[ROCm][DSv4] AITER MegaMoE Fixed-Slot Dispatch For DeepSeek-V4 (#61262)

Signed-off-by: Micah Williamson <micah.williamson@amd.com>
This commit is contained in:
Micah Williamson
2026-10-15 19:10:59 +08:00
committed by GitHub
parent 8b27d27402
commit e8e1e96196
3 changed files with 84 additions and 8 deletions
@@ -110,6 +110,31 @@ def test_experts_load_into_local_slots(monkeypatch):
assert torch.all(w2 == 10 * expert + 3)
@pytest.mark.cpu_test
@pytest.mark.parametrize(
"tokens_across_dp, expected",
[([4, 128], "decode"), ([4, 129], "prefill")],
)
def test_runtime_selected_by_max_tokens_across_dp(
monkeypatch, tokens_across_dp, expected
):
experts = _make_experts(monkeypatch)
experts._runtimes = [(128, "decode"), (4096, "prefill")]
monkeypatch.setattr(mega_moe_module, "is_forward_context_available", lambda: True)
monkeypatch.setattr(
mega_moe_module,
"get_forward_context",
lambda: SimpleNamespace(
dp_metadata=SimpleNamespace(
num_tokens_across_dp_cpu=torch.tensor(tokens_across_dp)
)
),
)
# Every rank must take the same runtime, so a small local batch still gets
# the prefill runtime when another DP rank is over the decode capacity.
assert experts.select_runtime(4) == expected
@pytest.mark.skipif(not current_platform.is_rocm(), reason="Requires ROCm")
def test_shared_expert_route_appended(monkeypatch):
experts = _make_experts(