[AuxOutput] Add block-keyed storage for routed-expert outputs (#45635)

Signed-off-by: aoshen02 <aoshen@inferact.ai>
Signed-off-by: Nick Hill <nickhill123@gmail.com>
Co-authored-by: aoshen02 <aoshen@inferact.ai>
Co-authored-by: OpenAI Codex <codex@openai.com>
Co-authored-by: Roger Wang <hey@rogerw.io>
Co-authored-by: Nick Hill <nickhill123@gmail.com>
Co-authored-by: Codex <noreply@openai.com>
This commit is contained in:
Hongxin Xu
2026-09-21 10:25:05 +00:00
committed by GitHub
co-authored by aoshen02 OpenAI Codex Roger Wang Nick Hill Codex
parent 0aee727ff6
commit 0b7f11a1ee
32 changed files with 3604 additions and 877 deletions
@@ -12,15 +12,13 @@ from ...utils import RemoteOpenAIServer
MODEL_NAME = "TitanML/tiny-mixtral"
# tiny-mixtral config: 8 local experts, top-2 routing, 2 hidden layers.
# The published config has sliding_window=4096, which produces
# SlidingWindowSpec kv-cache groups; RoutedExpertsManager requires a
# FullAttentionSpec group, so we override sliding_window=null below.
# Its published sliding window is incompatible with AuxOutput retention.
NUM_LOCAL_EXPERTS = 8
NUM_EXPERTS_PER_TOK = 2
NUM_HIDDEN_LAYERS = 2
def assert_valid_routed_experts(encoded: str | None) -> None:
def decode_routed_experts(encoded: str | None) -> np.ndarray:
assert encoded is not None
routed_experts = np.load(io.BytesIO(base64.b64decode(encoded)))
assert routed_experts.ndim == 3
@@ -30,6 +28,7 @@ def assert_valid_routed_experts(encoded: str | None) -> None:
assert topk == NUM_EXPERTS_PER_TOK
assert (routed_experts >= 0).all()
assert (routed_experts < NUM_LOCAL_EXPERTS).all()
return routed_experts
@pytest.fixture(scope="module")
@@ -44,7 +43,11 @@ def server():
"--hf-overrides",
'{"sliding_window": null}',
]
with RemoteOpenAIServer(MODEL_NAME, args) as remote_server:
with RemoteOpenAIServer(
MODEL_NAME,
args,
env_dict={"VLLM_USE_V2_MODEL_RUNNER": "1"},
) as remote_server:
yield remote_server
@@ -63,4 +66,8 @@ async def test_routed_experts(server):
choice = result.model_dump()["choices"][0]
assert choice["token_ids"] is not None
assert_valid_routed_experts(choice["routed_experts"])
assert result.usage is not None
routed_experts = decode_routed_experts(choice["routed_experts"])
assert len(routed_experts) == (
result.usage.prompt_tokens + len(choice["token_ids"]) - 1
)
@@ -15,9 +15,7 @@ MODEL_NAME = "TitanML/tiny-mixtral"
GEN_ENDPOINT = "/inference/v1/generate"
# tiny-mixtral config: 8 local experts, top-2 routing, 2 hidden layers.
# The published config has sliding_window=4096, which produces
# SlidingWindowSpec kv-cache groups; RoutedExpertsManager requires a
# FullAttentionSpec group, so we override sliding_window=null below.
# Its published sliding window is incompatible with AuxOutput retention.
NUM_LOCAL_EXPERTS = 8
NUM_EXPERTS_PER_TOK = 2
NUM_HIDDEN_LAYERS = 2
@@ -36,7 +34,13 @@ def server():
'{"sliding_window": null}',
"--enable-scale-out",
]
with RemoteOpenAIServer(MODEL_NAME, args) as remote_server:
with RemoteOpenAIServer(
MODEL_NAME,
args,
env_dict={
"VLLM_USE_V2_MODEL_RUNNER": "1",
},
) as remote_server:
yield remote_server
@@ -56,9 +60,10 @@ async def client(server: RemoteOpenAIServer):
@pytest.mark.asyncio
async def test_generate_routed_experts(client):
"""Test that /inference/v1/generate returns routed_experts when enabled."""
prompt_token_ids = [1, 2, 3]
payload = {
"model": MODEL_NAME,
"token_ids": [1, 2, 3],
"token_ids": prompt_token_ids,
"sampling_params": {"max_tokens": 10, "temperature": 0.0},
"stream": False,
}
@@ -80,3 +85,4 @@ async def test_generate_routed_experts(client):
assert topk == NUM_EXPERTS_PER_TOK
assert (routed_experts >= 0).all()
assert (routed_experts < NUM_LOCAL_EXPERTS).all()
assert num_tokens == len(prompt_token_ids) + len(choice["token_ids"]) - 1