Custom Ops: Unit Tests

[matthewdouglas]

Provide unit testing coverage for all new custom ops.

• Tests must support CUDA and, where applicable, CPU devices
• Must be device-agnostic, i.e. can add support for additional devices for new backends.
• Tests must pass the torch.library.opcheck tests.
• Tests must use all combinations of supported data types and options for each op.
• Operator tests should compare outputs with a slower, but known correct, PyTorch based implementation.

• Int8 blockwise quant/dequant
• LLM.int8 quant/dequant
• LLM.int8 matmul
• 4bit blockwise quant/dequant
• 4bit matmul