Skip to content

unitorch.cli.models.qwen¤

QWenProcessor¤

Tip

core/process/qwen is the section for configuration of QWenProcessor.

Bases: QWenProcessor

Processor for Bloom language models.

Source code in src/unitorch/cli/models/qwen/processing.py 
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
def __init__(
    self,
    tokenizer_file: str,
    tokenizer_config: Optional[str] = None,
    special_tokens_map: Optional[str] = None,
    chat_template: Optional[str] = None,
    max_seq_length: Optional[int] = 128,
    max_gen_seq_length: Optional[int] = 128,
):
    super().__init__(
        tokenizer_file=tokenizer_file,
        tokenizer_config=tokenizer_config,
        special_tokens_map=special_tokens_map,
        chat_template=chat_template,
        max_seq_length=max_seq_length,
        max_gen_seq_length=max_gen_seq_length,
    )

from_config classmethod ¤

from_config(config, **kwargs)
Source code in src/unitorch/cli/models/qwen/processing.py 
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
@classmethod
@config_defaults_init("core/process/qwen")
def from_config(cls, config, **kwargs):
    config.set_default_section("core/process/qwen")
    pretrained_name = config.getoption("pretrained_name", "qwen3-4b-thinking")
    tokenizer_file = config.getoption("tokenizer_file", None)
    tokenizer_file = pop_value(
        tokenizer_file,
        nested_dict_value(pretrained_qwen_infos, pretrained_name, "tokenizer"),
    )
    tokenizer_file = cached_path(tokenizer_file)

    tokenizer_config = config.getoption("tokenizer_config", None)
    tokenizer_config = pop_value(
        tokenizer_config,
        nested_dict_value(
            pretrained_qwen_infos, pretrained_name, "tokenizer_config"
        ),
        check_none=False,
    )
    tokenizer_config = (
        cached_path(tokenizer_config) if tokenizer_config is not None else None
    )

    special_tokens_map = config.getoption("special_tokens_map", None)
    special_tokens_map = pop_value(
        special_tokens_map,
        nested_dict_value(
            pretrained_qwen_infos, pretrained_name, "special_tokens_map"
        ),
        check_none=False,
    )
    special_tokens_map = (
        cached_path(special_tokens_map) if special_tokens_map is not None else None
    )

    chat_template = config.getoption("chat_template", None)
    chat_template = pop_value(
        chat_template,
        nested_dict_value(pretrained_qwen_infos, pretrained_name, "chat_template"),
        check_none=False,
    )
    chat_template = (
        cached_path(chat_template) if chat_template is not None else None
    )

    return {
        "tokenizer_file": tokenizer_file,
        "tokenizer_config": tokenizer_config,
        "special_tokens_map": special_tokens_map,
        "chat_template": chat_template,
    }

_chat_template ¤

_chat_template(messages: List[Dict[str, Any]])
Source code in src/unitorch/cli/models/qwen/processing.py 
 97
 98
 99
100
101
102
@register_process("core/process/qwen/chat_template")
def _chat_template(
    self,
    messages: List[Dict[str, Any]],
):
    return super().chat_template(messages=messages)

_generation_inputs ¤

_generation_inputs(
    text: str, max_seq_length: Optional[int] = None
)
Source code in src/unitorch/cli/models/qwen/processing.py 
104
105
106
107
108
109
110
111
112
113
114
@register_process("core/process/qwen/generation/inputs")
def _generation_inputs(
    self,
    text: str,
    max_seq_length: Optional[int] = None,
):
    outputs = super().generation_inputs(
        text=text,
        max_seq_length=max_seq_length,
    )
    return TensorInputs(input_ids=outputs.input_ids)

_generation_labels ¤

_generation_labels(
    text: str, max_gen_seq_length: Optional[int] = None
)
Source code in src/unitorch/cli/models/qwen/processing.py 
116
117
118
119
120
121
122
123
124
125
126
127
128
129
@register_process("core/process/qwen/generation/labels")
def _generation_labels(
    self,
    text: str,
    max_gen_seq_length: Optional[int] = None,
):
    outputs = super().generation_labels(
        text=text,
        max_gen_seq_length=max_gen_seq_length,
    )
    return GenerationTargets(
        refs=outputs.input_ids,
        masks=outputs.attention_mask,
    )

_generation ¤

_generation(
    text: str,
    text_pair: str,
    max_seq_length: Optional[int] = None,
    max_gen_seq_length: Optional[int] = None,
)
Source code in src/unitorch/cli/models/qwen/processing.py 
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
@register_process("core/process/qwen/generation")
def _generation(
    self,
    text: str,
    text_pair: str,
    max_seq_length: Optional[int] = None,
    max_gen_seq_length: Optional[int] = None,
):
    outputs = super().generation(
        text=text,
        text_pair=text_pair,
        max_seq_length=max_seq_length,
        max_gen_seq_length=max_gen_seq_length,
    )
    return TensorInputs(
        input_ids=outputs.input_ids,
        attention_mask=outputs.attention_mask,
    ), GenerationTargets(
        refs=outputs.input_ids_label,
        masks=outputs.attention_mask_label,
    )

_dpo_generation ¤

_dpo_generation(
    text: str,
    win_text_pair: str,
    lose_text_pair: str,
    max_seq_length: Optional[int] = None,
    max_gen_seq_length: Optional[int] = None,
)
Source code in src/unitorch/cli/models/qwen/processing.py 
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
@register_process("core/process/qwen/dpo/generation")
def _dpo_generation(
    self,
    text: str,
    win_text_pair: str,
    lose_text_pair: str,
    max_seq_length: Optional[int] = None,
    max_gen_seq_length: Optional[int] = None,
):
    inputs = super().generation_inputs(
        text=text,
        max_seq_length=max_seq_length,
    )
    win_labels = super().generation_labels(
        text=win_text_pair,
        max_gen_seq_length=max_gen_seq_length,
    )
    lose_labels = super().generation_labels(
        text=lose_text_pair,
        max_gen_seq_length=max_gen_seq_length,
    )
    return TensorInputs(
        input_ids=inputs.input_ids,
        attention_mask=inputs.attention_mask,
        win_input_ids=win_labels.input_ids,
        win_attention_mask=win_labels.attention_mask,
        lose_input_ids=lose_labels.input_ids,
        lose_attention_mask=lose_labels.attention_mask,
    )

_messages_generation ¤

_messages_generation(
    messages: List[Dict[str, Any]],
    max_seq_length: Optional[int] = None,
)
Source code in src/unitorch/cli/models/qwen/processing.py 
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
@register_process("core/process/qwen/messages/generation")
def _messages_generation(
    self,
    messages: List[Dict[str, Any]],
    max_seq_length: Optional[int] = None,
):
    outputs = super().messages_generation(
        messages=messages,
        max_seq_length=max_seq_length,
    )
    return TensorInputs(
        input_ids=outputs.input_ids,
        attention_mask=outputs.attention_mask,
    ), GenerationTargets(
        refs=outputs.input_ids_label,
        masks=outputs.attention_mask_label,
    )

_messages_dpo_generation ¤

_messages_dpo_generation(
    messages: List[Dict[str, Any]],
    win_messages: List[Dict[str, Any]],
    lose_messages: List[Dict[str, Any]],
    max_seq_length: Optional[int] = None,
    max_gen_seq_length: Optional[int] = None,
)
Source code in src/unitorch/cli/models/qwen/processing.py 
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
@register_process("core/process/qwen/messages/dpo/generation")
def _messages_dpo_generation(
    self,
    messages: List[Dict[str, Any]],
    win_messages: List[Dict[str, Any]],
    lose_messages: List[Dict[str, Any]],
    max_seq_length: Optional[int] = None,
    max_gen_seq_length: Optional[int] = None,
):
    if isinstance(messages, dict):
        messages = [messages]
    if isinstance(win_messages, dict):
        win_messages = [win_messages]
    if isinstance(lose_messages, dict):
        lose_messages = [lose_messages]
    inputs = super().generation_inputs(
        text=super().chat_template(messages=messages),
        max_seq_length=max_seq_length,
    )
    win_labels = super().generation_labels(
        text=super().chat_template(messages=win_messages),
        max_gen_seq_length=max_gen_seq_length,
    )
    lose_labels = super().generation_labels(
        text=super().chat_template(messages=lose_messages),
        max_gen_seq_length=max_gen_seq_length,
    )
    return TensorInputs(
        input_ids=inputs.input_ids,
        attention_mask=inputs.attention_mask,
        win_input_ids=win_labels.input_ids,
        win_attention_mask=win_labels.attention_mask,
        lose_input_ids=lose_labels.input_ids,
        lose_attention_mask=lose_labels.attention_mask,
    )

_messages_grpo_generation ¤

_messages_grpo_generation(
    messages: List[Dict[str, Any]],
    messages_labels: List[Dict[str, Any]],
    max_seq_length: Optional[int] = None,
    max_gen_seq_length: Optional[int] = None,
)
Source code in src/unitorch/cli/models/qwen/processing.py 
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
@register_process("core/process/qwen/messages/grpo/generation")
def _messages_grpo_generation(
    self,
    messages: List[Dict[str, Any]],
    messages_labels: List[Dict[str, Any]],
    max_seq_length: Optional[int] = None,
    max_gen_seq_length: Optional[int] = None,
):
    inputs = super().generation_inputs(
        text=super().chat_template(messages=messages),
        max_seq_length=max_seq_length,
    )
    labels = super().generation_labels(
        text=super().chat_template(messages=messages_labels),
        max_gen_seq_length=max_gen_seq_length,
    )
    sample_ids = labels.input_ids
    sample_attention_mask = labels.attention_mask
    sampled_rewards = torch.tensor(1.0).float()
    return TensorInputs(
        input_ids=inputs.input_ids,
        attention_mask=inputs.attention_mask,
        sampled_ids=torch.stack([sample_ids, sample_ids], dim=0),
        sampled_attention_mask=torch.stack(
            [sample_attention_mask, sample_attention_mask], dim=0
        ),
        sampled_rewards=torch.stack([sampled_rewards, sampled_rewards], dim=0),
    )

_detokenize ¤

_detokenize(outputs: GenerationOutputs)
Source code in src/unitorch/cli/models/qwen/processing.py 
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
@register_process("core/postprocess/qwen/detokenize")
def _detokenize(
    self,
    outputs: GenerationOutputs,
):
    results = outputs.to_pandas()
    assert results.shape[0] == 0 or results.shape[0] == outputs.sequences.shape[0]

    decoded = super().detokenize(sequences=outputs.sequences)

    def cleanup_string(text: str) -> str:
        # Strip any thinking-mode CoT blocks emitted by Qwen3 reasoning models.
        text = re.sub(r"<think>.*?</think>", "", text, flags=re.DOTALL)
        # Collapse newlines into spaces for TSV/CSV output compatibility.
        text = re.sub(r"\n", " ", text)
        return text.strip()

    if isinstance(decoded[0], list):
        decoded = [list(map(cleanup_string, sequence)) for sequence in decoded]
    elif isinstance(decoded[0], str):
        decoded = list(map(cleanup_string, decoded))
    else:
        raise ValueError(
            f"Unsupported type for Qwen detokenize: {type(decoded[0])}"
        )
    results["decoded"] = decoded
    return WriterOutputs(results)

QWenVLProcessor¤

Tip

core/process/qwen_vl is the section for configuration of QWenVLProcessor.

Bases: QWenVLProcessor

Processor for Bloom language models.

Source code in src/unitorch/cli/models/qwen/processing_vl.py 
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
def __init__(
    self,
    tokenizer_file: str,
    vision_config_path: str,
    tokenizer_config: Optional[str] = None,
    special_tokens_map: Optional[str] = None,
    chat_template: Optional[str] = None,
    max_seq_length: Optional[int] = 128,
    max_gen_seq_length: Optional[int] = 128,
):
    super().__init__(
        tokenizer_file=tokenizer_file,
        vision_config_path=vision_config_path,
        tokenizer_config=tokenizer_config,
        special_tokens_map=special_tokens_map,
        chat_template=chat_template,
        max_seq_length=max_seq_length,
        max_gen_seq_length=max_gen_seq_length,
    )

from_config classmethod ¤

from_config(config, **kwargs)
Source code in src/unitorch/cli/models/qwen/processing_vl.py 
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
@classmethod
@config_defaults_init("core/process/qwen_vl")
def from_config(cls, config, **kwargs):
    config.set_default_section("core/process/qwen_vl")
    pretrained_name = config.getoption("pretrained_name", "qwen3-vl-8b-instruct")
    tokenizer_file = config.getoption("tokenizer_file", None)
    tokenizer_file = pop_value(
        tokenizer_file,
        nested_dict_value(pretrained_qwen_infos, pretrained_name, "tokenizer"),
    )
    tokenizer_file = cached_path(tokenizer_file)

    vision_config_path = config.getoption("vision_config_path", None)
    vision_config_path = pop_value(
        vision_config_path,
        nested_dict_value(pretrained_qwen_infos, pretrained_name, "vision_config"),
    )
    vision_config_path = cached_path(vision_config_path)

    tokenizer_config = config.getoption("tokenizer_config", None)
    tokenizer_config = pop_value(
        tokenizer_config,
        nested_dict_value(
            pretrained_qwen_infos, pretrained_name, "tokenizer_config"
        ),
        check_none=False,
    )
    tokenizer_config = (
        cached_path(tokenizer_config) if tokenizer_config is not None else None
    )

    special_tokens_map = config.getoption("special_tokens_map", None)
    special_tokens_map = pop_value(
        special_tokens_map,
        nested_dict_value(
            pretrained_qwen_infos, pretrained_name, "special_tokens_map"
        ),
        check_none=False,
    )
    special_tokens_map = (
        cached_path(special_tokens_map) if special_tokens_map is not None else None
    )

    chat_template = config.getoption("chat_template", None)
    chat_template = pop_value(
        chat_template,
        nested_dict_value(pretrained_qwen_infos, pretrained_name, "chat_template"),
        check_none=False,
    )
    chat_template = (
        cached_path(chat_template) if chat_template is not None else None
    )

    return {
        "tokenizer_file": tokenizer_file,
        "vision_config_path": vision_config_path,
        "tokenizer_config": tokenizer_config,
        "special_tokens_map": special_tokens_map,
        "chat_template": chat_template,
    }

_chat_template ¤

_chat_template(messages: List[Dict[str, Any]])
Source code in src/unitorch/cli/models/qwen/processing_vl.py 
107
108
109
110
111
112
@register_process("core/process/qwen_vl/chat_template")
def _chat_template(
    self,
    messages: List[Dict[str, Any]],
):
    return super().chat_template(messages=messages)

_generation_inputs ¤

_generation_inputs(
    text: str,
    images: Union[Image, str, List[Image], List[str]],
    max_seq_length: Optional[int] = None,
)
Source code in src/unitorch/cli/models/qwen/processing_vl.py 
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
@register_process("core/process/qwen_vl/generation/inputs")
def _generation_inputs(
    self,
    text: str,
    images: Union[Image.Image, str, List[Image.Image], List[str]],
    max_seq_length: Optional[int] = None,
):
    outputs = super().generation_inputs(
        text=text,
        images=images,
        max_seq_length=max_seq_length,
    )
    return TensorInputs(
        input_ids=outputs.input_ids,
        pixel_values=outputs.pixel_values,
        image_grid_thw=outputs.image_grid_thw,
    )

_generation_labels ¤

_generation_labels(
    text: str, max_gen_seq_length: Optional[int] = None
)
Source code in src/unitorch/cli/models/qwen/processing_vl.py 
132
133
134
135
136
137
138
139
140
141
142
143
144
145
@register_process("core/process/qwen_vl/generation/labels")
def _generation_labels(
    self,
    text: str,
    max_gen_seq_length: Optional[int] = None,
):
    outputs = super().generation_labels(
        text=text,
        max_gen_seq_length=max_gen_seq_length,
    )
    return GenerationTargets(
        refs=outputs.input_ids,
        masks=outputs.attention_mask,
    )

_generation ¤

_generation(
    text: str,
    images: Union[Image, str, List[Image], List[str]],
    text_pair: str,
    max_seq_length: Optional[int] = None,
    max_gen_seq_length: Optional[int] = None,
)
Source code in src/unitorch/cli/models/qwen/processing_vl.py 
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
@register_process("core/process/qwen_vl/generation")
def _generation(
    self,
    text: str,
    images: Union[Image.Image, str, List[Image.Image], List[str]],
    text_pair: str,
    max_seq_length: Optional[int] = None,
    max_gen_seq_length: Optional[int] = None,
):
    outputs = super().generation(
        text=text,
        text_pair=text_pair,
        images=images,
        max_seq_length=max_seq_length,
        max_gen_seq_length=max_gen_seq_length,
    )
    return TensorInputs(
        input_ids=outputs.input_ids,
        attention_mask=outputs.attention_mask,
        pixel_values=outputs.pixel_values,
        image_grid_thw=outputs.image_grid_thw,
    ), GenerationTargets(
        refs=outputs.input_ids_label,
        masks=outputs.attention_mask_label,
    )

_dpo_generation ¤

_dpo_generation(
    text: str,
    images: Union[Image, str, List[Image], List[str]],
    win_text_pair: str,
    lose_text_pair: str,
    max_seq_length: Optional[int] = None,
    max_gen_seq_length: Optional[int] = None,
)
Source code in src/unitorch/cli/models/qwen/processing_vl.py 
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
@register_process("core/process/qwen_vl/dpo/generation")
def _dpo_generation(
    self,
    text: str,
    images: Union[Image.Image, str, List[Image.Image], List[str]],
    win_text_pair: str,
    lose_text_pair: str,
    max_seq_length: Optional[int] = None,
    max_gen_seq_length: Optional[int] = None,
):
    inputs = super().generation_inputs(
        text=text,
        images=images,
        max_seq_length=max_seq_length,
    )
    win_labels = super().generation_labels(
        text=win_text_pair,
        max_gen_seq_length=max_gen_seq_length,
    )
    lose_labels = super().generation_labels(
        text=lose_text_pair,
        max_gen_seq_length=max_gen_seq_length,
    )
    return TensorInputs(
        input_ids=inputs.input_ids,
        attention_mask=inputs.attention_mask,
        pixel_values=inputs.pixel_values,
        image_grid_thw=inputs.image_grid_thw,
        win_input_ids=win_labels.input_ids,
        win_attention_mask=win_labels.attention_mask,
        lose_input_ids=lose_labels.input_ids,
        lose_attention_mask=lose_labels.attention_mask,
    )

_messages_generation ¤

_messages_generation(
    messages: List[Dict[str, Any]],
    images: Union[Image, str, List[Image], List[str]],
    max_seq_length: Optional[int] = None,
)
Source code in src/unitorch/cli/models/qwen/processing_vl.py 
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
@register_process("core/process/qwen_vl/messages/generation")
def _messages_generation(
    self,
    messages: List[Dict[str, Any]],
    images: Union[Image.Image, str, List[Image.Image], List[str]],
    max_seq_length: Optional[int] = None,
):
    outputs = super().messages_generation(
        messages=messages,
        images=images,
        max_seq_length=max_seq_length,
    )
    return TensorInputs(
        input_ids=outputs.input_ids,
        attention_mask=outputs.attention_mask,
        pixel_values=outputs.pixel_values,
        image_grid_thw=outputs.image_grid_thw,
    ), GenerationTargets(
        refs=outputs.input_ids_label,
        masks=outputs.attention_mask_label,
    )

_messages_dpo_generation ¤

_messages_dpo_generation(
    messages: List[Dict[str, Any]],
    images: Union[Image, str, List[Image], List[str]],
    win_messages: List[Dict[str, Any]],
    lose_messages: List[Dict[str, Any]],
    max_seq_length: Optional[int] = None,
    max_gen_seq_length: Optional[int] = None,
)
Source code in src/unitorch/cli/models/qwen/processing_vl.py 
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
@register_process("core/process/qwen_vl/messages/dpo/generation")
def _messages_dpo_generation(
    self,
    messages: List[Dict[str, Any]],
    images: Union[Image.Image, str, List[Image.Image], List[str]],
    win_messages: List[Dict[str, Any]],
    lose_messages: List[Dict[str, Any]],
    max_seq_length: Optional[int] = None,
    max_gen_seq_length: Optional[int] = None,
):
    if isinstance(messages, dict):
        messages = [messages]
    if isinstance(win_messages, dict):
        win_messages = [win_messages]
    if isinstance(lose_messages, dict):
        lose_messages = [lose_messages]
    inputs = super().generation_inputs(
        text=super().chat_template(messages=messages),
        images=images,
        max_seq_length=max_seq_length,
    )
    win_labels = super().generation_labels(
        text=super().chat_template(messages=win_messages),
        max_gen_seq_length=max_gen_seq_length,
    )
    lose_labels = super().generation_labels(
        text=super().chat_template(messages=lose_messages),
        max_gen_seq_length=max_gen_seq_length,
    )
    return TensorInputs(
        input_ids=inputs.input_ids,
        attention_mask=inputs.attention_mask,
        pixel_values=inputs.pixel_values,
        image_grid_thw=inputs.image_grid_thw,
        win_input_ids=win_labels.input_ids,
        win_attention_mask=win_labels.attention_mask,
        lose_input_ids=lose_labels.input_ids,
        lose_attention_mask=lose_labels.attention_mask,
    )

_detokenize ¤

_detokenize(outputs: GenerationOutputs)
Source code in src/unitorch/cli/models/qwen/processing_vl.py 
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
@register_process("core/postprocess/qwen_vl/detokenize")
def _detokenize(
    self,
    outputs: GenerationOutputs,
):
    results = outputs.to_pandas()
    assert results.shape[0] == 0 or results.shape[0] == outputs.sequences.shape[0]

    decoded = super().detokenize(sequences=outputs.sequences)
    cleanup_string = lambda text: re.sub(r"\n", " ", text)
    if isinstance(decoded[0], list):
        decoded = [list(map(cleanup_string, sequence)) for sequence in decoded]
    elif isinstance(decoded[0], str):
        decoded = list(map(cleanup_string, decoded))
    else:
        raise ValueError(
            f"Unsupported type for Qwen detokenize: {type(decoded[0])}"
        )
    results["decoded"] = decoded
    return WriterOutputs(results)

QWen3ForGeneration¤

Tip

core/model/generation/qwen3 is the section for configuration of QWen3ForGeneration.

Bases: QWen3ForGeneration

Qwen3 model for text generation.

Source code in src/unitorch/cli/models/qwen/modeling.py 
29
30
31
32
33
34
35
36
37
def __init__(
    self,
    config_path: str,
    gradient_checkpointing: Optional[bool] = False,
):
    super().__init__(
        config_path=config_path,
        gradient_checkpointing=gradient_checkpointing,
    )

from_config classmethod ¤

from_config(config, **kwargs)
Source code in src/unitorch/cli/models/qwen/modeling.py 
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
@classmethod
@config_defaults_init("core/model/generation/qwen3")
def from_config(cls, config, **kwargs):
    config.set_default_section("core/model/generation/qwen3")
    pretrained_name = config.getoption("pretrained_name", "qwen3-4b-thinking")
    pretrained_lora_name = config.getoption("pretrained_lora_name", None)
    config_path = config.getoption("config_path", None)
    config_path = pop_value(
        config_path,
        nested_dict_value(pretrained_qwen_infos, pretrained_name, "config"),
    )

    config_path = cached_path(config_path)
    gradient_checkpointing = config.getoption("gradient_checkpointing", False)

    inst = cls(config_path, gradient_checkpointing)
    pretrained_weight_path = config.getoption("pretrained_weight_path", None)
    weight_path = pop_value(
        pretrained_weight_path,
        nested_dict_value(pretrained_qwen_infos, pretrained_name, "weight"),
        check_none=False,
    )

    if weight_path is not None:
        inst.from_pretrained(weight_path)

    pretrained_lora_weight_path = config.getoption(
        "pretrained_lora_weight_path", None
    )
    lora_weight_path = pop_value(
        pretrained_lora_weight_path,
        nested_dict_value(pretrained_qwen_extensions_infos, pretrained_lora_name),
        check_none=False,
    )
    pretrained_lora_weight = config.getoption("pretrained_lora_weight", 1.0)
    pretrained_lora_alpha = config.getoption("pretrained_lora_alpha", 32.0)
    if lora_weight_path is not None:
        inst.load_lora_weights(
            lora_weight_path,
            lora_weights=pretrained_lora_weight,
            lora_alphas=pretrained_lora_alpha,
            save_base_state=False,
        )

    return inst

forward ¤

forward(
    input_ids: Tensor,
    attention_mask: Optional[Tensor] = None,
)
Source code in src/unitorch/cli/models/qwen/modeling.py 
85
86
87
88
89
90
91
92
93
94
95
96
97
98
@autocast(
    device_type=("cuda" if torch.cuda.is_available() else "cpu"),
    dtype=(torch.bfloat16 if is_bfloat16_available() else torch.float32),
)
def forward(
    self,
    input_ids: torch.Tensor,
    attention_mask: Optional[torch.Tensor] = None,
):
    outputs = super().forward(
        input_ids=input_ids,
        attention_mask=attention_mask,
    )
    return GenerationOutputs(sequences=outputs)

generate ¤

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,
)
Source code in src/unitorch/cli/models/qwen/modeling.py 
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
@config_defaults_method("core/model/generation/qwen3")
@torch.no_grad()
@autocast(
    device_type=("cuda" if torch.cuda.is_available() else "cpu"),
    dtype=(torch.bfloat16 if is_bfloat16_available() else torch.float32),
)
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,
):
    outputs = super().generate(
        input_ids,
        num_beams=num_beams,
        decoder_start_token_id=decoder_start_token_id,
        decoder_end_token_id=decoder_end_token_id,
        decoder_pad_token_id=decoder_pad_token_id,
        num_return_sequences=num_return_sequences,
        min_gen_seq_length=min_gen_seq_length,
        max_gen_seq_length=max_gen_seq_length,
        repetition_penalty=repetition_penalty,
        no_repeat_ngram_size=no_repeat_ngram_size,
        early_stopping=early_stopping,
        length_penalty=length_penalty,
        num_beam_groups=num_beam_groups,
        diversity_penalty=diversity_penalty,
        do_sample=do_sample,
        temperature=temperature,
        top_k=top_k,
        top_p=top_p,
    )

    return GenerationOutputs(
        sequences=outputs.sequences,
        sequences_scores=outputs.sequences_scores,
    )

QWen3VLForGeneration¤

Tip

core/model/generation/qwen3_vl is the section for configuration of QWen3VLForGeneration.

Bases: QWen3VLForGeneration

Qwen3 model for text generation.

Source code in src/unitorch/cli/models/qwen/modeling_vl.py 
29
30
31
32
33
34
35
36
37
def __init__(
    self,
    config_path: str,
    gradient_checkpointing: Optional[bool] = False,
):
    super().__init__(
        config_path=config_path,
        gradient_checkpointing=gradient_checkpointing,
    )

from_config classmethod ¤

from_config(config, **kwargs)
Source code in src/unitorch/cli/models/qwen/modeling_vl.py 
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
@classmethod
@config_defaults_init("core/model/generation/qwen3_vl")
def from_config(cls, config, **kwargs):
    config.set_default_section("core/model/generation/qwen3_vl")
    pretrained_name = config.getoption("pretrained_name", "qwen3-vl-8b-instruct")
    pretrained_lora_name = config.getoption("pretrained_lora_name", None)
    config_path = config.getoption("config_path", None)
    config_path = pop_value(
        config_path,
        nested_dict_value(pretrained_qwen_infos, pretrained_name, "config"),
    )

    config_path = cached_path(config_path)
    gradient_checkpointing = config.getoption("gradient_checkpointing", False)

    inst = cls(config_path, gradient_checkpointing)
    pretrained_weight_path = config.getoption("pretrained_weight_path", None)
    weight_path = pop_value(
        pretrained_weight_path,
        nested_dict_value(pretrained_qwen_infos, pretrained_name, "weight"),
        check_none=False,
    )

    if weight_path is not None:
        inst.from_pretrained(weight_path)

    pretrained_lora_weight_path = config.getoption(
        "pretrained_lora_weight_path", None
    )
    lora_weight_path = pop_value(
        pretrained_lora_weight_path,
        nested_dict_value(pretrained_qwen_extensions_infos, pretrained_lora_name),
        check_none=False,
    )
    pretrained_lora_weight = config.getoption("pretrained_lora_weight", 1.0)
    pretrained_lora_alpha = config.getoption("pretrained_lora_alpha", 32.0)
    if lora_weight_path is not None:
        inst.load_lora_weights(
            lora_weight_path,
            lora_weights=pretrained_lora_weight,
            lora_alphas=pretrained_lora_alpha,
            save_base_state=False,
        )

    return inst

forward ¤

forward(
    input_ids: Tensor,
    pixel_values: Optional[Tensor] = None,
    image_grid_thw: Optional[Tensor] = None,
    attention_mask: Optional[Tensor] = None,
)
Source code in src/unitorch/cli/models/qwen/modeling_vl.py 
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
@autocast(
    device_type=("cuda" if torch.cuda.is_available() else "cpu"),
    dtype=(torch.bfloat16 if is_bfloat16_available() else torch.float32),
)
def forward(
    self,
    input_ids: torch.Tensor,
    pixel_values: Optional[torch.Tensor] = None,
    image_grid_thw: Optional[torch.Tensor] = None,
    attention_mask: Optional[torch.Tensor] = None,
):
    outputs = super().forward(
        input_ids=input_ids,
        pixel_values=pixel_values,
        image_grid_thw=image_grid_thw,
        attention_mask=attention_mask,
    )
    return GenerationOutputs(sequences=outputs)

generate ¤

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,
)
Source code in src/unitorch/cli/models/qwen/modeling_vl.py 
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
@config_defaults_method("core/model/generation/qwen3_vl")
@torch.no_grad()
@autocast(
    device_type=("cuda" if torch.cuda.is_available() else "cpu"),
    dtype=(torch.bfloat16 if is_bfloat16_available() else torch.float32),
)
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,
):
    outputs = super().generate(
        input_ids,
        pixel_values=pixel_values,
        image_grid_thw=image_grid_thw,
        num_beams=num_beams,
        decoder_start_token_id=decoder_start_token_id,
        decoder_end_token_id=decoder_end_token_id,
        decoder_pad_token_id=decoder_pad_token_id,
        num_return_sequences=num_return_sequences,
        min_gen_seq_length=min_gen_seq_length,
        max_gen_seq_length=max_gen_seq_length,
        repetition_penalty=repetition_penalty,
        no_repeat_ngram_size=no_repeat_ngram_size,
        early_stopping=early_stopping,
        length_penalty=length_penalty,
        num_beam_groups=num_beam_groups,
        diversity_penalty=diversity_penalty,
        do_sample=do_sample,
        temperature=temperature,
        top_k=top_k,
        top_p=top_p,
    )

    return GenerationOutputs(
        sequences=outputs.sequences,
        sequences_scores=outputs.sequences_scores,
    )