| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | from threading import Lock |
| 4 | |
| 5 | import pytest |
| 6 | import torch |
| 7 | |
| 8 | import vllm.lora.ops.triton_ops as triton_ops |
| 9 | from vllm.lora.ops.triton_ops import LoRAKernelMeta |
| 10 | from vllm.lora.ops.triton_ops.utils import _LORA_A_PTR_DICT, _LORA_B_PTR_DICT |
| 11 | from vllm.platforms import current_platform |
| 12 | from vllm.utils.torch_utils import set_random_seed |
| 13 | |
| 14 | from .utils import PunicaTensors, assert_close, generate_data_for_nslices |
| 15 | |
| 16 | DEVICE_TYPE = current_platform.device_type |
| 17 | |
| 18 | # On XPU, oneDNN/oneMKL can return wrong results for these reference matmuls |
| 19 | # after many Triton kernel launches, so the reference stays on CPU there. |
| 20 | _REF_ON_CPU = current_platform.is_xpu() |
| 21 | |
| 22 | |
| 23 | def _to_ref_device(tensor: torch.Tensor) -> torch.Tensor: |
| 24 | return tensor.cpu() if _REF_ON_CPU else tensor |
| 25 | |
| 26 | |
| 27 | @pytest.fixture(autouse=True) |
| 28 | def reset_device(reset_default_device): |
| 29 | pass |
| 30 | |
| 31 | |
| 32 | @pytest.fixture(autouse=True) |
| 33 | def cleanup_fixture(): |
| 34 | """Override conftest's cleanup_fixture— not needed for punica tests.""" |
| 35 | yield |
| 36 | |
| 37 | |
| 38 | @pytest.fixture(autouse=True) |
| 39 | def dynamo_reset(): |
| 40 | """Override conftest's dynamo_reset — not needed for punica tests.""" |
| 41 | yield |
| 42 | |
| 43 | |
| 44 | def _bgmv_shrink( |
| 45 | inputs, lora_weight, output, seq_len_tensor, lora_indices, scaling=1.0 |
| 46 | ): |
| 47 | """Memory-efficient shrink reference: per-LoRA matmul loop. |
| 48 | output[mask] = scaling * inputs[mask] @ weight.T""" |
| 49 | exploded = torch.repeat_interleave(lora_indices, seq_len_tensor) |
| 50 | for lid in exploded.unique(): |
| 51 | if lid < 0: |
| 52 | continue |
| 53 | mask = exploded == lid |
| 54 | inp = inputs[mask].to(output.dtype) |
| 55 | w = lora_weight[lid].to(output.dtype) |
| 56 | output[mask] = scaling * (inp @ w.T) |
| 57 | |
| 58 | |
| 59 | def _bgmv_expand( |
| 60 | inputs, |
| 61 | lora_weight, |
| 62 | output, |
| 63 | seq_len_tensor, |
| 64 | lora_indices, |
| 65 | offset=0, |
| 66 | add_inputs=False, |
| 67 | ): |
| 68 | """Memory-efficient expand reference: per-LoRA matmul loop. |
| 69 | output[mask, offset:offset+n] (+)= inputs[mask] @ weight.T""" |
| 70 | exploded = torch.repeat_interleave(lora_indices, seq_len_tensor) |
| 71 | for lid in exploded.unique(): |
| 72 | if lid < 0: |
| 73 | continue |
| 74 | mask = exploded == lid |
| 75 | inp = inputs[mask].to(output.dtype) |
| 76 | w = lora_weight[lid].to(output.dtype) |
| 77 | n = w.shape[0] |
| 78 | result = inp @ w.T |
| 79 | if add_inputs: |
| 80 | output[mask, offset : offset + n] += result |
| 81 | else: |
| 82 | output[mask, offset : offset + n] = result |
| 83 | |
| 84 | |
| 85 | # Utility shrink and expand operations used as reference implementations. |
| 86 | def sgmv_shrink_for_nslices( |
| 87 | nslices: int, |
| 88 | inputs_tensor: torch.Tensor, |
| 89 | lora_weights_lst: list[torch.Tensor], |
| 90 | out_tensor: torch.Tensor, |
| 91 | b_seq_start_loc: torch.Tensor, |
| 92 | seq_len_tensor: torch.Tensor, |
| 93 | prompt_lora_mapping: torch.Tensor, |
| 94 | batches: int, |
| 95 | max_seq_length: int, |
| 96 | num_tokens: int, |
| 97 | scaling: float, |
| 98 | ): |
| 99 | """Reference for sgmv_shrink using per-LoRA matmul loop.""" |
| 100 | inputs = _to_ref_device(inputs_tensor) |
| 101 | seq_len = _to_ref_device(seq_len_tensor) |
| 102 | mapping = _to_ref_device(prompt_lora_mapping) |
| 103 | out = _to_ref_device(out_tensor) |
| 104 | for index in range(nslices): |
| 105 | _bgmv_shrink( |
| 106 | inputs, |
| 107 | _to_ref_device(lora_weights_lst[index]), |
| 108 | out[index], |
| 109 | seq_len, |
| 110 | mapping, |
| 111 | scaling=scaling, |
| 112 | ) |
| 113 | if _REF_ON_CPU: |
| 114 | out_tensor.copy_(out) |
| 115 | |
| 116 | |
| 117 | def sgmv_expand_for_nslices( |
| 118 | nslices: int, |
| 119 | hidden_size: int, |
| 120 | inputs_tensor: torch.Tensor, |
| 121 | lora_weights_lst: list[torch.Tensor], |
| 122 | out_tensor: torch.Tensor, |
| 123 | b_seq_start_loc: torch.Tensor, |
| 124 | seq_len_tensor: torch.Tensor, |
| 125 | prompt_lora_mapping: torch.Tensor, |
| 126 | batches: int, |
| 127 | max_seq_length: int, |
| 128 | num_tokens: int, |
| 129 | add_inputs: bool, |
| 130 | ) -> None: |
| 131 | """Reference for sgmv_expand using per-LoRA matmul loop.""" |
| 132 | seq_len = _to_ref_device(seq_len_tensor) |
| 133 | mapping = _to_ref_device(prompt_lora_mapping) |
| 134 | out = _to_ref_device(out_tensor) |
| 135 | for index in range(nslices): |
| 136 | _bgmv_expand( |
| 137 | _to_ref_device(inputs_tensor[index]), |
| 138 | _to_ref_device(lora_weights_lst[index]), |
| 139 | out, |
| 140 | seq_len, |
| 141 | mapping, |
| 142 | offset=hidden_size * index, |
| 143 | add_inputs=add_inputs, |
| 144 | ) |
| 145 | if _REF_ON_CPU: |
| 146 | out_tensor.copy_(out) |
| 147 | |
| 148 | |
| 149 | _dict_lock = Lock() |
| 150 | |
| 151 | |
| 152 | def check_lora_shrink_kernel( |
| 153 | batches: int, |
| 154 | num_loras: int, |
| 155 | rank: int, |
| 156 | hidden_size: int, |
| 157 | nslices: int, |
| 158 | dtype: torch.dtype, |
| 159 | device: str, |
| 160 | seq_length: int, |
| 161 | scaling: float, |
| 162 | ): |
| 163 | """Compare outputs of torch_ops.sgmv_shrink and triton_ops.lora_shrink |
| 164 | kernels. |
| 165 | """ |
| 166 | data: PunicaTensors = generate_data_for_nslices( |
| 167 | batches, |
| 168 | hidden_size, |
| 169 | num_loras, |
| 170 | rank, |
| 171 | seq_length, |
| 172 | nslices, |
| 173 | dtype, |
| 174 | "shrink", |
| 175 | device, |
| 176 | ) |
| 177 | max_seq_length, token_nums = data.meta() |
| 178 | |
| 179 | # Setup metadata information for SGMV and reference kernels |
| 180 | sgmv_meta_args = ( |
| 181 | data.b_seq_start_loc, |
| 182 | data.seq_len_tensor, |
| 183 | data.prompt_lora_mapping, |
| 184 | batches, |
| 185 | max_seq_length, |
| 186 | token_nums, |
| 187 | ) |
| 188 | |
| 189 | # Setup metadata information for the LoRA kernel. |
| 190 | lora_meta = LoRAKernelMeta.make( |
| 191 | max_loras=num_loras, |
| 192 | max_num_tokens=token_nums, |
| 193 | device=DEVICE_TYPE, |
| 194 | ) |
| 195 | lora_meta.prepare_tensors(data.token_lora_mapping) |
| 196 | |
| 197 | ref_out_tensor = data.ref_out_tensor |
| 198 | out_tensor = data.our_out_tensor.clone() |
| 199 | |
| 200 | # Preventing cache error pointer. |
| 201 | with _dict_lock: |
| 202 | # lora_shrink kernel |
| 203 | _LORA_A_PTR_DICT.clear() |
| 204 | triton_ops.lora_shrink( |
| 205 | data.inputs_tensor, |
| 206 | data.lora_weights, |
| 207 | out_tensor, |
| 208 | *lora_meta.meta_args(token_nums=token_nums, specialize_active_lora=False), |
| 209 | scaling, |
| 210 | ) |
| 211 | |
| 212 | # Reference |
| 213 | sgmv_shrink_for_nslices( |
| 214 | nslices, |
| 215 | data.inputs_tensor, |
| 216 | data.lora_weights, |
| 217 | ref_out_tensor, |
| 218 | *sgmv_meta_args, |
| 219 | scaling, |
| 220 | ) |
| 221 | |
| 222 | assert_close(out_tensor, ref_out_tensor) |
| 223 | |
| 224 | |
| 225 | def check_lora_expand_kernel( |
| 226 | batches: int, |
| 227 | num_loras: int, |
| 228 | rank: int, |
| 229 | hidden_size: int, |
| 230 | nslices: int, |
| 231 | dtype: torch.dtype, |
| 232 | device: str, |
| 233 | seq_length: int, |
| 234 | add_inputs: bool, |
| 235 | ): |
| 236 | """Compare outputs of torch_ops.sgmv_expand and triton_ops.lora_expand |
| 237 | kernels. |
| 238 | """ |
| 239 | data: PunicaTensors = generate_data_for_nslices( |
| 240 | batches, |
| 241 | hidden_size, |
| 242 | num_loras, |
| 243 | rank, |
| 244 | seq_length, |
| 245 | nslices, |
| 246 | dtype, |
| 247 | "expand", |
| 248 | device, |
| 249 | ) |
| 250 | |
| 251 | max_seq_length, token_nums = data.meta() |
| 252 | |
| 253 | # Setup metadata information for SGMV and reference kernels |
| 254 | sgmv_meta_args = ( |
| 255 | data.b_seq_start_loc, |
| 256 | data.seq_len_tensor, |
| 257 | data.prompt_lora_mapping, |
| 258 | batches, |
| 259 | max_seq_length, |
| 260 | token_nums, |
| 261 | ) |
| 262 | |
| 263 | # Setup metadata information for the LoRA kernel. |
| 264 | lora_meta = LoRAKernelMeta.make( |
| 265 | max_loras=num_loras, |
| 266 | max_num_tokens=token_nums, |
| 267 | device=DEVICE_TYPE, |
| 268 | ) |
| 269 | lora_meta.prepare_tensors(data.token_lora_mapping) |
| 270 | |
| 271 | # Setup output tensors |
| 272 | ref_out_tensor = data.ref_out_tensor |
| 273 | out_tensor = data.our_out_tensor.clone() |
| 274 | |
| 275 | with _dict_lock: |
| 276 | # lora_expand kernel |
| 277 | _LORA_B_PTR_DICT.clear() |
| 278 | triton_ops.lora_expand( |
| 279 | data.inputs_tensor, |
| 280 | data.lora_weights, |
| 281 | out_tensor, |
| 282 | *lora_meta.meta_args(token_nums=token_nums, specialize_active_lora=False), |
| 283 | offset_start=0, |
| 284 | add_inputs=add_inputs, |
| 285 | ) |
| 286 | |
| 287 | # Reference |
| 288 | sgmv_expand_for_nslices( |
| 289 | nslices, |
| 290 | hidden_size, |
| 291 | data.inputs_tensor, |
| 292 | data.lora_weights, |
| 293 | ref_out_tensor, |
| 294 | *sgmv_meta_args, |
| 295 | add_inputs=add_inputs, |
| 296 | ) |
| 297 | |
| 298 | assert_close(out_tensor, ref_out_tensor) |
| 299 | |
| 300 | |
| 301 | # Tests |
| 302 | # We test the punica kernels along 2 verticals mainly. |
| 303 | # 1. Variations in hidden_dim size |
| 304 | # 2. Variations in all other parameters like (batch_size, max_rank, num_loras |
| 305 | # etc.) |
| 306 | |
| 307 | # We have collected the hidden_sizes included in the LoRA models |
| 308 | # currently supported by vLLM. It tests whether the corresponding Triton |
| 309 | # kernel can run normally when tensor parallelism is set to |
| 310 | # [1, 2, 4, 8, 16, 32, 64]. |
| 311 | HIDDEN_SIZES = [ |
| 312 | 128, |
| 313 | 256, |
| 314 | 512, |
| 315 | 896, |
| 316 | 1024, |
| 317 | 1152, |
| 318 | 1216, |
| 319 | 1280, |
| 320 | 1536, |
| 321 | 1664, |
| 322 | 2048, |
| 323 | 2240, |
| 324 | 2304, |
| 325 | 2368, |
| 326 | 2432, |
| 327 | 2560, |
| 328 | 2752, |
| 329 | 3072, |
| 330 | 3328, |
| 331 | 3456, |
| 332 | 3584, |
| 333 | 3712, |
| 334 | 4096, |
| 335 | 4480, |
| 336 | 4608, |
| 337 | 4736, |
| 338 | 4864, |
| 339 | 5120, |
| 340 | 5504, |
| 341 | 5632, |
| 342 | 5888, |
| 343 | 6144, |
| 344 | 6400, |
| 345 | 6848, |
| 346 | 6912, |
| 347 | 7168, |
| 348 | 7424, |
| 349 | 8192, |
| 350 | 8960, |
| 351 | 9216, |
| 352 | 9472, |
| 353 | 10240, |
| 354 | 11008, |
| 355 | 11264, |
| 356 | 13824, |
| 357 | 14336, |
| 358 | 14784, |
| 359 | 14848, |
| 360 | 15360, |
| 361 | 18944, |
| 362 | 22016, |
| 363 | 22528, |
| 364 | 24576, |
| 365 | 27392, |
| 366 | 27648, |
| 367 | 29568, |
| 368 | 29696, |
| 369 | 32000, |
| 370 | 32256, |
| 371 | 32512, |
| 372 | 32768, |
| 373 | 33024, |
| 374 | 36864, |
| 375 | 43264, |
| 376 | 49152, |
| 377 | 49408, |
| 378 | 60544, |
| 379 | 60672, |
| 380 | 64000, |
| 381 | 64256, |
| 382 | 102400, |
| 383 | 102656, |
| 384 | 128000, |
| 385 | 128256, |
| 386 | ] |
| 387 | # The size of TP |
| 388 | divisibility = [1, 2, 8, 16, 64] |
| 389 | |
| 390 | all_hidden_size = [] |
| 391 | for div in divisibility: |
| 392 | for hidden_size in HIDDEN_SIZES: |
| 393 | all_hidden_size.append(hidden_size // div) |
| 394 | |
| 395 | HIDDEN_SIZES = list(set(all_hidden_size)) |
| 396 | |
| 397 | # Test params that focuses on hidden_size variation. |
| 398 | hs_test_params = { |
| 399 | "hidden_sizes": HIDDEN_SIZES, |
| 400 | "batches": [4], |
| 401 | "num_loras": [4], |
| 402 | "max_ranks": [32], |
| 403 | } |
| 404 | |
| 405 | # General tests params that tests for variations in all dimensions |
| 406 | # except hidden_size. |
| 407 | test_params = { |
| 408 | "hidden_sizes": [2049], |
| 409 | "batches": [1, 4, 16, 32], |
| 410 | "num_loras": [1, 8, 32, 128], |
| 411 | "max_ranks": [1, 4, 8, 16, 32, 64, 128, 256], |
| 412 | } |
| 413 | |
| 414 | DTYPES = [torch.float16, torch.bfloat16] |
| 415 | DEVICES = [f"{DEVICE_TYPE}:{0}"] |
| 416 | SEED = [0] |
| 417 | |
| 418 | |
| 419 | @pytest.mark.parametrize("batches", test_params["batches"]) |
| 420 | @pytest.mark.parametrize("num_loras", test_params["num_loras"]) |
| 421 | @pytest.mark.parametrize("rank", test_params["max_ranks"]) |
| 422 | @pytest.mark.parametrize("hidden_size", test_params["hidden_sizes"]) |
| 423 | @pytest.mark.parametrize("nslices", [1, 2, 3]) |
| 424 | @pytest.mark.parametrize("dtype", DTYPES) |
| 425 | @pytest.mark.parametrize("device", DEVICES) |
| 426 | @pytest.mark.parametrize("seed", SEED) |
| 427 | @pytest.mark.parametrize("op_type", ["shrink", "expand"]) |
| 428 | def test_kernels( |
| 429 | batches: int, |
| 430 | num_loras: int, |
| 431 | rank: int, |
| 432 | hidden_size: int, |
| 433 | nslices: int, |
| 434 | dtype: torch.dtype, |
| 435 | device: str, |
| 436 | seed: int, |
| 437 | op_type: str, |
| 438 | ): |
| 439 | """Tests LoRA kernels.""" |
| 440 | torch.set_default_device(device) |
| 441 | torch.accelerator.set_device_index(device) |
| 442 | set_random_seed(seed) |
| 443 | |
| 444 | if op_type == "shrink": |
| 445 | check_lora_shrink_kernel( |
| 446 | batches=batches, |
| 447 | num_loras=num_loras, |
| 448 | rank=rank, |
| 449 | hidden_size=hidden_size, |
| 450 | nslices=nslices, |
| 451 | dtype=dtype, |
| 452 | device=device, |
| 453 | seq_length=128, |
| 454 | scaling=0.5, |
| 455 | ) |
| 456 | else: |
| 457 | check_lora_expand_kernel( |
| 458 | batches=batches, |
| 459 | num_loras=num_loras, |
| 460 | rank=rank, |
| 461 | hidden_size=hidden_size, |
| 462 | nslices=nslices, |
| 463 | dtype=dtype, |
| 464 | device=device, |
| 465 | seq_length=128, |
| 466 | add_inputs=True, |
| 467 | ) |
| 468 | |
| 469 | |
| 470 | @pytest.mark.parametrize("batches", hs_test_params["batches"]) |
| 471 | @pytest.mark.parametrize("num_loras", hs_test_params["num_loras"]) |
| 472 | @pytest.mark.parametrize("rank", hs_test_params["max_ranks"]) |
| 473 | @pytest.mark.parametrize("hidden_size", hs_test_params["hidden_sizes"]) |
| 474 | @pytest.mark.parametrize("nslices", [1, 2, 3]) |
| 475 | @pytest.mark.parametrize("dtype", DTYPES) |
| 476 | @pytest.mark.parametrize("device", DEVICES) |
| 477 | @pytest.mark.parametrize("seed", SEED) |
| 478 | @pytest.mark.parametrize("op_type", ["shrink", "expand"]) |
| 479 | def test_kernels_hidden_size( |
| 480 | batches: int, |
| 481 | num_loras: int, |
| 482 | rank: int, |
| 483 | hidden_size: int, |
| 484 | nslices: int, |
| 485 | dtype: torch.dtype, |
| 486 | device: str, |
| 487 | seed: int, |
| 488 | op_type: str, |
| 489 | ): |
| 490 | """Tests SGMV and LoRA kernels.""" |
| 491 | torch.set_default_device(device) |
| 492 | torch.accelerator.set_device_index(device) |
| 493 | set_random_seed(seed) |
| 494 | |
| 495 | if op_type == "shrink": |
| 496 | check_lora_shrink_kernel( |
| 497 | batches=batches, |
| 498 | num_loras=num_loras, |
| 499 | rank=rank, |
| 500 | hidden_size=hidden_size, |
| 501 | nslices=nslices, |
| 502 | dtype=dtype, |
| 503 | device=device, |
| 504 | seq_length=128, |
| 505 | scaling=0.5, |
| 506 | ) |
| 507 | else: |
| 508 | check_lora_expand_kernel( |
| 509 | batches=batches, |
| 510 | num_loras=num_loras, |
| 511 | rank=rank, |
| 512 | hidden_size=hidden_size, |
| 513 | nslices=nslices, |
| 514 | dtype=dtype, |
| 515 | device=device, |
| 516 | seq_length=128, |
| 517 | add_inputs=True, |
| 518 | ) |
| 519 | |
| 520 | |
| 521 | @pytest.mark.parametrize("device", DEVICES) |
| 522 | def test_add_lora_fused_moe_early_exit(device): |
| 523 | """Ensures add_lora_fused_moe does not invoke the LoRA kernel or |
| 524 | modify the output tensor when no_lora_flag_cpu is True |
| 525 | """ |
| 526 | from types import SimpleNamespace |
| 527 | |
| 528 | from vllm.lora.punica_wrapper.punica_gpu import PunicaWrapperGPU |
| 529 | |
| 530 | torch.set_default_device(device) |
| 531 | torch.accelerator.set_device_index(device) |
| 532 | |
| 533 | max_loras, num_tokens = 4, 16 |
| 534 | num_experts, top_k, max_lora_rank = 8, 2, 16 |
| 535 | K, N = 256, 128 |
| 536 | |
| 537 | # build PunicaWrapperGPU with minimal lora_config mock |
| 538 | lora_config = SimpleNamespace( |
| 539 | max_loras=max_loras, |
| 540 | specialize_active_lora=False, |
| 541 | ) |
| 542 | wrapper = PunicaWrapperGPU( |
| 543 | max_num_batched_tokens=num_tokens, |
| 544 | max_batches=num_tokens, |
| 545 | device=device, |
| 546 | lora_config=lora_config, |
| 547 | ) |
| 548 | |
| 549 | # simulate a prior LoRA batch so the internal mapping is |
| 550 | # populated with stale LoRA IDs |
| 551 | lora_mapping = torch.zeros( |
| 552 | num_tokens, |
| 553 | dtype=torch.int32, |
| 554 | device=device, |
| 555 | ) |
| 556 | lora_mapping[:8] = 1 |
| 557 | lora_mapping[8:] = 2 |
| 558 | wrapper.token_mapping_meta.prepare_tensors(lora_mapping) |
| 559 | |
| 560 | # simulate a base-model batch (all -1) |
| 561 | base_mapping = torch.full( |
| 562 | (num_tokens,), |
| 563 | -1, |
| 564 | dtype=torch.int32, |
| 565 | device=device, |
| 566 | ) |
| 567 | wrapper.token_mapping_meta.prepare_tensors(base_mapping) |
| 568 | |
| 569 | assert wrapper.token_mapping_meta.no_lora_flag_cpu[0].item() is True |
| 570 | |
| 571 | # dummy tensors for add_lora_fused_moe |
| 572 | y = torch.rand(num_tokens, top_k, N, dtype=torch.bfloat16, device=device) |
| 573 | y_snapshot = y.clone() |
| 574 | x = torch.rand(num_tokens, K, dtype=torch.bfloat16, device=device) |
| 575 | |
| 576 | lora_a_stacked = ( |
| 577 | torch.rand( |
| 578 | max_loras, |
| 579 | num_experts, |
| 580 | max_lora_rank, |
| 581 | K, |
| 582 | dtype=torch.bfloat16, |
| 583 | device=device, |
| 584 | ), |
| 585 | ) |
| 586 | lora_b_stacked = ( |
| 587 | torch.rand( |
| 588 | max_loras, |
| 589 | num_experts, |
| 590 | N, |
| 591 | max_lora_rank, |
| 592 | dtype=torch.bfloat16, |
| 593 | device=device, |
| 594 | ), |
| 595 | ) |
| 596 | topk_weights = torch.ones( |
| 597 | num_tokens, |
| 598 | top_k, |
| 599 | dtype=torch.float32, |
| 600 | device=device, |
| 601 | ) |
| 602 | adapter_enabled = torch.ones( |
| 603 | max_loras + 1, |
| 604 | dtype=torch.int32, |
| 605 | device=device, |
| 606 | ) |
| 607 | shrink_config = expand_config = { |
| 608 | "BLOCK_SIZE_M": 16, |
| 609 | "BLOCK_SIZE_N": 32, |
| 610 | "BLOCK_SIZE_K": 64, |
| 611 | "GROUP_SIZE_M": 1, |
| 612 | "NUM_WARPS": 4, |
| 613 | "NUM_STAGES": 3, |
| 614 | "SPLIT_K": 1, |
| 615 | } |
| 616 | |
| 617 | # call add_lora_fused_moe - the early exit should prevent any |
| 618 | # modification to the output |
| 619 | wrapper.add_lora_fused_moe( |
| 620 | y=y, |
| 621 | x=x, |
| 622 | lora_a_stacked=lora_a_stacked, |
| 623 | lora_b_stacked=lora_b_stacked, |
| 624 | topk_weights=topk_weights, |
| 625 | sorted_token_ids=None, |
| 626 | expert_ids=torch.zeros( |
| 627 | num_tokens * top_k, |
| 628 | dtype=torch.int32, |
| 629 | device=device, |
| 630 | ), |
| 631 | num_tokens_post_padded=None, |
| 632 | max_lora_rank=max_lora_rank, |
| 633 | top_k_num=top_k, |
| 634 | shrink_config=shrink_config, |
| 635 | expand_config=expand_config, |
| 636 | adapter_enabled=adapter_enabled, |
| 637 | ) |
| 638 | |
| 639 | assert torch.equal(y, y_snapshot), ( |
| 640 | "add_lora_fused_moe modified output tensor despite no_lora_flag_cpu=True" |
| 641 | ) |