{{ message }}
Repository navigation
Expand file tree
/
Copy pathloader.py
More file actions
506 lines (445 loc) · 20.7 KB
/
Copy pathloader.py
File metadata and controls
506 lines (445 loc) · 20.7 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
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
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
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
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
# (c) City96 || Apache-2.0 (apache.org/licenses/LICENSE-2.0)
import warnings
import logging
import torch
import gguf
import re
import os
from .ops import GGMLTensor
from .dequant import is_quantized, dequantize_tensor
IMG_ARCH_LIST = {"flux", "sd1", "sdxl", "sd3", "aura", "hidream", "cosmos", "ltxv", "hyvid", "wan", "lumina2", "qwen_image"}
TXT_ARCH_LIST = {"t5", "t5encoder", "llama", "qwen2vl", "qwen3", "qwen3vl", "gemma3"}
VIS_TYPE_LIST = {"clip-vision", "mmproj"}
def get_orig_shape(reader, tensor_name):
field_key = f"comfy.gguf.orig_shape.{tensor_name}"
field = reader.get_field(field_key)
if field is None:
return None
# Has original shape metadata, so we try to decode it.
if len(field.types) != 2 or field.types[0] != gguf.GGUFValueType.ARRAY or field.types[1] != gguf.GGUFValueType.INT32:
raise TypeError(f"Bad original shape metadata for {field_key}: Expected ARRAY of INT32, got {field.types}")
return torch.Size(tuple(int(field.parts[part_idx][0]) for part_idx in field.data))
def get_field(reader, field_name, field_type):
field = reader.get_field(field_name)
if field is None:
return None
elif field_type == str:
# extra check here as this is used for checking arch string
if len(field.types) != 1 or field.types[0] != gguf.GGUFValueType.STRING:
raise TypeError(f"Bad type for GGUF {field_name} key: expected string, got {field.types!r}")
return str(field.parts[field.data[-1]], encoding="utf-8")
elif field_type in [int, float, bool]:
return field_type(field.parts[field.data[-1]].item())
else:
raise TypeError(f"Unknown field type {field_type}")
def get_list_field(reader, field_name, field_type):
field = reader.get_field(field_name)
if field is None:
return None
elif field_type == str:
return tuple(str(field.parts[part_idx], encoding="utf-8") for part_idx in field.data)
elif field_type in [int, float, bool]:
return tuple(field_type(field.parts[part_idx][0]) for part_idx in field.data)
else:
raise TypeError(f"Unknown field type {field_type}")
def get_gguf_metadata(reader):
"""Extract all simple metadata fields like safetensors"""
metadata = {}
for field_name in reader.fields:
try:
field = reader.get_field(field_name)
if len(field.types) == 1: # Simple scalar fields only
if field.types[0] == gguf.GGUFValueType.STRING:
metadata[field_name] = str(field.parts[field.data[-1]], "utf-8")
elif field.types[0] == gguf.GGUFValueType.INT32:
metadata[field_name] = int(field.parts[field.data[-1]])
elif field.types[0] == gguf.GGUFValueType.F32:
metadata[field_name] = float(field.parts[field.data[-1]])
elif field.types[0] == gguf.GGUFValueType.BOOL:
metadata[field_name] = bool(field.parts[field.data[-1]])
except:
continue
return metadata
def gguf_sd_loader(path, handle_prefix="model.diffusion_model.", is_text_model=False):
"""
Read state dict as fake tensors
"""
reader = gguf.GGUFReader(path)
# filter and strip prefix
has_prefix = False
if handle_prefix is not None:
prefix_len = len(handle_prefix)
tensor_names = set(tensor.name for tensor in reader.tensors)
has_prefix = any(s.startswith(handle_prefix) for s in tensor_names)
tensors = []
for tensor in reader.tensors:
sd_key = tensor_name = tensor.name
if has_prefix:
if not tensor_name.startswith(handle_prefix):
continue
sd_key = tensor_name[prefix_len:]
tensors.append((sd_key, tensor))
# detect and verify architecture
compat = None
arch_str = get_field(reader, "general.architecture", str)
type_str = get_field(reader, "general.type", str)
if arch_str in [None, "pig", "cow"]:
if is_text_model:
raise ValueError(f"This gguf file is incompatible with llama.cpp!\nConsider using safetensors or a compatible gguf file\n({path})")
compat = "sd.cpp" if arch_str is None else arch_str
# import here to avoid changes to convert.py breaking regular models
from .tools.convert import detect_arch
try:
arch_str = detect_arch(set(val[0] for val in tensors)).arch
except Exception as e:
raise ValueError(f"This model is not currently supported - ({e})")
elif arch_str not in TXT_ARCH_LIST and is_text_model:
if type_str not in VIS_TYPE_LIST:
raise ValueError(f"Unexpected text model architecture type in GGUF file: {arch_str!r}")
elif arch_str not in IMG_ARCH_LIST and not is_text_model:
raise ValueError(f"Unexpected architecture type in GGUF file: {arch_str!r}")
if compat:
logging.warning(f"Warning: This gguf model file is loaded in compatibility mode '{compat}' [arch:{arch_str}]")
# main loading loop
state_dict = {}
qtype_dict = {}
for sd_key, tensor in tensors:
tensor_name = tensor.name
# torch_tensor = torch.from_numpy(tensor.data) # mmap
# NOTE: line above replaced with this block to avoid persistent numpy warning about mmap
with warnings.catch_warnings():
warnings.filterwarnings("ignore", message="The given NumPy array is not writable")
torch_tensor = torch.from_numpy(tensor.data) # mmap
shape = get_orig_shape(reader, tensor_name)
if shape is None:
shape = torch.Size(tuple(int(v) for v in reversed(tensor.shape)))
# Workaround for stable-diffusion.cpp SDXL detection.
if compat == "sd.cpp" and arch_str == "sdxl":
if any([tensor_name.endswith(x) for x in (".proj_in.weight", ".proj_out.weight")]):
while len(shape) > 2 and shape[-1] == 1:
shape = shape[:-1]
# add to state dict
if tensor.tensor_type in {gguf.GGMLQuantizationType.F32, gguf.GGMLQuantizationType.F16}:
torch_tensor = torch_tensor.view(*shape)
state_dict[sd_key] = GGMLTensor(torch_tensor, tensor_type=tensor.tensor_type, tensor_shape=shape)
# 1D tensors shouldn't be quantized, this is a fix for BF16
if len(shape) <= 1 and tensor.tensor_type == gguf.GGMLQuantizationType.BF16:
state_dict[sd_key] = dequantize_tensor(state_dict[sd_key], dtype=torch.float32)
# keep track of loaded tensor types
tensor_type_str = getattr(tensor.tensor_type, "name", repr(tensor.tensor_type))
qtype_dict[tensor_type_str] = qtype_dict.get(tensor_type_str, 0) + 1
# print loaded tensor type counts
logging.info("gguf qtypes: " + ", ".join(f"{k} ({v})" for k, v in qtype_dict.items()))
# mark largest tensor for vram estimation
qsd = {k:v for k,v in state_dict.items() if is_quantized(v)}
if len(qsd) > 0:
max_key = max(qsd.keys(), key=lambda k: qsd[k].numel())
state_dict[max_key].is_largest_weight = True
# extra info to return
extra = {
"arch_str": arch_str,
"metadata": get_gguf_metadata(reader)
}
return (state_dict, extra)
# for remapping llama.cpp -> original key names
T5_SD_MAP = {
"enc.": "encoder.",
".blk.": ".block.",
"token_embd": "shared",
"output_norm": "final_layer_norm",
"attn_q": "layer.0.SelfAttention.q",
"attn_k": "layer.0.SelfAttention.k",
"attn_v": "layer.0.SelfAttention.v",
"attn_o": "layer.0.SelfAttention.o",
"attn_norm": "layer.0.layer_norm",
"attn_rel_b": "layer.0.SelfAttention.relative_attention_bias",
"ffn_up": "layer.1.DenseReluDense.wi_1",
"ffn_down": "layer.1.DenseReluDense.wo",
"ffn_gate": "layer.1.DenseReluDense.wi_0",
"ffn_norm": "layer.1.layer_norm",
}
LLAMA_SD_MAP = {
"blk.": "model.layers.",
"attn_norm": "input_layernorm",
"attn_q_norm.": "self_attn.q_norm.",
"attn_k_norm.": "self_attn.k_norm.",
"attn_v_norm.": "self_attn.v_norm.",
"attn_q": "self_attn.q_proj",
"attn_k": "self_attn.k_proj",
"attn_v": "self_attn.v_proj",
"attn_output": "self_attn.o_proj",
"ffn_up": "mlp.up_proj",
"ffn_down": "mlp.down_proj",
"ffn_gate": "mlp.gate_proj",
"ffn_norm": "post_attention_layernorm",
"token_embd": "model.embed_tokens",
"output_norm": "model.norm",
"output.weight": "lm_head.weight",
}
GEMMA3_SD_MAP = LLAMA_SD_MAP.copy()
GEMMA3_SD_MAP.update({
"ffn_norm": "pre_feedforward_layernorm",
"post_ffw_norm": "post_feedforward_layernorm",
"post_attention_norm": "post_attention_layernorm",
})
CLIP_VISION_SD_MAP = {
"mm.": "visual.merger.mlp.",
"v.post_ln.": "visual.merger.ln_q.",
"v.patch_embd": "visual.patch_embed.proj",
"v.blk.": "visual.blocks.",
"ffn_up": "mlp.up_proj",
"ffn_down": "mlp.down_proj",
"ffn_gate": "mlp.gate_proj",
"attn_out.": "attn.proj.",
"ln1.": "norm1.",
"ln2.": "norm2.",
}
def sd_map_replace(raw_sd, key_map):
sd = {}
for k,v in raw_sd.items():
for s,d in key_map.items():
k = k.replace(s,d)
sd[k] = v
return sd
def llama_permute(raw_sd, n_head, n_head_kv):
# Reverse version of LlamaModel.permute in llama.cpp convert script
sd = {}
permute = lambda x,h: x.reshape(h, x.shape[0] // h // 2, 2, *x.shape[1:]).swapaxes(1, 2).reshape(x.shape)
for k,v in raw_sd.items():
if k.endswith(("q_proj.weight", "q_proj.bias")):
v.data = permute(v.data, n_head)
if k.endswith(("k_proj.weight", "k_proj.bias")):
v.data = permute(v.data, n_head_kv)
sd[k] = v
return sd
def gemma3_norm_corrections(sd):
# Reverse change from Gemma3Model modify_tensors in llama.cpp convert script
norm_patterns = [
"input_layernorm.weight",
"post_attention_layernorm.weight",
"pre_feedforward_layernorm.weight",
"post_feedforward_layernorm.weight",
"self_attn.q_norm.weight",
"self_attn.k_norm.weight",
"model.norm.weight"
]
corrected = 0
for key in list(sd.keys()):
if any(p in key for p in norm_patterns):
if is_quantized(sd[key]):
sd[key] = dequantize_tensor(sd[key], dtype=torch.float32) - 1.0
else:
sd[key] = sd[key].float() - 1.0
corrected += 1
#logging.info(f"Gemma3: Applied -1 norm correction to {corrected} tensors")
return sd
def strip_quant_suffix(name):
pattern = r"[-_]?(?:ud-)?i?q[0-9]_[a-z0-9_\-]{1,8}$"
match = re.search(pattern, name, re.IGNORECASE)
if match:
name = name[:match.start()]
return name
def gguf_mmproj_loader(path):
# Reverse version of Qwen2VLVisionModel.modify_tensors
logging.info("Attenpting to find mmproj file for text encoder...")
# get name to match w/o quant suffix
tenc_fname = os.path.basename(path)
tenc = os.path.splitext(tenc_fname)[0].lower()
tenc = strip_quant_suffix(tenc)
# try and find matching mmproj
target = []
root = os.path.dirname(path)
for fname in os.listdir(root):
name, ext = os.path.splitext(fname)
if ext.lower() != ".gguf":
continue
if "mmproj" not in name.lower():
continue
if tenc in name.lower():
target.append(fname)
if len(target) == 0:
logging.error(f"Error: Can't find mmproj file for '{tenc_fname}' (matching:'{tenc}')! Qwen-Image-Edit will be broken!")
return {}
if len(target) > 1:
logging.error(f"Ambiguous mmproj for text encoder '{tenc_fname}', will use first match.")
logging.info(f"Using mmproj '{target[0]}' for text encoder '{tenc_fname}'.")
target = os.path.join(root, target[0])
vsd, _ = gguf_sd_loader(target, is_text_model=True)
# concat 4D to 5D
if "v.patch_embd.weight.1" in vsd:
w1 = dequantize_tensor(vsd.pop("v.patch_embd.weight"), dtype=torch.float32)
w2 = dequantize_tensor(vsd.pop("v.patch_embd.weight.1"), dtype=torch.float32)
vsd["v.patch_embd.weight"] = torch.stack([w1, w2], dim=2)
# run main replacement
vsd = sd_map_replace(vsd, CLIP_VISION_SD_MAP)
# handle split Q/K/V
if "visual.blocks.0.attn_q.weight" in vsd:
attns = {}
# filter out attentions + group
for k,v in vsd.items():
if any(x in k for x in ["attn_q", "attn_k", "attn_v"]):
k_attn, k_name = k.rsplit(".attn_", 1)
k_attn += ".attn.qkv." + k_name.split(".")[-1]
if k_attn not in attns:
attns[k_attn] = {}
attns[k_attn][k_name] = dequantize_tensor(
v, dtype=(torch.bfloat16 if is_quantized(v) else torch.float16)
)
# recombine
for k,v in attns.items():
suffix = k.split(".")[-1]
vsd[k] = torch.cat([
v[f"q.{suffix}"],
v[f"k.{suffix}"],
v[f"v.{suffix}"],
], dim=0)
del attns
return vsd
def gguf_tokenizer_loader(path, temb_shape):
# convert gguf tokenizer to spiece
logging.info("Attempting to recreate sentencepiece tokenizer from GGUF file metadata...")
try:
from sentencepiece import sentencepiece_model_pb2 as model
except ImportError:
raise ImportError("Please make sure sentencepiece and protobuf are installed.\npip install sentencepiece protobuf")
spm = model.ModelProto()
reader = gguf.GGUFReader(path)
if get_field(reader, "tokenizer.ggml.model", str) == "t5":
if temb_shape == (256384, 4096): # probably UMT5
spm.trainer_spec.model_type == 1 # Unigram (do we have a T5 w/ BPE?)
else:
raise NotImplementedError("Unknown model, can't set tokenizer!")
else:
raise NotImplementedError("Unknown model, can't set tokenizer!")
spm.normalizer_spec.add_dummy_prefix = get_field(reader, "tokenizer.ggml.add_space_prefix", bool)
spm.normalizer_spec.remove_extra_whitespaces = get_field(reader, "tokenizer.ggml.remove_extra_whitespaces", bool)
tokens = get_list_field(reader, "tokenizer.ggml.tokens", str)
scores = get_list_field(reader, "tokenizer.ggml.scores", float)
toktypes = get_list_field(reader, "tokenizer.ggml.token_type", int)
for idx, (token, score, toktype) in enumerate(zip(tokens, scores, toktypes)):
# # These aren't present in the original?
# if toktype == 5 and idx >= temb_shape[0]%1000):
# continue
piece = spm.SentencePiece()
piece.piece = token
piece.score = score
piece.type = toktype
spm.pieces.append(piece)
# unsure if any of these are correct
spm.trainer_spec.byte_fallback = True
spm.trainer_spec.vocab_size = len(tokens) # split off unused?
spm.trainer_spec.max_sentence_length = 4096
spm.trainer_spec.eos_id = get_field(reader, "tokenizer.ggml.eos_token_id", int)
spm.trainer_spec.pad_id = get_field(reader, "tokenizer.ggml.padding_token_id", int)
logging.info(f"Created tokenizer with vocab size of {len(spm.pieces)}")
del reader
return torch.ByteTensor(list(spm.SerializeToString()))
def gguf_tekken_tokenizer_loader(path, temb_shape):
# convert ggml (hf) tokenizer metadata to tekken/comfy data
logging.info("Attempting to recreate tekken tokenizer from GGUF file metadata...")
import json
import base64
from transformers.convert_slow_tokenizer import bytes_to_unicode
reader = gguf.GGUFReader(path)
model_str = get_field(reader, "tokenizer.ggml.model", str)
if model_str == "gpt2":
if temb_shape == (131072, 5120): # probably Mistral
data = {
"config": {"num_vocab_tokens": 150000, "default_vocab_size": 131072},
"vocab": [],
"special_tokens": [],
}
else:
raise NotImplementedError("Unknown model, can't set tokenizer!")
else:
raise NotImplementedError("Unknown model, can't set tokenizer!")
tokens = get_list_field(reader, "tokenizer.ggml.tokens", str)
toktypes = get_list_field(reader, "tokenizer.ggml.token_type", int)
decoder = {v: k for k, v in bytes_to_unicode().items()}
for idx, (token, toktype) in enumerate(zip(tokens, toktypes)):
if toktype == 3:
data["special_tokens"].append(
{'rank': idx, 'token_str': token, 'is_control': True}
)
else:
tok = bytes([decoder[char] for char in token])
data["vocab"].append({
"rank": len(data["vocab"]),
"token_bytes": base64.b64encode(tok).decode("ascii"),
"token_str": tok.decode("utf-8", errors="replace") # ?
})
logging.info(f"Created tekken tokenizer with vocab size of {len(data['vocab'])} (+{len(data['special_tokens'])})")
del reader
return torch.ByteTensor(list(json.dumps(data).encode('utf-8')))
def gguf_gemma3_tokenizer_loader(path):
#TODO: merge into gguf_tokenizer_loader
logging.info("Attempting to recreate sentencepiece tokenizer from GGUF file metadata...")
try:
from sentencepiece import sentencepiece_model_pb2 as model
except ImportError:
raise ImportError("Please install sentencepiece and protobuf.\npip install sentencepiece protobuf")
spm = model.ModelProto()
reader = gguf.GGUFReader(path)
spm.normalizer_spec.name = "identity"
spm.normalizer_spec.add_dummy_prefix = False
spm.trainer_spec.model_type = 2
spm.trainer_spec.input_format = "tsv"
spm.trainer_spec.byte_fallback = True
spm.trainer_spec.max_sentence_length = 4192
spm.trainer_spec.bos_piece = "<bos>"
tokens = get_list_field(reader, "tokenizer.ggml.tokens", str)
scores = get_list_field(reader, "tokenizer.ggml.scores", float)
toktype = get_list_field(reader, "tokenizer.ggml.token_type", int)
if not tokens or not scores or not toktype:
raise ValueError("Missing tokenizer metadata")
for idx in range(len(tokens)):
piece = spm.SentencePiece()
piece.piece = tokens[idx]
if idx == 3: # UNK position
piece.type = 2 # UNK Token
piece.score = 0.0 # UNK Score
else:
piece.type = toktype[idx]
piece.score = scores[idx]
spm.pieces.append(piece)
spm.trainer_spec.vocab_size = len(spm.pieces)
logging.info(f"Created tokenizer with vocab size of {len(spm.pieces)}")
del reader
return torch.ByteTensor(list(spm.SerializeToString()))
def gguf_clip_loader(path):
sd, extra = gguf_sd_loader(path, is_text_model=True)
arch = extra.get("arch_str", None)
if arch in {"t5", "t5encoder"}:
temb_key = "token_embd.weight"
if temb_key in sd and sd[temb_key].shape == (256384, 4096):
# non-standard Comfy-Org tokenizer
sd["spiece_model"] = gguf_tokenizer_loader(path, sd[temb_key].shape)
# TODO: dequantizing token embed here is janky but otherwise we OOM due to tensor being massive.
logging.warning(f"Dequantizing {temb_key} to prevent runtime OOM.")
sd[temb_key] = dequantize_tensor(sd[temb_key], dtype=torch.float16)
sd = sd_map_replace(sd, T5_SD_MAP)
elif arch in {"llama", "qwen2vl", "qwen3", "qwen3vl", "gemma3"}:
# TODO: pass model_options["vocab_size"] to loader somehow
temb_key = "token_embd.weight"
if temb_key in sd and sd[temb_key].shape[0] >= (64 * 1024):
if arch == "llama" and sd[temb_key].shape == (131072, 5120):
# non-standard Comfy-Org tokenizer
sd["tekken_model"] = gguf_tekken_tokenizer_loader(path, sd[temb_key].shape)
elif arch == "gemma3":
sd["spiece_model"] = gguf_gemma3_tokenizer_loader(path)
# See note above for T5.
logging.warning(f"Dequantizing {temb_key} to prevent runtime OOM.")
sd[temb_key] = dequantize_tensor(sd[temb_key], dtype=torch.float16)
if arch == "gemma3":
sd = sd_map_replace(sd, GEMMA3_SD_MAP)
sd = gemma3_norm_corrections(sd)
else:
sd = sd_map_replace(sd, LLAMA_SD_MAP)
if arch == "llama":
sd = llama_permute(sd, 32, 8) # L3 / Mistral
if arch == "qwen2vl":
vsd = gguf_mmproj_loader(path)
sd.update(vsd)
else:
pass
return sd
You can’t perform that action at this time.