[Feature Request][Help Wanted] Convert AutoAWQ checkpoints to compressed-tensors

[dsikka]

Is your feature request related to a problem? Please describe.

• Add a conversion tool in compressed-tensors which when given an AutoAWQ checkpoint, can convert the model to the compressed-tensors format using the pack_quantized compressor
• The checkpoint should also contain an updated quantization_config in its config.json with metadata about the format. See example: https://huggingface.co/nm-testing/Qwen3-Coder-30B-A3B-Instruct-W4A16-awq/blob/main/config.json
• This can be done by representing the quantization params for the AutoAWQ model using QuantizationArgs and then applying the ModelCompressor to the model

Describe the solution you'd like

• A tool which when given an AutoAWQ checkpoint, produced a compressed-tensors formatted model
• The model should be able to run in vLLM without any drop in accuracy

[brian-dellabetta]

Working implementation will have to account for AutoAWQ's unique ordering for packing 8 int4 weight into a single int32. More detail here

• Can't using llmcompressor for awq quantization and using autoawq completely replace each other? #1909 (comment)

[aladerran]

I will be working on this!

[dsikka]

Amazing! I’ll assign it to you!
Thank you!