[CI][RL] Consolidate RL entrypoint tests under tests/entrypoints/rl (#59948)
Signed-off-by: wang.yuqi <yuqi.wang@daocloud.io> Signed-off-by: aoshen02 <aoshen@inferact.ai> Co-authored-by: wang.yuqi <yuqi.wang@daocloud.io> Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
co-authored by
wang.yuqi
Claude Opus 5.5
parent
bc8dd4ad27
commit
42ecc8eebe
@@ -1,73 +0,0 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
|
||||
import io
|
||||
|
||||
import numpy as np
|
||||
import pybase64 as base64
|
||||
import pytest
|
||||
|
||||
from ...utils import RemoteOpenAIServer
|
||||
|
||||
MODEL_NAME = "TitanML/tiny-mixtral"
|
||||
|
||||
# tiny-mixtral config: 8 local experts, top-2 routing, 2 hidden layers.
|
||||
# Its published sliding window is incompatible with AuxOutput retention.
|
||||
NUM_LOCAL_EXPERTS = 8
|
||||
NUM_EXPERTS_PER_TOK = 2
|
||||
NUM_HIDDEN_LAYERS = 2
|
||||
|
||||
|
||||
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
|
||||
num_tokens, num_layers, topk = routed_experts.shape
|
||||
assert num_tokens > 0
|
||||
assert num_layers == NUM_HIDDEN_LAYERS
|
||||
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")
|
||||
def server():
|
||||
args = [
|
||||
"--max-model-len",
|
||||
"256",
|
||||
"--max-num-seqs",
|
||||
"32",
|
||||
"--enforce-eager",
|
||||
"--enable-return-routed-experts",
|
||||
"--hf-overrides",
|
||||
'{"sliding_window": null}',
|
||||
]
|
||||
with RemoteOpenAIServer(
|
||||
MODEL_NAME,
|
||||
args,
|
||||
env_dict={"VLLM_USE_V2_MODEL_RUNNER": "1"},
|
||||
) as remote_server:
|
||||
yield remote_server
|
||||
|
||||
|
||||
@pytest.mark.asyncio
|
||||
async def test_routed_experts(server):
|
||||
"""Test that /v1/completions returns routed_experts when enabled."""
|
||||
async with server.get_async_client() as client:
|
||||
result = await client.completions.create(
|
||||
model=MODEL_NAME,
|
||||
prompt="Hello, world",
|
||||
max_tokens=10,
|
||||
temperature=0,
|
||||
extra_body={"return_token_ids": True},
|
||||
)
|
||||
|
||||
choice = result.model_dump()["choices"][0]
|
||||
|
||||
assert choice["token_ids"] is not None
|
||||
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
|
||||
)
|
||||
Reference in New Issue
Block a user