[New model][Multimodal] Add DeepSeek-V4-Flash-Vision-Exp support (#54566)

Signed-off-by: Isotr0py <Isotr0py@outlook.com>
This commit is contained in:
Isotr0py
2026-09-02 11:09:09 -07:00
committed by GitHub
parent 488e6fd53c
commit 1356635d83
37 changed files with 2918 additions and 152 deletions
@@ -164,6 +164,11 @@ def test_model_tensor_schema(model_id: str):
"Kimi-K2.5's offline inference has issues about vision chunks. Fix later."
)
if model_id == "deepseek-ai/DeepSeek-V4-Flash-Vision-Exp" and not (
current_platform.is_cuda()
):
pytest.skip("Deepseek V4 is only supported on CUDA")
model_info = HF_EXAMPLE_MODELS.find_hf_info(model_id)
model_info.check_available_online(on_fail="skip")
model_info.check_transformers_version(
+3
View File
@@ -833,6 +833,9 @@ _MULTIMODAL_EXAMPLE_MODELS = {
"DeepseekOCR2ForCausalLM": _HfExamplesInfo(
"deepseek-ai/DeepSeek-OCR-2",
),
"DeepseekV4ForConditionalGeneration": _HfExamplesInfo(
"deepseek-ai/DeepSeek-V4-Flash-Vision-Exp",
),
"Dots3NoteForCausalLM": _HfExamplesInfo(
"dots-studio/dots3-note-prev",
is_available_online=False,
+6
View File
@@ -132,6 +132,12 @@ def can_initialize(
f"capability {capability.major}.{capability.minor}"
)
if model_arch == "DeepseekV4ForConditionalGeneration":
from vllm.platforms import current_platform
if not current_platform.is_cuda():
pytest.skip("Deepseek V4 is only supported on CUDA")
with (
patch.object(V1EngineCore, "_initialize_kv_caches", _initialize_kv_caches_v1),
monkeypatch.context() as m,
+5
View File
@@ -64,6 +64,11 @@ def test_registry_imports(model_arch):
):
pytest.skip("HY V4 is only supported on CUDA")
if model_arch == "DeepseekV4ForConditionalGeneration" and (
current_platform.is_cuda()
):
pytest.skip("Deepseek V4 is only supported on CUDA")
# Ensure all model classes can be imported successfully
model_cls = ModelRegistry._try_load_model_cls(model_arch)
assert model_cls is not None