[Model] Add LLaVA-OneVision-2 (LlavaOnevision2ForConditionalGeneration) (#44785)

Signed-off-by: chengzheng345 <209475443+chengzheng345@users.noreply.github.com>
Co-authored-by: chengzheng345 <209475443+chengzheng345@users.noreply.github.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
This commit is contained in:
chengzheng345
2026-07-02 16:41:49 +08:00
committed by GitHub
co-authored by chengzheng345 mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent 3af8789559
commit b0b8a286dd
6 changed files with 2593 additions and 0 deletions
+11
View File
@@ -1048,6 +1048,17 @@ _MULTIMODAL_EXAMPLE_MODELS = {
"LlavaNextVideoForConditionalGeneration": _HfExamplesInfo(
"llava-hf/LLaVA-NeXT-Video-7B-hf"
),
"LlavaOnevision2ForConditionalGeneration": _HfExamplesInfo(
"lmms-lab-encoder/LLaVA-OneVision-2-8B-Instruct",
trust_remote_code=True,
# Keep the init/schema harness eager (these tests never run forward).
# NOTE: a separate H200-only hang exists in the core FA3 attention
# construction (after `FlashAttentionImpl.__init__` logs the FA3
# version, while building the Qwen3 backbone). enforce_eager does NOT
# gate that path, so this flag is hygiene, not the fix -- tracked with
# the maintainer.
enforce_eager=True,
),
"LlavaOnevisionForConditionalGeneration": _HfExamplesInfo(
"llava-hf/llava-onevision-qwen2-0.5b-ov-hf"
),