File size: 8,651 Bytes
2fee457 | 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 | import torch
from kernels.benchmark import Benchmark
_original_allclose = torch.allclose
def _fp8_dtype() -> torch.dtype:
if torch.version.hip is not None and hasattr(torch, "float8_e4m3fnuz"):
return torch.float8_e4m3fnuz
return torch.float8_e4m3fn
def _fp8_max() -> float:
return 240.0 if torch.version.hip is not None else 448.0
def _is_fp8_dtype(dtype: torch.dtype) -> bool:
return dtype == torch.float8_e4m3fn or (
hasattr(torch, "float8_e4m3fnuz") and dtype == torch.float8_e4m3fnuz
)
def _flashrt_allclose(input, other, rtol=1e-05, atol=1e-08, equal_nan=False):
if _is_fp8_dtype(input.dtype) or _is_fp8_dtype(other.dtype):
return _original_allclose(
input.float(),
other.float(),
rtol=0,
atol=0,
equal_nan=equal_nan,
)
return _original_allclose(input, other, rtol=rtol, atol=atol, equal_nan=equal_nan)
torch.allclose = _flashrt_allclose
QUANT_SHAPES = [
("decode_m1", 1, 4096),
("decode_m2", 2, 4096),
("decode_m4", 4, 4096),
("decode_m8", 8, 4096),
("small_m16", 16, 4096),
("small_m32", 32, 4096),
("prefill_m64", 64, 4096),
("prefill_m128", 128, 4096),
("prefill_m256", 256, 4096),
("wide_n8192_m16", 16, 8192),
("wide_n8192_m128", 128, 8192),
("vla_n12288_m16", 16, 12288),
("vla_n12288_m64", 64, 12288),
("vla_n16384_m16", 16, 16384),
("vla_n16384_m64", 64, 16384),
]
class BiasGeluFp8QuantizeBenchmark(Benchmark):
seed = 1
def _setup_shape(self, m: int, n: int) -> None:
self.input = torch.randn((m, n), device=self.device, dtype=torch.bfloat16)
self.bias = torch.randn((n,), device=self.device, dtype=torch.bfloat16)
self.scale = torch.tensor([0.25], device=self.device, dtype=torch.float32)
self.out = torch.empty_like(self.input, dtype=_fp8_dtype())
def _reference(self) -> torch.Tensor:
y = self.input.float() + self.bias.float()
y = torch.nn.functional.gelu(y, approximate="tanh")
limit = _fp8_max()
y = torch.clamp(y / self.scale.float(), -limit, limit)
return y.to(_fp8_dtype())
def setup_decode_m1(self) -> None:
self._setup_shape(1, 4096)
def benchmark_decode_m1(self) -> None:
self.kernel.bias_gelu_quantize_fp8_static_bf16(
self.input, self.bias, self.scale, out=self.out
)
def verify_decode_m1(self) -> torch.Tensor:
return self._reference()
def setup_decode_m8(self) -> None:
self._setup_shape(8, 4096)
def benchmark_decode_m8(self) -> None:
self.kernel.bias_gelu_quantize_fp8_static_bf16(
self.input, self.bias, self.scale, out=self.out
)
def verify_decode_m8(self) -> torch.Tensor:
return self._reference()
def setup_small_m16(self) -> None:
self._setup_shape(16, 4096)
def benchmark_small_m16(self) -> None:
self.kernel.bias_gelu_quantize_fp8_static_bf16(
self.input, self.bias, self.scale, out=self.out
)
def verify_small_m16(self) -> torch.Tensor:
return self._reference()
def setup_prefill_m64(self) -> None:
self._setup_shape(64, 4096)
def benchmark_prefill_m64(self) -> None:
self.kernel.bias_gelu_quantize_fp8_static_bf16(
self.input, self.bias, self.scale, out=self.out
)
def verify_prefill_m64(self) -> torch.Tensor:
return self._reference()
def setup_prefill_m128(self) -> None:
self._setup_shape(128, 4096)
def benchmark_prefill_m128(self) -> None:
self.kernel.bias_gelu_quantize_fp8_static_bf16(
self.input, self.bias, self.scale, out=self.out
)
def verify_prefill_m128(self) -> torch.Tensor:
return self._reference()
def setup_wide_n8192_m16(self) -> None:
self._setup_shape(16, 8192)
def benchmark_wide_n8192_m16(self) -> None:
self.kernel.bias_gelu_quantize_fp8_static_bf16(
self.input, self.bias, self.scale, out=self.out
)
def verify_wide_n8192_m16(self) -> torch.Tensor:
return self._reference()
def setup_wide_n8192_m128(self) -> None:
self._setup_shape(128, 8192)
def benchmark_wide_n8192_m128(self) -> None:
self.kernel.bias_gelu_quantize_fp8_static_bf16(
self.input, self.bias, self.scale, out=self.out
)
def verify_wide_n8192_m128(self) -> torch.Tensor:
return self._reference()
def _register_bias_gelu_shapes() -> None:
for label, m, n in QUANT_SHAPES:
def setup(self, m=m, n=n) -> None:
self._setup_shape(m, n)
def benchmark(self) -> None:
self.kernel.bias_gelu_quantize_fp8_static_bf16(
self.input, self.bias, self.scale, out=self.out
)
def verify(self) -> torch.Tensor:
return self._reference()
setattr(BiasGeluFp8QuantizeBenchmark, f"setup_{label}", setup)
setattr(BiasGeluFp8QuantizeBenchmark, f"benchmark_{label}", benchmark)
setattr(BiasGeluFp8QuantizeBenchmark, f"verify_{label}", verify)
class GeluFp8QuantizeBenchmark(Benchmark):
seed = 3
def _setup_shape(self, m: int, n: int) -> None:
self.input = torch.randn((m, n), device=self.device, dtype=torch.bfloat16)
self.scale = torch.tensor([0.25], device=self.device, dtype=torch.float32)
self.out = torch.empty_like(self.input, dtype=_fp8_dtype())
def _reference(self) -> torch.Tensor:
y = torch.nn.functional.gelu(self.input.float(), approximate="tanh")
limit = _fp8_max()
y = torch.clamp(y / self.scale.float(), -limit, limit)
return y.to(_fp8_dtype())
def setup_decode_m1(self) -> None:
self._setup_shape(1, 4096)
def benchmark_decode_m1(self) -> None:
self.kernel.gelu_quantize_fp8_static_bf16(
self.input, self.scale, out=self.out
)
def verify_decode_m1(self) -> torch.Tensor:
return self._reference()
def setup_decode_m8(self) -> None:
self._setup_shape(8, 4096)
def benchmark_decode_m8(self) -> None:
self.kernel.gelu_quantize_fp8_static_bf16(
self.input, self.scale, out=self.out
)
def verify_decode_m8(self) -> torch.Tensor:
return self._reference()
def setup_small_m16(self) -> None:
self._setup_shape(16, 4096)
def benchmark_small_m16(self) -> None:
self.kernel.gelu_quantize_fp8_static_bf16(
self.input, self.scale, out=self.out
)
def verify_small_m16(self) -> torch.Tensor:
return self._reference()
def setup_prefill_m64(self) -> None:
self._setup_shape(64, 4096)
def benchmark_prefill_m64(self) -> None:
self.kernel.gelu_quantize_fp8_static_bf16(
self.input, self.scale, out=self.out
)
def verify_prefill_m64(self) -> torch.Tensor:
return self._reference()
def setup_prefill_m128(self) -> None:
self._setup_shape(128, 4096)
def benchmark_prefill_m128(self) -> None:
self.kernel.gelu_quantize_fp8_static_bf16(
self.input, self.scale, out=self.out
)
def verify_prefill_m128(self) -> torch.Tensor:
return self._reference()
def setup_wide_n8192_m16(self) -> None:
self._setup_shape(16, 8192)
def benchmark_wide_n8192_m16(self) -> None:
self.kernel.gelu_quantize_fp8_static_bf16(
self.input, self.scale, out=self.out
)
def verify_wide_n8192_m16(self) -> torch.Tensor:
return self._reference()
def setup_wide_n8192_m128(self) -> None:
self._setup_shape(128, 8192)
def benchmark_wide_n8192_m128(self) -> None:
self.kernel.gelu_quantize_fp8_static_bf16(
self.input, self.scale, out=self.out
)
def verify_wide_n8192_m128(self) -> torch.Tensor:
return self._reference()
def _register_gelu_shapes() -> None:
for label, m, n in QUANT_SHAPES:
def setup(self, m=m, n=n) -> None:
self._setup_shape(m, n)
def benchmark(self) -> None:
self.kernel.gelu_quantize_fp8_static_bf16(
self.input, self.scale, out=self.out
)
def verify(self) -> torch.Tensor:
return self._reference()
setattr(GeluFp8QuantizeBenchmark, f"setup_{label}", setup)
setattr(GeluFp8QuantizeBenchmark, f"benchmark_{label}", benchmark)
setattr(GeluFp8QuantizeBenchmark, f"verify_{label}", verify)
_register_bias_gelu_shapes()
_register_gelu_shapes()
|