[Model] Add Qwen2 PRM model support

tests/models/embedding/language/test_embedding.py

@@ -17,14 +17,15 @@
                      marks=[pytest.mark.core_model, pytest.mark.cpu_model]),
         pytest.param("sentence-transformers/all-MiniLM-L12-v2"),
         pytest.param("intfloat/multilingual-e5-large"),
-        # [Encoder-decoder]
-        pytest.param("intfloat/e5-mistral-7b-instruct",
-                     marks=[pytest.mark.core_model, pytest.mark.cpu_model]),
+        # [Decoder-only]
         pytest.param("BAAI/bge-multilingual-gemma2",
                      marks=[pytest.mark.core_model]),
-        pytest.param("ssmits/Qwen2-7B-Instruct-embed-base"),
+        pytest.param("intfloat/e5-mistral-7b-instruct",
+                     marks=[pytest.mark.core_model, pytest.mark.cpu_model]),
         pytest.param("Alibaba-NLP/gte-Qwen2-1.5B-instruct"),
         pytest.param("Alibaba-NLP/gte-Qwen2-7B-instruct"),
+        pytest.param("ssmits/Qwen2-7B-Instruct-embed-base"),
+        # [Encoder-decoder]
         pytest.param("sentence-transformers/stsb-roberta-base-v2"),
     ],
 )

tests/models/registry.py

@@ -155,6 +155,7 @@ class _HfExamplesInfo:
     "MistralModel": _HfExamplesInfo("intfloat/e5-mistral-7b-instruct"),
     "Qwen2Model": _HfExamplesInfo("ssmits/Qwen2-7B-Instruct-embed-base"),
     "Qwen2ForRewardModel": _HfExamplesInfo("Qwen/Qwen2.5-Math-RM-72B"),
+    "Qwen2ForProcessRewardModel": _HfExamplesInfo("Qwen/Qwen2.5-Math-PRM-7B"),
     "Qwen2ForSequenceClassification": _HfExamplesInfo("jason9693/Qwen2.5-1.5B-apeach"),  # noqa: E501
     "RobertaModel": _HfExamplesInfo("sentence-transformers/stsb-roberta-base-v2"),  # noqa: E501
     "RobertaForMaskedLM": _HfExamplesInfo("sentence-transformers/all-roberta-large-v1"),  # noqa: E501

vllm/model_executor/models/qwen2_rm.py

@@ -12,7 +12,7 @@
 from vllm.config import VllmConfig
 from vllm.model_executor.layers.linear import (ColumnParallelLinear,
                                                RowParallelLinear)
-from vllm.model_executor.layers.pooler import Pooler, PoolingType
+from vllm.model_executor.layers.pooler import Pooler, PoolingType, SimplePooler
 from vllm.model_executor.pooling_metadata import PoolingMetadata
 from vllm.sequence import IntermediateTensors, PoolerOutput
 
@@ -32,7 +32,7 @@ def forward(self, input):
         return self.activation(input)
 
 
-class Qwen2ForRewardModel(nn.Module, SupportsLoRA, SupportsPP):
+class Qwen2RewardBaseModel(nn.Module, SupportsLoRA, SupportsPP):
     packed_modules_mapping = {
         "qkv_proj": [
             "q_proj",
@@ -60,7 +60,6 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
         config = vllm_config.model_config.hf_config
         quant_config = vllm_config.quant_config
         lora_config = vllm_config.lora_config
-        pooler_config = vllm_config.model_config.pooler_config
 
         self.config = config
         self.lora_config = lora_config
@@ -74,14 +73,11 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
                                  config.hidden_size,
                                  quant_config=quant_config),
             ReLU(),
-            RowParallelLinear(config.hidden_size, 1,
+            RowParallelLinear(config.hidden_size,
+                              config.num_labels,
                               quant_config=quant_config),
         )
-        self._pooler = Pooler.from_config_with_defaults(
-            pooler_config,
-            pooling_type=PoolingType.ALL,
-            normalize=False,
-            softmax=False)
+        self._pooler: SimplePooler
         self.make_empty_intermediate_tensors = (
             self.model.make_empty_intermediate_tensors)
 
@@ -115,3 +111,31 @@ def load_weights(self, weights: Iterable[Tuple[str,
         loader = AutoWeightsLoader(self,
                                    ignore_unexpected_prefixes=["lm_head."])
         return loader.load_weights(weights)
+
+
+class Qwen2ForRewardModel(Qwen2RewardBaseModel):
+
+    def __init__(self, *, vllm_config, prefix=""):
+        vllm_config.model_config.hf_config.num_labels = 1
+        super().__init__(vllm_config=vllm_config, prefix=prefix)
+        pooler_config = vllm_config.model_config.pooler_config
+        self._pooler = Pooler.from_config_with_defaults(
+            pooler_config,
+            pooling_type=PoolingType.ALL,
+            normalize=False,
+            softmax=False)
+
+
+class Qwen2ForProcessRewardModel(Qwen2RewardBaseModel):
+
+    def __init__(self, *, vllm_config, prefix=""):
+        vllm_config.model_config.hf_config.num_labels = 2
+        super().__init__(vllm_config=vllm_config, prefix=prefix)
+        pooler_config = vllm_config.model_config.pooler_config
+        self._pooler = Pooler.from_config_with_defaults(
+            pooler_config,
+            pooling_type=PoolingType.STEP,
+            normalize=False,
+            softmax=True,
+            step_tag_id=151651,
+        )

vllm/model_executor/models/registry.py

@@ -127,6 +127,7 @@
     "Qwen2Model": ("qwen2", "Qwen2EmbeddingModel"),
     "Qwen2ForCausalLM": ("qwen2", "Qwen2ForCausalLM"),
     "Qwen2ForRewardModel": ("qwen2_rm", "Qwen2ForRewardModel"),
+    "Qwen2ForProcessRewardModel": ("qwen2_rm", "Qwen2ForProcessRewardModel"),
     "TeleChat2ForCausalLM": ("telechat2", "TeleChat2ForCausalLM"),
     # [Multimodal]
     "LlavaNextForConditionalGeneration": ("llava_next", "LlavaNextForConditionalGeneration"),  # noqa: E501