| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | import pytest |
| 4 | import torch |
| 5 | |
| 6 | # This registers op implementations |
| 7 | import vllm.kernels # noqa: F401 |
| 8 | from tests.ir.ir_test_utils import ( |
| 9 | COMMON_HIDDEN_SIZES, |
| 10 | NUM_TOKENS, |
| 11 | assert_close, |
| 12 | clone_args, |
| 13 | supported_providers, |
| 14 | ) |
| 15 | from tests.utils import set_random_seed |
| 16 | from vllm import ir |
| 17 | from vllm.platforms import current_platform |
| 18 | |
| 19 | pytestmark = pytest.mark.skip_global_cleanup |
| 20 | |
| 21 | rms_norm_native = ir.ops.rms_norm.impls["native"].impl_fn |
| 22 | |
| 23 | IS_GPGPU_DEVICE = current_platform.is_cuda_alike() or current_platform.is_xpu() |
| 24 | |
| 25 | |
| 26 | @pytest.mark.skipif( |
| 27 | not IS_GPGPU_DEVICE, |
| 28 | reason="Currently only kernels on CUDA, ROCm and XPU", |
| 29 | ) |
| 30 | def test_rms_norm_registration(): |
| 31 | expected = { |
| 32 | "native": True, |
| 33 | "vllm_c": IS_GPGPU_DEVICE, |
| 34 | "aiter": current_platform.is_rocm(), |
| 35 | "oink": current_platform.has_device_capability(100) |
| 36 | and hasattr(torch.ops, "oink") |
| 37 | and hasattr(torch.ops.oink, "rmsnorm"), |
| 38 | } |
| 39 | |
| 40 | actual = { |
| 41 | provider: impl.supported for provider, impl in ir.ops.rms_norm.impls.items() |
| 42 | } |
| 43 | |
| 44 | assert actual == expected |
| 45 | |
| 46 | |
| 47 | @pytest.mark.parametrize("dtype", [torch.float16, torch.bfloat16, torch.float32]) |
| 48 | @pytest.mark.parametrize("n_tokens", NUM_TOKENS) |
| 49 | @pytest.mark.parametrize("hidden_size", COMMON_HIDDEN_SIZES) |
| 50 | @pytest.mark.parametrize("epsilon", [1e-6, 1e-5]) |
| 51 | @pytest.mark.skipif( |
| 52 | not IS_GPGPU_DEVICE, |
| 53 | reason="Currently only kernels on CUDA, ROCm and XPU", |
| 54 | ) |
| 55 | class TestRMSNorm: |
| 56 | def test_native_semantics(self, dtype, n_tokens, hidden_size, epsilon): |
| 57 | set_random_seed(0) |
| 58 | x, weight, epsilon = ir.ops.rms_norm.generate_inputs( |
| 59 | num_tokens=4, |
| 60 | hidden_size=8, |
| 61 | dtype=dtype, |
| 62 | epsilon=epsilon, |
| 63 | device=current_platform.device_type, |
| 64 | ) |
| 65 | out = rms_norm_native(x, weight, epsilon=epsilon) |
| 66 | |
| 67 | # Check shape, dtype, device |
| 68 | assert out.shape == x.shape |
| 69 | assert out.dtype == x.dtype |
| 70 | assert out.device == x.device |
| 71 | |
| 72 | # Check the scaling property of rms norm. This holds only |
| 73 | # approximately: epsilon does not scale with x, so |
| 74 | # rms_norm(2x) = x / sqrt(mean(x^2) + epsilon/4), which differs from |
| 75 | # rms_norm(x) by the epsilon term. Use the op's declared tolerance |
| 76 | # rather than a tighter hard-coded one. |
| 77 | out2 = rms_norm_native(x * 2.0, weight, epsilon=epsilon) |
| 78 | assert_close(ir.ops.rms_norm, out2, out) |
| 79 | |
| 80 | # Mean square should be approximately 1 (ignoring epsilon and weight scaling) |
| 81 | combined_norm = out.float() / weight.float() |
| 82 | variance = combined_norm.pow(2).mean(dim=-1) |
| 83 | # After RMS normalization, variance should be close to 1 |
| 84 | torch.testing.assert_close( |
| 85 | variance, torch.ones_like(variance), rtol=1e-2, atol=1e-2 |
| 86 | ) |
| 87 | |
| 88 | # Check behavior with and without weight |
| 89 | weight1 = torch.ones_like(weight) |
| 90 | out3 = rms_norm_native(x, weight1, epsilon=epsilon) |
| 91 | out4 = rms_norm_native(x, None, epsilon=epsilon) |
| 92 | torch.testing.assert_close(out3, out4) |
| 93 | |
| 94 | @pytest.mark.parametrize("provider", supported_providers(ir.ops.rms_norm)) |
| 95 | def test_impls(self, dtype, n_tokens, hidden_size, epsilon, provider): |
| 96 | impl = ir.ops.rms_norm.impls[provider] |
| 97 | x, weight, eps = ir.ops.rms_norm.generate_inputs( |
| 98 | num_tokens=n_tokens, |
| 99 | hidden_size=hidden_size, |
| 100 | dtype=dtype, |
| 101 | epsilon=epsilon, |
| 102 | device=current_platform.device_type, |
| 103 | ) |
| 104 | args = (x, weight, eps) |
| 105 | |
| 106 | if not impl.supports_args(*args): |
| 107 | pytest.skip(f"{provider} does not support args") |
| 108 | |
| 109 | ref_output = rms_norm_native(*clone_args(args)) |
| 110 | output = impl.impl_fn(*clone_args(args)) |
| 111 | assert_close(ir.ops.rms_norm, output, ref_output) |
| 112 | |
| 113 | # check that dispatched call matches direct call |
| 114 | with ir.ops.rms_norm.set_priority([provider, "native"]): |
| 115 | out_dispatched = ir.ops.rms_norm(*args) |
| 116 | out_direct = impl.impl_fn(*args) |
| 117 | torch.testing.assert_close(out_dispatched, out_direct, rtol=0.0, atol=0.0) |
| 118 | |
| 119 | # none of these support variance_size override |
| 120 | assert not impl.supports_args(x, weight, eps, 4) |
| 121 | assert not impl.supports_args(x, weight, eps, variance_size=4) |
| 122 | |
| 123 | # test weight=None behavior |
| 124 | out_no_weight = impl.impl_fn(x, None, eps) |
| 125 | out_unit_weight = impl.impl_fn(x, torch.ones_like(weight), eps) |
| 126 | assert_close(ir.ops.rms_norm, out_no_weight, out_unit_weight) |
| 127 | |
| 128 | @pytest.mark.parametrize("provider", ["vllm_c", "aiter", "native"]) |
| 129 | def test_torch_opcheck(self, dtype, n_tokens, hidden_size, epsilon, provider): |
| 130 | if not ir.ops.rms_norm.impls[provider].supported: |
| 131 | pytest.skip(f"{provider} impl not supported on this platform") |
| 132 | |
| 133 | args = ir.ops.rms_norm.generate_inputs( |
| 134 | num_tokens=n_tokens, |
| 135 | hidden_size=hidden_size, |
| 136 | dtype=dtype, |
| 137 | epsilon=epsilon, |
| 138 | device=current_platform.device_type, |
| 139 | ) |
| 140 | |
| 141 | # When checking the torch op, we have to set priority and use dispatch |
| 142 | with ir.ops.rms_norm.set_priority([provider, "native"]): |
| 143 | torch.library.opcheck(torch.ops.vllm_ir.rms_norm, args) |
| 144 | |
| 145 | |
| 146 | @pytest.mark.skipif( |
| 147 | not current_platform.is_rocm(), |
| 148 | reason="aiter is only supported on ROCm", |
| 149 | ) |
| 150 | def test_aiter_rejects_unsupported_dtypes(): |
| 151 | impl = ir.ops.rms_norm.impls["aiter"] |
| 152 | for dtype in [torch.float32, torch.float64]: |
| 153 | args = ir.ops.rms_norm.generate_inputs( |
| 154 | num_tokens=8, |
| 155 | hidden_size=4096, |
| 156 | dtype=dtype, |
| 157 | epsilon=1e-5, |
| 158 | device=current_platform.device_type, |
| 159 | ) |
| 160 | assert not impl.supports_args(*args), f"aiter should reject dtype={dtype}" |
| 161 | |
| 162 | |
| 163 | @pytest.mark.skipif( |
| 164 | not current_platform.is_rocm(), |
| 165 | reason="ROCm vllm_c RMSNorm needs explicit ND input handling", |
| 166 | ) |
| 167 | def test_vllm_c_rms_norm_accepts_nd_input(): |
| 168 | impl = ir.ops.rms_norm.impls["vllm_c"] |
| 169 | if not impl.supported: |
| 170 | pytest.skip("vllm_c impl not supported on this platform") |
| 171 | |
| 172 | base = torch.randn( |
| 173 | 3, 8, 192, dtype=torch.float16, device=current_platform.device_type |
| 174 | ) |
| 175 | x = base.split(64, dim=-1)[0].view(3, 8, 4, 16) |
| 176 | assert not x.is_contiguous() |
| 177 | weight = torch.randn(16, dtype=torch.float16, device=current_platform.device_type) |
| 178 | epsilon = 1e-5 |
| 179 | |
| 180 | output = impl.impl_fn(x, weight, epsilon) |
| 181 | ref_output = rms_norm_native(x, weight, epsilon) |
| 182 | |
| 183 | assert output.shape == x.shape |
| 184 | assert_close(ir.ops.rms_norm, output, ref_output) |
| 185 | |
| 186 | |
| 187 | @pytest.mark.skipif( |
| 188 | not current_platform.is_rocm(), |
| 189 | reason="ROCm vllm_c RMSNorm needs a contiguous output for strided inputs", |
| 190 | ) |
| 191 | def test_vllm_c_rms_norm_accepts_transposed_input(): |
| 192 | impl = ir.ops.rms_norm.impls["vllm_c"] |
| 193 | if not impl.supported: |
| 194 | pytest.skip("vllm_c impl not supported on this platform") |
| 195 | |
| 196 | x = torch.randn( |
| 197 | 1, 320, 120, dtype=torch.float16, device=current_platform.device_type |
| 198 | ).transpose(1, 2) |
| 199 | assert x.reshape(-1, x.shape[-1]).stride(-1) != 1 |
| 200 | weight = torch.randn(320, dtype=torch.float16, device=current_platform.device_type) |
| 201 | epsilon = 1e-5 |
| 202 | |
| 203 | output = impl.impl_fn(x, weight, epsilon) |
| 204 | ref_output = rms_norm_native(x, weight, epsilon) |
| 205 | |
| 206 | assert output.shape == x.shape |
| 207 | assert_close(ir.ops.rms_norm, output, ref_output) |
| 208 | |
| 209 | |
| 210 | fused_add_rms_norm_native = ir.ops.fused_add_rms_norm.impls["native"].impl_fn |
| 211 | |
| 212 | |
| 213 | @pytest.mark.skipif( |
| 214 | not IS_GPGPU_DEVICE, |
| 215 | reason="Currently only kernels on CUDA, ROCm and XPU", |
| 216 | ) |
| 217 | def test_fused_add_rms_norm_registration(): |
| 218 | expected = { |
| 219 | "native": True, |
| 220 | "vllm_c": IS_GPGPU_DEVICE, |
| 221 | "aiter": current_platform.is_rocm(), |
| 222 | "oink": current_platform.has_device_capability(100) |
| 223 | and hasattr(torch.ops, "oink") |
| 224 | and hasattr(torch.ops.oink, "fused_add_rms_norm"), |
| 225 | } |
| 226 | |
| 227 | actual = { |
| 228 | provider: impl.supported |
| 229 | for provider, impl in ir.ops.fused_add_rms_norm.impls.items() |
| 230 | } |
| 231 | |
| 232 | assert actual == expected |
| 233 | |
| 234 | |
| 235 | @pytest.mark.skipif( |
| 236 | not current_platform.is_rocm(), |
| 237 | reason="ROCm vllm_c fused_add_rms_norm needs explicit ND input handling", |
| 238 | ) |
| 239 | def test_vllm_c_fused_add_rms_norm_accepts_nd_input(): |
| 240 | impl = ir.ops.fused_add_rms_norm.impls["vllm_c"] |
| 241 | if not impl.supported: |
| 242 | pytest.skip("vllm_c impl not supported on this platform") |
| 243 | |
| 244 | base = torch.randn( |
| 245 | 3, 8, 192, dtype=torch.float16, device=current_platform.device_type |
| 246 | ) |
| 247 | residual_base = torch.randn( |
| 248 | 3, 8, 192, dtype=torch.float16, device=current_platform.device_type |
| 249 | ) |
| 250 | x = base.split(64, dim=-1)[0].view(3, 8, 4, 16) |
| 251 | x_residual = residual_base.split(64, dim=-1)[0].view(3, 8, 4, 16) |
| 252 | assert not x.is_contiguous() |
| 253 | assert not x_residual.is_contiguous() |
| 254 | weight = torch.randn(16, dtype=torch.float16, device=current_platform.device_type) |
| 255 | epsilon = 1e-5 |
| 256 | |
| 257 | output, residual = impl.impl_fn(x.clone(), x_residual.clone(), weight, epsilon) |
| 258 | ref_output, ref_residual = fused_add_rms_norm_native(x, x_residual, weight, epsilon) |
| 259 | |
| 260 | assert output.shape == x.shape |
| 261 | assert residual.shape == x_residual.shape |
| 262 | assert_close(ir.ops.fused_add_rms_norm, output, ref_output) |
| 263 | assert_close(ir.ops.fused_add_rms_norm, residual, ref_residual) |
| 264 | |
| 265 | |
| 266 | @pytest.mark.parametrize("dtype", [torch.float16, torch.bfloat16, torch.float32]) |
| 267 | @pytest.mark.parametrize("n_tokens", NUM_TOKENS) |
| 268 | @pytest.mark.parametrize("hidden_size", COMMON_HIDDEN_SIZES) |
| 269 | @pytest.mark.parametrize("epsilon", [1e-6, 1e-5]) |
| 270 | @pytest.mark.skipif( |
| 271 | not IS_GPGPU_DEVICE, |
| 272 | reason="Currently only kernels on CUDA, ROCm and XPU", |
| 273 | ) |
| 274 | class TestFusedAddRMSNorm: |
| 275 | def test_native_semantics(self, dtype, n_tokens, hidden_size, epsilon): |
| 276 | set_random_seed(0) |
| 277 | x, x_residual, weight, eps = ir.ops.fused_add_rms_norm.generate_inputs( |
| 278 | num_tokens=4, |
| 279 | hidden_size=8, |
| 280 | dtype=dtype, |
| 281 | epsilon=epsilon, |
| 282 | device=current_platform.device_type, |
| 283 | ) |
| 284 | out, residual_out = fused_add_rms_norm_native(x, x_residual, weight, eps) |
| 285 | |
| 286 | # Check shape, dtype, device |
| 287 | assert out.shape == x.shape |
| 288 | assert out.dtype == x.dtype |
| 289 | assert out.device == x.device |
| 290 | assert residual_out.shape == x_residual.shape |
| 291 | assert residual_out.dtype == x_residual.dtype |
| 292 | assert residual_out.device == x_residual.device |
| 293 | |
| 294 | # Check that residual_out = x + x_residual |
| 295 | expected_residual = (x.float() + x_residual.float()).to(dtype) |
| 296 | torch.testing.assert_close( |
| 297 | residual_out, expected_residual, rtol=1e-3, atol=1e-3 |
| 298 | ) |
| 299 | |
| 300 | # Verify that the output is RMS normalized version of (x + x_residual) |
| 301 | expected_out = rms_norm_native(expected_residual, weight, epsilon) |
| 302 | assert_close( |
| 303 | ir.ops.fused_add_rms_norm, |
| 304 | (out, residual_out), |
| 305 | (expected_out, expected_residual), |
| 306 | ) |
| 307 | |
| 308 | # Check the scaling property of rms norm |
| 309 | out1, _ = fused_add_rms_norm_native( |
| 310 | x, torch.zeros_like(x), weight, epsilon=epsilon |
| 311 | ) |
| 312 | out2, _ = fused_add_rms_norm_native( |
| 313 | x * 2.0, torch.zeros_like(x), weight, epsilon=epsilon |
| 314 | ) |
| 315 | assert_close(ir.ops.fused_add_rms_norm, out2, out1) |
| 316 | |
| 317 | # Check behavior with and without weight |
| 318 | weight1 = torch.ones_like(weight) |
| 319 | out3, _ = fused_add_rms_norm_native(x, x_residual, weight1, eps) |
| 320 | out4, _ = fused_add_rms_norm_native(x, x_residual, None, eps) |
| 321 | torch.testing.assert_close(out3, out4) |
| 322 | |
| 323 | @pytest.mark.parametrize("provider", supported_providers(ir.ops.fused_add_rms_norm)) |
| 324 | def test_impls(self, dtype, n_tokens, hidden_size, epsilon, provider): |
| 325 | impl = ir.ops.fused_add_rms_norm.impls[provider] |
| 326 | x, x_residual, weight, eps = ir.ops.fused_add_rms_norm.generate_inputs( |
| 327 | num_tokens=n_tokens, |
| 328 | hidden_size=hidden_size, |
| 329 | dtype=dtype, |
| 330 | epsilon=epsilon, |
| 331 | device=current_platform.device_type, |
| 332 | ) |
| 333 | args = (x, x_residual, weight, eps, None) |
| 334 | |
| 335 | if not impl.supports_args(*args): |
| 336 | pytest.skip(f"{provider} does not support args") |
| 337 | |
| 338 | ref_output, ref_residual = fused_add_rms_norm_native(*clone_args(args)) |
| 339 | output, residual = impl.impl_fn(*clone_args(args)) |
| 340 | assert_close(ir.ops.fused_add_rms_norm, output, ref_output) |
| 341 | assert_close(ir.ops.fused_add_rms_norm, residual, ref_residual) |
| 342 | |
| 343 | # check that dispatched call matches direct call |
| 344 | with ir.ops.fused_add_rms_norm.set_priority([provider, "native"]): |
| 345 | out_dispatched, residual_dispatched = ir.ops.fused_add_rms_norm(*args[:4]) |
| 346 | out_direct, residual_direct = impl.impl_fn(*clone_args(args)) |
| 347 | torch.testing.assert_close(out_dispatched, out_direct, rtol=0.0, atol=0.0) |
| 348 | torch.testing.assert_close( |
| 349 | residual_dispatched, residual_direct, rtol=0.0, atol=0.0 |
| 350 | ) |
| 351 | |
| 352 | # none of these support variance_size override |
| 353 | assert not impl.supports_args(x, x_residual, weight, epsilon, 4) |
| 354 | assert not impl.supports_args(x, x_residual, weight, epsilon, variance_size=4) |
| 355 | |
| 356 | # test weight=None behavior |
| 357 | out_no_weight, residual_no_weight = impl.impl_fn( |
| 358 | x.clone(), x_residual.clone(), None, epsilon |
| 359 | ) |
| 360 | out_unit_weight, residual_unit_weight = impl.impl_fn( |
| 361 | x.clone(), x_residual.clone(), torch.ones_like(weight), epsilon |
| 362 | ) |
| 363 | assert_close(ir.ops.fused_add_rms_norm, out_no_weight, out_unit_weight) |
| 364 | assert_close( |
| 365 | ir.ops.fused_add_rms_norm, residual_no_weight, residual_unit_weight |
| 366 | ) |
| 367 | |
| 368 | @pytest.mark.parametrize("provider", ["vllm_c"]) |
| 369 | def test_inplace_semantics(self, dtype, n_tokens, hidden_size, epsilon, provider): |
| 370 | """Test that inplace implementations reuse inputs, |
| 371 | for maybe_inplace overload but not for default overload.""" |
| 372 | impl = ir.ops.fused_add_rms_norm.impls[provider] |
| 373 | if not impl.supported: |
| 374 | pytest.skip(f"{provider} impl not supported on this platform") |
| 375 | |
| 376 | x, x_residual, weight, eps = ir.ops.fused_add_rms_norm.generate_inputs( |
| 377 | num_tokens=n_tokens, |
| 378 | hidden_size=hidden_size, |
| 379 | dtype=dtype, |
| 380 | epsilon=epsilon, |
| 381 | device=current_platform.device_type, |
| 382 | ) |
| 383 | |
| 384 | # Test default overload - should NOT modify inputs even with inplace impl |
| 385 | x_default = x.clone() |
| 386 | x_residual_default = x_residual.clone() |
| 387 | x_default_ptr = x_default.data_ptr() |
| 388 | x_residual_default_ptr = x_residual_default.data_ptr() |
| 389 | |
| 390 | with ir.ops.fused_add_rms_norm.set_priority([provider, "native"]): |
| 391 | out_default, residual_default = ir.ops.fused_add_rms_norm( |
| 392 | x_default, x_residual_default, weight, eps |
| 393 | ) |
| 394 | |
| 395 | # Default should NOT be inplace (even with inplace implementation) |
| 396 | assert out_default.data_ptr() != x_default_ptr |
| 397 | assert residual_default.data_ptr() != x_residual_default_ptr |
| 398 | torch.testing.assert_close(x, x_default, rtol=0.0, atol=0.0) |
| 399 | torch.testing.assert_close(x_residual, x_residual_default, rtol=0.0, atol=0.0) |
| 400 | |
| 401 | # Test maybe_inplace overload - should modify inputs with inplace impl |
| 402 | x_inplace = x.clone() |
| 403 | x_residual_inplace = x_residual.clone() |
| 404 | x_inplace_ptr = x_inplace.data_ptr() |
| 405 | x_residual_inplace_ptr = x_residual_inplace.data_ptr() |
| 406 | |
| 407 | with ir.ops.fused_add_rms_norm.set_priority([provider, "native"]): |
| 408 | out_inplace, residual_inplace = ir.ops.fused_add_rms_norm.maybe_inplace( |
| 409 | x_inplace, x_residual_inplace, weight, eps |
| 410 | ) |
| 411 | |
| 412 | # maybe_inplace should be inplace |
| 413 | assert out_inplace.data_ptr() == x_inplace_ptr |
| 414 | assert residual_inplace.data_ptr() == x_residual_inplace_ptr |
| 415 | |
| 416 | # Both should produce same results |
| 417 | torch.testing.assert_close(out_default, out_inplace, atol=0.0, rtol=0.0) |
| 418 | torch.testing.assert_close( |
| 419 | residual_default, residual_inplace, atol=0.0, rtol=0.0 |
| 420 | ) |
| 421 | |
| 422 | @pytest.mark.parametrize("provider", supported_providers(ir.ops.fused_add_rms_norm)) |
| 423 | def test_torch_opcheck(self, dtype, n_tokens, hidden_size, epsilon, provider): |
| 424 | args = ir.ops.fused_add_rms_norm.generate_inputs( |
| 425 | num_tokens=n_tokens, |
| 426 | hidden_size=hidden_size, |
| 427 | dtype=dtype, |
| 428 | epsilon=epsilon, |
| 429 | device=current_platform.device_type, |
| 430 | ) |
| 431 | args = args + (None,) # Add variance_size parameter |
| 432 | |
| 433 | # When checking the torch op, we have to set priority and use dispatch |
| 434 | with ir.ops.fused_add_rms_norm.set_priority([provider, "native"]): |
| 435 | torch.library.opcheck(torch.ops.vllm_ir.fused_add_rms_norm.default, args) |
| 436 | |
| 437 | # Only test maybe_inplace with non-inplace implementations |
| 438 | # Inplace implementations return aliases of inputs which is not allowed. |
| 439 | # We break this invariant, but we also convert maybe_inplace to the default |
| 440 | # overload during compilation, so maybe_inplace never reaches Inductor. |
| 441 | if not ir.ops.fused_add_rms_norm.impls[provider].inplace: |
| 442 | torch.library.opcheck( |
| 443 | torch.ops.vllm_ir.fused_add_rms_norm.maybe_inplace, args |
| 444 | ) |