feat: add QAT pipline

[ali-88123]

This PR introduces Quantization Aware Training (QAT) as a new compression method in AngelSlim. QAT enables models to learn quantization parameters during training by inserting fake-quantization operations into the forward pass, resulting in significantly better accuracy under low-bit quantization scenarios (e.g., W4A8, INT4) compared to post-training quantization (PTQ) alone.

Training Modes

1. End-to-End (end2end): Uses HuggingFace Seq2SeqTrainer with a custom AdamW optimizer targeting only quantization parameters (scale, zero_point).

2. Blockwise (blockwise): Trains each Transformer block independently using MSE loss between full-precision and quantized block outputs.

Model Conversion & Save

• convert(): Replaces all QuantLinear modules with inference-ready QDQModule (from the existing PTQ module), extracting learned weight_scale and input_scale from trained quantizers.

• save(): Supports two formats:

  • "fake": Saves raw state_dict via torch.save() — useful for checkpoint resumption.

  • "real": Delegates to the model-specific save function (e.g., vLLM/TRT-LLM compatible formats).

Key configuration sections:

• training.plugin_config: Plugin toggles (enable_scale, enable_rotation), per-plugin quant_config overrides

• training.hf_args (end2end): Full HuggingFace Seq2SeqTrainingArguments

• training.block_wise_config (blockwise): epochs, batch_size, quant_lr, weight_lr, min_lr_factor

Documentation

A comprehensive user guide has been added at docs/source/features/quantization/qat.md.

[yghstill]

mv to data folder

[yghstill]

-> quanter.py

[yghstill]

name: Union[str, List[str]]
QAT: Optional[QATTrainingConfig] = None

QATTrainingConfig里放QAT专用超参
其他共用