Paul Fultz II
Paul Fultz II
> but you need to push it to internal branch on Gerrit. I dont have access to push to that repo. Does this project not accept open source contributions?
> Are instructions available @ ROCm-Developer-Tools/HIP#2104 (comment) not working for you ? Those instructions of course do not work as they cannot be used by source-based package managers. This PR...
Here's a reduced test case: ```python p = migraphx.program() m = p.get_main_module() x_0 = m.add_literal(migraphx.generate_argument(migraphx.shape(type="float_type", lens=[5,784,768]), 0)) x_1 = m.add_literal(migraphx.generate_argument(migraphx.shape(type="float_type", lens=[1]), 1)) p_x = m.add_parameter("x",migraphx.shape(type="float_type", lens=[5,784,768])) x_3 = m.add_instruction(migraphx.op("reduce_mean", axes=[-1]),...
I would assume a test for `concat(relu, slice, relu, slice)` should reproduce the issue. If that doesnt trigger the issue then we can investigate why. We can use `MIGRAPHX_TRACE_MATCHES=1` to...
So `SoftmaxCrossEntropyLoss` is a onnx function which means it can be implemented using a composite of other onnx operators(like `DynamicQuantizeLinear`). So we shouldn't need to add a new operator for...
> > Furthermore `NegativeLogLikelihoodLoss` is also an ONNX function as well, so we dont need to write a custom operator for this as well, although I couldn't find a reference...
Can you paste the output of `migraphx-driver read ./test/onnx/gridsample_simple_test.onnx --py` here? This way we can debug this without needing GridSample branch. Very likely we need to add `logical_and` implementation [here](https://github.com/ROCm/AMDMIGraphX/blob/develop/src/targets/gpu/compile_gen.cpp#L201)....
Ah I know the issue, its uses more than 10 parameters, which messes up the sorting.
I added the `param_name` function in order to put a name that can be sorted correctly. Let me update all of our fusions to use that(and update it to handle...