AMDMIGraphX
AMDMIGraphX copied to clipboard
ROCmSoftwarePlatform
AMD's graph optimization engine.
Sigmoid not being merged into horizontally fused convolution because of multiple outputs. Minimal example program: ``` import migraphx import numpy as np p = migraphx.program() mm = p.get_main_module() shape_1 =...
In PR https://github.com/ROCm/AMDMIGraphX/pull/3568 it was noted that the sentence > Set to "1", "enable", "enabled", "yes", or "true" to use. It is not specific enough and anyone running it through...
Running with the same input values / seed can generate slightly different image pixels
- Partially resolves https://github.com/migraphx-benchmark/AMDMIGraphX/issues/200 - Reference implementation - ONNX parser
Reduce op onnx backend tests that have empty set inputs cause the operator to generate dynamic shapes, causing a failure in those tests cases. Currently we are just excluding any...
``` if(solution.is_null() or not tuning_cfg_path.empty()) get_module_tuned(); ``` The above code change allows rocMLIR team to run with a config even if solution is null (has been requested in the past),...
Spec here https://github.com/microsoft/onnxruntime/blob/main/docs/ContribOperators.md#com.microsoft.Attention Useful resources: https://towardsdatascience.com/transformers-explained-visually-part-2-how-it-works-step-by-step-b49fa4a64f34/ https://towardsdatascience.com/transformers-explained-visually-part-3-multi-head-attention-deep-dive-1c1ff1024853/
