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

[Bugfix] handle alignment of encoder_seq_lens in mllama.py (#14784)

Browse Source 
Signed-off-by: Travis Johnson <tsjohnso@us.ibm.com>
This commit is contained in:
Travis Johnson
2025-04-11 13:59:50 -06:00
committed by GitHub
parent 5285589f37
commit 71b9cde010
2 changed files with 82 additions and 22 deletions
Download Patch File Download Diff File Expand all files Collapse all files

59
tests/models/encoder_decoder/vision_language/test_mllama.py
View File

@@ -209,14 +209,15 @@ def _run_test(
# will hurt multiprocessing backend with fork method (the default method).
# max_model_len should be greater than image_feature_size
with vllm_runner(model,
dtype=dtype,
max_model_len=8192,
max_num_seqs=3,
tensor_parallel_size=tensor_parallel_size,
distributed_executor_backend=distributed_executor_backend,
limit_mm_per_prompt={"image": _LIMIT_IMAGE_PER_PROMPT
}) as vllm_model:
with vllm_runner(
model,
dtype=dtype,
max_model_len=19212, # 3 max size images
max_num_seqs=3,
tensor_parallel_size=tensor_parallel_size,
distributed_executor_backend=distributed_executor_backend,
limit_mm_per_prompt={"image":
_LIMIT_IMAGE_PER_PROMPT}) as vllm_model:
vllm_outputs_per_image = [
vllm_model.generate_greedy_logprobs(prompts,
max_tokens,
@@ -507,7 +508,7 @@ def test_regression(vllm_runner, image_assets, model, dtype, max_tokens,
model,
dtype=dtype,
max_model_len=8192,
max_num_seqs=2,
max_num_seqs=4,
tensor_parallel_size=1,
limit_mm_per_prompt={"image":
_LIMIT_IMAGE_PER_PROMPT}) as vllm_model:
@@ -552,6 +553,23 @@ def test_regression(vllm_runner, image_assets, model, dtype, max_tokens,
num_logprobs,
images=images)
# Mixed batch with text and images with different numbers of tiles
prompts = [
"<|begin_of_text|>Hello!",
"<|begin_of_text|>Some text before.<|image|>What is in the image?", # noqa: E501
"<|begin_of_text|>Some text before.<|image|>What is in the image?", # noqa: E501
]
images = [
None,
[stop_sign],
# smaller image must be 2nd for the repro
[stop_sign.resize((448, 448))],
]
vllm_model.generate_greedy_logprobs(prompts,
max_tokens,
num_logprobs,
images=images)
class DummyModel:
image_token_id = MLLAMA_IMAGE_TOKEN_ID
@@ -674,3 +692,26 @@ def test_get_full_text_row_masked_out_mask(input_indices) -> None:
f"full_text_row_masked_out_mask[{idx}] must be " \
f"'{must_be_masked}' "
idx += 1
@pytest.mark.core_model
@pytest.mark.parametrize("encoder_seq_lens, num_tiles, expected", [
([6404], [[4]], [6404]),
([0, 6404], [[4]], [6404]),
([0, 1601, 8005], [[1], [4, 1]], [1601, 8005]),
([0, 19212, 0, 3202], [[4, 4, 4], [2]], [19212, 3202]),
])
def test_parse_and_validate_encoder_lens(encoder_seq_lens, num_tiles,
expected) -> None:
dummy = DummyModel()
num_tokens_per_tile = 1601
actual_encoder_seq_lens = MllamaForConditionalGeneration \
._get_and_validate_encoder_lens(
dummy,
encoder_seq_lens,
num_tiles,
num_tokens_per_tile,
)
assert actual_encoder_seq_lens == expected, \
f"Expected {expected} but got {actual_encoder_seq_lens}"