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

[V1][VLM] V1 support for selected single-image models. (#11632)

Browse Source 
Signed-off-by: Roger Wang <ywang@roblox.com>
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
Signed-off-by: Isotr0py <2037008807@qq.com>
Co-authored-by: DarkLight1337 <tlleungac@connect.ust.hk>
Co-authored-by: Isotr0py <2037008807@qq.com>
This commit is contained in:
Roger Wang
2024-12-31 13:17:22 -08:00
committed by GitHub
parent 8c3230d8c1
commit e7c7c5e822
19 changed files with 575 additions and 621 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
tests/models/decoder_only/vision_language/test_models.py
View File

@@ -140,10 +140,7 @@ VLM_TEST_SETTINGS = {
"aria": VLMTestInfo(
models=["rhymes-ai/Aria"],
tokenizer_mode="slow",
test_type=(
VLMTestType.IMAGE,
VLMTestType.MULTI_IMAGE,
),
test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
dtype="bfloat16",
prompt_formatter=lambda img_prompt: f"<|im_start|>user\n{img_prompt}<|im_end|>\n<|im_start|>assistant\n ", # noqa: E501
img_idx_to_prompt=lambda idx: "<fim_prefix><|img|><fim_suffix>\n",
@@ -179,6 +176,7 @@ VLM_TEST_SETTINGS = {
test_type=VLMTestType.IMAGE,
prompt_formatter=lambda img_prompt: f"USER: {img_prompt}\nASSISTANT:",
max_model_len=4096,
max_num_seqs=2,
auto_cls=AutoModelForVision2Seq,
postprocess_inputs=model_utils.cast_dtype_post_processor(
"pixel_values"
@@ -201,7 +199,6 @@ VLM_TEST_SETTINGS = {
vllm_output_post_proc=model_utils.fuyu_vllm_to_hf_output,
num_logprobs=10,
image_size_factors=[(), (0.25,), (0.25, 0.25, 0.25), (0.25, 0.2, 0.15)],
marks=[large_gpu_mark(min_gb=48)],
),
"glm4": VLMTestInfo(
models=["THUDM/glm-4v-9b"],

29
tests/multimodal/test_processing.py
View File

@@ -528,7 +528,7 @@ def _rand_audio(
def _test_processing_cache_correctness(
model_id: str,
modalities: set[str],
modalities: dict[str, bool],
hit_rate: float,
num_batches: int,
simplify_rate: float,
@@ -583,9 +583,8 @@ def _test_processing_cache_correctness(
partial(_rand_audio, rng, min_len=256, max_len=512, sr=16000),
}
input_max_count = {
"image": 3,
"video": 3,
"audio": 3,
modality: 3 if supports_multi else 1
for modality, supports_multi in modalities.items()
}
for batch_idx in range(num_batches):
@@ -624,12 +623,16 @@ def _test_processing_cache_correctness(
# yapf: disable
@pytest.mark.parametrize(("model_id", "modalities"), [
("llava-hf/llava-1.5-7b-hf", {"image"}),
("TIGER-Lab/Mantis-8B-siglip-llama3", {"image"}),
("mistral-community/pixtral-12b", {"image"}),
("Qwen/Qwen2-VL-2B-Instruct", {"image", "video"}),
("Qwen/Qwen2-Audio-7B-Instruct", {"audio"}),
("fixie-ai/ultravox-v0_3", {"audio"}),
("rhymes-ai/Aria", {"image": True}),
("Salesforce/blip2-opt-2.7b", {"image": False}),
("facebook/chameleon-7b", {"image": True}),
("adept/fuyu-8b", {"image": False}),
("llava-hf/llava-1.5-7b-hf", {"image": True}),
("TIGER-Lab/Mantis-8B-siglip-llama3", {"image": True}),
("mistral-community/pixtral-12b", {"image": True}),
("Qwen/Qwen2-VL-2B-Instruct", {"image": True, "video": True}),
("Qwen/Qwen2-Audio-7B-Instruct", {"audio": True}),
("fixie-ai/ultravox-v0_3", {"audio": True}),
])
@pytest.mark.parametrize("hit_rate", [0.3, 0.5, 1.0])
@pytest.mark.parametrize("num_batches", [32])
@@ -637,7 +640,7 @@ def _test_processing_cache_correctness(
# yapf: enable
def test_processing_cache_correctness(
model_id: str,
modalities: set[str],
modalities: dict[str, bool],
hit_rate: float,
num_batches: int,
simplify_rate: float,
@@ -653,7 +656,7 @@ def test_processing_cache_correctness(
# yapf: disable
@pytest.mark.parametrize(("model_id", "modalities"), [
("microsoft/Phi-3-vision-128k-instruct", {"image"}),
("microsoft/Phi-3-vision-128k-instruct", {"image": True}),
])
@pytest.mark.parametrize("hit_rate", [0.3, 0.5, 1.0])
@pytest.mark.parametrize("num_batches", [32])
@@ -661,7 +664,7 @@ def test_processing_cache_correctness(
# yapf: enable
def test_processing_cache_correctness_phi3v(
model_id: str,
modalities: set[str],
modalities: dict[str, bool],
hit_rate: float,
num_batches: int,
simplify_rate: float,