Add WeightOnlyPTQ and GPTQ (PaddlePaddle#6572)

* Add WeightOnlyPTQ and GPTQ

* update

* update

llm/causallm/argument.py

@@ -52,6 +52,8 @@ class QuantArgument:
     # PTQ related parameters
     do_ptq: bool = field(default=False, metadata={"help": "Whether to use PTQ"})
     ptq_step: int = field(default=8, metadata={"help": "Step for PTQ"})
+    ptq_weight_only: bool = field(default=False, metadata={"help": "Whether to use PTQ weight only"})
+    quant_bits: int = field(default=8, metadata={"help": "Quantization bit size"})
 
     fused_qkv: bool = field(default=False, metadata={"help": "Whether to use Fused Quantized QKV"})
     parallel_ffn: bool = field(default=False, metadata={"help": "Whether to use Parallel FFN"})

llm/causallm/finetune_generation.py

@@ -48,9 +48,9 @@ def main():
     training_args.print_config(quant_args, "Quant")
     training_args.print_config(gen_args, "Generation")
 
-    if sum([quant_args.do_ptq, quant_args.do_qat, training_args.do_train]) > 1:
+    if sum([quant_args.do_ptq, quant_args.do_qat, quant_args.do_gptq, training_args.do_train]) > 1:
         raise ValueError(
-            "--do_train, --do_ptq and --do_qat cannot work at the same time. Please choose only one at a time"
+            "--do_train, --do_ptq, --do_gptq and --do_qat cannot work at the same time. Please choose only one at a time"
         )
 
     # Setup GPU & distributed training
@@ -257,6 +257,15 @@ def compute_metrics_do_generation(eval_preds):
 
         apply_ptq(quant_args, trainer, ptq_dataloader)
 
+    if quant_args.do_gptq:
+        if isinstance(model, LoRAModel):
+            raise NotImplementedError(
+                "PTQ strategy not supported for LoRA model. Please merge lora parameters to pretrain model first."
+            )
+        from quant import apply_gptq
+
+        apply_gptq(quant_args, trainer, ptq_dataloader)
+
     # Evaluation dev set
     if training_args.do_eval:
         eval_result = trainer.evaluate(dev_ds)

llm/causallm/quant.py

@@ -20,17 +20,19 @@
 from paddle.quantization import PTQ, QAT, QuantConfig
 from paddle.quantization.quanters.abs_max import FakeQuanterWithAbsMaxObserverLayer
 from paddleslim.quant.advanced import (
+    GPTQ,
     EMASampler,
     MultiStepSampler,
     PieceWiseSearch,
     Shift,
     Smooth,
 )
+from paddleslim.quant.advanced.utils import find_parent_layer_and_sub_name
 from paddleslim.quant.layers import (
     QuantizedColumnParallelLinear,
     QuantizedRowParallelLinear,
 )
-from paddleslim.quant.observers import AbsMaxChannelWiseWeightObserver, AbsmaxObserver
+from paddleslim.quant.observers import AbsMaxChannelWiseWeightObserver, AVGObserver
 from paddleslim.quant.quanters import PACTQuanter
 
 from paddlenlp.peft import PrefixModelForCausalLM
@@ -93,7 +95,7 @@ def apply_smooth(quant_args, trainer, ptq_dataloader, ptq_model_config):
             search_scale_min=1.0,
             search_scale_max=5.0,
             weight_quant_method="abs_max_channel_wise",
-            act_quant_method="abs_max",
+            act_quant_method="avg",
         )
     else:
         search_func = None
@@ -117,8 +119,8 @@ def apply_smooth(quant_args, trainer, ptq_dataloader, ptq_model_config):
 
 def apply_ptq(quant_args, trainer, ptq_dataloader):
     q_config = QuantConfig(activation=None, weight=None)
-    act_quanter = AbsmaxObserver()
-    weight_quanter = AbsMaxChannelWiseWeightObserver()
+    act_quanter = AVGObserver() if not quant_args.ptq_weight_only else None
+    weight_quanter = AbsMaxChannelWiseWeightObserver(quant_bits=quant_args.quant_bits)
     q_config.add_qat_layer_mapping(ColumnParallelLinear, QuantizedColumnParallelLinear)
     q_config.add_qat_layer_mapping(RowParallelLinear, QuantizedRowParallelLinear)
     q_config.add_type_config(
@@ -137,6 +139,26 @@ def apply_ptq(quant_args, trainer, ptq_dataloader):
     trainer.model = ptq.convert(trainer.model, inplace=True)
 
 
+def apply_gptq(quant_args, trainer, ptq_dataloader):
+    num_layer = 0
+    model = trainer.model
+    for cur_name, cur_layer in model.named_sublayers():
+        if type(cur_layer) in [paddle.nn.Linear, ColumnParallelLinear, RowParallelLinear]:
+            num_layer += 1
+            print("GPTQ layer", num_layer, cur_name)
+            parent_layer, sub_name = find_parent_layer_and_sub_name(model, cur_name)
+            cur_quant_layer = GPTQ(cur_layer)
+            setattr(parent_layer, sub_name, cur_quant_layer)
+            trainer.ptq_loop(
+                ptq_dataloader,
+                description="PTQ",
+                max_eval_iters=quant_args.ptq_step,
+            )
+            cur_quant_layer.fasterquant(percdamp=0.1, groupsize=-1, actorder=True)
+            del cur_quant_layer
+            setattr(parent_layer, sub_name, cur_layer)
+
+
 def get_ptq_model_config(model):
     if isinstance(model, PrefixModelForCausalLM):
         base_model_prefix = model.model.base_model_prefix