Daniel Hernandez-Juarez issues

Repositories
Issues
Comments

Results 3 issues of


                                            Daniel Hernandez-Juarez

[Issue]: accuracy_checker.py is not within tolerance for bert_large_uncased_1_fp16_gpu.onnx

### Problem Description I'm running (MIGraphX develop branch): ```bash python3 ../tools/accuracy/accuracy_checker.py --onnx ~/mlir-dev/AMDMIGraphX/build/bert_large_uncased_1_fp16_gpu.onnx --fill1 --input-dim input_ids:1,384 --disable-fast-math --tolerance 0.02 --verbose ``` And I get the error: ``` Max Difference: 0.13617822527885437...

Remove rocmlir unsupported reduce types

We are going to enable f16 for reduce in this rocmlir PR: https://github.com/ROCm/rocMLIR/pull/1722 However, the code here assumes bf16 and fp8 are supported and that's not the case.

test MFMA tiled layout in migraphx

### DOR (Definition of Ready) ready ### Description The purpose of this ticket is to try the tiled mfma-friendly layout POC from rocmlir (branch: mfma_layout_migraphx_integration). ## Layout explanation The idea...