[Model] Remove Tarsier, Tarsier2 (#47143)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
Harry Mellor
2026-06-30 13:20:33 +00:00
committed by GitHub
parent aab7af0bcb
commit 5dc36a4fa5
12 changed files with 2 additions and 872 deletions
@@ -993,31 +993,6 @@ VLM_TEST_SETTINGS = {
hf_output_post_proc=model_utils.smolvlm_trunc_hf_output,
num_logprobs=10,
),
"tarsier": VLMTestInfo(
models=["omni-research/Tarsier-7b"],
test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
prompt_formatter=lambda img_prompt: f"USER: {img_prompt} ASSISTANT:",
max_model_len=4096,
max_num_seqs=2,
auto_cls=AutoModelForImageTextToText,
patch_hf_runner=model_utils.tarsier_patch_hf_runner,
),
"tarsier2": VLMTestInfo(
models=["omni-research/Tarsier2-Recap-7b"],
test_type=(
VLMTestType.IMAGE,
VLMTestType.MULTI_IMAGE,
VLMTestType.VIDEO,
),
prompt_formatter=lambda img_prompt: f"<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\n{img_prompt}<|im_end|>\n<|im_start|>assistant\n", # noqa: E501
img_idx_to_prompt=lambda idx: "<|vision_start|><|image_pad|><|vision_end|>",
video_idx_to_prompt=lambda idx: "<|vision_start|><|video_pad|><|vision_end|>",
max_model_len=4096,
max_num_seqs=2,
auto_cls=AutoModelForImageTextToText,
image_size_factors=[(0.25,), (0.25, 0.25, 0.25), (0.25, 0.2, 0.15)],
marks=[pytest.mark.skip("Model initialization hangs")],
),
### Tensor parallel / multi-gpu broadcast tests
"chameleon-broadcast": VLMTestInfo(
models=["facebook/chameleon-7b"],
@@ -1220,18 +1220,6 @@ def qwen3_vl_patch_hf_runner(hf_model: HfRunner) -> HfRunner:
return hf_model
def tarsier_patch_hf_runner(hf_model: HfRunner) -> HfRunner:
from vllm.model_executor.models.tarsier import get_vision_encoder_info
vision_encoder_info = get_vision_encoder_info(hf_model.config)
hf_processor = hf_model.processor
if hf_processor.patch_size is None:
hf_processor.patch_size = vision_encoder_info.get_patch_size()
return hf_model
def voxtral_patch_hf_runner(hf_model: "HfRunner") -> "HfRunner":
"""Patch HfRunner for Voxtral's conversation-based processor.
-16
View File
@@ -1369,22 +1369,6 @@ _MULTIMODAL_EXAMPLE_MODELS = {
"fixie-ai/ultravox-v0_5-llama-3_2-1b",
trust_remote_code=True,
),
"TarsierForConditionalGeneration": _HfExamplesInfo("omni-research/Tarsier-7b"),
"Tarsier2ForConditionalGeneration": _HfExamplesInfo(
"omni-research/Tarsier2-Recap-7b",
hf_overrides={
"architectures": ["Tarsier2ForConditionalGeneration"],
"model_type": "tarsier2",
},
max_transformers_version="5.3",
transformers_version_reason={
"vllm": (
"Qwen2VLConfig was split into Qwen2VLConfig + "
"Qwen2VLTextConfig in transformers v5, breaking "
"attribute access (num_attention_heads, hidden_size, etc.)"
)
},
),
"VoxtralForConditionalGeneration": _HfExamplesInfo(
"mistralai/Voxtral-Mini-3B-2507",
tokenizer_mode="mistral",