Robert Knight
Robert Knight
The spec-mandated behavior here of returning the minimum or maximum value for the type does have an issue that it can mask errors, less so with floats where the min/max...
For reference https://developer.chrome.com/docs/css-ui/css-names discusses how `@property` is _supposed_ to work with Shadow DOM and how it actually behaves. https://github.com/w3c/csswg-drafts/issues/10541 is a specification issue concerning `@property` and Shadow DOM.
Some notes about 4-bit quantization via standard operators only: ONNX Runtime will [fuse DequantizeLinear + MatMul](https://github.com/microsoft/onnxruntime/blob/0463aa9fc3ef02d30d7177c0065cd4b7d36a39f7/onnxruntime/core/optimizer/qdq_transformer/selectors_actions/qdq_actions.cc#L282) into MatMulNBits. This makes it possible to create and distribute int4-quantized models using standard...
For the default [`DecodeMethod::Greedy`](https://docs.rs/ocrs/latest/ocrs/enum.DecodeMethod.html) decoding of the neural network outputs there is a score per character that could be exposed. If the beam search decoding method is used, it gets...
https://github.com/robertknight/rten/issues/911 changes the API for enabling subsets of operators so that it doesn't require referencing the operator types in `rten::ops`. Instead a separate set of "key" types have been created...
> I work with bitmasks quite a bit for highly parallel search algorithms, and looked for a way to get a bitmask from the rten-simd masks, like I can via...
> One possible solution could be to instead leverage the bitvec crate which can already abstract over different storage types and lengths with a consistent bitmask API. Thanks for the...
I think that could work. num_traits is already used indirectly by the main crate in this project. The next question is how the complexity/efficiency of a `to_bitmask` implementation for each...
Hmm... that's a bit more exotic. Something I didn't mention earlier is a general workaround for missing features in the portable APIs, which is to wrap and extend the base...
> The TensorProto messages in the ONNX model don't contain any metadata to identify the tensors as being packed 4-bit values Interestingly this is not the case when using "QDQ"...