Add autotune support for PT2E (#2110)

Add autotune support for PT2E and disable some conv1d-related test on HPU
---------

Signed-off-by: yiliu30 <[email protected]>
Co-authored-by: Xin He <[email protected]>

Author: yiliu30

neural_compressor/common/base_config.py

@@ -18,6 +18,7 @@
 
 from __future__ import annotations
 
+import copy
 import inspect
 import json
 import os
@@ -539,6 +540,7 @@ def expand(self) -> List[BaseConfig]:
                 tuning_param_pair = dict(zip(tuning_param_name_lst, params_values))
                 tmp_params_dict = {**not_tuning_param_pair, **tuning_param_pair}
                 new_config = self.__class__(**tmp_params_dict)
+                new_config.local_config = copy.deepcopy(self.local_config)
                 logger.info(new_config.to_dict())
                 config_list.append(new_config)
         logger.info("Expanded the %s and got %d configs.", self.__class__.name, len(config_list))
@@ -629,9 +631,13 @@ def __eq__(self, other: BaseConfig) -> bool:
         """
         if not isinstance(other, type(self)):
             return False
-        return self.params_list == other.params_list and all(
+
+        params_equal = self.params_list == other.params_list and all(
             getattr(self, str(attr)) == getattr(other, str(attr)) for attr in self.params_list
         )
+        local_config_equal = self.local_config == other.local_config
+        global_config_equal = self.global_config == other.global_config
+        return params_equal and local_config_equal and global_config_equal
 
 
 class ComposableConfig(BaseConfig):

neural_compressor/torch/algorithms/pt2e_quant/core.py

@@ -18,6 +18,7 @@
 
 from typing import Any
 
+import torch
 import torch.ao.quantization.quantizer.x86_inductor_quantizer as xiq
 from torch.ao.quantization.quantize_pt2e import convert_pt2e, prepare_pt2e
 from torch.ao.quantization.quantizer.x86_inductor_quantizer import X86InductorQuantizer
@@ -102,4 +103,5 @@ def half_precision_transformation(self, model, config):
         """
         half_precision_node_set = hp_rewriter.get_half_precision_node_set(model, config)
         logger.info("Try to convert %d nodes to half precision.", len(half_precision_node_set))
-        hp_rewriter.transformation(model, half_precision_node_set)
+        hp_rewriter.transformation(model, half_precision_node_set, torch.float16)
+        hp_rewriter.transformation(model, half_precision_node_set, torch.bfloat16)

neural_compressor/torch/algorithms/pt2e_quant/half_precision_rewriter.py

@@ -13,6 +13,7 @@
 # limitations under the License.
 """Rewrite the FP32 operators to FP16 or BF16 operators."""
 
+from collections import defaultdict
 from dataclasses import dataclass
 from functools import partial
 from typing import Any, Callable, Dict, List, Tuple
@@ -25,7 +26,7 @@
 from torch.fx.subgraph_rewriter import Match
 from typing_extensions import TypeAlias
 
-from neural_compressor.common import utils
+from neural_compressor.common import logger, utils
 
 # =============================================================================
 # Search and replace patterns
@@ -50,25 +51,44 @@ class PatternPair:
 
 # key: torch func
 # value: the tuple of args
-FuncArgsMappingType: TypeAlias = Dict[TorchFuncType, Tuple[torch.Tensor, ...]]
+FuncArgsMappingType: TypeAlias = Dict[TorchFuncType, List[Tuple[torch.Tensor, ...]]]
 
 
-# Align with https://pytorch.org/docs/stable/amp.html#cpu-ops-that-can-autocast-to-bfloat16
-# TODO: complete the mapping
+# Align with xiq, as it relay on xiq's set_module_xx capability
 FN_ARGS_MAPPING: FuncArgsMappingType = {
-    torch.nn.functional.linear: (torch.randn(0, 0), torch.randn(0, 0)),  # linear w/o bias
-    torch.nn.functional.linear: (torch.randn(0, 0), torch.randn(0, 0), torch.randn(0)),  # linear w/ bias
+    # Note: ORDER is matter
+    torch.nn.functional.linear: [
+        (torch.randn(0, 0), torch.randn(0, 0)),  # linear w/o bias
+        (torch.randn(0, 0), torch.randn(0, 0), torch.randn(0)),  # linear w/ bias
+    ],
+    torch.nn.functional.conv2d: [
+        (torch.randn(1, 1, 1, 1), torch.randn(1, 1, 1, 1)),  # conv2d w/o bias
+        (torch.randn(1, 1, 1, 1), torch.randn(1, 1, 1, 1), torch.randn(1)),  # conv2d w/ bias
+    ],
+    torch.matmul: [
+        (torch.randn(0, 0), torch.randn(0, 0)),
+        (torch.randn(0, 0, 0), torch.randn(0, 0, 0)),
+        (torch.randn(0, 0, 0, 0), torch.randn(0, 0, 0, 0)),
+    ],
 }
-# TODO: complete the mapping
-FN_ATEN_OPS_MAPPING = {
-    torch.nn.functional.linear: torch.ops.aten.linear.default,
+
+# module cls <-> function name
+NN_MODULES_TO_NN_FN = {
+    torch.nn.Linear: torch.nn.functional.linear,
+    torch.nn.Conv2d: torch.nn.functional.conv2d,
 }
 
+# Use the mapping from xiq
+FN_ATEN_OPS_MAPPING = xiq._map_module_function_to_aten_operator_type()
+
 SUPPORTED_OPERATORS = FN_ATEN_OPS_MAPPING.values()
 
 
 PatternRegistryType: TypeAlias = Dict[TorchFuncType, PatternPair]
-HALF_PRECISION_PATTERN_REGISTRY: Dict[torch.dtype, PatternRegistryType] = {torch.float16: {}, torch.bfloat16: {}}
+HALF_PRECISION_PATTERN_REGISTRY: Dict[torch.dtype, PatternRegistryType] = {
+    torch.float16: defaultdict(list),
+    torch.bfloat16: defaultdict(list),
+}
 
 # FP16_PATTERN_REGISTRY: PatternRegistryType = HALF_PRECISION_PATTERN_REGISTRY[torch.float16]
 # BF16_PATTERN_REGISTRY: PatternRegistryType = HALF_PRECISION_PATTERN_REGISTRY[torch.bfloat16]
@@ -98,15 +118,18 @@ def replace_fn_wrapper(fn_args, fn):
 
 
 def _register_pattern_pair(dtype: torch.dtype) -> None:
-    for fn, fn_args in FN_ARGS_MAPPING.items():
-        pattern_pair = pattern_factory(fn, fn_args)
-        HALF_PRECISION_PATTERN_REGISTRY[dtype][fn] = pattern_pair
-    utils.logger.info(
+    for fn, fn_args_lst in FN_ARGS_MAPPING.items():
+        for fn_args in fn_args_lst:
+            logger.debug(f"Registering search and replace patterns for {fn} with args: {fn_args}.")
+            pattern_pair = pattern_factory(fn, fn_args)
+            HALF_PRECISION_PATTERN_REGISTRY[dtype][fn].append(pattern_pair)
+    utils.logger.debug(
         f"Registered {len(HALF_PRECISION_PATTERN_REGISTRY[dtype])} search and replace patterns for {dtype}."
     )
 
 
 _register_pattern_pair(torch.float16)
+_register_pattern_pair(torch.bfloat16)
 
 
 def get_filter_fn(node_list, fn):
@@ -182,9 +205,10 @@ def get_unquantized_node_set(gm: torch.fx.GraphModule):
 
 def transformation(gm: torch.fx.GraphModule, node_candidate_list: List[str], target_dtype: torch.dtype = torch.float16):
     """Convert the nodes in `node_candidate_list` to `target_dtype` if possible."""
-    for pattern_pair in HALF_PRECISION_PATTERN_REGISTRY[target_dtype].values():
-        apply_single_pattern_pair(gm, pattern_pair, node_candidate_list)
-    utils.logger.info("Half precision conversion is done:")
+    for pattern_pair_lst in HALF_PRECISION_PATTERN_REGISTRY[target_dtype].values():
+        for pattern_pair in pattern_pair_lst:
+            apply_single_pattern_pair(gm, pattern_pair, node_candidate_list)
+    utils.logger.info(f"Half precision conversion({target_dtype}) completed.")
     if utils.level_name == "DEBUG":  # pragma: no cover
         gm.print_readable(True)
 
@@ -201,11 +225,11 @@ def _parse_node_candidate_set_from_user_config(config, gm):
     op_name_filters = []
     for op_type_name, config in op_type_configs.items():  # pragma: no cover
         op_type = getattr(torch.nn, op_type_name)
-        if config.act_dtype == "fp16":  # pragma: no cover
+        if config.act_dtype in ["fp16", "bf16"]:  # pragma: no cover
             filter = xpq._get_module_type_filter(op_type)
             op_type_filters.append(filter)
     for op_name, config in op_name_configs.items():
-        if config.act_dtype == "fp16":  # pragma: no cover
+        if config.act_dtype in ["fp16", "bf16"]:  # pragma: no cover
             filter = xpq._get_module_name_filter(op_name)
             op_name_filters.append(filter)
     node_set_from_user_config = set()
@@ -237,5 +261,7 @@ def get_half_precision_node_set(gm, config):
     for node in possible_node_set:
         if node.target in SUPPORTED_OPERATORS:
             half_precision_node_set.add(node)
-    utils.logger.info(f"Found {len(half_precision_node_set)} nodes to convert to half precision.")
+    utils.logger.info(
+        f"Found {len(half_precision_node_set)} nodes to convert to half precision: {half_precision_node_set}"
+    )
     return half_precision_node_set

neural_compressor/torch/algorithms/pt2e_quant/utility.py

@@ -26,7 +26,7 @@
 from torch.ao.quantization.quantizer import QuantizationSpec
 from torch.ao.quantization.quantizer.x86_inductor_quantizer import QuantizationConfig, X86InductorQuantizer
 
-from neural_compressor.torch.utils import GT_OR_EQUAL_TORCH_VERSION_2_5
+from neural_compressor.torch.utils import GT_OR_EQUAL_TORCH_VERSION_2_5, logger
 
 
 def create_quant_spec_from_config(dtype, sym, granularity, algo, is_dynamic=False) -> QuantizationSpec:
@@ -79,6 +79,7 @@ def create_quant_spec_from_config(dtype, sym, granularity, algo, is_dynamic=Fals
 def _map_inc_config_to_torch_quant_config(inc_config, is_dynamic=False) -> QuantizationConfig:
     NOT_QUANT_DTYPES = ["fp32", "fp16", "bf16"]
     if inc_config.act_dtype in NOT_QUANT_DTYPES and inc_config.w_dtype in NOT_QUANT_DTYPES:  # pragma: no cover
+        logger.debug("Got non-quantizable data types, skipping quantization.")
         return None
     default_quant_config = xiq.get_default_x86_inductor_quantization_config(is_dynamic=is_dynamic)
     input_act_quant_spec = create_quant_spec_from_config(

neural_compressor/torch/quantization/autotune.py

@@ -54,6 +54,15 @@ def get_all_config_set() -> Union[BaseConfig, List[BaseConfig]]:
     return get_all_config_set_from_config_registry(fwk_name=FRAMEWORK_NAME)
 
 
+def _deepcopy_warp(model):
+    additional_attr_lst = ["_exported", "dynamic_shapes"]
+    original_attr = {key: getattr(model, key, None) for key in additional_attr_lst}
+    new_model = deepcopy(model)
+    for key, value in original_attr.items():
+        setattr(new_model, key, value)
+    return new_model
+
+
 @dump_elapsed_time("Pass auto-tune")
 def autotune(
     model: torch.nn.Module,
@@ -81,7 +90,7 @@ def autotune(
     best_quant_model = None
     eval_func_wrapper = EvaluationFuncWrapper(eval_fn, eval_args)
     config_loader, tuning_logger, tuning_monitor = init_tuning(tuning_config=tune_config)
-    baseline: float = eval_func_wrapper.evaluate(deepcopy(model))
+    baseline: float = eval_func_wrapper.evaluate(_deepcopy_warp(model))
     tuning_monitor.set_baseline(baseline)
     tuning_logger.tuning_start()
     for trial_index, quant_config in enumerate(config_loader, 1):
@@ -90,7 +99,7 @@ def autotune(
         logger.info(quant_config.to_dict())
         # !!! Make sure to use deepcopy only when inplace is set to `True`.
         q_model = quantize(
-            deepcopy(model),
+            _deepcopy_warp(model),
             quant_config=quant_config,
             run_fn=run_fn,
             run_args=run_args,
@@ -112,7 +121,7 @@ def autotune(
                 best_quant_config: BaseConfig = best_trial_record.quant_config
                 # !!! Make sure to use deepcopy only when inplace is set to `True`.
                 q_model = quantize(
-                    deepcopy(model),
+                    _deepcopy_warp(model),
                     quant_config=best_quant_config,
                     run_fn=run_fn,
                     run_args=run_args,

test/3x/torch/quantization/test_pt2e_quant.py

@@ -29,7 +29,6 @@ def _is_ipex_imported():
     monkeypatch.setattr("neural_compressor.torch.quantization.algorithm_entry.is_ipex_imported", _is_ipex_imported)
     monkeypatch.setattr("neural_compressor.torch.export.pt2e_export.is_ipex_imported", _is_ipex_imported)
 
-
 class TestPT2EQuantization:
     def teardown_class(self):
         shutil.rmtree("saved_results", ignore_errors=True)
@@ -53,15 +52,15 @@ def forward(self, a: torch.Tensor, b: torch.Tensor) -> torch.Tensor:
         return bar, example_inputs
 
     @staticmethod
-    def build_model_include_conv_and_linear():
+    def build_model_include_conv_and_linear(bias=True):
         class Model(torch.nn.Module):
-            def __init__(self):
+            def __init__(self, bias=True):
                 super(Model, self).__init__()
-                self.conv1 = torch.nn.Conv2d(3, 6, 5)
+                self.conv1 = torch.nn.Conv2d(3, 6, 5, bias=bias)
                 self.pool = torch.nn.MaxPool2d(2, 2)
-                self.conv2 = torch.nn.Conv2d(6, 16, 5)
-                self.fc1 = torch.nn.Linear(16 * 5 * 5, 120)
-                self.fc2 = torch.nn.Linear(120, 84)
+                self.conv2 = torch.nn.Conv2d(6, 16, 5, bias=bias)
+                self.fc1 = torch.nn.Linear(16 * 5 * 5, 120, bias=bias)
+                self.fc2 = torch.nn.Linear(120, 84, bias=bias)
 
             def forward(self, x):
                 x = self.conv1(x)
@@ -74,7 +73,7 @@ def forward(self, x):
 
                 return x
 
-        model = Model()
+        model = Model(bias)
         example_inputs = (torch.randn(1, 3, 32, 32),)
         return model, example_inputs
 
@@ -283,3 +282,62 @@ def test_mixed_fp16_and_int8(self, force_not_import_ipex):
         opt_model = torch.compile(converted_model)
         out = opt_model(*example_inputs)
         assert out is not None
+
+    @pytest.mark.skipif(not GT_OR_EQUAL_TORCH_VERSION_2_5, reason="Requires torch>=2.5")
+    @pytest.mark.parametrize("half_precision_dtype", ["fp16", "bf16"])
+    @pytest.mark.parametrize("op_name_or_type", ["conv1", "fc1", torch.nn.Linear, torch.nn.Conv2d])
+    @pytest.mark.parametrize("bias", [True, False])
+    def test_auto_tune_mixed_int8_and_16bits(self, half_precision_dtype, op_name_or_type, bias, force_not_import_ipex):
+        # Just make sure the pattern matches, not the accuracy.
+        # config1: int8 for all
+        # config2: half precision for linear/conv
+        from neural_compressor.torch.quantization.config import INT8StaticQuantConfig
+        from neural_compressor.torch.quantization.autotune import autotune, TuningConfig
+
+        config1 = INT8StaticQuantConfig()
+        config2 = INT8StaticQuantConfig().set_local(
+            op_name_or_type, StaticQuantConfig(w_dtype=half_precision_dtype, act_dtype=half_precision_dtype)
+        )
+        tune_config = TuningConfig(config_set=[config1, config2], tolerable_loss=-0.1)
+        eval_result = [1, 1, 2]
+
+        def fake_eval_fn(model):
+            res = eval_result.pop(0)
+            return res
+
+        def run_fn(model):
+            for i in range(2):
+                model(*example_inputs)
+
+        model, example_inputs = self.build_model_include_conv_and_linear(bias)
+        model = export(model, example_inputs=example_inputs)
+        qmodel = autotune(
+            model=model, tune_config=tune_config, eval_fn=fake_eval_fn, run_fn=run_fn, example_inputs=example_inputs
+        )
+
+        # Calculate the expected number of `aten.to` operations based on bias and op_name_or_type
+        """
+        | Bias  | op_name | nn.Module |
+        |-------|---------|-----------|
+        | True  | 4       | 8         |
+        | False | 3       | 6         |
+        """
+        expected_node_occurrence = {
+            torch.ops.aten.to.dtype: (3 + int(bias)) * (1 if isinstance(op_name_or_type, str) else 2)
+        }
+
+        expected_node_occurrence = {
+            torch_test_quant_common.NodeSpec.call_function(k): v for k, v in expected_node_occurrence.items()
+        }
+        node_in_graph = self.get_node_in_graph(qmodel)
+        for node, cnt in expected_node_occurrence.items():
+            assert (
+                node_in_graph.get(node, 0) == cnt
+            ), f"Node {node} should occur {cnt} times, but {node_in_graph.get(node, 0)}"
+        # inference
+        from torch._inductor import config
+
+        config.freezing = True
+        opt_model = torch.compile(qmodel)
+        out = opt_model(*example_inputs)
+        assert out is not None

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

@@ -309,6 +309,8 @@ def test_rtn_with_quantize_API(self):
         ), "The results of calling `convert` + `prepare` and calling `quantize` should be equal."
 
     # TODO: (4, True, 32, 0), group_dim=0, format not supported
+    # TODO [SW-216127]: it's not in high priority, so we can implement it later.
+    @pytest.mark.skipif(is_hpex_available(), reason="These tests are not supported on HPU for now.")
     @pytest.mark.parametrize(
         "bits, use_sym, group_size, group_dim",
         [