-
Notifications
You must be signed in to change notification settings - Fork 3.6k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Add Gemm as an operator #47
Conversation
Rather than discuss this design choice for GEMM specifically, I continued the discussion started in #24. |
From offline discussion: add broadcasting to C |
also regenerated the operators doc
Why it isn't experiment initially? |
@ebarsoum It was experimental initially in this PR, in my most recent commit to this PR I have moved it out of experimental. |
@bddppq why we move it out of experiment? |
@ebarsoum From offline discussion with @yuanbyu @prasanthpul and @dzhulgakov, we will add Gemm (as non-experimental) and remove FC |
* Add error checks to file IO in onnx2trt * Fix handling of auto_pad for opset 7 * Fix handling of Add/Mul broadcasting for opset 7 * Add support for tensor Div/Sub weights * Prevent squeeze_trailing_dims removing all dims * Fix shape bugs in combineTensorsElementwise - This function incorrectly handled weights with rank != tensor_rank, particularly with respect to the batch dim, which a given weights array may or may not have. - It also incorrectly attempted to expand the dims of tensors that had insufficient rank, which cannot be done due to the batch dim always (implicitly) being the left-most dim in TRT.
Having Gemm would be super helpful for backend optimizing work. Also many popular frameworks have operator similar to Gemm, adding it will be convenient for their frontends to export to onnx.