Skip to content

vllm.model_executor.models.clip

Minimal implementation of CLIPVisionModel intended to be only used within a vision language model.

CLIPAttention

Bases: Module

Multi-headed attention from 'Attention Is All You Need' paper

Source code in vllm/model_executor/models/clip.py 
 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
135
136
137
138
139
140
141
142
143
class CLIPAttention(nn.Module):
    """Multi-headed attention from 'Attention Is All You Need' paper"""

    def __init__(
        self,
        config: CLIPVisionConfig,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ):
        super().__init__()
        self.config = config
        self.embed_dim = config.hidden_size
        self.num_heads = config.num_attention_heads
        self.head_dim = self.embed_dim // self.num_heads
        if self.head_dim * self.num_heads != self.embed_dim:
            raise ValueError(
                "embed_dim must be divisible by num_heads "
                f"(got `embed_dim`: {self.embed_dim} and `num_heads`:"
                f" {self.num_heads}).")
        self.scale = self.head_dim**-0.5

        self.qkv_proj = QKVParallelLinear(
            hidden_size=self.embed_dim,
            head_size=self.head_dim,
            total_num_heads=self.num_heads,
            quant_config=quant_config,
            prefix=f"{prefix}.qkv_proj",
        )

        self.out_proj = RowParallelLinear(
            input_size=self.embed_dim,
            output_size=self.embed_dim,
            quant_config=quant_config,
            prefix=f"{prefix}.out_proj",
        )

        self.tp_size = get_tensor_model_parallel_world_size()
        self.num_heads_per_partition = divide(self.num_heads, self.tp_size)

        self.attn = MultiHeadAttention(self.num_heads_per_partition,
                                       self.head_dim, self.scale)

    def forward(
        self,
        hidden_states: torch.Tensor,
    ):
        """Input shape: Batch x Time x Channel"""

        qkv_states, _ = self.qkv_proj(hidden_states)
        query_states, key_states, value_states = qkv_states.chunk(3, dim=-1)
        out = self.attn(query_states, key_states, value_states)
        attn_output, _ = self.out_proj(out)

        return attn_output, None

attn instance-attribute 

attn = MultiHeadAttention(
    num_heads_per_partition, head_dim, scale
)

config instance-attribute 

config = config

embed_dim instance-attribute 

embed_dim = hidden_size

head_dim instance-attribute 

head_dim = embed_dim // num_heads

num_heads instance-attribute 

num_heads = num_attention_heads

num_heads_per_partition instance-attribute 

num_heads_per_partition = divide(num_heads, tp_size)

out_proj instance-attribute 

out_proj = RowParallelLinear(
    input_size=embed_dim,
    output_size=embed_dim,
    quant_config=quant_config,
    prefix=f"{prefix}.out_proj",
)

qkv_proj instance-attribute 

qkv_proj = QKVParallelLinear(
    hidden_size=embed_dim,
    head_size=head_dim,
    total_num_heads=num_heads,
    quant_config=quant_config,
    prefix=f"{prefix}.qkv_proj",
)

scale instance-attribute 

scale = head_dim ** -0.5

tp_size instance-attribute 

tp_size = get_tensor_model_parallel_world_size()

__init__ 

__init__(
    config: CLIPVisionConfig,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
)
Source code in vllm/model_executor/models/clip.py 
 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
def __init__(
    self,
    config: CLIPVisionConfig,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
):
    super().__init__()
    self.config = config
    self.embed_dim = config.hidden_size
    self.num_heads = config.num_attention_heads
    self.head_dim = self.embed_dim // self.num_heads
    if self.head_dim * self.num_heads != self.embed_dim:
        raise ValueError(
            "embed_dim must be divisible by num_heads "
            f"(got `embed_dim`: {self.embed_dim} and `num_heads`:"
            f" {self.num_heads}).")
    self.scale = self.head_dim**-0.5

    self.qkv_proj = QKVParallelLinear(
        hidden_size=self.embed_dim,
        head_size=self.head_dim,
        total_num_heads=self.num_heads,
        quant_config=quant_config,
        prefix=f"{prefix}.qkv_proj",
    )

    self.out_proj = RowParallelLinear(
        input_size=self.embed_dim,
        output_size=self.embed_dim,
        quant_config=quant_config,
        prefix=f"{prefix}.out_proj",
    )

    self.tp_size = get_tensor_model_parallel_world_size()
    self.num_heads_per_partition = divide(self.num_heads, self.tp_size)

    self.attn = MultiHeadAttention(self.num_heads_per_partition,
                                   self.head_dim, self.scale)

forward 

forward(hidden_states: Tensor)

Input shape: Batch x Time x Channel

Source code in vllm/model_executor/models/clip.py 
132
133
134
135
136
137
138
139
140
141
142
143
def forward(
    self,
    hidden_states: torch.Tensor,
):
    """Input shape: Batch x Time x Channel"""

    qkv_states, _ = self.qkv_proj(hidden_states)
    query_states, key_states, value_states = qkv_states.chunk(3, dim=-1)
    out = self.attn(query_states, key_states, value_states)
    attn_output, _ = self.out_proj(out)

    return attn_output, None

CLIPEncoder

Bases: Module

Transformer encoder consisting of config.num_hidden_layers self attention layers. Each layer is a [CLIPEncoderLayer].

Parameters:

Name Type Description Default
config CLIPVisionConfig

CLIPConfig

 required
Source code in vllm/model_executor/models/clip.py 
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
250
251
252
253
254
255
256
257
258
259
class CLIPEncoder(nn.Module):
    """
    Transformer encoder consisting of `config.num_hidden_layers` self
    attention layers. Each layer is a [`CLIPEncoderLayer`].

    Args:
        config: CLIPConfig
    """

    def __init__(
        self,
        config: CLIPVisionConfig,
        quant_config: Optional[QuantizationConfig] = None,
        num_hidden_layers_override: Optional[int] = None,
        prefix: str = "",
    ) -> None:
        super().__init__()

        self.config = config

        if num_hidden_layers_override is None:
            num_hidden_layers = config.num_hidden_layers
        else:
            num_hidden_layers = num_hidden_layers_override
        self.layers = nn.ModuleList([
            CLIPEncoderLayer(config=config,
                             quant_config=quant_config,
                             prefix=f"{prefix}.layers.{layer_idx}")
            for layer_idx in range(num_hidden_layers)
        ])

    def forward(
        self, inputs_embeds: torch.Tensor, return_all_hidden_states: bool
    ) -> Union[torch.Tensor, list[torch.Tensor]]:
        hidden_states_pool = [inputs_embeds]
        hidden_states = inputs_embeds

        for encoder_layer in self.layers:
            hidden_states = encoder_layer(hidden_states)
            if return_all_hidden_states:
                hidden_states_pool.append(hidden_states)
        # If we have multiple feature sample layers, we return all hidden
        # states in order and grab the ones we need by index.
        if return_all_hidden_states:
            return hidden_states_pool
        return hidden_states

config instance-attribute 

config = config

layers instance-attribute 

layers = ModuleList(
    [
        CLIPEncoderLayer(
            config=config,
            quant_config=quant_config,
            prefix=f"{prefix}.layers.{layer_idx}",
        )
        for layer_idx in range(num_hidden_layers)
    ]
)

__init__ 

__init__(
    config: CLIPVisionConfig,
    quant_config: Optional[QuantizationConfig] = None,
    num_hidden_layers_override: Optional[int] = None,
    prefix: str = "",
) -> None
Source code in vllm/model_executor/models/clip.py 
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
def __init__(
    self,
    config: CLIPVisionConfig,
    quant_config: Optional[QuantizationConfig] = None,
    num_hidden_layers_override: Optional[int] = None,
    prefix: str = "",
) -> None:
    super().__init__()

    self.config = config

    if num_hidden_layers_override is None:
        num_hidden_layers = config.num_hidden_layers
    else:
        num_hidden_layers = num_hidden_layers_override
    self.layers = nn.ModuleList([
        CLIPEncoderLayer(config=config,
                         quant_config=quant_config,
                         prefix=f"{prefix}.layers.{layer_idx}")
        for layer_idx in range(num_hidden_layers)
    ])

forward 

forward(
    inputs_embeds: Tensor, return_all_hidden_states: bool
) -> Union[Tensor, list[Tensor]]
Source code in vllm/model_executor/models/clip.py 
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
def forward(
    self, inputs_embeds: torch.Tensor, return_all_hidden_states: bool
) -> Union[torch.Tensor, list[torch.Tensor]]:
    hidden_states_pool = [inputs_embeds]
    hidden_states = inputs_embeds

    for encoder_layer in self.layers:
        hidden_states = encoder_layer(hidden_states)
        if return_all_hidden_states:
            hidden_states_pool.append(hidden_states)
    # If we have multiple feature sample layers, we return all hidden
    # states in order and grab the ones we need by index.
    if return_all_hidden_states:
        return hidden_states_pool
    return hidden_states

CLIPEncoderInfo

Bases: VisionEncoderInfo[CLIPVisionConfig]

Source code in vllm/model_executor/models/clip.py 
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
class CLIPEncoderInfo(VisionEncoderInfo[CLIPVisionConfig]):

    def get_num_image_tokens(
        self,
        *,
        image_width: int,
        image_height: int,
    ) -> int:
        return self.get_patch_grid_length()**2 + 1

    def get_image_size(self) -> int:
        return self.vision_config.image_size

    def get_patch_size(self) -> int:
        return self.vision_config.patch_size

    def get_patch_grid_length(self) -> int:
        image_size, patch_size = self.get_image_size(), self.get_patch_size()
        assert image_size % patch_size == 0
        return image_size // patch_size

get_image_size 

get_image_size() -> int
Source code in vllm/model_executor/models/clip.py 
35
36
def get_image_size(self) -> int:
    return self.vision_config.image_size

get_num_image_tokens 

get_num_image_tokens(
    *, image_width: int, image_height: int
) -> int
Source code in vllm/model_executor/models/clip.py 
27
28
29
30
31
32
33
def get_num_image_tokens(
    self,
    *,
    image_width: int,
    image_height: int,
) -> int:
    return self.get_patch_grid_length()**2 + 1

get_patch_grid_length 

get_patch_grid_length() -> int
Source code in vllm/model_executor/models/clip.py 
41
42
43
44
def get_patch_grid_length(self) -> int:
    image_size, patch_size = self.get_image_size(), self.get_patch_size()
    assert image_size % patch_size == 0
    return image_size // patch_size

get_patch_size 

get_patch_size() -> int
Source code in vllm/model_executor/models/clip.py 
38
39
def get_patch_size(self) -> int:
    return self.vision_config.patch_size

CLIPEncoderLayer

Bases: Module

Source code in vllm/model_executor/models/clip.py 
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
class CLIPEncoderLayer(nn.Module):

    def __init__(
        self,
        config: CLIPVisionConfig,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.self_attn = CLIPAttention(
            config,
            quant_config=quant_config,
            prefix=f"{prefix}.self_attn",
        )
        self.layer_norm1 = nn.LayerNorm(config.hidden_size,
                                        eps=config.layer_norm_eps)
        self.mlp = CLIPMLP(config,
                           quant_config=quant_config,
                           prefix=f"{prefix}.mlp")
        self.layer_norm2 = nn.LayerNorm(config.hidden_size,
                                        eps=config.layer_norm_eps)

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:

        residual = hidden_states

        hidden_states = self.layer_norm1(hidden_states)
        hidden_states, _ = self.self_attn(hidden_states=hidden_states)
        hidden_states = residual + hidden_states

        residual = hidden_states
        hidden_states = self.layer_norm2(hidden_states)
        hidden_states = self.mlp(hidden_states)
        hidden_states = residual + hidden_states

        return hidden_states

layer_norm1 instance-attribute 

layer_norm1 = LayerNorm(hidden_size, eps=layer_norm_eps)

layer_norm2 instance-attribute 

layer_norm2 = LayerNorm(hidden_size, eps=layer_norm_eps)

mlp instance-attribute 

mlp = CLIPMLP(
    config,
    quant_config=quant_config,
    prefix=f"{prefix}.mlp",
)

self_attn instance-attribute 

self_attn = CLIPAttention(
    config,
    quant_config=quant_config,
    prefix=f"{prefix}.self_attn",
)

__init__ 

__init__(
    config: CLIPVisionConfig,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
) -> None
Source code in vllm/model_executor/models/clip.py 
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
def __init__(
    self,
    config: CLIPVisionConfig,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
) -> None:
    super().__init__()
    self.self_attn = CLIPAttention(
        config,
        quant_config=quant_config,
        prefix=f"{prefix}.self_attn",
    )
    self.layer_norm1 = nn.LayerNorm(config.hidden_size,
                                    eps=config.layer_norm_eps)
    self.mlp = CLIPMLP(config,
                       quant_config=quant_config,
                       prefix=f"{prefix}.mlp")
    self.layer_norm2 = nn.LayerNorm(config.hidden_size,
                                    eps=config.layer_norm_eps)

forward 

forward(hidden_states: Tensor) -> Tensor
Source code in vllm/model_executor/models/clip.py 
198
199
200
201
202
203
204
205
206
207
208
209
210
211
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:

    residual = hidden_states

    hidden_states = self.layer_norm1(hidden_states)
    hidden_states, _ = self.self_attn(hidden_states=hidden_states)
    hidden_states = residual + hidden_states

    residual = hidden_states
    hidden_states = self.layer_norm2(hidden_states)
    hidden_states = self.mlp(hidden_states)
    hidden_states = residual + hidden_states

    return hidden_states

CLIPMLP

Bases: Module

Source code in vllm/model_executor/models/clip.py 
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
class CLIPMLP(nn.Module):

    def __init__(
        self,
        config: CLIPVisionConfig,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.config = config
        self.activation_fn = get_act_fn(config.hidden_act)
        self.fc1 = ColumnParallelLinear(config.hidden_size,
                                        config.intermediate_size,
                                        bias=True,
                                        quant_config=quant_config,
                                        prefix=f"{prefix}.fc1")
        self.fc2 = RowParallelLinear(config.intermediate_size,
                                     config.hidden_size,
                                     bias=True,
                                     quant_config=quant_config,
                                     prefix=f"{prefix}.fc2")

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        hidden_states, _ = self.fc1(hidden_states)
        hidden_states = self.activation_fn(hidden_states)
        hidden_states, _ = self.fc2(hidden_states)

        return hidden_states

activation_fn instance-attribute 

activation_fn = get_act_fn(hidden_act)

config instance-attribute 

config = config

fc1 instance-attribute 

fc1 = ColumnParallelLinear(
    hidden_size,
    intermediate_size,
    bias=True,
    quant_config=quant_config,
    prefix=f"{prefix}.fc1",
)

fc2 instance-attribute 

fc2 = RowParallelLinear(
    intermediate_size,
    hidden_size,
    bias=True,
    quant_config=quant_config,
    prefix=f"{prefix}.fc2",
)

__init__ 

__init__(
    config: CLIPVisionConfig,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
) -> None
Source code in vllm/model_executor/models/clip.py 
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
def __init__(
    self,
    config: CLIPVisionConfig,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
) -> None:
    super().__init__()
    self.config = config
    self.activation_fn = get_act_fn(config.hidden_act)
    self.fc1 = ColumnParallelLinear(config.hidden_size,
                                    config.intermediate_size,
                                    bias=True,
                                    quant_config=quant_config,
                                    prefix=f"{prefix}.fc1")
    self.fc2 = RowParallelLinear(config.intermediate_size,
                                 config.hidden_size,
                                 bias=True,
                                 quant_config=quant_config,
                                 prefix=f"{prefix}.fc2")

forward 

forward(hidden_states: Tensor) -> Tensor
Source code in vllm/model_executor/models/clip.py 
168
169
170
171
172
173
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
    hidden_states, _ = self.fc1(hidden_states)
    hidden_states = self.activation_fn(hidden_states)
    hidden_states, _ = self.fc2(hidden_states)

    return hidden_states

CLIPVisionEmbeddings

Bases: Module

Source code in vllm/model_executor/models/clip.py 
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
class CLIPVisionEmbeddings(nn.Module):

    def __init__(self, config: CLIPVisionConfig):
        super().__init__()
        self.config = config
        self.embed_dim = config.hidden_size
        self.image_size = config.image_size
        self.patch_size = config.patch_size
        assert self.image_size % self.patch_size == 0

        self.class_embedding = nn.Parameter(torch.randn(self.embed_dim))

        self.patch_embedding = nn.Conv2d(
            in_channels=config.num_channels,
            out_channels=self.embed_dim,
            kernel_size=self.patch_size,
            stride=self.patch_size,
            bias=False,
        )

        self.num_patches = (self.image_size // self.patch_size)**2
        self.num_positions = self.num_patches + 1
        self.position_embedding = nn.Embedding(self.num_positions,
                                               self.embed_dim)
        self.register_buffer("position_ids",
                             torch.arange(self.num_positions).expand((1, -1)),
                             persistent=False)

    def forward(self, pixel_values: torch.Tensor) -> torch.Tensor:
        batch_size = pixel_values.shape[0]
        target_dtype = self.patch_embedding.weight.dtype
        patch_embeds = self.patch_embedding(pixel_values.to(
            dtype=target_dtype))  # shape = [*, width, grid, grid]
        patch_embeds = patch_embeds.flatten(2).transpose(1, 2)

        class_embeds = self.class_embedding.expand(batch_size, 1, -1)
        embeddings = torch.cat([class_embeds, patch_embeds], dim=1)
        embeddings = embeddings + self.position_embedding(self.position_ids)

        return embeddings

class_embedding instance-attribute 

class_embedding = Parameter(randn(embed_dim))

config instance-attribute 

config = config

embed_dim instance-attribute 

embed_dim = hidden_size

image_size instance-attribute 

image_size = image_size

num_patches instance-attribute 

num_patches = image_size // patch_size ** 2

num_positions instance-attribute 

num_positions = num_patches + 1

patch_embedding instance-attribute 

patch_embedding = Conv2d(
    in_channels=num_channels,
    out_channels=embed_dim,
    kernel_size=patch_size,
    stride=patch_size,
    bias=False,
)

patch_size instance-attribute 

patch_size = patch_size

position_embedding instance-attribute 

position_embedding = Embedding(num_positions, embed_dim)

__init__ 

__init__(config: CLIPVisionConfig)
Source code in vllm/model_executor/models/clip.py 
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
def __init__(self, config: CLIPVisionConfig):
    super().__init__()
    self.config = config
    self.embed_dim = config.hidden_size
    self.image_size = config.image_size
    self.patch_size = config.patch_size
    assert self.image_size % self.patch_size == 0

    self.class_embedding = nn.Parameter(torch.randn(self.embed_dim))

    self.patch_embedding = nn.Conv2d(
        in_channels=config.num_channels,
        out_channels=self.embed_dim,
        kernel_size=self.patch_size,
        stride=self.patch_size,
        bias=False,
    )

    self.num_patches = (self.image_size // self.patch_size)**2
    self.num_positions = self.num_patches + 1
    self.position_embedding = nn.Embedding(self.num_positions,
                                           self.embed_dim)
    self.register_buffer("position_ids",
                         torch.arange(self.num_positions).expand((1, -1)),
                         persistent=False)

forward 

forward(pixel_values: Tensor) -> Tensor
Source code in vllm/model_executor/models/clip.py 
76
77
78
79
80
81
82
83
84
85
86
87
def forward(self, pixel_values: torch.Tensor) -> torch.Tensor:
    batch_size = pixel_values.shape[0]
    target_dtype = self.patch_embedding.weight.dtype
    patch_embeds = self.patch_embedding(pixel_values.to(
        dtype=target_dtype))  # shape = [*, width, grid, grid]
    patch_embeds = patch_embeds.flatten(2).transpose(1, 2)

    class_embeds = self.class_embedding.expand(batch_size, 1, -1)
    embeddings = torch.cat([class_embeds, patch_embeds], dim=1)
    embeddings = embeddings + self.position_embedding(self.position_ids)

    return embeddings

CLIPVisionModel

Bases: Module, SupportsQuant

Source code in vllm/model_executor/models/clip.py 
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
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
class CLIPVisionModel(nn.Module, SupportsQuant):
    config_class = CLIPVisionConfig
    main_input_name = "pixel_values"
    packed_modules_mapping = {"qkv_proj": ["q_proj", "k_proj", "v_proj"]}

    def __init__(
        self,
        config: CLIPVisionConfig,
        quant_config: Optional[QuantizationConfig] = None,
        *,
        num_hidden_layers_override: Optional[int] = None,
        require_post_norm: Optional[bool] = None,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.vision_model = CLIPVisionTransformer(
            config=config,
            quant_config=quant_config,
            num_hidden_layers_override=num_hidden_layers_override,
            require_post_norm=require_post_norm,
            prefix=f"{prefix}.vision_model")

    def forward(
        self,
        pixel_values: torch.Tensor,
        feature_sample_layers: Optional[list[int]] = None,
    ) -> torch.Tensor:
        return self.vision_model(pixel_values, feature_sample_layers)

    @property
    def device(self):
        return next(self.parameters()).device

    # (TODO) Add prefix argument for filtering out weights to be loaded
    #        ref: https://github.com/vllm-project/vllm/pull/7186#discussion_r1734163986
    def load_weights(self, weights: Iterable[tuple[str,
                                                   torch.Tensor]]) -> set[str]:
        stacked_params_mapping = [
            # (param_name, shard_name, shard_id)
            ("qkv_proj", "q_proj", "q"),
            ("qkv_proj", "k_proj", "k"),
            ("qkv_proj", "v_proj", "v"),
        ]
        params_dict = dict(self.named_parameters())
        loaded_params: set[str] = set()
        layer_count = len(self.vision_model.encoder.layers)

        for name, loaded_weight in weights:
            # post_layernorm is not needed in CLIPVisionModel
            if (name.startswith("vision_model.post_layernorm")
                    and self.vision_model.post_layernorm is None):
                continue

            # omit layers when num_hidden_layers_override is set
            if name.startswith("vision_model.encoder.layers"):
                layer_idx = int(name.split(".")[3])
                if layer_idx >= layer_count:
                    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)

                param = params_dict[name]
                weight_loader = param.weight_loader
                weight_loader(param, loaded_weight, shard_id)
                break
            else:
                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_class class-attribute instance-attribute 

config_class = CLIPVisionConfig

device property 

device

main_input_name class-attribute instance-attribute 

main_input_name = 'pixel_values'

packed_modules_mapping class-attribute instance-attribute 

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

vision_model instance-attribute 

vision_model = CLIPVisionTransformer(
    config=config,
    quant_config=quant_config,
    num_hidden_layers_override=num_hidden_layers_override,
    require_post_norm=require_post_norm,
    prefix=f"{prefix}.vision_model",
)

__init__ 

__init__(
    config: CLIPVisionConfig,
    quant_config: Optional[QuantizationConfig] = None,
    *,
    num_hidden_layers_override: Optional[int] = None,
    require_post_norm: Optional[bool] = None,
    prefix: str = "",
) -> None
Source code in vllm/model_executor/models/clip.py 
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
def __init__(
    self,
    config: CLIPVisionConfig,
    quant_config: Optional[QuantizationConfig] = None,
    *,
    num_hidden_layers_override: Optional[int] = None,
    require_post_norm: Optional[bool] = None,
    prefix: str = "",
) -> None:
    super().__init__()
    self.vision_model = CLIPVisionTransformer(
        config=config,
        quant_config=quant_config,
        num_hidden_layers_override=num_hidden_layers_override,
        require_post_norm=require_post_norm,
        prefix=f"{prefix}.vision_model")

forward 

forward(
    pixel_values: Tensor,
    feature_sample_layers: Optional[list[int]] = None,
) -> Tensor
Source code in vllm/model_executor/models/clip.py 
355
356
357
358
359
360
def forward(
    self,
    pixel_values: torch.Tensor,
    feature_sample_layers: Optional[list[int]] = None,
) -> torch.Tensor:
    return self.vision_model(pixel_values, feature_sample_layers)

load_weights 

load_weights(
    weights: Iterable[tuple[str, Tensor]],
) -> set[str]
Source code in vllm/model_executor/models/clip.py 
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
def load_weights(self, weights: Iterable[tuple[str,
                                               torch.Tensor]]) -> set[str]:
    stacked_params_mapping = [
        # (param_name, shard_name, shard_id)
        ("qkv_proj", "q_proj", "q"),
        ("qkv_proj", "k_proj", "k"),
        ("qkv_proj", "v_proj", "v"),
    ]
    params_dict = dict(self.named_parameters())
    loaded_params: set[str] = set()
    layer_count = len(self.vision_model.encoder.layers)

    for name, loaded_weight in weights:
        # post_layernorm is not needed in CLIPVisionModel
        if (name.startswith("vision_model.post_layernorm")
                and self.vision_model.post_layernorm is None):
            continue

        # omit layers when num_hidden_layers_override is set
        if name.startswith("vision_model.encoder.layers"):
            layer_idx = int(name.split(".")[3])
            if layer_idx >= layer_count:
                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)

            param = params_dict[name]
            weight_loader = param.weight_loader
            weight_loader(param, loaded_weight, shard_id)
            break
        else:
            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

CLIPVisionTransformer

Bases: Module

Source code in vllm/model_executor/models/clip.py 
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
328
329
330
class CLIPVisionTransformer(nn.Module):

    def __init__(
        self,
        config: CLIPVisionConfig,
        quant_config: Optional[QuantizationConfig] = None,
        *,
        num_hidden_layers_override: Optional[int] = None,
        require_post_norm: Optional[bool] = None,
        prefix: str = "",
    ) -> None:
        super().__init__()

        self.config = config
        embed_dim = config.hidden_size

        self.embeddings = CLIPVisionEmbeddings(config)

        # NOTE: This typo of "layrnorm" is not fixed on purpose to match
        # the original transformers code and name of the model weights.
        self.pre_layrnorm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)

        self.encoder = CLIPEncoder(
            config=config,
            quant_config=quant_config,
            num_hidden_layers_override=num_hidden_layers_override,
            prefix=f"{prefix}.encoder",
        )

        num_hidden_layers = config.num_hidden_layers
        if len(self.encoder.layers) > config.num_hidden_layers:
            raise ValueError(
                f"The original encoder only has {num_hidden_layers} "
                f"layers, but you requested {len(self.encoder.layers)} layers."
            )

        # If possible, skip post_layernorm to conserve memory
        if require_post_norm is None:
            require_post_norm = len(self.encoder.layers) == num_hidden_layers

        if require_post_norm:
            self.post_layernorm = nn.LayerNorm(embed_dim,
                                               eps=config.layer_norm_eps)
        else:
            self.post_layernorm = None

    def forward(
        self,
        pixel_values: torch.Tensor,
        feature_sample_layers: Optional[list[int]] = None,
    ) -> torch.Tensor:

        hidden_states = self.embeddings(pixel_values)
        hidden_states = self.pre_layrnorm(hidden_states)

        return_all_hidden_states = feature_sample_layers is not None

        # Produces either the last layer output or all of the hidden states,
        # depending on if we have feature_sample_layers or not
        encoder_outputs = self.encoder(
            inputs_embeds=hidden_states,
            return_all_hidden_states=return_all_hidden_states)

        # Handle post-norm (if applicable) and stacks feature layers if needed
        encoder_outputs = resolve_visual_encoder_outputs(
            encoder_outputs, feature_sample_layers, self.post_layernorm,
            self.config.num_hidden_layers)

        return encoder_outputs

config instance-attribute 

config = config

embeddings instance-attribute 

embeddings = CLIPVisionEmbeddings(config)

encoder instance-attribute 

encoder = CLIPEncoder(
    config=config,
    quant_config=quant_config,
    num_hidden_layers_override=num_hidden_layers_override,
    prefix=f"{prefix}.encoder",
)

post_layernorm instance-attribute 

post_layernorm = LayerNorm(embed_dim, eps=layer_norm_eps)

pre_layrnorm instance-attribute 

pre_layrnorm = LayerNorm(embed_dim, eps=layer_norm_eps)

__init__ 

__init__(
    config: CLIPVisionConfig,
    quant_config: Optional[QuantizationConfig] = None,
    *,
    num_hidden_layers_override: Optional[int] = None,
    require_post_norm: Optional[bool] = None,
    prefix: str = "",
) -> None
Source code in vllm/model_executor/models/clip.py 
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
def __init__(
    self,
    config: CLIPVisionConfig,
    quant_config: Optional[QuantizationConfig] = None,
    *,
    num_hidden_layers_override: Optional[int] = None,
    require_post_norm: Optional[bool] = None,
    prefix: str = "",
) -> None:
    super().__init__()

    self.config = config
    embed_dim = config.hidden_size

    self.embeddings = CLIPVisionEmbeddings(config)

    # NOTE: This typo of "layrnorm" is not fixed on purpose to match
    # the original transformers code and name of the model weights.
    self.pre_layrnorm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)

    self.encoder = CLIPEncoder(
        config=config,
        quant_config=quant_config,
        num_hidden_layers_override=num_hidden_layers_override,
        prefix=f"{prefix}.encoder",
    )

    num_hidden_layers = config.num_hidden_layers
    if len(self.encoder.layers) > config.num_hidden_layers:
        raise ValueError(
            f"The original encoder only has {num_hidden_layers} "
            f"layers, but you requested {len(self.encoder.layers)} layers."
        )

    # If possible, skip post_layernorm to conserve memory
    if require_post_norm is None:
        require_post_norm = len(self.encoder.layers) == num_hidden_layers

    if require_post_norm:
        self.post_layernorm = nn.LayerNorm(embed_dim,
                                           eps=config.layer_norm_eps)
    else:
        self.post_layernorm = None

forward 

forward(
    pixel_values: Tensor,
    feature_sample_layers: Optional[list[int]] = None,
) -> Tensor
Source code in vllm/model_executor/models/clip.py 
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
def forward(
    self,
    pixel_values: torch.Tensor,
    feature_sample_layers: Optional[list[int]] = None,
) -> torch.Tensor:

    hidden_states = self.embeddings(pixel_values)
    hidden_states = self.pre_layrnorm(hidden_states)

    return_all_hidden_states = feature_sample_layers is not None

    # Produces either the last layer output or all of the hidden states,
    # depending on if we have feature_sample_layers or not
    encoder_outputs = self.encoder(
        inputs_embeds=hidden_states,
        return_all_hidden_states=return_all_hidden_states)

    # Handle post-norm (if applicable) and stacks feature layers if needed
    encoder_outputs = resolve_visual_encoder_outputs(
        encoder_outputs, feature_sample_layers, self.post_layernorm,
        self.config.num_hidden_layers)

    return encoder_outputs