Skip to content
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

[JAX] Scale sequence length in CP tests to avoid tiny sizes. #1347

Merged
merged 2 commits into from
Dec 4, 2024
Merged

[JAX] Scale sequence length in CP tests to avoid tiny sizes. #1347

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
0c4bb90
Scale sequence length in CP tests to avoid tiny sizes.
mgoldfarb-nvidia Nov 26, 2024
ce27b87
Merge branch 'main' into mgoldfarb/fix_cp_test_failure_at_8nodes
mgoldfarb-nvidia Dec 2, 2024
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
11 changes: 9 additions & 2 deletions tests/jax/test_distributed_fused_attn.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -341,8 +341,9 @@ def ref_func(query, kv, mask):
@pytest.mark.parametrize(
"data_shape",
[
pytest.param([2, 512, 12, 128], id="2-512-12-128"),
pytest.param([4, 1024, 16, 64], id="4-1024-16-64"),
# Sequence lengths will be scaled by CP so that we don't run with tiny sizes.
pytest.param([2, 128, 12, 128], id="2-128xCP-12-128"),
pytest.param([4, 256, 16, 64], id="4-256xCP-16-64"),
],
)
@pytest.mark.parametrize("kv_groups", [1, 4, 8, 12, 16])
Expand Down Expand Up @@ -423,6 +424,12 @@ def impl_test_contex_parallel_attn(
qkv_format = get_qkv_format(qkv_layout)

batch, seqlen, num_head, hidden = data_shape

# Scale the sequence length by 2*CP so its never too small as we scale up test.
# 2*CP is used since we split into two CP groups for load balancing.
seqlen = seqlen * cp_size * 2
data_shape = batch, seqlen, num_head, hidden

num_kv_heads = num_head // kv_groups
scaling_factor = 1.0 / np.sqrt(num_head)

Expand Down
Loading