Skip to content

vllm.model_executor.models.qwen2_rm

Inference-only Qwen2-RM model compatible with HuggingFace weights.

Qwen2ForProcessRewardModel

Bases: Qwen2RewardBaseModel

Source code in vllm/model_executor/models/qwen2_rm.py 
114
115
116
117
118
119
120
121
122
123
124
125
@default_pooling_type("STEP")
class Qwen2ForProcessRewardModel(Qwen2RewardBaseModel):
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        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
        assert pooler_config is not None

        self.pooler = DispatchPooler(
            {"token_classify": Pooler.for_token_classify(pooler_config)}
        )

pooler instance-attribute 

pooler = DispatchPooler(
    {"token_classify": for_token_classify(pooler_config)}
)

__init__ 

__init__(*, vllm_config: VllmConfig, prefix: str = '')
Source code in vllm/model_executor/models/qwen2_rm.py 
116
117
118
119
120
121
122
123
124
125
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
    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
    assert pooler_config is not None

    self.pooler = DispatchPooler(
        {"token_classify": Pooler.for_token_classify(pooler_config)}
    )

Qwen2ForRewardModel

Bases: Qwen2RewardBaseModel

Source code in vllm/model_executor/models/qwen2_rm.py 
100
101
102
103
104
105
106
107
108
109
110
111
@default_pooling_type("ALL")
class Qwen2ForRewardModel(Qwen2RewardBaseModel):
    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        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
        assert pooler_config is not None

        self.pooler = DispatchPooler(
            {"token_classify": Pooler.for_token_classify(pooler_config)}
        )

pooler instance-attribute 

pooler = DispatchPooler(
    {"token_classify": for_token_classify(pooler_config)}
)

__init__ 

__init__(*, vllm_config: VllmConfig, prefix: str = '')
Source code in vllm/model_executor/models/qwen2_rm.py 
102
103
104
105
106
107
108
109
110
111
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
    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
    assert pooler_config is not None

    self.pooler = DispatchPooler(
        {"token_classify": Pooler.for_token_classify(pooler_config)}
    )

Qwen2RewardBaseModel

Bases: Module, SupportsLoRA, SupportsPP

Source code in vllm/model_executor/models/qwen2_rm.py 
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
class Qwen2RewardBaseModel(nn.Module, SupportsLoRA, SupportsPP):
    is_pooling_model = True
    pooler: Pooler

    packed_modules_mapping = {
        "qkv_proj": [
            "q_proj",
            "k_proj",
            "v_proj",
        ],
        "gate_up_proj": [
            "gate_proj",
            "up_proj",
        ],
    }

    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        super().__init__()
        config = vllm_config.model_config.hf_config
        quant_config = vllm_config.quant_config
        lora_config = vllm_config.lora_config

        self.config = config
        self.lora_config = lora_config

        self.quant_config = quant_config
        self.model = Qwen2Model(
            vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
        )
        self.head_dtype = vllm_config.model_config.head_dtype

        self.score = nn.Sequential(
            ColumnParallelLinear(
                config.hidden_size,
                config.hidden_size,
                quant_config=quant_config,
                params_dtype=self.head_dtype,
                return_bias=False,
            ),
            nn.ReLU(),
            RowParallelLinear(
                config.hidden_size,
                config.num_labels,
                params_dtype=self.head_dtype,
                quant_config=quant_config,
                return_bias=False,
            ),
        )
        self.make_empty_intermediate_tensors = (
            self.model.make_empty_intermediate_tensors
        )

    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
        return self.model.get_input_embeddings(input_ids)

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        intermediate_tensors: IntermediateTensors | None = None,
        inputs_embeds: torch.Tensor | None = None,
    ) -> torch.Tensor | IntermediateTensors:
        hidden_states = self.model(
            input_ids, positions, intermediate_tensors, inputs_embeds
        )
        hidden_states = hidden_states.to(self.head_dtype)
        logits = self.score(hidden_states)
        return logits

    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
        loader = AutoWeightsLoader(self, ignore_unexpected_prefixes=["lm_head."])
        return loader.load_weights(weights)

config instance-attribute 

config = config

head_dtype instance-attribute 

head_dtype = head_dtype

is_pooling_model class-attribute instance-attribute 

is_pooling_model = True

lora_config instance-attribute 

lora_config = lora_config

make_empty_intermediate_tensors instance-attribute 

make_empty_intermediate_tensors = (
    make_empty_intermediate_tensors
)

model instance-attribute 

model = Qwen2Model(
    vllm_config=vllm_config,
    prefix=maybe_prefix(prefix, "model"),
)

packed_modules_mapping class-attribute instance-attribute 

packed_modules_mapping = {
    "qkv_proj": ["q_proj", "k_proj", "v_proj"],
    "gate_up_proj": ["gate_proj", "up_proj"],
}

pooler instance-attribute 

pooler: Pooler

quant_config instance-attribute 

quant_config = quant_config

score instance-attribute 

score = Sequential(
    ColumnParallelLinear(
        hidden_size,
        hidden_size,
        quant_config=quant_config,
        params_dtype=head_dtype,
        return_bias=False,
    ),
    ReLU(),
    RowParallelLinear(
        hidden_size,
        num_labels,
        params_dtype=head_dtype,
        quant_config=quant_config,
        return_bias=False,
    ),
)

__init__ 

__init__(*, vllm_config: VllmConfig, prefix: str = '')
Source code in vllm/model_executor/models/qwen2_rm.py 
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
    super().__init__()
    config = vllm_config.model_config.hf_config
    quant_config = vllm_config.quant_config
    lora_config = vllm_config.lora_config

    self.config = config
    self.lora_config = lora_config

    self.quant_config = quant_config
    self.model = Qwen2Model(
        vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")
    )
    self.head_dtype = vllm_config.model_config.head_dtype

    self.score = nn.Sequential(
        ColumnParallelLinear(
            config.hidden_size,
            config.hidden_size,
            quant_config=quant_config,
            params_dtype=self.head_dtype,
            return_bias=False,
        ),
        nn.ReLU(),
        RowParallelLinear(
            config.hidden_size,
            config.num_labels,
            params_dtype=self.head_dtype,
            quant_config=quant_config,
            return_bias=False,
        ),
    )
    self.make_empty_intermediate_tensors = (
        self.model.make_empty_intermediate_tensors
    )

forward 

forward(
    input_ids: Tensor,
    positions: Tensor,
    intermediate_tensors: IntermediateTensors | None = None,
    inputs_embeds: Tensor | None = None,
) -> Tensor | IntermediateTensors
Source code in vllm/model_executor/models/qwen2_rm.py 
81
82
83
84
85
86
87
88
89
90
91
92
93
def forward(
    self,
    input_ids: torch.Tensor,
    positions: torch.Tensor,
    intermediate_tensors: IntermediateTensors | None = None,
    inputs_embeds: torch.Tensor | None = None,
) -> torch.Tensor | IntermediateTensors:
    hidden_states = self.model(
        input_ids, positions, intermediate_tensors, inputs_embeds
    )
    hidden_states = hidden_states.to(self.head_dtype)
    logits = self.score(hidden_states)
    return logits

get_input_embeddings 

get_input_embeddings(input_ids: Tensor) -> Tensor
Source code in vllm/model_executor/models/qwen2_rm.py 
78
79
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
    return self.model.get_input_embeddings(input_ids)

load_weights 

load_weights(
    weights: Iterable[tuple[str, Tensor]],
) -> set[str]
Source code in vllm/model_executor/models/qwen2_rm.py 
95
96
97
def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
    loader = AutoWeightsLoader(self, ignore_unexpected_prefixes=["lm_head."])
    return loader.load_weights(weights)