| [FlagGems] fix IndexError: index 0 is out of bounds for dimension 0 with size 0 | 5 个月前 |
| Add FlagGems v4.2.0 | 5 个月前 |
| Add FlagGems v4.2.0 | 5 个月前 |
| Add --skip option to import skip list file for flaggems Usage: pytest test_xxx.py --skip='skip_tests.txt' The file contains a name list of testcases, e.g. test_accuracy_instancenorm[False-True-True-dtype2-shape1] test_accuracy_kron[dtype5-shape9] ... | 4 个月前 |
| Add FlagGems v4.2.0 | 5 个月前 |
| [FlagGems] Fix the flash_attention_forward operator | 19 天前 |
| [FlagGems]Fix for 'Tensor-like are not close!' in binary_pointwise test cases. Added equal_nan=True in the test case of addcdiv Ops. All division with bf16 and float16 is verified with equal_nan=True option. floordiv uses fmod Ops which is invoking fmod of sleef library. Removed Sleef call and it uses np.fmod. In the mlir it will transform to arith::remf. Corrected the test case. Upcasting and downcasting of reference input causes small precision difference. Fixed like other non GPU targets. | 4 个月前 |
| Add FlagGems v4.2.0 | 5 个月前 |
| Add FlagGems v4.2.0 | 5 个月前 |
| Add FlagGems v4.2.0 | 5 个月前 |
| Add FlagGems v4.2.0 | 5 个月前 |
| handle file races when pytest with more than one processes | 3 个月前 |
| Add FlagGems v4.2.0 | 5 个月前 |
| Add FlagGems v4.2.0 | 5 个月前 |
| Fix vector_norm test tolerance for large reduction dimensions Problem: test_accuracy_vectornorm failed for cases with large reduction dimensions (e.g., shape=(4096,256), dim=0) due to accumulated floating-point errors exceeding default tolerance. Root cause: The default tolerance (atol=1e-4, rtol=rtol*dtype) does not scale with the number of elements being reduced. When reducing many elements, the accumulated error grows proportionally. Solution: Add reduce_dim parameter to gems_assert_close, which scales atol by the number of reduced elements, following the pattern used in other reduction tests (e.g., sum, cross_entropy). | 4 个月前 |
| [FlagGems] fix the error like: 1. AttributeError: 'dict' object has no attribute 'total_memory' (test_pointwise_dynamic.py) 2. AttributeError: module 'torch.cpu' has no attribute 'get_device_capability' (test_pointwise_dynamic.py) 3. AssertionError: Torch not compiled with CUDA enabled (test_quant.py) 4. TypeError: unsupported operand type(s) for //: 'int' and 'NoneType' (test_general_reduction_ops.py) | 5 个月前 |
| Add FlagGems v4.2.0 | 5 个月前 |
| Fix for 'Tensor-like are not close!' Assertion in torch.add in kunpeng vendor torch.add has another definition under ARM Backend folder. And also marked as unused custom Ops. So the FlagGems invoked the default torch.add not from the FlagGems implementation. Added a new kunpeng vendor. The customized Add implementation is not there. | 4 个月前 |
| Add FlagGems v4.2.0 | 5 个月前 |
| Add FlagGems v4.2.0 | 5 个月前 |
| Fix hardcoded device type. | 1 个月前 |
| Add FlagGems v4.2.0 | 5 个月前 |
| [FlagGems] Update kernel name to fix syntax error in generated kernel | 4 个月前 |
| Add FlagGems v4.2.0 | 5 个月前 |