| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import pytest |
| 5 | import torch |
| 6 | import torch.nn.functional as F |
| 7 | |
| 8 | from vllm.platforms import current_platform |
| 9 | from vllm.third_party.flash_linear_attention.ops.layernorm_guard import ( |
| 10 | LayerNormFwdKernel, |
| 11 | calc_rows_per_block, |
| 12 | layer_norm_fwd, |
| 13 | layernorm_fn, |
| 14 | rms_norm_ref, |
| 15 | ) |
| 16 | from vllm.triton_utils import triton |
| 17 | from vllm.utils.torch_utils import set_random_seed |
| 18 | |
| 19 | DEVICE = "xpu" if current_platform.is_xpu() else "cuda" |
| 20 | |
| 21 | |
| 22 | def layer_norm_ref( |
| 23 | x, |
| 24 | weight, |
| 25 | bias, |
| 26 | z=None, |
| 27 | eps=1e-6, |
| 28 | group_size=None, |
| 29 | norm_before_gate=True, |
| 30 | is_rms_norm=False, |
| 31 | ): |
| 32 | """Reference implementation for both layer norm and RMS norm.""" |
| 33 | if is_rms_norm: |
| 34 | # Use the imported rms_norm_ref for RMS norm cases |
| 35 | return rms_norm_ref( |
| 36 | x, |
| 37 | weight, |
| 38 | bias, |
| 39 | z=z, |
| 40 | eps=eps, |
| 41 | group_size=group_size, |
| 42 | norm_before_gate=norm_before_gate, |
| 43 | upcast=True, |
| 44 | ) |
| 45 | |
| 46 | # Layer norm implementation |
| 47 | dtype = x.dtype |
| 48 | x = x.float() |
| 49 | weight = weight.float() |
| 50 | bias = bias.float() if bias is not None else None |
| 51 | z = z.float() if z is not None else None |
| 52 | |
| 53 | if z is not None and not norm_before_gate: |
| 54 | x = x * F.silu(z) |
| 55 | |
| 56 | if group_size is None: |
| 57 | # Layer norm: subtract mean |
| 58 | mean = x.mean(dim=-1, keepdim=True) |
| 59 | var = ((x - mean).square()).mean(dim=-1, keepdim=True) |
| 60 | rstd = 1 / torch.sqrt(var + eps) |
| 61 | out = (x - mean) * rstd * weight |
| 62 | if bias is not None: |
| 63 | out = out + bias |
| 64 | else: |
| 65 | # Group norm |
| 66 | from einops import rearrange |
| 67 | |
| 68 | x_group = rearrange(x, "... (g d) -> ... g d", d=group_size) |
| 69 | mean = x_group.mean(dim=-1, keepdim=True) |
| 70 | var = ((x_group - mean).square()).mean(dim=-1, keepdim=True) |
| 71 | rstd = 1 / torch.sqrt(var + eps) |
| 72 | x_group = (x_group - mean) * rstd |
| 73 | out = rearrange(x_group, "... g d -> ... (g d)") * weight |
| 74 | if bias is not None: |
| 75 | out = out + bias |
| 76 | |
| 77 | if z is not None and norm_before_gate: |
| 78 | out *= F.silu(z) |
| 79 | |
| 80 | return out.to(dtype) |
| 81 | |
| 82 | |
| 83 | DTYPES = [torch.float16, torch.bfloat16, torch.float32] |
| 84 | # Test various M sizes to ensure rows_per_block logic works correctly |
| 85 | NUM_TOKENS = [ |
| 86 | 1, |
| 87 | 7, |
| 88 | 16, |
| 89 | 63, |
| 90 | 128, |
| 91 | 256, |
| 92 | 512, |
| 93 | 1024, |
| 94 | 2048, |
| 95 | 4096, |
| 96 | 5789, |
| 97 | 8189, |
| 98 | 8191, |
| 99 | 16383, |
| 100 | 32767, |
| 101 | ] |
| 102 | HIDDEN_SIZES = [64, 128, 256, 1024] |
| 103 | GROUP_SIZES = [None, 64, 128] # None means full hidden size |
| 104 | NORM_BEFORE_GATE = [True, False] |
| 105 | IS_RMS_NORM = [True, False] |
| 106 | SEEDS = [0, 42] |
| 107 | |
| 108 | |
| 109 | @pytest.mark.parametrize("rows_per_token", [1, 2, 4, 8, 16]) |
| 110 | def test_layer_norm_fwd_warmup_keys_cover_qwen_gdn( |
| 111 | rows_per_token: int, |
| 112 | ) -> None: |
| 113 | device = torch.device(DEVICE) |
| 114 | group_size = 128 |
| 115 | max_num_tokens = 512 |
| 116 | kernel = LayerNormFwdKernel() |
| 117 | warmup_keys = set( |
| 118 | kernel.get_warmup_keys( |
| 119 | max_num_tokens=max_num_tokens, |
| 120 | rows_per_token=rows_per_token, |
| 121 | group_size=group_size, |
| 122 | x_dtype=torch.bfloat16, |
| 123 | weight_dtype=torch.bfloat16, |
| 124 | device=device, |
| 125 | norm_before_gate=True, |
| 126 | is_rms_norm=True, |
| 127 | activation="silu", |
| 128 | ) |
| 129 | ) |
| 130 | |
| 131 | for num_tokens in range(1, max_num_tokens + 1): |
| 132 | num_rows = num_tokens * rows_per_token |
| 133 | runtime_key = kernel.dispatch( |
| 134 | x_dtype=torch.bfloat16, |
| 135 | y_dtype=torch.bfloat16, |
| 136 | weight_dtype=torch.bfloat16, |
| 137 | bias_dtype=None, |
| 138 | z_dtype=torch.bfloat16, |
| 139 | mean_dtype=None, |
| 140 | rstd_dtype=torch.float32, |
| 141 | x_aligned=True, |
| 142 | y_aligned=True, |
| 143 | weight_aligned=True, |
| 144 | bias_aligned=True, |
| 145 | z_aligned=True, |
| 146 | mean_aligned=True, |
| 147 | rstd_aligned=True, |
| 148 | stride_x_row=group_size, |
| 149 | stride_y_row=group_size, |
| 150 | stride_z_row=group_size, |
| 151 | M=num_rows, |
| 152 | N=group_size, |
| 153 | BLOCK_N=triton.next_power_of_2(group_size), |
| 154 | ROWS_PER_BLOCK=calc_rows_per_block(num_rows, device), |
| 155 | norm_before_gate=True, |
| 156 | is_rms_norm=True, |
| 157 | activation="silu", |
| 158 | ) |
| 159 | assert runtime_key in warmup_keys |
| 160 | |
| 161 | |
| 162 | @pytest.mark.parametrize("num_tokens", NUM_TOKENS) |
| 163 | @pytest.mark.parametrize("hidden_size", HIDDEN_SIZES) |
| 164 | @pytest.mark.parametrize("dtype", DTYPES) |
| 165 | @pytest.mark.parametrize("seed", SEEDS) |
| 166 | @pytest.mark.parametrize("is_rms_norm", IS_RMS_NORM) |
| 167 | @torch.inference_mode() |
| 168 | def test_layer_norm_fwd_basic( |
| 169 | num_tokens: int, |
| 170 | hidden_size: int, |
| 171 | dtype: torch.dtype, |
| 172 | seed: int, |
| 173 | is_rms_norm: bool, |
| 174 | ) -> None: |
| 175 | """Test basic layer norm forward pass without z (gate) tensor.""" |
| 176 | set_random_seed(seed) |
| 177 | device = torch.device(DEVICE) |
| 178 | |
| 179 | # Create inputs |
| 180 | x = torch.randn(num_tokens, hidden_size, dtype=dtype, device=device) |
| 181 | weight = torch.randn(hidden_size, dtype=dtype, device=device) |
| 182 | bias = None if is_rms_norm else torch.randn(hidden_size, dtype=dtype, device=device) |
| 183 | eps = 1e-6 |
| 184 | |
| 185 | # Run the triton kernel |
| 186 | out, mean, rstd = layer_norm_fwd( |
| 187 | x, weight, bias, eps, z=None, is_rms_norm=is_rms_norm |
| 188 | ) |
| 189 | |
| 190 | # Run reference implementation |
| 191 | ref_out = layer_norm_ref(x, weight, bias, z=None, eps=eps, is_rms_norm=is_rms_norm) |
| 192 | |
| 193 | # Check outputs |
| 194 | assert out.shape == x.shape |
| 195 | assert out.dtype == x.dtype |
| 196 | torch.testing.assert_close(out, ref_out, atol=1e-2, rtol=1e-2) |
| 197 | |
| 198 | # Check mean and rstd shapes |
| 199 | if not is_rms_norm: |
| 200 | assert mean.shape == (num_tokens,) |
| 201 | assert rstd.shape == (num_tokens,) |
| 202 | |
| 203 | |
| 204 | @pytest.mark.parametrize("num_tokens", NUM_TOKENS) |
| 205 | @pytest.mark.parametrize("hidden_size", [128, 256, 1024]) |
| 206 | @pytest.mark.parametrize("dtype", DTYPES) |
| 207 | @pytest.mark.parametrize("norm_before_gate", NORM_BEFORE_GATE) |
| 208 | @pytest.mark.parametrize("is_rms_norm", IS_RMS_NORM) |
| 209 | @torch.inference_mode() |
| 210 | def test_layer_norm_fwd_with_gate( |
| 211 | num_tokens: int, |
| 212 | hidden_size: int, |
| 213 | dtype: torch.dtype, |
| 214 | norm_before_gate: bool, |
| 215 | is_rms_norm: bool, |
| 216 | ) -> None: |
| 217 | """Test layer norm forward pass with z (gate) tensor.""" |
| 218 | set_random_seed(42) |
| 219 | device = torch.device(DEVICE) |
| 220 | |
| 221 | # Create inputs |
| 222 | x = torch.randn(num_tokens, hidden_size, dtype=dtype, device=device) |
| 223 | z = torch.randn(num_tokens, hidden_size, dtype=dtype, device=device) |
| 224 | weight = torch.randn(hidden_size, dtype=dtype, device=device) |
| 225 | bias = None if is_rms_norm else torch.randn(hidden_size, dtype=dtype, device=device) |
| 226 | eps = 1e-6 |
| 227 | |
| 228 | # Run the triton kernel |
| 229 | out, mean, rstd = layer_norm_fwd( |
| 230 | x, |
| 231 | weight, |
| 232 | bias, |
| 233 | eps, |
| 234 | z=z, |
| 235 | norm_before_gate=norm_before_gate, |
| 236 | is_rms_norm=is_rms_norm, |
| 237 | ) |
| 238 | |
| 239 | # Run reference implementation |
| 240 | ref_out = layer_norm_ref( |
| 241 | x, |
| 242 | weight, |
| 243 | bias, |
| 244 | z=z, |
| 245 | eps=eps, |
| 246 | norm_before_gate=norm_before_gate, |
| 247 | is_rms_norm=is_rms_norm, |
| 248 | ) |
| 249 | |
| 250 | # Check outputs |
| 251 | assert out.shape == x.shape |
| 252 | assert out.dtype == x.dtype |
| 253 | torch.testing.assert_close(out, ref_out, atol=1e-2, rtol=1e-2) |
| 254 | |
| 255 | |
| 256 | @pytest.mark.parametrize("num_tokens", [128, 512]) |
| 257 | @pytest.mark.parametrize("hidden_size", [512, 1024]) |
| 258 | @pytest.mark.parametrize("group_size", [64, 128, 256]) |
| 259 | @pytest.mark.parametrize("dtype", [torch.float16, torch.bfloat16]) |
| 260 | @pytest.mark.parametrize("is_rms_norm", IS_RMS_NORM) |
| 261 | @torch.inference_mode() |
| 262 | def test_layer_norm_fwd_with_groups( |
| 263 | num_tokens: int, |
| 264 | hidden_size: int, |
| 265 | group_size: int, |
| 266 | dtype: torch.dtype, |
| 267 | is_rms_norm: bool, |
| 268 | ) -> None: |
| 269 | """Test layer norm forward pass with group normalization.""" |
| 270 | if hidden_size % group_size != 0: |
| 271 | pytest.skip( |
| 272 | f"hidden_size {hidden_size} not divisible by group_size {group_size}" |
| 273 | ) |
| 274 | |
| 275 | set_random_seed(42) |
| 276 | device = torch.device(DEVICE) |
| 277 | |
| 278 | # Create inputs |
| 279 | x = torch.randn(num_tokens, hidden_size, dtype=dtype, device=device) |
| 280 | weight = torch.randn(hidden_size, dtype=dtype, device=device) |
| 281 | bias = None if is_rms_norm else torch.randn(hidden_size, dtype=dtype, device=device) |
| 282 | eps = 1e-6 |
| 283 | |
| 284 | ngroups = hidden_size // group_size |
| 285 | |
| 286 | # Run the triton kernel |
| 287 | out, mean, rstd = layer_norm_fwd( |
| 288 | x, weight, bias, eps, z=None, group_size=group_size, is_rms_norm=is_rms_norm |
| 289 | ) |
| 290 | |
| 291 | # Run reference implementation |
| 292 | ref_out = layer_norm_ref( |
| 293 | x, weight, bias, z=None, eps=eps, group_size=group_size, is_rms_norm=is_rms_norm |
| 294 | ) |
| 295 | |
| 296 | # Check outputs |
| 297 | assert out.shape == x.shape |
| 298 | assert out.dtype == x.dtype |
| 299 | torch.testing.assert_close(out, ref_out, atol=1e-2, rtol=1e-2) |
| 300 | |
| 301 | # Check mean and rstd shapes for groups |
| 302 | if not is_rms_norm: |
| 303 | assert mean.shape == (ngroups * num_tokens,) |
| 304 | assert rstd.shape == (ngroups * num_tokens,) |
| 305 | |
| 306 | |
| 307 | @pytest.mark.parametrize("num_tokens", [7, 63, 128, 513, 1024, 2049]) |
| 308 | @pytest.mark.parametrize("dtype", [torch.float16, torch.bfloat16]) |
| 309 | @torch.inference_mode() |
| 310 | def test_layer_norm_rows_per_block( |
| 311 | num_tokens: int, |
| 312 | dtype: torch.dtype, |
| 313 | ) -> None: |
| 314 | """Test that rows_per_block logic works correctly for various M sizes.""" |
| 315 | set_random_seed(42) |
| 316 | device = torch.device(DEVICE) |
| 317 | hidden_size = 1024 |
| 318 | |
| 319 | # Create inputs |
| 320 | x = torch.randn(num_tokens, hidden_size, dtype=dtype, device=device) |
| 321 | weight = torch.randn(hidden_size, dtype=dtype, device=device) |
| 322 | bias = torch.randn(hidden_size, dtype=dtype, device=device) |
| 323 | eps = 1e-6 |
| 324 | |
| 325 | # Run the triton kernel |
| 326 | out, mean, rstd = layer_norm_fwd(x, weight, bias, eps, z=None, is_rms_norm=False) |
| 327 | |
| 328 | # Run reference implementation |
| 329 | ref_out = layer_norm_ref(x, weight, bias, z=None, eps=eps, is_rms_norm=False) |
| 330 | |
| 331 | # Check outputs |
| 332 | torch.testing.assert_close(out, ref_out, atol=1e-2, rtol=1e-2) |
| 333 | |
| 334 | |
| 335 | @pytest.mark.parametrize("dtype", [torch.bfloat16]) |
| 336 | @torch.inference_mode() |
| 337 | def test_strided_input(dtype: torch.dtype) -> None: |
| 338 | """Test that the kernel handles non-contiguous (strided) |
| 339 | inputs correctly.""" |
| 340 | set_random_seed(42) |
| 341 | device = torch.device(DEVICE) |
| 342 | num_tokens = 128 |
| 343 | hidden_size = 1024 |
| 344 | |
| 345 | # Create a larger tensor and take a strided slice |
| 346 | x_large = torch.randn(num_tokens, hidden_size * 2, dtype=dtype, device=device) |
| 347 | x = x_large[:, :hidden_size] |
| 348 | |
| 349 | # Make it contiguous for the kernel |
| 350 | x_contiguous = x.contiguous() |
| 351 | |
| 352 | weight = torch.randn(hidden_size, dtype=dtype, device=device) |
| 353 | bias = torch.randn(hidden_size, dtype=dtype, device=device) |
| 354 | eps = 1e-6 |
| 355 | |
| 356 | # Run the triton kernel with contiguous input |
| 357 | out, mean, rstd = layer_norm_fwd( |
| 358 | x_contiguous, weight, bias, eps, z=None, is_rms_norm=False |
| 359 | ) |
| 360 | |
| 361 | # Run reference implementation |
| 362 | ref_out = layer_norm_ref( |
| 363 | x_contiguous, weight, bias, z=None, eps=eps, is_rms_norm=False |
| 364 | ) |
| 365 | |
| 366 | # Check outputs |
| 367 | torch.testing.assert_close(out, ref_out, atol=1e-2, rtol=1e-2) |
| 368 | |
| 369 | |
| 370 | @pytest.mark.parametrize("num_tokens", [1, 128, 2048]) |
| 371 | @pytest.mark.parametrize("hidden_size", [768, 4096]) |
| 372 | @pytest.mark.parametrize("dtype", [torch.float16, torch.bfloat16]) |
| 373 | @torch.inference_mode() |
| 374 | def test_output_buffer_provided( |
| 375 | num_tokens: int, |
| 376 | hidden_size: int, |
| 377 | dtype: torch.dtype, |
| 378 | ) -> None: |
| 379 | """Test that the kernel works when an output buffer is provided.""" |
| 380 | set_random_seed(42) |
| 381 | device = torch.device(DEVICE) |
| 382 | |
| 383 | # Create inputs |
| 384 | x = torch.randn(num_tokens, hidden_size, dtype=dtype, device=device) |
| 385 | weight = torch.randn(hidden_size, dtype=dtype, device=device) |
| 386 | bias = torch.randn(hidden_size, dtype=dtype, device=device) |
| 387 | eps = 1e-6 |
| 388 | |
| 389 | # Pre-allocate output buffer |
| 390 | out_buffer = torch.empty_like(x) |
| 391 | |
| 392 | # Run the triton kernel with provided output |
| 393 | out, mean, rstd = layer_norm_fwd( |
| 394 | x, weight, bias, eps, z=None, out=out_buffer, is_rms_norm=False |
| 395 | ) |
| 396 | |
| 397 | # Check that the provided buffer was used |
| 398 | assert out.data_ptr() == out_buffer.data_ptr() |
| 399 | |
| 400 | # Run reference implementation |
| 401 | ref_out = layer_norm_ref(x, weight, bias, z=None, eps=eps, is_rms_norm=False) |
| 402 | |
| 403 | # Check outputs |
| 404 | torch.testing.assert_close(out, ref_out, atol=1e-2, rtol=1e-2) |
| 405 | |
| 406 | |
| 407 | @pytest.mark.parametrize( |
| 408 | "shape", |
| 409 | [ |
| 410 | (4, 16, 1024), # 3D tensor |
| 411 | (2, 8, 512, 256), # 4D tensor |
| 412 | ], |
| 413 | ) |
| 414 | @pytest.mark.parametrize("dtype", [torch.float16, torch.bfloat16]) |
| 415 | @torch.inference_mode() |
| 416 | def test_multidimensional_input( |
| 417 | shape: tuple, |
| 418 | dtype: torch.dtype, |
| 419 | ) -> None: |
| 420 | """Test that the autograd function handles multidimensional inputs.""" |
| 421 | set_random_seed(42) |
| 422 | device = torch.device(DEVICE) |
| 423 | hidden_size = shape[-1] |
| 424 | |
| 425 | # Create inputs |
| 426 | x = torch.randn(*shape, dtype=dtype, device=device) |
| 427 | weight = torch.randn(hidden_size, dtype=dtype, device=device) |
| 428 | bias = torch.randn(hidden_size, dtype=dtype, device=device) |
| 429 | eps = 1e-6 |
| 430 | |
| 431 | # Run through autograd function |
| 432 | out = layernorm_fn(x, weight, bias, z=None, eps=eps) |
| 433 | |
| 434 | # Run reference implementation |
| 435 | ref_out = layer_norm_ref(x, weight, bias, z=None, eps=eps, is_rms_norm=False) |
| 436 | |
| 437 | # Check outputs |
| 438 | assert out.shape == x.shape |
| 439 | torch.testing.assert_close(out, ref_out, atol=1e-2, rtol=1e-2) |
| 440 | |
| 441 | |
| 442 | @pytest.mark.parametrize("num_tokens", [1, 128, 1024]) |
| 443 | @pytest.mark.parametrize("hidden_size", [64, 256, 1024]) |
| 444 | @pytest.mark.parametrize("dtype", [torch.float16, torch.bfloat16, torch.float32]) |
| 445 | @pytest.mark.parametrize("has_gate", [True, False]) |
| 446 | @pytest.mark.parametrize("group_size", [None, 64]) |
| 447 | @pytest.mark.parametrize("norm_before_gate", [True, False]) |
| 448 | @torch.inference_mode() |
| 449 | def test_rmsnorm_gated_forward_native_dtype( |
| 450 | default_vllm_config, |
| 451 | num_tokens: int, |
| 452 | hidden_size: int, |
| 453 | dtype: torch.dtype, |
| 454 | has_gate: bool, |
| 455 | group_size: int | None, |
| 456 | norm_before_gate: bool, |
| 457 | ): |
| 458 | """Test that RMSNormGated.forward_native preserves input dtype.""" |
| 459 | if group_size is not None and hidden_size % group_size != 0: |
| 460 | pytest.skip( |
| 461 | f"hidden_size {hidden_size} not divisible by group_size {group_size}" |
| 462 | ) |
| 463 | |
| 464 | from vllm.model_executor.layers.layernorm import RMSNormGated |
| 465 | |
| 466 | device = torch.device(DEVICE) |
| 467 | set_random_seed(42) |
| 468 | |
| 469 | layer = RMSNormGated( |
| 470 | hidden_size, |
| 471 | eps=1e-5, |
| 472 | group_size=group_size, |
| 473 | norm_before_gate=norm_before_gate, |
| 474 | device=device, |
| 475 | dtype=dtype, |
| 476 | ) |
| 477 | |
| 478 | x = torch.randn(num_tokens, hidden_size, dtype=dtype, device=device) |
| 479 | z = ( |
| 480 | torch.randn(num_tokens, hidden_size, dtype=dtype, device=device) |
| 481 | if has_gate |
| 482 | else None |
| 483 | ) |
| 484 | |
| 485 | out = layer.forward_native(x, z) |
| 486 | |
| 487 | # Verify dtype preservation |
| 488 | assert out.dtype == dtype, f"Expected {dtype}, got {out.dtype}" |
| 489 | |
| 490 | # Verify numerical correctness against reference |
| 491 | ref_out = rms_norm_ref( |
| 492 | x, |
| 493 | layer.weight, |
| 494 | layer.bias, |
| 495 | z=z, |
| 496 | eps=1e-5, |
| 497 | group_size=group_size, |
| 498 | norm_before_gate=norm_before_gate, |
| 499 | upcast=True, |
| 500 | ) |
| 501 | torch.testing.assert_close(out, ref_out, atol=1e-2, rtol=1e-2) |
| 502 | |
| 503 | |
| 504 | @pytest.mark.parametrize("dtype", [torch.bfloat16]) |
| 505 | @torch.inference_mode() |
| 506 | def test_rmsnorm_gated_3d_matches_flattened_2d( |
| 507 | default_vllm_config, |
| 508 | dtype: torch.dtype, |
| 509 | ) -> None: |
| 510 | """Qwen GDN now norms (N, H, D) in place of a flatten to (N*H, D). |
| 511 | |
| 512 | Last-dim RMS + gate must match the flattened layout on both the native |
| 513 | and CUDA/HIP kernels, including a 3D z view that is not contiguous. |
| 514 | """ |
| 515 | from vllm.model_executor.layers.layernorm import RMSNormGated |
| 516 | |
| 517 | set_random_seed(42) |
| 518 | device = torch.device(DEVICE) |
| 519 | num_tokens, num_heads, head_dim = 5, 4, 64 |
| 520 | layer = RMSNormGated( |
| 521 | head_dim, |
| 522 | eps=1e-5, |
| 523 | group_size=None, |
| 524 | norm_before_gate=True, |
| 525 | device=device, |
| 526 | dtype=dtype, |
| 527 | ) |
| 528 | |
| 529 | packed = torch.randn(num_tokens, num_heads, 2, head_dim, dtype=dtype, device=device) |
| 530 | x3 = packed[:, :, 0, :] |
| 531 | z3 = packed[:, :, 1, :] |
| 532 | assert not x3.is_contiguous() |
| 533 | assert not z3.is_contiguous() |
| 534 | x2 = x3.reshape(-1, head_dim).contiguous() |
| 535 | z2 = z3.reshape(-1, head_dim).contiguous() |
| 536 | |
| 537 | native_3d = layer.forward_native(x3, z3) |
| 538 | native_2d = layer.forward_native(x2, z2).reshape(num_tokens, num_heads, head_dim) |
| 539 | torch.testing.assert_close(native_3d, native_2d, atol=0, rtol=0) |
| 540 | |
| 541 | cuda_3d = layer.forward_cuda(x3, z3) |
| 542 | cuda_2d = layer.forward_cuda(x2, z2).reshape(num_tokens, num_heads, head_dim) |
| 543 | torch.testing.assert_close(cuda_3d, cuda_2d, atol=1e-3, rtol=1e-2) |
| 544 | torch.testing.assert_close(cuda_3d, native_3d, atol=1e-2, rtol=1e-2) |
| 545 | assert cuda_3d.shape == (num_tokens, num_heads, head_dim) |
| 546 | |
| 547 | |
| 548 | if __name__ == "__main__": |
| 549 | # Run a quick smoke test |
| 550 | test_layer_norm_fwd_basic(128, 1024, torch.float16, 42, False) |
| 551 | test_layer_norm_fwd_with_gate(128, 1024, torch.float16, True, False) |
| 552 | test_layer_norm_rows_per_block(513, torch.float16) |
| 553 | print("All smoke tests passed!") |