AMDMIGraphX

AMDMIGraphX copied to clipboard

Model performing worse or similar to eager pytorch. Model Location: `/mnt/nas_share/migraphx/models/torch_benchmarks/dlrm` - [ ] Turn on the MLIR flags - [ ] Capture MIGX trace - [ ] Compare traces...

shivadbhavsar

Model performing worse or similar to eager pytorch. Model Location: `/mnt/nas_share/migraphx/models/torch_benchmarks/drq`

shivadbhavsar

Model performing worse or similar to eager pytorch. Model Location: `/mnt/nas_share/migraphx/models/torch_benchmarks/hf_Longformer`

shivadbhavsar

Torch benchmarks fail. Affected models: Torchbench: hf_BigBird

shivadbhavsar

Improve handling for adding more than 10 parameters to a submodule

4

Currently, we add parameters with names `x1`, `x2`, ... , `x9`, `x10`. When we sort them `x10` will come after `x1` instead of at the end. This updates the `param_name`...

pfultz2

Fix targets registration on Windows

10

Unfortunately, automatic target registration does not work on Windows. The linker will not add a specified DLL as a dependency if at least a symbol from the dynamic library is...

apwojcik

packed_or_broadcasted: gpu::mlir_op: Shapes are not packed nor broadcasted

3

Torch benchmarks fail. Affected models: TIMM: xcit_large_24_p8_224 HF: MobileBertForQuestionAnswering, MobileBertForMaskedLM

shivadbhavsar

Compute broadcastlen error in find_inner_broadcast

1

For mixed broadcasts, the pass fails when the shapes of the inputs are not broadcastable. Following program snippet reproduces the issue: ``` # broadcastlen.py import migraphx import numpy as np...

shivadbhavsar

This can be done by shifting zero point by -128.

umangyadav

pfultz2