[HWConfig] narrow_range parameter is introduced in hardware config (#3196)

### Changes


* Narrow range parameter is moved to hardware config
* Embedding / embedding bag nodes propagates the narrow_range attribute
of the activation quantizer up to the weight quantizer

### Reason for changes

* To extend the set of possible quantization configuration (like weights
+ symmetric + narrow_range=False) required for the some runtimes (ex:
XNNPACK)

### Related tickets



### Tests
* tests/cross_fw/test_templates/test_calculate_quantizer_parameters.py
is updated to check all possible combination of existing qconfigs with
narrow_range True/False
* tests/common/quantization/test_quantizer_propagation_solver.py is
updated to check requantization rule with narrow_range and the rule of
configuration merging with different narrow_range parameters
* tests/common/quantization/test_quantizer_propagation_graph.py is
updated to check subsequent quantizers with different narrow_range
values do not merge
* reference quantizers scales are updated to reflect the fix of
embedding weights quantization

post_training_quantization/586/ - Passed
job/weekly/job/ubuntu20_eval/245/ (+ job/ubuntu20_eval/246/) Passed
eval_tf/461/ - Passed
torch_weekly/100/ - Passed
job/nightly/job/torch_nightly/444/ - Passed
TriggerBetta/969/ - Passed

Author: daniil-lyakhov

nncf/common/hardware/config.py

@@ -20,7 +20,6 @@
 import nncf
 from nncf.common.graph.operator_metatypes import OperatorMetatype
 from nncf.common.logging import nncf_logger
-from nncf.common.quantization import quantizers as quant
 from nncf.common.quantization.structs import QuantizationScheme as QuantizationMode
 from nncf.common.quantization.structs import QuantizerConfig
 from nncf.common.utils.helpers import product_dict
@@ -158,28 +157,14 @@ def get_qconf_from_hw_config_subdict(quantization_subdict: Dict[str, Any]) -> Qu
         bits = quantization_subdict["bits"]
         mode = HWConfig.get_quantization_mode_from_config_value(quantization_subdict["mode"])
         is_per_channel = HWConfig.get_is_per_channel_from_config_value(quantization_subdict["granularity"])
-        signedness_to_force = None
-        if "level_low" in quantization_subdict and "level_high" in quantization_subdict:
-            signedness_to_force = False
-            if mode == QuantizationMode.SYMMETRIC:
-                if quantization_subdict["level_low"] < 0 < quantization_subdict["level_high"]:
-                    signedness_to_force = True
-                true_level_low, true_level_high = quant.calculate_symmetric_level_ranges(bits, signed=True)
-            else:
-                signedness_to_force = True
-                true_level_low, true_level_high = quant.calculate_asymmetric_level_ranges(bits)
-
-            assert (
-                quantization_subdict["level_low"] == true_level_low
-            ), "Invalid value of quantizer parameter `level_low`.\
-                         The parameter must be consistent with other parameters!"
-            assert (
-                quantization_subdict["level_high"] == true_level_high
-            ), "Invalid value of quantizer parameter `level_high`.\
-                         The parameter must be consistent with other parameters!"
-
+        signedness_to_force = quantization_subdict.get("signedness_to_force")
+        narrow_range = quantization_subdict["narrow_range"]
         return QuantizerConfig(
-            num_bits=bits, mode=mode, per_channel=is_per_channel, signedness_to_force=signedness_to_force
+            num_bits=bits,
+            mode=mode,
+            per_channel=is_per_channel,
+            signedness_to_force=signedness_to_force,
+            narrow_range=narrow_range,
         )
 
     @staticmethod

nncf/common/hardware/configs/cpu.json

@@ -7,15 +7,17 @@
                 "mode": [
                     "symmetric"
                 ],
-                "granularity": "pertensor"
+                "granularity": "pertensor",
+                "narrow_range": false
             },
             "q8_a": {
                 "bits": 8,
                 "mode": [
                     "symmetric",
                     "asymmetric"
                 ],
-                "granularity": "pertensor"
+                "granularity": "pertensor",
+                "narrow_range": false
             },
             "q8_a_ch": {
                 "bits": 8,
@@ -26,19 +28,28 @@
                 "granularity": [
                     "perchannel",
                     "pertensor"
-                ]
+                ],
+                "narrow_range": false
             },
             "q8_w_sym": {
                 "bits": 8,
                 "mode": "symmetric",
-                "level_low": -128,
-                "level_high": 127,
-                "granularity": ["perchannel", "pertensor"]
+                "signedness_to_force": true,
+                "granularity": ["perchannel", "pertensor"],
+                "narrow_range": true
+            },
+            "q8_w_sym_any_nr": {
+                "bits": 8,
+                "mode": "symmetric",
+                "signedness_to_force": true,
+                "granularity": ["perchannel", "pertensor"],
+                "narrow_range": [true, false]
             },
             "q8_w_asym": {
                 "bits": 8,
                 "mode": "asymmetric",
-                "granularity": ["perchannel", "pertensor"]
+                "granularity": ["perchannel", "pertensor"],
+                "narrow_range": false
             }
         }
     },
@@ -288,7 +299,7 @@
         {
             "type": "Embedding",
             "quantization": {
-                "weights": ["q8_w_sym", "q8_w_asym"]
+                "weights": ["q8_w_sym_any_nr", "q8_w_asym"]
             }
         },
         {"type": "EmbeddingBag"}

nncf/common/hardware/configs/gpu.json

@@ -8,7 +8,8 @@
                     "symmetric",
                     "asymmetric"
                 ],
-                "granularity": "pertensor"
+                "granularity": "pertensor",
+                "narrow_range": false
             },
             "q8_a_ch": {
                 "bits": 8,
@@ -19,25 +20,27 @@
                 "granularity": [
                     "perchannel",
                     "pertensor"
-                ]
+                ],
+                "narrow_range": false
             },
             "q8_w_sym": {
                 "bits": 8,
                 "mode": "symmetric",
-                "level_low": -128,
-                "level_high": 127,
+                "signedness_to_force": true,
                 "granularity":  [
                     "perchannel",
                     "pertensor"
-                ]
+                ],
+                "narrow_range": true
             },
             "q8_w_asym": {
                 "bits": 8,
                 "mode": "asymmetric",
                 "granularity":  [
                     "perchannel",
                     "pertensor"
-                ]
+                ],
+                "narrow_range": false
             }
         }
     },

nncf/common/hardware/configs/npu.json

@@ -7,15 +7,17 @@
                 "mode": [
                     "symmetric"
                 ],
-                "granularity": "pertensor"
+                "granularity": "pertensor",
+                "narrow_range": false
             },
             "q8_a": {
                 "bits": 8,
                 "mode": [
                     "symmetric",
                     "asymmetric"
                 ],
-                "granularity": "pertensor"
+                "granularity": "pertensor",
+                "narrow_range": false
             },
             "q8_a_ch": {
                 "bits": 8,
@@ -26,52 +28,66 @@
                 "granularity": [
                     "perchannel",
                     "pertensor"
-                ]
+                ],
+                "narrow_range": false
             },
             "q8_w_sym": {
                 "bits": 8,
                 "mode": "symmetric",
-                "level_low": -128,
-                "level_high": 127,
-                "granularity": ["perchannel", "pertensor"]
+                "signedness_to_force": true,
+                "granularity": ["perchannel", "pertensor"],
+                "narrow_range": true
+            },
+            "q8_w_sym_any_nr": {
+                "bits": 8,
+                "mode": "symmetric",
+                "signedness_to_force": true,
+                "granularity": ["perchannel", "pertensor"],
+                "narrow_range": [true, false]
             },
             "q8_w_asym": {
                 "bits": 8,
                 "mode": "asymmetric",
-                "granularity": ["perchannel", "pertensor"]
+                "granularity": ["perchannel", "pertensor"],
+                "narrow_range": false
             },
             // 4-bit configs
             "q4_tn": {
                 "bits": 4,
                 "mode": "symmetric",
-                "granularity": "pertensor"
+                "granularity": "pertensor",
+                "narrow_range": false
             },
             "q4_ch": {
                 "bits": 4,
                 "mode": "symmetric",
-                "granularity": "perchannel"
+                "granularity": "perchannel",
+                "narrow_range": false
             },
             "q4_w": {
                 "bits": 4,
                 "mode": "symmetric",
                 "granularity": [
                     "perchannel",
                     "pertensor"
-                ]
+                ],
+                "narrow_range": false
             },
             // 2-bit configs
             "q2_ch": {
                 "bits": 2,
                 "mode": "symmetric",
-                "granularity": "perchannel"
+                "granularity": "perchannel",
+                "narrow_range": false
             },
             "q2_w": {
                 "bits": 2,
                 "mode": "symmetric",
                 "granularity": [
                     "perchannel",
                     "pertensor"
-                ]
+                ],
+                "narrow_range": false
             }
         }
     },
@@ -382,15 +398,15 @@
             "type": "Embedding",
             "quantization": {
                 "weights": [
-                    "q8_w_sym", "q8_w_asym"
+                    "q8_w_sym_any_nr", "q8_w_asym"
                 ]
             }
         },
         {
             "type": "EmbeddingBag",
             "quantization": {
                 "weights": [
-                    "q8_w_sym", "q8_w_asym"
+                    "q8_w_sym_any_nr", "q8_w_asym"
                 ]
             }
         },

nncf/common/hardware/configs/template.json

@@ -16,14 +16,20 @@
                 "granularity": [
                     "pertensor",
                     "perchannel"
-                ]
+                ],
+                /*
+                 * Narrow range: should NNCF use 2**num_bits quants or 2**num_bits - 1
+                 */
+                "narrow_range": false
             },
-            "q8_sym_tnr_-128_127": { // Alias name for set of hyperparameters
+            "q8_sym_tnr_signed": { // Alias name for set of hyperparameters
                 "bits": 8, // Number of quantization bits
                 "mode": "symmetric", // Quantization mode
                 "granularity": "pertensor", // Granularity: one scale for output tensor
-                "level_low": -128, // Low quantization level
-                "level_high": 127  // High quantization level
+                "signedness_to_force": true, // signedness_to_force: True if the quantizer must be signed, False if must be unsigned,
+                                             // None if the signed/unsigned attribute should be determined based on the incoming activation
+                                             // statistics during range initialization.
+                "narrow_range": true
             }
         }
     },

nncf/common/quantization/config_assignment.py

@@ -47,6 +47,8 @@ def get_scoped_quantizer_config(
                 qconfig.per_channel = config_overrides["per_channel"]
             if config_overrides.get("signed") is not None:
                 qconfig.signedness_to_force = config_overrides["signed"]
+            if config_overrides.get("narrow_range") is not None:
+                qconfig.narrow_range = config_overrides["narrow_range"]
     return qconfig
 
 

nncf/common/quantization/quantizer_propagation/graph.py

@@ -1170,6 +1170,10 @@ def is_downstream_quantizer_redundant(
                 (ds_config.per_channel == us_config.per_channel)
                 or (ds_config.per_channel is True and us_config.per_channel is False)
             )
+
+            # Strictly prohibit merging of config with different narrow_range params
+            is_redundant = is_redundant and (ds_config.narrow_range == us_config.narrow_range)
+
             return is_redundant
 
         def merge_traverse_fn(

nncf/common/quantization/quantizer_propagation/solver.py

@@ -51,6 +51,7 @@
 from nncf.common.utils.debug import DEBUG_LOG_DIR
 from nncf.common.utils.debug import is_debug
 from nncf.common.utils.dot_file_rw import write_dot_graph
+from nncf.config.schemata.defaults import QUANTIZATION_NARROW_RANGE
 
 
 class TransitionStatus(Enum):
@@ -160,6 +161,7 @@ def is_final_qconfig_compatible_to_initial(initial_qconfig: QuantizerConfig) ->
                             final_qconfig.per_channel == initial_qconfig.per_channel
                             and final_qconfig.mode == initial_qconfig.mode
                             and final_qconfig.num_bits == initial_qconfig.num_bits
+                            and final_qconfig.narrow_range == initial_qconfig.narrow_range
                             and (
                                 final_qconfig.signedness_to_force == initial_qconfig.signedness_to_force
                                 or initial_qconfig.signedness_to_force is None
@@ -301,7 +303,13 @@ class QuantizerPropagationSolver:
     """
 
     DEFAULT_QUANTIZATION_TYPES = [
-        QuantizerConfig(num_bits=8, mode=QuantizationMode.SYMMETRIC, signedness_to_force=None, per_channel=False)
+        QuantizerConfig(
+            num_bits=8,
+            mode=QuantizationMode.SYMMETRIC,
+            signedness_to_force=None,
+            per_channel=False,
+            narrow_range=QUANTIZATION_NARROW_RANGE,
+        )
     ]
 
     DEFAULT_PROPAGATION_STRATEGY = QuantizerPropagationRule.MERGE_ALL_IN_ONE
@@ -1373,7 +1381,7 @@ def get_merged_qconfigs_for_downward_branching_case(
         Returns a tuple, of which the first node is the qconfig list for the quantizer to be placed
         above the branching node (i.e. that will affect all of the downward branches), and a list
         of nodes which are either None (which means that the corresponding branch quantizer has been successfully
-        merged, or qconfigs list to be set for the corresponding branch quantizer if it cannot be merged (e.g. if
+        merged), or qconfigs list to be set for the corresponding branch quantizer if it cannot be merged (e.g. if
         requantization to a lower bitwidth has to be done for this branch)
 
         :param potential_qconfigs_for_each_branch: For each branch defines the list of available configurations
@@ -1495,7 +1503,8 @@ def __disambiguate_config_list(
         """
         The input list should be sorted in descending order of priority. In case some qconfigs in the list have the
         same priority, this function will resolve the ambiguity in ordering these qconfigs in the final returned
-        list.
+        list. Quantization configs could not contain different narrow range parameters, so it does
+        not participate in __lt__ method of the QConfigComparator.
         """
 
         class QConfigComparator:

nncf/common/quantization/quantizer_setup.py

@@ -24,11 +24,16 @@
 from nncf.common.quantization.structs import UnifiedScaleType
 from nncf.common.quantization.structs import WeightQuantizerId
 from nncf.common.stateful_classes_registry import CommonStatefulClassesRegistry
+from nncf.config.schemata.defaults import QUANTIZATION_NARROW_RANGE
 
 QuantizationPointId = int
 
 DEFAULT_QUANTIZER_CONFIG = QuantizerConfig(
-    num_bits=8, mode=QuantizationMode.SYMMETRIC, signedness_to_force=None, per_channel=False
+    num_bits=8,
+    mode=QuantizationMode.SYMMETRIC,
+    signedness_to_force=None,
+    per_channel=False,
+    narrow_range=QUANTIZATION_NARROW_RANGE,
 )