Skip to content

Enable Partial Rowwise Adam in SSD TBE #4525

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

Enable Partial Rowwise Adam in SSD TBE #4525

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
21 changes: 20 additions & 1 deletion fbgemm_gpu/fbgemm_gpu/tbe/ssd/training.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -180,6 +180,7 @@ def __init__(
# Set the optimizer
assert optimizer in (
OptimType.EXACT_ROWWISE_ADAGRAD,
OptimType.PARTIAL_ROWWISE_ADAM,
), f"Optimizer {optimizer} is not supported by SSDTableBatchedEmbeddingBags"
self.optimizer = optimizer

Expand Down Expand Up @@ -2205,7 +2206,7 @@ def forward(
self.step += 1

# Increment the iteration (value is used for certain optimizers)
self._increment_iteration()
iter_int = self._increment_iteration()

if self.optimizer == OptimType.EXACT_SGD:
raise AssertionError(
Expand All @@ -2226,6 +2227,24 @@ def forward(
common_args, self.optimizer_args, momentum1
)

elif self.optimizer == OptimType.PARTIAL_ROWWISE_ADAM:
momentum2 = invokers.lookup_args_ssd.Momentum(
# pyre-ignore[6]
dev=self.momentum2_dev,
# pyre-ignore[6]
host=self.momentum2_host,
# pyre-ignore[6]
uvm=self.momentum2_uvm,
# pyre-ignore[6]
offsets=self.momentum2_offsets,
# pyre-ignore[6]
placements=self.momentum2_placements,
)

return invokers.lookup_partial_rowwise_adam_ssd.invoke(
common_args, self.optimizer_args, momentum1, momentum2, iter_int
)

@torch.jit.ignore
def _split_optimizer_states_non_kv_zch(
self,
Expand Down
176 changes: 176 additions & 0 deletions fbgemm_gpu/test/tbe/ssd/ssd_split_tbe_training_test.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -473,6 +473,9 @@ def generate_ssd_tbes(
weighted: bool,
lr: float = 0.01, # from SSDTableBatchedEmbeddingBags
eps: float = 1.0e-8, # from SSDTableBatchedEmbeddingBags
weight_decay: float = 0.0, # used by LARS-SGD, LAMB, ADAM, and Rowwise Adagrad
beta1: float = 0.9, # used by Partial Rowwise Adam
beta2: float = 0.999, # used by Partial Rowwise Adam
ssd_shards: int = 1, # from SSDTableBatchedEmbeddingBags
optimizer: OptimType = OptimType.EXACT_ROWWISE_ADAGRAD,
cache_set_scale: float = 1.0,
Expand All @@ -487,6 +490,7 @@ def generate_ssd_tbes(
lazy_bulk_init_enabled: bool = False,
backend_type: BackendType = BackendType.SSD,
enable_raw_embedding_streaming: bool = False,
optimizer_state_dtypes: Dict[str, SparseType] = {}, # noqa: B006
) -> Tuple[SSDTableBatchedEmbeddingBags, List[torch.nn.EmbeddingBag]]:
"""
Generate embedding modules (i,e., SSDTableBatchedEmbeddingBags and
Expand Down Expand Up @@ -561,6 +565,9 @@ def generate_ssd_tbes(
ssd_uniform_init_upper=0.1,
learning_rate=lr,
eps=eps,
weight_decay=weight_decay,
beta1=beta1,
beta2=beta2,
ssd_rocksdb_shards=ssd_shards,
optimizer=optimizer,
pooling_mode=pooling_mode,
Expand Down Expand Up @@ -796,6 +803,16 @@ def split_optimizer_states_(
bucket_asc_ids_list, no_snapshot=False, should_flush=True
)

def assert_close_(self, test: torch.Tensor, ref: torch.Tensor) -> None:
tolerance = 1.0e-4 if test.dtype == torch.float else 1.0e-2

torch.testing.assert_close(
test.float().cpu(),
ref.float().cpu(),
atol=tolerance,
rtol=tolerance,
)

@given(
**default_st, backend_type=st.sampled_from([BackendType.SSD, BackendType.DRAM])
)
Expand Down Expand Up @@ -1014,6 +1031,165 @@ def test_ssd_backward_adagrad(
rtol=tolerance,
)

@given(
**default_st,
backend_type=st.sampled_from([BackendType.SSD, BackendType.DRAM]),
)
@settings(verbosity=Verbosity.verbose, max_examples=MAX_EXAMPLES, deadline=None)
def test_ssd_backward_partial_rowwise_adam(
self,
T: int,
D: int,
B: int,
log_E: int,
L: int,
weighted: bool,
cache_set_scale: float,
pooling_mode: PoolingMode,
weights_precision: SparseType,
output_dtype: SparseType,
share_table: bool,
trigger_bounds_check: bool,
mixed_B: bool,
backend_type: BackendType,
) -> None:
assume(not weighted or pooling_mode == PoolingMode.SUM)
# VBE is currently not supported for PARTIAL_ROWWISE_ADAM optimizer
assume(not mixed_B)

# Constants
lr = 0.5
eps = 0.2
ssd_shards = 2
beta1 = 0.9
beta2 = 0.99
weight_decay = 0.01

# Generate embedding modules and inputs
(
emb,
emb_ref,
) = self.generate_ssd_tbes(
T,
D,
B,
log_E,
L,
weighted,
lr=lr,
eps=eps,
weight_decay=weight_decay,
beta1=beta1,
beta2=beta2,
ssd_shards=ssd_shards,
optimizer=OptimType.PARTIAL_ROWWISE_ADAM,
cache_set_scale=cache_set_scale,
pooling_mode=pooling_mode,
weights_precision=weights_precision,
output_dtype=output_dtype,
share_table=share_table,
backend_type=backend_type,
)

Es = [emb.embedding_specs[t][0] for t in range(T)]
(
indices_list,
per_sample_weights_list,
indices,
offsets,
per_sample_weights,
batch_size_per_feature_per_rank,
) = self.generate_inputs_(
B,
L,
Es,
emb.feature_table_map,
weights_precision=weights_precision,
trigger_bounds_check=trigger_bounds_check,
# mixed_B=mixed_B,
)

# Execute forward
output_ref_list, output = self.execute_ssd_forward_(
emb,
emb_ref,
indices_list,
per_sample_weights_list,
indices,
offsets,
per_sample_weights,
B,
L,
weighted,
batch_size_per_feature_per_rank=batch_size_per_feature_per_rank,
)

# Execute backward
self.execute_ssd_backward_(
output_ref_list,
output,
B,
D,
pooling_mode,
batch_size_per_feature_per_rank,
)

emb.flush()

tolerance = 1.0e-2

emb_test_weights = emb.debug_split_embedding_weights()
split_optimizer_states = self.split_optimizer_states_(emb)

for _, t in self.get_physical_table_arg_indices_(emb.feature_table_map):
(m1, m2) = split_optimizer_states[t]
print(f"{t=} {m1.shape=}, {m2.shape=}")

for f, t in self.get_physical_table_arg_indices_(emb.feature_table_map):
(m1, m2) = split_optimizer_states[t]
# Some optimizers have non-float momentum values
# pyre-ignore[16]
ref_grad = emb_ref[f].weight.grad.cpu().to_dense()
ref_weights = emb_ref[f].weight.cpu()

# Compare momentum2 values: (1 - beta2) * dL^2
m2_ref = (ref_grad.pow(2).mean(dim=1)) * (1.0 - beta2)
self.assert_close_(m2, m2_ref)

# Compare momentum1 values: (1 - beta1) * dL
m1_ref = ref_grad * (1.0 - beta1)
print(f"{m1_ref.shape=}, {m1.shape=}")
self.assert_close_(m1, m1_ref)

# Bias corrections
iter_ = emb.iter.item()
v_hat_t = m2_ref / (1 - beta2**iter_)
v_hat_t = v_hat_t.view(v_hat_t.numel(), 1)
m_hat_t = m1_ref / (1 - beta1**iter_)

# Weight update
ref_weights_updated = (
torch.addcdiv(
ref_weights,
value=-lr,
tensor1=m_hat_t,
tensor2=v_hat_t.sqrt_().add_(eps),
)
- lr * weight_decay * ref_weights
)

if weights_precision == SparseType.FP16:
# Round the reference weight the same way that TBE does
ref_weights_updated = ref_weights_updated.half().float()

# Compare weights
torch.testing.assert_close(
emb_test_weights[t].float().cuda(),
ref_weights_updated.cuda(),
atol=tolerance,
rtol=tolerance,
)

@given(
bulk_init_chunk_size=st.sampled_from([0, 204800]),
lazy_bulk_init_enabled=st.booleans(),
Expand Down
Loading