AMDMIGraphX
AMDMIGraphX copied to clipboard
ROCmSoftwarePlatform
AMD's graph optimization engine.
DequantizeLinear operations failed during compilation with Mistral model due to zero scale values for mistral-7b with mixed precision.
## Motivation Add cubic resize functionality for resize op ## Technical Details Related to customer support and or adding functionality to resize we're seeing in the depthanything model ## Changelog...
* allows for passing in dynamic input dims to existing tf models * modified some operators to make them dynamic, more remaining as TODOs
## Motivation Remove the GroupQueryAttention ref op and use equivalent ops in its place. ## Technical Details ## Changelog Category - - [ ] Added: New functionality. - - [x]...
Generate a raw mask from each mask_index number used for right padding. fix mask Fix to output conversion for operator as well as broadcast shapes ## Motivation ## Technical Details...
### Problem Description ONNX models take an exceedingly long amount of time to compile when using MIGraphX as the Execution Provider. ### Operating System Ubuntu 24.04.2 LTS (Noble Numbat) (WSL2...
### Problem Description migraphx for rocm-6.4.3 only builds with outdated oneDNN versions like [v2.7.3](https://github.com/uxlfoundation/oneDNN/releases/tag/v2.7.3). oneDNN has now reached v3.9 upstream. oneDNN 2.7.3 is incompatible with CMake 4 and can no...
- Resolves https://github.com/migraphx-benchmark/AMDMIGraphX/issues/197
## Motivation To be able to use amdmlss library selected kernels ## Technical Details Adds mlss_mha operator, fuse_mlss pass, and mlss jit compiler files as initial proposal of instruction replacement...
## Motivation Break out the list of supported Onnx operators (official and Contrib) that we support. This breaks out functionality seen within the MIGraphX driver via the `migraphx-driver onnx -l`...