已合并
【fix】set_tol_and_add_fallback #40178
Lu_G创建于 7月6日
【fix】set_tol_and_add_fallback #40178
已合并
共 2 个文件变更+33-0
| @@ -564,6 +564,14 @@ def _patch_model_21(): | |||
| 564 | # The current operator suffers from severe performance degradation. | 564 | # The current operator suffers from severe performance degradation. |
| 565 | # This patch will be removed after the issue is fixed in the future. | 565 | # This patch will be removed after the issue is fixed in the future. |
| 566 | patch_remove_decomposition(["aten._softmax"]) | 566 | patch_remove_decomposition(["aten._softmax"]) |
| 567 | + try: | ||
| 568 | + from torch_npu._inductor.ascend_npu_ir.ascend_npu_ir import ( | ||
| 569 | + config as anir_config, | ||
| 570 | + ) | ||
| 571 | + | ||
| 572 | + anir_config.force_fallback_kernel_names["mlir_fused_add_lt_neg_where_16"] = True | ||
| 573 | + except ImportError: | ||
| 574 | + log.warning("import config failed for hf_T5_base patch") | ||
 | |||
| 567 | 575 | ||
| 568 | 576 | ||
| 569 | 577 | ||
| @@ -571,6 +579,14 @@ def _patch_model_22(): | |||
| 571 | # The current operator suffers from severe performance degradation. | 579 | # The current operator suffers from severe performance degradation. |
| 572 | # This patch will be removed after the issue is fixed in the future. | 580 | # This patch will be removed after the issue is fixed in the future. |
| 573 | patch_remove_decomposition(["aten._softmax"]) | 581 | patch_remove_decomposition(["aten._softmax"]) |
| 582 | + try: | ||
| 583 | + from torch_npu._inductor.ascend_npu_ir.ascend_npu_ir import ( | ||
| 584 | + config as anir_config, | ||
| 585 | + ) | ||
| 586 | + | ||
| 587 | + anir_config.force_fallback_kernel_names["mlir_fused_add_lt_neg_where_16"] = True | ||
| 588 | + except ImportError: | ||
| 589 | + log.warning("import config failed for hf_T5_large patch") | ||
| 574 | 590 | ||
| 575 | 591 | ||
| 576 | 592 | ||
| @@ -45,9 +45,15 @@ from torch._dynamo.utils import clone_inputs | |||
| 45 | 45 | ||
| 46 | 46 | ||
| 47 | try: | 47 | try: |
| 48 | + import torch_npu | ||
| 48 | from npu_support import _patch_environment_variables | 49 | from npu_support import _patch_environment_variables |
| 49 | except ImportError: | 50 | except ImportError: |
| 50 | pass | 51 | pass |
| 52 | +try: | ||
| 53 | + from torch_npu.npu._backends import get_soc_version | ||
| 54 | + is_ascend950 = get_soc_version() >= 260 # Ascend950 = 260 | ||
| 55 | +except (ImportError, AttributeError): | ||
| 56 | + is_ascend950 = False | ||
| 51 | from benchmark.userbenchmark.dynamo.dynamobench.torchbench import ( | 57 | from benchmark.userbenchmark.dynamo.dynamobench.torchbench import ( |
| 52 | DONT_CHANGE_BATCH_SIZE, | 58 | DONT_CHANGE_BATCH_SIZE, |
| 53 | FORCE_AMP_FOR_FP16_BF16_MODELS, | 59 | FORCE_AMP_FOR_FP16_BF16_MODELS, |
| @@ -120,6 +126,15 @@ NPU_REQUIRE_HIGHER_TOLERANCE = { | |||
| 120 | "phlippe_resnet", | 126 | "phlippe_resnet", |
| 121 | } | 127 | } |
| 122 | 128 | ||
| 129 | +NPU_REQUIRE_HIGHER_TOLERANCE_ON_A5 = { | ||
| 130 | + "densenet121", | ||
| 131 | + "resnet50", | ||
| 132 | + "resnet18", | ||
| 133 | + "resnext50_32x4d", | ||
| 134 | + "resnet152", | ||
| 135 | +} | ||
| 136 | + | ||
| 137 | + | ||
| 123 | 138 | ||
| 124 | NPU_REQUIRE_EVEN_HIGHER_TOLERANCE = { | 139 | NPU_REQUIRE_EVEN_HIGHER_TOLERANCE = { |
| 125 | "shufflenet_v2_x1_0", | 140 | "shufflenet_v2_x1_0", |
| @@ -438,6 +453,8 @@ class TorchBenchmarkRunner(BenchmarkRunner): | |||
| 438 | tolerance = 1e-2 | 453 | tolerance = 1e-2 |
| 439 | elif name in NPU_REQUIRE_EVEN_HIGHER_TOLERANCE: | 454 | elif name in NPU_REQUIRE_EVEN_HIGHER_TOLERANCE: |
| 440 | tolerance = 2e-2 | 455 | tolerance = 2e-2 |
| 456 | + elif is_ascend950 and name in NPU_REQUIRE_HIGHER_TOLERANCE_ON_A5: | ||
| 457 | + tolerance = 2e-3 | ||
| 441 | return tolerance, cosine | 458 | return tolerance, cosine |
| 442 | 459 | ||
| 443 | def get_learning_rate(self, is_training, current_device, name): | 460 | def get_learning_rate(self, is_training, current_device, name): |
🟡 Medium Priority
register_patch装饰器通过字典赋值_patch_table[model_name] = fn来注册补丁，后注册的会覆盖先注册的。_patch_model_3修复了 T5Attention.forward 中两次连续_to_copy导致的精度问题（将scores += position_bias_masked替换为scores = scores.float() + position_bias_masked），是一个关键的精度修复。_patch_model_21仅包含patch_remove_decomposition(["aten._softmax"])和 force fallback 内核设置，不包含_patch_model_3的逻辑。后果：对
hf_T5_base模型，_patch_model_3不再执行，T5Attention.forward 不会被替换，可能导致该模型的精度检查失败。触发条件：在 NPU 上运行
hf_T5_base模型的 benchmark 精度检查时。建议：需要在 _patch_model_21 中集成 _patch_model_3 的逻辑，或者修改 register_patch 机制使其支持同一模型注册多个补丁。最直接的修复：在 _patch_model_21 函数开头添加 _patch_model_3 中的 T5Attention.forward 替换逻辑。或者，如果确认 _patch_model_3 的修复对 hf_T5_base 已不再需要（例如已被其他机制覆盖），应在代码注释中明确说明原因。