Skip to content

vllm.model_executor.models.qwen

Inference-only QWen model compatible with HuggingFace weights.

QWenAttention

Bases: Module

Source code in vllm/model_executor/models/qwen.py 
 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
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
class QWenAttention(nn.Module):

    def __init__(
        self,
        hidden_size: int,
        num_heads: int,
        max_position_embeddings: int,
        rope_theta: float = 10000,
        rope_scaling: Optional[dict[str, Any]] = None,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ):
        super().__init__()
        self.hidden_size = hidden_size
        tensor_model_parallel_world_size = get_tensor_model_parallel_world_size(
        )
        self.total_num_heads = num_heads
        assert self.total_num_heads % tensor_model_parallel_world_size == 0
        self.num_heads = (self.total_num_heads //
                          tensor_model_parallel_world_size)
        self.head_dim = hidden_size // self.total_num_heads
        self.c_attn = QKVParallelLinear(
            hidden_size,
            self.head_dim,
            self.total_num_heads,
            bias=True,
            quant_config=quant_config,
        )
        self.c_proj = RowParallelLinear(
            self.total_num_heads * self.head_dim,
            hidden_size,
            bias=False,
            quant_config=quant_config,
        )
        self.scaling = self.head_dim**-0.5

        self.rotary_emb = get_rope(
            self.head_dim,
            rotary_dim=self.head_dim,
            max_position=max_position_embeddings,
            base=rope_theta,
            rope_scaling=rope_scaling,
        )
        self.attn = Attention(self.num_heads,
                              self.head_dim,
                              self.scaling,
                              cache_config=cache_config,
                              quant_config=quant_config,
                              prefix=f"{prefix}.attn")

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
        qkv, _ = self.c_attn(hidden_states)
        q, k, v = qkv.chunk(chunks=3, dim=-1)
        q, k = self.rotary_emb(positions, q, k)
        attn_output = self.attn(q, k, v)
        output, _ = self.c_proj(attn_output)
        return output

attn instance-attribute 

attn = Attention(
    num_heads,
    head_dim,
    scaling,
    cache_config=cache_config,
    quant_config=quant_config,
    prefix=f"{prefix}.attn",
)

c_attn instance-attribute 

c_attn = QKVParallelLinear(
    hidden_size,
    head_dim,
    total_num_heads,
    bias=True,
    quant_config=quant_config,
)

c_proj instance-attribute 

c_proj = RowParallelLinear(
    total_num_heads * head_dim,
    hidden_size,
    bias=False,
    quant_config=quant_config,
)

head_dim instance-attribute 

head_dim = hidden_size // total_num_heads

hidden_size instance-attribute 

hidden_size = hidden_size

num_heads instance-attribute 

num_heads = (
    total_num_heads // tensor_model_parallel_world_size
)

rotary_emb instance-attribute 

rotary_emb = get_rope(
    head_dim,
    rotary_dim=head_dim,
    max_position=max_position_embeddings,
    base=rope_theta,
    rope_scaling=rope_scaling,
)

scaling instance-attribute 

scaling = head_dim ** -0.5

total_num_heads instance-attribute 

total_num_heads = num_heads

__init__ 

__init__(
    hidden_size: int,
    num_heads: int,
    max_position_embeddings: int,
    rope_theta: float = 10000,
    rope_scaling: Optional[dict[str, Any]] = None,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
)
Source code in vllm/model_executor/models/qwen.py 
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
def __init__(
    self,
    hidden_size: int,
    num_heads: int,
    max_position_embeddings: int,
    rope_theta: float = 10000,
    rope_scaling: Optional[dict[str, Any]] = None,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
):
    super().__init__()
    self.hidden_size = hidden_size
    tensor_model_parallel_world_size = get_tensor_model_parallel_world_size(
    )
    self.total_num_heads = num_heads
    assert self.total_num_heads % tensor_model_parallel_world_size == 0
    self.num_heads = (self.total_num_heads //
                      tensor_model_parallel_world_size)
    self.head_dim = hidden_size // self.total_num_heads
    self.c_attn = QKVParallelLinear(
        hidden_size,
        self.head_dim,
        self.total_num_heads,
        bias=True,
        quant_config=quant_config,
    )
    self.c_proj = RowParallelLinear(
        self.total_num_heads * self.head_dim,
        hidden_size,
        bias=False,
        quant_config=quant_config,
    )
    self.scaling = self.head_dim**-0.5

    self.rotary_emb = get_rope(
        self.head_dim,
        rotary_dim=self.head_dim,
        max_position=max_position_embeddings,
        base=rope_theta,
        rope_scaling=rope_scaling,
    )
    self.attn = Attention(self.num_heads,
                          self.head_dim,
                          self.scaling,
                          cache_config=cache_config,
                          quant_config=quant_config,
                          prefix=f"{prefix}.attn")

forward 

forward(positions: Tensor, hidden_states: Tensor) -> Tensor
Source code in vllm/model_executor/models/qwen.py 
124
125
126
127
128
129
130
131
132
133
134
def forward(
    self,
    positions: torch.Tensor,
    hidden_states: torch.Tensor,
) -> torch.Tensor:
    qkv, _ = self.c_attn(hidden_states)
    q, k, v = qkv.chunk(chunks=3, dim=-1)
    q, k = self.rotary_emb(positions, q, k)
    attn_output = self.attn(q, k, v)
    output, _ = self.c_proj(attn_output)
    return output

QWenBaseModel

Bases: Module

Source code in vllm/model_executor/models/qwen.py 
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
class QWenBaseModel(nn.Module):

    def __init__(
        self,
        *,
        vllm_config: VllmConfig,
        prefix: str = "",
        transformer_type: type[QWenModel] = QWenModel,
    ) -> None:
        super().__init__()
        config = vllm_config.model_config.hf_config
        quant_config = vllm_config.quant_config
        multimodal_config = vllm_config.model_config.multimodal_config
        self.config = config
        self.multimodal_config = multimodal_config
        self.quant_config = quant_config
        self.transformer = transformer_type(vllm_config=vllm_config,
                                            prefix=maybe_prefix(
                                                prefix, "transformer"))
        self.lm_head = ParallelLMHead(config.vocab_size,
                                      config.hidden_size,
                                      quant_config=quant_config)
        if self.config.tie_word_embeddings:
            self.lm_head.weight = self.transformer.wte.weight
        self.logits_processor = LogitsProcessor(config.vocab_size)
        self.make_empty_intermediate_tensors = (
            self.transformer.make_empty_intermediate_tensors)

    def compute_logits(
        self,
        hidden_states: torch.Tensor,
        sampling_metadata: SamplingMetadata,
    ) -> Optional[torch.Tensor]:
        logits = self.logits_processor(self.lm_head, hidden_states,
                                       sampling_metadata)
        return logits

    def load_weights(self, weights: Iterable[tuple[str,
                                                   torch.Tensor]]) -> set[str]:
        stacked_params_mapping = [
            # (param_name, shard_name, shard_id)
            ("gate_up_proj", "w2", 0),
            ("gate_up_proj", "w1", 1),
        ]
        params_dict = dict(self.named_parameters())
        loaded_params: set[str] = set()
        for name, loaded_weight in weights:
            if "rotary_emb.inv_freq" in name:
                continue
            for (param_name, weight_name, shard_id) in stacked_params_mapping:
                if weight_name not in name:
                    continue
                name = name.replace(weight_name, param_name)
                # Skip loading extra bias for GPTQ models.
                if name.endswith(".bias") and name not in params_dict:
                    continue
                # Skip layers on other devices.
                if is_pp_missing_parameter(name, self):
                    continue
                param = params_dict[name]
                weight_loader = param.weight_loader
                weight_loader(param, loaded_weight, shard_id)
                break
            else:
                # Skip loading extra bias for GPTQ models.
                if name.endswith(".bias") and name not in params_dict:
                    continue
                # Skip layers on other devices.
                if is_pp_missing_parameter(name, self):
                    continue
                param = params_dict[name]
                weight_loader = getattr(param, "weight_loader",
                                        default_weight_loader)
                weight_loader(param, loaded_weight)
            loaded_params.add(name)
        return loaded_params

config instance-attribute 

config = config

lm_head instance-attribute 

lm_head = ParallelLMHead(
    vocab_size, hidden_size, quant_config=quant_config
)

logits_processor instance-attribute 

logits_processor = LogitsProcessor(vocab_size)

make_empty_intermediate_tensors instance-attribute 

make_empty_intermediate_tensors = (
    make_empty_intermediate_tensors
)

multimodal_config instance-attribute 

multimodal_config = multimodal_config

quant_config instance-attribute 

quant_config = quant_config

transformer instance-attribute 

transformer = transformer_type(
    vllm_config=vllm_config,
    prefix=maybe_prefix(prefix, "transformer"),
)

__init__ 

__init__(
    *,
    vllm_config: VllmConfig,
    prefix: str = "",
    transformer_type: type[QWenModel] = QWenModel,
) -> None
Source code in vllm/model_executor/models/qwen.py 
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
def __init__(
    self,
    *,
    vllm_config: VllmConfig,
    prefix: str = "",
    transformer_type: type[QWenModel] = QWenModel,
) -> None:
    super().__init__()
    config = vllm_config.model_config.hf_config
    quant_config = vllm_config.quant_config
    multimodal_config = vllm_config.model_config.multimodal_config
    self.config = config
    self.multimodal_config = multimodal_config
    self.quant_config = quant_config
    self.transformer = transformer_type(vllm_config=vllm_config,
                                        prefix=maybe_prefix(
                                            prefix, "transformer"))
    self.lm_head = ParallelLMHead(config.vocab_size,
                                  config.hidden_size,
                                  quant_config=quant_config)
    if self.config.tie_word_embeddings:
        self.lm_head.weight = self.transformer.wte.weight
    self.logits_processor = LogitsProcessor(config.vocab_size)
    self.make_empty_intermediate_tensors = (
        self.transformer.make_empty_intermediate_tensors)

compute_logits 

compute_logits(
    hidden_states: Tensor,
    sampling_metadata: SamplingMetadata,
) -> Optional[Tensor]
Source code in vllm/model_executor/models/qwen.py 
280
281
282
283
284
285
286
287
def compute_logits(
    self,
    hidden_states: torch.Tensor,
    sampling_metadata: SamplingMetadata,
) -> Optional[torch.Tensor]:
    logits = self.logits_processor(self.lm_head, hidden_states,
                                   sampling_metadata)
    return logits

load_weights 

load_weights(
    weights: Iterable[tuple[str, Tensor]],
) -> set[str]
Source code in vllm/model_executor/models/qwen.py 
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
def load_weights(self, weights: Iterable[tuple[str,
                                               torch.Tensor]]) -> set[str]:
    stacked_params_mapping = [
        # (param_name, shard_name, shard_id)
        ("gate_up_proj", "w2", 0),
        ("gate_up_proj", "w1", 1),
    ]
    params_dict = dict(self.named_parameters())
    loaded_params: set[str] = set()
    for name, loaded_weight in weights:
        if "rotary_emb.inv_freq" in name:
            continue
        for (param_name, weight_name, shard_id) in stacked_params_mapping:
            if weight_name not in name:
                continue
            name = name.replace(weight_name, param_name)
            # Skip loading extra bias for GPTQ models.
            if name.endswith(".bias") and name not in params_dict:
                continue
            # Skip layers on other devices.
            if is_pp_missing_parameter(name, self):
                continue
            param = params_dict[name]
            weight_loader = param.weight_loader
            weight_loader(param, loaded_weight, shard_id)
            break
        else:
            # Skip loading extra bias for GPTQ models.
            if name.endswith(".bias") and name not in params_dict:
                continue
            # Skip layers on other devices.
            if is_pp_missing_parameter(name, self):
                continue
            param = params_dict[name]
            weight_loader = getattr(param, "weight_loader",
                                    default_weight_loader)
            weight_loader(param, loaded_weight)
        loaded_params.add(name)
    return loaded_params

QWenBlock

Bases: Module

Source code in vllm/model_executor/models/qwen.py 
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
class QWenBlock(nn.Module):

    def __init__(
        self,
        config: PretrainedConfig,
        cache_config: Optional[CacheConfig] = None,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ):
        super().__init__()
        self.ln_1 = RMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)

        rope_theta = getattr(config, "rope_theta", 10000)
        rope_scaling = getattr(config, "rope_scaling", None)
        self.attn = QWenAttention(config.hidden_size,
                                  config.num_attention_heads,
                                  config.max_position_embeddings,
                                  rope_theta=rope_theta,
                                  rope_scaling=rope_scaling,
                                  cache_config=cache_config,
                                  quant_config=quant_config,
                                  prefix=f"{prefix}.attn")

        self.ln_2 = RMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)

        self.mlp = QWenMLP(config.hidden_size,
                           config.intermediate_size // 2,
                           quant_config=quant_config)

    def forward(
        self,
        positions: torch.Tensor,
        hidden_states: torch.Tensor,
        residual: Optional[torch.Tensor],
    ) -> tuple[torch.Tensor, torch.Tensor]:
        # Self Attention
        if residual is None:
            residual = hidden_states
            hidden_states = self.ln_1(hidden_states)
        else:
            hidden_states, residual = self.ln_1(hidden_states, residual)
        hidden_states = self.attn(
            positions=positions,
            hidden_states=hidden_states,
        )

        # Fully Connected
        hidden_states, residual = self.ln_2(hidden_states, residual)
        hidden_states = self.mlp(hidden_states)
        return hidden_states, residual

attn instance-attribute 

attn = QWenAttention(
    hidden_size,
    num_attention_heads,
    max_position_embeddings,
    rope_theta=rope_theta,
    rope_scaling=rope_scaling,
    cache_config=cache_config,
    quant_config=quant_config,
    prefix=f"{prefix}.attn",
)

ln_1 instance-attribute 

ln_1 = RMSNorm(hidden_size, eps=layer_norm_epsilon)

ln_2 instance-attribute 

ln_2 = RMSNorm(hidden_size, eps=layer_norm_epsilon)

mlp instance-attribute 

mlp = QWenMLP(
    hidden_size,
    intermediate_size // 2,
    quant_config=quant_config,
)

__init__ 

__init__(
    config: PretrainedConfig,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
)
Source code in vllm/model_executor/models/qwen.py 
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
def __init__(
    self,
    config: PretrainedConfig,
    cache_config: Optional[CacheConfig] = None,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
):
    super().__init__()
    self.ln_1 = RMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)

    rope_theta = getattr(config, "rope_theta", 10000)
    rope_scaling = getattr(config, "rope_scaling", None)
    self.attn = QWenAttention(config.hidden_size,
                              config.num_attention_heads,
                              config.max_position_embeddings,
                              rope_theta=rope_theta,
                              rope_scaling=rope_scaling,
                              cache_config=cache_config,
                              quant_config=quant_config,
                              prefix=f"{prefix}.attn")

    self.ln_2 = RMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)

    self.mlp = QWenMLP(config.hidden_size,
                       config.intermediate_size // 2,
                       quant_config=quant_config)

forward 

forward(
    positions: Tensor,
    hidden_states: Tensor,
    residual: Optional[Tensor],
) -> tuple[Tensor, Tensor]
Source code in vllm/model_executor/models/qwen.py 
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
def forward(
    self,
    positions: torch.Tensor,
    hidden_states: torch.Tensor,
    residual: Optional[torch.Tensor],
) -> tuple[torch.Tensor, torch.Tensor]:
    # Self Attention
    if residual is None:
        residual = hidden_states
        hidden_states = self.ln_1(hidden_states)
    else:
        hidden_states, residual = self.ln_1(hidden_states, residual)
    hidden_states = self.attn(
        positions=positions,
        hidden_states=hidden_states,
    )

    # Fully Connected
    hidden_states, residual = self.ln_2(hidden_states, residual)
    hidden_states = self.mlp(hidden_states)
    return hidden_states, residual

QWenLMHeadModel

Bases: QWenBaseModel, SupportsPP, SupportsLoRA

Source code in vllm/model_executor/models/qwen.py 
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
class QWenLMHeadModel(QWenBaseModel, SupportsPP, SupportsLoRA):
    packed_modules_mapping = {
        "c_attn": ["c_attn"],
        "gate_up_proj": [
            "w2",
            "w1",
        ],
    }

    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        config = vllm_config.model_config.hf_config
        if hasattr(config, "visual"):
            hf_overrides = {
                "architectures": ["QwenVLForConditionalGeneration"]
            }
            raise RuntimeError(
                "The configuration of this model indicates that it supports "
                "vision inputs, but you instantiated the text-only version "
                "of this model. Please use the vision model by setting "
                f"`--hf-overrides '{json.dumps(hf_overrides)}'`")

        super().__init__(vllm_config=vllm_config, prefix=prefix)

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        intermediate_tensors: Optional[IntermediateTensors] = None,
        inputs_embeds: Optional[torch.Tensor] = None,
    ) -> Union[torch.Tensor, IntermediateTensors]:
        hidden_states = self.transformer(input_ids, positions,
                                         intermediate_tensors, inputs_embeds)
        return hidden_states

packed_modules_mapping class-attribute instance-attribute 

packed_modules_mapping = {
    "c_attn": ["c_attn"],
    "gate_up_proj": ["w2", "w1"],
}

__init__ 

__init__(*, vllm_config: VllmConfig, prefix: str = '')
Source code in vllm/model_executor/models/qwen.py 
339
340
341
342
343
344
345
346
347
348
349
350
351
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
    config = vllm_config.model_config.hf_config
    if hasattr(config, "visual"):
        hf_overrides = {
            "architectures": ["QwenVLForConditionalGeneration"]
        }
        raise RuntimeError(
            "The configuration of this model indicates that it supports "
            "vision inputs, but you instantiated the text-only version "
            "of this model. Please use the vision model by setting "
            f"`--hf-overrides '{json.dumps(hf_overrides)}'`")

    super().__init__(vllm_config=vllm_config, prefix=prefix)

forward 

forward(
    input_ids: Tensor,
    positions: Tensor,
    intermediate_tensors: Optional[
        IntermediateTensors
    ] = None,
    inputs_embeds: Optional[Tensor] = None,
) -> Union[Tensor, IntermediateTensors]
Source code in vllm/model_executor/models/qwen.py 
353
354
355
356
357
358
359
360
361
362
def forward(
    self,
    input_ids: torch.Tensor,
    positions: torch.Tensor,
    intermediate_tensors: Optional[IntermediateTensors] = None,
    inputs_embeds: Optional[torch.Tensor] = None,
) -> Union[torch.Tensor, IntermediateTensors]:
    hidden_states = self.transformer(input_ids, positions,
                                     intermediate_tensors, inputs_embeds)
    return hidden_states

QWenMLP

Bases: Module

MLP for the language component of the Qwen model, which contains a MergedColumnParallelLinear merging 2 outputs via silu activation.

Source code in vllm/model_executor/models/qwen.py 
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
class QWenMLP(nn.Module):
    """MLP for the language component of the Qwen model, which contains a
    MergedColumnParallelLinear merging 2 outputs via silu activation."""

    def __init__(
        self,
        hidden_size: int,
        intermediate_size: int,
        hidden_act: str = "silu",
        quant_config: Optional[QuantizationConfig] = None,
    ):
        super().__init__()
        self.gate_up_proj = MergedColumnParallelLinear(
            hidden_size, [intermediate_size] * 2,
            bias=False,
            quant_config=quant_config)
        self.c_proj = RowParallelLinear(intermediate_size,
                                        hidden_size,
                                        bias=False,
                                        quant_config=quant_config)
        if hidden_act != "silu":
            raise ValueError(f"Unsupported activation: {hidden_act}. "
                             "Only silu is supported for now.")
        self.act_fn = SiluAndMul()

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        gate_up, _ = self.gate_up_proj(x)
        x = self.act_fn(gate_up)
        x, _ = self.c_proj(x)
        return x

act_fn instance-attribute 

act_fn = SiluAndMul()

c_proj instance-attribute 

c_proj = RowParallelLinear(
    intermediate_size,
    hidden_size,
    bias=False,
    quant_config=quant_config,
)

gate_up_proj instance-attribute 

gate_up_proj = MergedColumnParallelLinear(
    hidden_size,
    [intermediate_size] * 2,
    bias=False,
    quant_config=quant_config,
)

__init__ 

__init__(
    hidden_size: int,
    intermediate_size: int,
    hidden_act: str = "silu",
    quant_config: Optional[QuantizationConfig] = None,
)
Source code in vllm/model_executor/models/qwen.py 
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
def __init__(
    self,
    hidden_size: int,
    intermediate_size: int,
    hidden_act: str = "silu",
    quant_config: Optional[QuantizationConfig] = None,
):
    super().__init__()
    self.gate_up_proj = MergedColumnParallelLinear(
        hidden_size, [intermediate_size] * 2,
        bias=False,
        quant_config=quant_config)
    self.c_proj = RowParallelLinear(intermediate_size,
                                    hidden_size,
                                    bias=False,
                                    quant_config=quant_config)
    if hidden_act != "silu":
        raise ValueError(f"Unsupported activation: {hidden_act}. "
                         "Only silu is supported for now.")
    self.act_fn = SiluAndMul()

forward 

forward(x: Tensor) -> Tensor
Source code in vllm/model_executor/models/qwen.py 
66
67
68
69
70
def forward(self, x: torch.Tensor) -> torch.Tensor:
    gate_up, _ = self.gate_up_proj(x)
    x = self.act_fn(gate_up)
    x, _ = self.c_proj(x)
    return x

QWenModel

Bases: Module

Source code in vllm/model_executor/models/qwen.py 
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
@support_torch_compile
class QWenModel(nn.Module):

    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
        super().__init__()

        config = vllm_config.model_config.hf_config
        cache_config = vllm_config.cache_config
        quant_config = vllm_config.quant_config

        self.config = config
        self.vocab_size = config.vocab_size

        self.wte = VocabParallelEmbedding(
            config.vocab_size,
            config.hidden_size,
        )
        self.start_layer, self.end_layer, self.h = make_layers(
            config.num_hidden_layers,
            lambda prefix: QWenBlock(
                config, cache_config, quant_config, prefix=prefix),
            prefix=f"{prefix}.h")
        self.ln_f = RMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)
        self.make_empty_intermediate_tensors = (
            make_empty_intermediate_tensors_factory(
                ["hidden_states", "residual"], config.hidden_size))

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

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        intermediate_tensors: Optional[IntermediateTensors],
        inputs_embeds: Optional[torch.Tensor] = None,
    ) -> Union[torch.Tensor, IntermediateTensors]:
        if get_pp_group().is_first_rank:
            if inputs_embeds is not None:
                hidden_states = inputs_embeds
            else:
                hidden_states = self.get_input_embeddings(input_ids)
            residual = None
        else:
            assert intermediate_tensors is not None
            hidden_states = intermediate_tensors["hidden_states"]
            residual = intermediate_tensors["residual"]

        for layer in self.h[self.start_layer:self.end_layer]:
            hidden_states, residual = layer(
                positions,
                hidden_states,
                residual,
            )
        if not get_pp_group().is_last_rank:
            return IntermediateTensors({
                "hidden_states": hidden_states,
                "residual": residual
            })
        hidden_states, _ = self.ln_f(hidden_states, residual)
        return hidden_states

config instance-attribute 

config = config

ln_f instance-attribute 

ln_f = RMSNorm(hidden_size, eps=layer_norm_epsilon)

make_empty_intermediate_tensors instance-attribute 

make_empty_intermediate_tensors = (
    make_empty_intermediate_tensors_factory(
        ["hidden_states", "residual"], hidden_size
    )
)

vocab_size instance-attribute 

vocab_size = vocab_size

wte instance-attribute 

wte = VocabParallelEmbedding(vocab_size, hidden_size)

__init__ 

__init__(*, vllm_config: VllmConfig, prefix: str = '')
Source code in vllm/model_executor/models/qwen.py 
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
    super().__init__()

    config = vllm_config.model_config.hf_config
    cache_config = vllm_config.cache_config
    quant_config = vllm_config.quant_config

    self.config = config
    self.vocab_size = config.vocab_size

    self.wte = VocabParallelEmbedding(
        config.vocab_size,
        config.hidden_size,
    )
    self.start_layer, self.end_layer, self.h = make_layers(
        config.num_hidden_layers,
        lambda prefix: QWenBlock(
            config, cache_config, quant_config, prefix=prefix),
        prefix=f"{prefix}.h")
    self.ln_f = RMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)
    self.make_empty_intermediate_tensors = (
        make_empty_intermediate_tensors_factory(
            ["hidden_states", "residual"], config.hidden_size))

forward 

forward(
    input_ids: Tensor,
    positions: Tensor,
    intermediate_tensors: Optional[IntermediateTensors],
    inputs_embeds: Optional[Tensor] = None,
) -> Union[Tensor, IntermediateTensors]
Source code in vllm/model_executor/models/qwen.py 
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
def forward(
    self,
    input_ids: torch.Tensor,
    positions: torch.Tensor,
    intermediate_tensors: Optional[IntermediateTensors],
    inputs_embeds: Optional[torch.Tensor] = None,
) -> Union[torch.Tensor, IntermediateTensors]:
    if get_pp_group().is_first_rank:
        if inputs_embeds is not None:
            hidden_states = inputs_embeds
        else:
            hidden_states = self.get_input_embeddings(input_ids)
        residual = None
    else:
        assert intermediate_tensors is not None
        hidden_states = intermediate_tensors["hidden_states"]
        residual = intermediate_tensors["residual"]

    for layer in self.h[self.start_layer:self.end_layer]:
        hidden_states, residual = layer(
            positions,
            hidden_states,
            residual,
        )
    if not get_pp_group().is_last_rank:
        return IntermediateTensors({
            "hidden_states": hidden_states,
            "residual": residual
        })
    hidden_states, _ = self.ln_f(hidden_states, residual)
    return hidden_states

get_input_embeddings 

get_input_embeddings(input_ids: Tensor) -> Tensor
Source code in vllm/model_executor/models/qwen.py 
216
217
def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
    return self.wte(input_ids)