Skip to content

unitorch.models.gemma¤

GemmaProcessor¤

Bases: HfLlmProcessor

Gemma tokenizer-backed processor for decoder-only generation tasks.

Source code in src/unitorch/models/gemma/processing.py 
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
def __init__(
    self,
    tokenizer_file: str,
    tokenizer_config: Optional[str] = None,
    chat_template: Optional[str] = None,
    max_seq_length: Optional[int] = 12800,
    max_gen_seq_length: Optional[int] = 512,
):
    tokenizer = _load_gemma_tokenizer(
        tokenizer_file=tokenizer_file,
        tokenizer_config=tokenizer_config,
        chat_template=chat_template,
    )
    super().__init__(
        tokenizer=tokenizer,
        max_seq_length=max_seq_length,
        max_gen_seq_length=max_gen_seq_length,
    )

chat_template ¤

chat_template(messages: List[Dict[str, Any]]) -> str
Source code in src/unitorch/models/gemma/processing.py 
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
def chat_template(
    self,
    messages: List[Dict[str, Any]],
) -> str:
    if getattr(self.tokenizer, "chat_template", None):
        return self.tokenizer.apply_chat_template(
            messages,
            tokenize=False,
            add_generation_prompt=True,
        )

    return _fallback_chat_template(
        messages,
        image_token=getattr(self.tokenizer, "image_token", None),
    )

GemmaVLProcessor¤

Bases: GemmaProcessor

Gemma processor for multimodal generation with image inputs.

Source code in src/unitorch/models/gemma/processing_vl.py 
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
def __init__(
    self,
    tokenizer_file: str,
    processor_config_path: str,
    tokenizer_config: Optional[str] = None,
    chat_template: Optional[str] = None,
    max_seq_length: Optional[int] = 12800,
    max_gen_seq_length: Optional[int] = 512,
):
    super().__init__(
        tokenizer_file=tokenizer_file,
        tokenizer_config=tokenizer_config,
        chat_template=chat_template,
        max_seq_length=max_seq_length,
        max_gen_seq_length=max_gen_seq_length,
    )

    self.vision_processor = Gemma4ImageProcessor.from_json_file(
        processor_config_path
    )
    self.image_token = getattr(self.tokenizer, "image_token", "<|image|>")
    self.boi_token = getattr(self.tokenizer, "boi_token", "<|image>")
    self.eoi_token = getattr(self.tokenizer, "eoi_token", "<image|>")
    self.image_token_id = getattr(
        self.tokenizer,
        "image_token_id",
        self.tokenizer.convert_tokens_to_ids(self.image_token),
    )

vision_processor instance-attribute ¤

vision_processor = from_json_file(processor_config_path)

image_token instance-attribute ¤

image_token = getattr(tokenizer, "image_token", "<|image|>")

boi_token instance-attribute ¤

boi_token = getattr(tokenizer, 'boi_token', '<|image>')

eoi_token instance-attribute ¤

eoi_token = getattr(tokenizer, 'eoi_token', '<image|>')

image_token_id instance-attribute ¤

image_token_id = getattr(
    tokenizer,
    "image_token_id",
    convert_tokens_to_ids(image_token),
)

_create_mm_token_type_ids ¤

_create_mm_token_type_ids(input_ids: Tensor) -> Tensor
Source code in src/unitorch/models/gemma/processing_vl.py 
48
49
50
51
52
53
54
def _create_mm_token_type_ids(
    self,
    input_ids: torch.Tensor,
) -> torch.Tensor:
    mm_token_type_ids = torch.zeros_like(input_ids)
    mm_token_type_ids[input_ids == self.image_token_id] = 1
    return mm_token_type_ids

processing_images ¤

processing_images(
    images: Union[Image, str, Sequence[Union[Image, str]]],
)
Source code in src/unitorch/models/gemma/processing_vl.py 
56
57
58
59
60
61
62
63
64
65
66
def processing_images(
    self,
    images: Union[Image.Image, str, Sequence[Union[Image.Image, str]]],
):
    if isinstance(images, (Image.Image, str)):
        images = [images]
    images = [
        image if isinstance(image, Image.Image) else Image.open(image).convert("RGB")
        for image in images
    ]
    return self.vision_processor(images=images, return_tensors="pt")

_prepare_text_with_images ¤

_prepare_text_with_images(
    text: str, num_soft_tokens_per_image: Sequence[int]
) -> str
Source code in src/unitorch/models/gemma/processing_vl.py 
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
def _prepare_text_with_images(
    self,
    text: str,
    num_soft_tokens_per_image: Sequence[int],
) -> str:
    text = str(text)
    image_count = len(num_soft_tokens_per_image)
    soft_token_placeholder = "<|gemma_image_soft_token|>"
    if image_count > 0 and self.image_token not in text:
        prefix = " ".join([self.image_token] * image_count)
        text = f"{prefix}\n{text}".strip()

    image_index = 0
    while self.image_token in text:
        if image_index >= image_count:
            raise ValueError(
                "More image placeholders were found in the prompt than image inputs."
            )
        replacement = (
            f"{self.boi_token}"
            f"{soft_token_placeholder * int(num_soft_tokens_per_image[image_index])}"
            f"{self.eoi_token}"
        )
        text = text.replace(self.image_token, replacement, 1)
        image_index += 1

    if image_index != image_count:
        raise ValueError(
            "The number of image placeholders in the prompt does not match the number of image inputs."
        )

    return text.replace(soft_token_placeholder, self.image_token)

generation_inputs ¤

generation_inputs(
    text: str,
    images: Optional[
        Union[Image, str, Sequence[Union[Image, str]]]
    ] = None,
    max_seq_length: Optional[int] = None,
) -> GenericOutputs
Source code in src/unitorch/models/gemma/processing_vl.py 
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
def generation_inputs(
    self,
    text: str,
    images: Optional[
        Union[Image.Image, str, Sequence[Union[Image.Image, str]]]
    ] = None,
    max_seq_length: Optional[int] = None,
) -> GenericOutputs:
    image_inputs = self.processing_images(images) if images else None
    num_soft_tokens_per_image = (
        image_inputs["num_soft_tokens_per_image"].tolist() if image_inputs else []
    )
    text = self._prepare_text_with_images(text, num_soft_tokens_per_image)
    text_inputs = super().generation_inputs(
        text=text,
        max_seq_length=max_seq_length,
    )
    return GenericOutputs(
        input_ids=text_inputs.input_ids,
        attention_mask=text_inputs.attention_mask,
        mm_token_type_ids=self._create_mm_token_type_ids(text_inputs.input_ids),
        pixel_values=(image_inputs["pixel_values"] if image_inputs else None),
        image_position_ids=(
            image_inputs["image_position_ids"] if image_inputs else None
        ),
    )

generation ¤

generation(
    text: str,
    images: Optional[
        Union[Image, str, Sequence[Union[Image, str]]]
    ],
    text_pair: str,
    max_seq_length: Optional[int] = None,
    max_gen_seq_length: Optional[int] = None,
) -> GenericOutputs
Source code in src/unitorch/models/gemma/processing_vl.py 
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
def generation(
    self,
    text: str,
    images: Optional[
        Union[Image.Image, str, Sequence[Union[Image.Image, str]]]
    ],
    text_pair: str,
    max_seq_length: Optional[int] = None,
    max_gen_seq_length: Optional[int] = None,
) -> GenericOutputs:
    image_inputs = self.processing_images(images) if images else None
    num_soft_tokens_per_image = (
        image_inputs["num_soft_tokens_per_image"].tolist() if image_inputs else []
    )
    text = self._prepare_text_with_images(text, num_soft_tokens_per_image)
    text_inputs = super().generation(
        text=text,
        text_pair=text_pair,
        max_seq_length=max_seq_length,
        max_gen_seq_length=max_gen_seq_length,
    )
    return GenericOutputs(
        input_ids=text_inputs.input_ids,
        attention_mask=text_inputs.attention_mask,
        mm_token_type_ids=self._create_mm_token_type_ids(text_inputs.input_ids),
        pixel_values=(image_inputs["pixel_values"] if image_inputs else None),
        image_position_ids=(
            image_inputs["image_position_ids"] if image_inputs else None
        ),
        input_ids_label=text_inputs.input_ids_label,
        attention_mask_label=text_inputs.attention_mask_label,
    )

messages_generation ¤

messages_generation(
    messages: List[Dict[str, Any]],
    images: Optional[
        Union[Image, str, Sequence[Union[Image, str]]]
    ] = None,
    max_seq_length: Optional[int] = None,
) -> GenericOutputs
Source code in src/unitorch/models/gemma/processing_vl.py 
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
def messages_generation(
    self,
    messages: List[Dict[str, Any]],
    images: Optional[
        Union[Image.Image, str, Sequence[Union[Image.Image, str]]]
    ] = None,
    max_seq_length: Optional[int] = None,
) -> GenericOutputs:
    while messages and messages[-1]["role"] != "assistant":
        messages.pop()

    text = self.chat_template(messages[:-1])
    text_pair = self.chat_template(messages[-1:])
    outputs = self.generation(
        text=text,
        images=images,
        text_pair=text_pair,
        max_seq_length=max_seq_length,
    )
    return GenericOutputs(
        input_ids=outputs.input_ids,
        attention_mask=outputs.attention_mask,
        mm_token_type_ids=outputs.mm_token_type_ids,
        pixel_values=outputs.pixel_values,
        image_position_ids=outputs.image_position_ids,
        input_ids_label=outputs.input_ids_label,
        attention_mask_label=outputs.attention_mask_label,
    )

GemmaForGeneration¤

Bases: GenericModel, PeftWeightLoaderMixin

Gemma text generation model backed by Gemma4's unified checkpoint.

Source code in src/unitorch/models/gemma/modeling.py 
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
def __init__(
    self,
    config_path: str,
    gradient_checkpointing: Optional[bool] = False,
):
    super().__init__()
    self.config = _get_gemma_text_config(config_path)
    if gradient_checkpointing:
        self.config.use_cache = False
        if getattr(self.config, "text_config", None) is not None:
            self.config.text_config.use_cache = False
    self.model = Gemma4ForConditionalGeneration(self.config)
    if gradient_checkpointing:
        self.model.gradient_checkpointing_enable()
    self.init_weights()
    self.model.to(dtype=_get_gemma_dtype(self.config))

prefix_keys_in_state_dict class-attribute instance-attribute ¤

prefix_keys_in_state_dict = {
    "^(?!model\\.model\\.).*": "model."
}

config instance-attribute ¤

config = _get_gemma_text_config(config_path)

model instance-attribute ¤

model = Gemma4ForConditionalGeneration(config)

forward ¤

forward(
    input_ids: Tensor,
    attention_mask: Optional[Tensor] = None,
) -> Tensor
Source code in src/unitorch/models/gemma/modeling.py 
59
60
61
62
63
64
65
66
67
68
69
def forward(
    self,
    input_ids: torch.Tensor,
    attention_mask: Optional[torch.Tensor] = None,
) -> torch.Tensor:
    outputs = self.model(
        input_ids=input_ids,
        attention_mask=attention_mask,
        return_dict=True,
    )
    return outputs.logits

generate ¤

generate(
    input_ids: Tensor,
    attention_mask: Optional[Tensor] = None,
    num_beams: Optional[int] = 5,
    decoder_start_token_id: Optional[int] = 2,
    decoder_end_token_id: Optional[
        Union[int, List[int]]
    ] = 1,
    decoder_pad_token_id: Optional[int] = 0,
    num_return_sequences: Optional[int] = 1,
    min_gen_seq_length: Optional[int] = 0,
    max_gen_seq_length: Optional[int] = 512,
    repetition_penalty: Optional[float] = 1.0,
    no_repeat_ngram_size: Optional[int] = 0,
    early_stopping: Optional[bool] = True,
    length_penalty: Optional[float] = 1.0,
    num_beam_groups: Optional[int] = 1,
    diversity_penalty: Optional[float] = 0.0,
    do_sample: Optional[bool] = False,
    temperature: Optional[float] = 1.0,
    top_k: Optional[int] = 50,
    top_p: Optional[float] = 1.0,
) -> GenericOutputs
Source code in src/unitorch/models/gemma/modeling.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
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
@torch.no_grad()
def generate(
    self,
    input_ids: torch.Tensor,
    attention_mask: Optional[torch.Tensor] = None,
    num_beams: Optional[int] = 5,
    decoder_start_token_id: Optional[int] = 2,
    decoder_end_token_id: Optional[Union[int, List[int]]] = 1,
    decoder_pad_token_id: Optional[int] = 0,
    num_return_sequences: Optional[int] = 1,
    min_gen_seq_length: Optional[int] = 0,
    max_gen_seq_length: Optional[int] = 512,
    repetition_penalty: Optional[float] = 1.0,
    no_repeat_ngram_size: Optional[int] = 0,
    early_stopping: Optional[bool] = True,
    length_penalty: Optional[float] = 1.0,
    num_beam_groups: Optional[int] = 1,
    diversity_penalty: Optional[float] = 0.0,
    do_sample: Optional[bool] = False,
    temperature: Optional[float] = 1.0,
    top_k: Optional[int] = 50,
    top_p: Optional[float] = 1.0,
) -> GenericOutputs:
    input_seq_length = input_ids.size(1)
    outputs = self.model.generate(
        input_ids=input_ids,
        attention_mask=attention_mask,
        max_length=max_gen_seq_length + input_seq_length,
        min_length=min_gen_seq_length + input_seq_length,
        num_beams=num_beams,
        do_sample=do_sample,
        no_repeat_ngram_size=no_repeat_ngram_size,
        early_stopping=early_stopping,
        length_penalty=length_penalty,
        repetition_penalty=repetition_penalty,
        num_return_sequences=num_return_sequences,
        bos_token_id=decoder_start_token_id,
        eos_token_id=decoder_end_token_id,
        pad_token_id=decoder_pad_token_id,
        num_beam_groups=num_beam_groups,
        diversity_penalty=diversity_penalty,
        temperature=temperature,
        top_k=top_k,
        top_p=top_p,
        return_dict_in_generate=True,
        output_scores=True,
    )

    sequences = outputs.sequences.reshape(
        -1, num_return_sequences, outputs.sequences.size(-1)
    )
    padded = torch.full(
        (sequences.size(0), num_return_sequences, max_gen_seq_length),
        fill_value=decoder_pad_token_id,
        device=sequences.device,
    )
    padded[:, :, : sequences.size(-1) - input_seq_length].copy_(
        sequences[:, :, input_seq_length : sequences.size(-1)]
    )

    if num_return_sequences == 1:
        padded = padded.reshape(-1, max_gen_seq_length)

    return GenericOutputs(
        sequences=padded.long(),
        sequences_scores=getattr(outputs, "sequences_scores", None),
    )

GemmaVLForGeneration¤

Bases: GenericModel, PeftWeightLoaderMixin

Gemma vision-language generation model backed by Gemma4's unified checkpoint.

Source code in src/unitorch/models/gemma/modeling_vl.py 
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
def __init__(
    self,
    config_path: str,
    gradient_checkpointing: Optional[bool] = False,
):
    super().__init__()
    self.config = Gemma4Config.from_json_file(config_path)
    self.vision_config = copy.deepcopy(self.config.vision_config)
    self.config.audio_config = None
    self.config.vision_config = None
    if gradient_checkpointing:
        self.config.use_cache = False
        if getattr(self.config, "text_config", None) is not None:
            self.config.text_config.use_cache = False
    self.model = Gemma4ForConditionalGeneration(self.config)
    self.model.model.vision_tower = GemmaUnifiedVisionTower(
        patch_size=self.vision_config.patch_size,
        pooling_kernel_size=self.vision_config.pooling_kernel_size,
        hidden_size=self.vision_config.output_proj_dims,
    )
    self.model.model.embed_vision = Gemma4MultimodalEmbedder(
        self.vision_config,
        self.config.text_config,
    )
    if gradient_checkpointing:
        self.model.gradient_checkpointing_enable()
    self.init_weights()
    self.model.to(dtype=_get_gemma_dtype(self.config))

prefix_keys_in_state_dict class-attribute instance-attribute ¤

prefix_keys_in_state_dict = {
    "^(?!model\\.model\\.).*": "model."
}

replace_keys_in_state_dict class-attribute instance-attribute ¤

replace_keys_in_state_dict = {
    "model\\.model\\.vision_embedder\\.": "model.model.vision_tower."
}

config instance-attribute ¤

config = from_json_file(config_path)

vision_config instance-attribute ¤

vision_config = deepcopy(vision_config)

model instance-attribute ¤

model = Gemma4ForConditionalGeneration(config)

forward ¤

forward(
    input_ids: Tensor,
    pixel_values: Optional[Tensor] = None,
    image_position_ids: Optional[Tensor] = None,
    attention_mask: Optional[Tensor] = None,
    mm_token_type_ids: Optional[Tensor] = None,
)
Source code in src/unitorch/models/gemma/modeling_vl.py 
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
def forward(
    self,
    input_ids: torch.Tensor,
    pixel_values: Optional[torch.Tensor] = None,
    image_position_ids: Optional[torch.Tensor] = None,
    attention_mask: Optional[torch.Tensor] = None,
    mm_token_type_ids: Optional[torch.Tensor] = None,
):
    pixel_values, image_position_ids = _reshape_image_inputs(
        pixel_values,
        image_position_ids,
    )
    outputs = self.model(
        input_ids=input_ids,
        pixel_values=pixel_values,
        image_position_ids=image_position_ids,
        attention_mask=attention_mask,
        mm_token_type_ids=mm_token_type_ids,
        return_dict=True,
    )
    return outputs.logits

generate ¤

generate(
    input_ids: Tensor,
    pixel_values: Optional[Tensor] = None,
    image_position_ids: Optional[Tensor] = None,
    attention_mask: Optional[Tensor] = None,
    mm_token_type_ids: Optional[Tensor] = None,
    num_beams: Optional[int] = 5,
    decoder_start_token_id: Optional[int] = 2,
    decoder_end_token_id: Optional[
        Union[int, List[int]]
    ] = 1,
    decoder_pad_token_id: Optional[int] = 0,
    num_return_sequences: Optional[int] = 1,
    min_gen_seq_length: Optional[int] = 0,
    max_gen_seq_length: Optional[int] = 512,
    repetition_penalty: Optional[float] = 1.0,
    no_repeat_ngram_size: Optional[int] = 0,
    early_stopping: Optional[bool] = True,
    length_penalty: Optional[float] = 1.0,
    num_beam_groups: Optional[int] = 1,
    diversity_penalty: Optional[float] = 0.0,
    do_sample: Optional[bool] = False,
    temperature: Optional[float] = 1.0,
    top_k: Optional[int] = 50,
    top_p: Optional[float] = 1.0,
)
Source code in src/unitorch/models/gemma/modeling_vl.py 
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
288
289
290
291
292
293
@torch.no_grad()
def generate(
    self,
    input_ids: torch.Tensor,
    pixel_values: Optional[torch.Tensor] = None,
    image_position_ids: Optional[torch.Tensor] = None,
    attention_mask: Optional[torch.Tensor] = None,
    mm_token_type_ids: Optional[torch.Tensor] = None,
    num_beams: Optional[int] = 5,
    decoder_start_token_id: Optional[int] = 2,
    decoder_end_token_id: Optional[Union[int, List[int]]] = 1,
    decoder_pad_token_id: Optional[int] = 0,
    num_return_sequences: Optional[int] = 1,
    min_gen_seq_length: Optional[int] = 0,
    max_gen_seq_length: Optional[int] = 512,
    repetition_penalty: Optional[float] = 1.0,
    no_repeat_ngram_size: Optional[int] = 0,
    early_stopping: Optional[bool] = True,
    length_penalty: Optional[float] = 1.0,
    num_beam_groups: Optional[int] = 1,
    diversity_penalty: Optional[float] = 0.0,
    do_sample: Optional[bool] = False,
    temperature: Optional[float] = 1.0,
    top_k: Optional[int] = 50,
    top_p: Optional[float] = 1.0,
):
    input_seq_length = input_ids.size(1)
    pixel_values, image_position_ids = _reshape_image_inputs(
        pixel_values,
        image_position_ids,
    )
    outputs = self.model.generate(
        input_ids=input_ids,
        pixel_values=pixel_values,
        image_position_ids=image_position_ids,
        attention_mask=attention_mask,
        mm_token_type_ids=mm_token_type_ids,
        max_length=max_gen_seq_length + input_seq_length,
        min_length=min_gen_seq_length + input_seq_length,
        num_beams=num_beams,
        do_sample=do_sample,
        no_repeat_ngram_size=no_repeat_ngram_size,
        early_stopping=early_stopping,
        length_penalty=length_penalty,
        repetition_penalty=repetition_penalty,
        num_return_sequences=num_return_sequences,
        bos_token_id=decoder_start_token_id,
        eos_token_id=decoder_end_token_id,
        pad_token_id=decoder_pad_token_id,
        num_beam_groups=num_beam_groups,
        diversity_penalty=diversity_penalty,
        temperature=temperature,
        top_k=top_k,
        top_p=top_p,
        return_dict_in_generate=True,
        output_scores=True,
    )

    sequences = outputs.sequences.reshape(
        -1, num_return_sequences, outputs.sequences.size(-1)
    )
    padded = torch.full(
        (sequences.size(0), num_return_sequences, max_gen_seq_length),
        fill_value=decoder_pad_token_id,
        device=sequences.device,
    )
    padded[:, :, : sequences.size(-1) - input_seq_length].copy_(
        sequences[:, :, input_seq_length : sequences.size(-1)]
    )

    if num_return_sequences == 1:
        padded = padded.reshape(-1, max_gen_seq_length)

    return GenericOutputs(
        sequences=padded.long(),
        sequences_scores=getattr(outputs, "sequences_scores", None),
    )