A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 8, 06:15 UTC
[CI/Build] Enable tests for recent day-0 new models (#34585)
Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>
Isotr0py committed
Feb 16, 2026 at 10:17 UTC
91ac5d9bfda99745ece40f5258f17a4c0585db40
4 files changed
+6
-16
tests/models/multimodal/processing/test_common.py
+3
-4
@@ -102,13 +102,13 @@ def glmasr_patch_mm_data(mm_data: MultiModalDataDict) -> MultiModalDataDict:
102
# incorrect token ids. So we need use `add_special_tokens=False` here
103
# to leave bos_token to be added by the processor.
104
_ADD_SPECIAL_TOKENS_OVERRIDES = {
105
+ "lfm2_vl": False,
106
"nemotron_parse": False,
107
"ovis": False,
108
"ovis2_5": False,
109
"paligemma": False,
110
"ultravox": False,
111
"whisper": False,
111
- "lfm2_vl": False,
112
}
113
114
_IGNORE_MM_KEYS = {
@@ -450,6 +450,8 @@ def test_processing_correctness(
450
num_batches: int,
451
simplify_rate: float,
452
):
453
+ if model_id == "allendou/Fun-ASR-Nano-2512-vllm":
454
+ pytest.skip("Cached audio `input_features` not matched. Fix later.")
455
if model_id == "google/gemma-3n-E2B-it":
456
pytest.skip("Fix later")
457
if model_id == "OpenGVLab/InternVL2-2B":
@@ -468,9 +470,6 @@ def test_processing_correctness(
470
"correctness test as is. Let's revisit adapting this "
471
"test once more realtime models exist."
472
)
471
- if model_id == "internlm/Intern-S1-Pro":
472
- # FIXME(Isotr0py): Fix later.
473
- pytest.skip("Tokenization issue. Fix later")
473
474
_test_processing_correctness(
475
model_id,
tests/models/multimodal/processing/test_tensor_schema.py
-3
@@ -160,9 +160,6 @@ def test_model_tensor_schema(model_id: str):
160
pytest.skip(
161
"Kimi-K2.5's offline inference has issues about vision chunks. Fix later."
162
)
163
- if model_id == "internlm/Intern-S1-Pro":
164
- # FIXME(Isotr0py): Fix later.
165
- pytest.skip("Intern-S1-Pro has issue to pass the test.")
163
164
model_info = HF_EXAMPLE_MODELS.find_hf_info(model_id)
165
model_info.check_available_online(on_fail="skip")
tests/models/registry.py
-2
@@ -730,7 +730,6 @@ _MULTIMODAL_EXAMPLE_MODELS = {
730
),
731
"FunASRForConditionalGeneration": _HfExamplesInfo(
732
"allendou/Fun-ASR-Nano-2512-vllm",
733
- is_available_online=False,
733
),
734
"FunAudioChatForConditionalGeneration": _HfExamplesInfo(
735
"funaudiochat", is_available_online=False
@@ -755,7 +754,6 @@ _MULTIMODAL_EXAMPLE_MODELS = {
754
"Glm4vMoeForConditionalGeneration": _HfExamplesInfo("zai-org/GLM-4.5V"),
755
"GlmOcrForConditionalGeneration": _HfExamplesInfo(
756
"zai-org/GLM-OCR",
758
- is_available_online=False,
757
min_transformers_version="5.1.0",
758
),
759
"H2OVLChatModel": _HfExamplesInfo(
vllm/model_executor/models/interns1_pro.py
+3
-7
@@ -85,11 +85,7 @@ class InternS1ProProcessingInfo(Qwen3VLProcessingInfo):
85
return self.ctx.get_hf_config()
86
87
def get_hf_processor(self, **kwargs: object) -> AutoProcessor:
88
- return AutoProcessor.from_pretrained(
89
- self.ctx.model_config.model,
90
- trust_remote_code=True,
91
- **kwargs,
92
- )
88
+ return self.ctx.get_hf_processor(**kwargs)
89
90
91
class InternS1ProMoeMLP(nn.Module):
@@ -497,7 +493,7 @@ class InternS1ProMoeLLMForCausalLM(Qwen3MoeForCausalLM):
493
)
494
495
500
-class Qwen3VLMoeMixtureOfExperts(MixtureOfExperts):
496
+class InternS1ProMoeMixtureOfExperts(MixtureOfExperts):
497
def update_physical_experts_metadata(
498
self,
499
num_physical_experts: int,
@@ -547,7 +543,7 @@ class Qwen3VLMoeMixtureOfExperts(MixtureOfExperts):
543
dummy_inputs=Qwen3VLDummyInputsBuilder,
544
)
545
class InternS1ProForConditionalGeneration(
550
- Qwen3VLForConditionalGeneration, Qwen3VLMoeMixtureOfExperts
546
+ Qwen3VLForConditionalGeneration, InternS1ProMoeMixtureOfExperts
547
):
548
is_3d_moe_weight: bool = True
549
packed_modules_mapping = {