gradio_demo.py 51.9 KB
Newer Older
gushiqiao's avatar
gushiqiao committed
1
2
3
4
5
6
7
8
9
10
import os
import gradio as gr
import argparse
import json
import torch
import gc
from easydict import EasyDict
from datetime import datetime
from loguru import logger

gushiqiao's avatar
gushiqiao committed
11
12
import importlib.util
import psutil
gushiqiao's avatar
gushiqiao committed
13
import random
gushiqiao's avatar
gushiqiao committed
14
import glob
gushiqiao's avatar
gushiqiao committed
15
16
17
18
19
20
21
22
23
24

logger.add(
    "inference_logs.log",
    rotation="100 MB",
    encoding="utf-8",
    enqueue=True,
    backtrace=True,
    diagnose=True,
)

gushiqiao's avatar
gushiqiao committed
25
26
27
MAX_NUMPY_SEED = 2**32 - 1


gushiqiao's avatar
gushiqiao committed
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
59
60
61
def find_hf_model_path(model_path, subdir=["original", "fp8", "int8"]):
    paths_to_check = [model_path]
    if isinstance(subdir, list):
        for sub in subdir:
            paths_to_check.append(os.path.join(model_path, sub))
    else:
        paths_to_check.append(os.path.join(model_path, subdir))

    for path in paths_to_check:
        safetensors_pattern = os.path.join(path, "*.safetensors")
        safetensors_files = glob.glob(safetensors_pattern)
        if safetensors_files:
            logger.info(f"Found Hugging Face model files in: {path}")
            return path
    raise FileNotFoundError(f"No Hugging Face model files (.safetensors) found.\nPlease download the model from: https://huggingface.co/lightx2v/ or specify the model path in the configuration file.")


def find_torch_model_path(model_path, filename=None, subdir=["original", "fp8", "int8"]):
    paths_to_check = [
        os.path.join(model_path, filename),
    ]
    if isinstance(subdir, list):
        for sub in subdir:
            paths_to_check.append(os.path.join(model_path, sub, filename))
    else:
        paths_to_check.append(os.path.join(model_path, subdir, filename))
    print(paths_to_check)
    for path in paths_to_check:
        if os.path.exists(path):
            logger.info(f"Found PyTorch model checkpoint: {path}")
            return path
    raise FileNotFoundError(f"PyTorch model file '{filename}' not found.\nPlease download the model from https://huggingface.co/lightx2v/ or specify the model path in the configuration file.")


gushiqiao's avatar
gushiqiao committed
62
63
64
def generate_random_seed():
    return random.randint(0, MAX_NUMPY_SEED)

gushiqiao's avatar
gushiqiao committed
65

gushiqiao's avatar
gushiqiao committed
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
def is_module_installed(module_name):
    try:
        spec = importlib.util.find_spec(module_name)
        return spec is not None
    except ModuleNotFoundError:
        return False


def get_available_quant_ops():
    available_ops = []

    vllm_installed = is_module_installed("vllm")
    if vllm_installed:
        available_ops.append(("vllm", True))
    else:
        available_ops.append(("vllm", False))

    sgl_installed = is_module_installed("sgl_kernel")
    if sgl_installed:
        available_ops.append(("sgl", True))
    else:
        available_ops.append(("sgl", False))

    q8f_installed = is_module_installed("q8_kernels")
    if q8f_installed:
        available_ops.append(("q8f", True))
    else:
        available_ops.append(("q8f", False))

    return available_ops


def get_available_attn_ops():
    available_ops = []

    vllm_installed = is_module_installed("flash_attn")
    if vllm_installed:
        available_ops.append(("flash_attn2", True))
    else:
        available_ops.append(("flash_attn2", False))

    sgl_installed = is_module_installed("flash_attn_interface")
    if sgl_installed:
        available_ops.append(("flash_attn3", True))
    else:
        available_ops.append(("flash_attn3", False))

    q8f_installed = is_module_installed("sageattention")
    if q8f_installed:
        available_ops.append(("sage_attn2", True))
    else:
        available_ops.append(("sage_attn2", False))

gushiqiao's avatar
gushiqiao committed
119
120
121
122
123
124
    torch_installed = is_module_installed("torch")
    if torch_installed:
        available_ops.append(("torch_sdpa", True))
    else:
        available_ops.append(("torch_sdpa", False))

gushiqiao's avatar
gushiqiao committed
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
    return available_ops


def get_gpu_memory(gpu_idx=0):
    if not torch.cuda.is_available():
        return 0
    try:
        with torch.cuda.device(gpu_idx):
            memory_info = torch.cuda.mem_get_info()
            total_memory = memory_info[1] / (1024**3)  # Convert bytes to GB
            return total_memory
    except Exception as e:
        logger.warning(f"Failed to get GPU memory: {e}")
        return 0


def get_cpu_memory():
    available_bytes = psutil.virtual_memory().available
    return available_bytes / 1024**3
gushiqiao's avatar
gushiqiao committed
144
145


gushiqiao's avatar
gushiqiao committed
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
def cleanup_memory():
    gc.collect()

    if torch.cuda.is_available():
        torch.cuda.empty_cache()
        torch.cuda.synchronize()

    try:
        if hasattr(psutil, "virtual_memory"):
            if os.name == "posix":
                try:
                    os.system("sync")
                except:  # noqa
                    pass
    except:  # noqa
        pass


gushiqiao's avatar
gushiqiao committed
164
165
def generate_unique_filename(output_dir):
    os.makedirs(output_dir, exist_ok=True)
gushiqiao's avatar
gushiqiao committed
166
    timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
gushiqiao's avatar
gushiqiao committed
167
    return os.path.join(output_dir, f"{model_cls}_{timestamp}.mp4")
gushiqiao's avatar
gushiqiao committed
168
169


gushiqiao's avatar
gushiqiao committed
170
171
172
173
174
175
176
177
def is_fp8_supported_gpu():
    if not torch.cuda.is_available():
        return False
    compute_capability = torch.cuda.get_device_capability(0)
    major, minor = compute_capability
    return (major == 8 and minor == 9) or (major >= 9)


gushiqiao's avatar
gushiqiao committed
178
179
180
181
182
183
184
185
186
187
188
189
def is_ada_architecture_gpu():
    if not torch.cuda.is_available():
        return False
    try:
        gpu_name = torch.cuda.get_device_name(0).upper()
        ada_keywords = ["RTX 40", "RTX40", "4090", "4080", "4070", "4060"]
        return any(keyword in gpu_name for keyword in ada_keywords)
    except Exception as e:
        logger.warning(f"Failed to get GPU name: {e}")
        return False


gushiqiao's avatar
gushiqiao committed
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
def get_quantization_options(model_path):
    """Get quantization options dynamically based on model_path"""
    import os

    # Check subdirectories
    subdirs = ["original", "fp8", "int8"]
    has_subdirs = {subdir: os.path.exists(os.path.join(model_path, subdir)) for subdir in subdirs}

    # Check original files in root directory
    t5_bf16_exists = os.path.exists(os.path.join(model_path, "models_t5_umt5-xxl-enc-bf16.pth"))
    clip_fp16_exists = os.path.exists(os.path.join(model_path, "models_clip_open-clip-xlm-roberta-large-vit-huge-14.pth"))

    # Generate options
    def get_choices(has_subdirs, original_type, fp8_type, int8_type, fallback_type, has_original_file=False):
        choices = []
        if has_subdirs["original"]:
            choices.append(original_type)
        if has_subdirs["fp8"]:
            choices.append(fp8_type)
        if has_subdirs["int8"]:
            choices.append(int8_type)

        # If no subdirectories but original file exists, add original type
        if not choices and has_original_file:
            choices.append(original_type)

        # If no options at all, use default value
        if not choices:
            choices = [fallback_type]

        return choices, choices[0]

    # DIT options
    dit_choices, dit_default = get_choices(has_subdirs, "bf16", "fp8", "int8", "bf16")

    # T5 options - check if original file exists
    t5_choices, t5_default = get_choices(has_subdirs, "bf16", "fp8", "int8", "bf16", t5_bf16_exists)

    # CLIP options - check if original file exists
    clip_choices, clip_default = get_choices(has_subdirs, "fp16", "fp8", "int8", "fp16", clip_fp16_exists)

    return {"dit_choices": dit_choices, "dit_default": dit_default, "t5_choices": t5_choices, "t5_default": t5_default, "clip_choices": clip_choices, "clip_default": clip_default}


gushiqiao's avatar
gushiqiao committed
234
235
global_runner = None
current_config = None
gushiqiao's avatar
gushiqiao committed
236
237
238
239
240
cur_dit_quant_scheme = None
cur_clip_quant_scheme = None
cur_t5_quant_scheme = None
cur_precision_mode = None
cur_enable_teacache = None
gushiqiao's avatar
gushiqiao committed
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256

available_quant_ops = get_available_quant_ops()
quant_op_choices = []
for op_name, is_installed in available_quant_ops:
    status_text = "✅ Installed" if is_installed else "❌ Not Installed"
    display_text = f"{op_name} ({status_text})"
    quant_op_choices.append((op_name, display_text))

available_attn_ops = get_available_attn_ops()
attn_op_choices = []
for op_name, is_installed in available_attn_ops:
    status_text = "✅ Installed" if is_installed else "❌ Not Installed"
    display_text = f"{op_name} ({status_text})"
    attn_op_choices.append((op_name, display_text))


gushiqiao's avatar
gushiqiao committed
257
258
259
260
261
262
263
264
265
266
267
268
def run_inference(
    prompt,
    negative_prompt,
    save_video_path,
    torch_compile,
    infer_steps,
    num_frames,
    resolution,
    seed,
    sample_shift,
    enable_teacache,
    teacache_thresh,
gushiqiao's avatar
gushiqiao committed
269
    use_ret_steps,
gushiqiao's avatar
gushiqiao committed
270
271
272
273
274
275
276
277
278
279
280
281
    enable_cfg,
    cfg_scale,
    dit_quant_scheme,
    t5_quant_scheme,
    clip_quant_scheme,
    fps,
    use_tiny_vae,
    use_tiling_vae,
    lazy_load,
    precision_mode,
    cpu_offload,
    offload_granularity,
gushiqiao's avatar
gushiqiao committed
282
    offload_ratio,
gushiqiao's avatar
gushiqiao committed
283
284
    t5_cpu_offload,
    unload_modules,
gushiqiao's avatar
gushiqiao committed
285
286
287
288
    t5_offload_granularity,
    attention_type,
    quant_op,
    rotary_chunk,
gushiqiao's avatar
gushiqiao committed
289
    rotary_chunk_size,
gushiqiao's avatar
gushiqiao committed
290
    clean_cuda_cache,
gushiqiao's avatar
gushiqiao committed
291
    image_path=None,
gushiqiao's avatar
gushiqiao committed
292
):
gushiqiao's avatar
gushiqiao committed
293
294
    cleanup_memory()

gushiqiao's avatar
gushiqiao committed
295
296
297
    quant_op = quant_op.split("(")[0].strip()
    attention_type = attention_type.split("(")[0].strip()

gushiqiao's avatar
gushiqiao committed
298
    global global_runner, current_config, model_path, task
gushiqiao's avatar
gushiqiao committed
299
    global cur_dit_quant_scheme, cur_clip_quant_scheme, cur_t5_quant_scheme, cur_precision_mode, cur_enable_teacache
gushiqiao's avatar
gushiqiao committed
300
301
302
303

    if os.path.exists(os.path.join(model_path, "config.json")):
        with open(os.path.join(model_path, "config.json"), "r") as f:
            model_config = json.load(f)
gushiqiao's avatar
gushiqiao committed
304
305
    else:
        model_config = {}
gushiqiao's avatar
gushiqiao committed
306
307

    if task == "t2v":
gushiqiao's avatar
gushiqiao committed
308
        if model_size == "1.3b":
gushiqiao's avatar
gushiqiao committed
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
            # 1.3B
            coefficient = [
                [
                    -5.21862437e04,
                    9.23041404e03,
                    -5.28275948e02,
                    1.36987616e01,
                    -4.99875664e-02,
                ],
                [
                    2.39676752e03,
                    -1.31110545e03,
                    2.01331979e02,
                    -8.29855975e00,
                    1.37887774e-01,
                ],
            ]
        else:
            # 14B
            coefficient = [
                [
                    -3.03318725e05,
                    4.90537029e04,
                    -2.65530556e03,
                    5.87365115e01,
                    -3.15583525e-01,
                ],
                [
                    -5784.54975374,
                    5449.50911966,
                    -1811.16591783,
                    256.27178429,
                    -13.02252404,
                ],
            ]
    elif task == "i2v":
        if resolution in [
            "1280x720",
            "720x1280",
            "1280x544",
            "544x1280",
            "1104x832",
            "832x1104",
            "960x960",
        ]:
            # 720p
            coefficient = [
                [
                    8.10705460e03,
                    2.13393892e03,
                    -3.72934672e02,
                    1.66203073e01,
                    -4.17769401e-02,
                ],
                [-114.36346466, 65.26524496, -18.82220707, 4.91518089, -0.23412683],
            ]
        else:
            # 480p
            coefficient = [
                [
                    2.57151496e05,
                    -3.54229917e04,
                    1.40286849e03,
                    -1.35890334e01,
                    1.32517977e-01,
                ],
                [
                    -3.02331670e02,
                    2.23948934e02,
                    -5.25463970e01,
                    5.87348440e00,
                    -2.01973289e-01,
                ],
            ]

gushiqiao's avatar
gushiqiao committed
384
    save_video_path = generate_unique_filename(output_dir)
gushiqiao's avatar
gushiqiao committed
385
386
387
388

    is_dit_quant = dit_quant_scheme != "bf16"
    is_t5_quant = t5_quant_scheme != "bf16"
    if is_t5_quant:
gushiqiao's avatar
gushiqiao committed
389
390
391
        t5_model_name = f"models_t5_umt5-xxl-enc-{t5_quant_scheme}.pth"
        t5_quant_ckpt = find_torch_model_path(model_path, t5_model_name, t5_quant_scheme)
        t5_original_ckpt = None
gushiqiao's avatar
gushiqiao committed
392
393
    else:
        t5_quant_ckpt = None
gushiqiao's avatar
gushiqiao committed
394
395
        t5_model_name = "models_t5_umt5-xxl-enc-bf16.pth"
        t5_original_ckpt = find_torch_model_path(model_path, t5_model_name, "original")
gushiqiao's avatar
gushiqiao committed
396

gushiqiao's avatar
gushiqiao committed
397
    is_clip_quant = clip_quant_scheme != "fp16"
gushiqiao's avatar
gushiqiao committed
398
    if is_clip_quant:
gushiqiao's avatar
gushiqiao committed
399
400
401
        clip_model_name = f"clip-{clip_quant_scheme}.pth"
        clip_quant_ckpt = find_torch_model_path(model_path, clip_model_name, clip_quant_scheme)
        clip_original_ckpt = None
gushiqiao's avatar
gushiqiao committed
402
403
    else:
        clip_quant_ckpt = None
gushiqiao's avatar
gushiqiao committed
404
405
        clip_model_name = "models_clip_open-clip-xlm-roberta-large-vit-huge-14.pth"
        clip_original_ckpt = find_torch_model_path(model_path, clip_model_name, "original")
gushiqiao's avatar
gushiqiao committed
406

gushiqiao's avatar
gushiqiao committed
407
408
    needs_reinit = (
        lazy_load
gushiqiao's avatar
gushiqiao committed
409
        or unload_modules
gushiqiao's avatar
gushiqiao committed
410
411
412
413
414
415
416
417
418
419
420
421
422
        or global_runner is None
        or current_config is None
        or cur_dit_quant_scheme is None
        or cur_dit_quant_scheme != dit_quant_scheme
        or cur_clip_quant_scheme is None
        or cur_clip_quant_scheme != clip_quant_scheme
        or cur_t5_quant_scheme is None
        or cur_t5_quant_scheme != t5_quant_scheme
        or cur_precision_mode is None
        or cur_precision_mode != precision_mode
        or cur_enable_teacache is None
        or cur_enable_teacache != enable_teacache
    )
gushiqiao's avatar
gushiqiao committed
423
424
425
426
427
428
429
430
431
432
433
434
435
436

    if torch_compile:
        os.environ["ENABLE_GRAPH_MODE"] = "true"
    else:
        os.environ["ENABLE_GRAPH_MODE"] = "false"
    if precision_mode == "bf16":
        os.environ["DTYPE"] = "BF16"
    else:
        os.environ.pop("DTYPE", None)

    if is_dit_quant:
        if quant_op == "vllm":
            mm_type = f"W-{dit_quant_scheme}-channel-sym-A-{dit_quant_scheme}-channel-sym-dynamic-Vllm"
        elif quant_op == "sgl":
gushiqiao's avatar
gushiqiao committed
437
438
439
440
            if dit_quant_scheme == "int8":
                mm_type = f"W-{dit_quant_scheme}-channel-sym-A-{dit_quant_scheme}-channel-sym-dynamic-Sgl-ActVllm"
            else:
                mm_type = f"W-{dit_quant_scheme}-channel-sym-A-{dit_quant_scheme}-channel-sym-dynamic-Sgl"
gushiqiao's avatar
gushiqiao committed
441
442
        elif quant_op == "q8f":
            mm_type = f"W-{dit_quant_scheme}-channel-sym-A-{dit_quant_scheme}-channel-sym-dynamic-Q8F"
gushiqiao's avatar
Fix  
gushiqiao committed
443
444
            t5_quant_scheme = f"{t5_quant_scheme}-q8f"
            clip_quant_scheme = f"{clip_quant_scheme}-q8f"
gushiqiao's avatar
gushiqiao committed
445

gushiqiao's avatar
gushiqiao committed
446
        dit_quantized_ckpt = find_hf_model_path(model_path, dit_quant_scheme)
gushiqiao's avatar
gushiqiao committed
447
448
449
        if os.path.exists(os.path.join(dit_quantized_ckpt, "config.json")):
            with open(os.path.join(dit_quantized_ckpt, "config.json"), "r") as f:
                quant_model_config = json.load(f)
gushiqiao's avatar
gushiqiao committed
450
451
        else:
            quant_model_config = {}
gushiqiao's avatar
gushiqiao committed
452
453
    else:
        mm_type = "Default"
gushiqiao's avatar
gushiqiao committed
454
        dit_quantized_ckpt = None
gushiqiao's avatar
gushiqiao committed
455
        quant_model_config = {}
gushiqiao's avatar
gushiqiao committed
456
457
458
459
460
461

    config = {
        "infer_steps": infer_steps,
        "target_video_length": num_frames,
        "target_width": int(resolution.split("x")[0]),
        "target_height": int(resolution.split("x")[1]),
gushiqiao's avatar
gushiqiao committed
462
463
464
        "self_attn_1_type": attention_type,
        "cross_attn_1_type": attention_type,
        "cross_attn_2_type": attention_type,
gushiqiao's avatar
gushiqiao committed
465
466
467
468
469
470
        "seed": seed,
        "enable_cfg": enable_cfg,
        "sample_guide_scale": cfg_scale,
        "sample_shift": sample_shift,
        "cpu_offload": cpu_offload,
        "offload_granularity": offload_granularity,
gushiqiao's avatar
gushiqiao committed
471
        "offload_ratio": offload_ratio,
gushiqiao's avatar
gushiqiao committed
472
        "t5_offload_granularity": t5_offload_granularity,
gushiqiao's avatar
gushiqiao committed
473
        "dit_quantized_ckpt": dit_quantized_ckpt,
gushiqiao's avatar
gushiqiao committed
474
475
476
477
478
        "mm_config": {
            "mm_type": mm_type,
        },
        "fps": fps,
        "feature_caching": "Tea" if enable_teacache else "NoCaching",
gushiqiao's avatar
gushiqiao committed
479
480
        "coefficients": coefficient[0] if use_ret_steps else coefficient[1],
        "use_ret_steps": use_ret_steps,
gushiqiao's avatar
gushiqiao committed
481
        "teacache_thresh": teacache_thresh,
gushiqiao's avatar
gushiqiao committed
482
483
        "t5_cpu_offload": t5_cpu_offload,
        "unload_modules": unload_modules,
gushiqiao's avatar
gushiqiao committed
484
        "t5_original_ckpt": t5_original_ckpt,
gushiqiao's avatar
gushiqiao committed
485
486
487
        "t5_quantized": is_t5_quant,
        "t5_quantized_ckpt": t5_quant_ckpt,
        "t5_quant_scheme": t5_quant_scheme,
gushiqiao's avatar
gushiqiao committed
488
        "clip_original_ckpt": clip_original_ckpt,
gushiqiao's avatar
gushiqiao committed
489
490
491
        "clip_quantized": is_clip_quant,
        "clip_quantized_ckpt": clip_quant_ckpt,
        "clip_quant_scheme": clip_quant_scheme,
gushiqiao's avatar
gushiqiao committed
492
        "vae_path": find_torch_model_path(model_path, "Wan2.1_VAE.pth"),
gushiqiao's avatar
gushiqiao committed
493
        "use_tiling_vae": use_tiling_vae,
helloyongyang's avatar
helloyongyang committed
494
        "use_tiny_vae": use_tiny_vae,
gushiqiao's avatar
gushiqiao committed
495
        "tiny_vae_path": (find_torch_model_path(model_path, "taew2_1.pth") if use_tiny_vae else None),
gushiqiao's avatar
gushiqiao committed
496
497
498
499
500
501
502
503
504
505
506
507
        "lazy_load": lazy_load,
        "do_mm_calib": False,
        "parallel_attn_type": None,
        "parallel_vae": False,
        "max_area": False,
        "vae_stride": (4, 8, 8),
        "patch_size": (1, 2, 2),
        "lora_path": None,
        "strength_model": 1.0,
        "use_prompt_enhancer": False,
        "text_len": 512,
        "rotary_chunk": rotary_chunk,
gushiqiao's avatar
gushiqiao committed
508
        "rotary_chunk_size": rotary_chunk_size,
gushiqiao's avatar
gushiqiao committed
509
        "clean_cuda_cache": clean_cuda_cache,
gushiqiao's avatar
gushiqiao committed
510
        "denoising_step_list": [1000, 750, 500, 250],
gushiqiao's avatar
gushiqiao committed
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
    }

    args = argparse.Namespace(
        model_cls=model_cls,
        task=task,
        model_path=model_path,
        prompt_enhancer=None,
        prompt=prompt,
        negative_prompt=negative_prompt,
        image_path=image_path,
        save_video_path=save_video_path,
    )

    config.update({k: v for k, v in vars(args).items()})
    config = EasyDict(config)
    config.update(model_config)
gushiqiao's avatar
gushiqiao committed
527
    config.update(quant_model_config)
gushiqiao's avatar
gushiqiao committed
528
529
530
531

    logger.info(f"Using model: {model_path}")
    logger.info(f"Inference configuration:\n{json.dumps(config, indent=4, ensure_ascii=False)}")

gushiqiao's avatar
gushiqiao committed
532
    # Initialize or reuse the runner
gushiqiao's avatar
gushiqiao committed
533
534
535
536
537
538
539
    runner = global_runner
    if needs_reinit:
        if runner is not None:
            del runner
            torch.cuda.empty_cache()
            gc.collect()

gushiqiao's avatar
gushiqiao committed
540
541
        from lightx2v.infer import init_runner  # noqa

gushiqiao's avatar
gushiqiao committed
542
543
        runner = init_runner(config)
        current_config = config
gushiqiao's avatar
gushiqiao committed
544
545
546
547
548
        cur_dit_quant_scheme = dit_quant_scheme
        cur_clip_quant_scheme = clip_quant_scheme
        cur_t5_quant_scheme = t5_quant_scheme
        cur_precision_mode = precision_mode
        cur_enable_teacache = enable_teacache
gushiqiao's avatar
gushiqiao committed
549
550
551

        if not lazy_load:
            global_runner = runner
gushiqiao's avatar
gushiqiao committed
552
553
    else:
        runner.config = config
gushiqiao's avatar
gushiqiao committed
554

555
    runner.run_pipeline()
gushiqiao's avatar
gushiqiao committed
556

gushiqiao's avatar
gushiqiao committed
557
558
559
560
561
562
563
564
565
    del config, args, model_config, quant_model_config
    if "dit_quantized_ckpt" in locals():
        del dit_quantized_ckpt
    if "t5_quant_ckpt" in locals():
        del t5_quant_ckpt
    if "clip_quant_ckpt" in locals():
        del clip_quant_ckpt

    cleanup_memory()
gushiqiao's avatar
gushiqiao committed
566
567
568
569

    return save_video_path


gushiqiao's avatar
gushiqiao committed
570
571
572
573
574
575
def handle_lazy_load_change(lazy_load_enabled):
    """Handle lazy_load checkbox change to automatically enable unload_modules"""
    return gr.update(value=lazy_load_enabled)


def auto_configure(enable_auto_config, resolution):
gushiqiao's avatar
gushiqiao committed
576
577
578
579
580
581
582
583
584
    default_config = {
        "torch_compile_val": False,
        "lazy_load_val": False,
        "rotary_chunk_val": False,
        "rotary_chunk_size_val": 100,
        "clean_cuda_cache_val": False,
        "cpu_offload_val": False,
        "offload_granularity_val": "block",
        "offload_ratio_val": 1,
gushiqiao's avatar
gushiqiao committed
585
586
        "t5_cpu_offload_val": False,
        "unload_modules_val": False,
gushiqiao's avatar
gushiqiao committed
587
588
589
590
591
592
593
594
595
596
597
598
599
        "t5_offload_granularity_val": "model",
        "attention_type_val": attn_op_choices[0][1],
        "quant_op_val": quant_op_choices[0][1],
        "dit_quant_scheme_val": "bf16",
        "t5_quant_scheme_val": "bf16",
        "clip_quant_scheme_val": "fp16",
        "precision_mode_val": "fp32",
        "use_tiny_vae_val": False,
        "use_tiling_vae_val": False,
        "enable_teacache_val": False,
        "teacache_thresh_val": 0.26,
        "use_ret_steps_val": False,
    }
gushiqiao's avatar
gushiqiao committed
600

gushiqiao's avatar
gushiqiao committed
601
602
603
604
605
606
607
608
609
610
611
    if not enable_auto_config:
        return tuple(gr.update(value=default_config[key]) for key in default_config)

    gpu_memory = round(get_gpu_memory())
    cpu_memory = round(get_cpu_memory())

    if is_fp8_supported_gpu():
        quant_type = "fp8"
    else:
        quant_type = "int8"

gushiqiao's avatar
gushiqiao committed
612
    attn_priority = ["sage_attn2", "flash_attn3", "flash_attn2", "torch_sdpa"]
gushiqiao's avatar
gushiqiao committed
613
614
615
616
617

    if is_ada_architecture_gpu():
        quant_op_priority = ["q8f", "vllm", "sgl"]
    else:
        quant_op_priority = ["sgl", "vllm", "q8f"]
gushiqiao's avatar
gushiqiao committed
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

    for op in attn_priority:
        if dict(available_attn_ops).get(op):
            default_config["attention_type_val"] = dict(attn_op_choices)[op]
            break

    for op in quant_op_priority:
        if dict(available_quant_ops).get(op):
            default_config["quant_op_val"] = dict(quant_op_choices)[op]
            break

    if resolution in [
        "1280x720",
        "720x1280",
        "1280x544",
        "544x1280",
        "1104x832",
        "832x1104",
        "960x960",
    ]:
        res = "720p"
    elif resolution in [
        "960x544",
        "544x960",
    ]:
        res = "540p"
    else:
        res = "480p"

gushiqiao's avatar
gushiqiao committed
647
    if model_size == "14b":
gushiqiao's avatar
gushiqiao committed
648
649
650
651
652
653
654
        is_14b = True
    else:
        is_14b = False

    if res == "720p" and is_14b:
        gpu_rules = [
            (80, {}),
gushiqiao's avatar
gushiqiao committed
655
656
657
            (48, {"cpu_offload_val": True, "offload_ratio_val": 0.5, "t5_cpu_offload_val": True}),
            (40, {"cpu_offload_val": True, "offload_ratio_val": 0.8, "t5_cpu_offload_val": True}),
            (32, {"cpu_offload_val": True, "offload_ratio_val": 1, "t5_cpu_offload_val": True}),
gushiqiao's avatar
gushiqiao committed
658
659
660
661
            (
                24,
                {
                    "cpu_offload_val": True,
gushiqiao's avatar
gushiqiao committed
662
                    "t5_cpu_offload_val": True,
gushiqiao's avatar
gushiqiao committed
663
664
665
666
667
668
669
670
671
672
                    "offload_ratio_val": 1,
                    "t5_offload_granularity_val": "block",
                    "precision_mode_val": "bf16",
                    "use_tiling_vae_val": True,
                },
            ),
            (
                16,
                {
                    "cpu_offload_val": True,
gushiqiao's avatar
gushiqiao committed
673
                    "t5_cpu_offload_val": True,
gushiqiao's avatar
gushiqiao committed
674
675
676
677
678
679
680
681
682
683
684
685
686
                    "offload_ratio_val": 1,
                    "t5_offload_granularity_val": "block",
                    "precision_mode_val": "bf16",
                    "use_tiling_vae_val": True,
                    "offload_granularity_val": "phase",
                    "rotary_chunk_val": True,
                    "rotary_chunk_size_val": 100,
                },
            ),
            (
                12,
                {
                    "cpu_offload_val": True,
gushiqiao's avatar
gushiqiao committed
687
                    "t5_cpu_offload_val": True,
gushiqiao's avatar
gushiqiao committed
688
689
690
691
692
693
694
695
                    "offload_ratio_val": 1,
                    "t5_offload_granularity_val": "block",
                    "precision_mode_val": "bf16",
                    "use_tiling_vae_val": True,
                    "offload_granularity_val": "phase",
                    "rotary_chunk_val": True,
                    "rotary_chunk_size_val": 100,
                    "clean_cuda_cache_val": True,
gushiqiao's avatar
gushiqiao committed
696
                    "use_tiny_vae_val": True,
gushiqiao's avatar
gushiqiao committed
697
698
699
700
701
702
                },
            ),
            (
                8,
                {
                    "cpu_offload_val": True,
gushiqiao's avatar
gushiqiao committed
703
                    "t5_cpu_offload_val": True,
gushiqiao's avatar
gushiqiao committed
704
705
706
707
708
709
710
711
712
713
714
715
                    "offload_ratio_val": 1,
                    "t5_offload_granularity_val": "block",
                    "precision_mode_val": "bf16",
                    "use_tiling_vae_val": True,
                    "offload_granularity_val": "phase",
                    "rotary_chunk_val": True,
                    "rotary_chunk_size_val": 100,
                    "clean_cuda_cache_val": True,
                    "t5_quant_scheme_val": quant_type,
                    "clip_quant_scheme_val": quant_type,
                    "dit_quant_scheme_val": quant_type,
                    "lazy_load_val": True,
gushiqiao's avatar
gushiqiao committed
716
                    "unload_modules_val": True,
gushiqiao's avatar
gushiqiao committed
717
                    "use_tiny_vae_val": True,
gushiqiao's avatar
gushiqiao committed
718
719
720
721
722
723
724
                },
            ),
        ]

    elif is_14b:
        gpu_rules = [
            (80, {}),
gushiqiao's avatar
gushiqiao committed
725
726
727
            (48, {"cpu_offload_val": True, "offload_ratio_val": 0.2, "t5_cpu_offload_val": True}),
            (40, {"cpu_offload_val": True, "offload_ratio_val": 0.5, "t5_cpu_offload_val": True}),
            (24, {"cpu_offload_val": True, "offload_ratio_val": 0.8, "t5_cpu_offload_val": True}),
gushiqiao's avatar
gushiqiao committed
728
729
730
731
            (
                16,
                {
                    "cpu_offload_val": True,
gushiqiao's avatar
gushiqiao committed
732
                    "t5_cpu_offload_val": True,
gushiqiao's avatar
gushiqiao committed
733
734
735
736
737
738
739
740
741
742
743
744
                    "offload_ratio_val": 1,
                    "t5_offload_granularity_val": "block",
                    "precision_mode_val": "bf16",
                    "use_tiling_vae_val": True,
                    "offload_granularity_val": "block",
                },
            ),
            (
                8,
                (
                    {
                        "cpu_offload_val": True,
gushiqiao's avatar
gushiqiao committed
745
                        "t5_cpu_offload_val": True,
gushiqiao's avatar
gushiqiao committed
746
747
748
749
750
751
752
753
754
                        "offload_ratio_val": 1,
                        "t5_offload_granularity_val": "block",
                        "precision_mode_val": "bf16",
                        "use_tiling_vae_val": True,
                        "offload_granularity_val": "phase",
                        "t5_quant_scheme_val": quant_type,
                        "clip_quant_scheme_val": quant_type,
                        "dit_quant_scheme_val": quant_type,
                        "lazy_load_val": True,
gushiqiao's avatar
gushiqiao committed
755
                        "unload_modules_val": True,
gushiqiao's avatar
gushiqiao committed
756
757
                        "rotary_chunk_val": True,
                        "rotary_chunk_size_val": 10000,
gushiqiao's avatar
gushiqiao committed
758
                        "use_tiny_vae_val": True,
gushiqiao's avatar
gushiqiao committed
759
760
761
762
                    }
                    if res == "540p"
                    else {
                        "cpu_offload_val": True,
gushiqiao's avatar
gushiqiao committed
763
                        "t5_cpu_offload_val": True,
gushiqiao's avatar
gushiqiao committed
764
765
766
767
768
769
770
771
772
                        "offload_ratio_val": 1,
                        "t5_offload_granularity_val": "block",
                        "precision_mode_val": "bf16",
                        "use_tiling_vae_val": True,
                        "offload_granularity_val": "phase",
                        "t5_quant_scheme_val": quant_type,
                        "clip_quant_scheme_val": quant_type,
                        "dit_quant_scheme_val": quant_type,
                        "lazy_load_val": True,
gushiqiao's avatar
gushiqiao committed
773
                        "unload_modules_val": True,
gushiqiao's avatar
gushiqiao committed
774
                        "use_tiny_vae_val": True,
gushiqiao's avatar
gushiqiao committed
775
776
777
778
                    }
                ),
            ),
        ]
gushiqiao's avatar
gushiqiao committed
779

gushiqiao's avatar
gushiqiao committed
780
    else:
gushiqiao's avatar
gushiqiao committed
781
782
783
784
785
786
787
788
789
790
791
        gpu_rules = [
            (24, {}),
            (
                8,
                {
                    "t5_cpu_offload_val": True,
                    "t5_offload_granularity_val": "block",
                    "t5_quant_scheme_val": quant_type,
                },
            ),
        ]
gushiqiao's avatar
gushiqiao committed
792

gushiqiao's avatar
gushiqiao committed
793
794
795
796
797
798
799
    if is_14b:
        cpu_rules = [
            (128, {}),
            (64, {"dit_quant_scheme_val": quant_type}),
            (32, {"dit_quant_scheme_val": quant_type, "lazy_load_val": True}),
            (
                16,
gushiqiao's avatar
gushiqiao committed
800
801
802
803
804
                {
                    "dit_quant_scheme_val": quant_type,
                    "t5_quant_scheme_val": quant_type,
                    "clip_quant_scheme_val": quant_type,
                    "lazy_load_val": True,
gushiqiao's avatar
gushiqiao committed
805
                    "unload_modules_val": True,
gushiqiao's avatar
gushiqiao committed
806
                },
gushiqiao's avatar
gushiqiao committed
807
808
            ),
        ]
gushiqiao's avatar
gushiqiao committed
809
    else:
gushiqiao's avatar
gushiqiao committed
810
811
812
813
814
815
816
817
818
819
820
        cpu_rules = [
            (64, {}),
            (
                16,
                {
                    "t5_quant_scheme_val": quant_type,
                    "unload_modules_val": True,
                    "use_tiny_vae_val": True,
                },
            ),
        ]
gushiqiao's avatar
gushiqiao committed
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835

    for threshold, updates in gpu_rules:
        if gpu_memory >= threshold:
            default_config.update(updates)
            break

    for threshold, updates in cpu_rules:
        if cpu_memory >= threshold:
            default_config.update(updates)
            break

    return tuple(gr.update(value=default_config[key]) for key in default_config)


def main():
gushiqiao's avatar
gushiqiao committed
836
    with gr.Blocks(
gushiqiao's avatar
gushiqiao committed
837
        title="Lightx2v (Lightweight Video Inference and Generation Engine)",
gushiqiao's avatar
gushiqiao committed
838
839
840
841
842
843
        css="""
        .main-content { max-width: 1400px; margin: auto; }
        .output-video { max-height: 650px; }
        .warning { color: #ff6b6b; font-weight: bold; }
        .advanced-options { background: #f9f9ff; border-radius: 10px; padding: 15px; }
        .tab-button { font-size: 16px; padding: 10px 20px; }
gushiqiao's avatar
gushiqiao committed
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
        .auto-config-title {
            background: linear-gradient(45deg, #ff6b6b, #4ecdc4);
            background-clip: text;
            -webkit-background-clip: text;
            color: transparent;
            text-align: center;
            margin: 0 !important;
            padding: 8px;
            border: 2px solid #4ecdc4;
            border-radius: 8px;
            background-color: #f0f8ff;
        }
        .auto-config-checkbox {
            border: 2px solid #ff6b6b !important;
            border-radius: 8px !important;
            padding: 10px !important;
            background: linear-gradient(135deg, #fff5f5, #f0fff0) !important;
            box-shadow: 0 2px 8px rgba(255, 107, 107, 0.2) !important;
        }
        .auto-config-checkbox label {
            font-size: 16px !important;
            font-weight: bold !important;
            color: #2c3e50 !important;
        }
gushiqiao's avatar
gushiqiao committed
868
869
870
    """,
    ) as demo:
        gr.Markdown(f"# 🎬 {model_cls} Video Generator")
gushiqiao's avatar
gushiqiao committed
871
        gr.Markdown(f"### Using Model: {model_path}")
gushiqiao's avatar
gushiqiao committed
872
873
874
875
876
877
878
879

        with gr.Tabs() as tabs:
            with gr.Tab("Basic Settings", id=1):
                with gr.Row():
                    with gr.Column(scale=4):
                        with gr.Group():
                            gr.Markdown("## 📥 Input Parameters")

gushiqiao's avatar
gushiqiao committed
880
881
882
883
884
885
886
887
888
                            if task == "i2v":
                                with gr.Row():
                                    image_path = gr.Image(
                                        label="Input Image",
                                        type="filepath",
                                        height=300,
                                        interactive=True,
                                        visible=True,
                                    )
gushiqiao's avatar
gushiqiao committed
889
890
891
892
893
894
895
896
897
898
899
900
901

                            with gr.Row():
                                with gr.Column():
                                    prompt = gr.Textbox(
                                        label="Prompt",
                                        lines=3,
                                        placeholder="Describe the video content...",
                                        max_lines=5,
                                    )
                                with gr.Column():
                                    negative_prompt = gr.Textbox(
                                        label="Negative Prompt",
                                        lines=3,
gushiqiao's avatar
gushiqiao committed
902
                                        placeholder="What you don't want to appear in the video...",
gushiqiao's avatar
gushiqiao committed
903
                                        max_lines=5,
gushiqiao's avatar
gushiqiao committed
904
                                        value="Bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality, low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, misshapen limbs, fused fingers, still picture, messy background, three legs, many people in the background, walking backwards",
gushiqiao's avatar
gushiqiao committed
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
                                    )
                                with gr.Column():
                                    resolution = gr.Dropdown(
                                        choices=[
                                            # 720p
                                            ("1280x720 (16:9, 720p)", "1280x720"),
                                            ("720x1280 (9:16, 720p)", "720x1280"),
                                            ("1280x544 (21:9, 720p)", "1280x544"),
                                            ("544x1280 (9:21, 720p)", "544x1280"),
                                            ("1104x832 (4:3, 720p)", "1104x832"),
                                            ("832x1104 (3:4, 720p)", "832x1104"),
                                            ("960x960 (1:1, 720p)", "960x960"),
                                            # 480p
                                            ("960x544 (16:9, 540p)", "960x544"),
                                            ("544x960 (9:16, 540p)", "544x960"),
                                            ("832x480 (16:9, 480p)", "832x480"),
                                            ("480x832 (9:16, 480p)", "480x832"),
                                            ("832x624 (4:3, 480p)", "832x624"),
                                            ("624x832 (3:4, 480p)", "624x832"),
                                            ("720x720 (1:1, 480p)", "720x720"),
                                            ("512x512 (1:1, 480p)", "512x512"),
                                        ],
gushiqiao's avatar
gushiqiao committed
927
928
                                        value="832x480",
                                        label="Maximum Resolution",
gushiqiao's avatar
gushiqiao committed
929
                                    )
gushiqiao's avatar
gushiqiao committed
930
931

                                with gr.Column():
gushiqiao's avatar
gushiqiao committed
932
933
934
935
936
937
938
939
                                    with gr.Group():
                                        gr.Markdown("### 🚀 **Smart Configuration Recommendation**", elem_classes=["auto-config-title"])
                                        enable_auto_config = gr.Checkbox(
                                            label="🎯 **Auto-configure Inference Options**",
                                            value=False,
                                            info="💡 **Automatically optimize GPU settings to match the current resolution. After changing the resolution, please re-check this option to prevent potential performance degradation or runtime errors.**",
                                            elem_classes=["auto-config-checkbox"],
                                        )
gushiqiao's avatar
gushiqiao committed
940
                                with gr.Column(scale=9):
gushiqiao's avatar
gushiqiao committed
941
942
                                    seed = gr.Slider(
                                        label="Random Seed",
gushiqiao's avatar
gushiqiao committed
943
944
                                        minimum=0,
                                        maximum=MAX_NUMPY_SEED,
gushiqiao's avatar
gushiqiao committed
945
                                        step=1,
gushiqiao's avatar
gushiqiao committed
946
                                        value=generate_random_seed(),
gushiqiao's avatar
gushiqiao committed
947
                                    )
gushiqiao's avatar
gushiqiao committed
948
949
950
951
952
953
                                with gr.Column(scale=1):
                                    randomize_btn = gr.Button("🎲 Randomize", variant="secondary")

                                randomize_btn.click(fn=generate_random_seed, inputs=None, outputs=seed)

                                with gr.Column():
gushiqiao's avatar
gushiqiao committed
954
                                    # Set default inference steps based on model class
gushiqiao's avatar
gushiqiao committed
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
                                    if model_cls == "wan2.1_distill":
                                        infer_steps = gr.Slider(
                                            label="Inference Steps",
                                            minimum=4,
                                            maximum=4,
                                            step=1,
                                            value=4,
                                            interactive=False,
                                            info="Inference steps fixed at 4 for optimal performance for distill model.",
                                        )
                                    elif model_cls == "wan2.1":
                                        if task == "i2v":
                                            infer_steps = gr.Slider(
                                                label="Inference Steps",
                                                minimum=1,
                                                maximum=100,
                                                step=1,
                                                value=40,
                                                info="Number of inference steps for video generation. Increasing steps may improve quality but reduce speed.",
                                            )
                                        elif task == "t2v":
                                            infer_steps = gr.Slider(
                                                label="Inference Steps",
                                                minimum=1,
                                                maximum=100,
                                                step=1,
                                                value=50,
                                                info="Number of inference steps for video generation. Increasing steps may improve quality but reduce speed.",
                                            )
gushiqiao's avatar
gushiqiao committed
984

gushiqiao's avatar
gushiqiao committed
985
986
                            # Set default CFG based on model class
                            default_enable_cfg = False if model_cls == "wan2.1_distill" else True
gushiqiao's avatar
gushiqiao committed
987
988
                            enable_cfg = gr.Checkbox(
                                label="Enable Classifier-Free Guidance",
gushiqiao's avatar
gushiqiao committed
989
                                value=default_enable_cfg,
gushiqiao's avatar
gushiqiao committed
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
                                info="Enable classifier-free guidance to control prompt strength",
                            )
                            cfg_scale = gr.Slider(
                                label="CFG Scale Factor",
                                minimum=1,
                                maximum=10,
                                step=1,
                                value=5,
                                info="Controls the influence strength of the prompt. Higher values give more influence to the prompt.",
                            )
                            sample_shift = gr.Slider(
                                label="Distribution Shift",
                                value=5,
                                minimum=0,
                                maximum=10,
                                step=1,
                                info="Controls the degree of distribution shift for samples. Larger values indicate more significant shifts.",
gushiqiao's avatar
gushiqiao committed
1007
1008
                            )

gushiqiao's avatar
gushiqiao committed
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
                            fps = gr.Slider(
                                label="Frames Per Second (FPS)",
                                minimum=8,
                                maximum=30,
                                step=1,
                                value=16,
                                info="Frames per second of the video. Higher FPS results in smoother videos.",
                            )
                            num_frames = gr.Slider(
                                label="Total Frames",
                                minimum=16,
                                maximum=120,
                                step=1,
                                value=81,
                                info="Total number of frames in the video. More frames result in longer videos.",
                            )
gushiqiao's avatar
gushiqiao committed
1025

gushiqiao's avatar
gushiqiao committed
1026
1027
                        save_video_path = gr.Textbox(
                            label="Output Video Path",
gushiqiao's avatar
gushiqiao committed
1028
                            value=generate_unique_filename(output_dir),
gushiqiao's avatar
gushiqiao committed
1029
1030
                            info="Must include .mp4 extension. If left blank or using the default value, a unique filename will be automatically generated.",
                        )
gushiqiao's avatar
gushiqiao committed
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
                    with gr.Column(scale=6):
                        gr.Markdown("## 📤 Generated Video")
                        output_video = gr.Video(
                            label="Result",
                            height=624,
                            width=360,
                            autoplay=True,
                            elem_classes=["output-video"],
                        )

gushiqiao's avatar
gushiqiao committed
1041
                        infer_btn = gr.Button("Generate Video", variant="primary", size="lg")
gushiqiao's avatar
gushiqiao committed
1042

gushiqiao's avatar
gushiqiao committed
1043
1044
            with gr.Tab("⚙️ Advanced Options", id=2):
                with gr.Group(elem_classes="advanced-options"):
gushiqiao's avatar
gushiqiao committed
1045
                    gr.Markdown("### GPU Memory Optimization")
gushiqiao's avatar
gushiqiao committed
1046
                    with gr.Row():
gushiqiao's avatar
gushiqiao committed
1047
1048
                        rotary_chunk = gr.Checkbox(
                            label="Chunked Rotary Position Embedding",
gushiqiao's avatar
gushiqiao committed
1049
                            value=False,
gushiqiao's avatar
gushiqiao committed
1050
                            info="When enabled, processes rotary position embeddings in chunks to save GPU memory.",
gushiqiao's avatar
gushiqiao committed
1051
1052
                        )

gushiqiao's avatar
gushiqiao committed
1053
1054
1055
1056
1057
1058
1059
                        rotary_chunk_size = gr.Slider(
                            label="Rotary Embedding Chunk Size",
                            value=100,
                            minimum=100,
                            maximum=10000,
                            step=100,
                            info="Controls the chunk size for applying rotary embeddings. Larger values may improve performance but increase memory usage. Only effective if 'rotary_chunk' is checked.",
gushiqiao's avatar
gushiqiao committed
1060
1061
                        )

gushiqiao's avatar
gushiqiao committed
1062
1063
1064
1065
1066
                        unload_modules = gr.Checkbox(
                            label="Unload Modules",
                            value=False,
                            info="Unload modules (T5, CLIP, DIT, etc.) after inference to reduce GPU/CPU memory usage",
                        )
gushiqiao's avatar
gushiqiao committed
1067
1068
1069
                        clean_cuda_cache = gr.Checkbox(
                            label="Clean CUDA Memory Cache",
                            value=False,
gushiqiao's avatar
gushiqiao committed
1070
                            info="When enabled, frees up GPU memory promptly but slows down inference.",
gushiqiao's avatar
gushiqiao committed
1071
1072
                        )

gushiqiao's avatar
gushiqiao committed
1073
                    gr.Markdown("### Asynchronous Offloading")
gushiqiao's avatar
gushiqiao committed
1074
1075
                    with gr.Row():
                        cpu_offload = gr.Checkbox(
gushiqiao's avatar
gushiqiao committed
1076
1077
1078
1079
1080
1081
1082
                            label="CPU Offloading",
                            value=False,
                            info="Offload parts of the model computation from GPU to CPU to reduce GPU memory usage",
                        )

                        lazy_load = gr.Checkbox(
                            label="Enable Lazy Loading",
gushiqiao's avatar
gushiqiao committed
1083
                            value=False,
gushiqiao's avatar
gushiqiao committed
1084
                            info="Lazy load model components during inference. Requires CPU loading and DIT quantization.",
gushiqiao's avatar
gushiqiao committed
1085
                        )
gushiqiao's avatar
gushiqiao committed
1086

gushiqiao's avatar
gushiqiao committed
1087
1088
1089
                        offload_granularity = gr.Dropdown(
                            label="Dit Offload Granularity",
                            choices=["block", "phase"],
gushiqiao's avatar
gushiqiao committed
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
                            value="phase",
                            info="Sets Dit model offloading granularity: blocks or computational phases",
                        )
                        offload_ratio = gr.Slider(
                            label="Offload ratio for Dit model",
                            minimum=0.0,
                            maximum=1.0,
                            step=0.1,
                            value=1.0,
                            info="Controls how much of the Dit model is offloaded to the CPU",
gushiqiao's avatar
gushiqiao committed
1100
                        )
gushiqiao's avatar
gushiqiao committed
1101
1102
1103
1104
1105
1106
                        t5_cpu_offload = gr.Checkbox(
                            label="T5 CPU Offloading",
                            value=False,
                            info="Offload the T5 Encoder model to CPU to reduce GPU memory usage",
                        )

gushiqiao's avatar
gushiqiao committed
1107
1108
1109
                        t5_offload_granularity = gr.Dropdown(
                            label="T5 Encoder Offload Granularity",
                            choices=["model", "block"],
gushiqiao's avatar
gushiqiao committed
1110
1111
                            value="model",
                            info="Controls the granularity when offloading the T5 Encoder model to CPU",
gushiqiao's avatar
gushiqiao committed
1112
1113
1114
1115
                        )

                    gr.Markdown("### Low-Precision Quantization")
                    with gr.Row():
gushiqiao's avatar
gushiqiao committed
1116
1117
1118
1119
1120
1121
                        torch_compile = gr.Checkbox(
                            label="Torch Compile",
                            value=False,
                            info="Use torch.compile to accelerate the inference process",
                        )

gushiqiao's avatar
gushiqiao committed
1122
1123
                        attention_type = gr.Dropdown(
                            label="Attention Operator",
gushiqiao's avatar
gushiqiao committed
1124
1125
1126
                            choices=[op[1] for op in attn_op_choices],
                            value=attn_op_choices[0][1],
                            info="Use appropriate attention operators to accelerate inference",
gushiqiao's avatar
gushiqiao committed
1127
1128
                        )
                        quant_op = gr.Dropdown(
gushiqiao's avatar
gushiqiao committed
1129
1130
1131
1132
1133
                            label="Quantization Matmul Operator",
                            choices=[op[1] for op in quant_op_choices],
                            value=quant_op_choices[0][1],
                            info="Select the quantization matrix multiplication operator to accelerate inference",
                            interactive=True,
gushiqiao's avatar
gushiqiao committed
1134
                        )
gushiqiao's avatar
gushiqiao committed
1135
1136
1137
                        # Get dynamic quantization options
                        quant_options = get_quantization_options(model_path)

gushiqiao's avatar
gushiqiao committed
1138
1139
                        dit_quant_scheme = gr.Dropdown(
                            label="Dit",
gushiqiao's avatar
gushiqiao committed
1140
1141
                            choices=quant_options["dit_choices"],
                            value=quant_options["dit_default"],
gushiqiao's avatar
gushiqiao committed
1142
                            info="Quantization precision for the Dit model",
gushiqiao's avatar
gushiqiao committed
1143
1144
1145
                        )
                        t5_quant_scheme = gr.Dropdown(
                            label="T5 Encoder",
gushiqiao's avatar
gushiqiao committed
1146
1147
                            choices=quant_options["t5_choices"],
                            value=quant_options["t5_default"],
gushiqiao's avatar
gushiqiao committed
1148
                            info="Quantization precision for the T5 Encoder model",
gushiqiao's avatar
gushiqiao committed
1149
1150
1151
                        )
                        clip_quant_scheme = gr.Dropdown(
                            label="Clip Encoder",
gushiqiao's avatar
gushiqiao committed
1152
1153
                            choices=quant_options["clip_choices"],
                            value=quant_options["clip_default"],
gushiqiao's avatar
gushiqiao committed
1154
                            info="Quantization precision for the Clip Encoder",
gushiqiao's avatar
gushiqiao committed
1155
1156
                        )
                        precision_mode = gr.Dropdown(
gushiqiao's avatar
gushiqiao committed
1157
                            label="Precision Mode for Sensitive Layers",
gushiqiao's avatar
gushiqiao committed
1158
                            choices=["fp32", "bf16"],
gushiqiao's avatar
gushiqiao committed
1159
                            value="fp32",
gushiqiao's avatar
gushiqiao committed
1160
                            info="Select the numerical precision for critical model components like normalization and embedding layers. FP32 offers higher accuracy, while BF16 improves performance on compatible hardware.",
gushiqiao's avatar
gushiqiao committed
1161
1162
1163
1164
1165
                        )

                    gr.Markdown("### Variational Autoencoder (VAE)")
                    with gr.Row():
                        use_tiny_vae = gr.Checkbox(
gushiqiao's avatar
gushiqiao committed
1166
                            label="Use Tiny VAE",
gushiqiao's avatar
gushiqiao committed
1167
1168
1169
1170
                            value=False,
                            info="Use a lightweight VAE model to accelerate the decoding process",
                        )
                        use_tiling_vae = gr.Checkbox(
gushiqiao's avatar
gushiqiao committed
1171
                            label="VAE Tiling Inference",
gushiqiao's avatar
gushiqiao committed
1172
                            value=False,
gushiqiao's avatar
gushiqiao committed
1173
                            info="Use VAE tiling inference to reduce GPU memory usage",
gushiqiao's avatar
gushiqiao committed
1174
1175
1176
1177
1178
                        )

                    gr.Markdown("### Feature Caching")
                    with gr.Row():
                        enable_teacache = gr.Checkbox(
gushiqiao's avatar
gushiqiao committed
1179
                            label="Tea Cache",
gushiqiao's avatar
gushiqiao committed
1180
1181
1182
1183
1184
1185
1186
1187
                            value=False,
                            info="Cache features during inference to reduce the number of inference steps",
                        )
                        teacache_thresh = gr.Slider(
                            label="Tea Cache Threshold",
                            value=0.26,
                            minimum=0,
                            maximum=1,
gushiqiao's avatar
gushiqiao committed
1188
1189
1190
1191
1192
1193
                            info="Higher acceleration may result in lower quality —— Setting to 0.1 provides ~2.0x acceleration, setting to 0.2 provides ~3.0x acceleration",
                        )
                        use_ret_steps = gr.Checkbox(
                            label="Cache Only Key Steps",
                            value=False,
                            info="When checked, cache is written only at key steps where the scheduler returns results; when unchecked, cache is written at all steps to ensure the highest quality",
gushiqiao's avatar
gushiqiao committed
1194
1195
                        )

gushiqiao's avatar
gushiqiao committed
1196
1197
                enable_auto_config.change(
                    fn=auto_configure,
gushiqiao's avatar
gushiqiao committed
1198
                    inputs=[enable_auto_config, resolution],
gushiqiao's avatar
gushiqiao committed
1199
1200
1201
1202
1203
1204
1205
1206
1207
                    outputs=[
                        torch_compile,
                        lazy_load,
                        rotary_chunk,
                        rotary_chunk_size,
                        clean_cuda_cache,
                        cpu_offload,
                        offload_granularity,
                        offload_ratio,
gushiqiao's avatar
gushiqiao committed
1208
1209
                        t5_cpu_offload,
                        unload_modules,
gushiqiao's avatar
gushiqiao committed
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
                        t5_offload_granularity,
                        attention_type,
                        quant_op,
                        dit_quant_scheme,
                        t5_quant_scheme,
                        clip_quant_scheme,
                        precision_mode,
                        use_tiny_vae,
                        use_tiling_vae,
                        enable_teacache,
                        teacache_thresh,
                        use_ret_steps,
                    ],
                )
gushiqiao's avatar
gushiqiao committed
1224
1225
1226
1227
1228
1229

                lazy_load.change(
                    fn=handle_lazy_load_change,
                    inputs=[lazy_load],
                    outputs=[unload_modules],
                )
gushiqiao's avatar
gushiqiao committed
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
        if task == "i2v":
            infer_btn.click(
                fn=run_inference,
                inputs=[
                    prompt,
                    negative_prompt,
                    save_video_path,
                    torch_compile,
                    infer_steps,
                    num_frames,
                    resolution,
                    seed,
                    sample_shift,
                    enable_teacache,
                    teacache_thresh,
                    use_ret_steps,
                    enable_cfg,
                    cfg_scale,
                    dit_quant_scheme,
                    t5_quant_scheme,
                    clip_quant_scheme,
                    fps,
                    use_tiny_vae,
                    use_tiling_vae,
                    lazy_load,
                    precision_mode,
                    cpu_offload,
                    offload_granularity,
                    offload_ratio,
gushiqiao's avatar
gushiqiao committed
1259
1260
                    t5_cpu_offload,
                    unload_modules,
gushiqiao's avatar
gushiqiao committed
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
                    t5_offload_granularity,
                    attention_type,
                    quant_op,
                    rotary_chunk,
                    rotary_chunk_size,
                    clean_cuda_cache,
                    image_path,
                ],
                outputs=output_video,
            )
        else:
            infer_btn.click(
                fn=run_inference,
                inputs=[
                    prompt,
                    negative_prompt,
                    save_video_path,
                    torch_compile,
                    infer_steps,
                    num_frames,
                    resolution,
                    seed,
                    sample_shift,
                    enable_teacache,
                    teacache_thresh,
                    use_ret_steps,
                    enable_cfg,
                    cfg_scale,
                    dit_quant_scheme,
                    t5_quant_scheme,
                    clip_quant_scheme,
                    fps,
                    use_tiny_vae,
                    use_tiling_vae,
                    lazy_load,
                    precision_mode,
                    cpu_offload,
                    offload_granularity,
                    offload_ratio,
gushiqiao's avatar
gushiqiao committed
1300
1301
                    t5_cpu_offload,
                    unload_modules,
gushiqiao's avatar
gushiqiao committed
1302
1303
1304
1305
1306
1307
1308
1309
1310
                    t5_offload_granularity,
                    attention_type,
                    quant_op,
                    rotary_chunk,
                    rotary_chunk_size,
                    clean_cuda_cache,
                ],
                outputs=output_video,
            )
gushiqiao's avatar
gushiqiao committed
1311

gushiqiao's avatar
gushiqiao committed
1312
    demo.launch(share=True, server_port=args.server_port, server_name=args.server_name, inbrowser=True, allowed_paths=[output_dir])
gushiqiao's avatar
gushiqiao committed
1313
1314
1315


if __name__ == "__main__":
gushiqiao's avatar
gushiqiao committed
1316
1317
1318
1319
1320
    parser = argparse.ArgumentParser(description="Light Video Generation")
    parser.add_argument("--model_path", type=str, required=True, help="Model folder path")
    parser.add_argument(
        "--model_cls",
        type=str,
gushiqiao's avatar
gushiqiao committed
1321
        choices=["wan2.1", "wan2.1_distill"],
gushiqiao's avatar
gushiqiao committed
1322
        default="wan2.1",
gushiqiao's avatar
gushiqiao committed
1323
        help="Model class to use (wan2.1: standard model, wan2.1_distill: distilled model for faster inference)",
gushiqiao's avatar
gushiqiao committed
1324
    )
gushiqiao's avatar
gushiqiao committed
1325
    parser.add_argument("--model_size", type=str, required=True, choices=["14b", "1.3b"], help="Model type to use")
gushiqiao's avatar
gushiqiao committed
1326
    parser.add_argument("--task", type=str, required=True, choices=["i2v", "t2v"], help="Specify the task type. 'i2v' for image-to-video translation, 't2v' for text-to-video generation.")
gushiqiao's avatar
gushiqiao committed
1327
1328
    parser.add_argument("--server_port", type=int, default=7862, help="Server port")
    parser.add_argument("--server_name", type=str, default="0.0.0.0", help="Server ip")
gushiqiao's avatar
gushiqiao committed
1329
    parser.add_argument("--output_dir", type=str, default="./outputs", help="Output video save directory")
gushiqiao's avatar
gushiqiao committed
1330
1331
    args = parser.parse_args()

gushiqiao's avatar
gushiqiao committed
1332
    global model_path, model_cls, model_size, output_dir
gushiqiao's avatar
gushiqiao committed
1333
1334
    model_path = args.model_path
    model_cls = args.model_cls
gushiqiao's avatar
gushiqiao committed
1335
    model_size = args.model_size
gushiqiao's avatar
gushiqiao committed
1336
    task = args.task
gushiqiao's avatar
gushiqiao committed
1337
    output_dir = args.output_dir
gushiqiao's avatar
gushiqiao committed
1338

gushiqiao's avatar
gushiqiao committed
1339
    main()