| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import random |
| 5 | from unittest.mock import patch |
| 6 | |
| 7 | import pytest |
| 8 | import torch |
| 9 | |
| 10 | from tests.kernels.allclose_default import get_default_atol, get_default_rtol |
| 11 | from tests.kernels.utils import opcheck |
| 12 | from vllm.model_executor.layers.activation import ( |
| 13 | FastGELU, |
| 14 | FatreluAndMul, |
| 15 | GeluAndMul, |
| 16 | MulAndSilu, |
| 17 | NewGELU, |
| 18 | QuickGELU, |
| 19 | ReLUSquaredActivation, |
| 20 | SiluAndMul, |
| 21 | SiluAndMulWithClamp, |
| 22 | SwigluOAIAndMul, |
| 23 | SwigluStepAndMul, |
| 24 | swiglustep_and_mul_triton, |
| 25 | ) |
| 26 | from vllm.model_executor.layers.fused_moe.activation import ( |
| 27 | ApplyMoEActivationConfig, |
| 28 | MoEActivation, |
| 29 | apply_moe_activation, |
| 30 | ) |
| 31 | from vllm.model_executor.layers.fused_moe.utils import swiglu_limit_func |
| 32 | from vllm.platforms import current_platform |
| 33 | from vllm.utils.torch_utils import set_random_seed |
| 34 | |
| 35 | DTYPES = [torch.half, torch.bfloat16, torch.float] |
| 36 | NUM_TOKENS = [7, 83, 2048] # Arbitrary values for testing |
| 37 | D = [512, 13824] # Arbitrary values for testing |
| 38 | SEEDS = [0] |
| 39 | CUDA_DEVICES = [ |
| 40 | f"cuda:{i}" for i in range(1 if torch.accelerator.device_count() == 1 else 2) |
| 41 | ] |
| 42 | |
| 43 | |
| 44 | def test_masked_moe_activation_rejects_unsupported_activation() -> None: |
| 45 | input = torch.empty(1, 1, 2) |
| 46 | output = torch.empty(1, 1, 1) |
| 47 | valid_token_counts = torch.ones(1, dtype=torch.int32) |
| 48 | |
| 49 | with pytest.raises(NotImplementedError, match="relu2"): |
| 50 | apply_moe_activation( |
| 51 | MoEActivation.RELU2, |
| 52 | output, |
| 53 | input, |
| 54 | valid_token_counts=valid_token_counts, |
| 55 | ) |
| 56 | |
| 57 | |
| 58 | def test_moe_silu_clamp_uses_native_xpu_fallback( |
| 59 | default_vllm_config, monkeypatch |
| 60 | ) -> None: |
| 61 | monkeypatch.setattr(current_platform, "is_xpu", lambda: True) |
| 62 | clamp_limit = 3.0 |
| 63 | input = torch.tensor([[12.0, -12.0, 8.0, -8.0], [-2.0, 2.0, -4.0, 4.0]]) |
| 64 | output = torch.empty(2, 2) |
| 65 | |
| 66 | apply_moe_activation( |
| 67 | MoEActivation.SILU, |
| 68 | output, |
| 69 | input, |
| 70 | activation_config=ApplyMoEActivationConfig(clamp_limit=clamp_limit), |
| 71 | ) |
| 72 | |
| 73 | expected = SiluAndMulWithClamp(clamp_limit, compile_native=False).forward_native( |
| 74 | input |
| 75 | ) |
| 76 | torch.testing.assert_close(output, expected) |
| 77 | |
| 78 | |
| 79 | def _assert_masked_moe_activation( |
| 80 | activation: MoEActivation, |
| 81 | activation_config: ApplyMoEActivationConfig, |
| 82 | *, |
| 83 | dtype: torch.dtype, |
| 84 | mask_layout: str, |
| 85 | d: int, |
| 86 | max_num_tokens: int, |
| 87 | ) -> None: |
| 88 | device = CUDA_DEVICES[0] |
| 89 | num_experts = 4 |
| 90 | input_dim = 2 * d if activation.is_gated else d |
| 91 | if mask_layout == "flat": |
| 92 | input = torch.randn(max_num_tokens, input_dim, dtype=dtype, device=device) |
| 93 | valid_token_counts = torch.tensor( |
| 94 | [max_num_tokens // 2], dtype=torch.int32, device=device |
| 95 | ) |
| 96 | output = torch.full((max_num_tokens, d), 42.0, dtype=dtype, device=device) |
| 97 | else: |
| 98 | input = torch.randn( |
| 99 | num_experts, max_num_tokens, input_dim, dtype=dtype, device=device |
| 100 | ) |
| 101 | valid_token_counts = torch.tensor( |
| 102 | [0, 1, max_num_tokens // 2, max_num_tokens], |
| 103 | dtype=torch.int32, |
| 104 | device=device, |
| 105 | ) |
| 106 | output = torch.full( |
| 107 | (num_experts, max_num_tokens, d), 42.0, dtype=dtype, device=device |
| 108 | ) |
| 109 | |
| 110 | apply_moe_activation( |
| 111 | activation, |
| 112 | output, |
| 113 | input, |
| 114 | activation_config=activation_config, |
| 115 | valid_token_counts=valid_token_counts, |
| 116 | ) |
| 117 | |
| 118 | batched_input = input.view(-1, max_num_tokens, input_dim) |
| 119 | batched_output = output.view(-1, max_num_tokens, d) |
| 120 | for expert, num_tokens in enumerate(valid_token_counts.cpu().tolist()): |
| 121 | if num_tokens: |
| 122 | expected = torch.empty((num_tokens, d), dtype=dtype, device=device) |
| 123 | apply_moe_activation( |
| 124 | activation, |
| 125 | expected, |
| 126 | batched_input[expert, :num_tokens].clone(), |
| 127 | activation_config=activation_config, |
| 128 | ) |
| 129 | torch.testing.assert_close( |
| 130 | batched_output[expert, :num_tokens], |
| 131 | expected, |
| 132 | atol=get_default_atol(output), |
| 133 | rtol=get_default_rtol(output), |
| 134 | ) |
| 135 | assert torch.all(batched_output[expert, num_tokens:] == 42.0) |
| 136 | |
| 137 | |
| 138 | @pytest.mark.parametrize( |
| 139 | "activation", |
| 140 | [ |
| 141 | "silu_and_mul", |
| 142 | "mul_and_silu", |
| 143 | "gelu", |
| 144 | "gelu_tanh", |
| 145 | "fatrelu", |
| 146 | "swigluoai_and_mul", |
| 147 | "swiglustep_and_mul", |
| 148 | ], |
| 149 | ) |
| 150 | @pytest.mark.parametrize("num_tokens", NUM_TOKENS) |
| 151 | @pytest.mark.parametrize("d", D) |
| 152 | @pytest.mark.parametrize("dtype", DTYPES) |
| 153 | @pytest.mark.parametrize("seed", SEEDS) |
| 154 | @pytest.mark.parametrize("device", CUDA_DEVICES) |
| 155 | @torch.inference_mode() |
| 156 | def test_act_and_mul( |
| 157 | default_vllm_config, |
| 158 | activation: str, |
| 159 | num_tokens: int, |
| 160 | d: int, |
| 161 | dtype: torch.dtype, |
| 162 | seed: int, |
| 163 | device: str, |
| 164 | ) -> None: |
| 165 | set_random_seed(seed) |
| 166 | torch.set_default_device(device) |
| 167 | x = torch.randn(num_tokens, 2 * d, dtype=dtype) |
| 168 | if activation == "silu_and_mul": |
| 169 | layer = SiluAndMul(compile_native=False) |
| 170 | fn = torch.ops._C.silu_and_mul |
| 171 | if activation == "mul_and_silu": |
| 172 | layer = MulAndSilu() |
| 173 | fn = torch.ops._C.mul_and_silu |
| 174 | elif activation == "gelu": |
| 175 | layer = GeluAndMul(approximate="none") |
| 176 | fn = torch.ops._C.gelu_and_mul |
| 177 | elif activation == "gelu_tanh": |
| 178 | layer = GeluAndMul(approximate="tanh") |
| 179 | fn = torch.ops._C.gelu_tanh_and_mul |
| 180 | elif activation == "fatrelu": |
| 181 | threshold = random.uniform(0, 1) |
| 182 | layer = FatreluAndMul(threshold) |
| 183 | fn = torch.ops._C.fatrelu_and_mul |
| 184 | elif activation == "swigluoai_and_mul": |
| 185 | layer = SwigluOAIAndMul() |
| 186 | fn = torch.ops._C.swigluoai_and_mul |
| 187 | elif activation == "swiglustep_and_mul": |
| 188 | layer = SwigluStepAndMul() |
| 189 | fn = swiglustep_and_mul_triton |
| 190 | out = layer(x) |
| 191 | ref_out = layer.forward_native(x) |
| 192 | if activation in ["swigluoai_and_mul", "swiglustep_and_mul"]: |
| 193 | rtol = { |
| 194 | # For fp16, change the relative tolerance from 1e-3 to 2e-3 |
| 195 | torch.float16: 2e-3, |
| 196 | torch.bfloat16: 2e-2, |
| 197 | torch.float: 1.3e-6, |
| 198 | } |
| 199 | |
| 200 | def _get_rtol(output) -> float: |
| 201 | return rtol[output.dtype] |
| 202 | |
| 203 | torch.testing.assert_close( |
| 204 | out, ref_out, atol=get_default_atol(out), rtol=_get_rtol(out) |
| 205 | ) |
| 206 | else: |
| 207 | # The SiluAndMul, MulAndSilu, GELU and FatReLU implementations are |
| 208 | # equivalent to the native PyTorch implementations, so we can do exact |
| 209 | # comparison. |
| 210 | torch.testing.assert_close(out, ref_out, atol=0.0, rtol=0.0) |
| 211 | |
| 212 | d = x.shape[-1] // 2 |
| 213 | output_shape = x.shape[:-1] + (d,) |
| 214 | out = torch.empty(output_shape, dtype=x.dtype, device=x.device) |
| 215 | if activation == "fatrelu": |
| 216 | opcheck(fn, (out, x, threshold)) |
| 217 | elif activation == "swigluoai_and_mul": |
| 218 | opcheck(fn, (out, x, layer.alpha, layer.limit)) |
| 219 | elif activation != "swiglustep_and_mul": |
| 220 | opcheck(fn, (out, x)) |
| 221 | |
| 222 | |
| 223 | SWIGLU_LIMITS = [3.0, 7.0, 15.0] |
| 224 | |
| 225 | |
| 226 | @torch.inference_mode() |
| 227 | def test_swiglu_limit_func_without_routing_uses_output_buffer() -> None: |
| 228 | x = torch.randn(7, 1024, dtype=torch.bfloat16, device="cuda") |
| 229 | output = torch.empty(7, 512, dtype=x.dtype, device=x.device) |
| 230 | |
| 231 | swiglu_limit_func(output, x, swiglu_limit=7.0) |
| 232 | gate, up = x.chunk(2, dim=-1) |
| 233 | expected = torch.nn.functional.silu(gate.clamp(max=7.0)) * up.clamp( |
| 234 | min=-7.0, max=7.0 |
| 235 | ) |
| 236 | |
| 237 | torch.testing.assert_close(output, expected, atol=2e-2, rtol=2e-2) |
| 238 | |
| 239 | |
| 240 | @pytest.mark.parametrize( |
| 241 | ("alpha", "beta"), |
| 242 | [(1.0, 0.0), (1.702, 0.0), (1.0, 1.0)], |
| 243 | ) |
| 244 | @pytest.mark.parametrize("swiglu_limit", SWIGLU_LIMITS) |
| 245 | @pytest.mark.parametrize("num_tokens", NUM_TOKENS) |
| 246 | @pytest.mark.parametrize("d", D) |
| 247 | @pytest.mark.parametrize("dtype", DTYPES) |
| 248 | @pytest.mark.parametrize("seed", SEEDS) |
| 249 | @pytest.mark.parametrize("device", CUDA_DEVICES) |
| 250 | @torch.inference_mode() |
| 251 | def test_silu_and_mul_with_clamp( |
| 252 | default_vllm_config, |
| 253 | alpha: float, |
| 254 | beta: float, |
| 255 | swiglu_limit: float, |
| 256 | num_tokens: int, |
| 257 | d: int, |
| 258 | dtype: torch.dtype, |
| 259 | seed: int, |
| 260 | device: str, |
| 261 | ) -> None: |
| 262 | """SiluAndMulWithClamp: cuda kernel must match native reference.""" |
| 263 | set_random_seed(seed) |
| 264 | torch.set_default_device(device) |
| 265 | # Use large values to ensure clamping is exercised. |
| 266 | x = torch.randn(num_tokens, 2 * d, dtype=dtype) * swiglu_limit * 2 |
| 267 | |
| 268 | default_vllm_config.compilation_config.custom_ops = [ |
| 269 | "none", |
| 270 | "+silu_and_mul_with_clamp", |
| 271 | ] |
| 272 | layer = SiluAndMulWithClamp( |
| 273 | swiglu_limit, |
| 274 | alpha=alpha, |
| 275 | beta=beta, |
| 276 | compile_native=False, |
| 277 | ) |
| 278 | if current_platform.is_rocm(): |
| 279 | # forward_hip is always dispatched; the alpha/beta gate is checked |
| 280 | # inside it at call time rather than picked at construction time, so |
| 281 | # verify the actual routing by spying on the two candidate methods. |
| 282 | assert layer._forward_method == layer.forward_hip |
| 283 | with ( |
| 284 | patch.object(layer, "forward_cuda", wraps=layer.forward_cuda) as cuda_spy, |
| 285 | patch.object( |
| 286 | layer, "forward_native", wraps=layer.forward_native |
| 287 | ) as native_spy, |
| 288 | ): |
| 289 | out = layer(x) |
| 290 | if alpha == 1.0 and beta == 0.0: |
| 291 | cuda_spy.assert_called_once() |
| 292 | native_spy.assert_not_called() |
| 293 | else: |
| 294 | native_spy.assert_called_once() |
| 295 | cuda_spy.assert_not_called() |
| 296 | else: |
| 297 | assert layer._forward_method == layer.forward_cuda |
| 298 | out = layer(x) |
| 299 | |
| 300 | ref_out = layer.forward_native(x) |
| 301 | |
| 302 | rtol = { |
| 303 | torch.float16: 2e-3, |
| 304 | torch.bfloat16: 2e-2, |
| 305 | torch.float: 1.3e-6, |
| 306 | } |
| 307 | torch.testing.assert_close( |
| 308 | out, ref_out, atol=get_default_atol(out), rtol=rtol[out.dtype] |
| 309 | ) |
| 310 | |
| 311 | # Verify clamping is actually being applied: the clamped output should |
| 312 | # differ from the unclamped SiluAndMul output when inputs are large. |
| 313 | if alpha == 1.0 and beta == 0.0: |
| 314 | unclamped_out = SiluAndMul.forward_native(x) |
| 315 | assert not torch.equal(ref_out.float(), unclamped_out.float()), ( |
| 316 | "Input was not large enough to exercise the clamp; increase scale" |
| 317 | ) |
| 318 | |
| 319 | # Verify gate clamping semantics with a controlled scalar case. |
| 320 | # gate=large_val is clamped to limit first, then silu(limit) * 1.0. |
| 321 | x_gate = torch.tensor( |
| 322 | [[swiglu_limit * 20.0, 1.0]], dtype=torch.float32, device=device |
| 323 | ) |
| 324 | out_gate = SiluAndMulWithClamp(swiglu_limit, compile_native=False)(x_gate) |
| 325 | expected_gate = torch.nn.functional.silu( |
| 326 | torch.tensor(swiglu_limit, dtype=torch.float32) |
| 327 | ).item() |
| 328 | torch.testing.assert_close( |
| 329 | out_gate, |
| 330 | torch.tensor([[expected_gate]], dtype=torch.float32, device=device), |
| 331 | atol=1e-3, |
| 332 | rtol=1e-3, |
| 333 | ) |
| 334 | |
| 335 | # Verify up clamping semantics: up >> limit gets clamped to limit. |
| 336 | x_up = torch.tensor( |
| 337 | [[1.0, swiglu_limit * 20.0]], dtype=torch.float32, device=device |
| 338 | ) |
| 339 | out_up = SiluAndMulWithClamp(swiglu_limit, compile_native=False)(x_up) |
| 340 | silu_1 = torch.nn.functional.silu(torch.tensor(1.0)).item() |
| 341 | torch.testing.assert_close( |
| 342 | out_up, |
| 343 | torch.tensor([[silu_1 * swiglu_limit]], dtype=torch.float32, device=device), |
| 344 | atol=1e-3, |
| 345 | rtol=1e-3, |
| 346 | ) |
| 347 | |
| 348 | # opcheck |
| 349 | out_buf = torch.empty(x.shape[:-1] + (d,), dtype=dtype, device=device) |
| 350 | opcheck( |
| 351 | torch.ops._C.silu_and_mul_with_clamp, |
| 352 | (out_buf, x, swiglu_limit, layer.alpha, layer.beta), |
| 353 | ) |
| 354 | |
| 355 | |
| 356 | @pytest.mark.parametrize("linear_beta", [-1.0, 2.0]) |
| 357 | @pytest.mark.parametrize("dtype", [torch.half, torch.bfloat16]) |
| 358 | @torch.inference_mode() |
| 359 | def test_masked_situ_and_mul( |
| 360 | default_vllm_config, |
| 361 | linear_beta: float, |
| 362 | dtype: torch.dtype, |
| 363 | ) -> None: |
| 364 | """Masked SITU computes valid expert rows and preserves padded zeros.""" |
| 365 | device = CUDA_DEVICES[0] |
| 366 | num_experts, max_num_tokens, d = 4, 7, 512 |
| 367 | beta = 1.5 |
| 368 | input = torch.randn(num_experts, max_num_tokens, 2 * d, dtype=dtype, device=device) |
| 369 | expert_num_tokens = torch.tensor([0, 1, 4, 7], dtype=torch.int32, device=device) |
| 370 | output = torch.zeros(num_experts, max_num_tokens, d, dtype=dtype, device=device) |
| 371 | |
| 372 | torch.ops._C.masked_situ_and_mul( |
| 373 | output, input, expert_num_tokens, beta, linear_beta |
| 374 | ) |
| 375 | |
| 376 | gate, up = input.float().chunk(2, dim=-1) |
| 377 | expected = beta * torch.tanh(gate / beta) * torch.sigmoid(gate) |
| 378 | if linear_beta > 0: |
| 379 | up = linear_beta * torch.tanh(up / linear_beta) |
| 380 | expected = (expected * up).to(dtype) |
| 381 | for expert, num_tokens in enumerate(expert_num_tokens.cpu().tolist()): |
| 382 | torch.testing.assert_close( |
| 383 | output[expert, :num_tokens], |
| 384 | expected[expert, :num_tokens], |
| 385 | atol=get_default_atol(output), |
| 386 | rtol=get_default_rtol(output), |
| 387 | ) |
| 388 | assert torch.count_nonzero(output[expert, num_tokens:]) == 0 |
| 389 | |
| 390 | opcheck( |
| 391 | torch.ops._C.masked_situ_and_mul, |
| 392 | (output, input, expert_num_tokens, beta, linear_beta), |
| 393 | ) |
| 394 | |
| 395 | |
| 396 | MOE_ACTIVATION_CASES = [ |
| 397 | pytest.param(MoEActivation.SILU, ApplyMoEActivationConfig(), id="silu"), |
| 398 | pytest.param( |
| 399 | MoEActivation.SILU, ApplyMoEActivationConfig(clamp_limit=3.0), id="silu_clamp" |
| 400 | ), |
| 401 | pytest.param(MoEActivation.GELU, ApplyMoEActivationConfig(), id="gelu"), |
| 402 | pytest.param(MoEActivation.GELU_TANH, ApplyMoEActivationConfig(), id="gelu_tanh"), |
| 403 | pytest.param( |
| 404 | MoEActivation.SITU, |
| 405 | ApplyMoEActivationConfig( |
| 406 | activation_situ_beta=1.5, |
| 407 | activation_situ_linear_beta=2.0, |
| 408 | ), |
| 409 | id="situ", |
| 410 | ), |
| 411 | pytest.param(MoEActivation.SWIGLUOAI, ApplyMoEActivationConfig(), id="swigluoai"), |
| 412 | pytest.param( |
| 413 | MoEActivation.SWIGLUOAI_UNINTERLEAVE, |
| 414 | ApplyMoEActivationConfig(clamp_limit=3.0, alpha=1.3, beta=0.5), |
| 415 | id="swigluoai_uninterleave", |
| 416 | ), |
| 417 | pytest.param(MoEActivation.SWIGLUSTEP, ApplyMoEActivationConfig(), id="swiglustep"), |
| 418 | pytest.param( |
| 419 | MoEActivation.SILU_NO_MUL, ApplyMoEActivationConfig(), id="silu_no_mul" |
| 420 | ), |
| 421 | pytest.param( |
| 422 | MoEActivation.GELU_NO_MUL, ApplyMoEActivationConfig(), id="gelu_no_mul" |
| 423 | ), |
| 424 | pytest.param( |
| 425 | MoEActivation.GELU_TANH_NO_MUL, |
| 426 | ApplyMoEActivationConfig(), |
| 427 | id="gelu_tanh_no_mul", |
| 428 | ), |
| 429 | pytest.param( |
| 430 | MoEActivation.RELU2_NO_MUL, ApplyMoEActivationConfig(), id="relu2_no_mul" |
| 431 | ), |
| 432 | ] |
| 433 | |
| 434 | |
| 435 | @pytest.mark.parametrize(("activation", "activation_config"), MOE_ACTIVATION_CASES) |
| 436 | @torch.inference_mode() |
| 437 | def test_masked_moe_activation_dispatch( |
| 438 | default_vllm_config, |
| 439 | activation: MoEActivation, |
| 440 | activation_config: ApplyMoEActivationConfig, |
| 441 | ) -> None: |
| 442 | _assert_masked_moe_activation( |
| 443 | activation, |
| 444 | activation_config, |
| 445 | dtype=torch.bfloat16, |
| 446 | mask_layout="batched_experts", |
| 447 | d=513, |
| 448 | max_num_tokens=7, |
| 449 | ) |
| 450 | |
| 451 | |
| 452 | @pytest.mark.parametrize( |
| 453 | ("activation", "activation_config", "mask_layout"), |
| 454 | [ |
| 455 | pytest.param( |
| 456 | MoEActivation.SILU, |
| 457 | ApplyMoEActivationConfig(), |
| 458 | "flat", |
| 459 | id="flat", |
| 460 | ), |
| 461 | pytest.param( |
| 462 | MoEActivation.SITU, |
| 463 | ApplyMoEActivationConfig( |
| 464 | activation_situ_beta=1.5, |
| 465 | activation_situ_linear_beta=2.0, |
| 466 | ), |
| 467 | "batched_experts", |
| 468 | id="batched-experts", |
| 469 | ), |
| 470 | ], |
| 471 | ) |
| 472 | @torch.inference_mode() |
| 473 | def test_masked_moe_activation_grid_stride( |
| 474 | default_vllm_config, |
| 475 | activation: MoEActivation, |
| 476 | activation_config: ApplyMoEActivationConfig, |
| 477 | mask_layout: str, |
| 478 | ) -> None: |
| 479 | _assert_masked_moe_activation( |
| 480 | activation, |
| 481 | activation_config, |
| 482 | dtype=torch.half, |
| 483 | mask_layout=mask_layout, |
| 484 | d=513, |
| 485 | max_num_tokens=67, |
| 486 | ) |
| 487 | |
| 488 | |
| 489 | @torch.inference_mode() |
| 490 | def test_masked_moe_activation_opcheck(default_vllm_config) -> None: |
| 491 | device = CUDA_DEVICES[0] |
| 492 | input = torch.randn(2, 3, 64, dtype=torch.half, device=device) |
| 493 | output = torch.empty(2, 3, 32, dtype=torch.half, device=device) |
| 494 | valid_token_counts = torch.tensor([1, 3], dtype=torch.int32, device=device) |
| 495 | opcheck( |
| 496 | torch.ops._C.masked_moe_activation, |
| 497 | (output, input, valid_token_counts, "silu", 0.0, 1.0, 0.0, 1.0, -1.0), |
| 498 | ) |
| 499 | |
| 500 | |
| 501 | @pytest.mark.parametrize( |
| 502 | "activation", |
| 503 | [ |
| 504 | (FastGELU, torch.ops._C.gelu_fast), |
| 505 | (NewGELU, torch.ops._C.gelu_new), |
| 506 | (QuickGELU, torch.ops._C.gelu_quick), |
| 507 | (ReLUSquaredActivation, torch.ops._C.relu_squared), |
| 508 | ], |
| 509 | ) |
| 510 | @pytest.mark.parametrize("num_tokens", NUM_TOKENS) |
| 511 | @pytest.mark.parametrize("d", D) |
| 512 | @pytest.mark.parametrize("dtype", DTYPES) |
| 513 | @pytest.mark.parametrize("seed", SEEDS) |
| 514 | @pytest.mark.parametrize("device", CUDA_DEVICES) |
| 515 | @torch.inference_mode() |
| 516 | def test_activation( |
| 517 | default_vllm_config, |
| 518 | activation: type[torch.nn.Module], |
| 519 | num_tokens: int, |
| 520 | d: int, |
| 521 | dtype: torch.dtype, |
| 522 | seed: int, |
| 523 | device: str, |
| 524 | ) -> None: |
| 525 | set_random_seed(seed) |
| 526 | torch.set_default_device(device) |
| 527 | x = torch.randn(num_tokens, d, dtype=dtype) |
| 528 | layer = activation[0]() |
| 529 | fn = activation[1] |
| 530 | out = layer(x) |
| 531 | ref_out = layer.forward_native(x) |
| 532 | torch.testing.assert_close( |
| 533 | out, ref_out, atol=get_default_atol(out), rtol=get_default_rtol(out) |
| 534 | ) |
| 535 | |
| 536 | out = torch.empty_like(x) |
| 537 | opcheck(fn, (out, x)) |
| 538 | |
| 539 | |
| 540 | HUMMING_ACTIVATION_CASES = MOE_ACTIVATION_CASES + [ |
| 541 | pytest.param(MoEActivation.RELU2, ApplyMoEActivationConfig(), id="relu2"), |
| 542 | ] |
| 543 | HUMMING_ACTIVATION_CASES += [ |
| 544 | pytest.param( |
| 545 | MoEActivation.SITU, |
| 546 | ApplyMoEActivationConfig( |
| 547 | activation_situ_beta=1.5, |
| 548 | activation_situ_linear_beta=linear_beta, |
| 549 | ), |
| 550 | id=f"situ-linear-beta-{linear_beta}", |
| 551 | ) |
| 552 | for linear_beta in (None, 0.0, -1.0) |
| 553 | ] |
| 554 | |
| 555 | |
| 556 | @pytest.mark.skipif(not current_platform.is_cuda(), reason="Humming requires CUDA") |
| 557 | @pytest.mark.parametrize(("activation", "activation_config"), HUMMING_ACTIVATION_CASES) |
| 558 | @pytest.mark.parametrize("dtype", DTYPES) |
| 559 | @pytest.mark.parametrize(("num_tokens", "d"), [(1, 512), (7, 768), (83, 512)]) |
| 560 | @torch.inference_mode() |
| 561 | def test_humming_activation_matches_framework( |
| 562 | activation: MoEActivation, |
| 563 | activation_config: ApplyMoEActivationConfig, |
| 564 | dtype: torch.dtype, |
| 565 | num_tokens: int, |
| 566 | d: int, |
| 567 | ) -> None: |
| 568 | """Compare activation math/layouts without quantization or Hadamard error.""" |
| 569 | pytest.importorskip("humming") |
| 570 | from humming.ops import process_input |
| 571 | |
| 572 | from vllm.model_executor.layers.quantization.utils.humming.activation import ( |
| 573 | get_humming_activation, |
| 574 | ) |
| 575 | |
| 576 | set_random_seed(0) |
| 577 | width = d * 2 if activation.is_gated else d |
| 578 | x = 4 * torch.randn(num_tokens, width, dtype=dtype, device=CUDA_DEVICES[0]) |
| 579 | # Exercise zero, saturation, and both sides of the clamp limits (3 and 7). |
| 580 | edges = x.new_tensor([0, 0.001, 1, 2.99, 3, 3.01, 6.99, 7, 7.01, 8, 16]) |
| 581 | edges = torch.cat((-edges[1:].flip(0), edges)) |
| 582 | x[0] = edges[torch.arange(width, device=x.device) % edges.numel()] |
| 583 | |
| 584 | actual, _, _ = process_input( |
| 585 | x, |
| 586 | quant_mode="none", |
| 587 | hadamard_block_size=0, |
| 588 | **get_humming_activation(activation, activation_config), |
| 589 | ) |
| 590 | expected = torch.empty(num_tokens, d, dtype=dtype, device=x.device) |
| 591 | if activation == MoEActivation.RELU2: |
| 592 | # apply_moe_activation has only the non-gated ReLU2 variant. |
| 593 | gate, up = x.float().chunk(2, dim=-1) |
| 594 | activated_gate = torch.empty_like(gate) |
| 595 | apply_moe_activation(MoEActivation.RELU2_NO_MUL, activated_gate, gate.clone()) |
| 596 | expected.copy_(activated_gate * up) |
| 597 | else: |
| 598 | # The framework's non-gated ReLU2 path modifies its input in place. |
| 599 | apply_moe_activation( |
| 600 | activation, expected, x.clone(), activation_config=activation_config |
| 601 | ) |
| 602 | torch.testing.assert_close( |
| 603 | actual, |
| 604 | expected, |
| 605 | atol=get_default_atol(expected), |
| 606 | rtol=get_default_rtol(expected), |
| 607 | ) |