model_management.py 13.9 KB
Newer Older
1
2
import psutil
from enum import Enum
comfyanonymous's avatar
comfyanonymous committed
3
from comfy.cli_args import args
4
import torch
5

6
class VRAMState(Enum):
7
    DISABLED = 0
8
9
10
11
    NO_VRAM = 1
    LOW_VRAM = 2
    NORMAL_VRAM = 3
    HIGH_VRAM = 4
12
13
14
15
16
17
    SHARED = 5

class CPUState(Enum):
    GPU = 0
    CPU = 1
    MPS = 2
18

19
20
21
# Determine VRAM State
vram_state = VRAMState.NORMAL_VRAM
set_vram_to = VRAMState.NORMAL_VRAM
22
cpu_state = CPUState.GPU
23

24
total_vram = 0
25

26
lowvram_available = True
藍+85CD's avatar
藍+85CD committed
27
xpu_available = False
28

29
directml_enabled = False
30
if args.directml is not None:
31
32
    import torch_directml
    directml_enabled = True
33
34
35
36
37
38
    device_index = args.directml
    if device_index < 0:
        directml_device = torch_directml.device()
    else:
        directml_device = torch_directml.device(device_index)
    print("Using directml with device:", torch_directml.device_name(device_index))
39
    # torch_directml.disable_tiled_resources(True)
40
    lowvram_available = False #TODO: need to find a way to get free memory in directml before this can be enabled by default.
41

42
try:
43
44
45
    import intel_extension_for_pytorch as ipex
    if torch.xpu.is_available():
        xpu_available = True
46
47
48
except:
    pass

49
50
51
52
53
54
55
56
57
try:
    if torch.backends.mps.is_available():
        cpu_state = CPUState.MPS
except:
    pass

if args.cpu:
    cpu_state = CPUState.CPU

58
59
60
def get_torch_device():
    global xpu_available
    global directml_enabled
61
    global cpu_state
62
63
64
    if directml_enabled:
        global directml_device
        return directml_device
65
    if cpu_state == CPUState.MPS:
66
        return torch.device("mps")
67
    if cpu_state == CPUState.CPU:
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
        return torch.device("cpu")
    else:
        if xpu_available:
            return torch.device("xpu")
        else:
            return torch.device(torch.cuda.current_device())

def get_total_memory(dev=None, torch_total_too=False):
    global xpu_available
    global directml_enabled
    if dev is None:
        dev = get_torch_device()

    if hasattr(dev, 'type') and (dev.type == 'cpu' or dev.type == 'mps'):
        mem_total = psutil.virtual_memory().total
        mem_total_torch = mem_total
    else:
        if directml_enabled:
            mem_total = 1024 * 1024 * 1024 #TODO
            mem_total_torch = mem_total
        elif xpu_available:
            mem_total = torch.xpu.get_device_properties(dev).total_memory
            mem_total_torch = mem_total
        else:
            stats = torch.cuda.memory_stats(dev)
            mem_reserved = stats['reserved_bytes.all.current']
            _, mem_total_cuda = torch.cuda.mem_get_info(dev)
            mem_total_torch = mem_reserved
            mem_total = mem_total_cuda

    if torch_total_too:
        return (mem_total, mem_total_torch)
    else:
        return mem_total

total_vram = get_total_memory(get_torch_device()) / (1024 * 1024)
total_ram = psutil.virtual_memory().total / (1024 * 1024)
print("Total VRAM {:0.0f} MB, total RAM {:0.0f} MB".format(total_vram, total_ram))
if not args.normalvram and not args.cpu:
    if lowvram_available and total_vram <= 4096:
        print("Trying to enable lowvram mode because your GPU seems to have 4GB or less. If you don't want this use: --normalvram")
        set_vram_to = VRAMState.LOW_VRAM
    elif total_vram > total_ram * 1.1 and total_vram > 14336:
        print("Enabling highvram mode because your GPU has more vram than your computer has ram. If you don't want this use: --normalvram")
        vram_state = VRAMState.HIGH_VRAM

114
115
116
117
118
try:
    OOM_EXCEPTION = torch.cuda.OutOfMemoryError
except:
    OOM_EXCEPTION = Exception

119
120
XFORMERS_VERSION = ""
XFORMERS_ENABLED_VAE = True
121
122
if args.disable_xformers:
    XFORMERS_IS_AVAILABLE = False
123
124
125
126
else:
    try:
        import xformers
        import xformers.ops
127
        XFORMERS_IS_AVAILABLE = True
128
129
130
131
132
133
134
135
136
137
138
        try:
            XFORMERS_VERSION = xformers.version.__version__
            print("xformers version:", XFORMERS_VERSION)
            if XFORMERS_VERSION.startswith("0.0.18"):
                print()
                print("WARNING: This version of xformers has a major bug where you will get black images when generating high resolution images.")
                print("Please downgrade or upgrade xformers to a different version.")
                print()
                XFORMERS_ENABLED_VAE = False
        except:
            pass
139
    except:
140
        XFORMERS_IS_AVAILABLE = False
141

142
143
ENABLE_PYTORCH_ATTENTION = args.use_pytorch_cross_attention
if ENABLE_PYTORCH_ATTENTION:
144
145
146
    torch.backends.cuda.enable_math_sdp(True)
    torch.backends.cuda.enable_flash_sdp(True)
    torch.backends.cuda.enable_mem_efficient_sdp(True)
147
    XFORMERS_IS_AVAILABLE = False
148

149
150
if args.lowvram:
    set_vram_to = VRAMState.LOW_VRAM
151
    lowvram_available = True
152
153
154
155
elif args.novram:
    set_vram_to = VRAMState.NO_VRAM
elif args.highvram:
    vram_state = VRAMState.HIGH_VRAM
156

157
158
159
160
161
FORCE_FP32 = False
if args.force_fp32:
    print("Forcing FP32, if this improves things please report it.")
    FORCE_FP32 = True

162
if lowvram_available:
163
164
    try:
        import accelerate
165
166
        if set_vram_to in (VRAMState.LOW_VRAM, VRAMState.NO_VRAM):
            vram_state = set_vram_to
167
168
169
    except Exception as e:
        import traceback
        print(traceback.format_exc())
170
171
        print("ERROR: LOW VRAM MODE NEEDS accelerate.")
        lowvram_available = False
172

173

174
175
if cpu_state != CPUState.GPU:
    vram_state = VRAMState.DISABLED
176

177
178
if cpu_state == CPUState.MPS:
    vram_state = VRAMState.SHARED
179

180
print(f"Set vram state to: {vram_state.name}")
181

182

183
184
def get_torch_device_name(device):
    if hasattr(device, 'type'):
185
186
187
188
189
190
        if device.type == "cuda":
            return "{} {}".format(device, torch.cuda.get_device_name(device))
        else:
            return "{}".format(device.type)
    else:
        return "CUDA {}: {}".format(device, torch.cuda.get_device_name(device))
191
192

try:
193
    print("Device:", get_torch_device_name(get_torch_device()))
194
195
196
except:
    print("Could not pick default device.")

197
198

current_loaded_model = None
comfyanonymous's avatar
comfyanonymous committed
199
current_gpu_controlnets = []
200

201
202
203
model_accelerated = False


204
205
def unload_model():
    global current_loaded_model
206
    global model_accelerated
comfyanonymous's avatar
comfyanonymous committed
207
    global current_gpu_controlnets
208
209
    global vram_state

210
    if current_loaded_model is not None:
211
212
213
214
        if model_accelerated:
            accelerate.hooks.remove_hook_from_submodules(current_loaded_model.model)
            model_accelerated = False

215
        #never unload models from GPU on high vram
216
        if vram_state != VRAMState.HIGH_VRAM:
217
            current_loaded_model.model.cpu()
218
            current_loaded_model.model_patches_to("cpu")
219
220
        current_loaded_model.unpatch_model()
        current_loaded_model = None
221

222
    if vram_state != VRAMState.HIGH_VRAM:
223
224
225
226
        if len(current_gpu_controlnets) > 0:
            for n in current_gpu_controlnets:
                n.cpu()
            current_gpu_controlnets = []
227
228
229
230


def load_model_gpu(model):
    global current_loaded_model
231
232
233
    global vram_state
    global model_accelerated

234
235
236
237
238
239
240
241
    if model is current_loaded_model:
        return
    unload_model()
    try:
        real_model = model.patch_model()
    except Exception as e:
        model.unpatch_model()
        raise e
242

243
244
245
246
247
248
249
    torch_dev = get_torch_device()
    model.model_patches_to(torch_dev)

    vram_set_state = vram_state
    if lowvram_available and (vram_set_state == VRAMState.LOW_VRAM or vram_set_state == VRAMState.NORMAL_VRAM):
        model_size = model.model_size()
        current_free_mem = get_free_memory(torch_dev)
250
        lowvram_model_memory = int(max(256 * (1024 * 1024), (current_free_mem - 1024 * (1024 * 1024)) / 1.3 ))
251
252
253
        if model_size > (current_free_mem - (512 * 1024 * 1024)): #only switch to lowvram if really necessary
            vram_set_state = VRAMState.LOW_VRAM

254
    current_loaded_model = model
255

256
    if vram_set_state == VRAMState.DISABLED:
Yurii Mazurevich's avatar
Yurii Mazurevich committed
257
        pass
258
    elif vram_set_state == VRAMState.NORMAL_VRAM or vram_set_state == VRAMState.HIGH_VRAM or vram_set_state == VRAMState.SHARED:
259
        model_accelerated = False
comfyanonymous's avatar
comfyanonymous committed
260
        real_model.to(get_torch_device())
261
    else:
262
        if vram_set_state == VRAMState.NO_VRAM:
263
            device_map = accelerate.infer_auto_device_map(real_model, max_memory={0: "256MiB", "cpu": "16GiB"})
264
265
        elif vram_set_state == VRAMState.LOW_VRAM:
            device_map = accelerate.infer_auto_device_map(real_model, max_memory={0: "{}MiB".format(lowvram_model_memory // (1024 * 1024)), "cpu": "16GiB"})
comfyanonymous's avatar
comfyanonymous committed
266

comfyanonymous's avatar
comfyanonymous committed
267
        accelerate.dispatch_model(real_model, device_map=device_map, main_device=get_torch_device())
268
        model_accelerated = True
269
    return current_loaded_model
270

271
def load_controlnet_gpu(control_models):
comfyanonymous's avatar
comfyanonymous committed
272
    global current_gpu_controlnets
273
    global vram_state
274
    if vram_state == VRAMState.DISABLED:
275
        return
276

277
    if vram_state == VRAMState.LOW_VRAM or vram_state == VRAMState.NO_VRAM:
278
279
280
        for m in control_models:
            if hasattr(m, 'set_lowvram'):
                m.set_lowvram(True)
281
282
283
        #don't load controlnets like this if low vram because they will be loaded right before running and unloaded right after
        return

284
285
286
287
    models = []
    for m in control_models:
        models += m.get_models()

comfyanonymous's avatar
comfyanonymous committed
288
289
290
291
    for m in current_gpu_controlnets:
        if m not in models:
            m.cpu()

292
    device = get_torch_device()
comfyanonymous's avatar
comfyanonymous committed
293
294
    current_gpu_controlnets = []
    for m in models:
295
        current_gpu_controlnets.append(m.to(device))
comfyanonymous's avatar
comfyanonymous committed
296

297

298
299
def load_if_low_vram(model):
    global vram_state
300
    if vram_state == VRAMState.LOW_VRAM or vram_state == VRAMState.NO_VRAM:
comfyanonymous's avatar
comfyanonymous committed
301
        return model.to(get_torch_device())
302
303
304
305
    return model

def unload_if_low_vram(model):
    global vram_state
306
    if vram_state == VRAMState.LOW_VRAM or vram_state == VRAMState.NO_VRAM:
307
308
309
        return model.cpu()
    return model

310
311
312
313
def get_autocast_device(dev):
    if hasattr(dev, 'type'):
        return dev.type
    return "cuda"
314

315

316
def xformers_enabled():
317
318
    global xpu_available
    global directml_enabled
319
320
    global cpu_state
    if cpu_state != CPUState.GPU:
321
        return False
322
323
324
325
    if xpu_available:
        return False
    if directml_enabled:
        return False
326
    return XFORMERS_IS_AVAILABLE
327

328
329
330
331
332

def xformers_enabled_vae():
    enabled = xformers_enabled()
    if not enabled:
        return False
333
334

    return XFORMERS_ENABLED_VAE
335

336
def pytorch_attention_enabled():
337
    global ENABLE_PYTORCH_ATTENTION
338
339
    return ENABLE_PYTORCH_ATTENTION

340
341
342
343
344
345
346
347
def pytorch_attention_flash_attention():
    global ENABLE_PYTORCH_ATTENTION
    if ENABLE_PYTORCH_ATTENTION:
        #TODO: more reliable way of checking for flash attention?
        if torch.version.cuda: #pytorch flash attention only works on Nvidia
            return True
    return False

348
def get_free_memory(dev=None, torch_free_too=False):
349
    global xpu_available
350
    global directml_enabled
351
    if dev is None:
352
        dev = get_torch_device()
353

Yurii Mazurevich's avatar
Yurii Mazurevich committed
354
    if hasattr(dev, 'type') and (dev.type == 'cpu' or dev.type == 'mps'):
355
356
357
        mem_free_total = psutil.virtual_memory().available
        mem_free_torch = mem_free_total
    else:
358
359
360
361
        if directml_enabled:
            mem_free_total = 1024 * 1024 * 1024 #TODO
            mem_free_torch = mem_free_total
        elif xpu_available:
362
363
364
365
366
367
368
369
370
            mem_free_total = torch.xpu.get_device_properties(dev).total_memory - torch.xpu.memory_allocated(dev)
            mem_free_torch = mem_free_total
        else:
            stats = torch.cuda.memory_stats(dev)
            mem_active = stats['active_bytes.all.current']
            mem_reserved = stats['reserved_bytes.all.current']
            mem_free_cuda, _ = torch.cuda.mem_get_info(dev)
            mem_free_torch = mem_reserved - mem_active
            mem_free_total = mem_free_cuda + mem_free_torch
371
372
373
374
375

    if torch_free_too:
        return (mem_free_total, mem_free_torch)
    else:
        return mem_free_total
376
377
378

def maximum_batch_area():
    global vram_state
379
    if vram_state == VRAMState.NO_VRAM:
380
381
382
        return 0

    memory_free = get_free_memory() / (1024 * 1024)
383
    if xformers_enabled() or pytorch_attention_flash_attention():
384
        #TODO: this needs to be tweaked
385
        area = 20 * memory_free
386
387
388
    else:
        #TODO: this formula is because AMD sucks and has memory management issues which might be fixed in the future
        area = ((memory_free - 1024) * 0.9) / (0.6)
389
    return int(max(area, 0))
390
391

def cpu_mode():
392
393
    global cpu_state
    return cpu_state == CPUState.CPU
394

Yurii Mazurevich's avatar
Yurii Mazurevich committed
395
def mps_mode():
396
397
    global cpu_state
    return cpu_state == CPUState.MPS
Yurii Mazurevich's avatar
Yurii Mazurevich committed
398

399
def should_use_fp16():
400
    global xpu_available
401
402
    global directml_enabled

403
404
405
    if FORCE_FP32:
        return False

406
407
408
    if directml_enabled:
        return False

409
    if cpu_mode() or mps_mode() or xpu_available:
410
411
412
413
414
        return False #TODO ?

    if torch.cuda.is_bf16_supported():
        return True

comfyanonymous's avatar
comfyanonymous committed
415
    props = torch.cuda.get_device_properties("cuda")
416
417
418
419
    if props.major < 7:
        return False

    #FP32 is faster on those cards?
420
    nvidia_16_series = ["1660", "1650", "1630", "T500", "T550", "T600"]
421
422
423
424
425
426
    for x in nvidia_16_series:
        if x in props.name:
            return False

    return True

427
428
def soft_empty_cache():
    global xpu_available
429
430
    global cpu_state
    if cpu_state == CPUState.MPS:
comfyanonymous's avatar
comfyanonymous committed
431
432
        torch.mps.empty_cache()
    elif xpu_available:
433
434
435
436
437
438
        torch.xpu.empty_cache()
    elif torch.cuda.is_available():
        if torch.version.cuda: #This seems to make things worse on ROCm so I only do it for cuda
            torch.cuda.empty_cache()
            torch.cuda.ipc_collect()

439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
#TODO: might be cleaner to put this somewhere else
import threading

class InterruptProcessingException(Exception):
    pass

interrupt_processing_mutex = threading.RLock()

interrupt_processing = False
def interrupt_current_processing(value=True):
    global interrupt_processing
    global interrupt_processing_mutex
    with interrupt_processing_mutex:
        interrupt_processing = value

def processing_interrupted():
    global interrupt_processing
    global interrupt_processing_mutex
    with interrupt_processing_mutex:
        return interrupt_processing

def throw_exception_if_processing_interrupted():
    global interrupt_processing
    global interrupt_processing_mutex
    with interrupt_processing_mutex:
        if interrupt_processing:
            interrupt_processing = False
            raise InterruptProcessingException()