Skip to content

vllm.model_executor.models.idefics2_vision_model

PyTorch Idefics2 model.

Idefics2Encoder

Bases: Module

Transformer encoder consisting of config.num_hidden_layers self attention layers. Each layer is a [Idefics2EncoderLayer].

Parameters:

Name Type Description Default
config Idefics2Config

Idefics2Config

 required
Source code in vllm/model_executor/models/idefics2_vision_model.py 
240
241
242
243
244
245
246
247
248
249
250
251
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
class Idefics2Encoder(nn.Module):
    """
    Transformer encoder consisting of `config.num_hidden_layers` self attention
    layers. Each layer is a
    [`Idefics2EncoderLayer`].

    Args:
        config: Idefics2Config
    """

    def __init__(
        self,
        config: Idefics2Config,
        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([
            Idefics2EncoderLayer(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,
    ) -> torch.Tensor:
        r"""
        Args:
            inputs_embeds (torch.Tensor):
                Optionally, instead of passing `input_ids` you can choose to
                directly pass an embedded representation.
                This is useful if you want more control over how to convert
                `input_ids` indices into associated vectorsthan the model's
                internal embedding lookup matrix.
        """
        hidden_states = inputs_embeds
        for encoder_layer in self.layers:
            layer_outputs = encoder_layer(hidden_states)
            hidden_states = layer_outputs
        return hidden_states

config instance-attribute 

config = config

layers instance-attribute 

layers = ModuleList(
    [
        Idefics2EncoderLayer(
            config,
            quant_config=quant_config,
            prefix=f"{prefix}.layers.{layer_idx}",
        )
        for layer_idx in range(num_hidden_layers)
    ]
)

__init__ 

__init__(
    config: Idefics2Config,
    quant_config: Optional[QuantizationConfig] = None,
    *,
    num_hidden_layers_override: Optional[int] = None,
    prefix: str = "",
) -> None
Source code in vllm/model_executor/models/idefics2_vision_model.py 
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
def __init__(
    self,
    config: Idefics2Config,
    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([
        Idefics2EncoderLayer(config,
                             quant_config=quant_config,
                             prefix=f"{prefix}.layers.{layer_idx}")
        for layer_idx in range(num_hidden_layers)
    ])

forward 

forward(inputs_embeds: Tensor) -> Tensor

Parameters:

Name Type Description Default
inputs_embeds Tensor

Optionally, instead of passing input_ids you can choose to directly pass an embedded representation. This is useful if you want more control over how to convert input_ids indices into associated vectorsthan the model's internal embedding lookup matrix.

 required
Source code in vllm/model_executor/models/idefics2_vision_model.py 
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
def forward(
    self,
    inputs_embeds: torch.Tensor,
) -> torch.Tensor:
    r"""
    Args:
        inputs_embeds (torch.Tensor):
            Optionally, instead of passing `input_ids` you can choose to
            directly pass an embedded representation.
            This is useful if you want more control over how to convert
            `input_ids` indices into associated vectorsthan the model's
            internal embedding lookup matrix.
    """
    hidden_states = inputs_embeds
    for encoder_layer in self.layers:
        layer_outputs = encoder_layer(hidden_states)
        hidden_states = layer_outputs
    return hidden_states

Idefics2EncoderLayer

Bases: Module

Source code in vllm/model_executor/models/idefics2_vision_model.py 
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
class Idefics2EncoderLayer(nn.Module):

    def __init__(
        self,
        config: Idefics2Config,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ) -> None:
        super().__init__()
        self.embed_dim = config.hidden_size
        self.self_attn = Idefics2VisionAttention(config,
                                                 quant_config=quant_config,
                                                 prefix=f"{prefix}.self_attn")
        self.layer_norm1 = nn.LayerNorm(self.embed_dim,
                                        eps=config.layer_norm_eps)
        self.mlp = Idefics2VisionMLP(config,
                                     quant_config=quant_config,
                                     prefix=f"{prefix}.mlp")
        self.layer_norm2 = nn.LayerNorm(self.embed_dim,
                                        eps=config.layer_norm_eps)

    def forward(
        self,
        hidden_states: torch.Tensor,
    ) -> torch.Tensor:
        """
        Args:
            hidden_states (`torch.FloatTensor`):
                Input to the layer of shape `(batch, seq_len, embed_dim)`.

        """
        residual = hidden_states
        hidden_states = self.layer_norm1(hidden_states)
        hidden_states = self.self_attn(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

embed_dim instance-attribute 

embed_dim = hidden_size

layer_norm1 instance-attribute 

layer_norm1 = LayerNorm(embed_dim, eps=layer_norm_eps)

layer_norm2 instance-attribute 

layer_norm2 = LayerNorm(embed_dim, eps=layer_norm_eps)

mlp instance-attribute 

mlp = Idefics2VisionMLP(
    config,
    quant_config=quant_config,
    prefix=f"{prefix}.mlp",
)

self_attn instance-attribute 

self_attn = Idefics2VisionAttention(
    config,
    quant_config=quant_config,
    prefix=f"{prefix}.self_attn",
)

__init__ 

__init__(
    config: Idefics2Config,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
) -> None
Source code in vllm/model_executor/models/idefics2_vision_model.py 
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
def __init__(
    self,
    config: Idefics2Config,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
) -> None:
    super().__init__()
    self.embed_dim = config.hidden_size
    self.self_attn = Idefics2VisionAttention(config,
                                             quant_config=quant_config,
                                             prefix=f"{prefix}.self_attn")
    self.layer_norm1 = nn.LayerNorm(self.embed_dim,
                                    eps=config.layer_norm_eps)
    self.mlp = Idefics2VisionMLP(config,
                                 quant_config=quant_config,
                                 prefix=f"{prefix}.mlp")
    self.layer_norm2 = nn.LayerNorm(self.embed_dim,
                                    eps=config.layer_norm_eps)

forward 

forward(hidden_states: Tensor) -> Tensor

Parameters:

Name Type Description Default
hidden_states `torch.FloatTensor`

Input to the layer of shape (batch, seq_len, embed_dim).

 required
Source code in vllm/model_executor/models/idefics2_vision_model.py 
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
def forward(
    self,
    hidden_states: torch.Tensor,
) -> torch.Tensor:
    """
    Args:
        hidden_states (`torch.FloatTensor`):
            Input to the layer of shape `(batch, seq_len, embed_dim)`.

    """
    residual = hidden_states
    hidden_states = self.layer_norm1(hidden_states)
    hidden_states = self.self_attn(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

Idefics2VisionAttention

Bases: Module

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

Source code in vllm/model_executor/models/idefics2_vision_model.py 
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
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
class Idefics2VisionAttention(nn.Module):
    """Multi-headed attention from 'Attention Is All You Need' paper"""

    def __init__(
        self,
        config: Idefics2VisionConfig,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = "",
    ) -> None:
        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(
                f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`:"  # noqa: E501
                f" {self.num_heads}).")
        self.scale = self.head_dim**-0.5
        self.dropout = config.attention_dropout
        self.qkv_proj = QKVParallelLinear(
            self.embed_dim,
            self.head_dim,
            self.num_heads,
            quant_config=quant_config,
            prefix=f"{prefix}.qkv_proj",
        )
        self.out_proj = RowParallelLinear(
            self.embed_dim,
            self.embed_dim,
            bias=True,
            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,
    ) -> torch.Tensor:
        qkv, _ = self.qkv_proj(
            hidden_states
        )  # batch_size, q_len, 3 * num_heads_per_partition * head_dim
        query_states, key_states, value_states = qkv.chunk(3, dim=-1)
        out = self.attn(query_states, key_states, value_states)
        attn_output, _ = self.out_proj(out)
        return attn_output

attn instance-attribute 

attn = MultiHeadAttention(
    num_heads_per_partition, head_dim, scale
)

config instance-attribute 

config = config

dropout instance-attribute 

dropout = attention_dropout

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(
    embed_dim,
    embed_dim,
    bias=True,
    quant_config=quant_config,
    prefix=f"{prefix}.out_proj",
)

qkv_proj instance-attribute 

qkv_proj = QKVParallelLinear(
    embed_dim,
    head_dim,
    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: Idefics2VisionConfig,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
) -> None
Source code in vllm/model_executor/models/idefics2_vision_model.py 
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
144
145
146
147
148
149
150
def __init__(
    self,
    config: Idefics2VisionConfig,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
) -> None:
    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(
            f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`:"  # noqa: E501
            f" {self.num_heads}).")
    self.scale = self.head_dim**-0.5
    self.dropout = config.attention_dropout
    self.qkv_proj = QKVParallelLinear(
        self.embed_dim,
        self.head_dim,
        self.num_heads,
        quant_config=quant_config,
        prefix=f"{prefix}.qkv_proj",
    )
    self.out_proj = RowParallelLinear(
        self.embed_dim,
        self.embed_dim,
        bias=True,
        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) -> Tensor
Source code in vllm/model_executor/models/idefics2_vision_model.py 
152
153
154
155
156
157
158
159
160
161
162
def forward(
    self,
    hidden_states: torch.Tensor,
) -> torch.Tensor:
    qkv, _ = self.qkv_proj(
        hidden_states
    )  # batch_size, q_len, 3 * num_heads_per_partition * head_dim
    query_states, key_states, value_states = qkv.chunk(3, dim=-1)
    out = self.attn(query_states, key_states, value_states)
    attn_output, _ = self.out_proj(out)
    return attn_output

Idefics2VisionEmbeddings

Bases: Module

This is a modified version of siglip.modelign_siglip.SiglipVisionEmbeddings to enable images of variable resolution.

The modifications are adapted from Patch n' Pack: NaViT, a Vision Transformer for any Aspect Ratio and Resolution which allows treating images in their native aspect ratio and without the need to resize them to the same fixed size. In particular, we start from the original pre-trained SigLIP model(which uses images of fixed-size square images) and adapt it by training on images of variable resolutions.

Source code in vllm/model_executor/models/idefics2_vision_model.py 
 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
 98
 99
100
101
102
103
104
105
106
107
108
109
110
class Idefics2VisionEmbeddings(nn.Module):
    """
    This is a modified version of `siglip.modelign_siglip.SiglipVisionEmbeddings
    ` to enable images of variable
    resolution.

    The modifications are adapted from [Patch n' Pack: NaViT, a Vision
    Transformer for any Aspect Ratio and Resolution](https://arxiv.org/abs/2307.06304)
    which allows treating images in their native aspect ratio and without the
    need to resize them to the same fixed size. In particular, we start from the
    original pre-trained SigLIP model(which uses images of fixed-size square
    images) and adapt it by training on images of variable resolutions.
    """

    def __init__(self, config: Idefics2VisionConfig):
        super().__init__()
        self.embed_dim = config.hidden_size
        self.image_size = config.image_size
        self.patch_size = config.patch_size
        self.patch_embedding = nn.Conv2d(
            in_channels=config.num_channels,
            out_channels=self.embed_dim,
            kernel_size=self.patch_size,
            stride=self.patch_size,
            padding="valid",
        )
        self.num_patches_per_side = self.image_size // self.patch_size
        self.num_patches = self.num_patches_per_side**2
        self.num_positions = self.num_patches
        self.position_embedding = nn.Embedding(self.num_positions,
                                               self.embed_dim)

    def forward(self,
                pixel_values: torch.FloatTensor,
                patch_attention_mask: torch.BoolTensor,
                tgt_sizes: Optional[torch.IntTensor] = None) -> torch.Tensor:
        batch_size, _, max_im_h, max_im_w = pixel_values.shape
        target_dtype = self.patch_embedding.weight.dtype
        patch_embeds = self.patch_embedding(pixel_values.to(target_dtype))
        embeddings = patch_embeds.flatten(2).transpose(1, 2)
        max_nb_patches_h, max_nb_patches_w = (
            max_im_h // self.patch_size,
            max_im_w // self.patch_size,
        )
        boundaries = torch.arange(1 / self.num_patches_per_side, 1.0,
                                  1 / self.num_patches_per_side)
        position_ids = torch.full(size=(batch_size,
                                        max_nb_patches_h * max_nb_patches_w),
                                  fill_value=0)

        for batch_idx, p_attn_mask in enumerate(patch_attention_mask):

            if tgt_sizes is not None:
                nb_patches_h = tgt_sizes[batch_idx][0]
                nb_patches_w = tgt_sizes[batch_idx][1]
            else:
                nb_patches_h = p_attn_mask[:, 0].sum()
                nb_patches_w = p_attn_mask[0].sum()
            fractional_coords_h = torch.arange(0, 1 - 1e-6, 1 / nb_patches_h)
            fractional_coords_w = torch.arange(0, 1 - 1e-6, 1 / nb_patches_w)
            bucket_coords_h = torch.bucketize(fractional_coords_h,
                                              boundaries,
                                              right=True)
            bucket_coords_w = torch.bucketize(fractional_coords_w,
                                              boundaries,
                                              right=True)
            pos_ids = (bucket_coords_h[:, None] * self.num_patches_per_side +
                       bucket_coords_w).flatten()
            position_ids[batch_idx][p_attn_mask.view(-1).cpu()] = pos_ids
        position_ids = position_ids.to(self.position_embedding.weight.device)
        embeddings = embeddings + self.position_embedding(position_ids)
        return embeddings

embed_dim instance-attribute 

embed_dim = hidden_size

image_size instance-attribute 

image_size = image_size

num_patches instance-attribute 

num_patches = num_patches_per_side ** 2

num_patches_per_side instance-attribute 

num_patches_per_side = image_size // patch_size

num_positions instance-attribute 

num_positions = num_patches

patch_embedding instance-attribute 

patch_embedding = Conv2d(
    in_channels=num_channels,
    out_channels=embed_dim,
    kernel_size=patch_size,
    stride=patch_size,
    padding="valid",
)

patch_size instance-attribute 

patch_size = patch_size

position_embedding instance-attribute 

position_embedding = Embedding(num_positions, embed_dim)

__init__ 

__init__(config: Idefics2VisionConfig)
Source code in vllm/model_executor/models/idefics2_vision_model.py 
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
def __init__(self, config: Idefics2VisionConfig):
    super().__init__()
    self.embed_dim = config.hidden_size
    self.image_size = config.image_size
    self.patch_size = config.patch_size
    self.patch_embedding = nn.Conv2d(
        in_channels=config.num_channels,
        out_channels=self.embed_dim,
        kernel_size=self.patch_size,
        stride=self.patch_size,
        padding="valid",
    )
    self.num_patches_per_side = self.image_size // self.patch_size
    self.num_patches = self.num_patches_per_side**2
    self.num_positions = self.num_patches
    self.position_embedding = nn.Embedding(self.num_positions,
                                           self.embed_dim)

forward 

forward(
    pixel_values: FloatTensor,
    patch_attention_mask: BoolTensor,
    tgt_sizes: Optional[IntTensor] = None,
) -> Tensor
Source code in vllm/model_executor/models/idefics2_vision_model.py 
 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
 98
 99
100
101
102
103
104
105
106
107
108
109
110
def forward(self,
            pixel_values: torch.FloatTensor,
            patch_attention_mask: torch.BoolTensor,
            tgt_sizes: Optional[torch.IntTensor] = None) -> torch.Tensor:
    batch_size, _, max_im_h, max_im_w = pixel_values.shape
    target_dtype = self.patch_embedding.weight.dtype
    patch_embeds = self.patch_embedding(pixel_values.to(target_dtype))
    embeddings = patch_embeds.flatten(2).transpose(1, 2)
    max_nb_patches_h, max_nb_patches_w = (
        max_im_h // self.patch_size,
        max_im_w // self.patch_size,
    )
    boundaries = torch.arange(1 / self.num_patches_per_side, 1.0,
                              1 / self.num_patches_per_side)
    position_ids = torch.full(size=(batch_size,
                                    max_nb_patches_h * max_nb_patches_w),
                              fill_value=0)

    for batch_idx, p_attn_mask in enumerate(patch_attention_mask):

        if tgt_sizes is not None:
            nb_patches_h = tgt_sizes[batch_idx][0]
            nb_patches_w = tgt_sizes[batch_idx][1]
        else:
            nb_patches_h = p_attn_mask[:, 0].sum()
            nb_patches_w = p_attn_mask[0].sum()
        fractional_coords_h = torch.arange(0, 1 - 1e-6, 1 / nb_patches_h)
        fractional_coords_w = torch.arange(0, 1 - 1e-6, 1 / nb_patches_w)
        bucket_coords_h = torch.bucketize(fractional_coords_h,
                                          boundaries,
                                          right=True)
        bucket_coords_w = torch.bucketize(fractional_coords_w,
                                          boundaries,
                                          right=True)
        pos_ids = (bucket_coords_h[:, None] * self.num_patches_per_side +
                   bucket_coords_w).flatten()
        position_ids[batch_idx][p_attn_mask.view(-1).cpu()] = pos_ids
    position_ids = position_ids.to(self.position_embedding.weight.device)
    embeddings = embeddings + self.position_embedding(position_ids)
    return embeddings

Idefics2VisionMLP

Bases: Module

Source code in vllm/model_executor/models/idefics2_vision_model.py 
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
class Idefics2VisionMLP(nn.Module):

    def __init__(
        self,
        config: Idefics2VisionConfig,
        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: Idefics2VisionConfig,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
) -> None
Source code in vllm/model_executor/models/idefics2_vision_model.py 
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
def __init__(
    self,
    config: Idefics2VisionConfig,
    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/idefics2_vision_model.py 
191
192
193
194
195
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

Idefics2VisionTransformer

Bases: Module

Source code in vllm/model_executor/models/idefics2_vision_model.py 
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
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
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
class Idefics2VisionTransformer(nn.Module):

    def __init__(
        self,
        config: Idefics2VisionConfig,
        quant_config: Optional[QuantizationConfig] = None,
        *,
        num_hidden_layers_override: Optional[int] = None,
        require_post_norm: bool = True,
        prefix: str = "",
    ) -> None:
        super().__init__()

        embed_dim = config.hidden_size
        self.config = config
        self.embeddings = Idefics2VisionEmbeddings(config)
        self.encoder = Idefics2Encoder(
            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."
            )

        self.require_post_norm = require_post_norm
        self.post_layernorm = nn.LayerNorm(
            embed_dim,
            eps=config.layer_norm_eps,
        ) if require_post_norm else nn.Identity()

    def get_input_embeddings(self):
        return self.embeddings

    def forward(
        self,
        pixel_values,
        patch_attention_mask: Optional[torch.BoolTensor] = None,
        tgt_sizes: Optional[torch.IntTensor] = None,
    ) -> torch.Tensor:
        hidden_states = self.embeddings(
            pixel_values=pixel_values,
            patch_attention_mask=patch_attention_mask,
            tgt_sizes=tgt_sizes,
        )
        encoder_outputs = self.encoder(hidden_states)
        last_hidden_state = self.post_layernorm(encoder_outputs)
        return last_hidden_state

    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.encoder.layers)

        for name, loaded_weight in weights:
            # skip pooling header
            if name.startswith("head."):
                continue

            # post_layernorm is optional
            if (name.startswith("post_layernorm.")
                    and not self.require_post_norm):
                continue

            # omit layers when num_hidden_layers_override is set
            if name.startswith("encoder.layers."):
                layer_idx = int(name.split(".")[2])
                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 instance-attribute 

config = config

embeddings instance-attribute 

embeddings = Idefics2VisionEmbeddings(config)

encoder instance-attribute 

encoder = Idefics2Encoder(
    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)
    if require_post_norm
    else Identity()
)

require_post_norm instance-attribute 

require_post_norm = require_post_norm

__init__ 

__init__(
    config: Idefics2VisionConfig,
    quant_config: Optional[QuantizationConfig] = None,
    *,
    num_hidden_layers_override: Optional[int] = None,
    require_post_norm: bool = True,
    prefix: str = "",
) -> None
Source code in vllm/model_executor/models/idefics2_vision_model.py 
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 __init__(
    self,
    config: Idefics2VisionConfig,
    quant_config: Optional[QuantizationConfig] = None,
    *,
    num_hidden_layers_override: Optional[int] = None,
    require_post_norm: bool = True,
    prefix: str = "",
) -> None:
    super().__init__()

    embed_dim = config.hidden_size
    self.config = config
    self.embeddings = Idefics2VisionEmbeddings(config)
    self.encoder = Idefics2Encoder(
        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."
        )

    self.require_post_norm = require_post_norm
    self.post_layernorm = nn.LayerNorm(
        embed_dim,
        eps=config.layer_norm_eps,
    ) if require_post_norm else nn.Identity()

forward 

forward(
    pixel_values,
    patch_attention_mask: Optional[BoolTensor] = None,
    tgt_sizes: Optional[IntTensor] = None,
) -> Tensor
Source code in vllm/model_executor/models/idefics2_vision_model.py 
332
333
334
335
336
337
338
339
340
341
342
343
344
345
def forward(
    self,
    pixel_values,
    patch_attention_mask: Optional[torch.BoolTensor] = None,
    tgt_sizes: Optional[torch.IntTensor] = None,
) -> torch.Tensor:
    hidden_states = self.embeddings(
        pixel_values=pixel_values,
        patch_attention_mask=patch_attention_mask,
        tgt_sizes=tgt_sizes,
    )
    encoder_outputs = self.encoder(hidden_states)
    last_hidden_state = self.post_layernorm(encoder_outputs)
    return last_hidden_state

get_input_embeddings 

get_input_embeddings()
Source code in vllm/model_executor/models/idefics2_vision_model.py 
329
330
def get_input_embeddings(self):
    return self.embeddings

load_weights 

load_weights(
    weights: Iterable[tuple[str, Tensor]],
) -> set[str]
Source code in vllm/model_executor/models/idefics2_vision_model.py 
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
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.encoder.layers)

    for name, loaded_weight in weights:
        # skip pooling header
        if name.startswith("head."):
            continue

        # post_layernorm is optional
        if (name.startswith("post_layernorm.")
                and not self.require_post_norm):
            continue

        # omit layers when num_hidden_layers_override is set
        if name.startswith("encoder.layers."):
            layer_idx = int(name.split(".")[2])
            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