[Model][CI] Let more pooling models support v1 (#21747)

Signed-off-by: wang.yuqi <noooop@126.com>

tests/models/language/pooling/test_classification.py

@@ -6,14 +6,6 @@ from transformers import AutoModelForSequenceClassification
 
 from vllm.platforms import current_platform
 
-# TODO: enable when float32 is supported by V1
-# @pytest.fixture(autouse=True)
-# def v1(run_with_both_engines):
-#     # Simple autouse wrapper to run both engines for each test
-#     # This can be promoted up to conftest.py to run for every
-#     # test in a package
-#     pass
-
 
 @pytest.mark.parametrize(
     "model",

tests/models/language/pooling/test_gte.py

@@ -56,17 +56,10 @@ MODELS = [
                    enable_test=False),
 ]
 
-V1FlashAttentionImpNotSupported = [
-    "Alibaba-NLP/gte-Qwen2-1.5B-instruct", "Alibaba-NLP/gte-modernbert-base"
-]
-
 
 @pytest.mark.parametrize("model_info", MODELS)
-def test_embed_models_mteb(hf_runner, vllm_runner, model_info: EmbedModelInfo,
-                           monkeypatch) -> None:
-    if model_info.name in V1FlashAttentionImpNotSupported:
-        monkeypatch.setenv("VLLM_USE_V1", "0")
-
+def test_embed_models_mteb(hf_runner, vllm_runner,
+                           model_info: EmbedModelInfo) -> None:
     vllm_extra_kwargs: dict[str, Any] = {}
     if model_info.architecture == "GteNewModel":
         vllm_extra_kwargs["hf_overrides"] = {"architectures": ["GteNewModel"]}
@@ -77,11 +70,8 @@ def test_embed_models_mteb(hf_runner, vllm_runner, model_info: EmbedModelInfo,
 
 @pytest.mark.parametrize("model_info", MODELS)
 def test_embed_models_correctness(hf_runner, vllm_runner,
-                                  model_info: EmbedModelInfo, example_prompts,
-                                  monkeypatch) -> None:
-    if model_info.name in V1FlashAttentionImpNotSupported:
-        monkeypatch.setenv("VLLM_USE_V1", "0")
-
+                                  model_info: EmbedModelInfo,
+                                  example_prompts) -> None:
     vllm_extra_kwargs: dict[str, Any] = {}
     if model_info.architecture == "GteNewModel":
         vllm_extra_kwargs["hf_overrides"] = {"architectures": ["GteNewModel"]}

tests/models/language/pooling/test_jina.py

@@ -4,7 +4,6 @@ from functools import partial
 
 import pytest
 
-import vllm.envs as envs
 from vllm import PoolingParams
 
 from ...utils import EmbedModelInfo, RerankModelInfo
@@ -24,14 +23,6 @@ RERANK_MODELS = [
 ]
 
 
-@pytest.fixture(autouse=True)
-def v1(run_with_both_engines):
-    # Simple autouse wrapper to run both engines for each test
-    # This can be promoted up to conftest.py to run for every
-    # test in a package
-    pass
-
-
 @pytest.mark.parametrize("model_info", EMBEDDING_MODELS)
 def test_embed_models_mteb(hf_runner, vllm_runner,
                            model_info: EmbedModelInfo) -> None:
@@ -63,10 +54,6 @@ def test_embed_models_correctness(hf_runner, vllm_runner,
 @pytest.mark.parametrize("model_info", RERANK_MODELS)
 def test_rerank_models_mteb(hf_runner, vllm_runner,
                             model_info: RerankModelInfo) -> None:
-    if (model_info.architecture == "XLMRobertaForSequenceClassification"
-            and envs.VLLM_USE_V1):
-        pytest.skip("Not supported yet")
-
     mteb_test_rerank_models(hf_runner, vllm_runner, model_info)
 
 

tests/models/language/pooling/test_qwen3_reranker.py

@@ -83,9 +83,6 @@ def test_rerank_models_mteb(vllm_runner, model_info: RerankModelInfo) -> None:
         }
     }
 
-    if model_info.name == "Qwen/Qwen3-Reranker-4B":
-        vllm_extra_kwargs["max_num_seqs"] = 1
-
     mteb_test_rerank_models(Qwen3RerankerHfRunner, vllm_runner, model_info,
                             vllm_extra_kwargs)
 
@@ -106,9 +103,6 @@ def test_rerank_models_mteb_tp(vllm_runner,
         "tensor_parallel_size": 2,
     }
 
-    if model_info.name == "Qwen/Qwen3-Reranker-4B":
-        vllm_extra_kwargs["max_num_seqs"] = 1
-
     mteb_test_rerank_models(Qwen3RerankerHfRunner,
                             vllm_runner,
                             model_info,

vllm/config.py

@@ -776,6 +776,9 @@ class ModelConfig:
             raise ValueError(
                 "`override_neuron_config` is only supported on Neuron.")
 
+        # Avoid running try_verify_and_update_config multiple times
+        self.config_updated = False
+
         self._verify_quantization()
         self._verify_cuda_graph()
         self._verify_bnb_config()
@@ -4914,6 +4917,11 @@ class VllmConfig:
         if self.model_config is None:
             return
 
+        # Avoid running try_verify_and_update_config multiple times
+        if getattr(self.model_config, "config_updated", False):
+            return
+        self.model_config.config_updated = True
+
         architecture = self.model_config.architecture
         if architecture is None:
             return

vllm/model_executor/models/bert_with_rope.py

@@ -8,7 +8,6 @@ from torch import nn
 from transformers import PretrainedConfig
 
 from vllm.attention import Attention, AttentionType
-from vllm.compilation.decorators import support_torch_compile
 from vllm.config import CacheConfig, VllmConfig
 from vllm.distributed import (divide, get_tensor_model_parallel_rank,
                               get_tensor_model_parallel_world_size,
@@ -26,7 +25,6 @@ from vllm.model_executor.layers.rotary_embedding import get_rope
 from vllm.model_executor.layers.vocab_parallel_embedding import (
     VocabParallelEmbedding)
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
-from vllm.model_executor.models import SupportsV0Only
 from vllm.model_executor.models.interfaces import SupportsQuant
 from vllm.model_executor.models.utils import WeightsMapper
 from vllm.model_executor.utils import set_weight_attrs
@@ -360,7 +358,6 @@ class BertWithRopeBlock(nn.Module):
         return hidden_states
 
 
-@support_torch_compile
 class BertWithRopeEncoder(nn.Module):
 
     def __init__(self,
@@ -394,7 +391,7 @@ class BertWithRopeEncoder(nn.Module):
         return hidden_states
 
 
-class BertWithRope(nn.Module, SupportsV0Only, SupportsQuant):
+class BertWithRope(nn.Module, SupportsQuant):
     hf_to_vllm_mapper = WeightsMapper(orig_to_new_prefix={"model.": ""})
 
     def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):

vllm/model_executor/models/config.py

@@ -93,7 +93,7 @@ class NomicBertModelConfig(VerifyAndUpdateConfig):
         config.num_hidden_layers = config.n_layer
 
         head_dim = config.hidden_size // config.num_attention_heads
-        rotary_emb_dim = head_dim * config.rotary_emb_fraction
+        rotary_emb_dim = int(head_dim * config.rotary_emb_fraction)
         max_trained_positions = getattr(config, "max_trained_positions", 2048)
         config.rotary_kwargs = {
             "head_size": head_dim,

vllm/model_executor/models/modernbert.py

@@ -8,7 +8,6 @@ from torch import nn
 from transformers import ModernBertConfig
 
 from vllm.attention import Attention, AttentionType
-from vllm.compilation.decorators import support_torch_compile
 from vllm.config import VllmConfig
 from vllm.distributed import get_tensor_model_parallel_world_size
 from vllm.model_executor.layers.linear import (QKVParallelLinear,
@@ -200,7 +199,6 @@ class ModernBertEncoderLayer(nn.Module):
         return hidden_states
 
 
-@support_torch_compile
 class ModernBertModel(nn.Module):
     hf_to_vllm_mapper = WeightsMapper(
         orig_to_new_prefix={"layers.": "encoder_layer.layers."})