Skip to content

unitorch.cli.models.diffusers¤

WanProcessor¤

Tip

core/process/diffusion/wan is the section for configuration of WanProcessor.

Bases: WanProcessor

Source code in src/unitorch/cli/models/diffusers/processing_wan.py 
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
def __init__(
    self,
    vocab_path: str,
    vae_config_path: Optional[str] = None,
    max_seq_length: Optional[int] = 512,
    position_start_id: Optional[int] = 0,
    video_size: Optional[Tuple[int, int]] = None,
):
    super().__init__(
        vocab_path=vocab_path,
        vae_config_path=vae_config_path,
        max_seq_length=max_seq_length,
        position_start_id=position_start_id,
        video_size=video_size,
    )

from_config classmethod ¤

from_config(config, **kwargs)
Source code in src/unitorch/cli/models/diffusers/processing_wan.py 
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
@classmethod
@config_defaults_init("core/process/diffusion/wan")
def from_config(cls, config, **kwargs):
    config.set_default_section("core/process/diffusion/wan")
    pretrained_name = config.getoption("pretrained_name", "wan-v2.2-i2v-14b")
    pretrained_infos = nested_dict_value(pretrained_stable_infos, pretrained_name)

    vocab_path = config.getoption("vocab_path", None)
    vocab_path = pop_value(
        vocab_path,
        nested_dict_value(pretrained_infos, "text", "vocab"),
    )
    vocab_path = cached_path(vocab_path)

    vae_config_path = config.getoption("vae_config_path", None)
    vae_config_path = pop_value(
        vae_config_path,
        nested_dict_value(pretrained_infos, "vae", "config"),
    )
    vae_config_path = cached_path(vae_config_path)

    return {
        "vocab_path": vocab_path,
        "vae_config_path": vae_config_path,
    }

_text2video ¤

_text2video(
    prompt: str,
    video: Union[VideoCapture, str, List[Image]],
    max_seq_length: Optional[int] = None,
)
Source code in src/unitorch/cli/models/diffusers/processing_wan.py 
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
@register_process("core/process/diffusion/wan/text2video")
def _text2video(
    self,
    prompt: str,
    video: Union[cv2.VideoCapture, str, List[Image.Image]],
    max_seq_length: Optional[int] = None,
):
    outputs = super().text2video(
        prompt=prompt,
        video=video,
        max_seq_length=max_seq_length,
    )
    return TensorInputs(
        pixel_values=outputs.pixel_values,
        input_ids=outputs.input_ids,
        attention_mask=outputs.attention_mask,
    )

_text2video_inputs ¤

_text2video_inputs(
    prompt: str,
    negative_prompt: Optional[str] = "",
    max_seq_length: Optional[int] = None,
)
Source code in src/unitorch/cli/models/diffusers/processing_wan.py 
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
@register_process("core/process/diffusion/wan/text2video/inputs")
def _text2video_inputs(
    self,
    prompt: str,
    negative_prompt: Optional[str] = "",
    max_seq_length: Optional[int] = None,
):
    outputs = super().text2video_inputs(
        prompt=prompt,
        negative_prompt=negative_prompt,
        max_seq_length=max_seq_length,
    )
    return TensorInputs(
        input_ids=outputs.input_ids,
        negative_input_ids=outputs.negative_input_ids,
        attention_mask=outputs.attention_mask,
        negative_attention_mask=outputs.negative_attention_mask,
    )

_image2video ¤

_image2video(
    prompt: str,
    video: Union[VideoCapture, str, List[Image]],
    image: Optional[Union[Image, str]] = None,
    max_seq_length: Optional[int] = None,
)
Source code in src/unitorch/cli/models/diffusers/processing_wan.py 
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
@register_process("core/process/diffusion/wan/image2video")
def _image2video(
    self,
    prompt: str,
    video: Union[cv2.VideoCapture, str, List[Image.Image]],
    image: Optional[Union[Image.Image, str]] = None,
    max_seq_length: Optional[int] = None,
):
    outputs = super().image2video(
        prompt=prompt,
        video=video,
        image=image,
        max_seq_length=max_seq_length,
    )
    return TensorInputs(
        pixel_values=outputs.pixel_values,
        input_ids=outputs.input_ids,
        attention_mask=outputs.attention_mask,
        vae_pixel_values=outputs.vae_pixel_values,
    )

_image2video_inputs ¤

_image2video_inputs(
    prompt: str,
    image: Union[Image, str],
    negative_prompt: Optional[str] = "",
    max_seq_length: Optional[int] = None,
)
Source code in src/unitorch/cli/models/diffusers/processing_wan.py 
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
@register_process("core/process/diffusion/wan/image2video/inputs")
def _image2video_inputs(
    self,
    prompt: str,
    image: Union[Image.Image, str],
    negative_prompt: Optional[str] = "",
    max_seq_length: Optional[int] = None,
):
    outputs = super().image2video_inputs(
        prompt=prompt,
        image=image,
        negative_prompt=negative_prompt,
        max_seq_length=max_seq_length,
    )
    return TensorInputs(
        vae_pixel_values=outputs.vae_pixel_values,
        input_ids=outputs.input_ids,
        attention_mask=outputs.attention_mask,
        negative_input_ids=outputs.negative_input_ids,
        negative_attention_mask=outputs.negative_attention_mask,
    )

WanForText2VideoGeneration¤

Tip

core/model/diffusers/text2video/wan is the section for configuration of WanForText2VideoGeneration.

Bases: WanForText2VideoGeneration

Source code in src/unitorch/cli/models/diffusers/modeling_wan.py 
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
60
61
62
63
64
def __init__(
    self,
    config_path: str,
    text_config_path: str,
    vae_config_path: str,
    scheduler_config_path: str,
    config2_path: Optional[str] = None,
    num_train_timesteps: Optional[int] = 1000,
    num_infer_timesteps: Optional[int] = 50,
    freeze_vae_encoder: Optional[bool] = True,
    freeze_text_encoder: Optional[bool] = True,
    snr_gamma: Optional[float] = 5.0,
    boundary_ratio: Optional[float] = 0.9,
    seed: Optional[int] = 1123,
    gradient_checkpointing: Optional[bool] = True,
):
    super().__init__(
        config_path=config_path,
        text_config_path=text_config_path,
        vae_config_path=vae_config_path,
        scheduler_config_path=scheduler_config_path,
        config2_path=config2_path,
        num_train_timesteps=num_train_timesteps,
        num_infer_timesteps=num_infer_timesteps,
        freeze_vae_encoder=freeze_vae_encoder,
        freeze_text_encoder=freeze_text_encoder,
        snr_gamma=snr_gamma,
        boundary_ratio=boundary_ratio,
        seed=seed,
        gradient_checkpointing=gradient_checkpointing,
    )

from_config classmethod ¤

from_config(config, **kwargs)
Source code in src/unitorch/cli/models/diffusers/modeling_wan.py 
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
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
188
189
190
191
192
193
194
195
@classmethod
@config_defaults_init("core/model/diffusers/text2video/wan")
def from_config(cls, config, **kwargs):
    config.set_default_section("core/model/diffusers/text2video/wan")
    pretrained_name = config.getoption("pretrained_name", "wan-v2.2-t2v-14b")
    pretrained_infos = nested_dict_value(pretrained_stable_infos, pretrained_name)

    config_path = config.getoption("config_path", None)
    config_path = pop_value(
        config_path,
        nested_dict_value(pretrained_infos, "transformer", "config"),
    )
    config_path = cached_path(config_path)

    config2_path = config.getoption("config2_path", None)
    config2_path = pop_value(
        config2_path,
        nested_dict_value(pretrained_infos, "transformer2", "config"),
    )

    if config2_path is not None:
        config2_path = cached_path(config2_path)

    text_config_path = config.getoption("text_config_path", None)
    text_config_path = pop_value(
        text_config_path,
        nested_dict_value(pretrained_infos, "text", "config"),
    )
    text_config_path = cached_path(text_config_path)

    vae_config_path = config.getoption("vae_config_path", None)
    vae_config_path = pop_value(
        vae_config_path,
        nested_dict_value(pretrained_infos, "vae", "config"),
    )
    vae_config_path = cached_path(vae_config_path)

    scheduler_config_path = config.getoption("scheduler_config_path", None)
    scheduler_config_path = pop_value(
        scheduler_config_path,
        nested_dict_value(pretrained_infos, "scheduler"),
    )
    scheduler_config_path = cached_path(scheduler_config_path)

    num_train_timesteps = config.getoption("num_train_timesteps", 1000)
    num_infer_timesteps = config.getoption("num_infer_timesteps", 50)
    freeze_vae_encoder = config.getoption("freeze_vae_encoder", True)
    freeze_text_encoder = config.getoption("freeze_text_encoder", True)
    snr_gamma = config.getoption("snr_gamma", 5.0)
    boundary_ratio = config.getoption(
        "boundary_ratio",
        nested_dict_value(pretrained_infos, "boundary_ratio") or 0.9,
    )
    seed = config.getoption("seed", 1123)
    gradient_checkpointing = config.getoption("gradient_checkpointing", True)

    inst = cls(
        config_path=config_path,
        text_config_path=text_config_path,
        vae_config_path=vae_config_path,
        scheduler_config_path=scheduler_config_path,
        config2_path=config2_path,
        num_train_timesteps=num_train_timesteps,
        num_infer_timesteps=num_infer_timesteps,
        freeze_vae_encoder=freeze_vae_encoder,
        freeze_text_encoder=freeze_text_encoder,
        snr_gamma=snr_gamma,
        boundary_ratio=boundary_ratio,
        seed=seed,
        gradient_checkpointing=gradient_checkpointing,
    )

    weight_path = config.getoption("pretrained_weight_path", None)

    state_dict = None
    if weight_path is None and pretrained_infos is not None:
        state_dict = [
            load_weight(
                nested_dict_value(pretrained_infos, "transformer", "weight"),
                prefix_keys={"": "transformer."},
            ),
            load_weight(
                nested_dict_value(pretrained_infos, "transformer2", "weight"),
                prefix_keys={"": "transformer2."},
            ),
            load_weight(
                nested_dict_value(pretrained_infos, "text", "weight"),
                prefix_keys={"": "text."},
            ),
            load_weight(
                nested_dict_value(pretrained_infos, "vae", "weight"),
                prefix_keys={"": "vae."},
            ),
        ]

    inst.from_pretrained(weight_path, state_dict=state_dict)

    pretrained_lora_names = config.getoption("pretrained_lora_names", None)
    pretrained_lora_weights = config.getoption("pretrained_lora_weights", 1.0)
    pretrained_lora_alphas = config.getoption("pretrained_lora_alphas", 32.0)

    if isinstance(pretrained_lora_names, str):
        pretrained_lora_weights_path = nested_dict_value(
            pretrained_stable_extensions_infos,
            pretrained_lora_names,
            "lora",
            "weight",
        )
    elif isinstance(pretrained_lora_names, list):
        pretrained_lora_weights_path = [
            nested_dict_value(
                pretrained_stable_extensions_infos, name, "lora", "weight"
            )
            for name in pretrained_lora_names
        ]
    else:
        pretrained_lora_weights_path = None

    lora_weights_path = config.getoption(
        "pretrained_lora_weights_path", pretrained_lora_weights_path
    )
    if lora_weights_path is not None:
        inst.load_lora_weights(
            lora_files=lora_weights_path,
            lora_weights=pretrained_lora_weights,
            lora_alphas=pretrained_lora_alphas,
            replace_keys={},
            save_base_state=False,
        )
    return inst

forward ¤

forward(
    input_ids: Tensor,
    pixel_values: Tensor,
    attention_mask: Optional[Tensor] = None,
)
Source code in src/unitorch/cli/models/diffusers/modeling_wan.py 
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
@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: torch.Tensor,
    attention_mask: Optional[torch.Tensor] = None,
):
    loss = super().forward(
        input_ids=input_ids,
        pixel_values=pixel_values,
        attention_mask=attention_mask,
    )
    return LossOutputs(loss=loss)

generate ¤

generate(
    input_ids: Tensor,
    negative_input_ids: Tensor,
    attention_mask: Optional[Tensor] = None,
    negative_attention_mask: Optional[Tensor] = None,
    height: Optional[int] = 480,
    width: Optional[int] = 832,
    num_frames: Optional[int] = 81,
    guidance_scale: Optional[float] = 5.0,
)
Source code in src/unitorch/cli/models/diffusers/modeling_wan.py 
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
@config_defaults_method("core/model/diffusers/text2video/wan")
@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,
    negative_input_ids: torch.Tensor,
    attention_mask: Optional[torch.Tensor] = None,
    negative_attention_mask: Optional[torch.Tensor] = None,
    height: Optional[int] = 480,
    width: Optional[int] = 832,
    num_frames: Optional[int] = 81,
    guidance_scale: Optional[float] = 5.0,
):
    outputs = super().generate(
        input_ids=input_ids,
        negative_input_ids=negative_input_ids,
        attention_mask=attention_mask,
        negative_attention_mask=negative_attention_mask,
        height=height,
        width=width,
        num_frames=num_frames,
        guidance_scale=guidance_scale,
    )

    return DiffusionOutputs(outputs=outputs.frames.float())

WanForImage2VideoGeneration¤

Tip

core/model/diffusers/image2video/wan is the section for configuration of WanForImage2VideoGeneration.

Bases: WanForImage2VideoGeneration

Source code in src/unitorch/cli/models/diffusers/modeling_wan.py 
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
def __init__(
    self,
    config_path: str,
    text_config_path: str,
    vae_config_path: str,
    scheduler_config_path: str,
    config2_path: Optional[str] = None,
    num_train_timesteps: Optional[int] = 1000,
    num_infer_timesteps: Optional[int] = 50,
    freeze_vae_encoder: Optional[bool] = True,
    freeze_text_encoder: Optional[bool] = True,
    snr_gamma: Optional[float] = 5.0,
    boundary_ratio: Optional[float] = 0.9,
    seed: Optional[int] = 1123,
    gradient_checkpointing: Optional[bool] = True,
):
    super().__init__(
        config_path=config_path,
        text_config_path=text_config_path,
        vae_config_path=vae_config_path,
        scheduler_config_path=scheduler_config_path,
        config2_path=config2_path,
        num_train_timesteps=num_train_timesteps,
        num_infer_timesteps=num_infer_timesteps,
        freeze_vae_encoder=freeze_vae_encoder,
        freeze_text_encoder=freeze_text_encoder,
        snr_gamma=snr_gamma,
        boundary_ratio=boundary_ratio,
        seed=seed,
        gradient_checkpointing=gradient_checkpointing,
    )

from_config classmethod ¤

from_config(config, **kwargs)
Source code in src/unitorch/cli/models/diffusers/modeling_wan.py 
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
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
@classmethod
@config_defaults_init("core/model/diffusers/image2video/wan")
def from_config(cls, config, **kwargs):
    config.set_default_section("core/model/diffusers/image2video/wan")
    pretrained_name = config.getoption("pretrained_name", "wan-v2.2-i2v-14b")
    pretrained_infos = nested_dict_value(pretrained_stable_infos, pretrained_name)

    config_path = config.getoption("config_path", None)
    config_path = pop_value(
        config_path,
        nested_dict_value(pretrained_infos, "transformer", "config"),
    )
    config_path = cached_path(config_path)

    config2_path = config.getoption("config2_path", None)
    config2_path = pop_value(
        config2_path,
        nested_dict_value(pretrained_infos, "transformer2", "config"),
    )

    if config2_path is not None:
        config2_path = cached_path(config2_path)

    text_config_path = config.getoption("text_config_path", None)
    text_config_path = pop_value(
        text_config_path,
        nested_dict_value(pretrained_infos, "text", "config"),
    )
    text_config_path = cached_path(text_config_path)

    vae_config_path = config.getoption("vae_config_path", None)
    vae_config_path = pop_value(
        vae_config_path,
        nested_dict_value(pretrained_infos, "vae", "config"),
    )
    vae_config_path = cached_path(vae_config_path)

    scheduler_config_path = config.getoption("scheduler_config_path", None)
    scheduler_config_path = pop_value(
        scheduler_config_path,
        nested_dict_value(pretrained_infos, "scheduler"),
    )
    scheduler_config_path = cached_path(scheduler_config_path)

    num_train_timesteps = config.getoption("num_train_timesteps", 1000)
    num_infer_timesteps = config.getoption("num_infer_timesteps", 50)
    freeze_vae_encoder = config.getoption("freeze_vae_encoder", True)
    freeze_text_encoder = config.getoption("freeze_text_encoder", True)
    snr_gamma = config.getoption("snr_gamma", 5.0)
    boundary_ratio = config.getoption(
        "boundary_ratio",
        nested_dict_value(pretrained_infos, "boundary_ratio") or 0.9,
    )
    seed = config.getoption("seed", 1123)
    gradient_checkpointing = config.getoption("gradient_checkpointing", True)

    inst = cls(
        config_path=config_path,
        text_config_path=text_config_path,
        vae_config_path=vae_config_path,
        scheduler_config_path=scheduler_config_path,
        config2_path=config2_path,
        num_train_timesteps=num_train_timesteps,
        num_infer_timesteps=num_infer_timesteps,
        freeze_vae_encoder=freeze_vae_encoder,
        freeze_text_encoder=freeze_text_encoder,
        snr_gamma=snr_gamma,
        boundary_ratio=boundary_ratio,
        seed=seed,
        gradient_checkpointing=gradient_checkpointing,
    )

    weight_path = config.getoption("pretrained_weight_path", None)

    state_dict = None
    if weight_path is None and pretrained_infos is not None:
        state_dict = [
            load_weight(
                nested_dict_value(pretrained_infos, "transformer", "weight"),
                prefix_keys={"": "transformer."},
            ),
            load_weight(
                nested_dict_value(pretrained_infos, "transformer2", "weight"),
                prefix_keys={"": "transformer2."},
            ),
            load_weight(
                nested_dict_value(pretrained_infos, "text", "weight"),
                prefix_keys={"": "text."},
            ),
            load_weight(
                nested_dict_value(pretrained_infos, "vae", "weight"),
                prefix_keys={"": "vae."},
            ),
        ]

    elif weight_path is not None:
        state_dict = load_weight(weight_path)

    if state_dict is not None:
        inst.from_pretrained(state_dict=state_dict)

    pretrained_lora_names = config.getoption("pretrained_lora_names", None)
    pretrained_lora_weights = config.getoption("pretrained_lora_weights", 1.0)
    pretrained_lora_alphas = config.getoption("pretrained_lora_alphas", 32.0)

    if isinstance(pretrained_lora_names, str):
        pretrained_lora_weights_path = nested_dict_value(
            pretrained_stable_extensions_infos,
            pretrained_lora_names,
            "lora",
            "weight",
        )
    elif isinstance(pretrained_lora_names, list):
        pretrained_lora_weights_path = [
            nested_dict_value(
                pretrained_stable_extensions_infos, name, "lora", "weight"
            )
            for name in pretrained_lora_names
        ]
    else:
        pretrained_lora_weights_path = None

    lora_weights_path = config.getoption(
        "pretrained_lora_weights_path", pretrained_lora_weights_path
    )
    if lora_weights_path is not None:
        inst.load_lora_weights(
            lora_files=lora_weights_path,
            lora_weights=pretrained_lora_weights,
            lora_alphas=pretrained_lora_alphas,
            replace_keys={},
            save_base_state=False,
        )
    return inst

forward ¤

forward(
    pixel_values: Tensor,
    vae_pixel_values: Tensor,
    input_ids: Tensor,
    attention_mask: Optional[Tensor] = None,
)
Source code in src/unitorch/cli/models/diffusers/modeling_wan.py 
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
@autocast(
    device_type=("cuda" if torch.cuda.is_available() else "cpu"),
    dtype=(torch.bfloat16 if is_bfloat16_available() else torch.float32),
)
def forward(
    self,
    pixel_values: torch.Tensor,
    vae_pixel_values: torch.Tensor,
    input_ids: torch.Tensor,
    attention_mask: Optional[torch.Tensor] = None,
):
    loss = super().forward(
        pixel_values=pixel_values,
        vae_pixel_values=vae_pixel_values,
        input_ids=input_ids,
        attention_mask=attention_mask,
    )
    return LossOutputs(loss=loss)

generate ¤

generate(
    input_ids: Tensor,
    negative_input_ids: Tensor,
    vae_pixel_values: Tensor,
    attention_mask: Optional[Tensor] = None,
    negative_attention_mask: Optional[Tensor] = None,
    num_frames: Optional[int] = 81,
    guidance_scale: Optional[float] = 5.0,
)
Source code in src/unitorch/cli/models/diffusers/modeling_wan.py 
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
@config_defaults_method("core/model/diffusers/image2video/wan")
@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,
    negative_input_ids: torch.Tensor,
    vae_pixel_values: torch.Tensor,
    attention_mask: Optional[torch.Tensor] = None,
    negative_attention_mask: Optional[torch.Tensor] = None,
    num_frames: Optional[int] = 81,
    guidance_scale: Optional[float] = 5.0,
):
    outputs = super().generate(
        input_ids=input_ids,
        negative_input_ids=negative_input_ids,
        vae_pixel_values=vae_pixel_values,
        attention_mask=attention_mask,
        negative_attention_mask=negative_attention_mask,
        num_frames=num_frames,
        guidance_scale=guidance_scale,
    )

    return DiffusionOutputs(outputs=outputs.frames.float())

QWenImageProcessor¤

Tip

core/process/diffusion/qwen_image is the section for configuration of QWenImageProcessor.

Bases: QWenImageProcessor

Source code in src/unitorch/cli/models/diffusers/processing_qwen_image.py 
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
def __init__(
    self,
    vocab_path: str,
    merge_path: str,
    vision_config_path: Optional[str] = None,
    vae_config_path: Optional[str] = None,
    tokenizer_config: Optional[str] = None,
    special_tokens_map: Optional[str] = None,
    max_seq_length: Optional[int] = 12800,
    image_size: Optional[Tuple[int, int]] = None,
    center_crop: Optional[bool] = False,
    random_flip: Optional[bool] = False,
):
    super().__init__(
        vocab_path=vocab_path,
        merge_path=merge_path,
        vision_config_path=vision_config_path,
        vae_config_path=vae_config_path,
        tokenizer_config=tokenizer_config,
        special_tokens_map=special_tokens_map,
        max_seq_length=max_seq_length,
        image_size=image_size,
        center_crop=center_crop,
        random_flip=random_flip,
    )

from_config classmethod ¤

from_config(config, **kwargs)
Source code in src/unitorch/cli/models/diffusers/processing_qwen_image.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
106
107
108
109
@classmethod
@config_defaults_init("core/process/diffusion/qwen_image")
def from_config(cls, config, **kwargs):
    config.set_default_section("core/process/diffusion/qwen_image")
    pretrained_name = config.getoption("pretrained_name", "qwen-image")
    pretrained_infos = nested_dict_value(pretrained_stable_infos, pretrained_name)

    vocab_path = config.getoption("vocab_path", None)
    vocab_path = pop_value(
        vocab_path,
        nested_dict_value(pretrained_infos, "text", "vocab"),
    )
    vocab_path = cached_path(vocab_path)

    merge_path = config.getoption("merge_path", None)
    merge_path = pop_value(
        merge_path,
        nested_dict_value(pretrained_infos, "text", "merge"),
    )
    merge_path = cached_path(merge_path)

    tokenizer_config = config.getoption("tokenizer_config", None)
    tokenizer_config = pop_value(
        tokenizer_config,
        nested_dict_value(pretrained_infos, "text", "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_infos, "text", "special_tokens_map"),
        check_none=False,
    )
    special_tokens_map = (
        cached_path(special_tokens_map) if special_tokens_map is not None else None
    )

    vision_config_path = config.getoption("vision_config_path", None)
    vision_config_path = pop_value(
        vision_config_path,
        nested_dict_value(pretrained_infos, "vision_config"),
        check_none=False,
    )
    if vision_config_path is not None:
        vision_config_path = cached_path(vision_config_path)

    vae_config_path = config.getoption("vae_config_path", None)
    vae_config_path = pop_value(
        vae_config_path,
        nested_dict_value(pretrained_infos, "vae", "config"),
    )
    vae_config_path = cached_path(vae_config_path)
    return {
        "vocab_path": vocab_path,
        "merge_path": merge_path,
        "vision_config_path": vision_config_path,
        "vae_config_path": vae_config_path,
        "tokenizer_config": tokenizer_config,
        "special_tokens_map": special_tokens_map,
    }

_text2image ¤

_text2image(
    prompt: str,
    image: Union[Image, str],
    max_seq_length: Optional[int] = None,
)
Source code in src/unitorch/cli/models/diffusers/processing_qwen_image.py 
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
@register_process("core/process/diffusion/qwen_image/text2image")
def _text2image(
    self,
    prompt: str,
    image: Union[Image.Image, str],
    max_seq_length: Optional[int] = None,
):
    outputs = super().text2image(
        prompt=prompt,
        image=image,
        max_seq_length=max_seq_length,
    )
    return TensorInputs(
        input_ids=outputs.input_ids,
        attention_mask=outputs.attention_mask,
        pixel_values=outputs.pixel_values,
    )

_text2image_inputs ¤

_text2image_inputs(
    prompt: str,
    negative_prompt: Optional[str] = "",
    max_seq_length: Optional[int] = None,
)
Source code in src/unitorch/cli/models/diffusers/processing_qwen_image.py 
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
@register_process("core/process/diffusion/qwen_image/text2image/inputs")
def _text2image_inputs(
    self,
    prompt: str,
    negative_prompt: Optional[str] = "",
    max_seq_length: Optional[int] = None,
):
    outputs = super().text2image_inputs(
        prompt=prompt,
        negative_prompt=negative_prompt,
        max_seq_length=max_seq_length,
    )
    return TensorInputs(
        input_ids=outputs.input_ids,
        attention_mask=outputs.attention_mask,
        negative_input_ids=outputs.negative_input_ids,
        negative_attention_mask=outputs.negative_attention_mask,
    )

_editing_inputs ¤

_editing_inputs(
    prompt: str,
    refer_image: Union[Image, str],
    negative_prompt: Optional[str] = "",
    max_seq_length: Optional[int] = None,
)
Source code in src/unitorch/cli/models/diffusers/processing_qwen_image.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
@register_process("core/process/diffusion/qwen_image/editing/inputs")
def _editing_inputs(
    self,
    prompt: str,
    refer_image: Union[Image.Image, str],
    negative_prompt: Optional[str] = "",
    max_seq_length: Optional[int] = None,
):
    outputs = super().editing_inputs(
        prompt=prompt,
        refer_image=refer_image,
        negative_prompt=negative_prompt,
        max_seq_length=max_seq_length,
    )

    return TensorInputs(
        input_ids=outputs.input_ids,
        attention_mask=outputs.attention_mask,
        refer_pixel_values=outputs.refer_pixel_values,
        refer_image_grid_thw=outputs.refer_image_grid_thw,
        refer_vae_pixel_values=outputs.refer_vae_pixel_values,
        negative_input_ids=outputs.negative_input_ids,
        negative_attention_mask=outputs.negative_attention_mask,
    )

_editing ¤

_editing(
    prompt: str,
    refer_image: Union[Image, str],
    image: Union[Image, str],
    max_seq_length: Optional[int] = None,
)
Source code in src/unitorch/cli/models/diffusers/processing_qwen_image.py 
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
@register_process("core/process/diffusion/qwen_image/editing")
def _editing(
    self,
    prompt: str,
    refer_image: Union[Image.Image, str],
    image: Union[Image.Image, str],
    max_seq_length: Optional[int] = None,
):
    outputs = super().editing(
        prompt=prompt,
        refer_image=refer_image,
        image=image,
        max_seq_length=max_seq_length,
    )
    return TensorInputs(
        input_ids=outputs.input_ids,
        attention_mask=outputs.attention_mask,
        pixel_values=outputs.pixel_values,
        refer_image_grid_thw=outputs.refer_image_grid_thw,
        refer_pixel_values=outputs.refer_pixel_values,
        refer_vae_pixel_values=outputs.refer_vae_pixel_values,
    )

QWenImageText2ImageGeneration¤

Tip

core/model/diffusers/text2image/qwen_image is the section for configuration of QWenImageText2ImageGeneration.

Bases: QWenImageText2ImageGeneration

Source code in src/unitorch/cli/models/diffusers/modeling_qwen_image.py 
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
60
61
62
63
64
def __init__(
    self,
    config_path: str,
    text_config_path: str,
    vae_config_path: str,
    scheduler_config_path: str,
    num_train_timesteps: Optional[int] = 1000,
    num_infer_timesteps: Optional[int] = 50,
    freeze_vae_encoder: Optional[bool] = True,
    freeze_text_encoder: Optional[bool] = True,
    snr_gamma: Optional[float] = 5.0,
    seed: Optional[int] = 1123,
    gradient_checkpointing: Optional[bool] = True,
    guidance_scale: Optional[float] = 1.0,
    prompt_start_index: Optional[int] = 34,
):
    super().__init__(
        config_path=config_path,
        text_config_path=text_config_path,
        vae_config_path=vae_config_path,
        scheduler_config_path=scheduler_config_path,
        num_train_timesteps=num_train_timesteps,
        num_infer_timesteps=num_infer_timesteps,
        freeze_vae_encoder=freeze_vae_encoder,
        freeze_text_encoder=freeze_text_encoder,
        snr_gamma=snr_gamma,
        seed=seed,
        gradient_checkpointing=gradient_checkpointing,
        guidance_scale=guidance_scale,
        prompt_start_index=prompt_start_index,
    )

from_config classmethod ¤

from_config(config, **kwargs)
Source code in src/unitorch/cli/models/diffusers/modeling_qwen_image.py 
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
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
@classmethod
@config_defaults_init("core/model/diffusers/text2image/qwen_image")
def from_config(cls, config, **kwargs):
    config.set_default_section("core/model/diffusers/text2image/qwen_image")
    pretrained_name = config.getoption("pretrained_name", "qwen-image")
    pretrained_infos = nested_dict_value(pretrained_stable_infos, pretrained_name)

    config_path = config.getoption("config_path", None)
    config_path = pop_value(
        config_path,
        nested_dict_value(pretrained_infos, "transformer", "config"),
    )
    config_path = cached_path(config_path)

    text_config_path = config.getoption("text_config_path", None)
    text_config_path = pop_value(
        text_config_path,
        nested_dict_value(pretrained_infos, "text", "config"),
    )
    text_config_path = cached_path(text_config_path)

    vae_config_path = config.getoption("vae_config_path", None)
    vae_config_path = pop_value(
        vae_config_path,
        nested_dict_value(pretrained_infos, "vae", "config"),
    )
    vae_config_path = cached_path(vae_config_path)

    scheduler_config_path = config.getoption("scheduler_config_path", None)
    scheduler_config_path = pop_value(
        scheduler_config_path,
        nested_dict_value(pretrained_infos, "scheduler"),
    )
    scheduler_config_path = cached_path(scheduler_config_path)

    num_train_timesteps = config.getoption("num_train_timesteps", 1000)
    num_infer_timesteps = config.getoption("num_infer_timesteps", 50)
    freeze_vae_encoder = config.getoption("freeze_vae_encoder", True)
    freeze_text_encoder = config.getoption("freeze_text_encoder", True)
    snr_gamma = config.getoption("snr_gamma", 5.0)
    seed = config.getoption("seed", 1123)
    gradient_checkpointing = config.getoption("gradient_checkpointing", True)
    guidance_scale = config.getoption("guidance_scale", 1.0)
    prompt_start_index = config.getoption("prompt_start_index", 34)

    inst = cls(
        config_path=config_path,
        text_config_path=text_config_path,
        vae_config_path=vae_config_path,
        scheduler_config_path=scheduler_config_path,
        num_train_timesteps=num_train_timesteps,
        num_infer_timesteps=num_infer_timesteps,
        freeze_vae_encoder=freeze_vae_encoder,
        freeze_text_encoder=freeze_text_encoder,
        snr_gamma=snr_gamma,
        seed=seed,
        gradient_checkpointing=gradient_checkpointing,
        guidance_scale=guidance_scale,
        prompt_start_index=prompt_start_index,
    )

    weight_path = config.getoption("pretrained_weight_path", None)

    state_dict = None
    if weight_path is None and pretrained_infos is not None:
        state_dict = [
            load_weight(
                nested_dict_value(pretrained_infos, "transformer", "weight"),
                prefix_keys={"": "transformer."},
            ),
            load_weight(
                nested_dict_value(pretrained_infos, "text", "weight"),
                prefix_keys={
                    "^visual.": "text.model.",
                    "^model(?!\.model).": "text.model.language_",
                    "^lm_head.": "text.",
                },
            ),
            load_weight(
                nested_dict_value(pretrained_infos, "vae", "weight"),
                prefix_keys={"": "vae."},
            ),
        ]

    inst.from_pretrained(weight_path, state_dict=state_dict)

    pretrained_lora_names = config.getoption("pretrained_lora_names", None)
    pretrained_lora_weights = config.getoption("pretrained_lora_weights", 1.0)
    pretrained_lora_alphas = config.getoption("pretrained_lora_alphas", 32.0)

    if isinstance(pretrained_lora_names, str):
        pretrained_lora_weights_path = nested_dict_value(
            pretrained_stable_extensions_infos,
            pretrained_lora_names,
            "lora",
            "weight",
        )
    elif isinstance(pretrained_lora_names, list):
        pretrained_lora_weights_path = [
            nested_dict_value(
                pretrained_stable_extensions_infos, name, "lora", "weight"
            )
            for name in pretrained_lora_names
        ]
    else:
        pretrained_lora_weights_path = None

    lora_weights_path = config.getoption(
        "pretrained_lora_weights_path", pretrained_lora_weights_path
    )
    if lora_weights_path is not None:
        inst.load_lora_weights(
            lora_files=lora_weights_path,
            lora_weights=pretrained_lora_weights,
            lora_alphas=pretrained_lora_alphas,
            replace_keys={},
            save_base_state=False,
        )
    return inst

forward ¤

forward(
    input_ids: Tensor,
    pixel_values: Tensor,
    attention_mask: Optional[Tensor] = None,
)
Source code in src/unitorch/cli/models/diffusers/modeling_qwen_image.py 
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
@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: torch.Tensor,
    attention_mask: Optional[torch.Tensor] = None,
):
    loss = super().forward(
        input_ids=input_ids,
        pixel_values=pixel_values,
        attention_mask=attention_mask,
    )
    return LossOutputs(loss=loss)

generate ¤

generate(
    input_ids: Tensor,
    negative_input_ids: Optional[Tensor] = None,
    attention_mask: Optional[Tensor] = None,
    negative_attention_mask: Optional[Tensor] = None,
    height: Optional[int] = 1024,
    width: Optional[int] = 1024,
    guidance_scale: Optional[float] = 4.0,
)
Source code in src/unitorch/cli/models/diffusers/modeling_qwen_image.py 
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
@config_defaults_method("core/model/diffusers/text2image/qwen_image")
@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,
    negative_input_ids: Optional[torch.Tensor] = None,
    attention_mask: Optional[torch.Tensor] = None,
    negative_attention_mask: Optional[torch.Tensor] = None,
    height: Optional[int] = 1024,
    width: Optional[int] = 1024,
    guidance_scale: Optional[float] = 4.0,
):
    outputs = super().generate(
        input_ids=input_ids,
        negative_input_ids=negative_input_ids,
        attention_mask=attention_mask,
        negative_attention_mask=negative_attention_mask,
        height=height,
        width=width,
        true_cfg_scale=guidance_scale,
    )

    return DiffusionOutputs(outputs=outputs.images)

QWenImageEditingGeneration¤

Tip

core/model/diffusers/editing/qwen_image is the section for configuration of QWenImageEditingGeneration.

Bases: QWenImageEditingGeneration

Source code in src/unitorch/cli/models/diffusers/modeling_qwen_image.py 
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
def __init__(
    self,
    config_path: str,
    text_config_path: str,
    vae_config_path: str,
    scheduler_config_path: str,
    num_train_timesteps: Optional[int] = 1000,
    num_infer_timesteps: Optional[int] = 50,
    freeze_vae_encoder: Optional[bool] = True,
    freeze_text_encoder: Optional[bool] = True,
    snr_gamma: Optional[float] = 5.0,
    seed: Optional[int] = 1123,
    gradient_checkpointing: Optional[bool] = True,
    guidance_scale: Optional[float] = 1.0,
    prompt_start_index: Optional[int] = 64,
):
    super().__init__(
        config_path=config_path,
        text_config_path=text_config_path,
        vae_config_path=vae_config_path,
        scheduler_config_path=scheduler_config_path,
        num_train_timesteps=num_train_timesteps,
        num_infer_timesteps=num_infer_timesteps,
        freeze_vae_encoder=freeze_vae_encoder,
        freeze_text_encoder=freeze_text_encoder,
        snr_gamma=snr_gamma,
        seed=seed,
        gradient_checkpointing=gradient_checkpointing,
        guidance_scale=guidance_scale,
        prompt_start_index=prompt_start_index,
    )

from_config classmethod ¤

from_config(config, **kwargs)
Source code in src/unitorch/cli/models/diffusers/modeling_qwen_image.py 
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
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
@classmethod
@config_defaults_init("core/model/diffusers/editing/qwen_image")
def from_config(cls, config, **kwargs):
    config.set_default_section("core/model/diffusers/editing/qwen_image")
    pretrained_name = config.getoption("pretrained_name", "qwen-image-editing")
    pretrained_infos = nested_dict_value(pretrained_stable_infos, pretrained_name)

    config_path = config.getoption("config_path", None)
    config_path = pop_value(
        config_path,
        nested_dict_value(pretrained_infos, "transformer", "config"),
    )
    config_path = cached_path(config_path)

    text_config_path = config.getoption("text_config_path", None)
    text_config_path = pop_value(
        text_config_path,
        nested_dict_value(pretrained_infos, "text", "config"),
    )
    text_config_path = cached_path(text_config_path)

    vae_config_path = config.getoption("vae_config_path", None)
    vae_config_path = pop_value(
        vae_config_path,
        nested_dict_value(pretrained_infos, "vae", "config"),
    )
    vae_config_path = cached_path(vae_config_path)

    scheduler_config_path = config.getoption("scheduler_config_path", None)
    scheduler_config_path = pop_value(
        scheduler_config_path,
        nested_dict_value(pretrained_infos, "scheduler"),
    )
    scheduler_config_path = cached_path(scheduler_config_path)

    num_train_timesteps = config.getoption("num_train_timesteps", 1000)
    num_infer_timesteps = config.getoption("num_infer_timesteps", 50)
    freeze_vae_encoder = config.getoption("freeze_vae_encoder", True)
    freeze_text_encoder = config.getoption("freeze_text_encoder", True)
    snr_gamma = config.getoption("snr_gamma", 5.0)
    seed = config.getoption("seed", 1123)
    gradient_checkpointing = config.getoption("gradient_checkpointing", True)
    guidance_scale = config.getoption("guidance_scale", 1.0)
    prompt_start_index = config.getoption("prompt_start_index", 64)

    inst = cls(
        config_path=config_path,
        text_config_path=text_config_path,
        vae_config_path=vae_config_path,
        scheduler_config_path=scheduler_config_path,
        num_train_timesteps=num_train_timesteps,
        num_infer_timesteps=num_infer_timesteps,
        freeze_vae_encoder=freeze_vae_encoder,
        freeze_text_encoder=freeze_text_encoder,
        snr_gamma=snr_gamma,
        seed=seed,
        gradient_checkpointing=gradient_checkpointing,
        guidance_scale=guidance_scale,
        prompt_start_index=prompt_start_index,
    )

    weight_path = config.getoption("pretrained_weight_path", None)

    state_dict = None
    if weight_path is None and pretrained_infos is not None:
        state_dict = [
            load_weight(
                nested_dict_value(pretrained_infos, "transformer", "weight"),
                prefix_keys={"": "transformer."},
            ),
            load_weight(
                nested_dict_value(pretrained_infos, "text", "weight"),
                prefix_keys={
                    "^visual.": "text.model.",
                    "^model(?!\.model).": "text.model.language_",
                    "^lm_head.": "text.",
                },
            ),
            load_weight(
                nested_dict_value(pretrained_infos, "vae", "weight"),
                prefix_keys={"": "vae."},
            ),
        ]

    elif weight_path is not None:
        state_dict = load_weight(weight_path)

    if state_dict is not None:
        inst.from_pretrained(state_dict=state_dict)

    pretrained_lora_names = config.getoption("pretrained_lora_names", None)
    pretrained_lora_weights = config.getoption("pretrained_lora_weights", 1.0)
    pretrained_lora_alphas = config.getoption("pretrained_lora_alphas", 32.0)

    if isinstance(pretrained_lora_names, str):
        pretrained_lora_weights_path = nested_dict_value(
            pretrained_stable_extensions_infos,
            pretrained_lora_names,
            "lora",
            "weight",
        )
    elif isinstance(pretrained_lora_names, list):
        pretrained_lora_weights_path = [
            nested_dict_value(
                pretrained_stable_extensions_infos, name, "lora", "weight"
            )
            for name in pretrained_lora_names
        ]
    else:
        pretrained_lora_weights_path = None

    lora_weights_path = config.getoption(
        "pretrained_lora_weights_path", pretrained_lora_weights_path
    )
    if lora_weights_path is not None:
        inst.load_lora_weights(
            lora_files=lora_weights_path,
            lora_weights=pretrained_lora_weights,
            lora_alphas=pretrained_lora_alphas,
            replace_keys={},
            save_base_state=False,
        )
    return inst

forward ¤

forward(
    input_ids: Tensor,
    pixel_values: Tensor,
    refer_pixel_values: Tensor,
    refer_image_grid_thw: Tensor,
    refer_vae_pixel_values: Tensor,
    attention_mask: Optional[Tensor] = None,
)
Source code in src/unitorch/cli/models/diffusers/modeling_qwen_image.py 
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
@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: torch.Tensor,
    refer_pixel_values: torch.Tensor,
    refer_image_grid_thw: torch.Tensor,
    refer_vae_pixel_values: torch.Tensor,
    attention_mask: Optional[torch.Tensor] = None,
):
    loss = super().forward(
        input_ids=input_ids,
        pixel_values=pixel_values,
        refer_pixel_values=refer_pixel_values,
        refer_image_grid_thw=refer_image_grid_thw,
        refer_vae_pixel_values=refer_vae_pixel_values,
        attention_mask=attention_mask,
    )
    return LossOutputs(loss=loss)

generate ¤

generate(
    input_ids: Tensor,
    negative_input_ids: Tensor,
    refer_pixel_values: Tensor,
    refer_image_grid_thw: Tensor,
    refer_vae_pixel_values: Tensor,
    attention_mask: Optional[Tensor] = None,
    negative_attention_mask: Optional[Tensor] = None,
    height: Optional[int] = 1024,
    width: Optional[int] = 1024,
    guidance_scale: Optional[float] = 1.0,
    true_cfg_scale: Optional[float] = 4.0,
)
Source code in src/unitorch/cli/models/diffusers/modeling_qwen_image.py 
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
@config_defaults_method("core/model/diffusers/editing/qwen_image")
@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,
    negative_input_ids: torch.Tensor,
    refer_pixel_values: torch.Tensor,
    refer_image_grid_thw: torch.Tensor,
    refer_vae_pixel_values: torch.Tensor,
    attention_mask: Optional[torch.Tensor] = None,
    negative_attention_mask: Optional[torch.Tensor] = None,
    height: Optional[int] = 1024,
    width: Optional[int] = 1024,
    guidance_scale: Optional[float] = 1.0,
    true_cfg_scale: Optional[float] = 4.0,
):
    outputs = super().generate(
        input_ids=input_ids,
        negative_input_ids=negative_input_ids,
        refer_pixel_values=refer_pixel_values,
        refer_image_grid_thw=refer_image_grid_thw,
        refer_vae_pixel_values=refer_vae_pixel_values,
        attention_mask=attention_mask,
        negative_attention_mask=negative_attention_mask,
        height=height,
        width=width,
        guidance_scale=guidance_scale,
        true_cfg_scale=true_cfg_scale,
    )

    return DiffusionOutputs(outputs=outputs.images)