 +245,11 @@ def test_sharding_rw(
SharderType.EMBEDDING_BAG_COLLECTION.value,
]
),
kernel_type=st.sampled_from(
[
EmbeddingComputeKernel.DENSE.value,
],
),
apply_optimizer_in_backward_config=st.sampled_from([None]),
kernel_type=st.just(EmbeddingComputeKernel.DENSE.value),
apply_optimizer_in_backward_config=st.just(None),
# TODO - need to enable optimizer overlapped behavior for data_parallel tables
)
@settings(verbosity=Verbosity.verbose, max_examples=3, deadline=None)
@settings(verbosity=Verbosity.verbose, max_examples=1, deadline=None)
def test_sharding_dp(
self,
sharder_type: str,
Expand Down Expand Up @@ -591,12 +587,13 @@ def test_sharding_twrw(
ShardingType.TABLE_WISE.value,
ShardingType.COLUMN_WISE.value,
ShardingType.ROW_WISE.value,
ShardingType.DATA_PARALLEL.value,
]
),
global_constant_batch=st.booleans(),
pooling=st.sampled_from([PoolingType.SUM, PoolingType.MEAN]),
)
@settings(verbosity=Verbosity.verbose, max_examples=6, deadline=None)
@settings(verbosity=Verbosity.verbose, max_examples=10, deadline=None)
def test_sharding_variable_batch(
self,
sharding_type: str,
Expand All @@ -608,13 +605,18 @@ def test_sharding_variable_batch(
self.skipTest(
"bounds_check_indices on CPU does not support variable length (batch size)"
)
kernel = (
EmbeddingComputeKernel.DENSE.value
if sharding_type == ShardingType.DATA_PARALLEL.value
else EmbeddingComputeKernel.FUSED.value
)
self._test_sharding(
# pyre-ignore[6]
sharders=[
create_test_sharder(
sharder_type=SharderType.EMBEDDING_BAG_COLLECTION.value,
sharding_type=sharding_type,
kernel_type=EmbeddingComputeKernel.FUSED.value,
kernel_type=kernel,
device=self.device,
),
],
Expand Down