Add save/load support for HQQ (#1913)

Signed-off-by: yiliu30 <[email protected]>
Co-authored-by: chen, suyue <[email protected]>

neural_compressor/torch/algorithms/weight_only/hqq/core.py

@@ -19,7 +19,7 @@
 # NOTICE: the original `Quantizer` has been modified to `HQQTensorHandle`
 # and `QTensor` to decouple the data structure and the quantization logic.
 
-from typing import Any, Dict, Tuple
+from typing import Any, Dict, Mapping, Tuple
 
 import torch
 
@@ -278,3 +278,61 @@ def from_float(
         # !!! Delete the float explicitly to save memory
         del float_module
         return new_mod
+
+    def state_dict(self, *args, **kwargs):  # nn.Module override compatible
+        state_dict = self.q_weight.to_state_dict()
+        if self.bias is not None:
+            state_dict["bias"] = self.bias
+        if "destination" in kwargs and "prefix" in kwargs:
+            for key, value in state_dict.items():
+                kwargs["destination"][kwargs["prefix"] + key] = value
+        return state_dict
+
+    def _load_from_state_dict(
+        self,
+        state_dict,
+        prefix,
+        local_metadata,
+        strict,
+        missing_keys,
+        unexpected_keys,
+        error_msgs,
+    ):
+        all_expected_keys = ["val", "scale_quantized", "zero_quantized", "meta_info"]
+        if self.bias is not None:
+            all_expected_keys.append("bias")
+
+        for key in all_expected_keys:
+            if prefix + key not in state_dict:
+                missing_keys.append(key)
+        if missing_keys:
+            return  # Can't load weights if either weight or meta is missing
+
+        cur_state_dict = {}
+        for key in all_expected_keys:
+            cur_state_dict[key] = state_dict.pop(prefix + key)
+
+        unexpected_keys += state_dict.keys()
+        self._assign_state_dict(cur_state_dict, strict)
+
+    def _assign_state_dict(self, state_dict: Mapping[str, Any], strict: bool = True, assign: bool = False):
+        _scale_quantized = state_dict["scale_quantized"]
+        _zero_quantized = state_dict["zero_quantized"]
+        scale_state = state_dict["meta_info"]["scale"]
+        zero_state = state_dict["meta_info"]["zero"]
+        if _scale_quantized:
+            scale = HQQTensorHandle._create_q_tensor(scale_state["val"], scale_state["meta_info"])
+        else:
+            scale = state_dict["meta_info"]["scale"]
+        if _zero_quantized:
+            zero = HQQTensorHandle._create_q_tensor(zero_state["val"], zero_state["meta_info"])
+        else:
+            zero = state_dict["meta_info"]["zero"]
+        meta = state_dict["meta_info"]
+        meta["scale"] = scale
+        meta["zero"] = zero
+        self.q_weight = HQQTensorHandle._create_q_tensor(state_dict["val"], meta)
+        if self.bias is not None:
+            self.bias = state_dict["bias"]
+        self.quantized = True
+        return self

neural_compressor/torch/algorithms/weight_only/hqq/qtensor.py

@@ -115,3 +115,19 @@ def half(self):
         if self.zero is not None:
             self.zero = self.zero.half()
         return self
+
+    def to_state_dict(self):
+        state = {}
+        state["val"] = self.val
+        state["meta_info"] = self.meta_info.to_dict()
+        state["scale_quantized"] = self.is_scale_quantized()
+        state["zero_quantized"] = self.is_zero_quantized()
+        if self.is_scale_quantized():
+            state["meta_info"]["scale"] = self.scale.to_state_dict()
+        else:
+            state["meta_info"]["scale"] = self.scale
+        if self.is_zero_quantized():
+            state["meta_info"]["zero"] = self.zero.to_state_dict()
+        else:
+            state["meta_info"]["zero"] = self.zero
+        return state

neural_compressor/torch/algorithms/weight_only/save_load.py

@@ -124,7 +124,6 @@ def load_inc_format_woq_model(self, qmodel_weight_file_path, qconfig_file_path):
 
         with open(qconfig_file_path, "r") as file:
             self.quantization_config = json.load(file)
-
         model = self._build_woq_model()
         model.load_state_dict(qweights, assign=True)
         model.eval()
@@ -157,8 +156,19 @@ def load_hf_format_woq_model(self):
 
         return model
 
+    def _is_hqq_model(self):
+        for name, module in self.original_model.named_modules():
+            pattern = rf"(\(.*{re.escape(name)}.*{re.escape(type(module).__name__)}.*\))"
+            for q_config_key, q_config_value in self.quantization_config.items():
+                if re.search(pattern, q_config_key):
+                    if isinstance(q_config_value, dict) and [algo for algo in q_config_value.keys()][0] == "hqq":
+                        return True
+
     def _build_woq_model(self):
         """Build weight-only quantization model."""
+        if self._is_hqq_model():
+            return self._build_hqq_model()
+
         from neural_compressor.torch.utils import set_module
 
         from .modules import MulLinear
@@ -228,6 +238,23 @@ def _build_woq_model(self):
         woq_model = self.original_model
         return woq_model
 
+    def _build_hqq_model(self):
+        """Replace quantized Linear with HQQLinear."""
+        from neural_compressor.torch.algorithms.weight_only.hqq.core import HQQLinear
+        from neural_compressor.torch.utils import set_module
+
+        for name, module in self.original_model.named_modules():
+            if isinstance(module, torch.nn.Linear):
+                loaded_state_dict_keys_set = set(self.loaded_state_dict_keys)
+                if name + ".val" not in loaded_state_dict_keys_set:
+                    continue
+                new_module = HQQLinear(
+                    in_features=module.in_features, out_features=module.out_features, bias=module.bias is not None
+                )
+                set_module(self.original_model, name, new_module)
+        woq_model = self.original_model
+        return woq_model
+
     def _get_model_class_and_config(self):
         from transformers import AutoConfig, AutoModelForCausalLM
         from transformers.dynamic_module_utils import get_class_from_dynamic_module, resolve_trust_remote_code

neural_compressor/torch/quantization/algorithm_entry.py

@@ -517,11 +517,14 @@ def hqq_entry(
     **kwargs,
 ) -> torch.nn.Module:
     from neural_compressor.torch.algorithms.weight_only.hqq import HQQuantizer
+    from neural_compressor.torch.algorithms.weight_only.save_load import save
 
     logger.info("Quantize model with the HQQ algorithm.")
 
     quantizer = get_quantizer(model, quantizer_cls=HQQuantizer, quant_config=configs_mapping)
     model = quantizer.execute(model, mode=mode)
+    model.qconfig = configs_mapping
+    model.save = MethodType(save, model)
     postprocess_model(model, mode, quantizer)
     dump_model_op_stats(mode, configs_mapping)
 

neural_compressor/torch/quantization/load_entry.py

@@ -22,6 +22,7 @@
     AWQConfig,
     FP8Config,
     GPTQConfig,
+    HQQConfig,
     RTNConfig,
     TEQConfig,
 )
@@ -89,7 +90,9 @@ def load(model_name_or_path, original_model=None, format="default", device="cpu"
             # select load function
             config_object = config_mapping[next(iter(config_mapping))]
 
-            if isinstance(config_object, (RTNConfig, GPTQConfig, AWQConfig, TEQConfig, AutoRoundConfig)):  # WOQ
+            if isinstance(
+                config_object, (RTNConfig, GPTQConfig, AWQConfig, TEQConfig, AutoRoundConfig, HQQConfig)
+            ):  # WOQ
                 from neural_compressor.torch.algorithms import weight_only
 
                 return weight_only.load(model_name_or_path, original_model, format=LoadFormat.DEFAULT)

test/3x/torch/quantization/weight_only/test_hqq.py

@@ -1,11 +1,14 @@
+import copy
 import os
+import time
 from copy import deepcopy
 
 import pytest
 import torch
 import transformers
 from transformers import AutoModelForCausalLM
 
+from neural_compressor.common import options
 from neural_compressor.common.utils import logger
 from neural_compressor.torch.algorithms.weight_only.hqq.config import HQQModuleConfig, QTensorConfig, hqq_global_option
 from neural_compressor.torch.algorithms.weight_only.hqq.core import HQQLinear
@@ -93,6 +96,27 @@ def test_hqq_quant(self, force_use_cpu, force_not_half):
             q_label_1.eq(q_label_2)
         ), "The results of calling `convert` + `prepare` and calling `quantize` should be equal."
 
+    def test_hqq_load_save(self, force_use_cpu, force_not_half):
+
+        hqq_global_option.use_half = False
+        fp32_model = AutoModelForCausalLM.from_pretrained("trl-internal-testing/tiny-random-OPTForCausalLM")
+        example_inputs = torch.tensor([[10, 20, 30, 40, 50, 60]], dtype=torch.long, device="cpu")
+        # test_default_config
+        quant_config = get_default_hqq_config()
+
+        # prepare + convert API
+        model = prepare(deepcopy(fp32_model), quant_config)
+        qmodel = convert(model)
+        qmodel_out_ref = model(example_inputs)[0]
+        save_path = options.workspace + f"/_hqq_model_{time.time()}.pth"
+        qmodel.save(save_path)
+        from neural_compressor.torch.quantization import load
+
+        # loading compressed model
+        loaded_model = load(save_path, copy.deepcopy(fp32_model))
+        loaded_model_out = loaded_model(example_inputs)[0]
+        assert torch.allclose(qmodel_out_ref, loaded_model_out), "Unexpected result. Please double check."
+
     def test_hqq_fallback(self, force_use_cpu, force_not_half):
 
         class ToyModel(torch.nn.Module):
@@ -181,3 +205,57 @@ def test_hqq_module(
             scale_quant_group_size=scale_quant_group_size,
             device=torch.device(device_name),
         )
+
+    @pytest.mark.parametrize(
+        "nbits, group_size, quant_zero, quant_scale, scale_quant_group_size",
+        [
+            (4, 64, True, False, 128),
+            (4, 64, False, False, 128),
+            (4, 64, True, True, 128),
+            (4, 64, False, True, 128),
+            (8, 64, True, False, 128),
+        ],
+    )
+    def test_hqq_linear_save_and_load(
+        self,
+        nbits,
+        group_size,
+        quant_zero,
+        quant_scale,
+        scale_quant_group_size,
+    ):
+        hqq_global_option.use_half = False
+        # Parse config
+        weight_qconfig = QTensorConfig(
+            nbits=nbits,
+            channel_wise=True,
+            group_size=group_size,
+            optimize=True,
+            round_zero=True if nbits == 4 else False,
+        )
+        zero_qconfig = None
+        if quant_zero:
+            zero_qconfig = QTensorConfig(nbits=8, channel_wise=False, group_size=None, optimize=False)
+        scale_qconfig = None
+        if quant_scale:
+            scale_qconfig = QTensorConfig(nbits=8, channel_wise=True, group_size=scale_quant_group_size, optimize=False)
+        hqq_quant_config = HQQModuleConfig(weight=weight_qconfig, scale=scale_qconfig, zero=zero_qconfig)
+        # Create HQQ Linear
+        bs = 4
+        in_features = 64
+        out_features = 128
+        float_linear = torch.nn.Linear(in_features=in_features, out_features=out_features)
+        float_linear.to(device)
+        float_linear_copy = deepcopy(float_linear)
+        input = torch.randn(bs, in_features, device=device)
+        hqq_linear = HQQLinear.from_float(float_linear_copy, quant_config=hqq_quant_config)
+        out_ref = hqq_linear(input)
+        state_dict = hqq_linear.state_dict()
+        hqq_module_path = options.workspace + f"/_hqq_linear_{time.time()}.pth"
+        torch.save(state_dict, hqq_module_path)
+        reload_state_dict = torch.load(hqq_module_path)
+        new_float = torch.nn.Linear(in_features=in_features, out_features=out_features)
+        new_hqq_linear = HQQLinear.from_float(new_float, quant_config=hqq_quant_config)
+        new_hqq_linear.load_state_dict(reload_state_dict)
+        out = new_hqq_linear(input)
+        assert torch.equal(out_ref, out), f"out_ref: {out_ref}, out: {out}"