Skip to content

vllm.model_executor.models.glm4v

Inference-only CogAgent model compatible with THUDM weights.

EVA2CLIPAttention

Bases: Module

Source code in vllm/model_executor/models/glm4v.py 
 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
class EVA2CLIPAttention(nn.Module):

    def __init__(
        self,
        config,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = '',
    ):
        super().__init__()
        self.hidden_size = config.hidden_size
        self.tp_size = get_tensor_model_parallel_world_size()
        self.num_heads_per_rank = config.num_heads // self.tp_size
        self.head_dim = config.hidden_size // config.num_heads
        self.scale = self.head_dim**-0.5

        self.query_key_value = QKVParallelLinear(
            config.hidden_size,
            self.head_dim,
            config.num_heads,
            quant_config=quant_config,
            prefix=f"{prefix}.query_key_value",
        )
        self.dense = RowParallelLinear(
            config.hidden_size,
            config.hidden_size,
            quant_config=quant_config,
            prefix=f"{prefix}.dense",
        )

        self.attn = MultiHeadAttention(self.num_heads_per_rank, self.head_dim,
                                       self.scale)
        self.output_dropout = torch.nn.Dropout(config.dropout_prob)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        qkv, _ = self.query_key_value(x)  # B, L, 3 * H * D
        q, k, v = qkv.chunk(3, dim=-1)

        out = self.attn(q, k, v)
        output, _ = self.dense(out)
        output = self.output_dropout(output)
        return output

attn instance-attribute 

attn = MultiHeadAttention(
    num_heads_per_rank, head_dim, scale
)

dense instance-attribute 

dense = RowParallelLinear(
    hidden_size,
    hidden_size,
    quant_config=quant_config,
    prefix=f"{prefix}.dense",
)

head_dim instance-attribute 

head_dim = hidden_size // num_heads

hidden_size instance-attribute 

hidden_size = hidden_size

num_heads_per_rank instance-attribute 

num_heads_per_rank = num_heads // tp_size

output_dropout instance-attribute 

output_dropout = Dropout(dropout_prob)

query_key_value instance-attribute 

query_key_value = QKVParallelLinear(
    hidden_size,
    head_dim,
    num_heads,
    quant_config=quant_config,
    prefix=f"{prefix}.query_key_value",
)

scale instance-attribute 

scale = head_dim ** -0.5

tp_size instance-attribute 

tp_size = get_tensor_model_parallel_world_size()

__init__ 

__init__(
    config,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
)
Source code in vllm/model_executor/models/glm4v.py 
 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
def __init__(
    self,
    config,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = '',
):
    super().__init__()
    self.hidden_size = config.hidden_size
    self.tp_size = get_tensor_model_parallel_world_size()
    self.num_heads_per_rank = config.num_heads // self.tp_size
    self.head_dim = config.hidden_size // config.num_heads
    self.scale = self.head_dim**-0.5

    self.query_key_value = QKVParallelLinear(
        config.hidden_size,
        self.head_dim,
        config.num_heads,
        quant_config=quant_config,
        prefix=f"{prefix}.query_key_value",
    )
    self.dense = RowParallelLinear(
        config.hidden_size,
        config.hidden_size,
        quant_config=quant_config,
        prefix=f"{prefix}.dense",
    )

    self.attn = MultiHeadAttention(self.num_heads_per_rank, self.head_dim,
                                   self.scale)
    self.output_dropout = torch.nn.Dropout(config.dropout_prob)

forward 

forward(x: Tensor) -> Tensor
Source code in vllm/model_executor/models/glm4v.py 
119
120
121
122
123
124
125
126
def forward(self, x: torch.Tensor) -> torch.Tensor:
    qkv, _ = self.query_key_value(x)  # B, L, 3 * H * D
    q, k, v = qkv.chunk(3, dim=-1)

    out = self.attn(q, k, v)
    output, _ = self.dense(out)
    output = self.output_dropout(output)
    return output

EVA2CLIPGLU

Bases: Module

Source code in vllm/model_executor/models/glm4v.py 
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
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
class EVA2CLIPGLU(nn.Module):

    def __init__(
        self,
        config,
        in_features,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = '',
    ):
        """
        The original implementation is the same as:
        ```python
        self.dense_h_to_4h = ColumnParallelLinear(
            config.hidden_size,
            config.ffn_hidden_size,
            bias=False,
            quant_config=quant_config
        )

        self.gate_proj = ColumnParallelLinear(
            config.hidden_size,
            config.ffn_hidden_size,
            bias=False,
            quant_config=quant_config
        )
        ```
        ```
        gate_proj_output, _ = self.gate_proj(x)
        dense_h_to_4h_output, _ = self.dense_h_to_4h(x)
        x = torch.cat([gate_proj_output, dense_h_to_4h_output], dim=-1)
        ```

        We merge two ColumnParallelLinear into one MergedColumnParallelLinear:
        ```
        self.merged_proj = MergedColumnParallelLinear(
            config.hidden_size,
            [config.ffn_hidden_size] * 2,
            bias=False,
            quant_config=quant_config
        )
        ```
        ```
        x, _ = self.merged_proj(x)
        ```
        """
        super().__init__()
        self.linear_proj = ReplicatedLinear(in_features,
                                            config.hidden_size,
                                            bias=False,
                                            quant_config=quant_config,
                                            prefix=f"{prefix}.linear_proj")
        self.norm1 = nn.LayerNorm(config.hidden_size)
        self.act1 = nn.GELU()
        self.act2 = SiluAndMul()

        self.merged_proj = MergedColumnParallelLinear(
            config.hidden_size, [config.ffn_hidden_size] * 2,
            bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.merged_proj")

        self.dense_4h_to_h = RowParallelLinear(
            config.ffn_hidden_size,
            config.hidden_size,
            bias=False,
            quant_config=quant_config,
            prefix=f"{prefix}.dense_4h_to_h")

    def forward(self, x):
        x, _ = self.linear_proj(x)
        x = self.act1(self.norm1(x))
        x, _ = self.merged_proj(x)
        x = self.act2(x)
        x, _ = self.dense_4h_to_h(x)
        return x

act1 instance-attribute 

act1 = GELU()

act2 instance-attribute 

act2 = SiluAndMul()

dense_4h_to_h instance-attribute 

dense_4h_to_h = RowParallelLinear(
    ffn_hidden_size,
    hidden_size,
    bias=False,
    quant_config=quant_config,
    prefix=f"{prefix}.dense_4h_to_h",
)

linear_proj instance-attribute 

linear_proj = ReplicatedLinear(
    in_features,
    hidden_size,
    bias=False,
    quant_config=quant_config,
    prefix=f"{prefix}.linear_proj",
)

merged_proj instance-attribute 

merged_proj = MergedColumnParallelLinear(
    hidden_size,
    [ffn_hidden_size] * 2,
    bias=False,
    quant_config=quant_config,
    prefix=f"{prefix}.merged_proj",
)

norm1 instance-attribute 

norm1 = LayerNorm(hidden_size)

__init__ 

__init__(
    config,
    in_features,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
)

The original implementation is the same as: 

self.dense_h_to_4h = ColumnParallelLinear(
    config.hidden_size,
    config.ffn_hidden_size,
    bias=False,
    quant_config=quant_config
)

self.gate_proj = ColumnParallelLinear(
    config.hidden_size,
    config.ffn_hidden_size,
    bias=False,
    quant_config=quant_config
)

gate_proj_output, _ = self.gate_proj(x)
dense_h_to_4h_output, _ = self.dense_h_to_4h(x)
x = torch.cat([gate_proj_output, dense_h_to_4h_output], dim=-1)

We merge two ColumnParallelLinear into one MergedColumnParallelLinear: 

self.merged_proj = MergedColumnParallelLinear(
    config.hidden_size,
    [config.ffn_hidden_size] * 2,
    bias=False,
    quant_config=quant_config
)

x, _ = self.merged_proj(x)

Source code in vllm/model_executor/models/glm4v.py 
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
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
def __init__(
    self,
    config,
    in_features,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = '',
):
    """
    The original implementation is the same as:
    ```python
    self.dense_h_to_4h = ColumnParallelLinear(
        config.hidden_size,
        config.ffn_hidden_size,
        bias=False,
        quant_config=quant_config
    )

    self.gate_proj = ColumnParallelLinear(
        config.hidden_size,
        config.ffn_hidden_size,
        bias=False,
        quant_config=quant_config
    )
    ```
    ```
    gate_proj_output, _ = self.gate_proj(x)
    dense_h_to_4h_output, _ = self.dense_h_to_4h(x)
    x = torch.cat([gate_proj_output, dense_h_to_4h_output], dim=-1)
    ```

    We merge two ColumnParallelLinear into one MergedColumnParallelLinear:
    ```
    self.merged_proj = MergedColumnParallelLinear(
        config.hidden_size,
        [config.ffn_hidden_size] * 2,
        bias=False,
        quant_config=quant_config
    )
    ```
    ```
    x, _ = self.merged_proj(x)
    ```
    """
    super().__init__()
    self.linear_proj = ReplicatedLinear(in_features,
                                        config.hidden_size,
                                        bias=False,
                                        quant_config=quant_config,
                                        prefix=f"{prefix}.linear_proj")
    self.norm1 = nn.LayerNorm(config.hidden_size)
    self.act1 = nn.GELU()
    self.act2 = SiluAndMul()

    self.merged_proj = MergedColumnParallelLinear(
        config.hidden_size, [config.ffn_hidden_size] * 2,
        bias=False,
        quant_config=quant_config,
        prefix=f"{prefix}.merged_proj")

    self.dense_4h_to_h = RowParallelLinear(
        config.ffn_hidden_size,
        config.hidden_size,
        bias=False,
        quant_config=quant_config,
        prefix=f"{prefix}.dense_4h_to_h")

forward 

forward(x)
Source code in vllm/model_executor/models/glm4v.py 
281
282
283
284
285
286
287
def forward(self, x):
    x, _ = self.linear_proj(x)
    x = self.act1(self.norm1(x))
    x, _ = self.merged_proj(x)
    x = self.act2(x)
    x, _ = self.dense_4h_to_h(x)
    return x

EVA2CLIPMLP

Bases: Module

Source code in vllm/model_executor/models/glm4v.py 
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
class EVA2CLIPMLP(nn.Module):

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

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

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,
    quant_config=quant_config,
    prefix=f"{prefix}.fc1",
)

fc2 instance-attribute 

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

__init__ 

__init__(
    config,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
)
Source code in vllm/model_executor/models/glm4v.py 
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
def __init__(
    self,
    config,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = '',
):
    super().__init__()
    self.config = config
    self.activation_fn = get_act_fn(config.hidden_act)
    self.fc1 = ColumnParallelLinear(
        config.hidden_size,
        config.intermediate_size,
        quant_config=quant_config,
        prefix=f"{prefix}.fc1",
    )
    self.fc2 = RowParallelLinear(
        config.intermediate_size,
        config.hidden_size,
        quant_config=quant_config,
        prefix=f"{prefix}.fc2",
    )

forward 

forward(x: Tensor) -> Tensor
Source code in vllm/model_executor/models/glm4v.py 
153
154
155
156
157
def forward(self, x: torch.Tensor) -> torch.Tensor:
    x, _ = self.fc1(x)
    x = self.activation_fn(x)
    x, _ = self.fc2(x)
    return x

EVA2CLIPModel

Bases: Module

Source code in vllm/model_executor/models/glm4v.py 
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
331
332
333
334
335
336
337
338
339
340
341
class EVA2CLIPModel(nn.Module):

    def __init__(
        self,
        config,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = '',
    ):
        super().__init__()
        vision_config = Namespace(**config.vision_config)
        self.patch_embedding = EVA2CLIPPatchEmbedding(vision_config)
        self.transformer = EVA2CLIPTransformer(vision_config,
                                               quant_config=quant_config,
                                               prefix=f"{prefix}.transformer")
        self.linear_proj = EVA2CLIPGLU(config,
                                       in_features=config.hidden_size,
                                       quant_config=quant_config,
                                       prefix=f"{prefix}.linear_proj")
        self.conv = nn.Conv2d(in_channels=vision_config.hidden_size,
                              out_channels=config.hidden_size,
                              kernel_size=2,
                              stride=2)
        self.boi = nn.Parameter(torch.zeros(1, 1, config.hidden_size))
        self.eoi = nn.Parameter(torch.zeros(1, 1, config.hidden_size))
        self.scaling_factor = vision_config.scaling_factor

    def forward(self, images: torch.Tensor) -> torch.Tensor:
        """
        Parameters:
        images : torch.Tensor
            Input image tensor with shape (B, C, H, W)

        Returns:
        torch.Tensor
            Transformed tensor with shape (B, L, D)
        """
        x = self.patch_embedding(images)
        x = self.transformer(x)
        x = x[:, 1:]

        b, s, h = x.shape
        grid_size = int(s**0.5)
        x = x.view(b, grid_size, grid_size, h).permute(0, 3, 1, 2)
        x = self.conv(x)

        x = x.flatten(2).transpose(1, 2)
        x = self.linear_proj(x)
        boi = self.boi.expand(x.shape[0], -1, -1)
        eoi = self.eoi.expand(x.shape[0], -1, -1)
        x = torch.cat((boi, x, eoi), dim=1)
        x = x / self.scaling_factor
        return x

boi instance-attribute 

boi = Parameter(zeros(1, 1, hidden_size))

conv instance-attribute 

conv = Conv2d(
    in_channels=hidden_size,
    out_channels=hidden_size,
    kernel_size=2,
    stride=2,
)

eoi instance-attribute 

eoi = Parameter(zeros(1, 1, hidden_size))

linear_proj instance-attribute 

linear_proj = EVA2CLIPGLU(
    config,
    in_features=hidden_size,
    quant_config=quant_config,
    prefix=f"{prefix}.linear_proj",
)

patch_embedding instance-attribute 

patch_embedding = EVA2CLIPPatchEmbedding(vision_config)

scaling_factor instance-attribute 

scaling_factor = scaling_factor

transformer instance-attribute 

transformer = EVA2CLIPTransformer(
    vision_config,
    quant_config=quant_config,
    prefix=f"{prefix}.transformer",
)

__init__ 

__init__(
    config,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
)
Source code in vllm/model_executor/models/glm4v.py 
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
def __init__(
    self,
    config,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = '',
):
    super().__init__()
    vision_config = Namespace(**config.vision_config)
    self.patch_embedding = EVA2CLIPPatchEmbedding(vision_config)
    self.transformer = EVA2CLIPTransformer(vision_config,
                                           quant_config=quant_config,
                                           prefix=f"{prefix}.transformer")
    self.linear_proj = EVA2CLIPGLU(config,
                                   in_features=config.hidden_size,
                                   quant_config=quant_config,
                                   prefix=f"{prefix}.linear_proj")
    self.conv = nn.Conv2d(in_channels=vision_config.hidden_size,
                          out_channels=config.hidden_size,
                          kernel_size=2,
                          stride=2)
    self.boi = nn.Parameter(torch.zeros(1, 1, config.hidden_size))
    self.eoi = nn.Parameter(torch.zeros(1, 1, config.hidden_size))
    self.scaling_factor = vision_config.scaling_factor

forward 

forward(images: Tensor) -> Tensor

images : torch.Tensor Input image tensor with shape (B, C, H, W)

torch.Tensor Transformed tensor with shape (B, L, D)

Source code in vllm/model_executor/models/glm4v.py 
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
def forward(self, images: torch.Tensor) -> torch.Tensor:
    """
    Parameters:
    images : torch.Tensor
        Input image tensor with shape (B, C, H, W)

    Returns:
    torch.Tensor
        Transformed tensor with shape (B, L, D)
    """
    x = self.patch_embedding(images)
    x = self.transformer(x)
    x = x[:, 1:]

    b, s, h = x.shape
    grid_size = int(s**0.5)
    x = x.view(b, grid_size, grid_size, h).permute(0, 3, 1, 2)
    x = self.conv(x)

    x = x.flatten(2).transpose(1, 2)
    x = self.linear_proj(x)
    boi = self.boi.expand(x.shape[0], -1, -1)
    eoi = self.eoi.expand(x.shape[0], -1, -1)
    x = torch.cat((boi, x, eoi), dim=1)
    x = x / self.scaling_factor
    return x

EVA2CLIPPatchEmbedding

Bases: Module

Source code in vllm/model_executor/models/glm4v.py 
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
class EVA2CLIPPatchEmbedding(nn.Module):

    def __init__(self, config):
        super().__init__()
        self.proj = nn.Conv2d(config.in_channels,
                              config.hidden_size,
                              kernel_size=config.patch_size,
                              stride=config.patch_size)
        self.cls_embedding = nn.Parameter(torch.zeros(1, config.hidden_size))
        self.position_embedding = nn.Embedding(config.num_positions,
                                               config.hidden_size)

    def forward(self, images: torch.Tensor) -> torch.Tensor:
        """
        Parameters:
        images : torch.Tensor
            Input image tensor with shape (B, C, H, W)

        Returns:
        torch.Tensor
            Transformed tensor with shape (B, L, D)
        """
        images = images.to(device=self.proj.weight.device,
                           dtype=self.proj.weight.dtype)
        x = self.proj(images)
        x = x.flatten(2).transpose(1, 2)
        cls_token = self.cls_embedding.expand(x.shape[0], -1, -1)
        x = torch.cat((cls_token, x), dim=1)
        x += self.position_embedding.weight.unsqueeze(0)
        return x

cls_embedding instance-attribute 

cls_embedding = Parameter(zeros(1, hidden_size))

position_embedding instance-attribute 

position_embedding = Embedding(num_positions, hidden_size)

proj instance-attribute 

proj = Conv2d(
    in_channels,
    hidden_size,
    kernel_size=patch_size,
    stride=patch_size,
)

__init__ 

__init__(config)
Source code in vllm/model_executor/models/glm4v.py 
56
57
58
59
60
61
62
63
64
def __init__(self, config):
    super().__init__()
    self.proj = nn.Conv2d(config.in_channels,
                          config.hidden_size,
                          kernel_size=config.patch_size,
                          stride=config.patch_size)
    self.cls_embedding = nn.Parameter(torch.zeros(1, config.hidden_size))
    self.position_embedding = nn.Embedding(config.num_positions,
                                           config.hidden_size)

forward 

forward(images: Tensor) -> Tensor

images : torch.Tensor Input image tensor with shape (B, C, H, W)

torch.Tensor Transformed tensor with shape (B, L, D)

Source code in vllm/model_executor/models/glm4v.py 
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
def forward(self, images: torch.Tensor) -> torch.Tensor:
    """
    Parameters:
    images : torch.Tensor
        Input image tensor with shape (B, C, H, W)

    Returns:
    torch.Tensor
        Transformed tensor with shape (B, L, D)
    """
    images = images.to(device=self.proj.weight.device,
                       dtype=self.proj.weight.dtype)
    x = self.proj(images)
    x = x.flatten(2).transpose(1, 2)
    cls_token = self.cls_embedding.expand(x.shape[0], -1, -1)
    x = torch.cat((cls_token, x), dim=1)
    x += self.position_embedding.weight.unsqueeze(0)
    return x

EVA2CLIPTransformer

Bases: Module

Source code in vllm/model_executor/models/glm4v.py 
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
class EVA2CLIPTransformer(nn.Module):

    def __init__(
        self,
        config,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = '',
    ):
        super().__init__()
        self.layers = nn.ModuleList([
            EVA2CLIPTransformerLayer(config,
                                     quant_config=quant_config,
                                     prefix=f"{prefix}.layers.{layer_idx}")
            for layer_idx in range(config.num_hidden_layers)
        ])

    def forward(self, hidden_states):
        for layer_module in self.layers:
            hidden_states = layer_module(hidden_states)
        return hidden_states

layers instance-attribute 

layers = ModuleList(
    [
        EVA2CLIPTransformerLayer(
            config,
            quant_config=quant_config,
            prefix=f"{prefix}.layers.{layer_idx}",
        )
        for layer_idx in range(num_hidden_layers)
    ]
)

__init__ 

__init__(
    config,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
)
Source code in vllm/model_executor/models/glm4v.py 
193
194
195
196
197
198
199
200
201
202
203
204
205
def __init__(
    self,
    config,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = '',
):
    super().__init__()
    self.layers = nn.ModuleList([
        EVA2CLIPTransformerLayer(config,
                                 quant_config=quant_config,
                                 prefix=f"{prefix}.layers.{layer_idx}")
        for layer_idx in range(config.num_hidden_layers)
    ])

forward 

forward(hidden_states)
Source code in vllm/model_executor/models/glm4v.py 
207
208
209
210
def forward(self, hidden_states):
    for layer_module in self.layers:
        hidden_states = layer_module(hidden_states)
    return hidden_states

EVA2CLIPTransformerLayer

Bases: Module

Source code in vllm/model_executor/models/glm4v.py 
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
187
188
class EVA2CLIPTransformerLayer(nn.Module):

    def __init__(
        self,
        config,
        quant_config: Optional[QuantizationConfig] = None,
        prefix: str = '',
    ):
        super().__init__()
        self.input_layernorm = LayerNorm(config.hidden_size,
                                         eps=config.layer_norm_eps)
        self.attention = EVA2CLIPAttention(config,
                                           quant_config=quant_config,
                                           prefix=f"{prefix}.attention")
        self.mlp = EVA2CLIPMLP(config,
                               quant_config=quant_config,
                               prefix=f"{prefix}.mlp")
        self.post_attention_layernorm = LayerNorm(config.hidden_size,
                                                  eps=config.layer_norm_eps)

    def forward(self, hidden_states):
        attention_input = hidden_states
        attention_output = self.input_layernorm(
            self.attention(attention_input))
        hidden_states = attention_input + attention_output
        mlp_input = hidden_states
        mlp_output = self.post_attention_layernorm(self.mlp(mlp_input))
        output = mlp_input + mlp_output
        return output

attention instance-attribute 

attention = EVA2CLIPAttention(
    config,
    quant_config=quant_config,
    prefix=f"{prefix}.attention",
)

input_layernorm instance-attribute 

input_layernorm = LayerNorm(hidden_size, eps=layer_norm_eps)

mlp instance-attribute 

mlp = EVA2CLIPMLP(
    config,
    quant_config=quant_config,
    prefix=f"{prefix}.mlp",
)

post_attention_layernorm instance-attribute 

post_attention_layernorm = LayerNorm(
    hidden_size, eps=layer_norm_eps
)

__init__ 

__init__(
    config,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = "",
)
Source code in vllm/model_executor/models/glm4v.py 
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
def __init__(
    self,
    config,
    quant_config: Optional[QuantizationConfig] = None,
    prefix: str = '',
):
    super().__init__()
    self.input_layernorm = LayerNorm(config.hidden_size,
                                     eps=config.layer_norm_eps)
    self.attention = EVA2CLIPAttention(config,
                                       quant_config=quant_config,
                                       prefix=f"{prefix}.attention")
    self.mlp = EVA2CLIPMLP(config,
                           quant_config=quant_config,
                           prefix=f"{prefix}.mlp")
    self.post_attention_layernorm = LayerNorm(config.hidden_size,
                                              eps=config.layer_norm_eps)

forward 

forward(hidden_states)
Source code in vllm/model_executor/models/glm4v.py 
180
181
182
183
184
185
186
187
188
def forward(self, hidden_states):
    attention_input = hidden_states
    attention_output = self.input_layernorm(
        self.attention(attention_input))
    hidden_states = attention_input + attention_output
    mlp_input = hidden_states
    mlp_output = self.post_attention_layernorm(self.mlp(mlp_input))
    output = mlp_input + mlp_output
    return output

GLM4VDummyInputsBuilder

Bases: BaseDummyInputsBuilder[GLM4VProcessingInfo]

Source code in vllm/model_executor/models/glm4v.py 
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
class GLM4VDummyInputsBuilder(BaseDummyInputsBuilder[GLM4VProcessingInfo]):

    def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str:
        num_images = mm_counts.get("image", 0)

        base_text = "<|begin_of_image|><|endoftext|><|end_of_image|>"

        return base_text * num_images

    def get_dummy_mm_data(
        self,
        seq_len: int,
        mm_counts: Mapping[str, int],
    ) -> MultiModalDataDict:
        hf_config = self.info.get_hf_config()
        vision_config = hf_config.vision_config

        target_width = target_height = vision_config["image_size"]
        num_images = mm_counts.get("image", 0)

        return {
            "image":
            self._get_dummy_images(width=target_width,
                                   height=target_height,
                                   num_images=num_images)
        }

get_dummy_mm_data 

get_dummy_mm_data(
    seq_len: int, mm_counts: Mapping[str, int]
) -> MultiModalDataDict
Source code in vllm/model_executor/models/glm4v.py 
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
def get_dummy_mm_data(
    self,
    seq_len: int,
    mm_counts: Mapping[str, int],
) -> MultiModalDataDict:
    hf_config = self.info.get_hf_config()
    vision_config = hf_config.vision_config

    target_width = target_height = vision_config["image_size"]
    num_images = mm_counts.get("image", 0)

    return {
        "image":
        self._get_dummy_images(width=target_width,
                               height=target_height,
                               num_images=num_images)
    }

get_dummy_text 

get_dummy_text(mm_counts: Mapping[str, int]) -> str
Source code in vllm/model_executor/models/glm4v.py 
451
452
453
454
455
456
def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str:
    num_images = mm_counts.get("image", 0)

    base_text = "<|begin_of_image|><|endoftext|><|end_of_image|>"

    return base_text * num_images

GLM4VForCausalLM

Bases: ChatGLMBaseModel, SupportsLoRA, SupportsPP, SupportsMultiModal

Source code in vllm/model_executor/models/glm4v.py 
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
@MULTIMODAL_REGISTRY.register_processor(GLM4VMultiModalProcessor,
                                        info=GLM4VProcessingInfo,
                                        dummy_inputs=GLM4VDummyInputsBuilder)
class GLM4VForCausalLM(ChatGLMBaseModel, SupportsLoRA, SupportsPP,
                       SupportsMultiModal):

    packed_modules_mapping = {
        "query_key_value": ["query_key_value"],
        "dense_h_to_4h": ["dense_h_to_4h"],
        "merged_proj": ["gate_proj", "dense_h_to_4h"]
    }

    def get_mm_mapping(self) -> MultiModelKeys:
        """
        Get the module prefix in multimodal models
        """
        return MultiModelKeys.from_string_field(
            language_model="transformer.encoder",
            connector="transformer.vision.linear_proj",
            tower_model="transformer.vision.transformer")

    @classmethod
    def get_placeholder_str(cls, modality: str, i: int) -> Optional[str]:
        if modality.startswith("image"):
            return "<|begin_of_image|><|endoftext|><|end_of_image|>"

        raise ValueError("Only image modality is supported")

    def __init__(
        self,
        *,
        vllm_config: VllmConfig,
        prefix: str = "",
        transformer_type: type[GLM4VModel] = GLM4VModel,
    ) -> None:
        super().__init__(
            vllm_config=vllm_config,
            prefix=prefix,
            transformer_type=transformer_type,
        )

        self.transformer: GLM4VModel

    def _validate_pixel_values(self, data: torch.Tensor) -> torch.Tensor:
        h = w = self.config.vision_config["image_size"]
        expected_dims = (3, h, w)
        actual_dims = tuple(data.shape[1:])

        if actual_dims != expected_dims:
            expected_expr = ("batch_size", *map(str, expected_dims))
            raise ValueError(
                f"The expected shape of pixel values is {expected_expr}. "
                f"You supplied {tuple(data.shape)}.")

        return data

    def _parse_and_validate_image_input(
            self, **kwargs: object) -> Optional[GLMVImagePixelInputs]:
        pixel_values = kwargs.pop("pixel_values", None)

        if pixel_values is not None:
            if not isinstance(pixel_values, (torch.Tensor, list)):
                raise ValueError("Incorrect type of pixel values. "
                                 f"Got type: {type(pixel_values)}")

            return GLMVImagePixelInputs(
                type="pixel_values",
                data=self._validate_pixel_values(
                    flatten_bn(pixel_values, concat=True)),
            )

        return None

    def _process_image_input(
            self, image_input: GLMVImagePixelInputs) -> torch.Tensor:
        pixel_values = image_input["data"].to(dtype=self.config.torch_dtype)

        return self.transformer.vision(pixel_values)

    def get_language_model(self) -> torch.nn.Module:
        return self.transformer

    def get_multimodal_embeddings(self,
                                  **kwargs: object) -> MultiModalEmbeddings:
        image_input = self._parse_and_validate_image_input(**kwargs)
        if image_input is None:
            return []

        vision_embeddings = self._process_image_input(image_input)
        return vision_embeddings

    def get_input_embeddings(
        self,
        input_ids: torch.Tensor,
        multimodal_embeddings: Optional[MultiModalEmbeddings] = None,
    ) -> torch.Tensor:
        inputs_embeds = self.transformer.get_input_embeddings(input_ids)

        if multimodal_embeddings is not None \
            and len(multimodal_embeddings) != 0:
            inputs_embeds = merge_multimodal_embeddings(
                input_ids=input_ids,
                inputs_embeds=inputs_embeds,
                multimodal_embeddings=multimodal_embeddings,
                placeholder_token_id=[
                    self.config.boi_token_id,
                    self.config.pad_token_id,
                    self.config.eoi_token_id,
                ],
            )

        return inputs_embeds

    def forward(
        self,
        input_ids: torch.Tensor,
        positions: torch.Tensor,
        intermediate_tensors: Optional[IntermediateTensors] = None,
        inputs_embeds: Optional[torch.Tensor] = None,
        **kwargs: object,
    ) -> Union[torch.Tensor, IntermediateTensors]:
        if intermediate_tensors is not None:
            inputs_embeds = None

        # NOTE: In v1, inputs_embeds is always generated at model runner, this
        # condition is for v0 compatibility.
        elif inputs_embeds is None:
            vision_embeddings = self.get_multimodal_embeddings(**kwargs)
            inputs_embeds = self.get_input_embeddings(input_ids,
                                                      vision_embeddings)
            input_ids = None

        hidden_states = self.transformer(input_ids, positions,
                                         intermediate_tensors, inputs_embeds)

        return hidden_states

packed_modules_mapping class-attribute instance-attribute 

packed_modules_mapping = {
    "query_key_value": ["query_key_value"],
    "dense_h_to_4h": ["dense_h_to_4h"],
    "merged_proj": ["gate_proj", "dense_h_to_4h"],
}

transformer instance-attribute 

transformer: GLM4VModel

__init__ 

__init__(
    *,
    vllm_config: VllmConfig,
    prefix: str = "",
    transformer_type: type[GLM4VModel] = GLM4VModel,
) -> None
Source code in vllm/model_executor/models/glm4v.py 
550
551
552
553
554
555
556
557
558
559
560
561
562
563
def __init__(
    self,
    *,
    vllm_config: VllmConfig,
    prefix: str = "",
    transformer_type: type[GLM4VModel] = GLM4VModel,
) -> None:
    super().__init__(
        vllm_config=vllm_config,
        prefix=prefix,
        transformer_type=transformer_type,
    )

    self.transformer: GLM4VModel

_parse_and_validate_image_input 

_parse_and_validate_image_input(
    **kwargs: object,
) -> Optional[GLMVImagePixelInputs]
Source code in vllm/model_executor/models/glm4v.py 
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
def _parse_and_validate_image_input(
        self, **kwargs: object) -> Optional[GLMVImagePixelInputs]:
    pixel_values = kwargs.pop("pixel_values", None)

    if pixel_values is not None:
        if not isinstance(pixel_values, (torch.Tensor, list)):
            raise ValueError("Incorrect type of pixel values. "
                             f"Got type: {type(pixel_values)}")

        return GLMVImagePixelInputs(
            type="pixel_values",
            data=self._validate_pixel_values(
                flatten_bn(pixel_values, concat=True)),
        )

    return None

_process_image_input 

_process_image_input(
    image_input: GLMVImagePixelInputs,
) -> Tensor
Source code in vllm/model_executor/models/glm4v.py 
595
596
597
598
599
def _process_image_input(
        self, image_input: GLMVImagePixelInputs) -> torch.Tensor:
    pixel_values = image_input["data"].to(dtype=self.config.torch_dtype)

    return self.transformer.vision(pixel_values)

_validate_pixel_values 

_validate_pixel_values(data: Tensor) -> Tensor
Source code in vllm/model_executor/models/glm4v.py 
565
566
567
568
569
570
571
572
573
574
575
576
def _validate_pixel_values(self, data: torch.Tensor) -> torch.Tensor:
    h = w = self.config.vision_config["image_size"]
    expected_dims = (3, h, w)
    actual_dims = tuple(data.shape[1:])

    if actual_dims != expected_dims:
        expected_expr = ("batch_size", *map(str, expected_dims))
        raise ValueError(
            f"The expected shape of pixel values is {expected_expr}. "
            f"You supplied {tuple(data.shape)}.")

    return data

forward 

forward(
    input_ids: Tensor,
    positions: Tensor,
    intermediate_tensors: Optional[
        IntermediateTensors
    ] = None,
    inputs_embeds: Optional[Tensor] = None,
    **kwargs: object,
) -> Union[Tensor, IntermediateTensors]
Source code in vllm/model_executor/models/glm4v.py 
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
def forward(
    self,
    input_ids: torch.Tensor,
    positions: torch.Tensor,
    intermediate_tensors: Optional[IntermediateTensors] = None,
    inputs_embeds: Optional[torch.Tensor] = None,
    **kwargs: object,
) -> Union[torch.Tensor, IntermediateTensors]:
    if intermediate_tensors is not None:
        inputs_embeds = None

    # NOTE: In v1, inputs_embeds is always generated at model runner, this
    # condition is for v0 compatibility.
    elif inputs_embeds is None:
        vision_embeddings = self.get_multimodal_embeddings(**kwargs)
        inputs_embeds = self.get_input_embeddings(input_ids,
                                                  vision_embeddings)
        input_ids = None

    hidden_states = self.transformer(input_ids, positions,
                                     intermediate_tensors, inputs_embeds)

    return hidden_states

get_input_embeddings 

get_input_embeddings(
    input_ids: Tensor,
    multimodal_embeddings: Optional[
        MultiModalEmbeddings
    ] = None,
) -> Tensor
Source code in vllm/model_executor/models/glm4v.py 
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
def get_input_embeddings(
    self,
    input_ids: torch.Tensor,
    multimodal_embeddings: Optional[MultiModalEmbeddings] = None,
) -> torch.Tensor:
    inputs_embeds = self.transformer.get_input_embeddings(input_ids)

    if multimodal_embeddings is not None \
        and len(multimodal_embeddings) != 0:
        inputs_embeds = merge_multimodal_embeddings(
            input_ids=input_ids,
            inputs_embeds=inputs_embeds,
            multimodal_embeddings=multimodal_embeddings,
            placeholder_token_id=[
                self.config.boi_token_id,
                self.config.pad_token_id,
                self.config.eoi_token_id,
            ],
        )

    return inputs_embeds

get_language_model 

get_language_model() -> Module
Source code in vllm/model_executor/models/glm4v.py 
601
602
def get_language_model(self) -> torch.nn.Module:
    return self.transformer

get_mm_mapping 

get_mm_mapping() -> MultiModelKeys

Get the module prefix in multimodal models

Source code in vllm/model_executor/models/glm4v.py 
534
535
536
537
538
539
540
541
def get_mm_mapping(self) -> MultiModelKeys:
    """
    Get the module prefix in multimodal models
    """
    return MultiModelKeys.from_string_field(
        language_model="transformer.encoder",
        connector="transformer.vision.linear_proj",
        tower_model="transformer.vision.transformer")

get_multimodal_embeddings 

get_multimodal_embeddings(
    **kwargs: object,
) -> MultiModalEmbeddings
Source code in vllm/model_executor/models/glm4v.py 
604
605
606
607
608
609
610
611
def get_multimodal_embeddings(self,
                              **kwargs: object) -> MultiModalEmbeddings:
    image_input = self._parse_and_validate_image_input(**kwargs)
    if image_input is None:
        return []

    vision_embeddings = self._process_image_input(image_input)
    return vision_embeddings

get_placeholder_str classmethod 

get_placeholder_str(modality: str, i: int) -> Optional[str]
Source code in vllm/model_executor/models/glm4v.py 
543
544
545
546
547
548
@classmethod
def get_placeholder_str(cls, modality: str, i: int) -> Optional[str]:
    if modality.startswith("image"):
        return "<|begin_of_image|><|endoftext|><|end_of_image|>"

    raise ValueError("Only image modality is supported")

GLM4VModel

Bases: ChatGLMModel

Source code in vllm/model_executor/models/glm4v.py 
344
345
346
347
348
349
350
351
352
353
class GLM4VModel(ChatGLMModel):

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

        quant_config = vllm_config.quant_config

        self.vision = EVA2CLIPModel(self.config,
                                    quant_config,
                                    prefix=f"{prefix}.vision")

vision instance-attribute 

vision = EVA2CLIPModel(
    config, quant_config, prefix=f"{prefix}.vision"
)

__init__ 

__init__(*, vllm_config: VllmConfig, prefix: str = '')
Source code in vllm/model_executor/models/glm4v.py 
346
347
348
349
350
351
352
353
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
    super().__init__(vllm_config=vllm_config, prefix=prefix)

    quant_config = vllm_config.quant_config

    self.vision = EVA2CLIPModel(self.config,
                                quant_config,
                                prefix=f"{prefix}.vision")

GLM4VMultiModalProcessor

Bases: BaseMultiModalProcessor[GLM4VProcessingInfo]

Source code in vllm/model_executor/models/glm4v.py 
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
class GLM4VMultiModalProcessor(BaseMultiModalProcessor[GLM4VProcessingInfo]):

    def _hf_processor_applies_updates(
        self,
        prompt_text: str,
        mm_items: MultiModalDataItems,
        hf_processor_mm_kwargs: Mapping[str, object],
        tokenization_kwargs: Mapping[str, object],
    ) -> bool:
        return False

    def _get_mm_fields_config(
        self,
        hf_inputs: BatchFeature,
        hf_processor_mm_kwargs: Mapping[str, object],
    ) -> Mapping[str, MultiModalFieldConfig]:
        return dict(pixel_values=MultiModalFieldConfig.batched("image"))

    def _get_prompt_updates(
        self,
        mm_items: MultiModalDataItems,
        hf_processor_mm_kwargs: Mapping[str, object],
        out_mm_kwargs: MultiModalKwargs,
    ) -> Sequence[PromptUpdate]:
        hf_config = self.info.get_hf_config()

        boi_token_id = hf_config.boi_token_id
        image_token_id = hf_config.pad_token_id
        eoi_token_id = hf_config.eoi_token_id

        def get_replacement(item_idx: int):
            num_image_tokens = self.info.get_num_image_tokens()
            image_tokens = [image_token_id] * num_image_tokens

            return [boi_token_id] + image_tokens + [eoi_token_id]

        return [
            PromptReplacement(
                modality="image",
                target=[boi_token_id, image_token_id, eoi_token_id],
                replacement=get_replacement,
            ),
        ]

_get_mm_fields_config 

_get_mm_fields_config(
    hf_inputs: BatchFeature,
    hf_processor_mm_kwargs: Mapping[str, object],
) -> Mapping[str, MultiModalFieldConfig]
Source code in vllm/model_executor/models/glm4v.py 
488
489
490
491
492
493
def _get_mm_fields_config(
    self,
    hf_inputs: BatchFeature,
    hf_processor_mm_kwargs: Mapping[str, object],
) -> Mapping[str, MultiModalFieldConfig]:
    return dict(pixel_values=MultiModalFieldConfig.batched("image"))

_get_prompt_updates 

_get_prompt_updates(
    mm_items: MultiModalDataItems,
    hf_processor_mm_kwargs: Mapping[str, object],
    out_mm_kwargs: MultiModalKwargs,
) -> Sequence[PromptUpdate]
Source code in vllm/model_executor/models/glm4v.py 
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
def _get_prompt_updates(
    self,
    mm_items: MultiModalDataItems,
    hf_processor_mm_kwargs: Mapping[str, object],
    out_mm_kwargs: MultiModalKwargs,
) -> Sequence[PromptUpdate]:
    hf_config = self.info.get_hf_config()

    boi_token_id = hf_config.boi_token_id
    image_token_id = hf_config.pad_token_id
    eoi_token_id = hf_config.eoi_token_id

    def get_replacement(item_idx: int):
        num_image_tokens = self.info.get_num_image_tokens()
        image_tokens = [image_token_id] * num_image_tokens

        return [boi_token_id] + image_tokens + [eoi_token_id]

    return [
        PromptReplacement(
            modality="image",
            target=[boi_token_id, image_token_id, eoi_token_id],
            replacement=get_replacement,
        ),
    ]

_hf_processor_applies_updates 

_hf_processor_applies_updates(
    prompt_text: str,
    mm_items: MultiModalDataItems,
    hf_processor_mm_kwargs: Mapping[str, object],
    tokenization_kwargs: Mapping[str, object],
) -> bool
Source code in vllm/model_executor/models/glm4v.py 
479
480
481
482
483
484
485
486
def _hf_processor_applies_updates(
    self,
    prompt_text: str,
    mm_items: MultiModalDataItems,
    hf_processor_mm_kwargs: Mapping[str, object],
    tokenization_kwargs: Mapping[str, object],
) -> bool:
    return False

GLM4VProcessingInfo

Bases: BaseProcessingInfo

Source code in vllm/model_executor/models/glm4v.py 
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
class GLM4VProcessingInfo(BaseProcessingInfo):

    def get_hf_config(self):
        return self.ctx.get_hf_config(ChatGLMConfig)

    def get_hf_processor(self, **kwargs: object) -> GLM4VProcessor:
        return self.ctx.init_processor(
            GLM4VProcessor,
            config=self.get_hf_config(),
            tokenizer=self.get_tokenizer(),
            **kwargs,
        )

    def get_supported_mm_limits(self) -> Mapping[str, Optional[int]]:
        return {"image": 1}

    def get_num_image_tokens(self) -> int:
        hf_config = self.get_hf_config()
        vision_config = hf_config.vision_config

        image_size = vision_config["image_size"]
        patch_size = vision_config["patch_size"]
        grid_length = image_size // patch_size // 2
        return grid_length * grid_length

    def get_num_image_feature_tokens(self) -> int:
        # EVA2CLIPModel has embeddings for boi and eoi tokens as well
        return self.get_num_image_tokens() + 2

get_hf_config 

get_hf_config()
Source code in vllm/model_executor/models/glm4v.py 
421
422
def get_hf_config(self):
    return self.ctx.get_hf_config(ChatGLMConfig)

get_hf_processor 

get_hf_processor(**kwargs: object) -> GLM4VProcessor
Source code in vllm/model_executor/models/glm4v.py 
424
425
426
427
428
429
430
def get_hf_processor(self, **kwargs: object) -> GLM4VProcessor:
    return self.ctx.init_processor(
        GLM4VProcessor,
        config=self.get_hf_config(),
        tokenizer=self.get_tokenizer(),
        **kwargs,
    )

get_num_image_feature_tokens 

get_num_image_feature_tokens() -> int
Source code in vllm/model_executor/models/glm4v.py 
444
445
446
def get_num_image_feature_tokens(self) -> int:
    # EVA2CLIPModel has embeddings for boi and eoi tokens as well
    return self.get_num_image_tokens() + 2

get_num_image_tokens 

get_num_image_tokens() -> int
Source code in vllm/model_executor/models/glm4v.py 
435
436
437
438
439
440
441
442
def get_num_image_tokens(self) -> int:
    hf_config = self.get_hf_config()
    vision_config = hf_config.vision_config

    image_size = vision_config["image_size"]
    patch_size = vision_config["patch_size"]
    grid_length = image_size // patch_size // 2
    return grid_length * grid_length

get_supported_mm_limits 

get_supported_mm_limits() -> Mapping[str, Optional[int]]
Source code in vllm/model_executor/models/glm4v.py 
432
433
def get_supported_mm_limits(self) -> Mapping[str, Optional[int]]:
    return {"image": 1}

GLM4VProcessor

This model doesn't define its own HF processor, so we implement our own one here.

Source code in vllm/model_executor/models/glm4v.py 
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
408
409
410
411
412
413
414
415
416
class GLM4VProcessor:
    """
    This model doesn't define its own HF processor,
    so we implement our own one here.
    """

    def __init__(
        self,
        config: ChatGLMConfig,
        tokenizer: PreTrainedTokenizer,
    ) -> None:
        super().__init__()

        self.config = config
        self.tokenizer = tokenizer

        vision_config = config.vision_config
        image_size = vision_config["image_size"]

        self.image_transform = transforms.Compose([
            transforms.Resize(
                (image_size, image_size),
                interpolation=InterpolationMode.BICUBIC,
            ),
            transforms.ToTensor(),
            transforms.Normalize(
                mean=(0.48145466, 0.4578275, 0.40821073),
                std=(0.26862954, 0.26130258, 0.27577711),
            ),
        ])

    def __call__(
        self,
        text: Optional[Union[TextInput, list[TextInput]]] = None,
        images: Optional[Union[ImageInput, list[ImageInput]]] = None,
        return_tensors: Optional[Union[str, TensorType]] = None,
    ) -> BatchFeature:
        if text is None:
            text = []
        if not isinstance(text, list):
            text = [text]
        if images is None:
            images = []
        if not isinstance(images, list):
            images = [images]

        text_inputs = self.tokenizer(text)

        if len(images) == 0:
            image_inputs = {}
        else:
            pixel_values = [self.image_transform(image) for image in images]
            image_inputs = {"pixel_values": torch.stack(pixel_values)}

        return BatchFeature(
            {
                **text_inputs,
                **image_inputs,
            },
            tensor_type=return_tensors,
        )

config instance-attribute 

config = config

image_transform instance-attribute 

image_transform = Compose(
    [
        Resize(
            (image_size, image_size), interpolation=BICUBIC
        ),
        ToTensor(),
        Normalize(
            mean=(0.48145466, 0.4578275, 0.40821073),
            std=(0.26862954, 0.26130258, 0.27577711),
        ),
    ]
)

tokenizer instance-attribute 

tokenizer = tokenizer

__call__ 

__call__(
    text: Optional[
        Union[TextInput, list[TextInput]]
    ] = None,
    images: Optional[
        Union[ImageInput, list[ImageInput]]
    ] = None,
    return_tensors: Optional[Union[str, TensorType]] = None,
) -> BatchFeature
Source code in vllm/model_executor/models/glm4v.py 
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
def __call__(
    self,
    text: Optional[Union[TextInput, list[TextInput]]] = None,
    images: Optional[Union[ImageInput, list[ImageInput]]] = None,
    return_tensors: Optional[Union[str, TensorType]] = None,
) -> BatchFeature:
    if text is None:
        text = []
    if not isinstance(text, list):
        text = [text]
    if images is None:
        images = []
    if not isinstance(images, list):
        images = [images]

    text_inputs = self.tokenizer(text)

    if len(images) == 0:
        image_inputs = {}
    else:
        pixel_values = [self.image_transform(image) for image in images]
        image_inputs = {"pixel_values": torch.stack(pixel_values)}

    return BatchFeature(
        {
            **text_inputs,
            **image_inputs,
        },
        tensor_type=return_tensors,
    )

__init__ 

__init__(
    config: ChatGLMConfig, tokenizer: PreTrainedTokenizer
) -> None
Source code in vllm/model_executor/models/glm4v.py 
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
def __init__(
    self,
    config: ChatGLMConfig,
    tokenizer: PreTrainedTokenizer,
) -> None:
    super().__init__()

    self.config = config
    self.tokenizer = tokenizer

    vision_config = config.vision_config
    image_size = vision_config["image_size"]

    self.image_transform = transforms.Compose([
        transforms.Resize(
            (image_size, image_size),
            interpolation=InterpolationMode.BICUBIC,
        ),
        transforms.ToTensor(),
        transforms.Normalize(
            mean=(0.48145466, 0.4578275, 0.40821073),
            std=(0.26862954, 0.26130258, 0.27577711),
        ),
    ])

GLMVImagePixelInputs

Bases: TypedDict

Source code in vllm/model_executor/models/glm4v.py 
48
49
50
51
class GLMVImagePixelInputs(TypedDict):
    type: Literal["pixel_values"]
    data: torch.Tensor
    """Shape: `(batch_size, num_channels, height, width)`"""

data instance-attribute 

data: Tensor

Shape: (batch_size, num_channels, height, width)

type instance-attribute 

type: Literal['pixel_values']