migration of quantize_ workflow configuration from callables to configs

[vkuzo]

summary

This issue tracks the migration of quantize_ per-workflow configuration from Callables to configs..

We are migrating the way quantize_ workflows are configured from callables (tensor subclass inserters) to direct configuration (config objects). Motivation: align with the rest of the ecosystem, enable inspection of configs after instantiation, remove a common source of confusion.

What is changing:

Specifically, here is how the signature of quantize_'s second argument will change:

#
# torchao v0.8.0 and before
#
def quantize(
    model: torch.nn.Module,
    apply_tensor_subclass: Callable[[torch.nn.Module], torch.nn.Module],
    ...,
): ...

#
# torchao v0.9.0
#
def quantize(
    model: torch.nn.Module,
    config: Union[AOBaseConfig, Callable[[torch.nn.Module], torch.nn.Module]],
    ...,
): ...

#
# torchao v0.10.0 or later (exact version TBD)
#
def quantize(
    model: torch.nn.Module,
    config: AOBaseConfig,
    ...,
): ...

1. the name of the second argument to quantize_ changed from apply_tensor_subclass to config. Since the vast majority of callsites today are passing in configuration with a positional argument, this change should not affect most people.
2. the type of the second argument to quantize_ will change from Callable[[torch.nn.Module], torch.nn.Module] to config: AOBaseConfig, following a deprecation process detailed below.
3. for individual workflows, the user facing API name changed from snake case (int8_weight_only) to camel case (Int8WeightOnlyConfig). All argument names for each config are kept as-is. We will keep the old snake case names (int8_weight_only) around and alias them to the new names (int8_weight_only = Int8WeightOnlyConfig), to avoid breaking callsites. We plan to keep the old names forever. Here are all the workflow config name changes:

old name (will keep working)	new name (recommended)
int4_weight_only	Int4WeightOnlyConfig
float8_dynamic_activation_float8_weight	Float8DynamicQuantizationFloat8WeightConfig
float8_static_activation_float8_weight	Float8StaticActivationFloat8WeightConfig
float8_weight_only	Float8WeightOnlyConfig
fpx_weight_only	FPXWeightOnlyConfig
gemlite_uintx_weight_only	GemliteUIntXWeightOnlyConfig
int4_dynamic_activation_int4_weight	Int4DynamicActivationInt4WeightConfig
int8_dynamic_activation_int4_weight	Int8DynamicActivationInt4WeightConfig
int8_dynamic_activation_int8_semi_sparse_weight	n/a (deprecated)
int8_dynamic_activation_int8_weight	Int8DynamicActivationInt8WeightConfig
int8_weight_only	Int8WeightOnlyConfig
uintx_weight_only	UIntXWeightOnlyConfig

Configuration for prototype workflows using quantize_ will be migrated at a later time. sparsify_ will be migrated in a similar fashion at a later time.

How these changes can affect you:

1. If you are a user of existing quantize_ API workflows and are passing in config by a positional argument (quantize_(model, int8_weight_only(group_size=128))), you are not affected. This syntax will keep working going forward. You have the option to migrate your callsite to the new config name (quantize_(model, Int8WeightOnlyConfig(group_size=128)) at your own pace.
2. If you are a user of existing quantize_ API workflows and are passing in config by a keyword argument (quantize_(model, tensor_subclass_inserter=int8_weight_only(group_size=128))), your callsite will break. You will need to change your callsite to quantize_(model, config=int8_weight_only(group_size=128)). We don't expect many people to be in this bucket.
3. If you are a developer writing new workflows for the quantize_ API, you will need to use the new configuration system. Please see migration of quantize_ workflow configuration from callables to configs #1690 for details.
4. If you are a user of sparsify_, you are not affected for now and a similar change will happen in a future version of torchao.

This migration will be a two step process:

• in torchao v0.9.0, we will enable the new syntax while starting the deprecation process for the old syntax.
• in torchao v.0.10.0 or later, we will remove the old syntax

We will keep the old callable syntax supported by quantize_ for one release cycle, and delete it afterwards. We will keep the old names as aliases for new names going forward (example: int4_weight_only as an alias of Int4WeightOnlyConfig) to keep existing callsites working without changes.

impact on API users

If you are just using the torchao quantize_ API as specified in the README, this is not BC-breaking. For example, the following syntax will keep working.

quantize_(model, int8_weight_only())

Note that the type of the object created by int8_weight_only() will change from a Callable to a config. You have the option to migrate to the explicit config creation, as follows:

quantize_(model, Int8WeightOnlyConfig())

user facing API changes

signature of quantize_

#
# before
#
def quantize(
    model: torch.nn.Module,
    apply_tensor_subclass: Callable[[torch.nn.Module], torch.nn.Module],
    ...,
): ...

#
# after - intermediate state, support both old and new for one release
#
def quantize(
    model: torch.nn.Module,
    config: Union[AOBaseConfig, Callable[[torch.nn.Module], torch.nn.Module]],
    ...,
): ...

#
# after - long term state
#
def quantize(
    model: torch.nn.Module,
    config: AOBaseConfig,
    ...,
): ...

usage example

An example for int4_weight_only

#
# before
#
quantize_(m, int4_weight_only(group_size=32))

#
# after, with new user facing names
#
quantize_(m, Int4WeightOnlyConfig(group_size=32))

#
# AND, after, with BC names
#
quantize_(m, int4_weight_only(group_size=32))

developer facing changes

See the PR details for examples, but they can be summarized as:

#
# old
#

# quantize_ calls the instance of calling this function on each module of the model
def int4_weight_only(group_size: int, ...) -> Callable:

    def new_callable(weight: torch.Tensor):
        # configuration is captured here via local variables
        ...
        
    # return type is a Callable
    return _get_linear_subclass_inserter(new_callable)

#
# new
#

# config base class
class AOBaseConfig(abc.ABC):
    pass

# user facing configuration of a workflow
@dataclass
class Int4WeightOnlyConfig(AOBaseConfig):
    group_size: int = 128
    ...

# not user facing transform of a module according to a worfklow's configuration
@register_quantize_module_handler(Int4WeightOnlyConfig)
def _int4_weight_only_transform(
    module: torch.nn.Module, 
    config: Int4WeightOnlyConfig,
) -> torch.nn.Module:
    # map to AQT, not user facing
    ...

migration status

quantize_ non-prototype workflow configuration

• int4_weight_only - [bc-breaking] enable direct configuration in quantize_ #1595
• qat: intx_quantization_aware_training - [bc-breaking] enable direct configuration in quantize_ #1595
• qat: from_intx_quantization_aware_training - [bc-breaking] enable direct configuration in quantize_ #1595
• float8_dynamic_activation_float8_weight - config migration: float8* #1694
• float8_static_activation_float8_weight - config migration: float8* #1694
• float8_weight_only - config migration: float8* #1694
• fpx_weight_only - config migration: fpx, gemlite, uintx #1697
• gemlite_uintx_weight_only - config migration: fpx, gemlite, uintx #1697
• int4_dynamic_activation_int4_weight - config migration: int* #1696
• int8_dynamic_activation_int4_weight - config migration: int* #1696
• int8_dynamic_activation_int8_semi_sparse_weight - marked for deprecation, confirmed with @jcaip we can delete this
• int8_dynamic_activation_int8_weight - config migration: int* #1696
• int8_weight_only - config migration: int* #1696
• uintx_weight_only - config migration: fpx, gemlite, uintx #1697
• nf4 - migrate nf4 to configs #1857

quantize_ prototype workflow configuration

Grep for callsites:

grep -r "quantize_(" torchao/prototype

• smoothquant: config migration: smoothquant #1851
• sparsity/superblock: using already migrated configs
• autoround: add assertion error about config migration to prototype/autoround #1852 (skipped)
• awq: migrate prototype/awq to configs #1853
• quantization/mixed_precision/scripts/mp_quant_eval: migrates prototype/mixed_precision to configs #1854
• quantization/mixed_precision/scripts/naive_intNwo: migrates prototype/mixed_precision to configs #1854
• quantization/mixed_precision/scripts/utils: migrates prototype/mixed_precision to configs #1854
• quantized_training: migrate prototype/quantized_training to configs #1855
• parq - the quantize_ used here is a different function, so nothing to do
• codebook: migrate prototype codebook quant to configs #1858

experimental

• everything: Migrate to config for Int8DynamicActivationIntxWeightConfig #1836

sparsify_

• everything: migrate sparsify_ to configs #1856

tutorials (replace with new registration API)

• calibration_flow/static_quant.py - migrate static quant tutorials to direct configuration #1710
• calibration_flow/gptq_like.py - migrate static quant tutorials to direct configuration #1710
• calibration_flow/awq_like.py - migrate static quant tutorials to direct configuration #1710

replace docblocks and public facing descriptions with new names

• README.md - update torchao READMEs with new configuration APIs #1711
• QAT README.md - update torchao READMEs with new configuration APIs #1711
• quantization README.md - update torchao READMEs with new configuration APIs #1711

verify partner integrations still work

• HF callsite: https://github.com/huggingface/transformers/blob/1feebb5b4150882deabddd190a541f336f3be817/src/transformers/quantizers/quantizer_torchao.py#L199
• SGLANG callsite: https://github.com/sgl-project/sglang/blob/2f47d710ae9cb1bdbbe0fe2392a0634827d257b3/python/sglang/srt/layers/torchao_utils.py#L39
• Diffusers callsite: https://github.com/huggingface/diffusers/blob/7fb481f840b5d73982cafd1affe89f21a5c0b20b/src/diffusers/quantizers/torchao/torchao_quantizer.py#L234

confirmed two out of three here: vkuzo/pytorch_scripts#28

delete old path (one version after migration)

• enforce AOBaseConfig type in quantize_'s config argument #1861