Support MX serialization

[stamalakhov]

What

Let's support mxint8 quantized model serialization. Right now tico/quantization/wrapq/examples/quantize_linear.py fails to save model with observer set to MXObserver with elem_format set to int8. No decomposition and serialization.

Why

Need this to produce quantized llm model in circle format.

[dayo09]

Just to note,

MXINT8 will be used as intermediate data format in TRIV3, but it isn't sure if it will be used as IO practically.
It will be good to have MXINT8 support I/O format, but the scenario isn't yet developed.

As you may know,
some operators(Normalization, Conv2d, etc) will use the feature to enhance precision at calculation of int8/int16 I/O operators.

(example)

INT16 input -> (conversion to MXINT8) -> RMS Normalization -> (conversion to INT16) -> INT16 output
# which is planned to be supported inside NPU_Compiler

But pre-quantized MXINT8 I/O hasn't come up yet.

/cc @parjong

[stamalakhov]

But pre-quantized MXINT8 I/O hasn't come up yet.
😢

@dayo09 Thank you for your reply.

[dayo09]

😢

We are still planning about TRIV3 support, so the situation could vary at any moment.
TICO has been referred as one option to support MXINT8 quantization, so let's see how it goes. 😃

[stamalakhov]

We are still planning about TRIV3 support, so the situation could vary at any moment.
TICO has been referred as one option to support MXINT8 quantization, so let's see how it goes. 😃

@dayo09 Got it. Thank you.