File size: 35,553 Bytes
473fcf2
 
 
 
 
 
374d38b
 
473fcf2
 
 
 
 
 
 
 
 
374d38b
 
 
 
 
 
 
473fcf2
 
af9b854
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
374d38b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
af9b854
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
374d38b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
af9b854
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
374d38b
473fcf2
 
 
 
 
 
 
 
 
374d38b
 
473fcf2
 
 
 
 
 
 
 
 
 
 
374d38b
 
 
 
 
473fcf2
 
374d38b
 
 
 
473fcf2
374d38b
 
473fcf2
 
 
 
 
 
 
 
 
374d38b
 
 
 
 
 
473fcf2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
374d38b
473fcf2
 
 
 
 
 
 
 
 
374d38b
 
473fcf2
 
 
 
 
 
 
 
 
 
 
374d38b
 
 
 
 
473fcf2
 
374d38b
 
 
 
473fcf2
374d38b
 
473fcf2
 
 
 
 
 
 
 
374d38b
 
 
 
 
 
 
473fcf2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
374d38b
 
af9b854
 
473fcf2
 
 
 
 
 
af9b854
 
 
 
 
473fcf2
 
 
 
 
af9b854
 
473fcf2
 
af9b854
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
473fcf2
 
 
 
 
 
 
 
 
 
af9b854
 
 
 
 
 
 
 
 
473fcf2
 
 
af9b854
 
473fcf2
 
 
 
374d38b
af9b854
 
374d38b
 
473fcf2
 
 
 
 
 
 
 
 
 
 
c30936f
 
 
 
 
473fcf2
 
30d461e
473fcf2
30d461e
473fcf2
 
30d461e
c30936f
 
 
 
 
 
 
30d461e
473fcf2
 
 
30d461e
c30936f
473fcf2
c30936f
30d461e
 
 
 
473fcf2
 
 
 
 
 
 
 
 
30d461e
c30936f
473fcf2
c30936f
30d461e
 
 
 
473fcf2
 
 
 
 
 
 
c30936f
 
473fcf2
 
 
c30936f
 
473fcf2
 
 
 
 
30d461e
 
473fcf2
 
30d461e
473fcf2
30d461e
 
473fcf2
 
 
 
30d461e
 
 
 
473fcf2
 
 
af9b854
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
473fcf2
af9b854
 
 
473fcf2
 
af9b854
 
473fcf2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
af9b854
473fcf2
 
 
 
 
 
 
 
 
 
 
 
 
af9b854
473fcf2
 
 
 
 
 
 
 
 
 
 
 
 
c30936f
 
 
 
 
 
 
 
 
 
473fcf2
c30936f
 
473fcf2
 
 
c30936f
 
473fcf2
 
 
 
 
 
 
 
 
c30936f
 
473fcf2
 
 
c30936f
 
473fcf2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
47751f7
473fcf2
c30936f
 
 
 
473fcf2
 
 
c30936f
 
473fcf2
30d461e
473fcf2
 
47751f7
c30936f
 
 
 
 
 
 
 
 
 
 
 
473fcf2
47751f7
 
 
 
 
 
 
fb9f765
473fcf2
47751f7
 
 
 
 
c30936f
47751f7
 
 
 
 
fb9f765
473fcf2
 
 
 
47751f7
 
 
 
 
 
 
 
 
 
 
473fcf2
 
47751f7
 
 
 
 
c30936f
47751f7
 
 
 
 
fb9f765
473fcf2
 
 
 
47751f7
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
473fcf2
30d461e
47751f7
fb9f765
47751f7
30d461e
473fcf2
 
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
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
"""
Prefill vs Decode phase comparison module.

Demonstrates the key difference between:
- Prefill: Process entire prompt in parallel (N² attention complexity)
- Decode: Generate one token at a time (N attention per token, but sequential)

Uses REAL HuggingFace model attention layers for accurate benchmarking.
"""

import torch
import torch.nn.functional as F
import numpy as np
import plotly.graph_objects as go
from plotly.subplots import make_subplots

from .constants import MODEL_CONFIGS, ATTENTION_BACKENDS
from .models import load_model
from .attention_utils import (
    extract_attention_layer,
    create_attention_inputs,
    benchmark_attention_layer,
    get_model_attention_info,
)


def get_real_model_config(model_name: str) -> dict:
    """
    Load model and extract ACTUAL config values from model.config.
    
    This function ensures we use real model architecture values,
    NOT hardcoded constants from MODEL_CONFIGS.
    
    Args:
        model_name: Key from MODEL_CONFIGS (e.g., "SmolLM2-360M")
        
    Returns:
        Dict with real model configuration values
    """
    model = load_model(model_name)
    config = model.config
    
    # Extract values directly from model.config
    num_heads = config.num_attention_heads
    num_kv_heads = getattr(config, 'num_key_value_heads', num_heads)
    head_dim = config.hidden_size // num_heads
    
    return {
        "num_layers": config.num_hidden_layers,
        "num_heads": num_heads,
        "num_kv_heads": num_kv_heads,
        "head_dim": head_dim,
        "hidden_size": config.hidden_size,
        "model_type": getattr(config, 'model_type', 'unknown'),
        "gqa_ratio": num_heads // num_kv_heads if num_kv_heads > 0 else 1,
    }


def run_prefill_with_real_model(
    model,
    attention_layer,
    seq_len: int,
    batch_size: int = 1,
    num_iterations: int = 5,
    use_flash: bool = True,
) -> dict:
    """
    Run prefill phase attention using a REAL model's attention layer.
    
    Prefill processes the entire prompt at once:
    - Hidden states have shape [batch, seq_len, hidden_dim]
    - Full N×N attention matrix computed via the real attention layer
    
    Args:
        model: Loaded HuggingFace model
        attention_layer: Extracted attention module
        seq_len: Sequence length
        batch_size: Batch size
        num_iterations: Number of timed iterations
        use_flash: Whether to use FlashAttention backend
    
    Returns:
        Dict with timing and memory stats
    """
    if not torch.cuda.is_available():
        return {"error": "CUDA not available"}
    
    device = torch.device("cuda")
    dtype = torch.float16
    
    # Create proper inputs for the attention layer
    hidden_states, position_ids = create_attention_inputs(
        model, batch_size, seq_len, device, dtype
    )
    
    # Backend configuration
    backend = "flash" if use_flash else "math"
    
    # Run benchmark using the utility function
    result = benchmark_attention_layer(
        attention_layer=attention_layer,
        hidden_states=hidden_states,
        position_ids=position_ids,
        backend=backend,
        num_iterations=num_iterations,
        warmup_iterations=2,
    )
    
    # Clean up
    del hidden_states, position_ids
    torch.cuda.empty_cache()
    
    # Add phase info to result
    result["seq_len"] = seq_len
    result["phase"] = "prefill"
    result["using_real_model"] = True
    
    return result


def run_prefill_benchmark(
    model_name: str,
    seq_len: int,
    batch_size: int = 1,
    num_iterations: int = 10,
    use_flash: bool = True,
) -> dict:
    """
    Benchmark prefill phase using F.scaled_dot_product_attention with REAL model dimensions.
    
    This function uses the model's actual configuration (from model.config) to create
    properly-sized Q, K, V tensors, then benchmarks the SDPA operation directly.
    This is more reliable than calling attention layer forward() methods.
    
    Args:
        model_name: Key from MODEL_CONFIGS (model will be loaded to get real config)
        seq_len: Sequence length (prompt tokens)
        batch_size: Batch size
        num_iterations: Number of timed iterations
        use_flash: Whether to use FlashAttention backend
        
    Returns:
        Dict with time_ms, memory_mb, and status
    """
    if not torch.cuda.is_available():
        return {"time_ms": 0, "memory_mb": 0, "status": "error: CUDA not available"}
    
    device = torch.device("cuda")
    dtype = torch.float16
    
    try:
        # Get REAL config from loaded model
        real_config = get_real_model_config(model_name)
        num_heads = real_config["num_heads"]
        head_dim = real_config["head_dim"]
        
        # Create Q, K, V tensors with REAL model dimensions
        # Shape: [batch, num_heads, seq_len, head_dim]
        Q = torch.randn(batch_size, num_heads, seq_len, head_dim, dtype=dtype, device=device)
        K = torch.randn(batch_size, num_heads, seq_len, head_dim, dtype=dtype, device=device)
        V = torch.randn(batch_size, num_heads, seq_len, head_dim, dtype=dtype, device=device)
        
        # Set backend flags
        if use_flash:
            enable_math, enable_flash, enable_mem_efficient = False, True, False
        else:
            enable_math, enable_flash, enable_mem_efficient = True, False, False
        
        # Warmup
        for _ in range(3):
            with torch.backends.cuda.sdp_kernel(
                enable_flash=enable_flash,
                enable_math=enable_math,
                enable_mem_efficient=enable_mem_efficient
            ):
                _ = F.scaled_dot_product_attention(Q, K, V, is_causal=True)
        
        torch.cuda.synchronize()
        torch.cuda.reset_peak_memory_stats()
        
        # Timed runs
        start = torch.cuda.Event(enable_timing=True)
        end = torch.cuda.Event(enable_timing=True)
        
        start.record()
        for _ in range(num_iterations):
            with torch.backends.cuda.sdp_kernel(
                enable_flash=enable_flash,
                enable_math=enable_math,
                enable_mem_efficient=enable_mem_efficient
            ):
                output = F.scaled_dot_product_attention(Q, K, V, is_causal=True)
        end.record()
        
        torch.cuda.synchronize()
        
        time_ms = start.elapsed_time(end) / num_iterations
        memory_mb = torch.cuda.max_memory_allocated() / (1024 * 1024)
        
        # Cleanup
        del Q, K, V, output
        torch.cuda.empty_cache()
        
        return {
            "time_ms": round(time_ms, 3),
            "memory_mb": round(memory_mb, 1),
            "seq_len": seq_len,
            "phase": "prefill",
            "backend": "flash" if use_flash else "math",
            "num_heads": num_heads,
            "head_dim": head_dim,
            "status": "success",
            "using_real_config": True,
        }
        
    except Exception as e:
        return {
            "time_ms": 0,
            "memory_mb": 0,
            "status": f"error: {str(e)[:100]}",
            "phase": "prefill",
        }


def run_decode_with_real_model(
    model,
    attention_layer,
    kv_cache_len: int,
    num_tokens: int = 10,
    batch_size: int = 1,
    num_iterations: int = 3,
    use_flash: bool = True,
) -> dict:
    """
    Run decode phase attention using a REAL model's attention layer.
    
    Decode generates one token at a time:
    - Single query token attending to all past keys/values
    - Simulates the memory-bound decode phase
    
    Args:
        model: Loaded HuggingFace model
        attention_layer: Extracted attention module
        kv_cache_len: Length of the KV cache (context)
        num_tokens: Number of tokens to simulate generating
        batch_size: Batch size
        num_iterations: Iterations for averaging
        use_flash: Whether to use FlashAttention backend
    
    Returns:
        Dict with per-token timing and memory stats
    """
    if not torch.cuda.is_available():
        return {"error": "CUDA not available"}
    
    device = torch.device("cuda")
    dtype = torch.float16
    
    # Create single-token query input (simulating decode)
    hidden_dim = model.config.hidden_size
    query_hidden = torch.randn(batch_size, 1, hidden_dim, dtype=dtype, device=device)
    position_ids = torch.tensor([[kv_cache_len]], device=device).expand(batch_size, 1)
    
    # Backend flags
    if use_flash:
        enable_math, enable_flash, enable_mem_efficient = False, True, False
    else:
        enable_math, enable_flash, enable_mem_efficient = True, False, False
    
    try:
        # Warmup
        with torch.backends.cuda.sdp_kernel(
            enable_flash=enable_flash,
            enable_math=enable_math,
            enable_mem_efficient=enable_mem_efficient
        ):
            with torch.no_grad():
                for _ in range(2):
                    _ = attention_layer(query_hidden, position_ids=position_ids)
        
        torch.cuda.synchronize()
        torch.cuda.reset_peak_memory_stats()
        
        # Time multiple tokens
        start = torch.cuda.Event(enable_timing=True)
        end = torch.cuda.Event(enable_timing=True)
        
        with torch.backends.cuda.sdp_kernel(
            enable_flash=enable_flash,
            enable_math=enable_math,
            enable_mem_efficient=enable_mem_efficient
        ):
            with torch.no_grad():
                start.record()
                for _ in range(num_tokens * num_iterations):
                    output = attention_layer(query_hidden, position_ids=position_ids)
                end.record()
        
        torch.cuda.synchronize()
        
        total_time_ms = start.elapsed_time(end)
        time_per_token_ms = total_time_ms / (num_tokens * num_iterations)
        memory_mb = torch.cuda.max_memory_allocated() / (1024 * 1024)
        
        # Clean up
        del query_hidden
        torch.cuda.empty_cache()
        
        return {
            "time_ms_per_token": round(time_per_token_ms, 4),
            "total_time_ms": round(total_time_ms / num_iterations, 3),
            "memory_mb": round(memory_mb, 1),
            "kv_cache_len": kv_cache_len,
            "num_tokens": num_tokens,
            "phase": "decode",
            "using_real_model": True,
            "status": "success",
        }
        
    except Exception as e:
        return {
            "time_ms_per_token": 0,
            "total_time_ms": 0,
            "memory_mb": 0,
            "kv_cache_len": kv_cache_len,
            "num_tokens": num_tokens,
            "phase": "decode",
            "status": f"error: {str(e)[:80]}",
        }


def run_decode_benchmark(
    model_name: str,
    kv_cache_len: int,
    num_tokens: int = 10,
    batch_size: int = 1,
    num_iterations: int = 5,
    use_flash: bool = True,
) -> dict:
    """
    Benchmark decode phase using F.scaled_dot_product_attention with REAL model dimensions.
    
    Properly simulates decode by:
    - Creating single query token (Q with seq_len=1)
    - Creating KV cache tensors with kv_cache_len tokens
    - Handling GQA by expanding KV heads to match Q heads
    
    Args:
        model_name: Key from MODEL_CONFIGS (model will be loaded to get real config)
        kv_cache_len: Length of KV cache (context length)
        num_tokens: Number of decode tokens to simulate
        batch_size: Batch size
        num_iterations: Iterations for timing
        use_flash: Whether to use FlashAttention backend
        
    Returns:
        Dict with time_ms_per_token, memory_mb, and status
    """
    if not torch.cuda.is_available():
        return {"time_ms_per_token": 0, "memory_mb": 0, "status": "error: CUDA not available"}
    
    device = torch.device("cuda")
    dtype = torch.float16
    
    try:
        # Get REAL config from loaded model
        real_config = get_real_model_config(model_name)
        num_heads = real_config["num_heads"]
        num_kv_heads = real_config["num_kv_heads"]
        head_dim = real_config["head_dim"]
        
        # Single query token: [batch, num_heads, 1, head_dim]
        Q = torch.randn(batch_size, num_heads, 1, head_dim, dtype=dtype, device=device)
        
        # KV cache with real model's KV head count: [batch, num_kv_heads, kv_cache_len, head_dim]
        K_cache = torch.randn(batch_size, num_kv_heads, kv_cache_len, head_dim, dtype=dtype, device=device)
        V_cache = torch.randn(batch_size, num_kv_heads, kv_cache_len, head_dim, dtype=dtype, device=device)
        
        # Handle GQA: expand KV heads to match Q heads if needed
        if num_kv_heads < num_heads:
            repeat_factor = num_heads // num_kv_heads
            K_cache = K_cache.repeat_interleave(repeat_factor, dim=1)
            V_cache = V_cache.repeat_interleave(repeat_factor, dim=1)
        
        # Set backend flags
        if use_flash:
            enable_math, enable_flash_flag, enable_mem_efficient = False, True, False
        else:
            enable_math, enable_flash_flag, enable_mem_efficient = True, False, False
        
        # Warmup
        for _ in range(3):
            with torch.backends.cuda.sdp_kernel(
                enable_flash=enable_flash_flag,
                enable_math=enable_math,
                enable_mem_efficient=enable_mem_efficient
            ):
                _ = F.scaled_dot_product_attention(Q, K_cache, V_cache)
        
        torch.cuda.synchronize()
        torch.cuda.reset_peak_memory_stats()
        
        # Timed runs - simulate generating num_tokens
        start = torch.cuda.Event(enable_timing=True)
        end = torch.cuda.Event(enable_timing=True)
        
        start.record()
        for _ in range(num_tokens * num_iterations):
            with torch.backends.cuda.sdp_kernel(
                enable_flash=enable_flash_flag,
                enable_math=enable_math,
                enable_mem_efficient=enable_mem_efficient
            ):
                output = F.scaled_dot_product_attention(Q, K_cache, V_cache)
        end.record()
        
        torch.cuda.synchronize()
        
        total_time_ms = start.elapsed_time(end)
        time_per_token_ms = total_time_ms / (num_tokens * num_iterations)
        memory_mb = torch.cuda.max_memory_allocated() / (1024 * 1024)
        
        # Cleanup
        del Q, K_cache, V_cache, output
        torch.cuda.empty_cache()
        
        return {
            "time_ms_per_token": round(time_per_token_ms, 4),
            "total_time_ms": round(total_time_ms / num_iterations, 3),
            "memory_mb": round(memory_mb, 1),
            "kv_cache_len": kv_cache_len,
            "num_tokens": num_tokens,
            "phase": "decode",
            "backend": "flash" if use_flash else "math",
            "num_heads": num_heads,
            "num_kv_heads": num_kv_heads,
            "head_dim": head_dim,
            "status": "success",
            "using_real_config": True,
        }
        
    except Exception as e:
        return {
            "time_ms_per_token": 0,
            "total_time_ms": 0,
            "memory_mb": 0,
            "kv_cache_len": kv_cache_len,
            "num_tokens": num_tokens,
            "phase": "decode",
            "status": f"error: {str(e)[:100]}",
        }


# Legacy function kept for backwards compatibility
def simulate_prefill_attention(
    batch_size: int,
    num_heads: int,
    seq_len: int,
    head_dim: int,
    num_iterations: int = 5,
    use_flash: bool = True,
) -> dict:
    """
    Legacy: Simulate prefill phase attention with random tensors.
    Use run_prefill_with_real_model() for real model benchmarks.
    """
    if not torch.cuda.is_available():
        return {"error": "CUDA not available"}
    
    device = torch.device("cuda")
    dtype = torch.float16
    
    Q = torch.randn(batch_size, num_heads, seq_len, head_dim, device=device, dtype=dtype)
    K = torch.randn(batch_size, num_heads, seq_len, head_dim, device=device, dtype=dtype)
    V = torch.randn(batch_size, num_heads, seq_len, head_dim, device=device, dtype=dtype)
    
    if use_flash:
        enable_math, enable_flash_flag, enable_mem_efficient = False, True, False
    else:
        enable_math, enable_flash_flag, enable_mem_efficient = True, False, False
    
    # Warmup
    for _ in range(2):
        with torch.backends.cuda.sdp_kernel(
            enable_flash=enable_flash_flag, enable_math=enable_math, enable_mem_efficient=enable_mem_efficient
        ):
            try:
                _ = F.scaled_dot_product_attention(Q, K, V)
            except Exception:
                pass
    
    torch.cuda.synchronize()
    torch.cuda.reset_peak_memory_stats()
    
    start = torch.cuda.Event(enable_timing=True)
    end = torch.cuda.Event(enable_timing=True)
    
    start.record()
    for _ in range(num_iterations):
        with torch.backends.cuda.sdp_kernel(
            enable_flash=enable_flash_flag, enable_math=enable_math, enable_mem_efficient=enable_mem_efficient
        ):
            try:
                output = F.scaled_dot_product_attention(Q, K, V)
            except Exception:
                output = F.scaled_dot_product_attention(Q, K, V)
    end.record()
    
    torch.cuda.synchronize()
    
    total_time_ms = start.elapsed_time(end)
    avg_time_ms = total_time_ms / num_iterations
    peak_memory_mb = torch.cuda.max_memory_allocated() / (1024 * 1024)
    
    del Q, K, V, output
    torch.cuda.empty_cache()
    
    return {
        "time_ms": avg_time_ms,
        "memory_mb": peak_memory_mb,
        "seq_len": seq_len,
        "phase": "prefill",
    }


# Legacy function kept for backwards compatibility
def simulate_decode_attention(
    batch_size: int,
    num_heads: int,
    kv_cache_len: int,
    head_dim: int,
    num_tokens: int = 10,
    use_flash: bool = True,
) -> dict:
    """
    Legacy: Simulate decode phase attention with random tensors.
    Use run_decode_with_real_model() for real model benchmarks.
    """
    if not torch.cuda.is_available():
        return {"error": "CUDA not available"}
    
    device = torch.device("cuda")
    dtype = torch.float16
    
    K_cache = torch.randn(batch_size, num_heads, kv_cache_len, head_dim, device=device, dtype=dtype)
    V_cache = torch.randn(batch_size, num_heads, kv_cache_len, head_dim, device=device, dtype=dtype)
    Q = torch.randn(batch_size, num_heads, 1, head_dim, device=device, dtype=dtype)
    
    if use_flash:
        enable_math, enable_flash_flag, enable_mem_efficient = False, True, False
    else:
        enable_math, enable_flash_flag, enable_mem_efficient = True, False, False
    
    # Warmup
    for _ in range(2):
        with torch.backends.cuda.sdp_kernel(
            enable_flash=enable_flash_flag, enable_math=enable_math, enable_mem_efficient=enable_mem_efficient
        ):
            try:
                _ = F.scaled_dot_product_attention(Q, K_cache, V_cache)
            except Exception:
                pass
    
    torch.cuda.synchronize()
    torch.cuda.reset_peak_memory_stats()
    
    start = torch.cuda.Event(enable_timing=True)
    end = torch.cuda.Event(enable_timing=True)
    
    start.record()
    for _ in range(num_tokens):
        with torch.backends.cuda.sdp_kernel(
            enable_flash=enable_flash_flag, enable_math=enable_math, enable_mem_efficient=enable_mem_efficient
        ):
            try:
                output = F.scaled_dot_product_attention(Q, K_cache, V_cache)
            except Exception:
                output = F.scaled_dot_product_attention(Q, K_cache, V_cache)
    end.record()
    
    torch.cuda.synchronize()
    
    total_time_ms = start.elapsed_time(end)
    avg_time_per_token_ms = total_time_ms / num_tokens
    peak_memory_mb = torch.cuda.max_memory_allocated() / (1024 * 1024)
    
    del Q, K_cache, V_cache, output
    torch.cuda.empty_cache()
    
    return {
        "time_ms_per_token": avg_time_per_token_ms,
        "total_time_ms": total_time_ms,
        "memory_mb": peak_memory_mb,
        "kv_cache_len": kv_cache_len,
        "num_tokens": num_tokens,
        "phase": "decode",
    }


def run_prefill_decode_comparison(
    model_name: str,
    context_length: int,
    decode_tokens: int = 32,
) -> tuple:
    """
    Run full comparison between prefill and decode phases using REAL HuggingFace model.
    
    Uses F.scaled_dot_product_attention with real model dimensions for reliable benchmarking.
    All config values come from model.config, not constants.
    
    Returns results dict, comparison chart, KV cache chart, and insight text.
    """
    if model_name not in MODEL_CONFIGS:
        return {"error": f"Unknown model: {model_name}"}, None, None, "Error: Unknown model"
    
    # Get REAL config from model.config (not constants)
    try:
        real_config = get_real_model_config(model_name)
    except Exception as e:
        return {"error": f"Failed to load model: {str(e)[:50]}"}, None, None, f"Error: {str(e)[:50]}"
    
    results = {
        "model": model_name,
        "context_length": context_length,
        "decode_tokens": decode_tokens,
        "real_config": real_config,
        "using_real_config": True,
    }
    
    # Run prefill benchmarks using SDPA with REAL model dimensions
    prefill_flash = run_prefill_benchmark(
        model_name=model_name,
        seq_len=context_length,
        batch_size=1,
        use_flash=True,
    )
    
    prefill_math = run_prefill_benchmark(
        model_name=model_name,
        seq_len=context_length,
        batch_size=1,
        use_flash=False,
    )
    
    # Run decode benchmarks using SDPA with proper KV cache simulation
    decode_flash = run_decode_benchmark(
        model_name=model_name,
        kv_cache_len=context_length,
        num_tokens=decode_tokens,
        batch_size=1,
        use_flash=True,
    )
    
    decode_math = run_decode_benchmark(
        model_name=model_name,
        kv_cache_len=context_length,
        num_tokens=decode_tokens,
        batch_size=1,
        use_flash=False,
    )
    
    results["prefill"] = {
        "flash": prefill_flash,
        "math": prefill_math,
    }
    results["decode"] = {
        "flash": decode_flash,
        "math": decode_math,
    }
    
    # Add model info for display
    results["model_info"] = {
        "num_heads": real_config["num_heads"],
        "num_kv_heads": real_config["num_kv_heads"],
        "head_dim": real_config["head_dim"],
        "num_layers": real_config["num_layers"],
        "gqa_ratio": real_config["gqa_ratio"],
    }
    
    # Create comparison chart
    comparison_chart = create_comparison_chart(results)
    
    # Create KV cache growth chart using REAL model config
    kv_cache_chart = create_kv_cache_chart(model_name, context_length, decode_tokens)
    
    # Generate insight
    insight = generate_phase_insight(results)
    
    # Add real model indicator to insight
    if results.get("using_real_config"):
        model_indicator = f"\n\n---\n\n*Benchmarked using real **{model_name}** config ({real_config['num_heads']} heads, {real_config['head_dim']}d, GQA {real_config['gqa_ratio']}:1)*"
        insight = insight + model_indicator
    
    return results, comparison_chart, kv_cache_chart, insight


def create_comparison_chart(results: dict) -> go.Figure:
    """Create bar chart comparing prefill vs decode timing."""
    
    prefill_flash = results["prefill"]["flash"]
    prefill_math = results["prefill"]["math"]
    decode_flash = results["decode"]["flash"]
    decode_math = results["decode"]["math"]
    
    # Helper to safely get numeric value (handles None)
    def safe_get(d, key, default=0):
        val = d.get(key, default)
        return val if val is not None else default
    
    fig = make_subplots(
        rows=1, cols=2,
        subplot_titles=("<b>Prefill Time</b> (Full Prompt)", "<b>Decode Time</b> (Per Token)"),
        horizontal_spacing=0.15,
        vertical_spacing=0.15,
    )
    
    # Get max values for proper y-axis scaling with headroom for labels
    prefill_math_time = safe_get(prefill_math, "time_ms", 0)
    prefill_flash_time = safe_get(prefill_flash, "time_ms", 0)
    decode_math_time = safe_get(decode_math, "time_ms_per_token", 0)
    decode_flash_time = safe_get(decode_flash, "time_ms_per_token", 0)
    
    prefill_max = max(prefill_math_time, prefill_flash_time)
    decode_max = max(decode_math_time, decode_flash_time)
    
    # Prefill comparison
    fig.add_trace(
        go.Bar(
            x=["Math<br>(Standard)", "Flash<br>Attention"],
            y=[prefill_math_time, prefill_flash_time],
            marker_color=["#ef4444", "#22c55e"],
            text=[f"<b>{prefill_math_time:.2f}ms</b>", f"<b>{prefill_flash_time:.2f}ms</b>"],
            textposition="inside",
            textangle=0,
            insidetextanchor="middle",
            textfont=dict(color="white", size=12),
            name="Prefill",
            showlegend=False,
        ),
        row=1, col=1
    )
    
    # Decode comparison (per token)
    fig.add_trace(
        go.Bar(
            x=["Math<br>(Standard)", "Flash<br>Attention"],
            y=[decode_math_time, decode_flash_time],
            marker_color=["#ef4444", "#22c55e"],
            text=[f"<b>{decode_math_time:.3f}ms</b>", f"<b>{decode_flash_time:.3f}ms</b>"],
            textposition="inside",
            textangle=0,
            insidetextanchor="middle",
            textfont=dict(color="white", size=12),
            name="Decode",
            showlegend=False,
        ),
        row=1, col=2
    )
    
    # Calculate speedups
    if prefill_math_time > 0 and prefill_flash_time > 0:
        prefill_speedup = prefill_math_time / prefill_flash_time
    else:
        prefill_speedup = 1.0
    
    if decode_math_time > 0 and decode_flash_time > 0:
        decode_speedup = decode_math_time / decode_flash_time
    else:
        decode_speedup = 1.0
    
    fig.update_layout(
        title=dict(
            text=f"<b>Prefill vs Decode: FlashAttention Speedup</b><br>"
                 f"<span style='font-size:13px;color:#16a34a'>"
                 f"Prefill: {prefill_speedup:.1f}× faster | Decode: {decode_speedup:.1f}× faster</span>",
            x=0.5,
            font=dict(size=15),
        ),
        height=380,
        margin=dict(l=60, r=40, t=100, b=60),
        yaxis_title="Time (ms)",
        yaxis2_title="Time (ms)",
    )
    
    # Add more y-axis headroom
    fig.update_yaxes(range=[0, prefill_max * 1.15], row=1, col=1)
    fig.update_yaxes(range=[0, decode_max * 1.15], row=1, col=2)
    
    return fig


def create_kv_cache_chart(model_name: str, context_length: int, decode_tokens: int) -> go.Figure:
    """
    Create chart showing KV cache growth during generation.
    
    Uses REAL model config values from model.config, not constants.
    
    Args:
        model_name: Model name to load config from
        context_length: Number of context tokens (prefill)
        decode_tokens: Number of decode tokens to generate
        
    Returns:
        Plotly figure showing KV cache growth
    """
    # Get REAL config from loaded model (no constants!)
    real_config = get_real_model_config(model_name)
    
    num_kv_heads = real_config["num_kv_heads"]
    head_dim = real_config["head_dim"]
    num_layers = real_config["num_layers"]
    
    # Calculate KV cache size at each step
    # KV cache per layer: 2 (K+V) × kv_heads × head_dim × 2 (FP16 bytes)
    bytes_per_token_per_layer = 2 * num_kv_heads * head_dim * 2
    total_bytes_per_token = bytes_per_token_per_layer * num_layers
    
    # Generate sequence of token counts
    token_counts = list(range(0, context_length + decode_tokens + 1, max(1, (context_length + decode_tokens) // 50)))
    if token_counts[-1] != context_length + decode_tokens:
        token_counts.append(context_length + decode_tokens)
    
    # Calculate cache sizes in MB
    cache_sizes_mb = [(t * total_bytes_per_token) / (1024 * 1024) for t in token_counts]
    
    fig = go.Figure()
    
    # Prefill region (0 to context_length)
    prefill_tokens = [t for t in token_counts if t <= context_length]
    prefill_sizes = [(t * total_bytes_per_token) / (1024 * 1024) for t in prefill_tokens]
    
    fig.add_trace(go.Scatter(
        x=prefill_tokens,
        y=prefill_sizes,
        mode="lines",
        name="Prefill Phase",
        fill="tozeroy",
        line=dict(color="#3b82f6", width=2),
        fillcolor="rgba(59, 130, 246, 0.3)",
    ))
    
    # Decode region (context_length to end)
    decode_tokens_list = [t for t in token_counts if t >= context_length]
    decode_sizes = [(t * total_bytes_per_token) / (1024 * 1024) for t in decode_tokens_list]
    
    fig.add_trace(go.Scatter(
        x=decode_tokens_list,
        y=decode_sizes,
        mode="lines",
        name="Decode Phase",
        fill="tozeroy",
        line=dict(color="#22c55e", width=2),
        fillcolor="rgba(34, 197, 94, 0.3)",
    ))
    
    # Add vertical line at context boundary
    cache_at_context = (context_length * total_bytes_per_token) / (1024 * 1024)
    fig.add_vline(
        x=context_length,
        line_dash="dash",
        line_color="rgba(0, 0, 0, 0.5)",
        annotation_text=f"Prefill→Decode<br>({cache_at_context:.1f} MB)",
        annotation_position="top",
    )
    
    fig.update_layout(
        title=dict(
            text=f"KV Cache Growth ({num_kv_heads} KV heads × {num_layers} layers)",
            x=0.5,
        ),
        xaxis_title="Tokens Processed",
        yaxis_title="KV Cache Size (MB)",
        height=300,
        margin=dict(l=50, r=50, t=60, b=50),
        legend=dict(
            orientation="h",
            yanchor="bottom",
            y=-0.25,
            xanchor="center",
            x=0.5,
        ),
        yaxis=dict(rangemode='tozero'),
    )
    
    return fig


def generate_phase_insight(results: dict) -> str:
    """Generate insight text from comparison results."""
    
    prefill_flash = results["prefill"]["flash"]
    prefill_math = results["prefill"]["math"]
    decode_flash = results["decode"]["flash"]
    decode_math = results["decode"]["math"]
    
    # Helper to safely get numeric value (handles None)
    def safe_get(d, key, default=0):
        val = d.get(key, default)
        return val if val is not None else default
    
    prefill_math_time = safe_get(prefill_math, "time_ms", 0)
    prefill_flash_time = safe_get(prefill_flash, "time_ms", 0)
    decode_math_time = safe_get(decode_math, "time_ms_per_token", 0)
    decode_flash_time = safe_get(decode_flash, "time_ms_per_token", 0)
    
    # Calculate speedups
    if prefill_math_time > 0 and prefill_flash_time > 0:
        prefill_speedup = prefill_math_time / prefill_flash_time
    else:
        prefill_speedup = 1.0
    
    if decode_math_time > 0 and decode_flash_time > 0:
        decode_speedup = decode_math_time / decode_flash_time
    else:
        decode_speedup = 1.0
    
    context_length = results["context_length"]
    decode_tokens = results["decode_tokens"]
    
    insight = f"""### Key Observations

**Prefill Phase** (processing {context_length} tokens):
- Standard attention: **{prefill_math_time:.2f}ms**
- FlashAttention: **{prefill_flash_time:.2f}ms**
- Speedup: **{prefill_speedup:.1f}×**

**Decode Phase** (generating {decode_tokens} tokens):
- Standard attention: **{decode_math_time:.3f}ms/token**
- FlashAttention: **{decode_flash_time:.3f}ms/token**
- Speedup: **{decode_speedup:.1f}×**

---

### Why the Difference?

1. **Prefill is compute-bound** with N² attention operations
   - FlashAttention's memory efficiency provides significant speedup
   - Larger contexts benefit more (quadratic scaling)

2. **Decode is memory-bound** with 1×N attention per token
   - Each decode step is fast but sequential
   - KV cache read dominates, limiting FlashAttention's advantage

3. **Optimal strategy**: FlashAttention helps most during prefill;
   decode phase benefits from KV cache optimizations (GQA/MQA)
"""
    
    return insight


def get_attention_pattern_chart(context_length: int) -> go.Figure:
    """Create visualization of prefill vs decode attention patterns using scatter."""
    
    # Calculate FLOPs for insight
    prefill_flops = context_length * context_length  # N² attention
    decode_flops_per_token = context_length  # 1×N per decode token
    
    fig = make_subplots(
        rows=1, cols=2,
        subplot_titles=(
            f"<b>Prefill:</b> {context_length}×{context_length} = {prefill_flops:,} ops",
            f"<b>Decode:</b> 1×{context_length} = {decode_flops_per_token:,} ops/token"
        ),
        horizontal_spacing=0.15,
    )
    
    # Prefill: Lower triangular pattern (causal mask)
    # Dynamic size based on context length for visual feedback
    if context_length <= 16:
        size = context_length
    elif context_length <= 128:
        size = 12
    elif context_length <= 512:
        size = 10
    else:
        size = 8  # Smaller grid for very large contexts
    
    # Adjust marker size based on grid size
    marker_size = max(10, 22 - size)
    
    # Generate coordinates for filled cells (lower triangular)
    prefill_x = []
    prefill_y = []
    for row in range(size):
        for col in range(row + 1):  # Only up to diagonal
            prefill_x.append(col)
            prefill_y.append(row)
    
    fig.add_trace(
        go.Scatter(
            x=prefill_x,
            y=prefill_y,
            mode="markers",
            marker=dict(
                size=marker_size,
                color="#3b82f6",
                symbol="square",
            ),
            name="Attends",
            showlegend=False,
            hovertemplate="Query %{y} → Key %{x}<extra></extra>",
        ),
        row=1, col=1
    )
    
    # Decode: Each step attends to growing sequence
    num_decode_steps = 6
    base_context = max(4, size - num_decode_steps)
    
    decode_x = []
    decode_y = []
    for step in range(num_decode_steps):
        attend_len = base_context + step + 1
        for col in range(min(attend_len, size)):
            decode_x.append(col)
            decode_y.append(step)
    
    fig.add_trace(
        go.Scatter(
            x=decode_x,
            y=decode_y,
            mode="markers",
            marker=dict(
                size=marker_size + 4,
                color="#22c55e",
                symbol="square",
            ),
            name="Attends",
            showlegend=False,
            hovertemplate="Decode step %{y} → Key %{x}<extra></extra>",
        ),
        row=1, col=2
    )
    
    # Update axes with proper ranges
    fig.update_xaxes(
        title_text="Key positions", 
        range=[-0.5, size - 0.5],
        dtick=2,
        row=1, col=1
    )
    fig.update_xaxes(
        title_text="Key positions (KV cache)", 
        range=[-0.5, size - 0.5],
        dtick=2,
        row=1, col=2
    )
    fig.update_yaxes(
        title_text="Query positions", 
        range=[-0.5, size - 0.5],
        dtick=2,
        row=1, col=1
    )
    fig.update_yaxes(
        title_text="Decode steps", 
        range=[-0.5, num_decode_steps - 0.5],
        dtick=1,
        row=1, col=2
    )
    
    fig.update_layout(
        height=380,
        margin=dict(l=60, r=30, t=70, b=50),
        plot_bgcolor="rgba(241, 245, 249, 0.5)",
    )
    
    return fig