feature/onnx-to-tosa #4
Draft
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
…odel has missing output types. Signed-off-by: Rickert, Jonas <[email protected]>
Signed-off-by: Rickert, Jonas <[email protected]>
Signed-off-by: Rickert, Jonas <[email protected]>
Signed-off-by: Rickert, Jonas <[email protected]>
Signed-off-by: Rickert, Jonas <[email protected]>
Signed-off-by: Rickert, Jonas <[email protected]>
Add optional output type inference
…composition Revert "chore: temporarily disable combine parallel conv recomposition."
…t is after all weights Signed-off-by: Rickert, Jonas <[email protected]>
…se_fixes Use none instead of nullptr for bias and make sure the insertion point is after all weights
Add missing cmake depends
…ith_dq Chaitany.handling convtranspose with dq
…ith_dq Fixing the Quant and Dequant node output shapes
[Backport] convtranspose qdq
…le uses. This is fine as it is the last required operation in the Layer/RMSNorm and its recomposition pattern. The optional bias gets added in a separate pattern Signed-off-by: Rickert, Jonas <[email protected]>
Recompose a Layer/RMSNorm even if the scale multiplication has multiple uses.
This is done by catching exceptions during shape inference (as they happen for example if the model uses custom ops) and by falling back to an onnx-mlir based type mapping for some kinds of invalid models. Pass ImportOptions by const reference instead of value to avoid unnecessay copying. Signed-off-by: Rickert, Jonas <[email protected]>
…rs/systems Signed-off-by: Rickert, Jonas <[email protected]>
…ypes Make parsing of models more robust if useOnnxModelTypes is enabled.
Fix problem of return type conversion that only works on some compilers/systems
Co-authored-by: Jonas Rickert <[email protected]>
Handle default attributes in decomposition
…ention_ops Add support for RotaryEmbedding and Attention ONNX ops
ONNXConstantOp: Allow extra attributes (parser/printer)
…pying the element type not only for the first result
Always try to take the element type from custom ops from the model, instead of guessing
Signed-off-by: Rickert, Jonas <[email protected]>
…md.quark Signed-off-by: Rickert, Jonas <[email protected]>
Signed-off-by: Rickert, Jonas <[email protected]>
Allow the import of not-builtin domains and add support for AMD Quarks BFPQuantizeDequantizeOp
According to the onnx documentation, the default domain is "ai.onnx", but in practice it seems like the empty-string domain is used by default. Add a canonicalization function to handle them in the same way. Signed-off-by: Rickert, Jonas <[email protected]>
Handle aliasing of "" domain and "ai.onnx" domain.
Signed-off-by: Rickert, Jonas <[email protected]>
Add BF16 support to AMDQuarkBFPQuantizeDequantizeOp
Add ResultNames attribute to imported ops
feat: sigmoid constprop
added slice decomposition, default off
…near-fix [AIESW-17979] ONNXToTOSA reject DequantizeLinear on dynamic input
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
DO NOT merge, just to look at our diff