Merged
fix(dequant_situ_quant): aclnn第一段补齐输入shape/组合约束校验 #14701
demoaugusteAdd on 4 days ago
fix(dequant_situ_quant): aclnn第一段补齐输入shape/组合约束校验 #14701
Merged
Merged fromdemoauguste/ops-nn-utmasterintocann/ops-nnmaster
Total: 2 files changed+874-7
| @@ -82,7 +82,7 @@ static bool CheckNotNull(const aclTensor* x, const aclTensor* yOut, const aclTen | |||
| 82 | } | 82 | } |
| 83 | 83 | ||
| 84 | // 2. 检查条件必选输入是否为空指针:INT8/INT32必选weightScaleOptional,INT32必选activationScaleOptional, | 84 | // 2. 检查条件必选输入是否为空指针:INT8/INT32必选weightScaleOptional,INT32必选activationScaleOptional, |
| 85 | -// static量化必选quantScaleOptional | 85 | +// INT8+static量化必选quantScaleOptional |
| 86 | static bool CheckConditionalInputNotNull(const aclTensor* x, const aclTensor* weightScaleOptional, | 86 | static bool CheckConditionalInputNotNull(const aclTensor* x, const aclTensor* weightScaleOptional, |
| 87 | const aclTensor* activationScaleOptional, const aclTensor* quantScaleOptional, | 87 | const aclTensor* activationScaleOptional, const aclTensor* quantScaleOptional, |
| 88 | QuantType quantType) | 88 | QuantType quantType) |
| @@ -94,7 +94,7 @@ static bool CheckConditionalInputNotNull(const aclTensor* x, const aclTensor* we | |||
| 94 | if (xDtype == op::DataType::DT_INT32) { | 94 | if (xDtype == op::DataType::DT_INT32) { |
| 95 | OP_CHECK_NULL(activationScaleOptional, return false); | 95 | OP_CHECK_NULL(activationScaleOptional, return false); |
| 96 | } | 96 | } |
| 97 | - if (quantType == QuantType::STATIC) { | 97 | + if (quantType == QuantType::STATIC && xDtype == op::DataType::DT_INT8) { |
| 98 | OP_CHECK_NULL(quantScaleOptional, return false); | 98 | OP_CHECK_NULL(quantScaleOptional, return false); |
| 99 | } | 99 | } |
| 100 | return true; | 100 | return true; |
| @@ -248,8 +248,8 @@ static bool CheckXShape(const aclTensor* x) | |||
| 248 | return true; | 248 | return true; |
| 249 | } | 249 | } |
| 250 | 250 | ||
| 251 | -// 9. 检查输出shape的维度与x匹配 | 251 | +// 9. 检查输出shape的维度与x匹配(含未知维(-1)时无法精确推导,仅校验rank) |
| 252 | -static bool CheckOutShape(const aclTensor* x, const aclTensor* yOut, const aclTensor* yScaleOut) | 252 | +static bool CheckOutShapeRank(const aclTensor* x, const aclTensor* yOut, const aclTensor* yScaleOut) |
| 253 | { | 253 | { |
| 254 | const size_t xDimNum = x->GetViewShape().GetDimNum(); | 254 | const size_t xDimNum = x->GetViewShape().GetDimNum(); |
| 255 | const size_t expectedYDimNum = (x->GetDataType() == op::DataType::DT_INT8) ? xDimNum : DIM_NUM_NOT_INT8; | 255 | const size_t expectedYDimNum = (x->GetDataType() == op::DataType::DT_INT8) ? xDimNum : DIM_NUM_NOT_INT8; |
| @@ -270,6 +270,284 @@ static bool CheckOutShape(const aclTensor* x, const aclTensor* yOut, const aclTe | |||
| 270 | return true; | 270 | return true; |
| 271 | } | 271 | } |
| 272 | 272 | ||
| 273 | +// 10. 检查可选输入是否应缺省(当前数据类型/量化模式下不使用) | ||
| 274 | +static bool IsTensorAbsent(const aclTensor* tensor, const char* paramName, const char* reason) | ||
| 275 | +{ | ||
| 276 | + if (tensor == nullptr) { | ||
| 277 | + return true; | ||
| 278 | + } | ||
| 279 | + ReportInvalidValue(paramName, "not null", reason); | ||
| 280 | + return false; | ||
| 281 | +} | ||
| 282 | + | ||
| 283 | +// 11. 检查tensor的shape size是否为指定取值之一 | ||
| 284 | +static bool CheckShapeSizeOneOf(const aclTensor* tensor, const char* paramName, int64_t firstSize, int64_t secondSize, | ||
| 285 | + const char* expectedRule) | ||
| 286 | +{ | ||
| 287 | + const int64_t size = tensor->GetViewShape().GetShapeSize(); | ||
| 288 | + if (size == firstSize || size == secondSize) { | ||
| 289 | + return true; | ||
| 290 | + } | ||
| 291 | + const std::string reason = (std::string(paramName) + " shape size must be " + expectedRule + ", but got " + | ||
| 292 | + std::to_string(size) + "."); | ||
| 293 | + ReportInvalidShape(paramName, op::ToString(tensor->GetViewShape()).GetString(), reason); | ||
| 294 | + return false; | ||
| 295 | +} | ||
| 296 | + | ||
| 297 | +// 12. 检查tensor的shape是否与期望shape一致 | ||
| 298 | +static bool CheckShapeExpected(const aclTensor* tensor, const char* paramName, const op::Shape& expectedShape, | ||
| 299 | + const char* expectedRule) | ||
| 300 | +{ | ||
| 301 | + if (tensor->GetViewShape() == expectedShape) { | ||
| 302 | + return true; | ||
| 303 | + } | ||
| 304 | + const std::string reason = (std::string(paramName) + " shape must be " + expectedRule + ", but got " + | ||
| 305 | + std::string(op::ToString(tensor->GetViewShape()).GetString()) + "."); | ||
| 306 | + ReportInvalidShape(paramName, op::ToString(tensor->GetViewShape()).GetString(), reason); | ||
| 307 | + return false; | ||
| 308 | +} | ||
| 309 | + | ||
| 310 | +// 13. 检查1维非空tensor(groupIndexOptional) | ||
| 311 | +static bool CheckNonEmptyOneDim(const aclTensor* tensor, const char* paramName) | ||
| 312 | +{ | ||
| 313 | + const op::Shape& shape = tensor->GetViewShape(); | ||
| 314 | + if (shape.GetDimNum() == 1 && shape.GetDim(0) > 0) { | ||
| 315 | + return true; | ||
| 316 | + } | ||
| 317 | + const std::string reason = (std::string(paramName) + " must be a non-empty 1-D tensor, but got " + | ||
| 318 | + std::string(op::ToString(shape).GetString()) + "."); | ||
| 319 | + ReportInvalidShape(paramName, op::ToString(shape).GetString(), reason); | ||
| 320 | + return false; | ||
| 321 | +} | ||
| 322 | + | ||
| 323 | +// 14. 检查INT32路径weightScaleOptional/biasOptional的shape:无group_index时为[2H]/[1]/[1,2H], | ||
| 324 | +// 有group_index时为[expertNum, 2H] | ||
| 325 | +static bool CheckMoeScaleShape(const aclTensor* tensor, const char* paramName, int64_t expertNum, int64_t inDimy, | ||
| 326 | + bool hasGroupIndex) | ||
| 327 | +{ | ||
| 328 | + const op::Shape& shape = tensor->GetViewShape(); | ||
| 329 | + const size_t dimNum = shape.GetDimNum(); | ||
| 330 | + bool valid = false; | ||
| 331 | + if (hasGroupIndex) { | ||
| 332 | + valid = (dimNum == static_cast<size_t>(NUM_TWO) && shape.GetDim(0) == expertNum && shape.GetDim(1) == inDimy); | ||
| 333 | + } else { | ||
| 334 | + valid = (dimNum == 1 && (shape.GetDim(0) == inDimy || shape.GetDim(0) == 1)) || | ||
| 335 | + (dimNum == static_cast<size_t>(NUM_TWO) && shape.GetDim(0) == 1 && shape.GetDim(1) == inDimy); | ||
| 336 | + } | ||
| 337 | + if (valid) { | ||
| 338 | + return true; | ||
| 339 | + } | ||
| 340 | + const std::string reason = (std::string(paramName) + " shape is invalid, expertNum=" + std::to_string(expertNum) + | ||
| 341 | + ", inputWidth=" + std::to_string(inDimy) + "."); | ||
| 342 | + ReportInvalidShape(paramName, op::ToString(shape).GetString(), reason); | ||
| 343 | + return false; | ||
| 344 | +} | ||
| 345 | + | ||
| 346 | +// 15. 检查INT32路径activationScaleOptional的shape:(1,)或每行一个值 | ||
| 347 | +static bool CheckInt32ActivationScaleShape(const aclTensor* tensor, int64_t rows) | ||
| 348 | +{ | ||
| 349 | + const op::Shape& shape = tensor->GetViewShape(); | ||
| 350 | + const bool rankValid = (shape.GetDimNum() == 1) || | ||
| 351 | + (shape.GetDimNum() == static_cast<size_t>(NUM_TWO) && shape.GetDim(1) == 1); | ||
| 352 | + const bool sizeValid = (shape.GetShapeSize() == rows) || (shape.GetShapeSize() == 1); | ||
| 353 | + if (rankValid && sizeValid) { | ||
| 354 | + return true; | ||
| 355 | + } | ||
| 356 | + const std::string reason = (std::string("activationScaleOptional shape must be (1,) or contain one FP32 value " | ||
| 357 | + "per row, but got ") + | ||
| 358 | + std::string(op::ToString(shape).GetString()) + "."); | ||
| 359 | + ReportInvalidShape("activationScaleOptional", op::ToString(shape).GetString(), reason); | ||
| 360 | + return false; | ||
| 361 | +} | ||
| 362 | + | ||
| 363 | +// 16. INT8路径的shape约束(与tiling侧入参校验保持一致) | ||
| 364 | +static bool CheckInt8ShapeConstraints(const aclTensor* x, const aclTensor* weightScaleOptional, | ||
| 365 | + const aclTensor* activationScaleOptional, const aclTensor* biasOptional, | ||
| 366 | + const aclTensor* quantScaleOptional, const aclTensor* quantOffsetOptional, | ||
| 367 | + const aclTensor* groupIndexOptional, QuantType quantType, const aclTensor* yOut, | ||
| 368 | + const aclTensor* yScaleOut) | ||
| 369 | +{ | ||
| 370 | + const op::Shape& xShape = x->GetViewShape(); | ||
| 371 | + const size_t dimNum = xShape.GetDimNum(); | ||
| 372 | + const int64_t inDimy = xShape.GetDim(dimNum - 1); | ||
| 373 | + const int64_t outDimy = inDimy / NUM_TWO; | ||
| 374 | + | ||
| 375 | + if (!CheckShapeSizeOneOf(weightScaleOptional, "weightScaleOptional", inDimy, 1, "(2H,) or (1,)")) { | ||
| 376 | + return false; | ||
| 377 | + } | ||
| 378 | + if (!IsTensorAbsent(activationScaleOptional, "activationScaleOptional", | ||
| 379 | + "activationScaleOptional is not used when the data type of x is INT8.")) { | ||
| 380 | + return false; | ||
| 381 | + } | ||
| 382 | + if (biasOptional != nullptr && !CheckShapeSizeOneOf(biasOptional, "biasOptional", inDimy, 1, "(2H,) or (1,)")) { | ||
| 383 | + return false; | ||
| 384 | + } | ||
| 385 | + if (quantScaleOptional != nullptr && | ||
| 386 | + !CheckShapeSizeOneOf(quantScaleOptional, "quantScaleOptional", outDimy, 1, "(H,) or (1,)")) { | ||
| 387 | + return false; | ||
| 388 | + } | ||
| 389 | + if (quantType == QuantType::STATIC) { | ||
| 390 | + if (quantOffsetOptional != nullptr && | ||
| 391 | + !CheckShapeSizeOneOf(quantOffsetOptional, "quantOffsetOptional", outDimy, 1, "(H,) or (1,)")) { | ||
| 392 | + return false; | ||
| 393 | + } | ||
| 394 | + } else if (!IsTensorAbsent(quantOffsetOptional, "quantOffsetOptional", | ||
| 395 | + "quantOffsetOptional is only supported when quantTypeOptional is static.")) { | ||
| 396 | + return false; | ||
| 397 | + } | ||
| 398 | + if (!IsTensorAbsent(groupIndexOptional, "groupIndexOptional", | ||
| 399 | + "groupIndexOptional is not used when the data type of x is INT8.")) { | ||
| 400 | + return false; | ||
| 401 | + } | ||
| 402 | + | ||
| 403 | + op::Shape expectedYShape; | ||
| 404 | + op::Shape expectedYScaleShape; | ||
| 405 | + for (size_t i = 0; i + 1 < dimNum; i++) { | ||
| 406 | + expectedYShape.AppendDim(xShape.GetDim(i)); | ||
| 407 | + expectedYScaleShape.AppendDim(xShape.GetDim(i)); | ||
| 408 | + } | ||
| 409 | + expectedYShape.AppendDim(outDimy); | ||
| 410 | + if (!CheckShapeExpected(yOut, "yOut", expectedYShape, "x.shape[:-1] + [H] (H = x.shape[-1] / 2)")) { | ||
| 411 | + return false; | ||
| 412 | + } | ||
| 413 | + return CheckShapeExpected(yScaleOut, "yScaleOut", expectedYScaleShape, "x.shape[:-1]"); | ||
| 414 | +} | ||
| 415 | + | ||
| 416 | +// 17. INT32路径的shape约束(与tiling侧入参校验保持一致) | ||
| 417 | +static bool CheckInt32ShapeConstraints(const aclTensor* x, const aclTensor* weightScaleOptional, | ||
| 418 | + const aclTensor* activationScaleOptional, const aclTensor* biasOptional, | ||
| 419 | + const aclTensor* quantScaleOptional, const aclTensor* quantOffsetOptional, | ||
| 420 | + const aclTensor* groupIndexOptional, QuantType quantType, const aclTensor* yOut, | ||
| 421 | + const aclTensor* yScaleOut) | ||
| 422 | +{ | ||
| 423 | + const op::Shape& xShape = x->GetViewShape(); | ||
| 424 | + const int64_t rows = xShape.GetDim(0); | ||
| 425 | + const int64_t inDimy = xShape.GetDim(1); | ||
| 426 | + const int64_t outDimy = inDimy / NUM_TWO; | ||
| 427 | + | ||
| 428 | + if (quantType != QuantType::DYNAMIC) { | ||
| 429 | + ReportInvalidValue("quantTypeOptional", "static", | ||
| 430 | + "quantTypeOptional must be dynamic when the data type of x is INT32."); | ||
| 431 | + return false; | ||
| 432 | + } | ||
| 433 | + if (!IsTensorAbsent(quantScaleOptional, "quantScaleOptional", | ||
| 434 | + "quantScaleOptional is not used when the data type of x is INT32.")) { | ||
| 435 | + return false; | ||
| 436 | + } | ||
| 437 | + if (!IsTensorAbsent(quantOffsetOptional, "quantOffsetOptional", | ||
| 438 | + "quantOffsetOptional is not used when the data type of x is INT32.")) { | ||
| 439 | + return false; | ||
| 440 | + } | ||
| 441 | + const bool hasGroupIndex = (groupIndexOptional != nullptr); | ||
| 442 | + int64_t expertNum = 1; | ||
| 443 | + if (hasGroupIndex) { | ||
| 444 | + if (!CheckNonEmptyOneDim(groupIndexOptional, "groupIndexOptional")) { | ||
| 445 | + return false; | ||
| 446 | + } | ||
| 447 | + expertNum = groupIndexOptional->GetViewShape().GetDim(0); | ||
| 448 | + } | ||
| 449 | + if (!CheckMoeScaleShape(weightScaleOptional, "weightScaleOptional", expertNum, inDimy, hasGroupIndex)) { | ||
| 450 | + return false; | ||
| 451 | + } | ||
| 452 | + if (!CheckInt32ActivationScaleShape(activationScaleOptional, rows)) { | ||
| 453 | + return false; | ||
| 454 | + } | ||
| 455 | + if (biasOptional != nullptr && | ||
| 456 | + !CheckMoeScaleShape(biasOptional, "biasOptional", expertNum, inDimy, hasGroupIndex)) { | ||
| 457 | + return false; | ||
| 458 | + } | ||
| 459 | + | ||
| 460 | + op::Shape expectedYShape; | ||
| 461 | + expectedYShape.AppendDim(rows); | ||
| 462 | + expectedYShape.AppendDim(outDimy); | ||
| 463 | + if (!CheckShapeExpected(yOut, "yOut", expectedYShape, "[rows, input_width / 2]")) { | ||
| 464 | + return false; | ||
| 465 | + } | ||
| 466 | + op::Shape expectedYScaleShape; | ||
| 467 | + expectedYScaleShape.AppendDim(rows); | ||
| 468 | + return CheckShapeExpected(yScaleOut, "yScaleOut", expectedYScaleShape, "[rows]"); | ||
| 469 | +} | ||
| 470 | + | ||
| 471 | +// 18. BFLOAT16/FLOAT16(预反量化)路径的shape约束(与tiling侧入参校验保持一致) | ||
| 472 | +static bool CheckPreDequantShapeConstraints(const aclTensor* x, const aclTensor* weightScaleOptional, | ||
| 473 | + const aclTensor* activationScaleOptional, const aclTensor* biasOptional, | ||
| 474 | + const aclTensor* quantScaleOptional, const aclTensor* quantOffsetOptional, | ||
| 475 | + const aclTensor* groupIndexOptional, QuantType quantType, | ||
| 476 | + const aclTensor* yOut, const aclTensor* yScaleOut) | ||
| 477 | +{ | ||
| 478 | + const op::Shape& xShape = x->GetViewShape(); | ||
| 479 | + const int64_t rows = xShape.GetDim(0); | ||
| 480 | + const int64_t inDimy = xShape.GetDim(1); | ||
| 481 | + const int64_t outDimy = inDimy / NUM_TWO; | ||
| 482 | + | ||
| 483 | + if (quantType != QuantType::DYNAMIC) { | ||
| 484 | + ReportInvalidValue("quantTypeOptional", "static", | ||
| 485 | + "quantTypeOptional must be dynamic when the data type of x is BFLOAT16/FLOAT16."); | ||
| 486 | + return false; | ||
| 487 | + } | ||
| 488 | + if (!IsTensorAbsent(weightScaleOptional, "weightScaleOptional", | ||
| 489 | + "weightScaleOptional is not used when the data type of x is BFLOAT16/FLOAT16.")) { | ||
| 490 | + return false; | ||
| 491 | + } | ||
| 492 | + if (!IsTensorAbsent(activationScaleOptional, "activationScaleOptional", | ||
| 493 | + "activationScaleOptional is not used when the data type of x is BFLOAT16/FLOAT16.")) { | ||
| 494 | + return false; | ||
| 495 | + } | ||
| 496 | + if (!IsTensorAbsent(biasOptional, "biasOptional", | ||
| 497 | + "biasOptional is not used when the data type of x is BFLOAT16/FLOAT16.")) { | ||
| 498 | + return false; | ||
| 499 | + } | ||
| 500 | + if (!IsTensorAbsent(quantScaleOptional, "quantScaleOptional", | ||
| 501 | + "quantScaleOptional is not used when the data type of x is BFLOAT16/FLOAT16.")) { | ||
| 502 | + return false; | ||
| 503 | + } | ||
| 504 | + if (!IsTensorAbsent(quantOffsetOptional, "quantOffsetOptional", | ||
| 505 | + "quantOffsetOptional is not used when the data type of x is BFLOAT16/FLOAT16.")) { | ||
| 506 | + return false; | ||
| 507 | + } | ||
| 508 | + if (groupIndexOptional != nullptr && !CheckNonEmptyOneDim(groupIndexOptional, "groupIndexOptional")) { | ||
| 509 | + return false; | ||
| 510 | + } | ||
| 511 | + | ||
| 512 | + op::Shape expectedYShape; | ||
| 513 | + expectedYShape.AppendDim(rows); | ||
| 514 | + expectedYShape.AppendDim(outDimy); | ||
| 515 | + if (!CheckShapeExpected(yOut, "yOut", expectedYShape, "[rows, input_width / 2]")) { | ||
| 516 | + return false; | ||
| 517 | + } | ||
| 518 | + op::Shape expectedYScaleShape; | ||
| 519 | + expectedYScaleShape.AppendDim(rows); | ||
| 520 | + return CheckShapeExpected(yScaleOut, "yScaleOut", expectedYScaleShape, "[rows]"); | ||
| 521 | +} | ||
| 522 | + | ||
| 523 | +// 19. 检查输入/输出shape约束(含输入之间的shape关系),与tiling侧入参校验口径一致: | ||
| 524 | +// 第一段接口即可判定的约束在此返回161002,避免落到tiling返回561002 | ||
| 525 | +static bool CheckShapeConstraints(const aclTensor* x, const aclTensor* weightScaleOptional, | ||
| 526 | + const aclTensor* activationScaleOptional, const aclTensor* biasOptional, | ||
| 527 | + const aclTensor* quantScaleOptional, const aclTensor* quantOffsetOptional, | ||
| 528 | + const aclTensor* groupIndexOptional, QuantType quantType, const aclTensor* yOut, | ||
| 529 | + const aclTensor* yScaleOut) | ||
| 530 | +{ | ||
| 531 | + const op::Shape& xShape = x->GetViewShape(); | ||
| 532 | + // 含未知维(-1)的动态shape无法在host侧精确推导,仅校验输出rank,交由执行阶段校验 | ||
| 533 | + if (HasUnknownDim(xShape) || HasUnknownDim(yOut->GetViewShape()) || HasUnknownDim(yScaleOut->GetViewShape())) { | ||
| 534 | + return CheckOutShapeRank(x, yOut, yScaleOut); | ||
| 535 | + } | ||
| 536 | + if (x->GetDataType() == op::DataType::DT_INT8) { | ||
| 537 | + return CheckInt8ShapeConstraints(x, weightScaleOptional, activationScaleOptional, biasOptional, | ||
| 538 | + quantScaleOptional, quantOffsetOptional, groupIndexOptional, quantType, yOut, | ||
| 539 | + yScaleOut); | ||
| 540 | + } | ||
| 541 | + if (x->GetDataType() == op::DataType::DT_INT32) { | ||
| 542 | + return CheckInt32ShapeConstraints(x, weightScaleOptional, activationScaleOptional, biasOptional, | ||
| 543 | + quantScaleOptional, quantOffsetOptional, groupIndexOptional, quantType, yOut, | ||
| 544 | + yScaleOut); | ||
| 545 | + } | ||
| 546 | + return CheckPreDequantShapeConstraints(x, weightScaleOptional, activationScaleOptional, biasOptional, | ||
| 547 | + quantScaleOptional, quantOffsetOptional, groupIndexOptional, quantType, yOut, | ||
| 548 | + yScaleOut); | ||
| 549 | +} | ||
| 550 | + | ||
| 273 | static aclnnStatus CheckParams(const aclTensor* x, const aclTensor* weightScaleOptional, | 551 | static aclnnStatus CheckParams(const aclTensor* x, const aclTensor* weightScaleOptional, |
| 274 | const aclTensor* activationScaleOptional, const aclTensor* biasOptional, | 552 | const aclTensor* activationScaleOptional, const aclTensor* biasOptional, |
| 275 | const aclTensor* quantScaleOptional, const aclTensor* quantOffsetOptional, | 553 | const aclTensor* quantScaleOptional, const aclTensor* quantOffsetOptional, |
| @@ -302,9 +580,11 @@ static aclnnStatus CheckParams(const aclTensor* x, const aclTensor* weightScaleO | |||
| 302 | // 7. 检查format是否符合要求 | 580 | // 7. 检查format是否符合要求 |
| 303 | CHECK_RET(CheckFormat(x, yOut, yScaleOut), ACLNN_ERR_PARAM_INVALID); | 581 | CHECK_RET(CheckFormat(x, yOut, yScaleOut), ACLNN_ERR_PARAM_INVALID); |
| 304 | 582 | ||
| 305 | - // 8. 检查x和输出的shape维度是否符合约束 | 583 | + // 8. 检查shape约束(含x维度、输入之间的shape关系及输出shape) |
| 306 | CHECK_RET(CheckXShape(x), ACLNN_ERR_PARAM_INVALID); | 584 | CHECK_RET(CheckXShape(x), ACLNN_ERR_PARAM_INVALID); |
| 307 | - CHECK_RET(CheckOutShape(x, yOut, yScaleOut), ACLNN_ERR_PARAM_INVALID); | 585 | + CHECK_RET(CheckShapeConstraints(x, weightScaleOptional, activationScaleOptional, biasOptional, quantScaleOptional, |
| 586 | + quantOffsetOptional, groupIndexOptional, quantType, yOut, yScaleOut), | ||
| 587 | + ACLNN_ERR_PARAM_INVALID); | ||
| 308 | return ACLNN_SUCCESS; | 588 | return ACLNN_SUCCESS; |
| 309 | } | 589 | } |
| 310 | } // namespace | 590 | } // namespace |
| @@ -487,7 +487,7 @@ TEST_F(l2_dequant_situ_quant_test, valid_int8_static_without_quant_offset) | |||
| 487 | TEST_F(l2_dequant_situ_quant_test, valid_int32_dynamic_params) | 487 | TEST_F(l2_dequant_situ_quant_test, valid_int32_dynamic_params) |
| 488 | { | 488 | { |
| 489 | auto x_desc = TensorDesc({16, 64}, ACL_INT32, ACL_FORMAT_ND).ValueRange(-2, 2); | 489 | auto x_desc = TensorDesc({16, 64}, ACL_INT32, ACL_FORMAT_ND).ValueRange(-2, 2); |
| 490 | - auto weight_scale_desc = TensorDesc({64}, ACL_FLOAT, ACL_FORMAT_ND); | 490 | + auto weight_scale_desc = TensorDesc({2, 64}, ACL_FLOAT, ACL_FORMAT_ND); |
| 491 | auto activation_scale_desc = TensorDesc({1}, ACL_FLOAT, ACL_FORMAT_ND); | 491 | auto activation_scale_desc = TensorDesc({1}, ACL_FLOAT, ACL_FORMAT_ND); |
| 492 | auto group_index_desc = TensorDesc({2}, ACL_INT64, ACL_FORMAT_ND); | 492 | auto group_index_desc = TensorDesc({2}, ACL_INT64, ACL_FORMAT_ND); |
| 493 | auto y_desc = TensorDesc({16, 32}, ACL_INT8, ACL_FORMAT_ND); | 493 | auto y_desc = TensorDesc({16, 32}, ACL_INT8, ACL_FORMAT_ND); |
【review】 问题分类:正确性 位置:quant/dequant_situ_quant/op_host/op_api/aclnn_dequant_situ_quant.cpp:533,代码:if (HasUnknownDim(xShape) || HasUnknownDim(yOut->GetViewShape()) || HasUnknownDim(yScaleOut->GetViewShape())) { 问题:这里发现未知维就只验输出 rank,跳过了后面与维度无关的 dtype 和量化模式组合校验。例如 INT32 x=[-1,64]、weightScale=[64]、activationScale=[1]、INT8 y=[16,32]、FLOAT yScale=[16]、quantType=static,且不传 quantScale。此前第一段会拦截静态量化缺少 quantScale,本 PR 将该前置约束限于 INT8 后,这组参数会通过 CheckParams;INT32 本身仍只支持 dynamic。BF16/FP16 路径也有同一问题。 可能导致的结果:无论未知维最终取什么值,静态模式都无法执行。调用方会在更晚的内部处理或 tiling 阶段才遇到失败,无法在第一段输入校验得到明确的参数错误。 建议的修改方法:先校验 dtype 与量化模式、以及与尺寸无关的可选输入组合,再根据未知维决定是否跳过依赖具体尺寸的检查。补一条 INT32 未知维且传 static 的负向用例。