refactor algorithm cli by n1ck-guo · Pull Request #2140 · intel/auto-round

n1ck-guo · 2026-08-11T01:23:20Z

Description

refactor algorithm config and cli

add register_args in alg config, automatically parses parameters, no need to manually add them in the CLI.
remove deprecated args: algorithm
remove rotation_config, pass by alg_config

old	new
AutoRound(m, rotation_config="quarot")	AutoRound(m, alg_configs=["auto_round", "quarot"])
AutoRound(m, rotation_config="default")	AutoRound(m, alg_configs=["auto_round", "hadamard"])
AutoRound(m, rotation_config="random_hadamard")	AutoRound(m, alg_configs=["auto_round", "random_hadamard"])
AutoRound(m, rotation_config=SpinQuantConfig(...))	AutoRound(m, alg_configs=["auto_round", SpinQuantConfig(...)])
AutoRound(m, rotation_config={"algorithm": "spinquant", ...})	dictionary format is no longer supported. construct `SpinQuantConfig(...)` directly and pass it to `alg_configs`.

Type of Change

Refactor

Related Issues

Fixes or relates to #

Checklist Before Submitting

My code has been tested locally.
Documentation has been updated as needed.
New or updated tests are included where applicable.
The CUDA CI has passed. You can trigger it by commenting /azp run Unit-Test-CUDA-AutoRound.

Signed-off-by: n1ck-guo <heng.guo@intel.com>

Signed-off-by: n1ck-guo <heng.guo@intel.com> # Conflicts: # test/unit/test_cuda/transform/test_spinquant.py

…d TypeError - Revert the strict TypeError raised for kwargs that don't match the selected alg_configs (e.g. disable_opt_rtn passed with a non-RTN algorithm). This broke real callers (llm-compressor's AutoRoundModifier always passes disable_opt_rtn regardless of iters) and our own test_audio_model.py test. Log via logger.error and ignore instead, same as before the earlier hardening pass. The auto-discovery mechanism (_discover_alg_config_fields / _owning_algorithm_names) is unchanged. - Remove the stray disable_opt_rtn=True left over from copy-paste in test_quantize_with_tuning (iters=1 selects SignRound, so the RTN-only flag never applied). - Fix pylint line-too-long (148/120) in AWQConfig.register_args help text. Signed-off-by: n1ck-guo <heng.guo@intel.com>

Signed-off-by: n1ck-guo <heng.guo@intel.com> # Conflicts: # auto_round/autoround.py # auto_round/cli/algorithms.py

AutoRoundBot · 2026-08-14T03:03:46Z

/azp run Unit-Test-CUDA-AutoRound

azure-pipelines · 2026-08-14T03:03:58Z

Azure Pipelines successfully started running 1 pipeline(s).

lkk12014402 · 2026-08-24T03:41:41Z

Description

refactor algorithm config and cli

add register_args in alg config, automatically parses parameters, no need to manually add them in the CLI.

remove deprecated args: algorithm

remove rotation_config, pass by alg_config

old new
AutoRound(m, rotation_config="quarot") AutoRound(m, alg_configs=["auto_round", "quarot"])
AutoRound(m, rotation_config="default") AutoRound(m, alg_configs=["auto_round", "hadamard"])
AutoRound(m, rotation_config="random_hadamard") AutoRound(m, alg_configs=["auto_round", "random_hadamard"])
AutoRound(m, rotation_config=SpinQuantConfig(...)) AutoRound(m, alg_configs=["auto_round", SpinQuantConfig(...)])
AutoRound(m, rotation_config={"algorithm": "spinquant", ...}) dictionary format is no longer supported. construct SpinQuantConfig(...) directly and pass it to alg_configs.

Type of Change

Refactor

Related Issues

Fixes or relates to #

Checklist Before Submitting

My code has been tested locally.

Documentation has been updated as needed.

New or updated tests are included where applicable.

The CUDA CI has passed. You can trigger it by commenting /azp run Unit-Test-CUDA-AutoRound.

Is rotation always applied before quantization regardless
of position (["quarot", "auto_round"] == ["auto_round", "quarot"])?

document the canonical execution order (pre-processors → rotation → block
quantizer) and make it independent of list position, or make position meaningful and say
so. This directly affects how rotation members are ordered in AlgorithmComposer.members().

lkk12014402 · 2026-08-24T07:32:01Z

+    def register_args(cls, registry: AlgorithmParameterRegistry) -> None:
+        mutex = registry.add_mutually_exclusive_group()
+        mutex.add_argument(
+            "--disable_opt_rtn",


Each config now declares parameters twice: once in __init__ (the kwargs) and again in
register_args (the CLI). For example RTNConfig declares disable_opt_rtn /
enable_opt_rtn in both places; SignRoundConfig re-lists ~15 fields.

These two lists must be kept in sync by hand; nothing enforces it. A field added to
__init__ but forgotten in register_args is silently unavailable from the CLI (and
vice-versa).

how about consider driving both from a single source of truth — pydantic Field(...) /
dataclass field(metadata=...) — so the CLI declaration is derived from the config
fields, not maintained alongside them.

for example

from pydantic import BaseModel, Field class RTNConfig(BaseModel): group_size: int = Field( default=128, description="quant group size" ) symmetric: bool = Field( default=True, description="symmetric quantization" )

auto generate cli args:

for field in RTNConfig.model_fields: parser.add_argument(...)

This is a good design, thanks. I will consider it. I had considered a similar approach before, but handling aliases is a bit difficult.
let me evaluate it.

no problem.

If the algorithm does not provide customized cli arguments, it would be better to fall back to this approach. The potential issue is that some parameters may have the same name, so we should automatically add a prefix to those parameters to avoid conflicts.

# Conflicts: # auto_round/autoround.py # auto_round/cli/algorithms.py # test/unit/test_cpu/core/test_autoround_entry.py

wenhuach21 · 2026-08-26T05:57:05Z

+    """
+    from auto_round.algorithms.registry import iter_algorithm_entries
+
+    result = []


As the entry of AutoRound, personally I don't like there are too many helper functions at the beginning of the file

wenhuach21 · 2026-08-26T05:59:40Z

-        seqlen: int = None,
+        alg_configs=None,
        **kwargs,
    ) -> "BaseCompressor":


We should annotate this class as thoroughly as possible, e.g., explaining the meaning of each argument, what can be passed in kwargs, and providing some examples.

wenhuach21 · 2026-08-26T06:04:14Z

+    from auto_round.algorithms.transforms.spinquant.preprocessor import SpinQuantConfig
+
+    register_algorithm("rtn", aliases=("rtn",), config_factory=RTNConfig, summary="Round-To-Nearest quantization.")
+    register_algorithm(


To follow the principle that each algorithm should only modify the files it owns, I’d suggest splitting this code into the respective algorithm files.

wenhuach21 · 2026-08-26T06:05:11Z

-        )
-        group.add_argument(
-            "--awq_clip_as_init",
-            dest="awq_clip_as_init",


Thanks for the refinements. The current version looks better.

Signed-off-by: n1ck-guo <heng.guo@intel.com>

AutoRoundBot · 2026-08-31T05:53:34Z

/azp run Unit-Test-CUDA-AutoRound

azure-pipelines · 2026-08-31T05:53:44Z

Azure Pipelines successfully started running 1 pipeline(s).

n1ck-guo added 4 commits August 11, 2026 09:21

refactor algorithm cli

30e5ee8

Signed-off-by: n1ck-guo <heng.guo@intel.com>

remove legance algorithm arg

2331ac4

Signed-off-by: n1ck-guo <heng.guo@intel.com>

remove class compressorbuilder

005d786

Signed-off-by: n1ck-guo <heng.guo@intel.com>

clean autoround api

8d331cd

Signed-off-by: n1ck-guo <heng.guo@intel.com>

chensuyue added this to the 0.16.0 milestone Aug 11, 2026

n1ck-guo added 3 commits August 11, 2026 15:50

remove rotation_config, pass by alg_configs

be5c0df

Signed-off-by: n1ck-guo <heng.guo@intel.com>

Merge branch 'main' into hengguo/refactor_cli

ec27d80

Signed-off-by: n1ck-guo <heng.guo@intel.com> # Conflicts: # test/unit/test_cuda/transform/test_spinquant.py

n1ck-guo force-pushed the hengguo/refactor_cli branch from 7cee70a to 8c8dee3 Compare August 12, 2026 05:29

Merge branch 'main' into hengguo/refactor_cli

1b38e95

Signed-off-by: n1ck-guo <heng.guo@intel.com> # Conflicts: # auto_round/autoround.py # auto_round/cli/algorithms.py

lkk12014402 reviewed Aug 24, 2026

View reviewed changes

Merge remote-tracking branch 'origin/main' into hengguo/refactor_cli

128a51c

# Conflicts: # auto_round/autoround.py # auto_round/cli/algorithms.py # test/unit/test_cpu/core/test_autoround_entry.py

wenhuach21 reviewed Aug 26, 2026

View reviewed changes

n1ck-guo added 2 commits August 31, 2026 08:49

Merge remote-tracking branch 'origin/main' into hengguo/refactor_cli

739b3d7

update by comments

a67796b

Signed-off-by: n1ck-guo <heng.guo@intel.com>

n1ck-guo mentioned this pull request Aug 31, 2026

[Feature]: Align args in all the AutoRound's interface #2076

Open

Uh oh!

Conversation

n1ck-guo commented Aug 11, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Description

Type of Change

Related Issues

Checklist Before Submitting

Uh oh!

AutoRoundBot commented Aug 14, 2026

Uh oh!

azure-pipelines Bot commented Aug 14, 2026

Uh oh!

lkk12014402 commented Aug 24, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Description

Type of Change

Related Issues

Checklist Before Submitting

Uh oh!

lkk12014402 Aug 24, 2026

Choose a reason for hiding this comment

Uh oh!

n1ck-guo Aug 25, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

lkk12014402 Aug 25, 2026

Choose a reason for hiding this comment

Uh oh!

wenhuach21 Aug 26, 2026

Choose a reason for hiding this comment

Uh oh!

wenhuach21 Aug 26, 2026

Choose a reason for hiding this comment

Uh oh!

wenhuach21 Aug 26, 2026

Choose a reason for hiding this comment

Uh oh!

wenhuach21 Aug 26, 2026

Choose a reason for hiding this comment

Uh oh!

wenhuach21 Aug 26, 2026

Choose a reason for hiding this comment

Uh oh!

AutoRoundBot commented Aug 31, 2026

Uh oh!

azure-pipelines Bot commented Aug 31, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

n1ck-guo commented Aug 11, 2026 •

edited

Loading

lkk12014402 commented Aug 24, 2026 •

edited

Loading

n1ck-guo Aug 25, 2026 •

edited

Loading