biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Bugfix] Proper input validation for multi-modal encoder-decoder models (#16156)

Browse Source 
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2025-04-09 00:45:21 +08:00
committed by GitHub
parent dc96fd54c6
commit 4ebc0b9640
10 changed files with 113 additions and 62 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
tests/engine/test_short_mm_context.py
View File

@@ -18,7 +18,8 @@ models = ["llava-hf/llava-1.5-7b-hf"]
def test_context_length_too_short(vllm_runner, image_assets, model):
images = [asset.pil_image for asset in image_assets]
with pytest.raises(ValueError, match="too long to fit into the model"):
with pytest.raises(ValueError,
match="longer than the maximum model length"):
vllm_model = vllm_runner(
model,
max_model_len=128, # LLaVA has a feature size of 576

2
tests/entrypoints/llm/test_prompt_validation.py
View File

@@ -15,7 +15,7 @@ def v1(run_with_both_engines):
def test_empty_prompt():
llm = LLM(model="openai-community/gpt2", enforce_eager=True)
with pytest.raises(ValueError, match='Prompt cannot be empty'):
with pytest.raises(ValueError, match='decoder prompt cannot be empty'):
llm.generate([""])

2
tests/entrypoints/openai/test_prompt_validation.py
View File

@@ -17,7 +17,7 @@ async def test_empty_prompt():
client = remote_server.get_async_client()
with pytest.raises(openai.BadRequestError,
match=re.compile('.+Prompt cannot be empty.+')):
match="decoder prompt cannot be empty"):
await client.completions.create(model=model_name,
prompt="",
max_tokens=5,

8
tests/models/encoder_decoder/vision_language/test_mllama.py
View File

@@ -211,7 +211,7 @@ def _run_test(
# max_model_len should be greater than image_feature_size
with vllm_runner(model,
dtype=dtype,
max_model_len=4096,
max_model_len=8192,
max_num_seqs=3,
tensor_parallel_size=tensor_parallel_size,
distributed_executor_backend=distributed_executor_backend,
@@ -422,7 +422,7 @@ def test_bnb_regression(
llm = LLM(
model=model,
dtype=dtype,
max_model_len=4096,
max_model_len=8192,
max_num_seqs=2,
quantization="bitsandbytes",
)
@@ -475,7 +475,7 @@ def test_explicit_implicit_prompt(
llm = LLM(
model=model,
dtype=dtype,
max_model_len=4096,
max_model_len=8192,
max_num_seqs=2,
tensor_parallel_size=1,
)
@@ -506,7 +506,7 @@ def test_regression(vllm_runner, image_assets, model, dtype, max_tokens,
with global_force_attn_backend_context_manager(attn_backend), vllm_runner(
model,
dtype=dtype,
max_model_len=4096,
max_model_len=8192,
max_num_seqs=2,
tensor_parallel_size=1,
limit_mm_per_prompt={"image":