Introduce INC 3.0 quantization API and port torch RTN into 3.0

[yiliu30]

Description

This PR includes the minimal sets of INC 3.0-related configs and algos to demonstrate its architecture and E2E pipeline. More tests, algos, and frameworks will be added by following PRs.

• INC 3.0 quantization API
• Port torch RTN to INC 3.0
• Several E2E UTs to demonstrate usage
• Add more args for quantize API (like calib_func, calib_func_args)
• Update the CI @chensuyue
• Docstring

Other TODOs by follow-up PRs.

• Support Config combination (RTNWeightOnlyConfig + GPTQWeightOnlyConfig)
• Validate the user config
• Port a LLM using RTN
• Document
• Other TODOs marked in the code

3.0 API Usage

Users can pass the quantization configs by a dict or instances of XxxAlgoConfig.
Below are demos for beginners, intermediate, and experts.

For beginners

from neural_compressor.torch import get_default_rtn_config, quantize
fp32_model = build_simple_torch_model()
qmodel = quantize(fp32_model, quant_config=get_default_rtn_config())

For intermediate

• Pass a dict as config

from neural_compressor.torch import quantize
quant_config = {
    "rtn_weight_only_quant": {
        "weight_dtype": "nf4",
        "weight_bits": 4,
        "weight_group_size": 32,
    },
}
fp32_model = build_simple_torch_model()
qmodel = quantize(fp32_model, quant_config)

• Pass instances of XxxAlgoConfig.

from neural_compressor.torch import RTNWeightQuantConfig, quantize
quant_config = RTNWeightQuantConfig(weight_bits=4, weight_dtype="nf4", weight_group_size=32)
fp32_model = build_simple_torch_model()
qmodel = quantize(fp32_model, quant_config)

For experts

• Pass a dict as config

from neural_compressor.torch import quantize
fp32_model = build_simple_torch_model()
quant_config = {
    "rtn_weight_only_quant": {
        "global": {
            "weight_dtype": "nf4",
            "weight_bits": 4,
            "weight_group_size": 32,
        },
        "operator_name": {
            "fc1": {
                "weight_dtype": "int8",
                "weight_bits": 4,
            }
        },
    }
}
qmodel = quantize(fp32_model, quant_config)

• Pass instances of XxxAlgoConfig.

from neural_compressor.torch import RTNWeightQuantConfig, quantize
quant_config = RTNWeightQuantConfig(weight_bits=4, weight_dtype="nf4")
# set operator instance
fc1_config = RTNWeightQuantConfig(weight_bits=4, weight_dtype="int8")
quant_config.set_operator_name("model.fc1", fc1_config)
# get model and quantize
fp32_model = build_simple_torch_model()
qmodel = quantize(fp32_model, quant_config)

The structure of INC 3.0

neural_compressor
  - common                 #  the common methods/configs across frameworks
    - base_config.py            # the base config for all algorithms
    - tunner(strategy)     # strategy-related code
  - torch
    - quantization
      - config.py             # the torch-specific config for its algorithms 
    - algorithms           # torch algo-related code
  - onnxrt                 # onnxrt-related code
  - tensorflow             # tensorflow-related code

E2E example

# copy the below code to try it :)
import torch
class Model(torch.nn.Module):
    def __init__(self):
        super(Model, self).__init__()
        self.fc1 = torch.nn.Linear(30, 50)
        self.fc2 = torch.nn.Linear(50, 30)
        self.fc3 = torch.nn.Linear(30, 5)

    def forward(self, x):
        out = self.fc1(x)
        out = self.fc2(out)
        out = self.fc3(out)
        return out

fp32_model = Model()

from neural_compressor.torch import quantize
quant_config = {
    "rtn_weight_only_quant": {
        "weight_dtype": "nf4",
        "weight_bits": 4,
        "weight_group_size": 32,
    },
}
qmodel = quantize(fp32_model, quant_config)

How has this PR been tested?

Pre-CI

Dependency Change?

• torch (add torch as a mandatory dependency package for neural-compressor[pt])

[ftian1]

looks good to me