Skip to content

unitorch.models.peftยค

ClipLoraForMatchingยค

Bases: GenericPeftModel, PeftWeightLoaderMixin

Source code in src/unitorch/models/peft/modeling_clip.py 
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
def __init__(
    self,
    config_path: str,
    projection_dim: Optional[int] = 512,
    lora_r: Optional[int] = 16,
    lora_alpha: Optional[int] = 32,
    lora_dropout: Optional[float] = 0.05,
    fan_in_fan_out: Optional[bool] = True,
    target_modules: Optional[Union[List[str], str]] = ["q_proj", "v_proj"],
):
    super().__init__()
    self.peft_config = LoraConfig(
        r=lora_r,
        lora_alpha=lora_alpha,
        lora_dropout=lora_dropout,
        fan_in_fan_out=fan_in_fan_out,
        target_modules=target_modules,
    )
    self.peft_model = PeftModelForSequenceClassification(
        ClipForMatching(config_path, projection_dim=projection_dim),
        self.peft_config,
    )
    self.classifier = nn.Linear(1, 1)

    self.init_weights()
    self.classifier.weight.data.fill_(5.0)

prefix_keys_in_state_dict class-attribute instance-attribute ยค

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

replace_keys_in_state_dict class-attribute instance-attribute ยค

replace_keys_in_state_dict = {
    "q_proj.weight": "q_proj.base_layer.weight",
    "q_proj.bias": "q_proj.base_layer.bias",
    "v_proj.weight": "v_proj.base_layer.weight",
    "v_proj.bias": "v_proj.base_layer.bias",
}

modules_to_save_checkpoints class-attribute instance-attribute ยค

modules_to_save_checkpoints = ['lora', 'classifier']

replace_keys_in_peft_state_dict class-attribute instance-attribute ยค

replace_keys_in_peft_state_dict = {
    ".weight": ".base_layer.weight",
    ".bias": ".base_layer.bias",
}

peft_config instance-attribute ยค

peft_config = LoraConfig(
    r=lora_r,
    lora_alpha=lora_alpha,
    lora_dropout=lora_dropout,
    fan_in_fan_out=fan_in_fan_out,
    target_modules=target_modules,
)

peft_model instance-attribute ยค

peft_model = PeftModelForSequenceClassification(
    ClipForMatching(
        config_path, projection_dim=projection_dim
    ),
    peft_config,
)

classifier instance-attribute ยค

classifier = Linear(1, 1)

forward ยค

forward(
    input_ids: Tensor,
    pixel_values: Tensor,
    attention_mask: Tensor,
    position_ids: Tensor,
)
Source code in src/unitorch/models/peft/modeling_clip.py 
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
def forward(
    self,
    input_ids: torch.Tensor,
    pixel_values: torch.Tensor,
    attention_mask: torch.Tensor,
    position_ids: torch.Tensor,
):
    text_embeds, image_embeds = self.peft_model(
        input_ids=input_ids,
        pixel_values=pixel_values,
        attention_mask=attention_mask,
        position_ids=position_ids,
    )

    text_embeds = text_embeds / text_embeds.norm(dim=-1, keepdim=True)
    image_embeds = image_embeds / image_embeds.norm(dim=-1, keepdim=True)
    scores = torch.sum(text_embeds * image_embeds, dim=-1, keepdim=True)

    outputs = self.classifier(scores)
    return outputs

ClipLoraForTextMatchingยค

Bases: GenericPeftModel, PeftWeightLoaderMixin

Source code in src/unitorch/models/peft/modeling_clip.py 
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
def __init__(
    self,
    config_path: str,
    projection_dim: Optional[int] = 512,
    lora_r: Optional[int] = 16,
    lora_alpha: Optional[int] = 32,
    lora_dropout: Optional[float] = 0.05,
    fan_in_fan_out: Optional[bool] = True,
    target_modules: Optional[Union[List[str], str]] = ["q_proj", "v_proj"],
):
    super().__init__()
    self.peft_config = LoraConfig(
        r=lora_r,
        lora_alpha=lora_alpha,
        lora_dropout=lora_dropout,
        fan_in_fan_out=fan_in_fan_out,
        target_modules=target_modules,
    )
    self.peft_model = PeftModelForSequenceClassification(
        ClipForTextMatching(config_path, projection_dim=projection_dim),
        self.peft_config,
    )
    self.classifier = nn.Linear(1, 1)

    self.init_weights()
    self.classifier.weight.data.fill_(5.0)

prefix_keys_in_state_dict class-attribute instance-attribute ยค

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

replace_keys_in_state_dict class-attribute instance-attribute ยค

replace_keys_in_state_dict = {
    "q_proj.weight": "q_proj.base_layer.weight",
    "q_proj.bias": "q_proj.base_layer.bias",
    "v_proj.weight": "v_proj.base_layer.weight",
    "v_proj.bias": "v_proj.base_layer.bias",
}

modules_to_save_checkpoints class-attribute instance-attribute ยค

modules_to_save_checkpoints = ['lora', 'classifier']

replace_keys_in_peft_state_dict class-attribute instance-attribute ยค

replace_keys_in_peft_state_dict = {
    ".weight": ".base_layer.weight",
    ".bias": ".base_layer.bias",
}

peft_config instance-attribute ยค

peft_config = LoraConfig(
    r=lora_r,
    lora_alpha=lora_alpha,
    lora_dropout=lora_dropout,
    fan_in_fan_out=fan_in_fan_out,
    target_modules=target_modules,
)

peft_model instance-attribute ยค

peft_model = PeftModelForSequenceClassification(
    ClipForTextMatching(
        config_path, projection_dim=projection_dim
    ),
    peft_config,
)

classifier instance-attribute ยค

classifier = Linear(1, 1)

forward ยค

forward(
    query_input_ids: Tensor,
    query_attention_mask: Tensor,
    query_position_ids: Tensor,
    doc_input_ids: Tensor,
    doc_attention_mask: Tensor,
    doc_position_ids: Tensor,
)
Source code in src/unitorch/models/peft/modeling_clip.py 
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
def forward(
    self,
    query_input_ids: torch.Tensor,
    query_attention_mask: torch.Tensor,
    query_position_ids: torch.Tensor,
    doc_input_ids: torch.Tensor,
    doc_attention_mask: torch.Tensor,
    doc_position_ids: torch.Tensor,
):
    query_embeds, doc_embeds = self.peft_model(
        query_input_ids=query_input_ids,
        query_attention_mask=query_attention_mask,
        query_position_ids=query_position_ids,
        doc_input_ids=doc_input_ids,
        doc_attention_mask=doc_attention_mask,
        doc_position_ids=doc_position_ids,
    )

    query_embeds = query_embeds / query_embeds.norm(dim=-1, keepdim=True)
    doc_embeds = doc_embeds / doc_embeds.norm(dim=-1, keepdim=True)
    scores = torch.sum(query_embeds * doc_embeds, dim=-1, keepdim=True)

    outputs = self.classifier(scores)
    return outputs

LlamaLoraForClassificationยค

Bases: GenericPeftModel

Source code in src/unitorch/models/peft/modeling_llama.py 
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
def __init__(
    self,
    config_path: str,
    lora_r: Optional[int] = 16,
    lora_alpha: Optional[int] = 32,
    lora_dropout: Optional[float] = 0.05,
    fan_in_fan_out: Optional[bool] = True,
    target_modules: Optional[Union[List[str], str]] = ["q_proj", "v_proj"],
    num_classes: Optional[int] = 1,
    hidden_dropout_prob: Optional[float] = 0.1,
    freeze_classifer: Optional[bool] = True,
    gradient_checkpointing: Optional[bool] = False,
):
    super().__init__()
    self.config = LlamaConfig.from_json_file(config_path)
    self.config.gradient_checkpointing = gradient_checkpointing
    self.peft_config = LoraConfig(
        r=lora_r,
        lora_alpha=lora_alpha,
        lora_dropout=lora_dropout,
        fan_in_fan_out=fan_in_fan_out,
        target_modules=target_modules,
    )
    self.model = LlamaModel(self.config)
    self.model.add_adapter(self.peft_config)
    self.dropout = nn.Dropout(hidden_dropout_prob)
    self.classifier = nn.Linear(self.config.hidden_size, num_classes)
    if freeze_classifer:
        for param in self.classifier.parameters():
            param.requires_grad = False
    self.init_weights()

replace_keys_in_state_dict class-attribute instance-attribute ยค

replace_keys_in_state_dict = {
    "q_proj.weight": "q_proj.base_layer.weight",
    "v_proj.weight": "v_proj.base_layer.weight",
}

modules_to_save_checkpoints class-attribute instance-attribute ยค

modules_to_save_checkpoints = ['lora', 'classifier']

config instance-attribute ยค

config = from_json_file(config_path)

peft_config instance-attribute ยค

peft_config = LoraConfig(
    r=lora_r,
    lora_alpha=lora_alpha,
    lora_dropout=lora_dropout,
    fan_in_fan_out=fan_in_fan_out,
    target_modules=target_modules,
)

model instance-attribute ยค

model = LlamaModel(config)

dropout instance-attribute ยค

dropout = Dropout(hidden_dropout_prob)

classifier instance-attribute ยค

classifier = Linear(hidden_size, num_classes)

forward ยค

forward(
    input_ids: Tensor,
    attention_mask: Optional[Tensor] = None,
    position_ids: Optional[Tensor] = None,
)

Forward pass of the classification model.

Parameters:

Name Type Description Default
input_ids Tensor

Input tensor of shape (batch_size, sequence_length).

 required
attention_mask Tensor

Attention mask tensor of shape (batch_size, sequence_length). Defaults to None.

 None
position_ids Tensor

Position IDs tensor of shape (batch_size, sequence_length). Defaults to None.

 None

Returns:

Type Description

torch.Tensor: Output logits of shape (batch_size, num_classes).
Source code in src/unitorch/models/peft/modeling_llama.py 
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
def forward(
    self,
    input_ids: torch.Tensor,
    attention_mask: Optional[torch.Tensor] = None,
    position_ids: Optional[torch.Tensor] = None,
):
    """
    Forward pass of the classification model.

    Args:
        input_ids (torch.Tensor): Input tensor of shape (batch_size, sequence_length).
        attention_mask (torch.Tensor, optional): Attention mask tensor of shape (batch_size, sequence_length). Defaults to None.
        position_ids (torch.Tensor, optional): Position IDs tensor of shape (batch_size, sequence_length). Defaults to None.

    Returns:
        torch.Tensor: Output logits of shape (batch_size, num_classes).
    """
    outputs = self.model(
        input_ids,
        attention_mask=attention_mask,
        position_ids=position_ids,
    )[0]
    pooled_output = outputs[:, -1]
    pooled_output = self.dropout(pooled_output)
    logits = self.classifier(pooled_output)
    return logits

LlamaLoraForGenerationยค

Bases: GenericPeftModel

Source code in src/unitorch/models/peft/modeling_llama.py 
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
def __init__(
    self,
    config_path: str,
    lora_r: Optional[int] = 16,
    lora_alpha: Optional[int] = 32,
    lora_dropout: Optional[float] = 0.05,
    fan_in_fan_out: Optional[bool] = True,
    target_modules: Optional[Union[List[str], str]] = ["q_proj", "v_proj"],
    gradient_checkpointing: Optional[bool] = False,
):
    super().__init__()
    self.config = LlamaConfig.from_json_file(config_path)
    self.config.gradient_checkpointing = gradient_checkpointing
    self.peft_config = LoraConfig(
        r=lora_r,
        lora_alpha=lora_alpha,
        lora_dropout=lora_dropout,
        fan_in_fan_out=fan_in_fan_out,
        target_modules=target_modules,
    )
    self.base_model = LlamaForCausalLM(self.config)
    self.base_model.add_adapter(self.peft_config)
    self.init_weights()

prefix_keys_in_state_dict class-attribute instance-attribute ยค

prefix_keys_in_state_dict = {
    "^model.": "base_model.",
    "^lm_head.": "base_model.",
}

replace_keys_in_state_dict class-attribute instance-attribute ยค

replace_keys_in_state_dict = {
    "q_proj.weight": "q_proj.base_layer.weight",
    "v_proj.weight": "v_proj.base_layer.weight",
}

config instance-attribute ยค

config = from_json_file(config_path)

peft_config instance-attribute ยค

peft_config = LoraConfig(
    r=lora_r,
    lora_alpha=lora_alpha,
    lora_dropout=lora_dropout,
    fan_in_fan_out=fan_in_fan_out,
    target_modules=target_modules,
)

base_model instance-attribute ยค

base_model = LlamaForCausalLM(config)

forward ยค

forward(
    input_ids: Tensor,
    attention_mask: Optional[Tensor] = None,
    position_ids: Optional[Tensor] = None,
)

Forward pass of the generation model.

Parameters:

Name Type Description Default
input_ids Tensor

Input tensor of shape (batch_size, sequence_length). Defaults to None.

 required
attention_mask Tensor

Attention mask tensor of shape (batch_size, sequence_length). Defaults to None.

 None
position_ids Tensor

Position IDs tensor of shape (batch_size, sequence_length). Defaults to None.

 None

Returns:

Type Description

torch.Tensor: Output logits of shape (batch_size, sequence_length, vocab_size).
Source code in src/unitorch/models/peft/modeling_llama.py 
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
def forward(
    self,
    input_ids: torch.Tensor,
    attention_mask: Optional[torch.Tensor] = None,
    position_ids: Optional[torch.Tensor] = None,
):
    """
    Forward pass of the generation model.

    Args:
        input_ids (torch.Tensor, optional): Input tensor of shape (batch_size, sequence_length). Defaults to None.
        attention_mask (torch.Tensor, optional): Attention mask tensor of shape (batch_size, sequence_length). Defaults to None.
        position_ids (torch.Tensor, optional): Position IDs tensor of shape (batch_size, sequence_length). Defaults to None.

    Returns:
        torch.Tensor: Output logits of shape (batch_size, sequence_length, vocab_size).
    """
    outputs = self.base_model(
        input_ids=input_ids,
        attention_mask=attention_mask,
        position_ids=position_ids,
        return_dict=True,
    )
    logits = outputs.logits
    return logits

generate ยค

generate(
    input_ids: Tensor,
    num_beams: Optional[int] = 5,
    decoder_start_token_id: Optional[int] = 1,
    decoder_end_token_id: Optional[
        Union[int, List[int]]
    ] = 2,
    num_return_sequences: Optional[int] = 1,
    min_gen_seq_length: Optional[int] = 0,
    max_gen_seq_length: Optional[int] = 48,
    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,
)

Generate text using the generation model.

Parameters:

Name Type Description Default
input_ids Tensor

Input tensor of shape (batch_size, sequence_length).

 required
num_beams int

Number of beams for beam search. Defaults to 5.

 5
decoder_start_token_id int

The ID of the decoder start token. Defaults to 2.

 1
decoder_end_token_id int or List[int]

The ID(s) of the decoder end token(s). Defaults to 2.

 2
num_return_sequences int

Number of generated sequences to return. Defaults to 1.

 1
min_gen_seq_length int

Minimum length of generated sequences. Defaults to 0.

 0
max_gen_seq_length int

Maximum length of generated sequences. Defaults to 48.

 48
repetition_penalty float

Penalty for repeated tokens. Defaults to 1.0.

 1.0
no_repeat_ngram_size int

Size of n-grams to avoid repeating. Defaults to 0.

 0
early_stopping bool

Whether to stop generation early. Defaults to True.

 True
length_penalty float

Penalty for longer sequences. Defaults to 1.0.

 1.0
num_beam_groups int

Number of beam groups for diverse beam search. Defaults to 1.

 1
diversity_penalty float

Penalty for diverse sequences in diverse beam search. Defaults to 0.0.

 0.0
do_sample bool

Whether to use sampling for generation. Defaults to False.

 False
temperature float

Sampling temperature. Defaults to 1.0.

 1.0
top_k int

Top-k value for sampling. Defaults to 50.

 50
top_p float

Top-p value for sampling. Defaults to 1.0.

 1.0

Returns:

Name Type Description
GenericOutputs

Generated sequences and their scores.
Source code in src/unitorch/models/peft/modeling_llama.py 
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
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
@torch.no_grad()
def generate(
    self,
    input_ids: torch.Tensor,
    num_beams: Optional[int] = 5,
    decoder_start_token_id: Optional[int] = 1,
    decoder_end_token_id: Optional[Union[int, List[int]]] = 2,
    num_return_sequences: Optional[int] = 1,
    min_gen_seq_length: Optional[int] = 0,
    max_gen_seq_length: Optional[int] = 48,
    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,
):
    """
    Generate text using the generation model.

    Args:
        input_ids: Input tensor of shape (batch_size, sequence_length).
        num_beams (int, optional): Number of beams for beam search. Defaults to 5.
        decoder_start_token_id (int, optional): The ID of the decoder start token. Defaults to 2.
        decoder_end_token_id (int or List[int], optional): The ID(s) of the decoder end token(s). Defaults to 2.
        num_return_sequences (int, optional): Number of generated sequences to return. Defaults to 1.
        min_gen_seq_length (int, optional): Minimum length of generated sequences. Defaults to 0.
        max_gen_seq_length (int, optional): Maximum length of generated sequences. Defaults to 48.
        repetition_penalty (float, optional): Penalty for repeated tokens. Defaults to 1.0.
        no_repeat_ngram_size (int, optional): Size of n-grams to avoid repeating. Defaults to 0.
        early_stopping (bool, optional): Whether to stop generation early. Defaults to True.
        length_penalty (float, optional): Penalty for longer sequences. Defaults to 1.0.
        num_beam_groups (int, optional): Number of beam groups for diverse beam search. Defaults to 1.
        diversity_penalty (float, optional): Penalty for diverse sequences in diverse beam search. Defaults to 0.0.
        do_sample (bool, optional): Whether to use sampling for generation. Defaults to False.
        temperature (float, optional): Sampling temperature. Defaults to 1.0.
        top_k (int, optional): Top-k value for sampling. Defaults to 50.
        top_p (float, optional): Top-p value for sampling. Defaults to 1.0.

    Returns:
        GenericOutputs: Generated sequences and their scores.
    """
    input_seq_length = input_ids.size(1)
    outputs = self.base_model.generate(
        input_ids=input_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,
        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)
    )
    outputs.sequences = (
        torch.zeros(sequences.size(0), num_return_sequences, max_gen_seq_length).to(
            device=sequences.device
        )
        + decoder_start_token_id
    )
    outputs.sequences[:, :, : sequences.size(-1) - input_seq_length].copy_(
        sequences[:, :, input_seq_length : sequences.size(-1)]
    )

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

    return GenericOutputs(
        sequences=outputs.sequences.long(),
        sequences_scores=outputs.sequences_scores,
    )

LlavaMistralClipLoraForClassificationยค

Bases: GenericPeftModel

Source code in src/unitorch/models/peft/modeling_llava.py 
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
def __init__(
    self,
    config_path: str,
    image_token_index: Optional[int] = 32000,
    lora_r: Optional[int] = 16,
    lora_alpha: Optional[int] = 32,
    lora_dropout: Optional[float] = 0.05,
    fan_in_fan_out: Optional[bool] = True,
    target_modules: Optional[Union[List[str], str]] = ["q_proj", "v_proj"],
    num_classes: Optional[int] = 1,
    hidden_dropout_prob: Optional[float] = 0.1,
    freeze_multi_modal_projector: Optional[bool] = True,
    freeze_classifer: Optional[bool] = True,
    gradient_checkpointing: Optional[bool] = False,
):
    super().__init__()
    self.config = LlavaNextConfig.from_json_file(config_path)
    self.config.gradient_checkpointing = gradient_checkpointing
    self.peft_config = LoraConfig(
        r=lora_r,
        lora_alpha=lora_alpha,
        lora_dropout=lora_dropout,
        fan_in_fan_out=fan_in_fan_out,
        target_modules=target_modules,
    )
    self.vision_tower = CLIPVisionModel(self.config.vision_config)
    self.multi_modal_projector = LlavaNextMultiModalProjector(self.config)
    embed_std = 1 / math.sqrt(self.config.text_config.hidden_size)
    self.image_newline = nn.Parameter(
        torch.randn(self.config.text_config.hidden_size, dtype=self.dtype)
        * embed_std
    )
    self.language_model = MistralModel(self.config.text_config)

    self.language_model.add_adapter(self.peft_config)
    self.dropout = nn.Dropout(hidden_dropout_prob)
    self.classifier = nn.Linear(self.config.text_config.hidden_size, num_classes)
    self.init_weights()

    if freeze_classifer:
        for param in self.classifier.parameters():
            param.requires_grad = False

    for param in self.vision_tower.parameters():
        param.requires_grad = False

    if freeze_multi_modal_projector:
        for param in self.multi_modal_projector.parameters():
            param.requires_grad = False

    self.image_token_index = image_token_index

replace_keys_in_state_dict class-attribute instance-attribute ยค

replace_keys_in_state_dict = {
    "(language_model.*?)q_proj\\.weight": "\\1q_proj.base_layer.weight",
    "(language_model.*?)v_proj\\.weight": "\\1v_proj.base_layer.weight",
    "language_model.": "",
}

modules_to_save_checkpoints class-attribute instance-attribute ยค

modules_to_save_checkpoints = [
    "lora",
    "multi_modal_projector",
    "classifier",
]

config instance-attribute ยค

config = from_json_file(config_path)

peft_config instance-attribute ยค

peft_config = LoraConfig(
    r=lora_r,
    lora_alpha=lora_alpha,
    lora_dropout=lora_dropout,
    fan_in_fan_out=fan_in_fan_out,
    target_modules=target_modules,
)

vision_tower instance-attribute ยค

vision_tower = CLIPVisionModel(vision_config)

multi_modal_projector instance-attribute ยค

multi_modal_projector = LlavaNextMultiModalProjector(config)

image_newline instance-attribute ยค

image_newline = Parameter(
    randn(hidden_size, dtype=dtype) * embed_std
)

language_model instance-attribute ยค

language_model = MistralModel(text_config)

dropout instance-attribute ยค

dropout = Dropout(hidden_dropout_prob)

classifier instance-attribute ยค

classifier = Linear(hidden_size, num_classes)

image_token_index instance-attribute ยค

image_token_index = image_token_index

forward ยค

forward(
    input_ids: Tensor,
    pixel_values: Tensor,
    attention_mask: Optional[Tensor] = None,
)

Forward pass of the classification model.

Parameters:

Name Type Description Default
input_ids Tensor

Input tensor of shape (batch_size, sequence_length).

 required
attention_mask Tensor

Attention mask tensor of shape (batch_size, sequence_length). Defaults to None.

 None
position_ids Tensor

Position IDs tensor of shape (batch_size, sequence_length). Defaults to None.

 required

Returns:

Type Description

torch.Tensor: Output logits of shape (batch_size, num_classes).
Source code in src/unitorch/models/peft/modeling_llava.py 
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
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
163
164
165
166
167
168
169
170
171
def forward(
    self,
    input_ids: torch.Tensor,
    pixel_values: torch.Tensor,
    attention_mask: Optional[torch.Tensor] = None,
):
    """
    Forward pass of the classification model.

    Args:
        input_ids (torch.Tensor): Input tensor of shape (batch_size, sequence_length).
        attention_mask (torch.Tensor, optional): Attention mask tensor of shape (batch_size, sequence_length). Defaults to None.
        position_ids (torch.Tensor, optional): Position IDs tensor of shape (batch_size, sequence_length). Defaults to None.

    Returns:
        torch.Tensor: Output logits of shape (batch_size, num_classes).
    """
    vision_outputs = self.vision_tower(pixel_values, output_hidden_states=True)
    image_embeds = vision_outputs.hidden_states[-2][:, 1:]
    image_embeds = self.multi_modal_projector(image_embeds)
    image_embeds = torch.cat(
        [
            image_embeds,
            self.image_newline.expand(
                image_embeds.shape[0], 1, image_embeds.shape[-1]
            ),
        ],
        dim=1,
    )

    image_seq_length = image_embeds.size(1)
    batch_indices, text_indices = torch.where(input_ids != self.image_token_index)
    image_masks = (input_ids == self.image_token_index).long() * (
        image_seq_length - 1
    )
    new_positions = torch.cumsum(image_masks + 1, dim=1) - 1
    new_text_indices = new_positions[batch_indices, text_indices]

    input_ids[input_ids == self.image_token_index] = 0
    text_embeds = self.language_model.get_input_embeddings()(input_ids)

    batch_size, text_seq_length, text_dim = text_embeds.size()

    if attention_mask is None:
        attention_mask = torch.ones(batch_size, text_seq_length).to(
            text_embeds.device
        )

    final_embeds = torch.zeros(
        batch_size, text_seq_length + image_seq_length - 1, text_dim
    ).to(text_embeds.device)
    overwrite_masks = torch.ones(
        batch_size, text_seq_length + image_seq_length - 1
    ).to(text_embeds.device)
    overwrite_masks[batch_indices, new_text_indices] = 0
    final_embeds[overwrite_masks == 0] = text_embeds[
        batch_indices, text_indices
    ].to(final_embeds)
    final_embeds[overwrite_masks == 1] = (
        image_embeds.contiguous().view(-1, text_dim).to(final_embeds)
    )
    final_masks = torch.zeros(
        batch_size, text_seq_length + image_seq_length - 1
    ).to(attention_mask)
    final_masks[overwrite_masks == 0] = attention_mask[
        batch_indices, text_indices
    ].to(final_masks)
    final_masks[overwrite_masks == 1] = 1
    position_ids = (final_masks.cumsum(dim=1) - 1).masked_fill(final_masks == 0, -1)

    outputs = self.language_model(
        inputs_embeds=final_embeds,
        attention_mask=final_masks,
        position_ids=position_ids,
    )[0]
    pooled_output = outputs[:, -1]
    pooled_output = self.dropout(pooled_output)
    logits = self.classifier(pooled_output)
    return logits

LlavaMistralClipLoraForGenerationยค

Bases: GenericPeftModel

Source code in src/unitorch/models/peft/modeling_llava.py 
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
def __init__(
    self,
    config_path: str,
    image_token_index: Optional[int] = 32000,
    lora_r: Optional[int] = 16,
    lora_alpha: Optional[int] = 32,
    lora_dropout: Optional[float] = 0.05,
    fan_in_fan_out: Optional[bool] = True,
    target_modules: Optional[Union[List[str], str]] = ["q_proj", "v_proj"],
    freeze_multi_modal_projector: Optional[bool] = True,
    gradient_checkpointing: Optional[bool] = False,
    pad_token_id: Optional[int] = 0,
):
    super().__init__()
    self.config = LlavaNextConfig.from_json_file(config_path)
    self.config.gradient_checkpointing = gradient_checkpointing
    self.config.pad_token_id = pad_token_id
    self.peft_config = LoraConfig(
        r=lora_r,
        lora_alpha=lora_alpha,
        lora_dropout=lora_dropout,
        fan_in_fan_out=fan_in_fan_out,
        target_modules=target_modules,
    )
    self.vision_tower = CLIPVisionModel(self.config.vision_config)
    self.multi_modal_projector = LlavaNextMultiModalProjector(self.config)
    embed_std = 1 / math.sqrt(self.config.text_config.hidden_size)
    self.image_newline = nn.Parameter(
        torch.randn(self.config.text_config.hidden_size, dtype=self.dtype)
        * embed_std
    )
    self.language_model = MistralForCausalLM(self.config.text_config)

    self.language_model.add_adapter(self.peft_config)
    self.init_weights()

    for param in self.vision_tower.parameters():
        param.requires_grad = False

    if freeze_multi_modal_projector:
        for param in self.multi_modal_projector.parameters():
            param.requires_grad = False

    self.image_token_index = image_token_index

replace_keys_in_state_dict class-attribute instance-attribute ยค

replace_keys_in_state_dict = {
    "(language_model.*?)q_proj\\.weight": "\\1q_proj.base_layer.weight",
    "(language_model.*?)v_proj\\.weight": "\\1v_proj.base_layer.weight",
    "language_model\\.": "model.",
}

config instance-attribute ยค

config = from_json_file(config_path)

peft_config instance-attribute ยค

peft_config = LoraConfig(
    r=lora_r,
    lora_alpha=lora_alpha,
    lora_dropout=lora_dropout,
    fan_in_fan_out=fan_in_fan_out,
    target_modules=target_modules,
)

vision_tower instance-attribute ยค

vision_tower = CLIPVisionModel(vision_config)

multi_modal_projector instance-attribute ยค

multi_modal_projector = LlavaNextMultiModalProjector(config)

image_newline instance-attribute ยค

image_newline = Parameter(
    randn(hidden_size, dtype=dtype) * embed_std
)

language_model instance-attribute ยค

language_model = MistralForCausalLM(text_config)

image_token_index instance-attribute ยค

image_token_index = image_token_index

forward ยค

forward(
    input_ids: Tensor,
    pixel_values: Tensor,
    attention_mask: Optional[Tensor] = None,
    position_ids: Optional[Tensor] = None,
)

Forward pass of the generation model.

Parameters:

Name Type Description Default
input_ids Tensor

Input tensor of shape (batch_size, sequence_length). Defaults to None.

 required
attention_mask Tensor

Attention mask tensor of shape (batch_size, sequence_length). Defaults to None.

 None
position_ids Tensor

Position IDs tensor of shape (batch_size, sequence_length). Defaults to None.

 None

Returns:

Type Description

torch.Tensor: Output logits of shape (batch_size, sequence_length, vocab_size).
Source code in src/unitorch/models/peft/modeling_llava.py 
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
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
def forward(
    self,
    input_ids: torch.Tensor,
    pixel_values: torch.Tensor,
    attention_mask: Optional[torch.Tensor] = None,
    position_ids: Optional[torch.Tensor] = None,
):
    """
    Forward pass of the generation model.

    Args:
        input_ids (torch.Tensor, optional): Input tensor of shape (batch_size, sequence_length). Defaults to None.
        attention_mask (torch.Tensor, optional): Attention mask tensor of shape (batch_size, sequence_length). Defaults to None.
        position_ids (torch.Tensor, optional): Position IDs tensor of shape (batch_size, sequence_length). Defaults to None.

    Returns:
        torch.Tensor: Output logits of shape (batch_size, sequence_length, vocab_size).
    """
    vision_outputs = self.vision_tower(pixel_values, output_hidden_states=True)
    image_embeds = vision_outputs.hidden_states[-2][:, 1:]
    image_embeds = self.multi_modal_projector(image_embeds)
    image_embeds = torch.cat(
        [
            image_embeds,
            self.image_newline.expand(
                image_embeds.shape[0], 1, image_embeds.shape[-1]
            ),
        ],
        dim=1,
    )

    image_seq_length = image_embeds.size(1)
    batch_indices, text_indices = torch.where(input_ids != self.image_token_index)
    image_masks = (input_ids == self.image_token_index).long() * (
        image_seq_length - 1
    )
    new_positions = torch.cumsum(image_masks + 1, dim=1) - 1
    new_text_indices = new_positions[batch_indices, text_indices]

    input_ids[input_ids == self.image_token_index] = 0
    text_embeds = self.language_model.get_input_embeddings()(input_ids)

    batch_size, text_seq_length, text_dim = text_embeds.size()

    if attention_mask is None:
        attention_mask = torch.ones(batch_size, text_seq_length).to(
            text_embeds.device
        )

    final_embeds = torch.zeros(
        batch_size, text_seq_length + image_seq_length - 1, text_dim
    ).to(text_embeds.device)
    overwrite_masks = torch.ones(
        batch_size, text_seq_length + image_seq_length - 1
    ).to(text_embeds.device)
    overwrite_masks[batch_indices, new_text_indices] = 0
    final_embeds[overwrite_masks == 0] = text_embeds[
        batch_indices, text_indices
    ].to(final_embeds)
    final_embeds[overwrite_masks == 1] = (
        image_embeds.contiguous().view(-1, text_dim).to(final_embeds)
    )
    final_masks = torch.zeros(
        batch_size, text_seq_length + image_seq_length - 1
    ).to(attention_mask)
    final_masks[overwrite_masks == 0] = attention_mask[
        batch_indices, text_indices
    ].to(final_masks)
    final_masks[overwrite_masks == 1] = 1
    position_ids = (final_masks.cumsum(dim=1) - 1).masked_fill(final_masks == 0, -1)

    outputs = self.language_model(
        inputs_embeds=final_embeds,
        attention_mask=final_masks,
        position_ids=position_ids,
    )
    logits = torch.zeros(batch_size, text_seq_length, outputs.logits.size(-1)).to(
        outputs.logits.device
    )
    logits[batch_indices, text_indices] = outputs.logits[
        batch_indices, new_text_indices
    ]
    return logits

generate ยค

generate(
    input_ids: Tensor,
    pixel_values: Tensor,
    attention_mask: Optional[Tensor] = None,
    num_beams: Optional[int] = 5,
    decoder_start_token_id: Optional[int] = 1,
    decoder_end_token_id: Optional[
        Union[int, List[int]]
    ] = 2,
    num_return_sequences: Optional[int] = 1,
    min_gen_seq_length: Optional[int] = 0,
    max_gen_seq_length: Optional[int] = 48,
    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,
)

Generate text using the generation model.

Parameters:

Name Type Description Default
input_ids Tensor

Input tensor of shape (batch_size, sequence_length).

 required
num_beams int

Number of beams for beam search. Defaults to 5.

 5
decoder_start_token_id int

The ID of the decoder start token. Defaults to 2.

 1
decoder_end_token_id int or List[int]

The ID(s) of the decoder end token(s). Defaults to 2.

 2
num_return_sequences int

Number of generated sequences to return. Defaults to 1.

 1
min_gen_seq_length int

Minimum length of generated sequences. Defaults to 0.

 0
max_gen_seq_length int

Maximum length of generated sequences. Defaults to 48.

 48
repetition_penalty float

Penalty for repeated tokens. Defaults to 1.0.

 1.0
no_repeat_ngram_size int

Size of n-grams to avoid repeating. Defaults to 0.

 0
early_stopping bool

Whether to stop generation early. Defaults to True.

 True
length_penalty float

Penalty for longer sequences. Defaults to 1.0.

 1.0
num_beam_groups int

Number of beam groups for diverse beam search. Defaults to 1.

 1
diversity_penalty float

Penalty for diverse sequences in diverse beam search. Defaults to 0.0.

 0.0
do_sample bool

Whether to use sampling for generation. Defaults to False.

 False
temperature float

Sampling temperature. Defaults to 1.0.

 1.0
top_k int

Top-k value for sampling. Defaults to 50.

 50
top_p float

Top-p value for sampling. Defaults to 1.0.

 1.0

Returns:

Name Type Description
GenericOutputs

Generated sequences and their scores.
Source code in src/unitorch/models/peft/modeling_llava.py 
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
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
417
418
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
447
448
449
450
451
452
453
454
@torch.no_grad()
def generate(
    self,
    input_ids: torch.Tensor,
    pixel_values: torch.Tensor,
    attention_mask: Optional[torch.Tensor] = None,
    num_beams: Optional[int] = 5,
    decoder_start_token_id: Optional[int] = 1,
    decoder_end_token_id: Optional[Union[int, List[int]]] = 2,
    num_return_sequences: Optional[int] = 1,
    min_gen_seq_length: Optional[int] = 0,
    max_gen_seq_length: Optional[int] = 48,
    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,
):
    """
    Generate text using the generation model.

    Args:
        input_ids: Input tensor of shape (batch_size, sequence_length).
        num_beams (int, optional): Number of beams for beam search. Defaults to 5.
        decoder_start_token_id (int, optional): The ID of the decoder start token. Defaults to 2.
        decoder_end_token_id (int or List[int], optional): The ID(s) of the decoder end token(s). Defaults to 2.
        num_return_sequences (int, optional): Number of generated sequences to return. Defaults to 1.
        min_gen_seq_length (int, optional): Minimum length of generated sequences. Defaults to 0.
        max_gen_seq_length (int, optional): Maximum length of generated sequences. Defaults to 48.
        repetition_penalty (float, optional): Penalty for repeated tokens. Defaults to 1.0.
        no_repeat_ngram_size (int, optional): Size of n-grams to avoid repeating. Defaults to 0.
        early_stopping (bool, optional): Whether to stop generation early. Defaults to True.
        length_penalty (float, optional): Penalty for longer sequences. Defaults to 1.0.
        num_beam_groups (int, optional): Number of beam groups for diverse beam search. Defaults to 1.
        diversity_penalty (float, optional): Penalty for diverse sequences in diverse beam search. Defaults to 0.0.
        do_sample (bool, optional): Whether to use sampling for generation. Defaults to False.
        temperature (float, optional): Sampling temperature. Defaults to 1.0.
        top_k (int, optional): Top-k value for sampling. Defaults to 50.
        top_p (float, optional): Top-p value for sampling. Defaults to 1.0.

    Returns:
        GenericOutputs: Generated sequences and their scores.
    """
    vision_outputs = self.vision_tower(pixel_values, output_hidden_states=True)
    image_embeds = vision_outputs.hidden_states[-2][:, 1:]
    image_embeds = self.multi_modal_projector(image_embeds)
    image_embeds = torch.cat(
        [
            image_embeds,
            self.image_newline.expand(
                image_embeds.shape[0], 1, image_embeds.shape[-1]
            ),
        ],
        dim=1,
    )

    image_seq_length = image_embeds.size(1)
    batch_indices, text_indices = torch.where(input_ids != self.image_token_index)
    image_masks = (input_ids == self.image_token_index).long() * (
        image_seq_length - 1
    )
    new_positions = torch.cumsum(image_masks + 1, dim=1) - 1
    new_text_indices = new_positions[batch_indices, text_indices]

    input_ids[input_ids == self.image_token_index] = 0
    text_embeds = self.language_model.get_input_embeddings()(input_ids)

    batch_size = text_embeds.size(0)
    text_seq_length, image_seq_length = text_embeds.size(1), image_embeds.size(1)
    text_dim = text_embeds.size(2)

    if attention_mask is None:
        attention_mask = torch.ones(batch_size, text_seq_length).to(
            text_embeds.device
        )

    final_embeds = torch.zeros(
        batch_size, text_seq_length + image_seq_length - 1, text_dim
    ).to(text_embeds.device)
    overwrite_masks = torch.ones(
        batch_size, text_seq_length + image_seq_length - 1
    ).to(text_embeds.device)
    overwrite_masks[batch_indices, new_text_indices] = 0
    final_embeds[overwrite_masks == 0] = text_embeds[
        batch_indices, text_indices
    ].to(final_embeds)
    final_embeds[overwrite_masks == 1] = (
        image_embeds.contiguous().view(-1, text_dim).to(final_embeds)
    )
    final_masks = torch.zeros(
        batch_size, text_seq_length + image_seq_length - 1
    ).to(attention_mask)
    final_masks[overwrite_masks == 0] = attention_mask[
        batch_indices, text_indices
    ].to(final_masks)
    final_masks[overwrite_masks == 1] = 1
    input_seq_length = final_embeds.size(1)
    outputs = self.language_model.generate(
        inputs_embeds=final_embeds,
        attention_mask=final_masks,
        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,
        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)
    )
    outputs.sequences = (
        torch.zeros(sequences.size(0), num_return_sequences, max_gen_seq_length).to(
            device=sequences.device
        )
        + decoder_start_token_id
    )
    outputs.sequences[:, :, : sequences.size(-1) - input_seq_length].copy_(
        sequences[:, :, input_seq_length : sequences.size(-1)]
    )

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

    return GenericOutputs(
        sequences=outputs.sequences.long(),
        sequences_scores=outputs.sequences_scores,
    )

LlavaLlamaSiglipLoraForGenerationยค

Bases: GenericPeftModel

Source code in src/unitorch/models/peft/modeling_llava.py 
464
465
466
467
468
469
470
471
472
473
474
475
476
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
def __init__(
    self,
    config_path: str,
    image_token_index: Optional[int] = 128077,
    lora_r: Optional[int] = 16,
    lora_alpha: Optional[int] = 32,
    lora_dropout: Optional[float] = 0.05,
    fan_in_fan_out: Optional[bool] = True,
    target_modules: Optional[Union[List[str], str]] = ["q_proj", "v_proj"],
    freeze_multi_modal_projector: Optional[bool] = True,
    gradient_checkpointing: Optional[bool] = False,
    pad_token_id: Optional[int] = 0,
):
    super().__init__()
    self.config = LlavaConfig.from_json_file(config_path)
    self.config.gradient_checkpointing = gradient_checkpointing
    self.config.pad_token_id = pad_token_id
    self.peft_config = LoraConfig(
        r=lora_r,
        lora_alpha=lora_alpha,
        lora_dropout=lora_dropout,
        fan_in_fan_out=fan_in_fan_out,
        target_modules=target_modules,
    )
    self.vision_tower = SiglipVisionModel(self.config.vision_config)
    self.multi_modal_projector = LlavaMultiModalProjector(self.config)
    self.language_model = LlamaForCausalLM(self.config.text_config)

    self.language_model.add_adapter(self.peft_config)
    self.init_weights()

    for param in self.vision_tower.parameters():
        param.requires_grad = False

    if freeze_multi_modal_projector:
        for param in self.multi_modal_projector.parameters():
            param.requires_grad = False

    self.image_token_index = image_token_index

replace_keys_in_state_dict class-attribute instance-attribute ยค

replace_keys_in_state_dict = {
    "(language_model.*?)q_proj\\.weight": "\\1q_proj.base_layer.weight",
    "(language_model.*?)v_proj\\.weight": "\\1v_proj.base_layer.weight",
    "language_model\\.": "model.",
}

config instance-attribute ยค

config = from_json_file(config_path)

peft_config instance-attribute ยค

peft_config = LoraConfig(
    r=lora_r,
    lora_alpha=lora_alpha,
    lora_dropout=lora_dropout,
    fan_in_fan_out=fan_in_fan_out,
    target_modules=target_modules,
)

vision_tower instance-attribute ยค

vision_tower = SiglipVisionModel(vision_config)

multi_modal_projector instance-attribute ยค

multi_modal_projector = LlavaMultiModalProjector(config)

language_model instance-attribute ยค

language_model = LlamaForCausalLM(text_config)

image_token_index instance-attribute ยค

image_token_index = image_token_index

forward ยค

forward(
    input_ids: Tensor,
    pixel_values: Tensor,
    attention_mask: Optional[Tensor] = None,
    position_ids: Optional[Tensor] = None,
)

Forward pass of the generation model.

Parameters:

Name Type Description Default
input_ids Tensor

Input tensor of shape (batch_size, sequence_length). Defaults to None.

 required
attention_mask Tensor

Attention mask tensor of shape (batch_size, sequence_length). Defaults to None.

 None
position_ids Tensor

Position IDs tensor of shape (batch_size, sequence_length). Defaults to None.

 None

Returns:

Type Description

torch.Tensor: Output logits of shape (batch_size, sequence_length, vocab_size).
Source code in src/unitorch/models/peft/modeling_llava.py 
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
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
def forward(
    self,
    input_ids: torch.Tensor,
    pixel_values: torch.Tensor,
    attention_mask: Optional[torch.Tensor] = None,
    position_ids: Optional[torch.Tensor] = None,
):
    """
    Forward pass of the generation model.

    Args:
        input_ids (torch.Tensor, optional): Input tensor of shape (batch_size, sequence_length). Defaults to None.
        attention_mask (torch.Tensor, optional): Attention mask tensor of shape (batch_size, sequence_length). Defaults to None.
        position_ids (torch.Tensor, optional): Position IDs tensor of shape (batch_size, sequence_length). Defaults to None.

    Returns:
        torch.Tensor: Output logits of shape (batch_size, sequence_length, vocab_size).
    """
    vision_outputs = self.vision_tower(pixel_values, output_hidden_states=True)
    image_embeds = vision_outputs.hidden_states[-2]
    image_embeds = self.multi_modal_projector(image_embeds)
    image_seq_length = image_embeds.size(1)
    batch_indices, text_indices = torch.where(input_ids != self.image_token_index)
    image_masks = (input_ids == self.image_token_index).long() * (
        image_seq_length - 1
    )
    new_positions = torch.cumsum(image_masks + 1, dim=1) - 1
    new_text_indices = new_positions[batch_indices, text_indices]

    input_ids[input_ids == self.image_token_index] = 0
    text_embeds = self.language_model.get_input_embeddings()(input_ids)

    batch_size, text_seq_length, text_dim = text_embeds.size()

    if attention_mask is None:
        attention_mask = torch.ones(batch_size, text_seq_length).to(
            text_embeds.device
        )

    final_embeds = torch.zeros(
        batch_size, text_seq_length + image_seq_length - 1, text_dim
    ).to(text_embeds.device)
    overwrite_masks = torch.ones(
        batch_size, text_seq_length + image_seq_length - 1
    ).to(text_embeds.device)
    overwrite_masks[batch_indices, new_text_indices] = 0
    final_embeds[overwrite_masks == 0] = text_embeds[
        batch_indices, text_indices
    ].to(final_embeds)
    final_embeds[overwrite_masks == 1] = (
        image_embeds.contiguous().view(-1, text_dim).to(final_embeds)
    )
    final_masks = torch.zeros(
        batch_size, text_seq_length + image_seq_length - 1
    ).to(attention_mask)
    final_masks[overwrite_masks == 0] = attention_mask[
        batch_indices, text_indices
    ].to(final_masks)
    final_masks[overwrite_masks == 1] = 1
    position_ids = (final_masks.cumsum(dim=1) - 1).masked_fill(final_masks == 0, -1)

    outputs = self.language_model(
        inputs_embeds=final_embeds,
        attention_mask=final_masks,
        position_ids=position_ids,
    )
    logits = torch.zeros(batch_size, text_seq_length, outputs.logits.size(-1)).to(
        outputs.logits.device
    )
    logits[batch_indices, text_indices] = outputs.logits[
        batch_indices, new_text_indices
    ]
    return logits

generate ยค

generate(
    input_ids: Tensor,
    pixel_values: Tensor,
    attention_mask: Optional[Tensor] = None,
    num_beams: Optional[int] = 5,
    decoder_start_token_id: Optional[int] = 128000,
    decoder_end_token_id: Optional[
        Union[int, List[int]]
    ] = None,
    num_return_sequences: Optional[int] = 1,
    min_gen_seq_length: Optional[int] = 0,
    max_gen_seq_length: Optional[int] = 48,
    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,
)

Generate text using the generation model.

Parameters:

Name Type Description Default
input_ids Tensor

Input tensor of shape (batch_size, sequence_length).

 required
num_beams int

Number of beams for beam search. Defaults to 5.

 5
decoder_start_token_id int

The ID of the decoder start token. Defaults to 2.

 128000
decoder_end_token_id int or List[int]

The ID(s) of the decoder end token(s). Defaults to 2.

 None
num_return_sequences int

Number of generated sequences to return. Defaults to 1.

 1
min_gen_seq_length int

Minimum length of generated sequences. Defaults to 0.

 0
max_gen_seq_length int

Maximum length of generated sequences. Defaults to 48.

 48
repetition_penalty float

Penalty for repeated tokens. Defaults to 1.0.

 1.0
no_repeat_ngram_size int

Size of n-grams to avoid repeating. Defaults to 0.

 0
early_stopping bool

Whether to stop generation early. Defaults to True.

 True
length_penalty float

Penalty for longer sequences. Defaults to 1.0.

 1.0
num_beam_groups int

Number of beam groups for diverse beam search. Defaults to 1.

 1
diversity_penalty float

Penalty for diverse sequences in diverse beam search. Defaults to 0.0.

 0.0
do_sample bool

Whether to use sampling for generation. Defaults to False.

 False
temperature float

Sampling temperature. Defaults to 1.0.

 1.0
top_k int

Top-k value for sampling. Defaults to 50.

 50
top_p float

Top-p value for sampling. Defaults to 1.0.

 1.0

Returns:

Name Type Description
GenericOutputs

Generated sequences and their scores.
Source code in src/unitorch/models/peft/modeling_llava.py 
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
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
@torch.no_grad()
def generate(
    self,
    input_ids: torch.Tensor,
    pixel_values: torch.Tensor,
    attention_mask: Optional[torch.Tensor] = None,
    num_beams: Optional[int] = 5,
    decoder_start_token_id: Optional[int] = 128000,
    decoder_end_token_id: Optional[Union[int, List[int]]] = None,
    num_return_sequences: Optional[int] = 1,
    min_gen_seq_length: Optional[int] = 0,
    max_gen_seq_length: Optional[int] = 48,
    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,
):
    """
    Generate text using the generation model.

    Args:
        input_ids: Input tensor of shape (batch_size, sequence_length).
        num_beams (int, optional): Number of beams for beam search. Defaults to 5.
        decoder_start_token_id (int, optional): The ID of the decoder start token. Defaults to 2.
        decoder_end_token_id (int or List[int], optional): The ID(s) of the decoder end token(s). Defaults to 2.
        num_return_sequences (int, optional): Number of generated sequences to return. Defaults to 1.
        min_gen_seq_length (int, optional): Minimum length of generated sequences. Defaults to 0.
        max_gen_seq_length (int, optional): Maximum length of generated sequences. Defaults to 48.
        repetition_penalty (float, optional): Penalty for repeated tokens. Defaults to 1.0.
        no_repeat_ngram_size (int, optional): Size of n-grams to avoid repeating. Defaults to 0.
        early_stopping (bool, optional): Whether to stop generation early. Defaults to True.
        length_penalty (float, optional): Penalty for longer sequences. Defaults to 1.0.
        num_beam_groups (int, optional): Number of beam groups for diverse beam search. Defaults to 1.
        diversity_penalty (float, optional): Penalty for diverse sequences in diverse beam search. Defaults to 0.0.
        do_sample (bool, optional): Whether to use sampling for generation. Defaults to False.
        temperature (float, optional): Sampling temperature. Defaults to 1.0.
        top_k (int, optional): Top-k value for sampling. Defaults to 50.
        top_p (float, optional): Top-p value for sampling. Defaults to 1.0.

    Returns:
        GenericOutputs: Generated sequences and their scores.
    """
    vision_outputs = self.vision_tower(pixel_values, output_hidden_states=True)
    image_embeds = vision_outputs.hidden_states[-2]
    image_embeds = self.multi_modal_projector(image_embeds)
    image_seq_length = image_embeds.size(1)
    batch_indices, text_indices = torch.where(input_ids != self.image_token_index)
    image_masks = (input_ids == self.image_token_index).long() * (
        image_seq_length - 1
    )
    new_positions = torch.cumsum(image_masks + 1, dim=1) - 1
    new_text_indices = new_positions[batch_indices, text_indices]

    input_ids[input_ids == self.image_token_index] = 0
    text_embeds = self.language_model.get_input_embeddings()(input_ids)

    batch_size = text_embeds.size(0)
    text_seq_length, image_seq_length = text_embeds.size(1), image_embeds.size(1)
    text_dim = text_embeds.size(2)

    if attention_mask is None:
        attention_mask = torch.ones(batch_size, text_seq_length).to(
            text_embeds.device
        )

    final_embeds = torch.zeros(
        batch_size, text_seq_length + image_seq_length - 1, text_dim
    ).to(text_embeds.device)
    overwrite_masks = torch.ones(
        batch_size, text_seq_length + image_seq_length - 1
    ).to(text_embeds.device)
    overwrite_masks[batch_indices, new_text_indices] = 0
    final_embeds[overwrite_masks == 0] = text_embeds[
        batch_indices, text_indices
    ].to(final_embeds)
    final_embeds[overwrite_masks == 1] = (
        image_embeds.contiguous().view(-1, text_dim).to(final_embeds)
    )
    final_masks = torch.zeros(
        batch_size, text_seq_length + image_seq_length - 1
    ).to(attention_mask)
    final_masks[overwrite_masks == 0] = attention_mask[
        batch_indices, text_indices
    ].to(final_masks)
    final_masks[overwrite_masks == 1] = 1
    input_seq_length = final_embeds.size(1)
    outputs = self.language_model.generate(
        inputs_embeds=final_embeds,
        attention_mask=final_masks,
        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,
        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)
    )
    outputs.sequences = (
        torch.zeros(sequences.size(0), num_return_sequences, max_gen_seq_length).to(
            device=sequences.device
        )
        + decoder_start_token_id
    )
    outputs.sequences[:, :, : sequences.size(-1) - input_seq_length].copy_(
        sequences[:, :, input_seq_length : sequences.size(-1)]
    )

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

    return GenericOutputs(
        sequences=outputs.sequences.long(),
        sequences_scores=outputs.sequences_scores,
    )

MistralLoraForClassificationยค

Bases: GenericPeftModel

Source code in src/unitorch/models/peft/modeling_mistral.py 
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
def __init__(
    self,
    config_path: str,
    lora_r: Optional[int] = 16,
    lora_alpha: Optional[int] = 32,
    lora_dropout: Optional[float] = 0.05,
    fan_in_fan_out: Optional[bool] = True,
    target_modules: Optional[Union[List[str], str]] = ["q_proj", "v_proj"],
    num_classes: Optional[int] = 1,
    hidden_dropout_prob: Optional[float] = 0.1,
    freeze_classifer: Optional[bool] = True,
    gradient_checkpointing: Optional[bool] = False,
):
    super().__init__()
    self.config = MistralConfig.from_json_file(config_path)
    self.config.gradient_checkpointing = gradient_checkpointing
    self.peft_config = LoraConfig(
        r=lora_r,
        lora_alpha=lora_alpha,
        lora_dropout=lora_dropout,
        fan_in_fan_out=fan_in_fan_out,
        target_modules=target_modules,
    )
    self.model = MistralModel(self.config)
    self.model.add_adapter(self.peft_config)
    self.dropout = nn.Dropout(hidden_dropout_prob)
    self.classifier = nn.Linear(self.config.hidden_size, num_classes)
    if freeze_classifer:
        for param in self.classifier.parameters():
            param.requires_grad = False
    self.init_weights()

replace_keys_in_state_dict class-attribute instance-attribute ยค

replace_keys_in_state_dict = {
    "q_proj.weight": "q_proj.base_layer.weight",
    "v_proj.weight": "v_proj.base_layer.weight",
}

modules_to_save_checkpoints class-attribute instance-attribute ยค

modules_to_save_checkpoints = ['lora', 'classifier']

config instance-attribute ยค

config = from_json_file(config_path)

peft_config instance-attribute ยค

peft_config = LoraConfig(
    r=lora_r,
    lora_alpha=lora_alpha,
    lora_dropout=lora_dropout,
    fan_in_fan_out=fan_in_fan_out,
    target_modules=target_modules,
)

model instance-attribute ยค

model = MistralModel(config)

dropout instance-attribute ยค

dropout = Dropout(hidden_dropout_prob)

classifier instance-attribute ยค

classifier = Linear(hidden_size, num_classes)

forward ยค

forward(
    input_ids: Tensor,
    attention_mask: Optional[Tensor] = None,
    position_ids: Optional[Tensor] = None,
)

Forward pass of the classification model.

Parameters:

Name Type Description Default
input_ids Tensor

Input tensor of shape (batch_size, sequence_length).

 required
attention_mask Tensor

Attention mask tensor of shape (batch_size, sequence_length). Defaults to None.

 None
position_ids Tensor

Position IDs tensor of shape (batch_size, sequence_length). Defaults to None.

 None

Returns:

Type Description

torch.Tensor: Output logits of shape (batch_size, num_classes).
Source code in src/unitorch/models/peft/modeling_mistral.py 
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
def forward(
    self,
    input_ids: torch.Tensor,
    attention_mask: Optional[torch.Tensor] = None,
    position_ids: Optional[torch.Tensor] = None,
):
    """
    Forward pass of the classification model.

    Args:
        input_ids (torch.Tensor): Input tensor of shape (batch_size, sequence_length).
        attention_mask (torch.Tensor, optional): Attention mask tensor of shape (batch_size, sequence_length). Defaults to None.
        position_ids (torch.Tensor, optional): Position IDs tensor of shape (batch_size, sequence_length). Defaults to None.

    Returns:
        torch.Tensor: Output logits of shape (batch_size, num_classes).
    """
    outputs = self.model(
        input_ids,
        attention_mask=attention_mask,
        position_ids=position_ids,
    )[0]
    pooled_output = outputs[:, -1]
    pooled_output = self.dropout(pooled_output)
    logits = self.classifier(pooled_output)
    return logits

MistralLoraForGenerationยค

Bases: GenericPeftModel

Source code in src/unitorch/models/peft/modeling_mistral.py 
 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
def __init__(
    self,
    config_path: str,
    lora_r: Optional[int] = 16,
    lora_alpha: Optional[int] = 32,
    lora_dropout: Optional[float] = 0.05,
    fan_in_fan_out: Optional[bool] = True,
    target_modules: Optional[Union[List[str], str]] = ["q_proj", "v_proj"],
    gradient_checkpointing: Optional[bool] = False,
    pad_token_id: Optional[int] = 0,
):
    super().__init__()
    self.config = MistralConfig.from_json_file(config_path)
    self.config.gradient_checkpointing = gradient_checkpointing
    self.config.pad_token_id = pad_token_id
    self.peft_config = LoraConfig(
        r=lora_r,
        lora_alpha=lora_alpha,
        lora_dropout=lora_dropout,
        fan_in_fan_out=fan_in_fan_out,
        target_modules=target_modules,
    )
    self.base_model = MistralForCausalLM(self.config)
    self.base_model.add_adapter(self.peft_config)
    self.init_weights()

prefix_keys_in_state_dict class-attribute instance-attribute ยค

prefix_keys_in_state_dict = {
    "^model.": "base_model.",
    "^lm_head.": "base_model.",
}

replace_keys_in_state_dict class-attribute instance-attribute ยค

replace_keys_in_state_dict = {
    "q_proj.weight": "q_proj.base_layer.weight",
    "v_proj.weight": "v_proj.base_layer.weight",
}

config instance-attribute ยค

config = from_json_file(config_path)

peft_config instance-attribute ยค

peft_config = LoraConfig(
    r=lora_r,
    lora_alpha=lora_alpha,
    lora_dropout=lora_dropout,
    fan_in_fan_out=fan_in_fan_out,
    target_modules=target_modules,
)

base_model instance-attribute ยค

base_model = MistralForCausalLM(config)

forward ยค

forward(
    input_ids: Tensor,
    attention_mask: Optional[Tensor] = None,
    position_ids: Optional[Tensor] = None,
)

Forward pass of the generation model.

Parameters:

Name Type Description Default
input_ids Tensor

Input tensor of shape (batch_size, sequence_length). Defaults to None.

 required
attention_mask Tensor

Attention mask tensor of shape (batch_size, sequence_length). Defaults to None.

 None
position_ids Tensor

Position IDs tensor of shape (batch_size, sequence_length). Defaults to None.

 None

Returns:

Type Description

torch.Tensor: Output logits of shape (batch_size, sequence_length, vocab_size).
Source code in src/unitorch/models/peft/modeling_mistral.py 
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
def forward(
    self,
    input_ids: torch.Tensor,
    attention_mask: Optional[torch.Tensor] = None,
    position_ids: Optional[torch.Tensor] = None,
):
    """
    Forward pass of the generation model.

    Args:
        input_ids (torch.Tensor, optional): Input tensor of shape (batch_size, sequence_length). Defaults to None.
        attention_mask (torch.Tensor, optional): Attention mask tensor of shape (batch_size, sequence_length). Defaults to None.
        position_ids (torch.Tensor, optional): Position IDs tensor of shape (batch_size, sequence_length). Defaults to None.

    Returns:
        torch.Tensor: Output logits of shape (batch_size, sequence_length, vocab_size).
    """
    outputs = self.base_model(
        input_ids=input_ids,
        attention_mask=attention_mask,
        position_ids=position_ids,
        return_dict=True,
    )
    logits = outputs.logits
    return logits

generate ยค

generate(
    input_ids: Tensor,
    num_beams: Optional[int] = 5,
    decoder_start_token_id: Optional[int] = 1,
    decoder_end_token_id: Optional[
        Union[int, List[int]]
    ] = 2,
    num_return_sequences: Optional[int] = 1,
    min_gen_seq_length: Optional[int] = 0,
    max_gen_seq_length: Optional[int] = 48,
    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,
)

Generate text using the generation model.

Parameters:

Name Type Description Default
input_ids Tensor

Input tensor of shape (batch_size, sequence_length).

 required
num_beams int

Number of beams for beam search. Defaults to 5.

 5
decoder_start_token_id int

The ID of the decoder start token. Defaults to 2.

 1
decoder_end_token_id int or List[int]

The ID(s) of the decoder end token(s). Defaults to 2.

 2
num_return_sequences int

Number of generated sequences to return. Defaults to 1.

 1
min_gen_seq_length int

Minimum length of generated sequences. Defaults to 0.

 0
max_gen_seq_length int

Maximum length of generated sequences. Defaults to 48.

 48
repetition_penalty float

Penalty for repeated tokens. Defaults to 1.0.

 1.0
no_repeat_ngram_size int

Size of n-grams to avoid repeating. Defaults to 0.

 0
early_stopping bool

Whether to stop generation early. Defaults to True.

 True
length_penalty float

Penalty for longer sequences. Defaults to 1.0.

 1.0
num_beam_groups int

Number of beam groups for diverse beam search. Defaults to 1.

 1
diversity_penalty float

Penalty for diverse sequences in diverse beam search. Defaults to 0.0.

 0.0
do_sample bool

Whether to use sampling for generation. Defaults to False.

 False
temperature float

Sampling temperature. Defaults to 1.0.

 1.0
top_k int

Top-k value for sampling. Defaults to 50.

 50
top_p float

Top-p value for sampling. Defaults to 1.0.

 1.0

Returns:

Name Type Description
GenericOutputs

Generated sequences and their scores.
Source code in src/unitorch/models/peft/modeling_mistral.py 
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
@torch.no_grad()
def generate(
    self,
    input_ids: torch.Tensor,
    num_beams: Optional[int] = 5,
    decoder_start_token_id: Optional[int] = 1,
    decoder_end_token_id: Optional[Union[int, List[int]]] = 2,
    num_return_sequences: Optional[int] = 1,
    min_gen_seq_length: Optional[int] = 0,
    max_gen_seq_length: Optional[int] = 48,
    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,
):
    """
    Generate text using the generation model.

    Args:
        input_ids: Input tensor of shape (batch_size, sequence_length).
        num_beams (int, optional): Number of beams for beam search. Defaults to 5.
        decoder_start_token_id (int, optional): The ID of the decoder start token. Defaults to 2.
        decoder_end_token_id (int or List[int], optional): The ID(s) of the decoder end token(s). Defaults to 2.
        num_return_sequences (int, optional): Number of generated sequences to return. Defaults to 1.
        min_gen_seq_length (int, optional): Minimum length of generated sequences. Defaults to 0.
        max_gen_seq_length (int, optional): Maximum length of generated sequences. Defaults to 48.
        repetition_penalty (float, optional): Penalty for repeated tokens. Defaults to 1.0.
        no_repeat_ngram_size (int, optional): Size of n-grams to avoid repeating. Defaults to 0.
        early_stopping (bool, optional): Whether to stop generation early. Defaults to True.
        length_penalty (float, optional): Penalty for longer sequences. Defaults to 1.0.
        num_beam_groups (int, optional): Number of beam groups for diverse beam search. Defaults to 1.
        diversity_penalty (float, optional): Penalty for diverse sequences in diverse beam search. Defaults to 0.0.
        do_sample (bool, optional): Whether to use sampling for generation. Defaults to False.
        temperature (float, optional): Sampling temperature. Defaults to 1.0.
        top_k (int, optional): Top-k value for sampling. Defaults to 50.
        top_p (float, optional): Top-p value for sampling. Defaults to 1.0.

    Returns:
        GenericOutputs: Generated sequences and their scores.
    """
    input_seq_length = input_ids.size(1)
    outputs = self.base_model.generate(
        input_ids=input_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,
        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)
    )
    outputs.sequences = (
        torch.zeros(sequences.size(0), num_return_sequences, max_gen_seq_length).to(
            device=sequences.device
        )
        + decoder_start_token_id
    )
    outputs.sequences[:, :, : sequences.size(-1) - input_seq_length].copy_(
        sequences[:, :, input_seq_length : sequences.size(-1)]
    )

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

    return GenericOutputs(
        sequences=outputs.sequences.long(),
        sequences_scores=outputs.sequences_scores,
    )

QWen3LoraForGenerationยค

Bases: GenericPeftModel

QWen3 LoRA model for text generation.

Parameters:

Name Type Description Default
config_path str

Path to the model configuration file.

 required
gradient_checkpointing bool

Whether to use gradient checkpointing. Defaults to False.

 False
Source code in src/unitorch/models/peft/modeling_qwen.py 
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
def __init__(
    self,
    config_path: str,
    lora_r: Optional[int] = 16,
    lora_alpha: Optional[int] = 32,
    lora_dropout: Optional[float] = 0.05,
    fan_in_fan_out: Optional[bool] = True,
    target_modules: Optional[Union[List[str], str]] = ["q_proj", "v_proj"],
    gradient_checkpointing: Optional[bool] = False,
):
    """
    QWen3 LoRA model for text generation.

    Args:
        config_path (str): Path to the model configuration file.
        gradient_checkpointing (bool, optional): Whether to use gradient checkpointing. Defaults to False.
    """
    super().__init__()
    self.config = Qwen3Config.from_json_file(config_path)
    self.config.gradient_checkpointing = gradient_checkpointing
    self.peft_config = LoraConfig(
        r=lora_r,
        lora_alpha=lora_alpha,
        lora_dropout=lora_dropout,
        fan_in_fan_out=fan_in_fan_out,
        target_modules=target_modules,
    )
    self.model = Qwen3ForCausalLM(self.config)
    self.model.add_adapter(self.peft_config)
    self.init_weights()

prefix_keys_in_state_dict class-attribute instance-attribute ยค

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

replace_keys_in_state_dict class-attribute instance-attribute ยค

replace_keys_in_state_dict = {
    "q_proj.weight": "q_proj.base_layer.weight",
    "q_proj.bias": "q_proj.base_layer.bias",
    "v_proj.weight": "v_proj.base_layer.weight",
    "v_proj.bias": "v_proj.base_layer.bias",
}

config instance-attribute ยค

config = from_json_file(config_path)

peft_config instance-attribute ยค

peft_config = LoraConfig(
    r=lora_r,
    lora_alpha=lora_alpha,
    lora_dropout=lora_dropout,
    fan_in_fan_out=fan_in_fan_out,
    target_modules=target_modules,
)

model instance-attribute ยค

model = Qwen3ForCausalLM(config)

forward ยค

forward(
    input_ids: Tensor,
    attention_mask: Optional[Tensor] = None,
)

Forward pass of the generation model.

Parameters:

Name Type Description Default
input_ids Tensor

Input tensor of shape (batch_size, sequence_length).

 required
attention_mask Tensor

Attention mask tensor. Defaults to None.

 None

Returns:

Type Description

torch.Tensor: Output logits of shape (batch_size, sequence_length, vocab_size).
Source code in src/unitorch/models/peft/modeling_qwen.py 
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
def forward(
    self,
    input_ids: torch.Tensor,
    attention_mask: Optional[torch.Tensor] = None,
):
    """
    Forward pass of the generation model.

    Args:
        input_ids (torch.Tensor): Input tensor of shape (batch_size, sequence_length).
        attention_mask (torch.Tensor, optional): Attention mask tensor. Defaults to None.

    Returns:
        torch.Tensor: Output logits of shape (batch_size, sequence_length, vocab_size).
    """
    outputs = self.model(
        input_ids=input_ids,
        attention_mask=attention_mask,
        return_dict=True,
    )
    logits = outputs.logits
    return logits

generate ยค

generate(
    input_ids: Tensor,
    num_beams: Optional[int] = 5,
    decoder_start_token_id: Optional[int] = 151643,
    decoder_end_token_id: Optional[
        Union[int, List[int]]
    ] = 151645,
    decoder_pad_token_id: Optional[int] = 151643,
    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,
)

Generate text using the generation model.

Parameters:

Name Type Description Default
input_ids Tensor

Input tensor of shape (batch_size, sequence_length).

 required
num_beams int

Number of beams for beam search. Defaults to 5.

 5
decoder_start_token_id int

The ID of the decoder start token. Defaults to 2.

 151643
decoder_end_token_id int or List[int]

The ID(s) of the decoder end token(s). Defaults to 2.

 151645
num_return_sequences int

Number of generated sequences to return. Defaults to 1.

 1
min_gen_seq_length int

Minimum length of generated sequences. Defaults to 0.

 0
max_gen_seq_length int

Maximum length of generated sequences. Defaults to 48.

 512
repetition_penalty float

Penalty for repeated tokens. Defaults to 1.0.

 1.0
no_repeat_ngram_size int

Size of n-grams to avoid repeating. Defaults to 0.

 0
early_stopping bool

Whether to stop generation early. Defaults to True.

 True
length_penalty float

Penalty for longer sequences. Defaults to 1.0.

 1.0
num_beam_groups int

Number of beam groups for diverse beam search. Defaults to 1.

 1
diversity_penalty float

Penalty for diverse sequences in diverse beam search. Defaults to 0.0.

 0.0
do_sample bool

Whether to use sampling for generation. Defaults to False.

 False
temperature float

Sampling temperature. Defaults to 1.0.

 1.0
top_k int

Top-k value for sampling. Defaults to 50.

 50
top_p float

Top-p value for sampling. Defaults to 1.0.

 1.0

Returns:

Name Type Description
GenericOutputs

Generated sequences and their scores.
Source code in src/unitorch/models/peft/modeling_qwen.py 
 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
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
@torch.no_grad()
def generate(
    self,
    input_ids: torch.Tensor,
    num_beams: Optional[int] = 5,
    decoder_start_token_id: Optional[int] = 151643,
    decoder_end_token_id: Optional[Union[int, List[int]]] = 151645,
    decoder_pad_token_id: Optional[int] = 151643,
    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,
):
    """
    Generate text using the generation model.

    Args:
        input_ids: Input tensor of shape (batch_size, sequence_length).
        num_beams (int, optional): Number of beams for beam search. Defaults to 5.
        decoder_start_token_id (int, optional): The ID of the decoder start token. Defaults to 2.
        decoder_end_token_id (int or List[int], optional): The ID(s) of the decoder end token(s). Defaults to 2.
        num_return_sequences (int, optional): Number of generated sequences to return. Defaults to 1.
        min_gen_seq_length (int, optional): Minimum length of generated sequences. Defaults to 0.
        max_gen_seq_length (int, optional): Maximum length of generated sequences. Defaults to 48.
        repetition_penalty (float, optional): Penalty for repeated tokens. Defaults to 1.0.
        no_repeat_ngram_size (int, optional): Size of n-grams to avoid repeating. Defaults to 0.
        early_stopping (bool, optional): Whether to stop generation early. Defaults to True.
        length_penalty (float, optional): Penalty for longer sequences. Defaults to 1.0.
        num_beam_groups (int, optional): Number of beam groups for diverse beam search. Defaults to 1.
        diversity_penalty (float, optional): Penalty for diverse sequences in diverse beam search. Defaults to 0.0.
        do_sample (bool, optional): Whether to use sampling for generation. Defaults to False.
        temperature (float, optional): Sampling temperature. Defaults to 1.0.
        top_k (int, optional): Top-k value for sampling. Defaults to 50.
        top_p (float, optional): Top-p value for sampling. Defaults to 1.0.

    Returns:
        GenericOutputs: Generated sequences and their scores.
    """
    input_seq_length = input_ids.size(1)
    outputs = self.model.generate(
        input_ids=input_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)
    )
    outputs.sequences = (
        torch.zeros(sequences.size(0), num_return_sequences, max_gen_seq_length).to(
            device=sequences.device
        )
        + decoder_start_token_id
    )
    outputs.sequences[:, :, : sequences.size(-1) - input_seq_length].copy_(
        sequences[:, :, input_seq_length : sequences.size(-1)]
    )

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

    return GenericOutputs(
        sequences=outputs.sequences.long(),
        sequences_scores=outputs.sequences_scores,
    )

QWen3DPOLoraForGenerationยค

Bases: GenericPeftModel

QWen3 DPO LoRA model for preference optimization.

Parameters:

Name Type Description Default
config_path str

Path to the model configuration file.

 required
gradient_checkpointing bool

Whether to use gradient checkpointing. Defaults to False.

 False
dpo_beta float

DPO beta coefficient. Defaults to 0.1.

 0.1
Source code in src/unitorch/models/peft/modeling_qwen.py 
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
def __init__(
    self,
    config_path: str,
    lora_r: Optional[int] = 16,
    lora_alpha: Optional[int] = 32,
    lora_dropout: Optional[float] = 0.05,
    fan_in_fan_out: Optional[bool] = True,
    target_modules: Optional[Union[List[str], str]] = ["q_proj", "v_proj"],
    gradient_checkpointing: Optional[bool] = False,
    dpo_beta: Optional[float] = 0.1,
):
    """
    QWen3 DPO LoRA model for preference optimization.

    Args:
        config_path (str): Path to the model configuration file.
        gradient_checkpointing (bool, optional): Whether to use gradient checkpointing. Defaults to False.
        dpo_beta (float, optional): DPO beta coefficient. Defaults to 0.1.
    """
    super().__init__()
    self.config = Qwen3Config.from_json_file(config_path)
    self.config.gradient_checkpointing = gradient_checkpointing
    self.peft_config = LoraConfig(
        r=lora_r,
        lora_alpha=lora_alpha,
        lora_dropout=lora_dropout,
        fan_in_fan_out=fan_in_fan_out,
        target_modules=target_modules,
    )
    self.model = Qwen3ForCausalLM(self.config)
    self.model.add_adapter(self.peft_config)
    self.init_weights()
    self.dpo_beta = dpo_beta

prefix_keys_in_state_dict class-attribute instance-attribute ยค

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

replace_keys_in_state_dict class-attribute instance-attribute ยค

replace_keys_in_state_dict = {
    "q_proj.weight": "q_proj.base_layer.weight",
    "q_proj.bias": "q_proj.base_layer.bias",
    "v_proj.weight": "v_proj.base_layer.weight",
    "v_proj.bias": "v_proj.base_layer.bias",
}

config instance-attribute ยค

config = from_json_file(config_path)

peft_config instance-attribute ยค

peft_config = LoraConfig(
    r=lora_r,
    lora_alpha=lora_alpha,
    lora_dropout=lora_dropout,
    fan_in_fan_out=fan_in_fan_out,
    target_modules=target_modules,
)

model instance-attribute ยค

model = Qwen3ForCausalLM(config)

dpo_beta instance-attribute ยค

dpo_beta = dpo_beta

forward ยค

forward(
    input_ids: Tensor,
    win_input_ids: Tensor,
    lose_input_ids: Tensor,
    attention_mask: Optional[Tensor] = None,
    win_attention_mask: Optional[Tensor] = None,
    lose_attention_mask: Optional[Tensor] = None,
)

Forward pass computing DPO loss.

Parameters:

Name Type Description Default
input_ids Tensor

Prompt input IDs.

 required
win_input_ids Tensor

Winning response IDs.

 required
lose_input_ids Tensor

Losing response IDs.

 required
attention_mask Tensor

Prompt attention mask. Defaults to None.

 None
win_attention_mask Tensor

Win response attention mask. Defaults to None.

 None
lose_attention_mask Tensor

Lose response attention mask. Defaults to None.

 None

Returns:

Type Description

torch.Tensor: DPO loss scalar.
Source code in src/unitorch/models/peft/modeling_qwen.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
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
def forward(
    self,
    input_ids: torch.Tensor,
    win_input_ids: torch.Tensor,
    lose_input_ids: torch.Tensor,
    attention_mask: Optional[torch.Tensor] = None,
    win_attention_mask: Optional[torch.Tensor] = None,
    lose_attention_mask: Optional[torch.Tensor] = None,
):
    """
    Forward pass computing DPO loss.

    Args:
        input_ids (torch.Tensor): Prompt input IDs.
        win_input_ids (torch.Tensor): Winning response IDs.
        lose_input_ids (torch.Tensor): Losing response IDs.
        attention_mask (torch.Tensor, optional): Prompt attention mask. Defaults to None.
        win_attention_mask (torch.Tensor, optional): Win response attention mask. Defaults to None.
        lose_attention_mask (torch.Tensor, optional): Lose response attention mask. Defaults to None.

    Returns:
        torch.Tensor: DPO loss scalar.
    """
    win_input_ids = torch.cat([input_ids, win_input_ids], dim=1)
    lose_input_ids = torch.cat([input_ids, lose_input_ids], dim=1)
    if attention_mask is not None and win_attention_mask is not None:
        win_attention_mask = torch.cat([attention_mask, win_attention_mask], dim=1)
    if attention_mask is not None and lose_attention_mask is not None:
        lose_attention_mask = torch.cat(
            [attention_mask, lose_attention_mask], dim=1
        )
    win_outputs = self.model(
        input_ids=win_input_ids,
        attention_mask=win_attention_mask,
        return_dict=True,
    )
    lose_outputs = self.model(
        input_ids=lose_input_ids,
        attention_mask=lose_attention_mask,
        return_dict=True,
    )
    input_seq_length = input_ids.size(1)
    win_logits = win_outputs.logits[:, input_seq_length - 1 : -1, :]
    lose_logits = lose_outputs.logits[:, input_seq_length - 1 : -1, :]
    win_labels = win_input_ids[:, input_seq_length:]
    lose_labels = lose_input_ids[:, input_seq_length:]
    win_labels_mask = win_attention_mask[:, input_seq_length:]
    lose_labels_mask = lose_attention_mask[:, input_seq_length:]
    win_nll_loss = F.cross_entropy(
        win_logits.reshape(-1, win_logits.size(-1)),
        win_labels.reshape(-1),
        reduction="none",
    ).reshape(win_labels.size(0), -1)
    win_logprobs = -win_nll_loss * win_labels_mask
    lose_nll_loss = F.cross_entropy(
        lose_logits.reshape(-1, lose_logits.size(-1)),
        lose_labels.reshape(-1),
        reduction="none",
    ).reshape(lose_labels.size(0), -1)
    lose_logprobs = -lose_nll_loss * lose_labels_mask

    with torch.no_grad():
        self.model.disable_adapters()
        ref_win_outputs = self.model(
            input_ids=win_input_ids,
            attention_mask=win_attention_mask,
            return_dict=True,
        )
        ref_lose_outputs = self.model(
            input_ids=lose_input_ids,
            attention_mask=lose_attention_mask,
            return_dict=True,
        )
        ref_win_logits = ref_win_outputs.logits[:, input_seq_length - 1 : -1, :]
        ref_lose_logits = ref_lose_outputs.logits[:, input_seq_length - 1 : -1, :]
        ref_win_nll_loss = F.cross_entropy(
            ref_win_logits.reshape(-1, ref_win_logits.size(-1)),
            win_labels.reshape(-1),
            reduction="none",
        ).reshape(win_labels.size(0), -1)
        ref_win_logprobs = -ref_win_nll_loss * win_labels_mask
        ref_lose_nll_loss = F.cross_entropy(
            ref_lose_logits.reshape(-1, ref_lose_logits.size(-1)),
            lose_labels.reshape(-1),
            reduction="none",
        ).reshape(lose_labels.size(0), -1)
        ref_lose_logprobs = -ref_lose_nll_loss * lose_labels_mask
        self.model.enable_adapters()

    logratios = win_logprobs - lose_logprobs
    ref_logratios = ref_win_logprobs - ref_lose_logprobs
    logits = logratios - ref_logratios
    loss = -F.logsigmoid(self.dpo_beta * logits).mean()
    return loss

generate ยค

generate(
    input_ids: Tensor,
    num_beams: Optional[int] = 5,
    decoder_start_token_id: Optional[int] = 151643,
    decoder_end_token_id: Optional[
        Union[int, List[int]]
    ] = 151645,
    decoder_pad_token_id: Optional[int] = 151643,
    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,
)

Generate text using the generation model.

Parameters:

Name Type Description Default
input_ids Tensor

Input tensor of shape (batch_size, sequence_length).

 required
num_beams int

Number of beams for beam search. Defaults to 5.

 5
decoder_start_token_id int

The ID of the decoder start token. Defaults to 2.

 151643
decoder_end_token_id int or List[int]

The ID(s) of the decoder end token(s). Defaults to 2.

 151645
num_return_sequences int

Number of generated sequences to return. Defaults to 1.

 1
min_gen_seq_length int

Minimum length of generated sequences. Defaults to 0.

 0
max_gen_seq_length int

Maximum length of generated sequences. Defaults to 48.

 512
repetition_penalty float

Penalty for repeated tokens. Defaults to 1.0.

 1.0
no_repeat_ngram_size int

Size of n-grams to avoid repeating. Defaults to 0.

 0
early_stopping bool

Whether to stop generation early. Defaults to True.

 True
length_penalty float

Penalty for longer sequences. Defaults to 1.0.

 1.0
num_beam_groups int

Number of beam groups for diverse beam search. Defaults to 1.

 1
diversity_penalty float

Penalty for diverse sequences in diverse beam search. Defaults to 0.0.

 0.0
do_sample bool

Whether to use sampling for generation. Defaults to False.

 False
temperature float

Sampling temperature. Defaults to 1.0.

 1.0
top_k int

Top-k value for sampling. Defaults to 50.

 50
top_p float

Top-p value for sampling. Defaults to 1.0.

 1.0

Returns:

Name Type Description
GenericOutputs

Generated sequences and their scores.
Source code in src/unitorch/models/peft/modeling_qwen.py 
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
390
391
392
393
394
395
396
397
398
@torch.no_grad()
def generate(
    self,
    input_ids: torch.Tensor,
    num_beams: Optional[int] = 5,
    decoder_start_token_id: Optional[int] = 151643,
    decoder_end_token_id: Optional[Union[int, List[int]]] = 151645,
    decoder_pad_token_id: Optional[int] = 151643,
    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,
):
    """
    Generate text using the generation model.

    Args:
        input_ids: Input tensor of shape (batch_size, sequence_length).
        num_beams (int, optional): Number of beams for beam search. Defaults to 5.
        decoder_start_token_id (int, optional): The ID of the decoder start token. Defaults to 2.
        decoder_end_token_id (int or List[int], optional): The ID(s) of the decoder end token(s). Defaults to 2.
        num_return_sequences (int, optional): Number of generated sequences to return. Defaults to 1.
        min_gen_seq_length (int, optional): Minimum length of generated sequences. Defaults to 0.
        max_gen_seq_length (int, optional): Maximum length of generated sequences. Defaults to 48.
        repetition_penalty (float, optional): Penalty for repeated tokens. Defaults to 1.0.
        no_repeat_ngram_size (int, optional): Size of n-grams to avoid repeating. Defaults to 0.
        early_stopping (bool, optional): Whether to stop generation early. Defaults to True.
        length_penalty (float, optional): Penalty for longer sequences. Defaults to 1.0.
        num_beam_groups (int, optional): Number of beam groups for diverse beam search. Defaults to 1.
        diversity_penalty (float, optional): Penalty for diverse sequences in diverse beam search. Defaults to 0.0.
        do_sample (bool, optional): Whether to use sampling for generation. Defaults to False.
        temperature (float, optional): Sampling temperature. Defaults to 1.0.
        top_k (int, optional): Top-k value for sampling. Defaults to 50.
        top_p (float, optional): Top-p value for sampling. Defaults to 1.0.

    Returns:
        GenericOutputs: Generated sequences and their scores.
    """
    input_seq_length = input_ids.size(1)
    outputs = self.model.generate(
        input_ids=input_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)
    )
    outputs.sequences = (
        torch.zeros(sequences.size(0), num_return_sequences, max_gen_seq_length).to(
            device=sequences.device
        )
        + decoder_start_token_id
    )
    outputs.sequences[:, :, : sequences.size(-1) - input_seq_length].copy_(
        sequences[:, :, input_seq_length : sequences.size(-1)]
    )

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

    return GenericOutputs(
        sequences=outputs.sequences.long(),
        sequences_scores=outputs.sequences_scores,
    )

QWen3GRPOLoraForGenerationยค

Bases: GenericPeftModel

QWen3 GRPO LoRA model for group relative policy optimization.

Parameters:

Name Type Description Default
config_path str

Path to the model configuration file.

 required
gradient_checkpointing bool

Whether to use gradient checkpointing. Defaults to False.

 False
Source code in src/unitorch/models/peft/modeling_qwen.py 
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
def __init__(
    self,
    config_path: str,
    lora_r: Optional[int] = 16,
    lora_alpha: Optional[int] = 32,
    lora_dropout: Optional[float] = 0.05,
    fan_in_fan_out: Optional[bool] = True,
    target_modules: Optional[Union[List[str], str]] = ["q_proj", "v_proj"],
    gradient_checkpointing: Optional[bool] = False,
):
    """
    QWen3 GRPO LoRA model for group relative policy optimization.

    Args:
        config_path (str): Path to the model configuration file.
        gradient_checkpointing (bool, optional): Whether to use gradient checkpointing. Defaults to False.
    """
    super().__init__()
    self.config = Qwen3Config.from_json_file(config_path)
    self.config.gradient_checkpointing = gradient_checkpointing
    self.peft_config = LoraConfig(
        r=lora_r,
        lora_alpha=lora_alpha,
        lora_dropout=lora_dropout,
        fan_in_fan_out=fan_in_fan_out,
        target_modules=target_modules,
    )
    self.model = Qwen3ForCausalLM(self.config)
    self.model.add_adapter(self.peft_config)
    self.init_weights()

prefix_keys_in_state_dict class-attribute instance-attribute ยค

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

replace_keys_in_state_dict class-attribute instance-attribute ยค

replace_keys_in_state_dict = {
    "q_proj.weight": "q_proj.base_layer.weight",
    "q_proj.bias": "q_proj.base_layer.bias",
    "v_proj.weight": "v_proj.base_layer.weight",
    "v_proj.bias": "v_proj.base_layer.bias",
}

config instance-attribute ยค

config = from_json_file(config_path)

peft_config instance-attribute ยค

peft_config = LoraConfig(
    r=lora_r,
    lora_alpha=lora_alpha,
    lora_dropout=lora_dropout,
    fan_in_fan_out=fan_in_fan_out,
    target_modules=target_modules,
)

model instance-attribute ยค

model = Qwen3ForCausalLM(config)

forward ยค

forward(
    input_ids: Tensor,
    sampled_ids: Tensor,
    sampled_rewards: Tensor,
    attention_mask: Optional[Tensor] = None,
    sampled_attention_mask: Optional[Tensor] = None,
)

Forward pass of the generation model.

Parameters:

Name Type Description Default
input_ids Tensor

Input tensor of shape (batch_size, sequence_length). Defaults to None.

 required
attention_mask Tensor

Attention mask tensor of shape (batch_size, sequence_length). Defaults to None.

 None
position_ids Tensor

Position IDs tensor of shape (batch_size, sequence_length). Defaults to None.

 required

Returns:

Type Description

torch.Tensor: Output logits of shape (batch_size, sequence_length, vocab_size).
Source code in src/unitorch/models/peft/modeling_qwen.py 
441
442
443
444
445
446
447
448
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
475
476
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
520
521
522
523
524
525
526
527
528
529
530
531
def forward(
    self,
    input_ids: torch.Tensor,
    sampled_ids: torch.Tensor,
    sampled_rewards: torch.Tensor,
    attention_mask: Optional[torch.Tensor] = None,
    sampled_attention_mask: Optional[torch.Tensor] = None,
):
    """
    Forward pass of the generation model.

    Args:
        input_ids (torch.Tensor, optional): Input tensor of shape (batch_size, sequence_length). Defaults to None.
        attention_mask (torch.Tensor, optional): Attention mask tensor of shape (batch_size, sequence_length). Defaults to None.
        position_ids (torch.Tensor, optional): Position IDs tensor of shape (batch_size, sequence_length). Defaults to None.

    Returns:
        torch.Tensor: Output logits of shape (batch_size, sequence_length, vocab_size).
    """
    mean_rewards = sampled_rewards.mean(dim=-1)
    std_rewards = sampled_rewards.std(dim=-1)
    adv_rewards = sampled_rewards - mean_rewards.unsqueeze(-1)
    adv_rewards = adv_rewards / (std_rewards.unsqueeze(-1) + 1e-5)
    adv_rewards = adv_rewards.detach().reshape(-1)

    # ้‡ๅค input_ids ๅ’Œ attention_mask
    G = sampled_ids.size(1)
    repeated_input_ids = input_ids.repeat_interleave(G, dim=0)  # [B*G, L_in]
    repeated_attention = (
        attention_mask.repeat_interleave(G, dim=0)
        if attention_mask is not None
        else None
    )
    sampled_seq_length = sampled_ids.size(-1)
    all_input_ids = torch.cat(
        [repeated_input_ids, sampled_ids.view(-1, sampled_seq_length)], dim=1
    )  # [B*G, L_in + L_out]

    if repeated_attention is not None:
        all_attention_mask = torch.cat(
            [
                repeated_attention,
                sampled_attention_mask.view(-1, sampled_seq_length),
            ],
            dim=1,
        )
    else:
        all_attention_mask = None

    all_outputs = self.model(
        input_ids=all_input_ids,
        attention_mask=all_attention_mask,
        return_dict=True,
    )

    input_seq_length = input_ids.size(1)
    logits = all_outputs.logits[:, input_seq_length - 1 : -1, :]

    labels = all_input_ids[:, input_seq_length:]
    labels_mask = all_attention_mask[:, input_seq_length:]
    nll_loss = F.cross_entropy(
        logits.reshape(-1, logits.size(-1)),
        labels.reshape(-1),
        reduction="none",
    ).reshape(labels.size(0), -1)
    logprobs = -nll_loss * labels_mask

    with torch.no_grad():
        self.model.disable_adapters()
        ref_outputs = self.model(
            input_ids=all_input_ids,
            attention_mask=all_attention_mask,
            return_dict=True,
        )
        ref_logits = ref_outputs.logits[:, input_seq_length - 1 : -1, :]
        ref_nll_loss = F.cross_entropy(
            ref_logits.reshape(-1, ref_logits.size(-1)),
            labels.reshape(-1),
            reduction="none",
        ).reshape(labels.size(0), -1)
        ref_logprobs = -ref_nll_loss * labels_mask
        self.model.enable_adapters()

    logprob_mean = logprobs.sum(-1) / (labels_mask.sum(-1) + 1e-5)
    ref_logprob_mean = ref_logprobs.sum(-1) / (labels_mask.sum(-1) + 1e-5)

    log_ratio = logprob_mean - ref_logprob_mean.detach()
    kl_div = 0.02 * (log_ratio**2)
    loss = (-adv_rewards.reshape(-1) * log_ratio + kl_div).mean()

    return loss

generate ยค

generate(
    input_ids: Tensor,
    num_beams: Optional[int] = 5,
    decoder_start_token_id: Optional[int] = 151643,
    decoder_end_token_id: Optional[
        Union[int, List[int]]
    ] = 151645,
    decoder_pad_token_id: Optional[int] = 151643,
    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,
)

Generate text using the generation model.

Parameters:

Name Type Description Default
input_ids Tensor

Input tensor of shape (batch_size, sequence_length).

 required
num_beams int

Number of beams for beam search. Defaults to 5.

 5
decoder_start_token_id int

The ID of the decoder start token. Defaults to 2.

 151643
decoder_end_token_id int or List[int]

The ID(s) of the decoder end token(s). Defaults to 2.

 151645
num_return_sequences int

Number of generated sequences to return. Defaults to 1.

 1
min_gen_seq_length int

Minimum length of generated sequences. Defaults to 0.

 0
max_gen_seq_length int

Maximum length of generated sequences. Defaults to 48.

 512
repetition_penalty float

Penalty for repeated tokens. Defaults to 1.0.

 1.0
no_repeat_ngram_size int

Size of n-grams to avoid repeating. Defaults to 0.

 0
early_stopping bool

Whether to stop generation early. Defaults to True.

 True
length_penalty float

Penalty for longer sequences. Defaults to 1.0.

 1.0
num_beam_groups int

Number of beam groups for diverse beam search. Defaults to 1.

 1
diversity_penalty float

Penalty for diverse sequences in diverse beam search. Defaults to 0.0.

 0.0
do_sample bool

Whether to use sampling for generation. Defaults to False.

 False
temperature float

Sampling temperature. Defaults to 1.0.

 1.0
top_k int

Top-k value for sampling. Defaults to 50.

 50
top_p float

Top-p value for sampling. Defaults to 1.0.

 1.0

Returns:

Name Type Description
GenericOutputs

Generated sequences and their scores.
Source code in src/unitorch/models/peft/modeling_qwen.py 
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
@torch.no_grad()
def generate(
    self,
    input_ids: torch.Tensor,
    num_beams: Optional[int] = 5,
    decoder_start_token_id: Optional[int] = 151643,
    decoder_end_token_id: Optional[Union[int, List[int]]] = 151645,
    decoder_pad_token_id: Optional[int] = 151643,
    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,
):
    """
    Generate text using the generation model.

    Args:
        input_ids: Input tensor of shape (batch_size, sequence_length).
        num_beams (int, optional): Number of beams for beam search. Defaults to 5.
        decoder_start_token_id (int, optional): The ID of the decoder start token. Defaults to 2.
        decoder_end_token_id (int or List[int], optional): The ID(s) of the decoder end token(s). Defaults to 2.
        num_return_sequences (int, optional): Number of generated sequences to return. Defaults to 1.
        min_gen_seq_length (int, optional): Minimum length of generated sequences. Defaults to 0.
        max_gen_seq_length (int, optional): Maximum length of generated sequences. Defaults to 48.
        repetition_penalty (float, optional): Penalty for repeated tokens. Defaults to 1.0.
        no_repeat_ngram_size (int, optional): Size of n-grams to avoid repeating. Defaults to 0.
        early_stopping (bool, optional): Whether to stop generation early. Defaults to True.
        length_penalty (float, optional): Penalty for longer sequences. Defaults to 1.0.
        num_beam_groups (int, optional): Number of beam groups for diverse beam search. Defaults to 1.
        diversity_penalty (float, optional): Penalty for diverse sequences in diverse beam search. Defaults to 0.0.
        do_sample (bool, optional): Whether to use sampling for generation. Defaults to False.
        temperature (float, optional): Sampling temperature. Defaults to 1.0.
        top_k (int, optional): Top-k value for sampling. Defaults to 50.
        top_p (float, optional): Top-p value for sampling. Defaults to 1.0.

    Returns:
        GenericOutputs: Generated sequences and their scores.
    """
    input_seq_length = input_ids.size(1)
    outputs = self.model.generate(
        input_ids=input_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)
    )
    outputs.sequences = (
        torch.zeros(sequences.size(0), num_return_sequences, max_gen_seq_length).to(
            device=sequences.device
        )
        + decoder_start_token_id
    )
    outputs.sequences[:, :, : sequences.size(-1) - input_seq_length].copy_(
        sequences[:, :, input_seq_length : sequences.size(-1)]
    )

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

    return GenericOutputs(
        sequences=outputs.sequences.long(),
        sequences_scores=outputs.sequences_scores,
    )

QWen3VLLoraForGenerationยค

Bases: GenericPeftModel

QWen3-VL LoRA model for multimodal generation.

Parameters:

Name Type Description Default
config_path str

Path to the model configuration file.

 required
gradient_checkpointing bool

Whether to use gradient checkpointing. Defaults to False.

 False
Source code in src/unitorch/models/peft/modeling_qwen_vl.py 
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
def __init__(
    self,
    config_path: str,
    lora_r: Optional[int] = 16,
    lora_alpha: Optional[int] = 32,
    lora_dropout: Optional[float] = 0.05,
    fan_in_fan_out: Optional[bool] = True,
    target_modules: Optional[Union[List[str], str]] = ["q_proj", "v_proj"],
    gradient_checkpointing: Optional[bool] = False,
):
    """
    QWen3-VL LoRA model for multimodal generation.

    Args:
        config_path (str): Path to the model configuration file.
        gradient_checkpointing (bool, optional): Whether to use gradient checkpointing. Defaults to False.
    """
    super().__init__()
    self.config = Qwen3VLConfig.from_json_file(config_path)
    self.config.gradient_checkpointing = gradient_checkpointing
    self.peft_config = LoraConfig(
        r=lora_r,
        lora_alpha=lora_alpha,
        lora_dropout=lora_dropout,
        fan_in_fan_out=fan_in_fan_out,
        target_modules=target_modules,
    )
    self.model = Qwen3VLForConditionalGeneration(self.config)
    self.model.add_adapter(self.peft_config)
    self.init_weights()

prefix_keys_in_state_dict class-attribute instance-attribute ยค

prefix_keys_in_state_dict = {
    "^visual.": "model.model.",
    "^model(?!\\.model).": "model.model.language_",
}

replace_keys_in_state_dict class-attribute instance-attribute ยค

replace_keys_in_state_dict = {
    "q_proj.weight": "q_proj.base_layer.weight",
    "q_proj.bias": "q_proj.base_layer.bias",
    "v_proj.weight": "v_proj.base_layer.weight",
    "v_proj.bias": "v_proj.base_layer.bias",
}

config instance-attribute ยค

config = from_json_file(config_path)

peft_config instance-attribute ยค

peft_config = LoraConfig(
    r=lora_r,
    lora_alpha=lora_alpha,
    lora_dropout=lora_dropout,
    fan_in_fan_out=fan_in_fan_out,
    target_modules=target_modules,
)

model instance-attribute ยค

model = Qwen3VLForConditionalGeneration(config)

forward ยค

forward(
    input_ids: Tensor,
    pixel_values: Tensor,
    image_grid_thw: Tensor,
    attention_mask: Optional[Tensor] = None,
)

Forward pass of the generation model.

Parameters:

Name Type Description Default
input_ids Tensor

Input tensor of shape (batch_size, sequence_length). Defaults to None.

 required
attention_mask Tensor

Attention mask tensor of shape (batch_size, sequence_length). Defaults to None.

 None
position_ids Tensor

Position IDs tensor of shape (batch_size, sequence_length). Defaults to None.

 required

Returns:

Type Description

torch.Tensor: Output logits of shape (batch_size, sequence_length, vocab_size).
Source code in src/unitorch/models/peft/modeling_qwen_vl.py 
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
def forward(
    self,
    input_ids: torch.Tensor,
    pixel_values: torch.Tensor,
    image_grid_thw: torch.Tensor,
    attention_mask: Optional[torch.Tensor] = None,
):
    """
    Forward pass of the generation model.

    Args:
        input_ids (torch.Tensor, optional): Input tensor of shape (batch_size, sequence_length). Defaults to None.
        attention_mask (torch.Tensor, optional): Attention mask tensor of shape (batch_size, sequence_length). Defaults to None.
        position_ids (torch.Tensor, optional): Position IDs tensor of shape (batch_size, sequence_length). Defaults to None.

    Returns:
        torch.Tensor: Output logits of shape (batch_size, sequence_length, vocab_size).
    """
    image_grid_thw = image_grid_thw.view(-1, image_grid_thw.size(-1))
    pixel_values = pixel_values.view(-1, pixel_values.size(-1))
    outputs = self.model(
        input_ids=input_ids,
        pixel_values=pixel_values,
        image_grid_thw=image_grid_thw,
        attention_mask=attention_mask,
        return_dict=True,
    )
    logits = outputs.logits
    return logits

generate ยค

generate(
    input_ids: Tensor,
    pixel_values: Tensor,
    image_grid_thw: Tensor,
    num_beams: Optional[int] = 5,
    decoder_start_token_id: Optional[int] = 151643,
    decoder_end_token_id: Optional[
        Union[int, List[int]]
    ] = 151645,
    decoder_pad_token_id: Optional[int] = 151643,
    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,
)

Generate text using the generation model.

Parameters:

Name Type Description Default
input_ids Tensor

Input tensor of shape (batch_size, sequence_length).

 required
num_beams int

Number of beams for beam search. Defaults to 5.

 5
decoder_start_token_id int

The ID of the decoder start token. Defaults to 2.

 151643
decoder_end_token_id int or List[int]

The ID(s) of the decoder end token(s). Defaults to 2.

 151645
num_return_sequences int

Number of generated sequences to return. Defaults to 1.

 1
min_gen_seq_length int

Minimum length of generated sequences. Defaults to 0.

 0
max_gen_seq_length int

Maximum length of generated sequences. Defaults to 48.

 512
repetition_penalty float

Penalty for repeated tokens. Defaults to 1.0.

 1.0
no_repeat_ngram_size int

Size of n-grams to avoid repeating. Defaults to 0.

 0
early_stopping bool

Whether to stop generation early. Defaults to True.

 True
length_penalty float

Penalty for longer sequences. Defaults to 1.0.

 1.0
num_beam_groups int

Number of beam groups for diverse beam search. Defaults to 1.

 1
diversity_penalty float

Penalty for diverse sequences in diverse beam search. Defaults to 0.0.

 0.0
do_sample bool

Whether to use sampling for generation. Defaults to False.

 False
temperature float

Sampling temperature. Defaults to 1.0.

 1.0
top_k int

Top-k value for sampling. Defaults to 50.

 50
top_p float

Top-p value for sampling. Defaults to 1.0.

 1.0

Returns:

Name Type Description
GenericOutputs

Generated sequences and their scores.
Source code in src/unitorch/models/peft/modeling_qwen_vl.py 
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
@torch.no_grad()
def generate(
    self,
    input_ids: torch.Tensor,
    pixel_values: torch.Tensor,
    image_grid_thw: torch.Tensor,
    num_beams: Optional[int] = 5,
    decoder_start_token_id: Optional[int] = 151643,
    decoder_end_token_id: Optional[Union[int, List[int]]] = 151645,
    decoder_pad_token_id: Optional[int] = 151643,
    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,
):
    """
    Generate text using the generation model.

    Args:
        input_ids: Input tensor of shape (batch_size, sequence_length).
        num_beams (int, optional): Number of beams for beam search. Defaults to 5.
        decoder_start_token_id (int, optional): The ID of the decoder start token. Defaults to 2.
        decoder_end_token_id (int or List[int], optional): The ID(s) of the decoder end token(s). Defaults to 2.
        num_return_sequences (int, optional): Number of generated sequences to return. Defaults to 1.
        min_gen_seq_length (int, optional): Minimum length of generated sequences. Defaults to 0.
        max_gen_seq_length (int, optional): Maximum length of generated sequences. Defaults to 48.
        repetition_penalty (float, optional): Penalty for repeated tokens. Defaults to 1.0.
        no_repeat_ngram_size (int, optional): Size of n-grams to avoid repeating. Defaults to 0.
        early_stopping (bool, optional): Whether to stop generation early. Defaults to True.
        length_penalty (float, optional): Penalty for longer sequences. Defaults to 1.0.
        num_beam_groups (int, optional): Number of beam groups for diverse beam search. Defaults to 1.
        diversity_penalty (float, optional): Penalty for diverse sequences in diverse beam search. Defaults to 0.0.
        do_sample (bool, optional): Whether to use sampling for generation. Defaults to False.
        temperature (float, optional): Sampling temperature. Defaults to 1.0.
        top_k (int, optional): Top-k value for sampling. Defaults to 50.
        top_p (float, optional): Top-p value for sampling. Defaults to 1.0.

    Returns:
        GenericOutputs: Generated sequences and their scores.
    """
    input_seq_length = input_ids.size(1)
    image_grid_thw = image_grid_thw.view(-1, image_grid_thw.size(-1))
    pixel_values = pixel_values.view(-1, pixel_values.size(-1))
    outputs = self.model.generate(
        input_ids=input_ids,
        pixel_values=pixel_values,
        image_grid_thw=image_grid_thw,
        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)
    )
    outputs.sequences = (
        torch.zeros(sequences.size(0), num_return_sequences, max_gen_seq_length).to(
            device=sequences.device
        )
        + decoder_start_token_id
    )
    outputs.sequences[:, :, : sequences.size(-1) - input_seq_length].copy_(
        sequences[:, :, input_seq_length : sequences.size(-1)]
    )

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

    return GenericOutputs(
        sequences=outputs.sequences.long(),
        sequences_scores=outputs.sequences_scores,
    )

QWen3VLDPOLoraForGenerationยค

Bases: GenericPeftModel

QWen3-VL LoRA model for multimodal generation.

Parameters:

Name Type Description Default
config_path str

Path to the model configuration file.

 required
gradient_checkpointing bool

Whether to use gradient checkpointing. Defaults to False.

 False
Source code in src/unitorch/models/peft/modeling_qwen_vl.py 
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
def __init__(
    self,
    config_path: str,
    lora_r: Optional[int] = 16,
    lora_alpha: Optional[int] = 32,
    lora_dropout: Optional[float] = 0.05,
    fan_in_fan_out: Optional[bool] = True,
    target_modules: Optional[Union[List[str], str]] = ["q_proj", "v_proj"],
    gradient_checkpointing: Optional[bool] = False,
    dpo_beta: Optional[float] = 0.1,
):
    """
    QWen3-VL LoRA model for multimodal generation.

    Args:
        config_path (str): Path to the model configuration file.
        gradient_checkpointing (bool, optional): Whether to use gradient checkpointing. Defaults to False.
    """
    super().__init__()
    self.config = Qwen3VLConfig.from_json_file(config_path)
    self.config.gradient_checkpointing = gradient_checkpointing
    self.peft_config = LoraConfig(
        r=lora_r,
        lora_alpha=lora_alpha,
        lora_dropout=lora_dropout,
        fan_in_fan_out=fan_in_fan_out,
        target_modules=target_modules,
    )
    self.model = Qwen3VLForConditionalGeneration(self.config)
    self.model.add_adapter(self.peft_config)
    self.init_weights()
    self.dpo_beta = dpo_beta

prefix_keys_in_state_dict class-attribute instance-attribute ยค

prefix_keys_in_state_dict = {
    "^visual.": "model.model.",
    "^model(?!\\.model).": "model.model.language_",
}

replace_keys_in_state_dict class-attribute instance-attribute ยค

replace_keys_in_state_dict = {
    "q_proj.weight": "q_proj.base_layer.weight",
    "q_proj.bias": "q_proj.base_layer.bias",
    "v_proj.weight": "v_proj.base_layer.weight",
    "v_proj.bias": "v_proj.base_layer.bias",
}

config instance-attribute ยค

config = from_json_file(config_path)

peft_config instance-attribute ยค

peft_config = LoraConfig(
    r=lora_r,
    lora_alpha=lora_alpha,
    lora_dropout=lora_dropout,
    fan_in_fan_out=fan_in_fan_out,
    target_modules=target_modules,
)

model instance-attribute ยค

model = Qwen3VLForConditionalGeneration(config)

dpo_beta instance-attribute ยค

dpo_beta = dpo_beta

forward ยค

forward(
    input_ids: Tensor,
    pixel_values: Tensor,
    image_grid_thw: Tensor,
    win_input_ids: Tensor,
    lose_input_ids: Tensor,
    attention_mask: Optional[Tensor] = None,
    win_attention_mask: Optional[Tensor] = None,
    lose_attention_mask: Optional[Tensor] = None,
)

Forward pass of the generation model.

Parameters:

Name Type Description Default
input_ids Tensor

Input tensor of shape (batch_size, sequence_length). Defaults to None.

 required
attention_mask Tensor

Attention mask tensor of shape (batch_size, sequence_length). Defaults to None.

 None
position_ids Tensor

Position IDs tensor of shape (batch_size, sequence_length). Defaults to None.

 required

Returns:

Type Description

torch.Tensor: Output logits of shape (batch_size, sequence_length, vocab_size).
Source code in src/unitorch/models/peft/modeling_qwen_vl.py 
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
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
def forward(
    self,
    input_ids: torch.Tensor,
    pixel_values: torch.Tensor,
    image_grid_thw: torch.Tensor,
    win_input_ids: torch.Tensor,
    lose_input_ids: torch.Tensor,
    attention_mask: Optional[torch.Tensor] = None,
    win_attention_mask: Optional[torch.Tensor] = None,
    lose_attention_mask: Optional[torch.Tensor] = None,
):
    """
    Forward pass of the generation model.

    Args:
        input_ids (torch.Tensor, optional): Input tensor of shape (batch_size, sequence_length). Defaults to None.
        attention_mask (torch.Tensor, optional): Attention mask tensor of shape (batch_size, sequence_length). Defaults to None.
        position_ids (torch.Tensor, optional): Position IDs tensor of shape (batch_size, sequence_length). Defaults to None.

    Returns:
        torch.Tensor: Output logits of shape (batch_size, sequence_length, vocab_size).
    """
    image_grid_thw = image_grid_thw.view(-1, image_grid_thw.size(-1))
    pixel_values = pixel_values.view(-1, pixel_values.size(-1))
    win_input_ids = torch.cat([input_ids, win_input_ids], dim=1)
    lose_input_ids = torch.cat([input_ids, lose_input_ids], dim=1)
    if attention_mask is not None and win_attention_mask is not None:
        win_attention_mask = torch.cat([attention_mask, win_attention_mask], dim=1)
    if attention_mask is not None and lose_attention_mask is not None:
        lose_attention_mask = torch.cat(
            [attention_mask, lose_attention_mask], dim=1
        )
    win_outputs = self.model(
        input_ids=win_input_ids,
        pixel_values=pixel_values,
        image_grid_thw=image_grid_thw,
        attention_mask=win_attention_mask,
        return_dict=True,
    )
    lose_outputs = self.model(
        input_ids=lose_input_ids,
        pixel_values=pixel_values,
        image_grid_thw=image_grid_thw,
        attention_mask=lose_attention_mask,
        return_dict=True,
    )
    input_seq_length = input_ids.size(1)
    win_logits = win_outputs.logits[:, input_seq_length - 1 : -1, :]
    lose_logits = lose_outputs.logits[:, input_seq_length - 1 : -1, :]
    win_labels = win_input_ids[:, input_seq_length:]
    lose_labels = lose_input_ids[:, input_seq_length:]
    win_labels_mask = win_attention_mask[:, input_seq_length:]
    lose_labels_mask = lose_attention_mask[:, input_seq_length:]
    win_nll_loss = F.cross_entropy(
        win_logits.reshape(-1, win_logits.size(-1)),
        win_labels.reshape(-1),
        reduction="none",
    ).reshape(win_labels.size(0), -1)
    win_logprobs = -win_nll_loss * win_labels_mask
    lose_nll_loss = F.cross_entropy(
        lose_logits.reshape(-1, lose_logits.size(-1)),
        lose_labels.reshape(-1),
        reduction="none",
    ).reshape(lose_labels.size(0), -1)
    lose_logprobs = -lose_nll_loss * lose_labels_mask

    with torch.no_grad():
        self.model.disable_adapters()
        ref_win_outputs = self.model(
            input_ids=win_input_ids,
            attention_mask=win_attention_mask,
            return_dict=True,
        )
        ref_lose_outputs = self.model(
            input_ids=lose_input_ids,
            attention_mask=lose_attention_mask,
            return_dict=True,
        )
        ref_win_logits = ref_win_outputs.logits[:, input_seq_length - 1 : -1, :]
        ref_lose_logits = ref_lose_outputs.logits[:, input_seq_length - 1 : -1, :]
        ref_win_nll_loss = F.cross_entropy(
            ref_win_logits.reshape(-1, ref_win_logits.size(-1)),
            win_labels.reshape(-1),
            reduction="none",
        ).reshape(win_labels.size(0), -1)
        ref_win_logprobs = -ref_win_nll_loss * win_labels_mask
        ref_lose_nll_loss = F.cross_entropy(
            ref_lose_logits.reshape(-1, ref_lose_logits.size(-1)),
            lose_labels.reshape(-1),
            reduction="none",
        ).reshape(lose_labels.size(0), -1)
        ref_lose_logprobs = -ref_lose_nll_loss * lose_labels_mask
        self.model.enable_adapters()

    logratios = win_logprobs - lose_logprobs
    ref_logratios = ref_win_logprobs - ref_lose_logprobs
    logits = logratios - ref_logratios
    loss = -F.logsigmoid(self.dpo_beta * logits).mean()
    return loss

generate ยค

generate(
    input_ids: Tensor,
    pixel_values: Tensor,
    image_grid_thw: Tensor,
    num_beams: Optional[int] = 5,
    decoder_start_token_id: Optional[int] = 151643,
    decoder_end_token_id: Optional[
        Union[int, List[int]]
    ] = 151645,
    decoder_pad_token_id: Optional[int] = 151643,
    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,
)

Generate text using the generation model.

Parameters:

Name Type Description Default
input_ids Tensor

Input tensor of shape (batch_size, sequence_length).

 required
num_beams int

Number of beams for beam search. Defaults to 5.

 5
decoder_start_token_id int

The ID of the decoder start token. Defaults to 2.

 151643
decoder_end_token_id int or List[int]

The ID(s) of the decoder end token(s). Defaults to 2.

 151645
num_return_sequences int

Number of generated sequences to return. Defaults to 1.

 1
min_gen_seq_length int

Minimum length of generated sequences. Defaults to 0.

 0
max_gen_seq_length int

Maximum length of generated sequences. Defaults to 48.

 512
repetition_penalty float

Penalty for repeated tokens. Defaults to 1.0.

 1.0
no_repeat_ngram_size int

Size of n-grams to avoid repeating. Defaults to 0.

 0
early_stopping bool

Whether to stop generation early. Defaults to True.

 True
length_penalty float

Penalty for longer sequences. Defaults to 1.0.

 1.0
num_beam_groups int

Number of beam groups for diverse beam search. Defaults to 1.

 1
diversity_penalty float

Penalty for diverse sequences in diverse beam search. Defaults to 0.0.

 0.0
do_sample bool

Whether to use sampling for generation. Defaults to False.

 False
temperature float

Sampling temperature. Defaults to 1.0.

 1.0
top_k int

Top-k value for sampling. Defaults to 50.

 50
top_p float

Top-p value for sampling. Defaults to 1.0.

 1.0

Returns:

Name Type Description
GenericOutputs

Generated sequences and their scores.
Source code in src/unitorch/models/peft/modeling_qwen_vl.py 
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
@torch.no_grad()
def generate(
    self,
    input_ids: torch.Tensor,
    pixel_values: torch.Tensor,
    image_grid_thw: torch.Tensor,
    num_beams: Optional[int] = 5,
    decoder_start_token_id: Optional[int] = 151643,
    decoder_end_token_id: Optional[Union[int, List[int]]] = 151645,
    decoder_pad_token_id: Optional[int] = 151643,
    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,
):
    """
    Generate text using the generation model.

    Args:
        input_ids: Input tensor of shape (batch_size, sequence_length).
        num_beams (int, optional): Number of beams for beam search. Defaults to 5.
        decoder_start_token_id (int, optional): The ID of the decoder start token. Defaults to 2.
        decoder_end_token_id (int or List[int], optional): The ID(s) of the decoder end token(s). Defaults to 2.
        num_return_sequences (int, optional): Number of generated sequences to return. Defaults to 1.
        min_gen_seq_length (int, optional): Minimum length of generated sequences. Defaults to 0.
        max_gen_seq_length (int, optional): Maximum length of generated sequences. Defaults to 48.
        repetition_penalty (float, optional): Penalty for repeated tokens. Defaults to 1.0.
        no_repeat_ngram_size (int, optional): Size of n-grams to avoid repeating. Defaults to 0.
        early_stopping (bool, optional): Whether to stop generation early. Defaults to True.
        length_penalty (float, optional): Penalty for longer sequences. Defaults to 1.0.
        num_beam_groups (int, optional): Number of beam groups for diverse beam search. Defaults to 1.
        diversity_penalty (float, optional): Penalty for diverse sequences in diverse beam search. Defaults to 0.0.
        do_sample (bool, optional): Whether to use sampling for generation. Defaults to False.
        temperature (float, optional): Sampling temperature. Defaults to 1.0.
        top_k (int, optional): Top-k value for sampling. Defaults to 50.
        top_p (float, optional): Top-p value for sampling. Defaults to 1.0.

    Returns:
        GenericOutputs: Generated sequences and their scores.
    """
    input_seq_length = input_ids.size(1)
    image_grid_thw = image_grid_thw.view(-1, image_grid_thw.size(-1))
    pixel_values = pixel_values.view(-1, pixel_values.size(-1))
    outputs = self.model.generate(
        input_ids=input_ids,
        pixel_values=pixel_values,
        image_grid_thw=image_grid_thw,
        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)
    )
    outputs.sequences = (
        torch.zeros(sequences.size(0), num_return_sequences, max_gen_seq_length).to(
            device=sequences.device
        )
        + decoder_start_token_id
    )
    outputs.sequences[:, :, : sequences.size(-1) - input_seq_length].copy_(
        sequences[:, :, input_seq_length : sequences.size(-1)]
    )

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

    return GenericOutputs(
        sequences=outputs.sequences.long(),
        sequences_scores=outputs.sequences_scores,
    )