| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | import itertools |
| 4 | from collections.abc import Callable |
| 5 | from unittest.mock import patch |
| 6 | |
| 7 | import pandas as pd |
| 8 | import torch |
| 9 | |
| 10 | from vllm.benchmarks.lib.utils import default_vllm_config |
| 11 | from vllm.model_executor.layers.quantization.input_quant_fp8 import QuantFP8 |
| 12 | from vllm.model_executor.layers.quantization.utils.quant_utils import GroupShape |
| 13 | from vllm.triton_utils import triton |
| 14 | from vllm.utils.argparse_utils import FlexibleArgumentParser |
| 15 | from vllm.utils.torch_utils import STR_DTYPE_TO_TORCH_DTYPE |
| 16 | |
| 17 | |
| 18 | def with_triton_mode(fn): |
| 19 | """Temporarily force the Triton fallback path.""" |
| 20 | |
| 21 | def wrapped(*args, **kwargs): |
| 22 | with patch("vllm.platforms.current_platform.is_cuda", return_value=False): |
| 23 | return fn(*args, **kwargs) |
| 24 | |
| 25 | return wrapped |
| 26 | |
| 27 | |
| 28 | # TODO(luka): use standalone_compile utility |
| 29 | def with_dyn_arg(fn: Callable, arg_index: int, dim_index: int): |
| 30 | def inner(*args): |
| 31 | torch._dynamo.mark_dynamic(args[arg_index], dim_index) |
| 32 | return fn(*args) |
| 33 | |
| 34 | return inner |
| 35 | |
| 36 | |
| 37 | def bench_compile(fn: Callable): |
| 38 | # recompile for different shapes |
| 39 | fwd = torch.compile(fn, fullgraph=True, dynamic=False) |
| 40 | |
| 41 | # First dim is explicitly dynamic to simulate vLLM usage |
| 42 | return with_dyn_arg(fwd, 0, 0) |
| 43 | |
| 44 | |
| 45 | torch._dynamo.config.recompile_limit = 8888 |
| 46 | |
| 47 | |
| 48 | def calculate_diff( |
| 49 | batch_size: int, |
| 50 | hidden_size: int, |
| 51 | group_shape: GroupShape, |
| 52 | dtype: torch.dtype, |
| 53 | ): |
| 54 | """Calculate the difference between Inductor and CUDA implementations.""" |
| 55 | device = torch.device("cuda") |
| 56 | x = torch.randn((batch_size, hidden_size), dtype=dtype, device=device) |
| 57 | |
| 58 | quant_fp8 = QuantFP8(False, group_shape, column_major_scales=False) |
| 59 | |
| 60 | torch_out, torch_scale = bench_compile(quant_fp8.forward_native)(x) |
| 61 | torch_eager_out, torch_eager_scale = quant_fp8.forward_native(x) |
| 62 | cuda_out, cuda_scale = quant_fp8.forward_cuda(x) |
| 63 | |
| 64 | try: |
| 65 | torch.testing.assert_close( |
| 66 | cuda_out.to(torch.float32), |
| 67 | torch_out.to(torch.float32), |
| 68 | rtol=1e-3, |
| 69 | atol=1e-5, |
| 70 | ) |
| 71 | torch.testing.assert_close(cuda_scale, torch_scale, rtol=1e-3, atol=1e-5) |
| 72 | torch.testing.assert_close( |
| 73 | cuda_out.to(torch.float32), |
| 74 | torch_eager_out.to(torch.float32), |
| 75 | rtol=1e-3, |
| 76 | atol=1e-5, |
| 77 | ) |
| 78 | torch.testing.assert_close(cuda_scale, torch_eager_scale, rtol=1e-3, atol=1e-5) |
| 79 | print("✅ All implementations match") |
| 80 | except AssertionError as e: |
| 81 | print("❌ Implementations differ") |
| 82 | print(e) |
| 83 | |
| 84 | |
| 85 | configs = [] |
| 86 | |
| 87 | |
| 88 | @default_vllm_config() |
| 89 | def benchmark_quantization( |
| 90 | batch_size, |
| 91 | hidden_size, |
| 92 | provider, |
| 93 | group_shape: GroupShape, |
| 94 | col_major: bool, |
| 95 | dtype: torch.dtype, |
| 96 | ): |
| 97 | device = torch.device("cuda") |
| 98 | |
| 99 | x = torch.randn(batch_size, hidden_size, device=device, dtype=dtype) |
| 100 | |
| 101 | quantiles = [0.5, 0.2, 0.8] |
| 102 | quant_fp8 = QuantFP8(False, group_shape, column_major_scales=col_major) |
| 103 | |
| 104 | if provider == "torch": |
| 105 | fn = lambda: bench_compile(quant_fp8.forward_native)(x.clone()) |
| 106 | elif provider == "cuda": |
| 107 | fn = lambda: quant_fp8.forward_cuda(x.clone()) |
| 108 | elif provider == "triton": |
| 109 | if not group_shape.is_per_group(): |
| 110 | # Triton only supported for per-group |
| 111 | return 0, 0, 0 |
| 112 | |
| 113 | fn = lambda: with_triton_mode(quant_fp8.forward_cuda)(x.clone()) |
| 114 | |
| 115 | ms, min_ms, max_ms = triton.testing.do_bench_cudagraph(fn, quantiles=quantiles) |
| 116 | |
| 117 | return 1000 * ms, 1000 * max_ms, 1000 * min_ms |
| 118 | |
| 119 | |
| 120 | # TODO(luka) extract to utils |
| 121 | def compute_geomean_speedups( |
| 122 | df: pd.DataFrame, |
| 123 | baseline_col: str, |
| 124 | speedup_cols: list[str], |
| 125 | groupby_cols: list[str] | None = None, |
| 126 | ) -> pd.DataFrame: |
| 127 | """Compute geometric mean speedups over a baseline column. |
| 128 | |
| 129 | Args: |
| 130 | df: Input dataframe |
| 131 | baseline_col: Column to use as baseline |
| 132 | speedup_cols: Columns to compute speedups for |
| 133 | groupby_cols: Columns to group by. If None, compute over entire df. |
| 134 | |
| 135 | Returns: |
| 136 | pd.DataFrame with geometric mean speedups |
| 137 | |
| 138 | """ |
| 139 | from scipy.stats import gmean |
| 140 | |
| 141 | def geo_speedup(group: pd.DataFrame) -> pd.Series: |
| 142 | ratios = { |
| 143 | col: (group[baseline_col] / group[col]).values for col in speedup_cols |
| 144 | } |
| 145 | return pd.Series({col: gmean(vals) for col, vals in ratios.items()}) |
| 146 | |
| 147 | if groupby_cols is None: |
| 148 | result = geo_speedup(df).to_frame().T |
| 149 | else: |
| 150 | result = ( |
| 151 | df.groupby(groupby_cols) |
| 152 | .apply(geo_speedup, include_groups=False) |
| 153 | .reset_index() |
| 154 | ) |
| 155 | |
| 156 | return result |
| 157 | |
| 158 | |
| 159 | if __name__ == "__main__": |
| 160 | parser = FlexibleArgumentParser( |
| 161 | description="Benchmark the various implementations of QuantFP8 (dynamic-only)" |
| 162 | ) |
| 163 | parser.add_argument("-c", "--check", action="store_true") |
| 164 | parser.add_argument( |
| 165 | "--dtype", type=str, choices=["half", "bfloat16", "float"], default="bfloat16" |
| 166 | ) |
| 167 | parser.add_argument( |
| 168 | "--hidden-sizes", |
| 169 | type=int, |
| 170 | nargs="+", |
| 171 | default=[896, 1024, 2048, 4096, 7168], |
| 172 | help="Hidden sizes to benchmark", |
| 173 | ) |
| 174 | parser.add_argument( |
| 175 | "--batch-sizes", |
| 176 | type=int, |
| 177 | nargs="+", |
| 178 | default=[1, 16, 128, 512, 1024], |
| 179 | help="Batch sizes to benchmark", |
| 180 | ) |
| 181 | parser.add_argument( |
| 182 | "--group-sizes", |
| 183 | type=int, |
| 184 | nargs="+", |
| 185 | default=None, |
| 186 | help="Group sizes for GroupShape(1,N) to benchmark. " |
| 187 | "Use 0 for PER_TENSOR, -1 for PER_TOKEN (default: 0,-1,64,128)", |
| 188 | ) |
| 189 | parser.add_argument( |
| 190 | "--no-column-major", |
| 191 | action="store_true", |
| 192 | help="Disable column-major scales testing", |
| 193 | ) |
| 194 | |
| 195 | args = parser.parse_args() |
| 196 | assert args |
| 197 | |
| 198 | dtype = STR_DTYPE_TO_TORCH_DTYPE[args.dtype] |
| 199 | |
| 200 | hidden_sizes = args.hidden_sizes |
| 201 | batch_sizes = args.batch_sizes |
| 202 | |
| 203 | if args.group_sizes is not None: |
| 204 | group_shapes = [] |
| 205 | for size in args.group_sizes: |
| 206 | if size == 0: |
| 207 | group_shapes.append(GroupShape.PER_TENSOR) |
| 208 | elif size == -1: |
| 209 | group_shapes.append(GroupShape.PER_TOKEN) |
| 210 | else: |
| 211 | group_shapes.append(GroupShape(1, size)) |
| 212 | else: |
| 213 | group_shapes = [ |
| 214 | GroupShape.PER_TENSOR, |
| 215 | GroupShape.PER_TOKEN, |
| 216 | GroupShape(1, 64), |
| 217 | GroupShape(1, 128), |
| 218 | ] |
| 219 | |
| 220 | column_major_scales = [False] if args.no_column_major else [True, False] |
| 221 | |
| 222 | config_gen = itertools.product( |
| 223 | group_shapes, |
| 224 | column_major_scales, |
| 225 | batch_sizes, |
| 226 | hidden_sizes, |
| 227 | ) |
| 228 | |
| 229 | # filter out column-major scales for non-group, reverse order |
| 230 | configs.extend(c[::-1] for c in config_gen if (c[0].is_per_group() or not c[1])) |
| 231 | |
| 232 | print(f"Running {len(configs)} configurations:") |
| 233 | print(f" Hidden sizes: {hidden_sizes}") |
| 234 | print(f" Batch sizes: {batch_sizes}") |
| 235 | print(f" Group shapes: {[str(g) for g in group_shapes]}") |
| 236 | print(f" Column major scales: {column_major_scales}") |
| 237 | print() |
| 238 | |
| 239 | if args.check: |
| 240 | for group_shape in group_shapes: |
| 241 | group_size = group_shape[1] |
| 242 | print(f"{group_size=}") |
| 243 | calculate_diff( |
| 244 | batch_size=4, hidden_size=4096, group_shape=group_shape, dtype=dtype |
| 245 | ) |
| 246 | |
| 247 | benchmark = triton.testing.perf_report( |
| 248 | triton.testing.Benchmark( |
| 249 | x_names=["hidden_size", "batch_size", "col_major", "group_shape"], |
| 250 | x_vals=configs, |
| 251 | line_arg="provider", |
| 252 | line_vals=["torch", "cuda", "triton"], |
| 253 | line_names=["Torch (Compiled)", "CUDA", "Triton"], |
| 254 | styles=[("blue", "-"), ("green", "-"), ("black", "-")], |
| 255 | ylabel="us", |
| 256 | plot_name="QuantFP8 performance", |
| 257 | args={}, |
| 258 | ) |
| 259 | )(benchmark_quantization) |
| 260 | |
| 261 | df = benchmark.run(print_data=True, dtype=dtype, return_df=True) |
| 262 | |
| 263 | # Print geomean speedups |
| 264 | geo_table_grouped = compute_geomean_speedups( |
| 265 | df, |
| 266 | baseline_col="Torch (Compiled)", |
| 267 | speedup_cols=["CUDA", "Triton"], |
| 268 | groupby_cols=["col_major", "group_shape"], |
| 269 | ) |
| 270 | |
| 271 | print("Speedup over Torch (Compiled)") |
| 272 | print(geo_table_grouped.to_string(index=False)) |