| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | # ruff: noqa: E501 |
| 4 | import time |
| 5 | |
| 6 | import torch |
| 7 | |
| 8 | from vllm import _custom_ops as ops |
| 9 | from vllm.model_executor.layers.quantization.utils.fp8_utils import ( |
| 10 | per_token_group_quant_fp8, |
| 11 | w8a8_triton_block_scaled_mm, |
| 12 | ) |
| 13 | from vllm.triton_utils import triton |
| 14 | from vllm.utils.deep_gemm import ( |
| 15 | calc_diff, |
| 16 | fp8_gemm_nt, |
| 17 | per_block_cast_to_fp8, |
| 18 | ) |
| 19 | from vllm.utils.torch_utils import set_random_seed |
| 20 | |
| 21 | |
| 22 | def benchmark_shape( |
| 23 | m: int, |
| 24 | n: int, |
| 25 | k: int, |
| 26 | warmup: int = 100, |
| 27 | repeat: int = 10000, |
| 28 | verbose: bool = False, |
| 29 | ) -> dict: |
| 30 | """Benchmark all implementations for a specific (m, n, k) shape.""" |
| 31 | if verbose: |
| 32 | print(f"\n=== Benchmarking shape: m={m}, n={n}, k={k} ===") |
| 33 | |
| 34 | # Create test tensors |
| 35 | A = torch.randn((m, k), device="cuda", dtype=torch.bfloat16) |
| 36 | B = torch.randn((n, k), device="cuda", dtype=torch.bfloat16) |
| 37 | |
| 38 | # Reference result in BF16 |
| 39 | torch.accelerator.synchronize() |
| 40 | C_ref = A @ B.t() |
| 41 | |
| 42 | # Pre-quantize B for all implementations |
| 43 | # (weights can be pre-quantized offline) |
| 44 | B_deepgemm, B_scale_deepgemm = per_block_cast_to_fp8(B, [128, 128], use_ue8m0=True) |
| 45 | B_vllm, B_scale_vllm = per_block_cast_to_fp8(B, [128, 128], use_ue8m0=True) |
| 46 | |
| 47 | # Block size configuration |
| 48 | block_size = [128, 128] |
| 49 | |
| 50 | # Pre-quantize A for all implementations |
| 51 | A_deepgemm, A_scale_deepgemm = per_token_group_quant_fp8( |
| 52 | A, block_size[1], column_major_scales=True, tma_aligned_scales=True |
| 53 | ) |
| 54 | C_deepgemm = torch.empty((m, n), device="cuda", dtype=torch.bfloat16) |
| 55 | A_vllm, A_scale_vllm = per_token_group_quant_fp8(A, block_size[1]) |
| 56 | A_vllm_cutlass, A_scale_vllm_cutlass = per_token_group_quant_fp8( |
| 57 | A, block_size[1], column_major_scales=True |
| 58 | ) |
| 59 | |
| 60 | # === DeepGEMM Implementation === |
| 61 | def deepgemm_gemm(): |
| 62 | fp8_gemm_nt( |
| 63 | (A_deepgemm, A_scale_deepgemm), (B_deepgemm, B_scale_deepgemm), C_deepgemm |
| 64 | ) |
| 65 | return C_deepgemm |
| 66 | |
| 67 | # === vLLM Triton Implementation === |
| 68 | def vllm_triton_gemm(): |
| 69 | return w8a8_triton_block_scaled_mm( |
| 70 | A_vllm, |
| 71 | B_vllm, |
| 72 | A_scale_vllm, |
| 73 | B_scale_vllm, |
| 74 | block_size, |
| 75 | output_dtype=torch.bfloat16, |
| 76 | ) |
| 77 | |
| 78 | # === vLLM CUTLASS Implementation === |
| 79 | def vllm_cutlass_gemm(): |
| 80 | return ops.cutlass_scaled_mm( |
| 81 | A_vllm_cutlass, |
| 82 | B_vllm.T, |
| 83 | scale_a=A_scale_vllm_cutlass, |
| 84 | scale_b=B_scale_vllm.T, |
| 85 | out_dtype=torch.bfloat16, |
| 86 | ) |
| 87 | |
| 88 | # Run correctness check first |
| 89 | if verbose: |
| 90 | print("Running correctness check...") |
| 91 | C_deepgemm = deepgemm_gemm() |
| 92 | C_vllm_triton = vllm_triton_gemm() |
| 93 | C_vllm_cutlass = vllm_cutlass_gemm() |
| 94 | |
| 95 | deepgemm_diff = calc_diff(C_deepgemm, C_ref) |
| 96 | vllm_triton_diff = calc_diff(C_vllm_triton, C_ref) |
| 97 | vllm_cutlass_diff = calc_diff(C_vllm_cutlass, C_ref) |
| 98 | |
| 99 | if verbose: |
| 100 | print(f"DeepGEMM vs Reference difference: {deepgemm_diff:.6f}") |
| 101 | print(f"vLLM Triton vs Reference difference: {vllm_triton_diff:.6f}") |
| 102 | print(f"vLLM CUTLASS vs Reference difference: {vllm_cutlass_diff:.6f}") |
| 103 | print( |
| 104 | "vLLM Triton vs DeepGEMM difference: " |
| 105 | f"{calc_diff(C_vllm_triton, C_deepgemm):.6f}" |
| 106 | ) |
| 107 | print( |
| 108 | "vLLM CUTLASS vs DeepGEMM difference: " |
| 109 | f"{calc_diff(C_vllm_cutlass, C_deepgemm):.6f}" |
| 110 | ) |
| 111 | |
| 112 | # Benchmark implementations |
| 113 | implementations = { |
| 114 | "DeepGEMM": deepgemm_gemm, |
| 115 | "vLLM Triton": vllm_triton_gemm, |
| 116 | "vLLM CUTLASS": vllm_cutlass_gemm, |
| 117 | } |
| 118 | |
| 119 | benchmark_results = {"shape": {"m": m, "n": n, "k": k}, "implementations": {}} |
| 120 | |
| 121 | for name, func in implementations.items(): |
| 122 | # Warmup |
| 123 | for _ in range(warmup): |
| 124 | func() |
| 125 | torch.accelerator.synchronize() |
| 126 | |
| 127 | # Timing loop |
| 128 | torch.accelerator.synchronize() |
| 129 | start = time.time() |
| 130 | for _ in range(repeat): |
| 131 | func() |
| 132 | torch.accelerator.synchronize() |
| 133 | end = time.time() |
| 134 | |
| 135 | # Calculate timing and TFLOPS |
| 136 | avg_time_ms = (end - start) / repeat * 1000 |
| 137 | avg_time_us = avg_time_ms * 1000 |
| 138 | tflops = 2 * m * n * k / (avg_time_ms * 1e-3) / 1e12 |
| 139 | gb_s = (m * k + k * n + m * n * 2) / 1e9 / (avg_time_ms * 1e-3) |
| 140 | |
| 141 | benchmark_results["implementations"][name] = { |
| 142 | "time_ms": avg_time_ms, |
| 143 | "time_us": avg_time_us, |
| 144 | "tflops": tflops, |
| 145 | "gb_s": gb_s, |
| 146 | "diff": { |
| 147 | "DeepGEMM": 0.0 |
| 148 | if name == "DeepGEMM" |
| 149 | else calc_diff(func(), C_deepgemm), |
| 150 | "Reference": deepgemm_diff |
| 151 | if name == "DeepGEMM" |
| 152 | else (vllm_triton_diff if name == "vLLM Triton" else vllm_cutlass_diff), |
| 153 | }, |
| 154 | } |
| 155 | |
| 156 | if verbose: |
| 157 | print(f"{name}: {avg_time_ms:.3f} ms, {tflops:.2f} TFLOPS, {gb_s:.2f} GB/s") |
| 158 | |
| 159 | # Calculate speedups |
| 160 | baseline = benchmark_results["implementations"]["DeepGEMM"]["time_ms"] |
| 161 | for name, data in benchmark_results["implementations"].items(): |
| 162 | if name != "DeepGEMM": |
| 163 | speedup = baseline / data["time_ms"] |
| 164 | benchmark_results["implementations"][name]["speedup_vs_deepgemm"] = speedup |
| 165 | if verbose: |
| 166 | print( |
| 167 | f"DeepGEMM is {1 / speedup:.2f}x " |
| 168 | f"{'faster' if 1 / speedup > 1 else 'slower'} than {name}" |
| 169 | ) |
| 170 | |
| 171 | vllm_triton_time = benchmark_results["implementations"]["vLLM Triton"]["time_ms"] |
| 172 | vllm_cutlass_time = benchmark_results["implementations"]["vLLM CUTLASS"]["time_ms"] |
| 173 | cutlass_vs_triton = vllm_triton_time / vllm_cutlass_time |
| 174 | benchmark_results["implementations"]["vLLM CUTLASS"]["speedup_vs_triton"] = ( |
| 175 | cutlass_vs_triton |
| 176 | ) |
| 177 | if verbose: |
| 178 | print( |
| 179 | f"vLLM CUTLASS is {cutlass_vs_triton:.2f}x " |
| 180 | f"{'faster' if cutlass_vs_triton > 1 else 'slower'} than vLLM Triton" |
| 181 | ) |
| 182 | |
| 183 | return benchmark_results |
| 184 | |
| 185 | |
| 186 | def format_table_row(values, widths): |
| 187 | """Format a row with specified column widths.""" |
| 188 | return "| " + " | ".join(f"{val:{w}}" for val, w in zip(values, widths)) + " |" |
| 189 | |
| 190 | |
| 191 | def print_table(headers, rows, title=None): |
| 192 | """Print a table with headers and rows.""" |
| 193 | if title: |
| 194 | print(f"\n{title}") |
| 195 | |
| 196 | # Calculate column widths based on headers and data |
| 197 | widths = [ |
| 198 | max(len(str(h)), max(len(str(row[i])) for row in rows)) |
| 199 | for i, h in enumerate(headers) |
| 200 | ] |
| 201 | |
| 202 | # Create separator line |
| 203 | separator = "+-" + "-+-".join("-" * w for w in widths) + "-+" |
| 204 | |
| 205 | # Print table |
| 206 | print(separator) |
| 207 | print(format_table_row(headers, widths)) |
| 208 | print(separator) |
| 209 | for row in rows: |
| 210 | print(format_table_row(row, widths)) |
| 211 | print(separator) |
| 212 | |
| 213 | |
| 214 | def format_speedup(value): |
| 215 | """Format speedup value with indicator if it's faster or slower.""" |
| 216 | return f"{value:.2f}x {'faster' if value > 1.0 else 'slower'}" |
| 217 | |
| 218 | |
| 219 | def run_benchmarks(verbose: bool = False): |
| 220 | """Run benchmarks for a set of common shapes.""" |
| 221 | print("===== STARTING FP8 GEMM BENCHMARK =====") |
| 222 | |
| 223 | # Make sure we're using the GPU |
| 224 | if not torch.cuda.is_available(): |
| 225 | print("CUDA not available! Tests require GPU.") |
| 226 | return |
| 227 | |
| 228 | # Print system information |
| 229 | print(f"PyTorch version: {torch.__version__}") |
| 230 | print(f"CUDA version: {torch.version.cuda}") |
| 231 | print(f"Triton version: {triton.__version__}") |
| 232 | print(f"Using device: {torch.cuda.get_device_name()}") |
| 233 | |
| 234 | # Enable TF32 for better performance |
| 235 | torch.backends.cuda.matmul.allow_tf32 = True |
| 236 | torch.backends.cudnn.allow_tf32 = True |
| 237 | |
| 238 | # Set seeds for reproducibility |
| 239 | set_random_seed(42) |
| 240 | # Define benchmark shapes (m, n, k) |
| 241 | shapes = [ |
| 242 | (8, 4096, 7168), |
| 243 | (8, 7168, 18432), |
| 244 | (8, 18432, 7168), |
| 245 | (64, 4096, 7168), |
| 246 | (64, 7168, 18432), |
| 247 | (64, 18432, 7168), |
| 248 | (64, 24576, 1536), |
| 249 | (64, 32768, 512), |
| 250 | (64, 7168, 16384), |
| 251 | (128, 4096, 7168), |
| 252 | (128, 7168, 18432), |
| 253 | (128, 18432, 7168), |
| 254 | (1024, 4096, 7168), |
| 255 | (1024, 18432, 7168), |
| 256 | (2048, 4096, 7168), |
| 257 | (4096, 4096, 7168), |
| 258 | ] |
| 259 | shapes = [ |
| 260 | # (64, 2112, 7168), |
| 261 | (64, 24576, 1536), |
| 262 | (64, 32768, 512), |
| 263 | (64, 7168, 16384), |
| 264 | (64, 4096, 7168), |
| 265 | (64, 7168, 2048), |
| 266 | # (128, 2112, 7168), |
| 267 | (128, 24576, 1536), |
| 268 | (128, 32768, 512), |
| 269 | (128, 7168, 16384), |
| 270 | (128, 4096, 7168), |
| 271 | (128, 7168, 2048), |
| 272 | # (4096, 2112, 7168), |
| 273 | (4096, 24576, 1536), |
| 274 | (4096, 32768, 512), |
| 275 | (4096, 7168, 16384), |
| 276 | (4096, 4096, 7168), |
| 277 | (4096, 7168, 2048), |
| 278 | ] |
| 279 | |
| 280 | all_results = [] |
| 281 | for m, n, k in shapes: |
| 282 | result = benchmark_shape(m, n, k, verbose=verbose) |
| 283 | all_results.append(result) |
| 284 | |
| 285 | # Print results in a nicely formatted table |
| 286 | print("\n===== PERFORMANCE COMPARISON =====") |
| 287 | |
| 288 | # Print DeepGEMM table |
| 289 | deepgemm_headers = ["m", "n", "k", "Time (μs)", "TFLOPS", "GB/s"] |
| 290 | deepgemm_rows = [] |
| 291 | for result in all_results: |
| 292 | shape = result["shape"] |
| 293 | impl_data = result["implementations"]["DeepGEMM"] |
| 294 | deepgemm_rows.append( |
| 295 | [ |
| 296 | shape["m"], |
| 297 | shape["n"], |
| 298 | shape["k"], |
| 299 | f"{impl_data['time_us']:.1f}", |
| 300 | f"{impl_data['tflops']:.1f}", |
| 301 | f"{impl_data['gb_s']:.1f}", |
| 302 | ] |
| 303 | ) |
| 304 | |
| 305 | print_table(deepgemm_headers, deepgemm_rows, title="DeepGEMM Implementation:") |
| 306 | |
| 307 | # Print vLLM Triton table |
| 308 | triton_headers = ["m", "n", "k", "Time (μs)", "TFLOPS", "GB/s", "vs DeepGEMM"] |
| 309 | triton_rows = [] |
| 310 | for result in all_results: |
| 311 | shape = result["shape"] |
| 312 | impl_data = result["implementations"]["vLLM Triton"] |
| 313 | speedup = impl_data.get("speedup_vs_deepgemm", 1.0) |
| 314 | triton_rows.append( |
| 315 | [ |
| 316 | shape["m"], |
| 317 | shape["n"], |
| 318 | shape["k"], |
| 319 | f"{impl_data['time_us']:.1f}", |
| 320 | f"{impl_data['tflops']:.1f}", |
| 321 | f"{impl_data['gb_s']:.1f}", |
| 322 | format_speedup(speedup), |
| 323 | ] |
| 324 | ) |
| 325 | |
| 326 | print_table(triton_headers, triton_rows, title="vLLM Triton Implementation:") |
| 327 | |
| 328 | # Print vLLM CUTLASS table |
| 329 | cutlass_headers = [ |
| 330 | "m", |
| 331 | "n", |
| 332 | "k", |
| 333 | "Time (μs)", |
| 334 | "TFLOPS", |
| 335 | "GB/s", |
| 336 | "vs DeepGEMM", |
| 337 | "vs Triton", |
| 338 | ] |
| 339 | cutlass_rows = [] |
| 340 | for result in all_results: |
| 341 | shape = result["shape"] |
| 342 | impl_data = result["implementations"]["vLLM CUTLASS"] |
| 343 | vs_deepgemm = impl_data.get("speedup_vs_deepgemm", 1.0) |
| 344 | vs_triton = impl_data.get("speedup_vs_triton", 1.0) |
| 345 | cutlass_rows.append( |
| 346 | [ |
| 347 | shape["m"], |
| 348 | shape["n"], |
| 349 | shape["k"], |
| 350 | f"{impl_data['time_us']:.1f}", |
| 351 | f"{impl_data['tflops']:.1f}", |
| 352 | f"{impl_data['gb_s']:.1f}", |
| 353 | format_speedup(vs_deepgemm), |
| 354 | format_speedup(vs_triton), |
| 355 | ] |
| 356 | ) |
| 357 | |
| 358 | print_table(cutlass_headers, cutlass_rows, title="vLLM CUTLASS Implementation:") |
| 359 | |
| 360 | # Calculate and print averages |
| 361 | print("\n===== AVERAGE PERFORMANCE =====") |
| 362 | |
| 363 | implementations = ["DeepGEMM", "vLLM Triton", "vLLM CUTLASS"] |
| 364 | avg_metrics = { |
| 365 | impl: {"tflops": 0, "gb_s": 0, "time_ms": 0} for impl in implementations |
| 366 | } |
| 367 | |
| 368 | for result in all_results: |
| 369 | for impl in implementations: |
| 370 | impl_data = result["implementations"][impl] |
| 371 | avg_metrics[impl]["tflops"] += impl_data["tflops"] |
| 372 | avg_metrics[impl]["gb_s"] += impl_data["gb_s"] |
| 373 | avg_metrics[impl]["time_ms"] += impl_data["time_ms"] |
| 374 | |
| 375 | num_shapes = len(all_results) |
| 376 | avg_headers = ["Implementation", "Avg TFLOPS", "Avg GB/s", "Avg Time (ms)"] |
| 377 | avg_rows = [] |
| 378 | |
| 379 | for impl in implementations: |
| 380 | avg_tflops = avg_metrics[impl]["tflops"] / num_shapes |
| 381 | avg_mem_bw = avg_metrics[impl]["gb_s"] / num_shapes |
| 382 | avg_time = avg_metrics[impl]["time_ms"] / num_shapes |
| 383 | avg_rows.append( |
| 384 | [impl, f"{avg_tflops:.2f}", f"{avg_mem_bw:.2f}", f"{avg_time:.2f}"] |
| 385 | ) |
| 386 | |
| 387 | print_table(avg_headers, avg_rows) |
| 388 | |
| 389 | # Calculate average speedups |
| 390 | avg_speedups = { |
| 391 | "DeepGEMM vs vLLM Triton": 0, |
| 392 | "DeepGEMM vs vLLM CUTLASS": 0, |
| 393 | "vLLM CUTLASS vs vLLM Triton": 0, |
| 394 | } |
| 395 | |
| 396 | for result in all_results: |
| 397 | deepgemm_time = result["implementations"]["DeepGEMM"]["time_ms"] |
| 398 | vllm_triton_time = result["implementations"]["vLLM Triton"]["time_ms"] |
| 399 | vllm_cutlass_time = result["implementations"]["vLLM CUTLASS"]["time_ms"] |
| 400 | |
| 401 | avg_speedups["DeepGEMM vs vLLM Triton"] += vllm_triton_time / deepgemm_time |
| 402 | avg_speedups["DeepGEMM vs vLLM CUTLASS"] += vllm_cutlass_time / deepgemm_time |
| 403 | avg_speedups["vLLM CUTLASS vs vLLM Triton"] += ( |
| 404 | vllm_triton_time / vllm_cutlass_time |
| 405 | ) |
| 406 | |
| 407 | print("\n===== AVERAGE SPEEDUPS =====") |
| 408 | speedup_headers = ["Comparison", "Speedup"] |
| 409 | speedup_rows = [] |
| 410 | for comparison, total in avg_speedups.items(): |
| 411 | avg_speedup = total / num_shapes |
| 412 | status = "faster" if avg_speedup > 1 else "slower" |
| 413 | speedup_rows.append([comparison, f"{avg_speedup:.2f}x {status}"]) |
| 414 | |
| 415 | print_table(speedup_headers, speedup_rows) |
| 416 | |
| 417 | # Average accuracy comparison |
| 418 | print("\n===== ACCURACY COMPARISON =====") |
| 419 | avg_diff = {impl: 0 for impl in implementations} |
| 420 | |
| 421 | for result in all_results: |
| 422 | for impl in implementations: |
| 423 | avg_diff[impl] += result["implementations"][impl]["diff"]["Reference"] |
| 424 | |
| 425 | diff_headers = ["Implementation", "Avg Diff vs Reference"] |
| 426 | diff_rows = [] |
| 427 | for impl in implementations: |
| 428 | diff_rows.append([impl, f"{avg_diff[impl] / num_shapes:.6f}"]) |
| 429 | |
| 430 | print_table(diff_headers, diff_rows) |
| 431 | |
| 432 | |
| 433 | if __name__ == "__main__": |
| 434 | run_benchmarks(verbose=False) |