As mentioned here: https://onnxruntime.ai/docs/performance/model-optimizations/float16.html#float16-tool-arguments
the float16.DEFAULT_OP_BLOCK_LIST should contain unsupported ops for given datatypes, this is no longer up to date as newer upsets for example do support float16 for resize.
Would it be possible to have DEFAULT_OP_BLOCK_LIST be a function instead of a static variable that returns different lists depending on the given opset ?