[GPTQ] Change actorder default to "static"

[kylesayrs]

Purpose

• Use best defaults for GPTQ quantization

Prerequisites

• [GPTQ] Fix actorder resolution, add sentinel #1453
• [Tests] GPTQ Actorder Resolution Tests #1468

Changes

• Set gptq actorder default to "static"

Testing

• Ran llama w4a16 example to completion and validated the correct activation ordering

[dsikka]

nice!

[markurtz]

@kylesayrs quick question from my side. Since the old default was None, do we risk defaulting to an incorrect value for older recipes that don't include it? Especially for ones that specified it in the quantization scheme?

The base branch was changed.

[kylesayrs]

@markurtz

1. "older" recipes which did not specify will now use "static". This is (more or less) inevitable and imho acceptable, since a user which does not specify actorder probably does not care and just wants the best configuration
2. For recipes which specify anything except "static", they will see this error encouraging them to modify their recipe to disable global actorder

The base branch was changed.

[kylesayrs]

Waiting for next weekly to run before merging

The base branch was changed.

[dsikka]

FYI this will incorrectly save models for our e2e testing:
https://github.com/vllm-project/llm-compressor/tree/main/tests/e2e/vLLM/configs (specifically, essentially remove the cases where there is no act order while duplicating some cases)

Same for lm-eval tests:
https://github.com/vllm-project/llm-compressor/tree/main/tests/lmeval/configs

[kylesayrs]

@dsikka

specifically, essentially remove the cases where there is no act order while duplicating some cases

Turning all non-specified actorder cases into actorder cases is the intention of this PR.

As for duplicating tests, I don't think any are actually duplicated? In order for a duplication to occur, it must conflict with an existing weight actorder config

w4a16_actorder_weight.yaml
w4a16_actorder_weight_qwen.yaml

vl_w4a16_actorder_weight.yaml
w4a16_actorder_weight.yaml

These are all w4a16. If you grep for non-specified w4a16 configs, you get

w4a16_channel_quant.yaml  # channelwise, not duplicating
w4a16_channel_quant_qwen.yaml  # channelwise, not duplicating
w4a16_grouped_quant.yaml  # different calibration and dataset, not duplicating
w4a16_grouped_quant_asym_awq.yaml  # AWQ, not duplicating
w4a16_2of4_channel_quant.yaml  # 2of4, not duplicating
w4a16_2of4_grouped_quant.yaml  # 2of4, not duplicating

All of the other w4a16 configs test for different things, except for maybe w4a16_grouped_quant.yaml which essentially tests for the same thing but has slightly different model and dataset

[dsikka]

w4a16_actorder_weight.yaml

1. We lose all non-act order cases. We still need to test this in e2e and lm-eval
2. We need to update the naming in the model produced/pushed to nm-testing so that we indicate act order being present
3. We have two identical lm-eval cases?

tests/lmeval/configs/w4a16_actorder_weight.yaml and tests/lmeval/configs/w4a16_grouped_quant.yaml

[kylesayrs]

https://github.com/neuralmagic/llm-compressor-testing/actions/runs/15934262421

[kylesayrs]

We lose all non-act order cases. We still need to test this in e2e and lm-eval

Just posted a run above

We need to update the naming in the model produced/pushed to nm-testing so that we indicate act order being present

Since actorder is now the default, we do not need to include it in model names, since the convention is that default arguments are not included in the model name

We have two identical lm-eval cases?

As I mentioned previously, these two tests are slightly different. After the above nightly finishes, I'll delete tests/lmeval/configs/w4a16_actorder_weight.yaml

[dsikka]

We lose all non-act order cases. We still need to test this in e2e and lm-eval

Just posted a run above

We need to update the naming in the model produced/pushed to nm-testing so that we indicate act order being present

Since actorder is now the default, we do not need to include it in model names, since the convention is that default arguments are not included in the model name

We have two identical lm-eval cases?

As I mentioned previously, these two tests are slightly different. After the above nightly finishes, I'll delete tests/lmeval/configs/w4a16_actorder_weight.yaml

This is for our internal testing. Please keep the naming convention until we have adaptability of this case being our default.

And as I mentioned, we still need to test a case with no act order.