[EPLB] Mask padding in EPLB load recording (#38128)

Signed-off-by: ilmarkov <markovilya197@gmail.com>
Signed-off-by: Markov Ilya <markovilya19@gmail.com>
Co-authored-by: Markov Ilya <markovilya19@gmail.com>
This commit is contained in:
Ilya Markov
2026-06-28 19:43:58 -07:00
committed by GitHub
co-authored by Markov Ilya
parent 311ad689ad
commit a2abce646f
17 changed files with 289 additions and 14 deletions
+3
View File
@@ -1332,6 +1332,9 @@ def _test_body_eplb(
eplb_moe_layer.router.eplb_state.should_record_tensor = torch.ones(
(), dtype=torch.bool, device=device
)
eplb_moe_layer.router.eplb_state.num_unpadded_tokens_tensors = [
torch.tensor(0, dtype=torch.int32, device=device)
]
# Get "after" output with rearranged weights and EPLB routing
with set_forward_context(
+66
View File
@@ -61,12 +61,14 @@ def setup_eplb_state(
global_num_experts, dtype=torch.int64, device="cuda"
)
should_record_tensor = torch.ones((), dtype=torch.bool, device="cuda")
num_unpadded_tokens_tensors = [torch.tensor(0, dtype=torch.int32, device="cuda")]
return EplbLayerState(
expert_load_view=expert_load_view,
logical_to_physical_map=logical_to_physical_map,
logical_replica_count=logical_replica_count,
should_record_tensor=should_record_tensor,
num_unpadded_tokens_tensors=num_unpadded_tokens_tensors,
)
@@ -782,3 +784,67 @@ def test_eplb_map_with_redundancy(
torch.testing.assert_close(load, exp_load)
else:
assert load.sum().item() == 0
@pytest.mark.parametrize(
"l2p_map, replica_count, num_physical, topk_ids, "
"num_unpadded, expected_out, expected_load",
[
pytest.param(
[[0], [1], [2], [3]],
[1, 1, 1, 1],
4,
[[0, 1], [2, 3], [0, 2], [1, 3]],
2,
[[0, 1], [2, 3], [0, 2], [1, 3]],
# only rows 0,1 counted: expert 0→1, 1→1, 2→1, 3→1
[1, 1, 1, 1],
id="half_padded",
),
pytest.param(
# record everything (None = no padding info)
[[0], [1], [2], [3]],
[1, 1, 1, 1],
4,
[[0, 1], [2, 3], [0, 2], [1, 3]],
None,
[[0, 1], [2, 3], [0, 2], [1, 3]],
[2, 2, 2, 2],
id="no_padding_info",
),
],
)
def test_eplb_map_num_unpadded_tokens(
l2p_map,
replica_count,
num_physical,
topk_ids,
num_unpadded,
expected_out,
expected_load,
):
l2p = torch.tensor(l2p_map, dtype=torch.int64, device="cuda")
rc = torch.tensor(replica_count, dtype=torch.int64, device="cuda")
load = torch.zeros(num_physical, dtype=torch.int32, device="cuda")
rec = torch.tensor(True, dtype=torch.bool, device="cuda")
ids = torch.tensor(topk_ids, dtype=torch.int32, device="cuda")
num_unpadded_t = (
torch.tensor(num_unpadded, dtype=torch.int32, device="cuda")
if num_unpadded is not None
else None
)
out = eplb_map_to_physical_and_record(
topk_ids=ids,
expert_load_view=load,
logical_to_physical_map=l2p,
logical_replica_count=rc,
record_enabled=rec,
num_unpadded_tokens=num_unpadded_t,
)
exp_out = torch.tensor(expected_out, dtype=out.dtype, device="cuda")
torch.testing.assert_close(out, exp_out)
exp_load = torch.tensor(expected_load, dtype=torch.int32, device="cuda")
torch.testing.assert_close(load, exp_load)