[Bugfix]: v1 engine - consider lora adapters in allowed_token_ids (#17855)

Signed-off-by: Ben Browning <bbrownin@redhat.com>
This commit is contained in:
Ben Browning
2025-05-11 00:53:58 -07:00
committed by GitHub
parent eea22a56ab
commit 8132365b74
3 changed files with 154 additions and 5 deletions
+14 -2
View File
@@ -139,6 +139,12 @@ def dummy_model_gate_up() -> nn.Module:
return model
@pytest.fixture(scope="session")
def llama_2_7b_base_huggingface_id():
# used as a base model for testing with sql lora adapter
return "meta-llama/Llama-2-7b-hf"
@pytest.fixture(scope="session")
def sql_lora_huggingface_id():
# huggingface repo id is used to test lora runtime downloading.
@@ -198,6 +204,12 @@ def qwen2vl_lora_files():
return snapshot_download(repo_id="jeeejeee/qwen2-vl-lora-pokemon")
@pytest.fixture(scope="session")
def qwen25vl_base_huggingface_id():
# used as a base model for testing with qwen25vl lora adapter
return "Qwen/Qwen2.5-VL-3B-Instruct"
@pytest.fixture(scope="session")
def qwen25vl_lora_files():
return snapshot_download(repo_id="jeeejeee/qwen25-vl-lora-pokemon")
@@ -261,8 +273,8 @@ def run_with_both_engines_lora(request, monkeypatch):
@pytest.fixture
def reset_default_device():
"""
Some tests, such as `test_punica_ops.py`, explicitly set the
default device, which can affect subsequent tests. Adding this fixture
Some tests, such as `test_punica_ops.py`, explicitly set the
default device, which can affect subsequent tests. Adding this fixture
helps avoid this problem.
"""
original_device = torch.get_default_device()