PTQ-Level Efficiency, QAT-Level Accuracy | 训练后量化的效率,量化感知训练的精度
Gate large FP4 packing test by GPU memory (#1696)
Yi Liua4dee9862881
diff --git a/test/test_cuda/quantization/test_packing.py b/test/test_cuda/quantization/test_packing.pyindex cffd36293..8c346b5a3 100644--- a/test/test_cuda/quantization/test_packing.py+++ b/test/test_cuda/quantization/test_packing.py@@ -4,6 +4,15 @@ import torchfrom auto_round.export.export_to_autoround.qlinear_fp import FLOAT_TO_E2M1, pack_fp4_to_uint8+def _cuda_device_has_more_than_40gb_memory():+ if not torch.cuda.is_available():+ return False++ device_index = torch.cuda.current_device()+ total_memory = torch.cuda.get_device_properties(device_index).total_memory+ return total_memory > 40_000_000_000++# Random sampling from FLOAT_TO_E2M1def _create_random_e2m1_tensor(shape):"""Create a tensor of the given shape with random values from FLOAT_TO_E2M1."""@@ -66,7 +75,12 @@ qwen_weight_shapes = [torch.Size([128, 2048]),torch.Size([512, 2048]),torch.Size([4096, 2048]),- torch.Size([151936, 2048]),+ pytest.param(+ torch.Size([151936, 2048]),+ marks=pytest.mark.skipif(+ not _cuda_device_has_more_than_40gb_memory(), reason="requires a CUDA device with more than 40 GB memory"+ ),+ ),torch.Size([2048, 4096]),]