Remove QuantizationScheme.default_scheme

[kylesayrs]

Purpose

• Remove unused function. I believe that this function was initially intended to provide defaults to QuantizationModifier while not overwriting existing configs, but since all these values are all written to configs and the QuantizationModifier is now also the default for configs which do not specify values, this function is no longer necessary.

Prerequisites

• Explicit defaults for QuantizationModifier targets vllm-project/llm-compressor#889

Changes

• Remove QuantizationScheme.default_scheme

Testing

• Grepped codebases for uses

[dsikka]

Can you verify behaviour for cases that where we're only running kv_cache quantization? That is the only case I can recall where this function was relevant.

[kylesayrs]

@dsikka Good idea. I tested a modified examples/quantization_kv_cache/llama3_fp8_kv_example.py with only kv cache quantization e2e with vllm
LC = kylesayrs/quantization-modifier-defaults
CT = this branch

[dsikka]

For completeness, could we validate that the example loads properly in vllm?
Alternatively, just make sure the keys/values are attached to the attention block in the state dict

[kylesayrs]

@dsikka Yep, I validated that both kvcache quantization with weight&input quantization and without weight&input quantization load and produce valid results in vllm.

[horheynm]

I remember hitting issues with this since it was setting default scheme which was not intended.
I do think we can get rid of this