File size: 40,232 Bytes
b701455
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
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
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
"""Flux2 Klein model adapter for LightDiffusion-Next.

Provides a clean interface to the Flux2 Klein 4B model that inherits from
AbstractModel and integrates with the LightDiffusion-Next model factory.

This implementation uses ONLY native LightDiffusion-Next components,
without any ComfyUI imports.

File structure expected:
    - include/diffusion_model/flux-2-klein-4b.safetensors  (or similar)
    - include/text_encoder/qwen_3_4b.safetensors
    - include/vae/ae.safetensors (Flux VAE)
"""

import logging
import os
from typing import TYPE_CHECKING, Any, Callable, Optional

import torch

from src.Core.AbstractModel import AbstractModel, ModelCapabilities
from src.Utilities import util
from src.Device import Device

# Import modules that were previously lazy-loaded inside methods
# This avoids KeyError: 'src' when running via uv run streamlit
from src.NeuralNetwork.flux2.model import Flux2, Flux2Params
from src.Model.ModelPatcher import ModelPatcher
from src.clip.KleinEncoder import KleinCLIP, Qwen3_4BModel
from src.AutoEncoders import VariationalAE
from src.sample import sampling
from src.Utilities import Latent
from src.Model import LoRas

if TYPE_CHECKING:
    from src.Core.Context import Context


logger = logging.getLogger(__name__)

# Default paths for Flux2 Klein components
DEFAULT_DIFFUSION_MODEL_DIR = "./include/diffusion_model"
DEFAULT_TEXT_ENCODER_DIR = "./include/text_encoder"  
DEFAULT_VAE_DIR = "./include/vae"


class Flux2KleinModel(AbstractModel):
    """Flux2 Klein 4B model implementation.
    
    Wraps the Flux2 Klein model with the clean AbstractModel interface
    for use with the LightDiffusion-Next pipeline system.
    
    The Flux2 Klein model is a distilled version of the Flux2 architecture
    using the Klein (Qwen3 4B) text encoder.
    
    Unlike SD1.5/SDXL which use combined checkpoints, Flux2 Klein loads
    components separately:
    - Diffusion model from include/diffusion_model/
    - Text encoder (Qwen3 4B) from include/text_encoder/
    - VAE from include/vae/
    """
    
    def __init__(
        self, 
        model_path: str = None,
        text_encoder_path: str = None,
        vae_path: str = None,
        quantization: str = None,  # "fp8", "nvfp4", or None
    ):
        """Initialize the Flux2 Klein model adapter.
        
        Args:
            model_path: Path to diffusion model (safetensors)
            text_encoder_path: Path to Qwen3 text encoder (optional, auto-detected)
            vae_path: Path to VAE (optional, auto-detected)
            quantization: Quantization format to use ("fp8", "nvfp4", or None)
        """
        super().__init__(model_path)
        self._text_encoder = None
        self._tokenizer = None
        self._model_config = None
        self._text_encoder_path = text_encoder_path
        self._vae_path = vae_path
        self._raw_model = None  # The raw Flux2 nn.Module
        self.quantization = quantization
        
        # Device management
        self.load_device = Device.get_torch_device()
        self.offload_device = torch.device("cpu")
    
    def _create_capabilities(self) -> ModelCapabilities:
        """Create capabilities for Flux2 Klein model."""
        return ModelCapabilities(
            min_resolution=256,
            max_resolution=4096,
            preferred_resolution=1024,
            requires_resolution_multiple=16,  # Flux2 uses 16-pixel patches
            supports_hires_fix=True, 
            supports_img2img=True,
            supports_inpainting=False,  # Not yet implemented for Flux2
            supports_controlnet=False,  # ControlNet support pending
            supports_stable_fast=False,  # May need special handling
            supports_deepcache=False,  # Architecture differs from UNet
            supports_tome=False,  # Token merging needs special implementation
            supports_lora=False,  # Flux2 LoRA format differs from SD
            uses_dual_clip=False,  # Uses single Klein (Qwen3) encoder
            requires_size_conditioning=False,
            is_flux=True,
            is_flux2=True,
        )

    def _find_diffusion_model(self) -> Optional[str]:
        """Auto-detect Flux2 diffusion model in default directory."""
        if os.path.exists(DEFAULT_DIFFUSION_MODEL_DIR):
            for f in os.listdir(DEFAULT_DIFFUSION_MODEL_DIR):
                f_lower = f.lower()
                if ("flux" in f_lower or "klein" in f_lower) and f.endswith((".safetensors", ".pt", ".pth")):
                    return os.path.join(DEFAULT_DIFFUSION_MODEL_DIR, f)
        return None

    def _find_text_encoder(self) -> Optional[str]:
        """Auto-detect Qwen3 text encoder in default directory."""
        if os.path.exists(DEFAULT_TEXT_ENCODER_DIR):
            for f in os.listdir(DEFAULT_TEXT_ENCODER_DIR):
                f_lower = f.lower()
                if ("qwen" in f_lower or "klein" in f_lower) and f.endswith((".safetensors", ".pt", ".pth")):
                    return os.path.join(DEFAULT_TEXT_ENCODER_DIR, f)
        return None

    def _find_vae(self) -> Optional[str]:
        """Auto-detect VAE in default directory."""
        if os.path.exists(DEFAULT_VAE_DIR):
            # Look for Flux-compatible VAE (ae.safetensors)
            for f in os.listdir(DEFAULT_VAE_DIR):
                if f.endswith((".safetensors", ".pt", ".pth")):
                    return os.path.join(DEFAULT_VAE_DIR, f)
        return None
    
    def load(self, model_path: str = None) -> "Flux2KleinModel":
        """Load the Flux2 Klein model components from disk.
        
        Components are loaded separately:
        - Diffusion model (Flux2 transformer)
        - Text encoder (Qwen3 4B via Klein tokenizer)
        - VAE
        
        Args:
            model_path: Optional override for the diffusion model path
            
        Returns:
            Self for method chaining
        """
        # Resolve paths
        diffusion_path = model_path or self.model_path or self._find_diffusion_model()
        
        # Guard: Don't reload if already loaded with same diffusion model
        if self._loaded and self.model_path == diffusion_path:
            logger.info("Flux2KleinModel: Already loaded, skipping redundant load")
            return self
        
        if diffusion_path is None:
            raise ValueError(
                "No Flux2 diffusion model found. Please place the model in "
                f"{DEFAULT_DIFFUSION_MODEL_DIR}/ with 'flux' or 'klein' in the filename."
            )
        
        self.model_path = diffusion_path
        
        # Resolve other paths only when loading is actually needed
        text_encoder_path = self._text_encoder_path or self._find_text_encoder()
        vae_path = self._vae_path or self._find_vae()
        
        logger.info(f"Flux2KleinModel: Loading components...")
        logger.info(f"  Diffusion model: {diffusion_path}")
        logger.info(f"  Text encoder: {text_encoder_path}")
        logger.info(f"  VAE: {vae_path}")
        
        try:
            # Load diffusion model
            # self.model = self._load_diffusion_model(diffusion_path) # Original line
            
            # New FP8 loading logic
            from src.NeuralNetwork.flux2.model import create_flux2_klein
            from src.Device import Device
            from src.FileManaging import Loader
            
            # Check for FP8 support and user preference/environment
            use_fp8 = Device.is_fp8_supported(self.load_device)
            # For 8GB cards, we force FP8 for Flux2 Klein 4B to avoid swapping
            total_vram = Device.get_total_memory(self.load_device) / (1024**3)
            if total_vram < 12.0: # If less than 12GB, FP8 is highly recommended for Flux
                use_fp8 = use_fp8 and True
            
            dtype = torch.bfloat16 # Base weight dtype
            
            # Create model with detected config
            config = self._detect_flux2_config(util.load_torch_file(diffusion_path, device=torch.device("cpu"))) # Load temporarily to detect config
            params = Flux2Params(**config)
            self.model = Flux2(params=params, dtype=dtype, device=torch.device("cpu")) # Create on CPU first
            self.model.eval()
            
            # Attach config for compatibility
            self._model_config = self._create_model_config() # Ensure _model_config is set
            
            # Load weights
            sd = util.load_torch_file(diffusion_path, device=self.offload_device)
            # Sanitize NaN values in weights (some Flux2 checkpoints have NaN biases)
            nan_keys = []
            for key, value in sd.items():
                if isinstance(value, torch.Tensor) and torch.isnan(value).any():
                    nan_keys.append(key)
                    sd[key] = torch.where(torch.isnan(value), torch.zeros_like(value), value)
            if nan_keys:
                logger.warning(f"Sanitized NaN values in {len(nan_keys)} keys: {nan_keys[:5]}...")
            
            self.model.load_state_dict(sd, strict=False)
            del sd
            
            self._raw_model = self.model # Store raw model
            
            # Create ModelPatcher
            self.model = ModelPatcher(self.model, self.load_device, self.offload_device)
            
            # Apply quantization if requested or needed
            quant_format = self.quantization
            if quant_format is None and use_fp8:
                quant_format = "fp8"
                
            if quant_format == "nvfp4":
                logging.info("Flux2: Applying NVFP4 (4-bit) weight-only quantization")
                self.model.weight_only_quantize("nvfp4")
                self.model.model_dtype = lambda: torch.float16 # Compute in FP16 for dequantization
            elif quant_format == "fp8":
                logging.info("Flux2: Applying FP8 weight-only quantization")
                self.model.weight_only_quantize(torch.float8_e4m3fn)
                self.model.model_dtype = lambda: torch.float8_e4m3fn # Override
            
            # Load text encoder
            if text_encoder_path:
                self.clip = self._load_klein_text_encoder(text_encoder_path, quantize=quant_format)
                self._text_encoder = self.clip # For internal reference
                self._tokenizer = self.clip.tokenizer
            else:
                logger.warning("No Qwen3 text encoder found - prompt encoding may fail")
                self.clip = None
            
            # Load VAE
            if vae_path:
                self.vae = self._load_vae(vae_path)
            else:
                logger.warning("No VAE found - image decoding may fail")
                self.vae = None
            
            # Store config for sampling
            self._model_config = self._create_model_config()
            
            # Attach model_sampling for sampler infrastructure
            from src.sample import sampling
            self.model.model_sampling = sampling.model_sampling(self._model_config, "flux2", flux=True, flux2=True)
            
            self._loaded = True
            logger.info(f"Flux2KleinModel: Successfully loaded all components")
            
        except Exception as e:
            logger.exception(f"Flux2KleinModel: Failed to load: {e}")
            raise
        
        return self

    def _load_diffusion_model(self, path: str):
        """Load the Flux2 diffusion model using native LightDiffusion-Next.
        
        Args:
            path: Path to diffusion model safetensors
            
        Returns:
            ModelPatcher wrapping the Flux2 model
        """
        logger.info(f"Loading Flux2 diffusion model: {path}")
        
        # Load state dict using native utility
        sd = util.load_torch_file(path)
        
        # Sanitize NaN values in weights (some Flux2 checkpoints have NaN biases)
        nan_keys = []
        for key, value in sd.items():
            if isinstance(value, torch.Tensor) and torch.isnan(value).any():
                nan_keys.append(key)
                sd[key] = torch.where(torch.isnan(value), torch.zeros_like(value), value)
        if nan_keys:
            logger.warning(f"Sanitized NaN values in {len(nan_keys)} keys: {nan_keys[:5]}...")
        
        # Detect model configuration from state dict
        config = self._detect_flux2_config(sd)
        
        # Determine dtype and device
        load_device = Device.get_torch_device()
        offload_device = Device.unet_offload_device()
        
        # Infer dtype from weights
        dtype = torch.bfloat16
        for k, v in sd.items():
            if isinstance(v, torch.Tensor) and v.dtype in (torch.float16, torch.bfloat16, torch.float32):
                dtype = v.dtype
                break
        
        logger.info(f"Flux2 model dtype: {dtype}")
        
        # Create model with detected config
        params = Flux2Params(**config)
        model = Flux2(params=params, dtype=dtype, device="cpu")
        
        # Attach config for compatibility
        model.model_config = self._create_model_config()
        
        # Load weights
        missing, unexpected = model.load_state_dict(sd, strict=False)
        if missing:
            logger.debug(f"Missing keys: {len(missing)}")
        if unexpected:
            logger.debug(f"Unexpected keys: {len(unexpected)}")
        
        self._raw_model = model
        
        # Wrap in ModelPatcher for compatibility with sampling infrastructure
        model_patcher = ModelPatcher.ModelPatcher(
            model,
            load_device=load_device,
            offload_device=offload_device,
            current_device=torch.device("cpu"),
        )
        
        return model_patcher

    def _detect_flux2_config(self, sd: dict) -> dict:
        """Detect Flux2 model configuration from state dict.
        
        Args:
            sd: Model state dictionary
            
        Returns:
            Configuration dict for Flux2Params
        """
        # Detect if this is Flux2 (has double_stream_modulation) or Flux1
        is_flux2 = any("double_stream_modulation" in k for k in sd.keys())
        
        if is_flux2:
            # Flux2 / Klein defaults (patch_size=1 unlike Flux1!)
            config = {
                "patch_size": 1,  # CRITICAL: Flux2 uses patch_size=1 (no spatial patchification)
                "in_channels": 128,  # Direct channel input (no patch_size division)
                "out_channels": 128,  # Direct channel output
                "vec_in_dim": 768,
                "context_in_dim": 7680,  # Klein uses concatenated multi-layer output
                "hidden_size": 3072,
                "mlp_ratio": 3.0,  # Klein uses 3.0 with gated MLP
                "num_heads": 24,  # Flux2: hidden_size/sum(axes_dim) = 3072/128 = 24
                "depth": 19,
                "depth_single_blocks": 38,
                "axes_dim": [32, 32, 32, 32],  # Flux2 specific - sum=128
                "theta": 2000,  # Flux2 uses lower theta
                "qkv_bias": False,
                "guidance_embed": False,
                "gated_mlp": True,  # Klein uses gated MLP (SwiGLU)
                "global_modulation": True,  # Flux2 feature
                "mlp_silu_act": True,  # Flux2 feature
                "ops_bias": False,  # Flux2 feature
                "use_vector_in": False,  # Flux2/Klein doesn't use pooled conditioning
            }
            logger.info("Detected Flux2 model (has double_stream_modulation)")
        else:
            # Flux1 defaults
            config = {
                "in_channels": 16,
                "out_channels": 16,
                "vec_in_dim": 768,
                "context_in_dim": 7680,
                "hidden_size": 3072,
                "mlp_ratio": 4.0,
                "num_heads": 24,
                "depth": 19,
                "depth_single_blocks": 38,
                "axes_dim": [16, 56, 56],  # Flux1 specific
                "theta": 10000,
                "qkv_bias": True,
                "guidance_embed": True,
                "gated_mlp": False,
            }
            logger.info("Detected Flux1 model")
        
        # Detect depth from double_blocks
        double_blocks = [k for k in sd.keys() if "double_blocks" in k]
        if double_blocks:
            max_block = max(
                int(k.split("double_blocks.")[1].split(".")[0])
                for k in double_blocks
                if "double_blocks." in k
            )
            config["depth"] = max_block + 1
        
        # Detect single blocks depth
        single_blocks = [k for k in sd.keys() if "single_blocks" in k]
        if single_blocks:
            max_single = max(
                int(k.split("single_blocks.")[1].split(".")[0])
                for k in single_blocks
                if "single_blocks." in k
            )
            config["depth_single_blocks"] = max_single + 1
        
        # Detect hidden size and in_channels from img_in
        if "img_in.weight" in sd:
            config["hidden_size"] = sd["img_in.weight"].shape[0]
            # img_in input dim = in_channels * patch_size^2
            # For Flux2 with patch_size=1: in_channels = img_in_dim directly
            img_in_dim = sd["img_in.weight"].shape[1]
            patch_size = config.get("patch_size", 2)
            config["in_channels"] = img_in_dim // (patch_size ** 2)
            logger.info(f"Detected in_channels={config['in_channels']} from img_in (patch_size={patch_size})")
        
        # Detect out_channels from final_layer
        if "final_layer.linear.weight" in sd:
            # final_layer.linear maps hidden -> patch_size * patch_size * out_channels
            # For Flux2 with patch_size=1: out_channels = final.shape[0] directly
            final_out = sd["final_layer.linear.weight"].shape[0]
            patch_size = config.get("patch_size", 2)
            config["out_channels"] = final_out // (patch_size ** 2)
            logger.info(f"Detected out_channels={config['out_channels']} from final_layer")
        
        # Detect mlp_ratio and gated_mlp from double_blocks MLP weights
        # For gated MLP: img_mlp.0 maps hidden -> 2*intermediate (gate+up)
        #                img_mlp.2 maps intermediate -> hidden
        # So: mlp_0_out = 2 * intermediate, intermediate = mlp_2_in
        # mlp_ratio = intermediate / hidden
        if "double_blocks.0.img_mlp.0.weight" in sd and "double_blocks.0.img_mlp.2.weight" in sd:
            mlp_0_out = sd["double_blocks.0.img_mlp.0.weight"].shape[0]
            mlp_2_in = sd["double_blocks.0.img_mlp.2.weight"].shape[1]
            hidden = config["hidden_size"]
            
            # Check if it's gated MLP: mlp_0_out should be 2 * mlp_2_in
            if abs(mlp_0_out - 2 * mlp_2_in) < 10:  # Small tolerance
                # Gated MLP detected
                config["gated_mlp"] = True
                intermediate = mlp_2_in
                config["mlp_ratio"] = intermediate / hidden
                logger.info(f"Detected gated MLP: intermediate={intermediate}, mlp_ratio={config['mlp_ratio']}")
            else:
                # Standard MLP: mlp_0_out = mlp_2_in = hidden * mlp_ratio
                config["gated_mlp"] = False
                config["mlp_ratio"] = mlp_0_out / hidden
        
        # Calculate num_heads from hidden_size and axes_dim (ComfyUI approach)
        # num_heads = hidden_size // sum(axes_dim)
        axes_sum = sum(config["axes_dim"])
        config["num_heads"] = config["hidden_size"] // axes_sum
        logger.info(f"Calculated num_heads={config['num_heads']} from hidden_size={config['hidden_size']} / axes_sum={axes_sum}")
        
        # Detect context_in_dim from txt_in
        if "txt_in.weight" in sd:
            config["context_in_dim"] = sd["txt_in.weight"].shape[1]
        
        # Detect vec_in_dim from vector_in
        if "vector_in.in_layer.weight" in sd:
            config["vec_in_dim"] = sd["vector_in.in_layer.weight"].shape[1]
            config["use_vector_in"] = True  # Enable vector_in if weights exist
            logger.info(f"Detected vector_in with dim {config['vec_in_dim']}")
        
        # Detect guidance embedding
        if any("guidance_in" in k for k in sd.keys()):
            config["guidance_embed"] = True
            
        # Detect txt_norm (critical for some Flux2 variants)
        if any("txt_norm.scale" in k for k in sd.keys()):
            config["txt_norm"] = True
            logger.info("Detected txt_norm in model weights")
        
        logger.info(f"Detected Flux2 config: depth={config['depth']}, "
                   f"single_blocks={config['depth_single_blocks']}, "
                   f"hidden={config['hidden_size']}, mlp_ratio={config['mlp_ratio']}, "
                   f"gated_mlp={config.get('gated_mlp', False)}")
        
        return config

    def _load_klein_text_encoder(self, path: str, quantize: str = None):
        """Load the Klein (Qwen3-4B) text encoder.
        
        Args:
            path: Path to text encoder safetensors
            quantize: Quantization format ("fp8", "nvfp4", or None)
            
        Returns:
            KleinCLIP wrapper
        """
        logger.info(f"Loading Text Encoder: {path}")
        from src.clip.KleinEncoder import KleinCLIP, KleinTokenizer, Qwen3_4BModel, get_ops
        from src.Model.ModelPatcher import ModelPatcher
        
        # Determine paths
        sd_path = path
        tokenizer_path = os.path.join(os.path.dirname(path), "qwen25_tokenizer")
        if not os.path.exists(tokenizer_path):
             tokenizer_path = None # Let KleinTokenizer find its default
             
        # Load weights
        sd = util.load_torch_file(sd_path, device=torch.device("cpu"))
        
        # Create model structure
        # Base dtype is BF16
        dtype = torch.bfloat16
        model = Qwen3_4BModel(dtype=dtype, device="cpu")
        
        # Load state dict
        model_sd = {}
        for k, v in sd.items():
            if k.startswith("model."):
                model_sd[k[6:]] = v
            else:
                model_sd[k] = v
        
        missing, unexpected = model.load_state_dict(model_sd, strict=False)
        
        # Apply quantization BEFORE moving to offload device if requested
        if quantize:
            logger.info(f"Flux2KleinModel: Quantizing Klein (Qwen3-4B) to {quantize}")
            # We must use ModelPatcher to correctly update comfy_cast_weights flags
            te_patcher = ModelPatcher(model, self.load_device, self.offload_device)
            
            if quantize == "nvfp4":
                te_patcher.weight_only_quantize("nvfp4")
            else:
                te_patcher.weight_only_quantize(torch.float8_e4m3fn)
                
            model = te_patcher.model

        # IMPORTANT: Keep model on CPU to save VRAM for diffusion model
        offload_device = Device.text_encoder_offload_device()
        model = model.to(offload_device)
        
        # Create wrapper
        tokenizer = KleinTokenizer(tokenizer_path)
        clip = KleinCLIP(tokenizer=tokenizer, model=model, dtype=dtype, device=self.load_device, offload_device=offload_device)
        
        return clip

    def _load_vae(self, path: str):
        """Load the VAE for decoding latents using native LightDiffusion-Next.
        
        Following ComfyUI's VAE loading approach:
        - Detects z_channels from decoder.conv_in.weight.shape[1]
        - Uses post_quant_conv/quant_conv (flux=False) for standard VAE structure
        
        Args:
            path: Path to VAE safetensors
            
        Returns:
            VAE model
        """
        logger.info(f"Loading VAE: {path}")
        
        # Load state dict
        sd = util.load_torch_file(path)
        
        # Check for diffusers format and convert if needed (ComfyUI approach)
        if 'decoder.up_blocks.0.resnets.0.norm1.weight' in sd:
            logger.info("Converting diffusers VAE format to SD format")
            sd = self._convert_diffusers_vae(sd)
        
        # Log VAE structure
        is_flux_vae = False
        if 'decoder.conv_in.weight' in sd:
            z_ch = sd['decoder.conv_in.weight'].shape[1]
            logger.info(f"VAE z_channels: {z_ch}")
        
        if 'post_quant_conv.weight' in sd:
            embed_dim = sd['post_quant_conv.weight'].shape[1]
            logger.info(f"VAE embed_dim: {embed_dim} (Standard VAE)")
            is_flux_vae = False
        else:
            logger.info("VAE missing post_quant_conv (Flux VAE)")
            is_flux_vae = True
        
        # Create VAE using native implementation
        # Set flux=True if it's a Flux VAE (skips post_quant_conv)
        # Use bfloat16 for better precision/memory balance on modern GPUs
        vae = VariationalAE.VAE(sd=sd, flux=is_flux_vae, dtype=torch.bfloat16)
        
        return vae
    
    def _convert_diffusers_vae(self, sd: dict) -> dict:
        """Convert diffusers VAE format to SD format (ComfyUI approach)."""
        # VAE conversion map from ComfyUI's diffusers_convert.py
        vae_conversion_map = [
            ("nin_shortcut", "conv_shortcut"),
            ("norm_out", "conv_norm_out"),
            ("mid.attn_1.", "mid_block.attentions.0."),
        ]
        
        for i in range(4):
            for j in range(2):
                hf_down_prefix = f"encoder.down_blocks.{i}.resnets.{j}."
                sd_down_prefix = f"encoder.down.{i}.block.{j}."
                vae_conversion_map.append((sd_down_prefix, hf_down_prefix))
            
            if i < 3:
                hf_downsample_prefix = f"down_blocks.{i}.downsamplers.0."
                sd_downsample_prefix = f"down.{i}.downsample."
                vae_conversion_map.append((sd_downsample_prefix, hf_downsample_prefix))
                
                hf_upsample_prefix = f"up_blocks.{i}.upsamplers.0."
                sd_upsample_prefix = f"up.{3 - i}.upsample."
                vae_conversion_map.append((sd_upsample_prefix, hf_upsample_prefix))
            
            for j in range(3):
                hf_up_prefix = f"decoder.up_blocks.{i}.resnets.{j}."
                sd_up_prefix = f"decoder.up.{3 - i}.block.{j}."
                vae_conversion_map.append((sd_up_prefix, hf_up_prefix))
        
        for i in range(2):
            hf_mid_res_prefix = f"mid_block.resnets.{i}."
            sd_mid_res_prefix = f"mid.block_{i + 1}."
            vae_conversion_map.append((sd_mid_res_prefix, hf_mid_res_prefix))
        
        vae_conversion_map_attn = [
            ("norm.", "group_norm."),
            ("q.", "query."), ("k.", "key."), ("v.", "value."),
            ("q.", "to_q."), ("k.", "to_k."), ("v.", "to_v."),
            ("proj_out.", "to_out.0."), ("proj_out.", "proj_attn."),
        ]
        
        mapping = {k: k for k in sd.keys()}
        for k, v in mapping.items():
            for sd_part, hf_part in vae_conversion_map:
                v = v.replace(hf_part, sd_part)
            mapping[k] = v
        
        for k, v in mapping.items():
            if "attentions" in k:
                for sd_part, hf_part in vae_conversion_map_attn:
                    v = v.replace(hf_part, sd_part)
                mapping[k] = v
        
        new_state_dict = {v: sd[k] for k, v in mapping.items()}
        
        # Reshape attention weights
        weights_to_convert = ["q", "k", "v", "proj_out"]
        for k, v in new_state_dict.items():
            for weight_name in weights_to_convert:
                if f"mid.attn_1.{weight_name}.weight" in k:
                    new_state_dict[k] = v.reshape(*v.shape, 1, 1)
        
        return new_state_dict
    
    def _create_model_config(self):
        """Create a model config object for sampling."""
        class Flux2KleinConfig:
            """Configuration for Flux2 Klein sampling."""
            sampling_settings = {
                "shift": 2.02,  # Flux2 default shift (different from Flux1's 1.15)
            }
            latent_format = Latent.Flux2()
            recommended_steps = 4
            recommended_cfg = 1.0
        
        return Flux2KleinConfig()
    
    def encode_prompt(
        self,
        prompt: str | list[str],
        negative_prompt: str | list[str] = "",
        clip_skip: int = None,
    ) -> tuple[Any, Any]:
        """Encode text prompts into conditioning tensors.
        
        For Flux2 Klein, this uses the Qwen3-based Klein text encoder
        which does not use traditional CLIP skip.
        
        CRITICAL: ComfyUI LEFT-PADS text embeddings to 512 tokens before passing
        to the diffusion model. This is essential for matching image quality because:
        1. The positional encoding (RoPE) depends on sequence length
        2. The model was trained with fixed 512-token text sequences
        
        Args:
            prompt: Positive prompt(s) to encode
            negative_prompt: Negative prompt(s) (may be ignored for Flux2)
            clip_skip: Not used for Klein encoder
            
        Returns:
            Tuple of (positive_conditioning, negative_conditioning)
        """
        if not self._loaded:
            raise RuntimeError("Model must be loaded before encoding prompts")
        
        if self.clip is None:
            raise RuntimeError("No text encoder loaded")
        
        try:
            import torch
            
            # Use Klein encoder directly
            if isinstance(prompt, list):
                # Encode each prompt in the batch
                all_hidden = []
                all_pooled = []
                for p in prompt:
                    tokens = self.clip.tokenizer.tokenize_with_weights(p)
                    h, pol, _ = self.clip.encode_token_weights(tokens)
                    all_hidden.append(h)
                    # Handle cases where pooled output might be None (common in Klein/Qwen encoders)
                    if pol is not None:
                        all_pooled.append(pol)
                
                hidden_states = torch.cat(all_hidden, dim=0)
                pooled = torch.cat(all_pooled, dim=0) if all_pooled else None
            else:
                # Single prompt
                tokens = self.clip.tokenizer.tokenize_with_weights(prompt)
                hidden_states, pooled, extra = self.clip.encode_token_weights(tokens)
            
            # Encode negative (or empty)
            neg_prompt = negative_prompt
            if neg_prompt:
                if isinstance(neg_prompt, list):
                    # We usually only need one negative for the whole batch or match batch size
                    if len(neg_prompt) == 1:
                        neg_prompt = neg_prompt[0]
                    else:
                        # Encode all negatives
                        all_neg_hidden = []
                        all_neg_pooled = []
                        for np in neg_prompt:
                            ntokens = self.clip.tokenizer.tokenize_with_weights(np)
                            nh, npol, _ = self.clip.encode_token_weights(ntokens)
                            all_neg_hidden.append(nh)
                            if npol is not None:
                                all_neg_pooled.append(npol)
                        neg_hidden = torch.cat(all_neg_hidden, dim=0)
                        neg_pooled = torch.cat(all_neg_pooled, dim=0) if all_neg_pooled else None
                        neg_prompt = None # Mark as processed
            
            if neg_prompt is not None:
                neg_tokens = self.clip.tokenizer.tokenize_with_weights(neg_prompt or "")
                neg_hidden, neg_pooled, neg_extra = self.clip.encode_token_weights(neg_tokens)
            
            # Embeddings are already padded to 512 tokens by the tokenizer
            # Format as conditioning
            # Note: ComfyUI does NOT pass attention_mask to diffusion model for Flux2
            # The zero-padded tokens don't contribute meaningfully to cross-attention
            cond_dict = {"pooled_output": pooled}
            positive = [[hidden_states, cond_dict]]
            
            neg_cond_dict = {"pooled_output": neg_pooled}
            negative = [[neg_hidden, neg_cond_dict]]
            
            return positive, negative
            
        except Exception as e:
            logger.exception(f"Prompt encoding failed: {e}")
            raise
    
    def generate(
        self,
        ctx: "Context",
        positive: Any,
        negative: Any,
        latent_image: Optional[Any] = None,
        start_step: Optional[int] = None,
        last_step: Optional[int] = None,
        disable_noise: bool = False,
        callback: Optional[Callable] = None,
    ) -> dict:
        """Generate latents using the Flux2 sampler.
        
        Args:
            ctx: Context with generation parameters
            positive: Positive conditioning
            negative: Negative conditioning (may be ignored)
            
        Returns:
            Dictionary with 'samples' key containing generated latents
        """
        if not self._loaded:
            raise RuntimeError("Model must be loaded before generating")
        
        # Log recommendation if CFG is high for this distilled model
        if ctx.sampling.cfg > 2.0:
            logger.info(f"Tip: Flux2 Klein works best with CFG 1.0. "
                       f"You are currently using CFG {ctx.sampling.cfg}.")
        
        try:
            # Use provided latent or create empty one for Flux2
            if latent_image is not None:
                latent = latent_image
            else:
                latent = self._create_flux2_latent(
                    ctx.width,
                    ctx.height,
                    ctx.generation.batch,
                )
                # Add seeds for deterministic noise
                latent["seeds"] = ctx.seeds[:ctx.generation.batch] if ctx.seeds else [ctx.seed]
            
            # CRITICAL: Force-disable multi-scale for Flux2 models
            # Multi-scale is designed for UNet architectures (SD1.5/SDXL) and
            # causes significant performance overhead for Flux2's DiT architecture
            enable_multiscale = False  # Always disable for Flux2
            if ctx.sampling.enable_multiscale:
                logger.info("Multi-scale disabled: not compatible with Flux2 architecture")
            
            # Run sampling with flux=True AND flux2=True for resolution-aware scheduler
            ksampler = sampling.KSampler()
            result = ksampler.sample(
                seed=ctx.seed,
                steps=ctx.sampling.steps,
                cfg=ctx.sampling.cfg,
                sampler_name=ctx.sampling.sampler,
                scheduler=ctx.sampling.scheduler,
                denoise=ctx.sampling.denoise,
                pipeline=True,
                model=self.model,
                positive=positive,
                negative=negative,
                latent_image=latent,
                start_step=start_step,
                last_step=last_step,
                disable_noise=disable_noise,
                callback=callback or ctx.callback,
                flux=True,  # Enable Flux sampling mode
                flux2=True,  # Enable Flux2-specific resolution-aware scheduler (matches ComfyUI Flux2Scheduler)
                enable_multiscale=enable_multiscale,  # Force disabled for Flux2
                multiscale_factor=ctx.sampling.multiscale_factor,
                multiscale_fullres_start=ctx.sampling.multiscale_fullres_start,
                multiscale_fullres_end=ctx.sampling.multiscale_fullres_end,
                multiscale_intermittent_fullres=ctx.sampling.multiscale_intermittent_fullres,
                cfg_free_enabled=ctx.sampling.cfg_free_enabled,
                cfg_free_start_percent=ctx.sampling.cfg_free_start_percent,
                batched_cfg=ctx.sampling.batched_cfg,
                dynamic_cfg_rescaling=ctx.sampling.dynamic_cfg_rescaling,
                dynamic_cfg_method=ctx.sampling.dynamic_cfg_method,
                dynamic_cfg_percentile=ctx.sampling.dynamic_cfg_percentile,
                dynamic_cfg_target_scale=ctx.sampling.dynamic_cfg_target_scale,
                adaptive_noise_enabled=ctx.sampling.adaptive_noise_enabled,
                adaptive_noise_method=ctx.sampling.adaptive_noise_method,
            )
            
            return result[0]
            
        except Exception as e:
            logger.exception(f"Generation failed: {e}")
            raise
    
    def _create_flux2_latent(self, width: int, height: int, batch_size: int) -> dict:
        """Create an empty latent tensor for Flux2.
        
        Flux2 uses 32-channel VAE-shaped latents in the pipeline.
        
        Args:
            width: Image width
            height: Image height
            batch_size: Batch size
            
        Returns:
            Dict with 'samples' key containing latent tensor
        """
        # Flux VAE uses 8x downscaling 
        latent_height = height // 8
        latent_width = width // 8
        
        latent = torch.zeros(
            batch_size,
            32,
            latent_height,
            latent_width,
            dtype=torch.float32,
        )
        
        return {"samples": latent}
    
    def decode(self, latents: torch.Tensor) -> torch.Tensor:
        """Decode latents to pixel space using the VAE.
        
        Args:
            latents: Latent tensor or dict with 'samples' key
            
        Returns:
            Decoded image tensor in [0, 1] range
        """
        if not self._loaded:
            raise RuntimeError("Model must be loaded before decoding")
        
        try:
            # Handle both raw tensor and dict input
            if isinstance(latents, dict):
                samples_tensor = latents["samples"]
            else:
                samples_tensor = latents
            
            # Use the Flux2 latent format
            # Apply process_latent_out (undo scale/shift from sampling) is now handled by KSAMPLER
            
            # Decode with VAE
            decoder = VariationalAE.VAEDecode()
            result = decoder.decode(
                vae=self.vae,
                samples={"samples": samples_tensor},
            )
            
            return result[0]
            
        except Exception as e:
            logger.exception(f"Decoding failed: {e}")
            raise
    
    def get_model_object(self, name):
        """Get an attribute from the model or its patcher."""
        if name == "latent_format":
            return self._model_config.latent_format
        if self.model:
            return self.model.get_model_object(name)
        return None

    def apply_lora(
        self,
        lora_name: str,
        strength_model: float = 1.0,
        strength_clip: float = 1.0,
    ) -> "Flux2KleinModel":
        """Apply a LoRA to the Flux2 Klein model.
        
        Note: LoRA support for Flux2 may be limited.
        
        Args:
            lora_name: Name/path of the LoRA file
            strength_model: Strength to apply to the model
            strength_clip: Strength to apply to CLIP
            
        Returns:
            Self for method chaining
        """
        if not self._loaded:
            raise RuntimeError("Model must be loaded before applying LoRA")
        
        try:
            loader = LoRas.LoraLoader()
            result = loader.load_lora(
                lora_name=lora_name,
                strength_model=strength_model,
                strength_clip=strength_clip,
                model=self.model,
                clip=self.clip,
            )
            self.model = result[0]
            self.clip = result[1]
            logger.info(f"Applied LoRA: {lora_name}")
        except Exception as e:
            logger.warning(f"Failed to apply LoRA {lora_name}: {e}")
        
        return self