mppq 0.7.4

PPQ is an offline quantization tools

Minimal PPQ (mPPQ) PPQ 裁剪版

PPQ Readme

This repo will continue the development of PPQ, maintain features and fix bugs. I pruned most deprecated APIs and sample codes, functions in original PPQ and make it more extensible and easier to use.

Key Changes

1. TargetPlatform

TargetPlatform in PPQ mixed the concept of Precision and Device. We can see TargetPlatform.OPENVINO_INT8 comes with TargetPlatform.INT8.

In mPPQ, I separate the concept of Precision and Device and all calls to TargetPlatform are now changed to TargetPrecision.

I also removed platforms support in PPQ because they are lacking of correct maintenance and support. Instead of that, I add a new api to register your customized platform as well as your frontend parser, dispatcher and quantizer.

2. Extension

Users need to register their own platforms to perform a correct quantization. There is no pre-defined platforms in mPPQ, just a sample quantizer for example.

Let me explain the way users need to follow.

1. Parser and Exporter

I keep onnx_parser and onnxruntime_exporter as the default graph serialization and de-serialization methods. Users can create a new parser by inheriting GraphBuilder and a new exporter by inheriting GraphExporter.

class MyParser(GraphBuilder):
    def build(self, model_object: Any, **kwargs) -> BaseGraph:
        """Parser offers the way how to read from a model object and turn it into PPQ BaseGraph.
        """
        ...

class MyExporter(GraphExporter):
    def export(
        self,
        file_path: str | os.PathLike,
        graph: BaseGraph,
        config_path: Optional[str] = None,
        **kwargs,
    ):
        """Exporter offers the way how to serialize a PPQ BaseGraph into a model object.
        ...

The model object can be onnx or others like openvino .xml or qnn .dlc.

2. Dispatcher

Dispatcher is the core concept from PPQ, it will analysis the whole graph and decide which op should be quantized in which precision.

In mPPQ, all pre-defined dispatchers in PPQ are still kept, and they are:

• allin
• aggressive
• conservative
• pointwise
• perseus

Users can create a new dispatcher by inheriting GraphDispatcher and implement its own logic.

3. Quantizer

Quantizer is a fundamental component in PPQ, it will control all the operations that needed to quantize a model.

In mPPQ quantizer, it will offer 3 abstract methods for users to implement:

@abstractmethod
def init_quantize_config(self, operation: Operation) -> OperationQuantizationConfig:
    r"""Return a query to the operation how it should be quantized."""
    raise NotImplementedError

@property
def default_prequant_pipeline(self) -> QuantizationOptimizationPipeline:
    r"""A simplified API to return a default quantization pipeline."""
    return QuantizationOptimizationPipeline([])

@property
def default_quant_pipeline(self) -> QuantizationOptimizationPipeline:
    r"""A simplified API to return a default quantization pipeline."""
    raise NotImplementedError

Quantizer doesn't work with dispatcher any more, now user's dispatcher (or builtin ones) must dispatch all operations in the graph to a specific precision, if one of the operation is Precision.UNSPECIFIED, quantizer will raise an error now.

4. Platform

In mPPQ, I provide a register to support add new parsers, exporters, dispatchers and quantizers to a specific platform from external codebase.

from mppq.api import load_quantizer, register_platform, export_ppq_graph

MyPlatformID = 1

register_platform(
    MyPlatformID,
    dispatcher={"mydisp": MyDispatcher},
    quantizer={"myquant": MyQuantizer},
    parsers={"myparser": MyParser},
    exporters={"myexporter": MyExporter},
)

quantizer = load_quantizer("mymodel.onnx", MyPlatformID)
quantized_graph = quantizer.quantize()
export_ppq_graph(quantized_graph, "quantized_mymodel.onnx")

Users can use builtin dispatcher by specifying a name in load_quantizer, but I highly recommend to know very detail of your platform and design your own dispatcher and quantizer.

register_platform(
    MyPlatformID,
    dispatcher={},  # need to specify a builtin dispatcher name in quantize api
    quantizer={"myquant": MyQuantizer},
    parsers={},  # use builtin onnx parser
    exporters={},  # use builtin onnx exporter
)

quantizer = load_quantizer("mymodel.onnx", MyPlatformID, dispatcher="allin")

5. Operation

In mPPQ, most ONNX operators up to opset 19 are supported. In order to add a new support from users code base, users can register a new operation to a specific platform.

from mppq.api import register_operation

@register_operation("myop", MyPlatformID)
def myop_forward(
    op: Operation,
    values: Sequence[torch.Tensor],
    ctx: Optional[TorchBackendContext] = None,
    **kwargs,
) -> torch.Tensor | Tuple[torch.Tensor, ...]:
    """
    Args:
        op: operation information (precision, attributes, types, etc.)
        values: input tensors
        ctx: execution device information

    Returns:
        a result tensor or a tuple of result tensors
    """
    ...

Contribution

All mPPQ python codes are clean with flake8, black, and pyright (except for mppq/executor/op/default.py).

Test coverage: 51% (v0.7.1)

Acknowledgement

PPQ