Spaces:
Sleeping
Sleeping
File size: 37,329 Bytes
05810bb 374d38b 05810bb 374d38b d5ef20e e2de6cd d5ef20e e2de6cd d5ef20e e2de6cd d5ef20e e2de6cd d5ef20e e2de6cd d5ef20e e2de6cd d5ef20e e2de6cd d5ef20e 5642188 e2de6cd d5ef20e e2de6cd 5642188 e2de6cd 5642188 d5ef20e 05810bb 374d38b 05810bb 374d38b 05810bb 374d38b 05810bb 374d38b 05810bb 374d38b 05810bb 685194e 374d38b 685194e 374d38b 685194e 374d38b 685194e 374d38b 685194e 374d38b 685194e 374d38b 685194e 374d38b 05810bb 374d38b 05810bb 374d38b 05810bb 374d38b 05810bb 374d38b 05810bb 374d38b 05810bb 374d38b 05810bb 374d38b 05810bb 3fec038 05810bb d5ef20e 3fec038 05810bb 3fec038 d5ef20e 3fec038 05810bb d5ef20e 05810bb 3fec038 05810bb 3fec038 05810bb 3fec038 05810bb 3fec038 e2de6cd 3fec038 e2de6cd 3fec038 e2de6cd 3fec038 e2de6cd 3fec038 e2de6cd 3fec038 e2de6cd 3fec038 e2de6cd 3fec038 e2de6cd 3fec038 e2de6cd 3fec038 e2de6cd 3fec038 e2de6cd 3fec038 e2de6cd 05810bb e2de6cd 05810bb e2de6cd 05810bb e2de6cd 05810bb e2de6cd 05810bb e2de6cd 05810bb e2de6cd 05810bb e2de6cd 05810bb e2de6cd 05810bb e2de6cd 05810bb e2de6cd 05810bb e2de6cd 05810bb e2de6cd 05810bb 3fec038 05810bb 3fec038 05810bb 3fec038 | 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 | """
Benchmark module for FlashAttention Explorer.
GPU benchmark functions for comparing attention backends using real HuggingFace models.
"""
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 GPU_SPECS, ATTENTION_BACKENDS, MODEL_CONFIGS, DEFAULT_GPU, DEFAULT_MODEL
from .models import load_model, clear_model_cache
from .attention_utils import (
extract_attention_layer,
create_attention_inputs,
benchmark_attention_layer,
get_model_attention_info,
)
def detect_gpu() -> dict:
"""
Detect the actual GPU and return its specs.
Returns:
Dict with GPU name and specs
"""
if not torch.cuda.is_available():
return {"name": "CPU (No GPU)", "detected": False, **GPU_SPECS[DEFAULT_GPU]}
gpu_name_raw = torch.cuda.get_device_name(0)
gpu_name = gpu_name_raw.lower()
# Get memory in GB for dynamic spec estimation
try:
mem_gb = torch.cuda.get_device_properties(0).total_memory / (1024**3)
except Exception:
mem_gb = 24 # fallback
# Match against known GPUs (ordered from newest to oldest)
if "h200" in gpu_name:
# H200 specs - HBM3e memory, very high bandwidth
return {
"detected": True,
"detected_name": gpu_name_raw,
"name": "NVIDIA H200",
"tflops_fp16": 989, # Same compute as H100
"bandwidth_gbps": 4800, # HBM3e: 4.8 TB/s
"memory_gb": round(mem_gb),
"sram_kb": 256,
}
elif "h100" in gpu_name:
return {"detected": True, "detected_name": gpu_name_raw, **GPU_SPECS["H100"]}
elif "a100" in gpu_name:
return {"detected": True, "detected_name": gpu_name_raw, **GPU_SPECS["A100_80GB"]}
elif "a10" in gpu_name:
return {"detected": True, "detected_name": gpu_name_raw, **GPU_SPECS["A10G"]}
elif "l40" in gpu_name:
# L40S specs
return {
"detected": True,
"detected_name": gpu_name_raw,
"name": "NVIDIA L40S",
"tflops_fp16": 362,
"bandwidth_gbps": 864,
"memory_gb": round(mem_gb),
"sram_kb": 192,
}
elif "l4" in gpu_name:
# L4 specs
return {
"detected": True,
"detected_name": gpu_name_raw,
"name": "NVIDIA L4",
"tflops_fp16": 121,
"bandwidth_gbps": 300,
"memory_gb": round(mem_gb),
"sram_kb": 96,
}
elif "t4" in gpu_name:
return {
"detected": True,
"detected_name": gpu_name_raw,
"name": "NVIDIA T4",
"tflops_fp16": 65,
"bandwidth_gbps": 320,
"memory_gb": round(mem_gb),
"sram_kb": 64,
}
elif "v100" in gpu_name:
return {
"detected": True,
"detected_name": gpu_name_raw,
"name": "NVIDIA V100",
"tflops_fp16": 125,
"bandwidth_gbps": 900,
"memory_gb": round(mem_gb),
"sram_kb": 128,
}
elif "rtx 4090" in gpu_name or "4090" in gpu_name:
return {
"detected": True,
"detected_name": gpu_name_raw,
"name": "NVIDIA RTX 4090",
"tflops_fp16": 330,
"bandwidth_gbps": 1008,
"memory_gb": round(mem_gb),
"sram_kb": 128,
}
else:
# Unknown GPU - estimate specs using compute capability and SM count
# These are the best indicators of performance we can query
try:
props = torch.cuda.get_device_properties(0)
sm_count = props.multi_processor_count
major, minor = torch.cuda.get_device_capability(0)
# FP16 FLOPs per SM per cycle varies by architecture
# Ampere (8.x): 256 FP16 ops/SM/cycle, Hopper (9.x): 512
# Clock speed ~1.5-2 GHz typically
if major >= 9: # Hopper/Ada
flops_per_sm = 512
clock_ghz = 1.8
bw_per_gb_mem = 50 # Rough: HBM3 ~50 GB/s per GB capacity
elif major >= 8: # Ampere
flops_per_sm = 256
clock_ghz = 1.5
bw_per_gb_mem = 25 # HBM2e
elif major >= 7: # Volta/Turing
flops_per_sm = 128
clock_ghz = 1.4
bw_per_gb_mem = 28
else: # Older
flops_per_sm = 64
clock_ghz = 1.2
bw_per_gb_mem = 20
# Estimate TFLOPS: SMs × FLOPs/SM/cycle × clock × 2 (FMA)
est_tflops = (sm_count * flops_per_sm * clock_ghz * 2) / 1000
est_bw = mem_gb * bw_per_gb_mem
except Exception:
# Fallback if properties query fails
est_tflops = 125
est_bw = 600
return {
"detected": True,
"detected_name": gpu_name_raw,
"name": gpu_name_raw,
"tflops_fp16": round(est_tflops),
"bandwidth_gbps": round(est_bw),
"memory_gb": round(mem_gb),
"sram_kb": 128,
"estimated": True, # Flag that these are estimated from compute capability
"compute_capability": f"{major}.{minor}" if 'major' in dir() else "unknown",
}
def run_attention_benchmark(
model_name: str = None,
seq_len: int = 1024,
batch_size: int = 1,
num_iterations: int = 10,
warmup_iterations: int = 3,
# Legacy parameters (used if model_name is None)
num_heads: int = 16,
head_dim: int = 64,
) -> dict:
"""
Benchmark three SDPA backends using a real HuggingFace model's attention layer.
Args:
model_name: Name of the model from MODEL_CONFIGS (e.g., "SmolLM2-360M")
If None, falls back to legacy random tensor mode
seq_len: Sequence length (number of tokens)
batch_size: Batch size
num_iterations: Number of timed iterations
warmup_iterations: Number of warmup iterations
num_heads: (Legacy) Number of attention heads if model_name is None
head_dim: (Legacy) Dimension per head if model_name is None
Returns:
Dict with timing and memory results per backend
"""
if not torch.cuda.is_available():
return {"error": "CUDA not available"}
device = torch.device("cuda")
dtype = torch.float16
# If model_name is provided, use real model dimensions for benchmarking
if model_name is not None and model_name in MODEL_CONFIGS:
try:
# Load the real HuggingFace model
model = load_model(model_name)
# Get model attention info for real dimensions
attn_info = get_model_attention_info(model)
# Extract dimensions from real model
model_num_heads = attn_info["num_attention_heads"]
model_head_dim = attn_info["head_dim"]
results = {"model_name": model_name, "using_real_model": True}
results["model_info"] = attn_info
# First try: Use actual attention layer forward pass
attention_layer_works = False
try:
attention_layer = extract_attention_layer(model, layer_idx=0)
hidden_states, position_ids = create_attention_inputs(
model, batch_size, seq_len, device, dtype
)
# Test if attention layer works with first backend
test_result = benchmark_attention_layer(
attention_layer=attention_layer,
hidden_states=hidden_states,
position_ids=position_ids,
backend="flash",
num_iterations=2,
warmup_iterations=1,
)
if test_result.get("time_ms") is not None:
attention_layer_works = True
del hidden_states, position_ids
torch.cuda.empty_cache()
except Exception as layer_error:
print(f"[run_attention_benchmark] Attention layer extraction failed: {layer_error}")
attention_layer_works = False
if attention_layer_works:
# Use actual attention layer
hidden_states, position_ids = create_attention_inputs(
model, batch_size, seq_len, device, dtype
)
for backend in ["math", "flash", "mem_efficient"]:
result = benchmark_attention_layer(
attention_layer=attention_layer,
hidden_states=hidden_states,
position_ids=position_ids,
backend=backend,
num_iterations=num_iterations,
warmup_iterations=warmup_iterations,
)
results[backend] = result
del hidden_states, position_ids
torch.cuda.empty_cache()
else:
# Fallback: Use F.scaled_dot_product_attention with real model dimensions
print(f"[run_attention_benchmark] Falling back to SDPA with model dimensions")
results["fallback_mode"] = True
# Create Q, K, V tensors with real model dimensions
Q = torch.randn(batch_size, model_num_heads, seq_len, model_head_dim, device=device, dtype=dtype)
K = torch.randn(batch_size, model_num_heads, seq_len, model_head_dim, device=device, dtype=dtype)
V = torch.randn(batch_size, model_num_heads, seq_len, model_head_dim, device=device, dtype=dtype)
backends = [
("math", True, False, False),
("flash", False, True, False),
("mem_efficient", False, False, True),
]
for backend_name, enable_math, enable_flash, enable_mem_efficient in backends:
try:
torch.cuda.reset_peak_memory_stats()
torch.cuda.synchronize()
with torch.backends.cuda.sdp_kernel(
enable_flash=enable_flash,
enable_math=enable_math,
enable_mem_efficient=enable_mem_efficient
):
# Warmup
for _ in range(warmup_iterations):
_ = F.scaled_dot_product_attention(Q, K, V)
torch.cuda.synchronize()
# Timed runs
start = torch.cuda.Event(enable_timing=True)
end = torch.cuda.Event(enable_timing=True)
start.record()
for _ in range(num_iterations):
_ = F.scaled_dot_product_attention(Q, K, V)
end.record()
torch.cuda.synchronize()
time_ms = start.elapsed_time(end) / num_iterations
memory_mb = torch.cuda.max_memory_allocated() / 1e6
results[backend_name] = {
"time_ms": round(time_ms, 3),
"memory_mb": round(memory_mb, 1),
"status": "success"
}
except Exception as e:
results[backend_name] = {
"time_ms": None,
"memory_mb": None,
"status": f"error: {str(e)[:50]}"
}
del Q, K, V
torch.cuda.empty_cache()
# Calculate speedups
if results.get("math", {}).get("time_ms"):
base_time = results["math"]["time_ms"]
for backend in ["math", "flash", "mem_efficient"]:
if results.get(backend, {}).get("time_ms"):
results[backend]["speedup"] = round(base_time / results[backend]["time_ms"], 2)
return results
except Exception as e:
return {"error": f"Failed to load model: {str(e)[:100]}"}
# Legacy mode: Use raw SDPA with random tensors (fallback)
results = {"using_real_model": False}
# Create input tensors
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)
# Test each backend
backends = [
("math", True, False, False),
("flash", False, True, False),
("mem_efficient", False, False, True),
]
for backend_name, enable_math, enable_flash, enable_mem_efficient in backends:
try:
torch.cuda.reset_peak_memory_stats()
torch.cuda.synchronize()
with torch.backends.cuda.sdp_kernel(
enable_flash=enable_flash,
enable_math=enable_math,
enable_mem_efficient=enable_mem_efficient
):
# Warmup
for _ in range(warmup_iterations):
_ = F.scaled_dot_product_attention(Q, K, V)
torch.cuda.synchronize()
# Timed runs
start = torch.cuda.Event(enable_timing=True)
end = torch.cuda.Event(enable_timing=True)
start.record()
for _ in range(num_iterations):
_ = F.scaled_dot_product_attention(Q, K, V)
end.record()
torch.cuda.synchronize()
time_ms = start.elapsed_time(end) / num_iterations
memory_mb = torch.cuda.max_memory_allocated() / 1e6
results[backend_name] = {
"time_ms": round(time_ms, 3),
"memory_mb": round(memory_mb, 1),
"status": "success"
}
except Exception as e:
results[backend_name] = {
"time_ms": None,
"memory_mb": None,
"status": f"error: {str(e)[:50]}"
}
# Calculate speedups relative to math backend
if results.get("math", {}).get("time_ms"):
base_time = results["math"]["time_ms"]
for backend in results:
if isinstance(results[backend], dict) and results[backend].get("time_ms"):
results[backend]["speedup"] = round(base_time / results[backend]["time_ms"], 2)
# Clean up
del Q, K, V
torch.cuda.empty_cache()
return results
def run_scaling_benchmark(
model_name: str = None,
seq_lengths: list = None,
batch_size: int = 1,
# Legacy parameters (used if model_name is None)
num_heads: int = 16,
head_dim: int = 64,
) -> dict:
"""
Benchmark attention backends across multiple sequence lengths using a real model.
Args:
model_name: Name of the model from MODEL_CONFIGS (e.g., "SmolLM2-360M")
seq_lengths: List of sequence lengths to test
batch_size: Batch size
num_heads: (Legacy) Number of attention heads if model_name is None
head_dim: (Legacy) Dimension per head if model_name is None
Returns:
Dict with arrays of timing and memory results for each backend
"""
if seq_lengths is None:
seq_lengths = [512, 1024, 2048, 4096]
if not torch.cuda.is_available():
return {"error": "CUDA not available"}
results = {
"seq_lengths": seq_lengths,
"model_name": model_name,
"math": {"time_ms": [], "memory_mb": []},
"flash": {"time_ms": [], "memory_mb": []},
"mem_efficient": {"time_ms": [], "memory_mb": []},
}
for seq_len in seq_lengths:
bench_result = run_attention_benchmark(
model_name=model_name,
seq_len=seq_len,
batch_size=batch_size,
num_iterations=5, # Fewer iterations for scaling test
warmup_iterations=2,
# Legacy params (ignored if model_name is set)
num_heads=num_heads,
head_dim=head_dim,
)
for backend in ["math", "flash", "mem_efficient"]:
if bench_result.get(backend, {}).get("time_ms"):
results[backend]["time_ms"].append(bench_result[backend]["time_ms"])
results[backend]["memory_mb"].append(bench_result[backend]["memory_mb"])
else:
results[backend]["time_ms"].append(None)
results[backend]["memory_mb"].append(None)
return results
def create_benchmark_results_table(results: dict) -> str:
"""Create a markdown table from benchmark results."""
if "error" in results:
return f"**Error:** {results['error']}"
# Build table
lines = [
"| Backend | Time (ms) | Memory (MB) | Speedup |",
"|---------|-----------|-------------|---------|",
]
for backend in ["math", "flash", "mem_efficient"]:
if backend in results:
r = results[backend]
name = ATTENTION_BACKENDS.get(backend, backend)
time_str = f"{r['time_ms']:.2f}" if r.get('time_ms') else "N/A"
mem_str = f"{r['memory_mb']:.0f}" if r.get('memory_mb') else "N/A"
speedup_str = f"{r.get('speedup', 1.0):.1f}×"
lines.append(f"| {name} | {time_str} | {mem_str} | {speedup_str} |")
return "\n".join(lines)
def create_benchmark_insight(results: dict) -> str:
"""Create insight text from benchmark results."""
if "error" in results:
return ""
flash = results.get("flash", {})
math = results.get("math", {})
if not flash.get("time_ms") or not math.get("time_ms"):
return "**Note:** Some backends may not be available on this GPU."
speedup = math["time_ms"] / flash["time_ms"]
mem_reduction = math["memory_mb"] / flash["memory_mb"] if flash["memory_mb"] > 0 else 1
return f"""**Key Insight:**
FlashAttention is **{speedup:.1f}× faster** and uses **{mem_reduction:.1f}× less memory**!
This improvement comes from:
- Tiling attention into SRAM-sized blocks
- Never materializing the full N×N attention matrix in HBM
- Fused kernel avoiding multiple HBM round-trips"""
def create_scaling_chart(results: dict) -> go.Figure:
"""Create a scaling chart showing time and memory vs sequence length."""
if "error" in results:
fig = go.Figure()
fig.add_annotation(
x=0.5, y=0.5,
text=f"Error: {results['error']}",
showarrow=False,
font=dict(size=16, color="red")
)
return fig
seq_lengths = results["seq_lengths"]
# Create subplot with two y-axes
fig = make_subplots(
rows=1, cols=2,
subplot_titles=("Execution Time", "Peak Memory"),
horizontal_spacing=0.12,
)
colors = {
"math": "rgba(239, 68, 68, 0.8)", # Red
"flash": "rgba(34, 197, 94, 0.8)", # Green
"mem_efficient": "rgba(59, 130, 246, 0.8)", # Blue
}
# Plot time
for backend in ["math", "flash", "mem_efficient"]:
times = results[backend]["time_ms"]
name = ATTENTION_BACKENDS.get(backend, backend)
# Filter out None values
valid_points = [(s, t) for s, t in zip(seq_lengths, times) if t is not None]
if valid_points:
x_vals, y_vals = zip(*valid_points)
fig.add_trace(
go.Scatter(
x=list(x_vals),
y=list(y_vals),
mode="lines+markers",
name=name,
line=dict(color=colors[backend], width=2),
marker=dict(size=8),
legendgroup=backend,
),
row=1, col=1
)
# Plot memory
for backend in ["math", "flash", "mem_efficient"]:
memory = results[backend]["memory_mb"]
name = ATTENTION_BACKENDS.get(backend, backend)
valid_points = [(s, m) for s, m in zip(seq_lengths, memory) if m is not None]
if valid_points:
x_vals, y_vals = zip(*valid_points)
fig.add_trace(
go.Scatter(
x=list(x_vals),
y=list(y_vals),
mode="lines+markers",
name=name,
line=dict(color=colors[backend], width=2),
marker=dict(size=8),
legendgroup=backend,
showlegend=False,
),
row=1, col=2
)
fig.update_xaxes(title_text="Sequence Length", row=1, col=1)
fig.update_xaxes(title_text="Sequence Length", row=1, col=2)
fig.update_yaxes(title_text="Time (ms)", row=1, col=1)
fig.update_yaxes(title_text="Memory (MB)", row=1, col=2)
fig.update_layout(
height=350,
margin=dict(l=50, r=50, t=50, b=50),
legend=dict(
orientation="h",
yanchor="bottom",
y=-0.3,
xanchor="center",
x=0.5
),
)
return fig
def calculate_attention_flops(seq_len: int, num_heads: int, head_dim: int, batch_size: int = 1) -> float:
"""
Calculate FLOPs for scaled dot-product attention.
FLOPs breakdown:
- Q @ K^T: 2 * batch * heads * seq * seq * head_dim
- Softmax: ~5 * batch * heads * seq * seq (exp, sum, div)
- P @ V: 2 * batch * heads * seq * seq * head_dim
Total: ~4 * batch * heads * seq² * head_dim + 5 * batch * heads * seq²
"""
qk_flops = 2 * batch_size * num_heads * seq_len * seq_len * head_dim
softmax_flops = 5 * batch_size * num_heads * seq_len * seq_len
pv_flops = 2 * batch_size * num_heads * seq_len * seq_len * head_dim
return qk_flops + softmax_flops + pv_flops
def calculate_memory_traffic(
seq_len: int,
num_heads: int,
head_dim: int,
batch_size: int = 1,
is_flash: bool = False,
dtype_bytes: int = 2, # FP16
) -> float:
"""
Calculate memory traffic in bytes for attention.
Standard Attention:
- Read Q, K, V: 3 * batch * heads * seq * head_dim * dtype_bytes
- Write S = Q @ K^T: batch * heads * seq * seq * dtype_bytes
- Read S for softmax: batch * heads * seq * seq * dtype_bytes
- Write P = softmax(S): batch * heads * seq * seq * dtype_bytes
- Read P and V: batch * heads * seq * seq + batch * heads * seq * head_dim
- Write O: batch * heads * seq * head_dim * dtype_bytes
FlashAttention:
- Read Q, K, V once: 3 * batch * heads * seq * head_dim * dtype_bytes
- Write O once: batch * heads * seq * head_dim * dtype_bytes
- No attention matrix written to HBM!
"""
qkv_size = 3 * batch_size * num_heads * seq_len * head_dim * dtype_bytes
output_size = batch_size * num_heads * seq_len * head_dim * dtype_bytes
if is_flash:
# FlashAttention: Only Q, K, V reads + O write
return qkv_size + output_size
else:
# Standard: Also materializes attention matrix (read + write twice)
attention_matrix_size = batch_size * num_heads * seq_len * seq_len * dtype_bytes
return qkv_size + output_size + 3 * attention_matrix_size
def calculate_roofline_metrics(
results: dict,
seq_len: int,
num_heads: int,
head_dim: int,
batch_size: int = 1,
) -> dict:
"""
Calculate arithmetic intensity and achieved TFLOPS from benchmark results.
Returns dict with measured metrics for each backend.
"""
flops = calculate_attention_flops(seq_len, num_heads, head_dim, batch_size)
metrics = {}
for backend in ["math", "flash", "mem_efficient"]:
if backend not in results or results[backend].get("time_ms") is None:
continue
time_ms = results[backend]["time_ms"]
time_s = time_ms / 1000.0
# Calculate achieved TFLOPS
achieved_tflops = (flops / time_s) / 1e12
# Calculate memory traffic (approximation)
is_flash = backend in ["flash", "mem_efficient"]
memory_bytes = calculate_memory_traffic(
seq_len, num_heads, head_dim, batch_size, is_flash=is_flash
)
# Arithmetic intensity = FLOPs / bytes
arith_intensity = flops / memory_bytes
metrics[backend] = {
"flops": flops,
"memory_bytes": memory_bytes,
"time_ms": time_ms,
"achieved_tflops": achieved_tflops,
"arith_intensity": arith_intensity,
}
return metrics
def create_roofline_chart(
results: dict,
gpu_specs: dict = None,
benchmark_metrics: dict = None,
) -> go.Figure:
"""
Create a roofline chart showing where different attention implementations fall.
The roofline model shows:
- X-axis: Arithmetic intensity (FLOPs per byte of memory traffic)
- Y-axis: Performance (TFLOPS)
- The roofline is min(peak_compute, bandwidth * intensity)
Args:
results: Benchmark results dict (can be empty)
gpu_specs: GPU specifications dict (from detect_gpu() or GPU_SPECS)
benchmark_metrics: Roofline metrics from calculate_roofline_metrics()
If benchmark_metrics is provided, plots MEASURED values.
Otherwise, plots theoretical approximations.
"""
# Use provided specs or default to A10G
if gpu_specs is None:
gpu = GPU_SPECS[DEFAULT_GPU]
else:
gpu = gpu_specs
peak_tflops = gpu["tflops_fp16"]
bandwidth_gbps = gpu["bandwidth_gbps"]
# Ridge point: where memory-bound meets compute-bound
ridge_point = (peak_tflops * 1e12) / (bandwidth_gbps * 1e9)
# Create figure
fig = go.Figure()
# Roofline curve
x_range = np.logspace(0, 3, 100)
y_roofline = np.minimum(
peak_tflops,
bandwidth_gbps * x_range / 1000
)
fig.add_trace(go.Scatter(
x=x_range,
y=y_roofline,
mode="lines",
name="Roofline",
line=dict(color="rgba(0, 0, 0, 0.6)", width=2),
))
# Memory-bound region (dashed)
fig.add_trace(go.Scatter(
x=[1, ridge_point],
y=[bandwidth_gbps / 1000, peak_tflops],
mode="lines",
name="Memory Bound",
line=dict(color="rgba(239, 68, 68, 0.5)", width=3, dash="dash"),
))
# Compute-bound region (dashed)
fig.add_trace(go.Scatter(
x=[ridge_point, 1000],
y=[peak_tflops, peak_tflops],
mode="lines",
name="Compute Bound",
line=dict(color="rgba(34, 197, 94, 0.5)", width=3, dash="dash"),
))
# Determine if we have measured data or should use theoretical
use_measured = benchmark_metrics is not None and len(benchmark_metrics) > 0
if use_measured:
# Plot MEASURED data points
title_suffix = " (Measured)"
# Math/Standard backend
if "math" in benchmark_metrics:
m = benchmark_metrics["math"]
fig.add_trace(go.Scatter(
x=[m["arith_intensity"]],
y=[m["achieved_tflops"]],
mode="markers",
name=f"Math ({m['achieved_tflops']:.1f} TFLOPS, {m['time_ms']:.1f}ms)",
marker=dict(size=16, color="#dc2626", symbol="circle",
line=dict(color="white", width=2)),
))
# Add label as annotation for better visibility
fig.add_annotation(
x=np.log10(m["arith_intensity"]),
y=m["achieved_tflops"],
text=f"<b>Math</b><br>{m['time_ms']:.1f}ms",
showarrow=True,
arrowhead=2,
arrowsize=1,
arrowwidth=1,
arrowcolor="#dc2626",
ax=0,
ay=-40,
font=dict(size=10, color="#dc2626"),
bgcolor="rgba(255, 255, 255, 0.95)",
bordercolor="#dc2626",
borderwidth=1,
borderpad=3,
)
# Flash backend
if "flash" in benchmark_metrics:
m = benchmark_metrics["flash"]
fig.add_trace(go.Scatter(
x=[m["arith_intensity"]],
y=[m["achieved_tflops"]],
mode="markers",
name=f"Flash ({m['achieved_tflops']:.1f} TFLOPS, {m['time_ms']:.1f}ms)",
marker=dict(size=16, color="#16a34a", symbol="circle",
line=dict(color="white", width=2)),
))
fig.add_annotation(
x=np.log10(m["arith_intensity"]),
y=m["achieved_tflops"],
text=f"<b>Flash</b><br>{m['time_ms']:.1f}ms",
showarrow=True,
arrowhead=2,
arrowsize=1,
arrowwidth=1,
arrowcolor="#16a34a",
ax=0,
ay=-40,
font=dict(size=10, color="#16a34a"),
bgcolor="rgba(255, 255, 255, 0.95)",
bordercolor="#16a34a",
borderwidth=1,
borderpad=3,
)
# Memory-efficient backend
if "mem_efficient" in benchmark_metrics:
m = benchmark_metrics["mem_efficient"]
fig.add_trace(go.Scatter(
x=[m["arith_intensity"]],
y=[m["achieved_tflops"]],
mode="markers",
name=f"MemEff ({m['achieved_tflops']:.1f} TFLOPS, {m['time_ms']:.1f}ms)",
marker=dict(size=16, color="#2563eb", symbol="circle",
line=dict(color="white", width=2)),
))
fig.add_annotation(
x=np.log10(m["arith_intensity"]),
y=m["achieved_tflops"],
text=f"<b>MemEff</b><br>{m['time_ms']:.1f}ms",
showarrow=True,
arrowhead=2,
arrowsize=1,
arrowwidth=1,
arrowcolor="#2563eb",
ax=30, # Offset to avoid overlap
ay=-30,
font=dict(size=10, color="#2563eb"),
bgcolor="rgba(255, 255, 255, 0.95)",
bordercolor="#2563eb",
borderwidth=1,
borderpad=3,
)
else:
# Plot THEORETICAL approximations
title_suffix = " (Theoretical)"
# Standard attention - memory bound
std_intensity = 10
std_achieved = min(peak_tflops * 0.15, bandwidth_gbps * std_intensity / 1000)
fig.add_trace(go.Scatter(
x=[std_intensity],
y=[std_achieved],
mode="markers",
name="Standard (Theoretical)",
marker=dict(size=15, color="rgba(220, 38, 38, 0.6)", symbol="circle-open",
line=dict(width=2)),
))
fig.add_annotation(
x=np.log10(std_intensity),
y=std_achieved,
text="<b>Standard</b><br>(theoretical)",
showarrow=True,
arrowhead=2,
ax=0,
ay=-35,
font=dict(size=10, color="#dc2626"),
bgcolor="rgba(255, 255, 255, 0.9)",
bordercolor="rgba(220, 38, 38, 0.5)",
borderwidth=1,
borderpad=3,
)
# FlashAttention - compute bound
flash_intensity = 200
flash_achieved = min(peak_tflops * 0.7, bandwidth_gbps * flash_intensity / 1000)
fig.add_trace(go.Scatter(
x=[flash_intensity],
y=[flash_achieved],
mode="markers",
name="Flash (Theoretical)",
marker=dict(size=15, color="rgba(22, 163, 74, 0.6)", symbol="circle-open",
line=dict(width=2)),
))
fig.add_annotation(
x=np.log10(flash_intensity),
y=flash_achieved,
text="<b>FlashAttention</b><br>(theoretical)",
showarrow=True,
arrowhead=2,
ax=0,
ay=-35,
font=dict(size=10, color="#16a34a"),
bgcolor="rgba(255, 255, 255, 0.9)",
bordercolor="rgba(22, 163, 74, 0.5)",
borderwidth=1,
borderpad=3,
)
# Add ridge point marker
fig.add_trace(go.Scatter(
x=[ridge_point],
y=[peak_tflops],
mode="markers",
name=f"Ridge Point ({ridge_point:.0f} FLOPs/byte)",
marker=dict(size=10, color="rgba(0, 0, 0, 0.6)", symbol="diamond"),
))
# Add annotations with better visibility (white background)
fig.add_annotation(
x=np.log10(5),
y=peak_tflops * 0.1,
text="<b>Memory Bound</b><br>(limited by bandwidth)",
showarrow=False,
font=dict(size=11, color="#dc2626"), # Solid red
bgcolor="rgba(255, 255, 255, 0.9)",
bordercolor="#dc2626",
borderwidth=1,
borderpad=4,
)
fig.add_annotation(
x=np.log10(300),
y=peak_tflops * 0.65,
text="<b>Compute Bound</b><br>(limited by TFLOPS)",
showarrow=False,
font=dict(size=11, color="#16a34a"), # Solid green
bgcolor="rgba(255, 255, 255, 0.9)",
bordercolor="#16a34a",
borderwidth=1,
borderpad=4,
)
# Use detected_name if available, otherwise use name
display_name = gpu.get("detected_name", gpu.get("name", "GPU"))
# Add estimated indicator if specs were estimated
estimated_note = " (estimated specs)" if gpu.get("estimated") else ""
fig.update_layout(
title=dict(
text=f"Roofline Model: {display_name}{title_suffix}{estimated_note}<br>"
f"<span style='font-size:12px;color:#666'>"
f"Peak: {peak_tflops} TFLOPS | Bandwidth: {bandwidth_gbps} GB/s</span>",
x=0.5,
font=dict(size=14),
),
xaxis=dict(
title="Arithmetic Intensity (FLOPs/byte)",
type="log",
range=[0, 3],
),
yaxis=dict(
title="Performance (TFLOPS)",
range=[0, peak_tflops * 1.2], # More headroom for text
),
height=420,
margin=dict(l=60, r=40, t=80, b=80), # More room for title and legend
legend=dict(
orientation="h",
yanchor="bottom",
y=-0.30,
xanchor="center",
x=0.5,
font=dict(size=10),
),
showlegend=True,
)
return fig
def get_roofline_insight(benchmark_metrics: dict = None) -> str:
"""Return insight text for the roofline chart."""
base_insight = """**Why FlashAttention is Faster:**
The roofline model reveals the key insight:
1. **Standard Attention** sits in the **memory-bound** region (left of ridge point)
- Limited by HBM bandwidth, not compute
- Reading/writing the N×N attention matrix dominates runtime
2. **FlashAttention** moves to the **compute-bound** region (right of ridge point)
- By never materializing the full attention matrix
- Arithmetic intensity increases ~20-50×
- Can now utilize most of the GPU's TFLOPS
*The same FLOPs, but 10× less memory traffic = faster execution!*"""
if benchmark_metrics and "math" in benchmark_metrics and "flash" in benchmark_metrics:
math_m = benchmark_metrics["math"]
flash_m = benchmark_metrics["flash"]
speedup = math_m["time_ms"] / flash_m["time_ms"]
intensity_ratio = flash_m["arith_intensity"] / math_m["arith_intensity"]
measured_insight = f"""
---
**📊 Measured Results:**
- **Math backend:** {math_m['achieved_tflops']:.1f} TFLOPS @ {math_m['arith_intensity']:.0f} FLOPs/byte
- **Flash backend:** {flash_m['achieved_tflops']:.1f} TFLOPS @ {flash_m['arith_intensity']:.0f} FLOPs/byte
- **Speedup:** {speedup:.1f}× faster
- **Intensity increase:** {intensity_ratio:.0f}× higher arithmetic intensity"""
return base_insight + measured_insight
return base_insight + "\n\n*Run a benchmark to see measured values on the chart!*"
|