Skip to content

fix(attention test): give the mask tests the axis rules the cp gate reads #5134

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

fix(attention test): give the mask tests the axis rules the cp gate reads #5134

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
111 changes: 61 additions & 50 deletions tests/unit/attention_test.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -622,6 +622,8 @@ def test_dot_product_mask_uses_original_load_balanced_positions(self):
causal_block_size=4,
context_parallel_load_balance=True,
context_sharding="context",
ulysses_context_sharding="context_usp_ulysses",
logical_axis_rules=[["activation_q_length", ["context"]]],
shard_mode="auto",
debug_sharding=False,
eval_interval=-1,
Expand All @@ -630,23 +632,24 @@ def test_dot_product_mask_uses_original_load_balanced_positions(self):
if len(devices) < 2:
self.skipTest("Need at least 2 devices to test chunk mask")
mesh = Mesh(devices[:2], ["context"])
op = AttentionOp(
config=config,
num_query_heads=1,
num_kv_heads=1,
max_target_length=sequence_length,
mesh=mesh,
attention_kernel="dot_product",
attention_type=AttentionType.BLOCK_DIFFUSION,
)
with nn_partitioning.axis_rules(config.logical_axis_rules):
op = AttentionOp(
config=config,
num_query_heads=1,
num_kv_heads=1,
max_target_length=sequence_length,
mesh=mesh,
attention_kernel="dot_product",
attention_type=AttentionType.BLOCK_DIFFUSION,
)

mask = op.generate_attention_mask(
query,
key,
segment_ids,
MODEL_MODE_TRAIN,
segment_positions=positions,
)
mask = op.generate_attention_mask(
query,
key,
segment_ids,
MODEL_MODE_TRAIN,
segment_positions=positions,
)

expected = np.asarray(positions[0])[:, None] // 4 >= np.asarray(positions[0])[None, :] // 4
np.testing.assert_array_equal(np.asarray(mask == 0.0)[0, 0, 0], expected)
Expand Down Expand Up @@ -920,6 +923,7 @@ def test_dot_product_local_mask_uses_segment_positions(self):
config = types.SimpleNamespace(
context_parallel_load_balance=True,
context_sharding="context",
ulysses_context_sharding="context_usp_ulysses",
using_pipeline_parallelism=False,
logical_axis_rules=[["segment_ids_batch", ["context"]]],
shard_mode="auto",
Expand All @@ -936,24 +940,27 @@ def test_dot_product_local_mask_uses_segment_positions(self):
query = jnp.zeros((1, seq_len, 1, 128))
key = jnp.zeros((1, seq_len, 1, 128))
decoder_segment_ids = jnp.ones((1, seq_len), dtype=jnp.int32)
op = AttentionOp(
config=config,
num_query_heads=1,
num_kv_heads=1,
max_target_length=seq_len,
mesh=mesh,
attention_kernel="dot_product",
attention_type=AttentionType.LOCAL_SLIDING,
sliding_window_size=sliding_window_size,
)
# Only the query-length rule: `segment_ids_batch` -> context would shard this batch of 1
# across the 4-way context mesh.
with nn_partitioning.axis_rules([["activation_q_length", ["context"]]]):
op = AttentionOp(
config=config,
num_query_heads=1,
num_kv_heads=1,
max_target_length=seq_len,
mesh=mesh,
attention_kernel="dot_product",
attention_type=AttentionType.LOCAL_SLIDING,
sliding_window_size=sliding_window_size,
)

mask = op.generate_attention_mask(
query,
key,
decoder_segment_ids,
MODEL_MODE_TRAIN,
segment_positions=positions,
)
mask = op.generate_attention_mask(
query,
key,
decoder_segment_ids,
MODEL_MODE_TRAIN,
segment_positions=positions,
)

expected_mask = np.zeros((seq_len, seq_len), dtype=np.bool_)
for r, q_pos in enumerate(np.asarray(positions[0])):
Expand All @@ -967,6 +974,7 @@ def test_dot_product_chunk_mask_uses_segment_positions(self):
config = types.SimpleNamespace(
context_parallel_load_balance=True,
context_sharding="context",
ulysses_context_sharding="context_usp_ulysses",
using_pipeline_parallelism=False,
logical_axis_rules=[["segment_ids_batch", ["context"]]],
shard_mode="auto",
Expand All @@ -983,24 +991,27 @@ def test_dot_product_chunk_mask_uses_segment_positions(self):
query = jnp.zeros((1, seq_len, 1, 128))
key = jnp.zeros((1, seq_len, 1, 128))
decoder_segment_ids = jnp.ones((1, seq_len), dtype=jnp.int32)
op = AttentionOp(
config=config,
num_query_heads=1,
num_kv_heads=1,
max_target_length=seq_len,
mesh=mesh,
attention_kernel="dot_product",
attention_type=AttentionType.CHUNK,
chunk_attn_window_size=chunk_size,
)
# Only the query-length rule: `segment_ids_batch` -> context would shard this batch of 1
# across the 4-way context mesh.
with nn_partitioning.axis_rules([["activation_q_length", ["context"]]]):
op = AttentionOp(
config=config,
num_query_heads=1,
num_kv_heads=1,
max_target_length=seq_len,
mesh=mesh,
attention_kernel="dot_product",
attention_type=AttentionType.CHUNK,
chunk_attn_window_size=chunk_size,
)

mask = op.generate_attention_mask(
query,
key,
decoder_segment_ids,
MODEL_MODE_TRAIN,
segment_positions=positions,
)
mask = op.generate_attention_mask(
query,
key,
decoder_segment_ids,
MODEL_MODE_TRAIN,
segment_positions=positions,
)

expected_mask = np.zeros((seq_len, seq_len), dtype=np.bool_)
for r, q_pos in enumerate(np.asarray(positions[0])):
Expand Down
Loading