[Model] Add support for deepseek-vl2-tiny model (#12068)

Signed-off-by: Isotr0py <2037008807@qq.com>
This commit is contained in:
Isotr0py
2025-01-16 17:14:48 +00:00
committed by GitHub
parent 5fd24ec02e
commit 62b06ba23d
6 changed files with 22 additions and 21 deletions
@@ -9,6 +9,7 @@ from typing import Type
import pytest
from transformers import AutoModelForVision2Seq
from transformers import __version__ as TRANSFORMERS_VERSION
from transformers.utils import is_flash_attn_2_available
from vllm.platforms import current_platform
@@ -189,30 +190,27 @@ VLM_TEST_SETTINGS = {
dtype="bfloat16",
),
"deepseek_vl_v2": VLMTestInfo(
models=["deepseek-ai/deepseek-vl2-small"],
models=["deepseek-ai/deepseek-vl2-tiny"],
test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
dtype="bfloat16",
prompt_formatter=lambda img_prompt: f"<|User|>: {img_prompt}\n\n<|Assistant|>: ", # noqa: E501
max_model_len=4096,
max_num_seqs=2,
single_image_prompts=IMAGE_ASSETS.prompts({
"stop_sign": "<image>\nWhat's the color of the stop sign and car?",
"cherry_blossom": "<image>\nWhat's the color of the tower?",
"stop_sign": "<image>\nWhat's the content in the center of the image?", # noqa: E501
"cherry_blossom": "<image>\nPlease infer the season with reason in details.", # noqa: E501
}),
multi_image_prompt="image_1:<image>\nimage_2:<image>\nDescribe the two images shortly.", # noqa: E501
multi_image_prompt="image_1:<image>\nimage_2:<image>\nWhich image can we see the car and the tower?", # noqa: E501
vllm_runner_kwargs={"hf_overrides": {"architectures": ["DeepseekVLV2ForCausalLM"]}}, # noqa: E501
image_size_factors=[(0.10, 0.15)],
patch_hf_runner=model_utils.deepseekvl2_patch_hf_runner,
postprocess_inputs=model_utils.cast_dtype_post_processor("images"),
hf_output_post_proc=model_utils.deepseekvl2_trunc_hf_output,
stop_str=["<|end▁of▁sentence|>", "<|begin▁of▁sentence|>"], # noqa: E501
num_logprobs=5,
image_size_factors=[(), (1.0, ), (1.0, 1.0, 1.0), (0.1, 0.5, 1.0)],
marks=[
pytest.mark.skipif(
not is_flash_attn_2_available(),
reason="Model needs flash-attn for numeric convergence.",
),
large_gpu_mark(min_gb=48),
TRANSFORMERS_VERSION >= "4.48.0",
reason="HF model is not compatible with transformers>=4.48.0",
)
],
),
"fuyu": VLMTestInfo(
+1 -2
View File
@@ -181,8 +181,7 @@ _MULTIMODAL_EXAMPLE_MODELS = {
trust_remote_code=True),
"ChatGLMForConditionalGeneration": _HfExamplesInfo("chatglm2-6b",
is_available_online=False),
# TODO(Isotr0py): Use deepseek-vl2-tiny for test after it's supported
"DeepseekVLV2ForCausalLM": _HfExamplesInfo("deepseek-ai/deepseek-vl2-small"), # noqa: E501
"DeepseekVLV2ForCausalLM": _HfExamplesInfo("deepseek-ai/deepseek-vl2-tiny"), # noqa: E501
"FuyuForCausalLM": _HfExamplesInfo("adept/fuyu-8b"),
"H2OVLChatModel": _HfExamplesInfo("h2oai/h2ovl-mississippi-800m"),
"InternVLChatModel": _HfExamplesInfo("OpenGVLab/InternVL2-1B",