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

[Model] Broadcast Ovis2 implementation to fit Ovis1.6 (#17861)

Browse Source 
Signed-off-by: Isotr0py <2037008807@qq.com>
This commit is contained in:
Isotr0py
2025-05-12 08:56:30 +08:00
committed by GitHub
parent 7de18d541b
commit 021c16c7ca
16 changed files with 330 additions and 212 deletions
Download Patch File Download Diff File Expand all files Collapse all files

21
examples/offline_inference/vision_language.py
View File

@@ -725,8 +725,8 @@ def run_nvlm_d(questions: list[str], modality: str) -> ModelRequestData:
)
# Ovis2
def run_ovis2(questions: list[str], modality: str) -> ModelRequestData:
# Ovis
def run_ovis(questions: list[str], modality: str) -> ModelRequestData:
assert modality == "image"
model_name = "AIDC-AI/Ovis2-1B"
@@ -737,15 +737,18 @@ def run_ovis2(questions: list[str], modality: str) -> ModelRequestData:
max_num_seqs=2,
trust_remote_code=True,
dtype="half",
hf_overrides={"architectures": ["Ovis2ForConditionalGeneration"]},
limit_mm_per_prompt={modality: 1},
)
placeholder = "<image>\n"
prompts = [("<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n"
f"<|im_start|>user\n{placeholder}"
f"{question}<|im_end|>\n"
"<|im_start|>assistant\n") for question in questions]
tokenizer = AutoTokenizer.from_pretrained(model_name,
trust_remote_code=True)
messages = [[{
'role': 'user',
'content': f"<image>\n{question}"
}] for question in questions]
prompts = tokenizer.apply_chat_template(messages,
tokenize=False,
add_generation_prompt=True)
return ModelRequestData(
engine_args=engine_args,
@@ -1069,7 +1072,7 @@ model_example_map = {
"llama4": run_llama4,
"molmo": run_molmo,
"NVLM_D": run_nvlm_d,
"ovis2": run_ovis2,
"ovis": run_ovis,
"paligemma": run_paligemma,
"paligemma2": run_paligemma2,
"phi3_v": run_phi3v,

22
examples/offline_inference/vision_language_multi_image.py
View File

@@ -436,8 +436,8 @@ def load_nvlm_d(question: str, image_urls: list[str]) -> ModelRequestData:
)
# Ovis2
def load_ovis2(question: str, image_urls: list[str]) -> ModelRequestData:
# Ovis
def load_ovis(question: str, image_urls: list[str]) -> ModelRequestData:
model_name = "AIDC-AI/Ovis2-1B"
engine_args = EngineArgs(
@@ -447,15 +447,17 @@ def load_ovis2(question: str, image_urls: list[str]) -> ModelRequestData:
trust_remote_code=True,
dtype="half",
limit_mm_per_prompt={"image": len(image_urls)},
hf_overrides={"architectures": ["Ovis2ForConditionalGeneration"]},
)
placeholder = '\n'.join(
[f'Image {i+1}: <image>' for i in range(len(image_urls))]) + '\n'
prompt = ("<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n"
f"<|im_start|>user\n{placeholder}"
f"{question}<|im_end|>\n"
"<|im_start|>assistant\n")
placeholders = "\n".join(f"Image-{i}: <image>\n"
for i, _ in enumerate(image_urls, start=1))
messages = [{'role': 'user', 'content': f"{placeholders}\n{question}"}]
tokenizer = AutoTokenizer.from_pretrained(model_name,
trust_remote_code=True)
prompt = tokenizer.apply_chat_template(messages,
tokenize=False,
add_generation_prompt=True)
return ModelRequestData(
engine_args=engine_args,
@@ -713,7 +715,7 @@ model_example_map = {
"mistral3": load_mistral3,
"mllama": load_mllama,
"NVLM_D": load_nvlm_d,
"ovis2": load_ovis2,
"ovis": load_ovis,
"phi3_v": load_phi3v,
"phi4_mm": load_phi4mm,
"pixtral_hf": load_pixtral_hf,