Write a guide for quantizing models using frameworks outside of Compressed-Tensors quantization

[kylesayrs]

Background

• Some users want to quantize models and encode using the compressed-tensors format, but do not necessarily want to use compressed-tensors primitives

Proposed Changes

• Add documentation for encoding a model using compressed tensors format using, for example, QAT

[xiao-yang25]

Hi, I’d be happy to contribute to this issue.
I reviewed the current checkpoint conversion workflow. It supports converting already-quantized external checkpoints into compressed-tensors layouts, with ModelOptNvfp4Converter and AutoAWQConverter providing concrete examples.
Would it be useful to scope the guide around:
producing a quantized checkpoint with an external framework;
adapting its tensor names, layouts, scales, and zero-points through the Converter interface;
documenting how the corresponding QuantizationConfig is generated and automatically written to the converted checkpoint; and
verifying that the resulting checkpoint can be loaded for inference?
The issue also mentions QAT as an example. Since the existing converters operate on already-quantized checkpoints rather than covering the external quantization or training step itself, should the guide include a concrete end-to-end QAT workflow using a particular framework—and, if necessary, a new converter—or should it focus on documenting how externally produced checkpoints can be encoded through the existing conversion API?
I also noticed that the conversion documentation was migrated to the llm-compressor repository as the single source of truth. Should this guide live there or in compressed-tensors?
If this scope sounds appropriate, I’d be happy to take it.